Problem:
When a provider or proxy drops a streaming response mid-flight (httpcore
raises RemoteProtocolError: "incomplete chunked read", "peer closed
connection", "response ended prematurely", etc.), _generate_summary
would not classify it as a transient error. Instead of retrying on the
main model, it entered the generic 60-second cooldown, leaving context
growing unbounded until the cooldown expired. Issue #18458.
Root cause:
_is_connection_error in auxiliary_client.py did not match httpcore's
streaming premature-close error substrings. context_compressor.py's
_generate_summary except block never called _is_connection_error, so
those errors fell through to the 60-second generic cooldown rather than
triggering the retry-on-main fallback path used for timeouts.
Fix:
1. auxiliary_client.py — extend _is_connection_error keyword list with:
"incomplete chunked read", "peer closed connection",
"response ended prematurely", "unexpected eof",
"remoteprotocolerror", "localprotocolerror".
Also guard the `from openai import ...` with try/except ImportError
so the function works in environments without the openai package.
2. context_compressor.py — import _is_connection_error and call it in
_generate_summary's except block as _is_streaming_closed. Include
_is_streaming_closed in the fallback-to-main condition (alongside
_is_model_not_found, _is_timeout, _is_json_decode) and use the
shorter 30s transient cooldown for streaming-closed errors.
Tests:
4 new regression tests in TestStreamingClosedFallback:
- test_incomplete_chunked_read_falls_back_to_main
- test_peer_closed_connection_falls_back_to_main
- test_streaming_closed_on_main_uses_short_cooldown (stash-verified)
- test_non_streaming_unknown_error_still_uses_long_cooldown
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Pick openrouter/pareto-code as your model and OpenRouter auto-routes each
request to the cheapest model meeting your coding-quality bar (ranked by
Artificial Analysis). The new openrouter.min_coding_score config key (0.0-1.0,
default 0.65) tunes the floor.
- hermes_cli/models.py: add openrouter/pareto-code to OPENROUTER_MODELS so
it shows up in the picker with a description
- hermes_cli/config.py: add openrouter.min_coding_score (default 0.65 — lands
on a mid-tier coder on the current Pareto frontier)
- plugins/model-providers/openrouter: emit extra_body.plugins =
[{id: pareto-router, min_coding_score: X}] when model is openrouter/pareto-code
AND the score is a valid float in [0.0, 1.0]
- agent/transports/chat_completions.py: same emission on the legacy flag
path (when no provider profile is loaded)
- run_agent.py: openrouter_min_coding_score kwarg + storage; plumbed into
both build_kwargs() invocations and the context-summary extra_body path
- cli.py: read openrouter.min_coding_score once at init, validate float in
[0,1], pass to AIAgent constructions (CLI + background-task paths)
- cron/scheduler.py, batch_runner.py, tools/delegate_tool.py,
tui_gateway/server.py: propagate the kwarg (mirrors providers_order
plumbing — subagents inherit, cron/batch read from config)
- tests: profile-level + transport-level coverage of the model gating,
unset/empty/out-of-range handling, and the legacy flag path
- docs: new 'OpenRouter Pareto Code Router' section in providers.md
Verified end-to-end against api.openrouter.ai: at score=0.65 we land on a
mid-tier coder, at omission we get the strongest. Score is silently dropped
on any model other than openrouter/pareto-code, so it's safe to leave set.
Pass session_id through to provider profile build_api_kwargs_extras so
the OpenRouter profile can attach an xAI cache-affinity header
(x-grok-conv-id: <session-id>) for x-ai/grok-* models. xAI prompt
cache requires server affinity via this header — without it the cache
is poisoned and Grok prompt-cache hit rates drop dramatically on
multi-turn sessions.
Carve-out of #22708 by Ninso112. The original PR bundled a /diff
slash command, a zsh completion fix (already on main via #22802),
and holographic memory null-guards. This salvage keeps just the
Grok header work — small, targeted, and well-tested. Other
contributors and changes preserved for separate review.
Closes#22705.
Two co-located fixes:
1. agent/model_metadata.py: bump hy3-preview static fallback from
256000 to 262144 (256 * 1024) to match OpenRouter live metadata
so cache and offline both agree (issue #22268).
2. tests/hermes_cli/test_tencent_tokenhub_provider.py: replace the
exact-value change-detector (assert ctx == 256000) with an
invariant assertion (registered + >= 4096). Per AGENTS.md
'Don't write change-detector tests': pinning the upstream-controlled
context length is exactly the test class the rule forbids — it
breaks every time the provider bumps the published value, with
zero behavioral coverage gained.
Salvage of #22574 with a redirect on the test approach. The
contributor's diff bumped the integer and added a SECOND
change-detector pinning DEFAULT_CONTEXT_LENGTHS[hy3-preview] == 262144,
which would re-break on the next published bump. We instead delete
the change-detector entirely and assert the relationship.
Closes#22268.
`fetch_models_dev()` is on the hot path of every `AIAgent.__init__`
(via `context_compressor → get_model_context_length`). The previous
policy was "always try network first, only fall back to disk if
network fails," so every fresh `hermes chat` / `hermes gateway` /
batch / cron process paid 250-500 ms re-fetching a 2 MB JSON registry
that was already on disk from earlier runs.
Add a stage 2 between in-mem and network: if
`models_dev_cache.json` exists and its mtime is younger than the
existing `_MODELS_DEV_CACHE_TTL` (1 hour, same TTL the in-mem cache
already uses), load from disk and skip the network call.
The in-mem TTL is anchored to the disk file's age, so a 50-min-old
cache stays in-memory for only 10 more minutes — no surprise
extension of staleness window.
Invariants preserved:
- `force_refresh=True` still always hits the network and only falls
back to disk on failure (`hermes config refresh` semantics).
- Missing disk cache → fall through to network (first-ever run).
- Stale disk cache (mtime > TTL) → fall through to network.
- Negative file age (clock skew) → fall through to network.
- Network failure → existing stage-4 stale-disk fallback unchanged.
Measured impact (3-run medians, 9950X3D, fresh process per run):
fetch_models_dev cold: 256 → 17 ms (-93%)
hermes chat -q wall: 4.00 → 3.73 s (-7% median)
3.99 → 3.60 s (-10% min)
The chat-end-to-end win is bounded below by API latency variance, but
the fetch_models_dev microbenchmark is the cleanest signal: 239 ms
shaved off every fresh-process agent construction.
Win compounds with the previous perf PRs:
#22681 google_chat lazy-load
#22766 doctor parallel + IMDS off
#22790 gateway.platforms PEP 562
Tests: all 30 `tests/agent/test_models_dev.py` pass (added 4 new ones
covering the new disk-cache-first path, force_refresh override, stale
disk fallback, and missing-disk-cache fall-through). Full `tests/agent/`
suite: 2560 passed, 0 failed.
The is_xai_responses branch only sent include=[reasoning.encrypted_content]
without forwarding the resolved reasoning_effort. Other Responses providers
(OpenAI, GitHub) already get effort forwarded — this aligns the xAI path.
Without this, agent.reasoning_effort is silently dropped on the xAI direct
path, making Hermes unable to control reasoning depth on grok-4.x via
api.x.ai. Tests added to TestCodexBuildKwargs cover effort passthrough,
disabled state, and minimal-clamp parity with non-xAI.
The model regularly writes session-outcome facts to MEMORY.md despite
the existing 'Do NOT save task progress' line — entries like
'Submitted PR #22577 for the kanban dedup fix' or 'Fixed bug X in
file Y'. These are stale within days, pollute the system prompt,
and crowd out durable user preferences (the issue #22563 reporter
saw 9 sections of bug-fix notes injected on a brand-new task).
Add explicit examples of what NOT to save (PR numbers, issue
numbers, commit SHAs, 'fixed/submitted/Phase N done', file counts)
plus the 7-day-staleness heuristic so the model has a concrete
calibration target rather than guessing what counts as 'task progress'.
Closes#22563 (the prompt-side, low-risk portion). The bigger
relevance-based-injection / vector-retrieval feature requested in
#22563 is tracked under #2184 (Richer local memory). Per skill rule
on prompt caching, dynamic memory injection breaks the frozen-snapshot
invariant and needs a separate design call.
`ToolCall.extra_content` was annotated `Optional[Dict[str, Any]]`,
but neither `Optional` nor `Dict` are imported at the top of
`agent/transports/types.py` — only `Any` is. The rest of the file
consistently uses PEP 604 / 585 syntax (e.g. `str | None`,
`dict[str, Any] | None`).
The file has `from __future__ import annotations`, so the missing
names don't crash class definition. But the annotation IS evaluated
when anything calls `typing.get_type_hints(ToolCall)` —
introspection raises `NameError: name 'Optional' is not defined`.
ruff catches it cleanly:
F821 Undefined name `Optional` agent/transports/types.py:65:32
F821 Undefined name `Dict` agent/transports/types.py:65:41
Switch the annotation to `dict[str, Any] | None` to match the
rest of the file's style. No new imports needed.
Verified:
- ruff F-checks now pass on the file
- `typing.get_type_hints(ToolCall)` succeeds where it raised before
- 166/166 tests in tests/agent/transports/ pass on Windows + Python 3.12
WebUI sessions construct AIAgent(platform="webui") but PLATFORM_HINTS
had no "webui" entry, so the agent received no platform hint at all.
The WebUI frontend supports rich MEDIA:/absolute/path previews for
images, audio, video, PDF, HTML, CSV, diffs, and Excalidraw, but
without a hint the agent either ignores MEDIA: or falls back to
Markdown image syntax which silently fails for local files.
Add a webui hint that documents the MEDIA: render path and warns
against  for local files.
Fixes#21883
When an auxiliary LLM provider (or an upstream proxy) returns a non-JSON
body with `Content-Type: application/json` — e.g. an HTML 502 page from a
misconfigured gateway — the OpenAI SDK's `response.json()` raises a raw
`json.JSONDecodeError` (or wraps it in `APIResponseValidationError` whose
message contains "expecting value"). Previously this fell through to the
unknown-error branch and entered a 60s cooldown without retrying on the
main model, dropping the middle conversation turns instead.
This change folds JSON-decode detection into the existing fast-path
fallback chain: detect by `isinstance(e, JSONDecodeError)` OR substring
match for "expecting value", retry once on the main model, and use a
shorter 30s cooldown when already on main (the body shape tends to flip
back to valid quickly when the upstream proxy recovers).
The three duplicated fallback bodies (model-not-found, unknown-error,
JSON-decode) are consolidated into a single `_fallback_to_main_for_compression`
helper that handles the shared bookkeeping (record aux-model failure for
`/usage`-style callers, clear summary_model, clear cooldown).
Also adds three unit tests covering: raw `JSONDecodeError` retries on main,
substring-match for wrapped exceptions, and the 30s cooldown when already
on main.
Salvage of #22248 by @0xharryriddle. Closes#22244.
Co-authored-by: Harry Riddle <ntconguit@gmail.com>
Interactive `hermes` launch drops from ~21s to ~2.5s. Three independent
fixes, each targets a distinct hot spot in the banner / tool-registration
path that fires on every CLI invocation.
1. `get_external_skills_dirs()` in-process mtime cache (~10s saved)
The function re-read + YAML-parsed the full ~/.hermes/config.yaml on
every call. Banner build invokes it once per skill to resolve the
category column, which on a 120-skill install meant ~120 reparses of
a 15 KB config (~85 ms each). Added a
`(config_path, mtime_ns) -> list[Path]` memo; stat() is ~2 us vs
~85 ms for the parse. Edits to config.yaml invalidate the cache on
the next call via mtime.
2. Feishu availability probe uses `importlib.util.find_spec` (~5.2s saved)
`tools/feishu_doc_tool.py::_check_feishu` and the identical helper in
`feishu_drive_tool.py` were calling `import lark_oapi` purely to
detect whether the SDK was installed. Executing the real import pulls
in websockets + dispatcher + every v2 API model — ~5 seconds of work
that fires at every tool-registry bootstrap. `find_spec` answers the
same question ("is lark_oapi importable?") without executing the
module. The actual tool handlers still do the real import on invoke,
so runtime behavior is unchanged.
3. `_web_requires_env` no longer triggers Nous portal refresh (~800ms saved)
`tools/web_tools.py::_web_requires_env` used
`managed_nous_tools_enabled()` to gate four gateway env-var names in
the returned list. The gate called `get_nous_auth_status()` ->
`resolve_nous_runtime_credentials()` -> live HTTP POST to the portal
on every tool-registry bootstrap. But the list is pure metadata — if
the env var is set at runtime, the tool lights up; otherwise it
doesn't. Including the four names unconditionally is harmless for
unsubscribed users (vars just aren't set) and eliminates the sync
HTTP round trip from startup.
Test:
- tests/agent/test_external_skills_dirs_cache.py (new, 6 cases):
returns config'd dir, caches on second call (yaml_load patched to
raise — never invoked), invalidates on mtime bump, empty when config
missing, returned list is a defensive copy, per-HERMES_HOME cache key
isolation.
- Existing tests/agent/test_external_skills.py and tests/tools/
continue to pass modulo pre-existing flakes on main (test_delegate,
test_send_message — unrelated, pass in isolation).
Measured: bare `hermes` (cold → REPL ready) 21,519ms -> 2,618ms on
Teknium's install (119 skills, 15 KB config.yaml, Nous auth logged in,
lark_oapi installed). 8x faster.
## Why
Hermes supports Linux, macOS, and native Windows, but the codebase grew up
POSIX-first and has accumulated patterns that silently break (or worse,
silently kill!) on Windows:
- `os.kill(pid, 0)` as a liveness probe — on Windows this maps to
CTRL_C_EVENT and broadcasts Ctrl+C to the target's entire console
process group (bpo-14484, open since 2012).
- `os.killpg` — doesn't exist on Windows at all (AttributeError).
- `os.setsid` / `os.getuid` / `os.geteuid` — same.
- `signal.SIGKILL` / `signal.SIGHUP` / `signal.SIGUSR1` — module-attr
errors at runtime on Windows.
- `open(path)` / `open(path, "r")` without explicit encoding= — inherits
the platform default, which is cp1252/mbcs on Windows (UTF-8 on POSIX),
causing mojibake round-tripping between hosts.
- `wmic` — removed from Windows 10 21H1+.
This commit does three things:
1. Makes `psutil` a core dependency and migrates critical callsites to it.
2. Adds a grep-based CI gate (`scripts/check-windows-footguns.py`) that
blocks new instances of any of the above patterns.
3. Fixes every existing instance in the codebase so the baseline is clean.
## What changed
### 1. psutil as a core dependency (pyproject.toml)
Added `psutil>=5.9.0,<8` to core deps. psutil is the canonical
cross-platform answer for "is this PID alive" and "kill this process
tree" — its `pid_exists()` uses `OpenProcess + GetExitCodeProcess` on
Windows (NOT a signal call), and its `Process.children(recursive=True)`
+ `.kill()` combo replaces `os.killpg()` portably.
### 2. `gateway/status.py::_pid_exists`
Rewrote to call `psutil.pid_exists()` first, falling back to the
hand-rolled ctypes `OpenProcess + WaitForSingleObject` dance on Windows
(and `os.kill(pid, 0)` on POSIX) only if psutil is somehow missing —
e.g. during the scaffold phase of a fresh install before pip finishes.
### 3. `os.killpg` migration to psutil (7 callsites, 5 files)
- `tools/code_execution_tool.py`
- `tools/process_registry.py`
- `tools/tts_tool.py`
- `tools/environments/local.py` (3 sites kept as-is, suppressed with
`# windows-footgun: ok` — the pgid semantics psutil can't replicate,
and the calls are already Windows-guarded at the outer branch)
- `gateway/platforms/whatsapp.py`
### 4. `scripts/check-windows-footguns.py` (NEW, 500 lines)
Grep-based checker with 11 rules covering every Windows cross-platform
footgun we've hit so far:
1. `os.kill(pid, 0)` — the silent killer
2. `os.setsid` without guard
3. `os.killpg` (recommends psutil)
4. `os.getuid` / `os.geteuid` / `os.getgid`
5. `os.fork`
6. `signal.SIGKILL`
7. `signal.SIGHUP/SIGUSR1/SIGUSR2/SIGALRM/SIGCHLD/SIGPIPE/SIGQUIT`
8. `subprocess` shebang script invocation
9. `wmic` without `shutil.which` guard
10. Hardcoded `~/Desktop` (OneDrive trap)
11. `asyncio.add_signal_handler` without try/except
12. `open()` without `encoding=` on text mode
Features:
- Triple-quoted-docstring aware (won't flag prose inside docstrings)
- Trailing-comment aware (won't flag mentions in `# os.kill(pid, 0)` comments)
- Guard-hint aware (skips lines with `hasattr(os, ...)`,
`shutil.which(...)`, `if platform.system() != 'Windows'`, etc.)
- Inline suppression with `# windows-footgun: ok — <reason>`
- `--list` to print all rules with fixes
- `--all` / `--diff <ref>` / staged-files (default) modes
- Scans 380 files in under 2 seconds
### 5. CI integration
A GitHub Actions workflow that runs the checker on every PR and push is
staged at `/tmp/hermes-stash/windows-footguns.yml` — not included in this
commit because the GH token on the push machine lacks `workflow` scope.
A maintainer with `workflow` permissions should add it as
`.github/workflows/windows-footguns.yml` in a follow-up. Content:
```yaml
name: Windows footgun check
on:
push:
branches: [main]
pull_request:
branches: [main]
jobs:
check:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: actions/setup-python@v5
with: {python-version: "3.11"}
- run: python scripts/check-windows-footguns.py --all
```
### 6. CONTRIBUTING.md — "Cross-Platform Compatibility" expansion
Expanded from 5 to 16 rules, each with message, example, and fix.
Recommends psutil as the preferred API for PID / process-tree operations.
### 7. Baseline cleanup (91 → 0 findings)
- 14 `open()` sites → added `encoding='utf-8'` (internal logs/caches) or
`encoding='utf-8-sig'` (user-editable files that Notepad may BOM)
- 23 POSIX-only callsites in systemd helpers, pty_bridge, and plugin
tool subprocess management → annotated with
`# windows-footgun: ok — <reason>`
- 7 `os.killpg` sites → migrated to psutil (see §3 above)
## Verification
```
$ python scripts/check-windows-footguns.py --all
✓ No Windows footguns found (380 file(s) scanned).
$ python -c "from gateway.status import _pid_exists; import os
> print('self:', _pid_exists(os.getpid())); print('bogus:', _pid_exists(999999))"
self: True
bogus: False
```
Proof-of-repro that `os.kill(pid, 0)` was actually killing processes
before this fix — see commit `1cbe39914` and bpo-14484. This commit
removes the last hand-rolled ctypes path from the hot liveness-check
path and defers to the best-maintained cross-platform answer.
build_environment_hints() now emits a factual block describing the
execution environment on every prompt build:
* Local backend: host OS, $HOME, and cwd — so the agent stops guessing
paths from the hostname. Windows also gets two specific callouts:
- hostname != username (prevents C:\Users\<hostname>\... bugs)
- `terminal` shells out to bash (git-bash/MSYS), not PowerShell
* Remote backend (docker/singularity/modal/daytona/ssh/vercel_sandbox):
host info is SUPPRESSED — the agent's tools can't touch the host, so
showing it is misleading. Instead we probe the backend once per
process with `uname/whoami/pwd` and cache the result. On probe
failure, fall back to a per-backend description that states only what
we know from the backend choice itself (container type + likely OS
family) without inventing user/cwd/$HOME.
Linux/Mac local users now get a small helpful 3-line host block instead
of an empty string. Zero change to the existing WSL hint paragraph.
Tests: 8 new/updated in TestEnvironmentHints, including a regression
guard that fails if a new remote backend is added without listing it in
_REMOTE_TERMINAL_BACKENDS.
Closes the last Python-on-Windows UTF-8 exposure by making every
text-mode open() call explicit about its encoding.
Before: on Windows, bare open(path, 'r') defaults to the system
locale encoding (cp1252 on US-locale installs). That means reading
any config/yaml/markdown/json file with non-ASCII content either
crashes with UnicodeDecodeError or silently mis-decodes bytes.
After: all 89 affected call sites in production code now pass
encoding='utf-8' explicitly. Works identically on every platform
and every locale, no surprise behavior.
Mechanical sweep via:
ruff check --preview --extend-select PLW1514 --unsafe-fixes --fix --exclude 'tests,venv,.venv,node_modules,website,optional-skills, skills,tinker-atropos,plugins' .
All 89 fixes have the same shape: open(x) or open(x, mode) became
open(x, encoding='utf-8') or open(x, mode, encoding='utf-8'). Nothing
else changed. Every modified file still parses and the Windows/sandbox
test suite is still green (85 passed, 14 skipped, 0 failed across
tests/tools/test_code_execution_windows_env.py +
tests/tools/test_code_execution_modes.py + tests/tools/test_env_passthrough.py +
tests/test_hermes_bootstrap.py).
Scope notes:
- tests/ excluded: test fixtures can use locale encoding intentionally
(exercising edge cases). If we want to tighten tests later that's
a separate PR.
- plugins/ excluded: plugin-specific conventions may differ; plugin
authors own their code.
- optional-skills/ and skills/ excluded: skill scripts are user-authored
and we don't want to mass-edit them.
- website/ and tinker-atropos/ excluded: vendored / generated content.
46 files touched, 89 +/- lines (symmetric replacement). No behavior
change on POSIX or on Windows when the file is ASCII; bug fix on
Windows when the file contains non-ASCII.
Extends the cua-driver computer-use backend to drive backgrounded macOS
windows without stealing keyboard or mouse focus from the foreground app.
All changes target the cua-driver MCP backend and the shared dispatcher.
## cua_backend.py
**Window-aware capture**: capture() now calls list_windows + get_window_state
instead of the removed capture tool. Prefers structuredContent.windows
(MCP 2024-11-05+ cua-driver) for zero-parse window enumeration; falls back
to regex-parsed text for older builds. Stores the selected (pid, window_id)
as sticky context so subsequent action calls do not need a redundant round-trip.
**Action routing**: click/scroll/type_text/key all carry the sticky pid
(and window_id for element-indexed clicks). type_text routes through
type_text_chars (individual key events) rather than AX attribute write --
WebKit AXTextFields reject attribute writes from backgrounded processes.
**Key parsing**: _parse_key_combo splits cmd+s-style strings into
(key, [modifiers]) and routes to hotkey (modifier present) or
press_key (bare key) -- cua-driver actual tool names.
**set_value method**: new set_value(value, element) calls the cua-driver
set_value MCP tool. For AXPopUpButton / HTML select in a backgrounded Safari,
AXPress opens the native macOS popup which closes immediately when the app is
non-frontmost; set_value AX-presses the matching child option directly
(no menu required, no focus steal).
**focus_app**: reimplemented as a pure window-selector (enumerates
list_windows, sets sticky pid/window_id) without ever raising the window
or stealing focus.
**list_apps**: fixed tool name from listApps to list_apps; handles plain-text
response via regex when structured data is absent.
**Structured-content extraction**: _extract_tool_result now surfaces
structuredContent from MCP results, enabling the list_windows window array
without text parsing.
**Helpers**: _parse_windows_from_text, _parse_elements_from_tree,
_split_tree_text, _parse_key_combo extracted as module-level functions.
## schema.py
Added set_value to the action enum with a description explaining when to
prefer it over click (select/popup elements, sliders, no focus steal).
Added value field for set_value payloads.
## tool.py
Routed set_value action through _dispatch to backend.set_value.
Added set_value to _DESTRUCTIVE_ACTIONS (approval-gated).
Fixed MIME-type detection in _capture_response: cua-driver may return
JPEG; detect from base64 magic bytes (/9j/ -> image/jpeg, else image/png)
rather than hardcoding image/png.
## agent/display.py + run_agent.py
Guard _detect_tool_failure and result-preview logic against non-string
function_result values: multimodal tool results (dicts with _multimodal=True)
are not string-sliceable; treat them as successes and fall back to str()
for length/preview.
Background macOS desktop control via cua-driver MCP — does NOT steal the
user's cursor or keyboard focus, works with any tool-capable model.
Replaces the Anthropic-native `computer_20251124` approach from the
abandoned #4562 with a generic OpenAI function-calling schema plus SOM
(set-of-mark) captures so Claude, GPT, Gemini, and open models can all
drive the desktop via numbered element indices.
- `tools/computer_use/` package — swappable ComputerUseBackend ABC +
CuaDriverBackend (stdio MCP client to trycua/cua's cua-driver binary).
- Universal `computer_use` tool with one schema for all providers.
Actions: capture (som/vision/ax), click, double_click, right_click,
middle_click, drag, scroll, type, key, wait, list_apps, focus_app.
- Multimodal tool-result envelope (`_multimodal=True`, OpenAI-style
`content: [text, image_url]` parts) that flows through
handle_function_call into the tool message. Anthropic adapter converts
into native `tool_result` image blocks; OpenAI-compatible providers
get the parts list directly.
- Image eviction in convert_messages_to_anthropic: only the 3 most
recent screenshots carry real image data; older ones become text
placeholders to cap per-turn token cost.
- Context compressor image pruning: old multimodal tool results have
their image parts stripped instead of being skipped.
- Image-aware token estimation: each image counts as a flat 1500 tokens
instead of its base64 char length (~1MB would have registered as
~250K tokens before).
- COMPUTER_USE_GUIDANCE system-prompt block — injected when the toolset
is active.
- Session DB persistence strips base64 from multimodal tool messages.
- Trajectory saver normalises multimodal messages to text-only.
- `hermes tools` post-setup installs cua-driver via the upstream script
and prints permission-grant instructions.
- CLI approval callback wired so destructive computer_use actions go
through the same prompt_toolkit approval dialog as terminal commands.
- Hard safety guards at the tool level: blocked type patterns
(curl|bash, sudo rm -rf, fork bomb), blocked key combos (empty trash,
force delete, lock screen, log out).
- Skill `apple/macos-computer-use/SKILL.md` — universal (model-agnostic)
workflow guide.
- Docs: `user-guide/features/computer-use.md` plus reference catalog
entries.
44 new tests in tests/tools/test_computer_use.py covering schema
shape (universal, not Anthropic-native), dispatch routing, safety
guards, multimodal envelope, Anthropic adapter conversion, screenshot
eviction, context compressor pruning, image-aware token estimation,
run_agent helpers, and universality guarantees.
469/469 pass across tests/tools/test_computer_use.py + the affected
agent/ test suites.
- `model_tools.py` provider-gating: the tool is available to every
provider. Providers without multi-part tool message support will see
text-only tool results (graceful degradation via `text_summary`).
- Anthropic server-side `clear_tool_uses_20250919` — deferred;
client-side eviction + compressor pruning cover the same cost ceiling
without a beta header.
- macOS only. cua-driver uses private SkyLight SPIs
(SLEventPostToPid, SLPSPostEventRecordTo,
_AXObserverAddNotificationAndCheckRemote) that can break on any macOS
update. Pin with HERMES_CUA_DRIVER_VERSION.
- Requires Accessibility + Screen Recording permissions — the post-setup
prints the Settings path.
Supersedes PR #4562 (pyautogui/Quartz foreground backend, Anthropic-
native schema). Credit @0xbyt4 for the original #3816 groundwork whose
context/eviction/token design is preserved here in generic form.
The previous revision of this PR added six GMI-specific branches
(`elif base_url_host_matches(..., 'api.gmi-serving.com')`) across
run_agent.py and agent/auxiliary_client.py, plus a _HERMES_UA_HEADERS
constant in auxiliary_client.py.
ProviderProfile already has a `default_headers: dict[str, str]` field
commented as 'Client-level quirks (set once at client construction)'.
Other plugins (ai-gateway, kimi-coding) already use it. Two of the four
auxiliary_client sites we previously patched already had a generic
`else: profile.default_headers` fallback that picked it up (so did
both run_agent sites).
This revision:
* Sets `default_headers={'User-Agent': 'HermesAgent/<ver>'}` on the
GMI profile in plugins/model-providers/gmi/__init__.py.
* Reverts all six GMI-specific branches in run_agent.py and
auxiliary_client.py.
* Adds the generic profile-fallback `else` block to the two
auxiliary_client sites (`_to_async_client`, `resolve_provider_client`)
that didn't have it yet. This benefits every provider whose profile
declares default_headers, not just GMI — e.g. Vercel AI Gateway's
HTTP-Referer/X-Title now flow through the async client path too.
* Replaces the GMI-specific URL-branch tests with a profile-level
assertion and keeps the run_agent integration test (with
`provider='gmi'` so the fallback picks up the profile).
Net diff vs main: +82/-0 across 5 files, touching only the GMI plugin,
two generic fallback blocks in auxiliary_client.py, AUTHOR_MAP, and
tests. No core files change.
Based on #20907 by @isaachuangGMICLOUD.
- Add pricing entries for Claude Opus 4.5/4.6/4.7, Sonnet 4.5/4.6, and
Haiku 4.5 with updated source URLs (platform.claude.com)
- Add _normalize_anthropic_model_name() to handle dot-notation variants
(e.g. claude-opus-4.7 → claude-opus-4-7) for pricing lookups
- Fix silent token loss: ensure session row exists before UPDATE in both
run_agent.py and hermes_state.py (INSERT OR IGNORE is idempotent)
- Log token persistence failures at DEBUG level instead of swallowing
them silently — makes undercounted analytics diagnosable
- Surface reasoning tokens in CLI /usage and TUI usage panel
- Add 'reasoning' and 'cost_status' fields to TUI Usage type
## Summary
- Forwards chat-completions `timeout` into the Codex Responses stream call.
- Adds total elapsed-time enforcement while the Responses stream is still yielding events.
- Closes the underlying client on timeout to unblock stalled streams, then raises `TimeoutError`.
- Adds focused tests for timeout forwarding and total timeout enforcement.
## Why
The Codex auxiliary adapter can be used by non-interactive auxiliary work such as context compression. If the stream keeps yielding progress-like events but never completes, SDK socket/read timeouts do not necessarily protect the full operation. This makes the CLI look stuck until the user force-interrupts the whole session.
This is a refreshed upstream-ready version of the earlier fork fix around `d3f08e9a0` / PR #3.
## Verification
- `python -m py_compile agent/auxiliary_client.py tests/agent/test_auxiliary_client.py`
- `python -m pytest -o addopts='' tests/agent/test_auxiliary_client.py::TestCodexAuxiliaryAdapterTimeout -q`
- `git diff --check`
Z.AI (智谱 GLM) vision models (glm-4v-flash, glm-4v-plus, etc.) have two
compatibility issues when used through the Anthropic-compatible endpoint:
1. **Error 1210 — max_tokens rejected on multimodal calls**: Z.AI rejects
the max_tokens parameter for vision model requests with error code 1210
("API 调用参数有误"). The error string does not contain "max_tokens",
so the existing unsupported-parameter retry logic never fires.
2. **Wrong endpoint inheritance**: When the main runtime provider uses Z.AI's
Anthropic-compatible endpoint (open.bigmodel.cn/api/anthropic), the vision
client inherits this endpoint. But Z.AI's Anthropic wire cannot properly
handle image content — models silently fail ("I can't see the image") or
reject max_tokens.
Changes:
- resolve_vision_provider_client(): force Z.AI vision to use OpenAI-compatible
endpoint (open.bigmodel.cn/api/paas/v4) instead of inheriting Anthropic wire
- _build_call_kwargs(): skip max_tokens for Z.AI vision models (4v/5v/-v suffix)
- _AnthropicCompletionsAdapter: support _skip_zai_max_tokens flag
- _to_openai_base_url(): rewrite Z.AI Anthropic URLs to OpenAI-compatible path
- call_llm() retry: detect Z.AI error 1210 and strip max_tokens before retry
Discord (and similar platforms) can serve a PNG image cached as
discord_xxx.webp because the CDN reports content_type=image/webp for
proxied stickers, custom emoji, and certain bot-uploaded images even
when the actual bytes are PNG. Hermes' agent.image_routing._guess_mime
trusted the file suffix and declared media_type=image/webp to
Anthropic, which strict-validates and returns:
HTTP 400 messages.N.content.M.image.source.base64:
The image was specified using the image/webp media type,
but the image appears to be a image/png image
The Discord image attachment never reaches the model; the whole turn
fails with no salvage path.
Fix: sniff magic bytes in _file_to_data_url before declaring MIME.
Suffix-based detection is kept as a fallback when bytes aren't
available. New helper _sniff_mime_from_bytes covers PNG, JPEG, GIF,
WEBP, BMP, and HEIC/HEIF.
Tests:
- Two existing tests asserted the old broken behaviour (PNG bytes in
a .jpg/.webp file should report jpeg/webp); rewritten with real
jpeg/webp magic bytes so they still cover suffix-aligned cases.
- New regression test test_mime_sniff_overrides_misleading_extension
reproduces the exact Discord scenario (PNG bytes, .webp suffix) and
asserts the data URL comes back as image/png.
All 28 tests in tests/agent/test_image_routing.py pass.
When multiple custom_providers share the same base_url but have different API keys,
get_custom_provider_pool_key() always returned the first match, causing wrong-key
unauthorized errors. Add provider_name parameter to prefer exact name matches
over base_url-only matching, with fallback for backward compatibility.
Fixes#19083
Flip the default for HERMES_REDACT_SECRETS from off to on so the redactor
already wired into send_message_tool, logs, and tool output actually runs
on a fresh install.
- agent/redact.py: env-var default "" → "true"
- hermes_cli/config.py: DEFAULT_CONFIG security.redact_secrets True;
two config-template comments rewritten
- gateway/run.py + cli.py: startup log / banner warning when the user
has explicitly opted out, so the downgrade is visible in agent.log
and at CLI banner time
- docs/reference/environment-variables.md: description reconciled
- tests: flipped the default-pin, restructured the force=True
regression test to explicit-false instead of unset
Users who need raw credential values (redactor development) can still
opt out via security.redact_secrets: false in config.yaml or
HERMES_REDACT_SECRETS=false in .env.
Closes#17691.
Addresses #20785 (short-term output-pipeline recommendation).
Widen PR #20314's fix to the other timeout-polling sites in the codebase
that share the same wall-clock-jump bug class. All of these measure elapsed
timeout duration, not civil time, so they belong on time.monotonic().
- hermes_cli/auth.py: auth-store file-lock timeout, Spotify OAuth callback
wait, Nous portal device-auth token poll.
- hermes_cli/copilot_auth.py: Copilot OAuth device-flow token poll.
- hermes_cli/gateway.py: gateway systemd restart wait.
- hermes_cli/web_server.py: dashboard Codex device-auth user_code wait,
dashboard Nous device-auth token poll. (sess["expires_at"] stays on
time.time() — it's a persisted absolute timestamp, not a local
deadline-polling variable.)
- agent/copilot_acp_client.py: Copilot ACP JSON-RPC request timeout.
In native image mode (vision-capable models like gpt-4o, claude-sonnet-4),
build_native_content_parts() previously emitted only the user's caption
plus image_url parts. The local file path of each attached image never
appeared in the conversation text, so the model could see the pixels but
had no string handle for tools that take image_url: str (custom MCP
tools, vision_analyze on a re-look, attach-to-tracker workflows).
The text-mode path already injects an equivalent hint via
Runner._enrich_message_with_vision ("...vision_analyze using image_url:
<path>..."). This brings native mode to parity by appending one
"[Image attached at: <path>]" line per successfully attached image to
the user-text part of the multimodal turn. Skipped (unreadable) paths
are NOT advertised, so the model is never told a non-existent file is
attached.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Fix /compact → /compress in context-overflow tips (closes#20020)
- Evict cached agent after session hygiene and /compress so system
prompt refreshes with current SOUL.md, memory, and skills
- Restore memory authority across compaction: change 'informational
background data' to 'authoritative reference data' in memory block
and SUMMARY_PREFIX, with backward-compatible regex
Based on:
- PR #20027 by @LeonSGP43
- PR #18767 by @MacroAnarchy
- PR #17380 by @vominh1919
PR #17121 boundary marker fix already merged to main (2eef395e1).
PR #9262 user-message anchoring already on main via _ensure_last_user_message_in_tail().
- Add locales/tr.yaml with Turkish translations for all approval.* and gateway.* keys
- Register 'tr' in SUPPORTED_LANGUAGES
- Add Turkish aliases: turkish, türkçe, tr-tr
- Add fr.yaml with French translations for approval prompts and gateway messages
- Register 'fr' in SUPPORTED_LANGUAGES
- Add French aliases: french, français, fr-fr, fr-be, fr-ca, fr-ch
- Update locale sync comment in en.yaml
Introduces providers/ package — single source of truth for every
inference provider. Adding a simple api-key provider now requires one
providers/<name>.py file with zero edits anywhere else.
What this PR ships:
- providers/ package (ProviderProfile ABC + 33 profiles across 4 api_modes)
- ProviderProfile declarative fields: name, api_mode, aliases, display_name,
env_vars, base_url, models_url, auth_type, fallback_models, hostname,
default_headers, fixed_temperature, default_max_tokens, default_aux_model
- 4 overridable hooks: prepare_messages, build_extra_body,
build_api_kwargs_extras, fetch_models
- chat_completions.build_kwargs: profile path via _build_kwargs_from_profile,
legacy flag path retained for lmstudio/tencent-tokenhub (which have
session-aware reasoning probing that doesn't map cleanly to hooks yet)
- run_agent.py: profile path for all registered providers; legacy path
variable scoping fixed (all flags defined before branching)
- Auto-wires: auth.PROVIDER_REGISTRY, models.CANONICAL_PROVIDERS,
doctor health checks, config.OPTIONAL_ENV_VARS, model_metadata._URL_TO_PROVIDER
- GeminiProfile: thinking_config translation (native + openai-compat nested)
- New tests/providers/ (79 tests covering profile declarations, transport
parity, hook overrides, e2e kwargs assembly)
Deltas vs original PR (salvaged onto current main):
- Added profiles: alibaba-coding-plan, azure-foundry, minimax-oauth
(were added to main since original PR)
- Skipped profiles: lmstudio, tencent-tokenhub stay on legacy path (their
reasoning_effort probing has no clean hook equivalent yet)
- Removed lmstudio alias from custom profile (it's a separate provider now)
- Skipped openrouter/custom from PROVIDER_REGISTRY auto-extension
(resolve_provider special-cases them; adding breaks runtime resolution)
- runtime_provider: profile.api_mode only as fallback when URL detection
finds nothing (was breaking minimax /v1 override)
- Preserved main's legacy-path improvements: deepseek reasoning_content
preserve, gemini Gemma skip, OpenRouter response caching, Anthropic 1M
beta recovery, etc.
- Kept agent/copilot_acp_client.py in place (rejected PR's relocation —
main has 7 fixes landed since; relocation would revert them)
- _API_KEY_PROVIDER_AUX_MODELS alias kept for backward compat with existing
test imports
Co-authored-by: kshitijk4poor <82637225+kshitijk4poor@users.noreply.github.com>
Closes#14418
The BuiltinMemoryProvider class was removed from the codebase but its
name lingered in the module-level docstrings of memory_manager.py and
memory_provider.py, creating false expectations:
- memory_manager.py docstring showed example code doing
add_provider(BuiltinMemoryProvider(...)) which ImportError at runtime
- memory_provider.py docstring listed BuiltinMemoryProvider as
'always present, not removable' — misleading for new contributors
The regression test (test_memory_user_id.py) already passes without
any reference to BuiltinMemoryProvider; it uses RecordingProvider
instances directly. The stale references were docs-only drift.
Update both docstrings to reflect the actual current architecture:
MemoryManager accepts external plugin providers only (one at a time).
Closes#14402
When a provider returns a 429 rate-limit error (not billing-related),
the auxiliary client's call_llm/async_call_llm previously did NOT trigger
the fallback chain. This caused auxiliary tasks like session_search to
exhaust all 3 retries against the same rate-limited endpoint, losing
session metadata that depended on the summarization completing.
Root cause: `_is_payment_error()` only matched 429s containing billing
keywords ("credits", "insufficient funds", etc.). Provider-specific
rate-limit messages like Nous's "Hold up for a bit, you've exceeded the
rate limit on your API key" didn't match, so `_is_payment_error` returned
False, `_is_connection_error` returned False, and `should_fallback` was
False — all retries hit the same rate-limited provider.
Fix:
- New `_is_rate_limit_error()` function that detects 429 + rate-limit
keywords, generic 429 without billing keywords, and OpenAI SDK
`RateLimitError` class instances (which may omit .status_code).
- Updated `should_fallback` in both `call_llm` and `async_call_llm` to
include `_is_rate_limit_error`.
- Updated the max_tokens retry path to also check for rate-limit errors.
- Updated the reason string to include "rate limit".
This complements the Nous rate guard (PR #10568) which prevents new calls
to Nous when already rate-limited — this fix handles the case where a
request is already in flight when the 429 arrives.
Related: #8023, #12554, #11034
Co-authored-by: Zeejay <zjtan1@gmail.com>
OpenRouter's dashboard attributes usage via the `X-Title` header.
Hermes was sending `X-OpenRouter-Title`, which OpenRouter does not
recognize, so Hermes usage showed up unlabeled. Rename to `X-Title`
to match the canonical header (already used elsewhere in the same
file via _AI_GATEWAY_HEADERS).
Salvages the core fix from @JTroyerOvermatch's PR #13649. Dropped the
PR's `HERMES_OPENROUTER_TITLE` / `HERMES_OPENROUTER_REFERER` env-var
override plumbing per the '.env is for secrets only' policy — if
per-deployment attribution is needed later it should go under
`openrouter.title` / `openrouter.referer` in config.yaml instead.
* revert(gateway): remove stale-code self-check and auto-restart
Removes the _detect_stale_code / _trigger_stale_code_restart mechanism
introduced in #17648 and iterated in #19740. On every incoming message
the gateway compared the boot-time git HEAD SHA to the current SHA on
disk, and if they differed it would reply with
Gateway code was updated in the background --
restarting this gateway so your next message runs
on the new code. Please retry in a moment.
and then kick off a graceful restart. This is unwanted behaviour:
users who run a long-lived gateway and do their own ad-hoc git
operations on the checkout end up with their chat interrupted and
the current message dropped every time HEAD moves, with no way to
opt out.
If an operator really needs the old protection against stale
sys.modules after "hermes update", the SIGKILL-survivor sweep in
hermes update (hermes_cli/main.py, also tagged #17648) already
handles the supervisor-respawn case on its own.
Removed:
gateway/run.py:
- _STALE_CODE_SENTINELS, _GIT_SHA_CACHE_TTL_SECS
- _read_git_head_sha(), _compute_repo_mtime() module helpers
- class-level _boot_wall_time / _boot_repo_mtime / _boot_git_sha /
_stale_code_restart_triggered defaults
- __init__ boot-snapshot block (_boot_*, _cached_current_sha*,
_repo_root_for_staleness, _stale_code_notified)
- _current_git_sha_cached(), _detect_stale_code(),
_trigger_stale_code_restart() methods
- stale-code check + user-facing restart notice at the top of
_handle_message()
tests/gateway/test_stale_code_self_check.py (deleted, 412 lines)
No new logic added. Zero remaining references to any removed
symbol. Gateway test suite passes the same 4589 tests it passed
before; the 3 pre-existing unrelated failures (discord free-channel,
feishu bot admission, teams typing) are unchanged by this commit.
* feat(i18n): add display.language for static message translation (zh/ja/de/es)
Adds a thin-slice i18n layer covering the highest-impact static user-facing
messages: the CLI dangerous-command approval prompt and a handful of gateway
slash-command replies (restart-drain, goal cleared, approval expired, config
read/save errors).
Out of scope (stays English): agent responses, log lines, tool outputs,
slash-command descriptions, error tracebacks.
Infrastructure:
- agent/i18n.py: catalog loader, t() helper, language resolution
(HERMES_LANGUAGE env var > display.language config > en)
- locales/{en,zh,ja,de,es}.yaml: ~19 translated strings per language
- display.language in DEFAULT_CONFIG (hermes_cli/config.py)
Tests:
- tests/agent/test_i18n.py: 21 tests covering catalog parity, placeholder
parity across locales, fallback behavior, env-var override, alias
normalization, missing-key graceful degradation.
Docs:
- website/docs/user-guide/configuration.md: display.language entry plus a
short section explaining scope so users don't expect agent responses to
translate via this knob.
When auxiliary.<task> config has base_url set but api_key is empty
(common when user expects env var fallback), _resolve_task_provider_model()
returned provider="custom" with api_key=None. This caused downstream
client construction to make API calls without an Authorization header,
resulting in HTTP 401 errors.
Fix: only return "custom" when BOTH cfg_base_url AND cfg_api_key are
non-empty. When base_url is set without api_key but with a known
provider (e.g. "openrouter"), pass through to that provider so it can
resolve credentials from environment variables.
Fixes#16829
auxiliary.<task>.extra_body.reasoning, but the new translation path in
_CodexCompletionsAdapter.create() reads the effort with
``reasoning_cfg.get("effort", "medium")``. That returns the configured
value verbatim when the key is present, so ``effort: null`` /
``effort: ""`` (both common YAML shapes) flow through as
``{"effort": null, "summary": "auto"}`` and Codex rejects the request
with "Invalid value for parameter ``reasoning.effort``".
agent/transports/codex.py::build_kwargs() — which the new adapter is
documented to mirror — uses a truthy check (``elif
reasoning_config.get("effort"):``) so the same falsy values keep the
"medium" default. Switch the auxiliary adapter to the same
``or "medium"`` truthy form so identical config produces identical
requests on both paths.
- [x] Two new regression tests cover ``effort: None`` and
``effort: ""`` and assert the request goes out as
``{"effort": "medium", "summary": "auto"}``.
- [x] Old behaviour fails the new tests (``{'effort': None} !=
{'effort': 'medium'}``); fixed behaviour passes all 11 tests in the
``TestCodexAdapterReasoningTranslation`` class.
- [x] Adjacent suites green: ``tests/agent/test_auxiliary_client.py``
(108 passed) and ``tests/agent/transports/test_codex_transport.py +
test_chat_completions.py`` (73 passed).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The API server is a documented, first-class messaging platform with its own
gateway adapter, docs pages, and toolset. But it's the only messaging
platform missing from PLATFORM_HINTS in agent/prompt_builder.py.
Without a platform hint, the agent has no context about the API server's
rendering environment and defaults to markdown-heavy document-style outputs
(code fences, bold, bullet points) — which break on the plain-text frontends
most API server consumers wrap (Open WebUI, custom agents, third-party
bridges).
Adds a generic api_server entry that describes the medium (unknown rendering,
assume plain text) without encoding any specific use case. Individual consumers
can layer additional style guidance via ephemeral system prompts.
Before (DeepSeek V4 Pro via API server, no hint):
**Sendblue bridge** at /opt/sendblue-bridge - **68MB** on disk
After (same prompt, with hint):
Sendblue bridge at /opt/sendblue-bridge, 68MB on disk
No breaking changes — new dict entry only. Existing API server consumers see
no behavioral change except for models that previously defaulted to markdown
formatting, which now produce cleaner plain-text output.
When the head ends with assistant/tool and the tail starts with assistant,
the summary is inserted as a standalone role="user" message. The body's
verbatim "## Active Task" quote then gets read as fresh user input by
weak/local models (#11475, #14521).
The merge-into-tail path already appends an explicit end-of-summary marker
for this reason. Mirror it on the standalone path so both insertion routes
give the model the same "summary above, not new input" signal.
* revert(gateway): remove stale-code self-check and auto-restart
Removes the _detect_stale_code / _trigger_stale_code_restart mechanism
introduced in #17648 and iterated in #19740. On every incoming message
the gateway compared the boot-time git HEAD SHA to the current SHA on
disk, and if they differed it would reply with
Gateway code was updated in the background --
restarting this gateway so your next message runs
on the new code. Please retry in a moment.
and then kick off a graceful restart. This is unwanted behaviour:
users who run a long-lived gateway and do their own ad-hoc git
operations on the checkout end up with their chat interrupted and
the current message dropped every time HEAD moves, with no way to
opt out.
If an operator really needs the old protection against stale
sys.modules after "hermes update", the SIGKILL-survivor sweep in
hermes update (hermes_cli/main.py, also tagged #17648) already
handles the supervisor-respawn case on its own.
Removed:
gateway/run.py:
- _STALE_CODE_SENTINELS, _GIT_SHA_CACHE_TTL_SECS
- _read_git_head_sha(), _compute_repo_mtime() module helpers
- class-level _boot_wall_time / _boot_repo_mtime / _boot_git_sha /
_stale_code_restart_triggered defaults
- __init__ boot-snapshot block (_boot_*, _cached_current_sha*,
_repo_root_for_staleness, _stale_code_notified)
- _current_git_sha_cached(), _detect_stale_code(),
_trigger_stale_code_restart() methods
- stale-code check + user-facing restart notice at the top of
_handle_message()
tests/gateway/test_stale_code_self_check.py (deleted, 412 lines)
No new logic added. Zero remaining references to any removed
symbol. Gateway test suite passes the same 4589 tests it passed
before; the 3 pre-existing unrelated failures (discord free-channel,
feishu bot admission, teams typing) are unchanged by this commit.
* fix(agent): stateful streaming scrubber for reasoning-block leaks (#17924)
Per-delta _strip_think_blocks ran at _fire_stream_delta and destroyed
downstream state. When MiniMax-M2.7 / DeepSeek / Qwen3 streamed a tag
split across deltas (delta1='<think>', delta2='Let me check'), the
regex case-2 match erased delta1 entirely, so CLI/gateway state
machines never learned a block was open and leaked delta2 as content.
Raw consumers (ACP, api_server, TTS) had no downstream defense at all.
Replace the per-delta regex with a stateful StreamingThinkScrubber
that survives delta boundaries:
- Closed <tag>X</tag> pairs always stripped (matches _strip_think_blocks
case 1).
- Unterminated open at block boundary enters a block; content
discarded until close tag arrives. At end-of-stream, held
content is dropped.
- Orphan close tags stripped without boundary gating.
- Partial tags at delta boundaries held back until resolved.
- Block-boundary rule (start-of-stream, after \n, or
whitespace-only since last \n) preserves prose that mentions
tag names.
Reset at turn start alongside the existing context scrubber; flush at
turn end so a benign '<' held back at end-of-stream reaches the UI.
E2E-verified on live OpenRouter->MiniMax-m2 streams: closed pairs
strip cleanly, first word of post-block content is preserved, pure
content passes through unchanged. Stefan's screenshot case (#17924)
— 'Let me check' getting chopped to ' me check' — no longer happens.
Final _strip_think_blocks calls on completed strings (final_response,
replay, compression) are preserved; only the streaming per-delta call
site switched to the scrubber.
MCP servers commonly emit JSON Schema `pattern` (e.g. `\\d{4}-\\d{2}-\\d{2}`
for date-time params) and `format` keywords. llama.cpp's
`json-schema-to-grammar` converter rejects regex escape classes
(\\d/\\w/\\s) and most format values, returning HTTP 400
"parse: error parsing grammar: unknown escape at \\d" — the whole request
fails.
Cloud providers (OpenAI, Anthropic, OpenRouter, Gemini) accept these
keywords fine and use them as prompting hints. Stripping unconditionally
loses useful hints for every cloud user to fix a llama.cpp-only bug.
Approach: classify the llama.cpp grammar-parse 400 in the error
classifier, and on match do a one-shot in-place strip of pattern/format
from `self.tools`, then retry. Follows the existing
`thinking_signature` recovery pattern. Cloud users hit zero overhead;
llama.cpp users pay one failed request per session.
Changes
- agent/error_classifier.py: new `FailoverReason.llama_cpp_grammar_pattern`
+ narrow HTTP-400 branch matching "error parsing grammar",
"json-schema-to-grammar", or "unable to generate parser ... template".
- tools/schema_sanitizer.py: new `strip_pattern_and_format()` helper —
reactive, walks schema nodes, skips property names (search_files.pattern
survives). Returns strip count for logging.
- run_agent.py: new one-shot recovery block in the retry loop. Strips,
logs, continues. Falls through to normal retry if nothing to strip.
- tests: 4 classifier tests (3 variants + 1 non-400 negative), 7 strip
tests including the property-name preservation and idempotency checks.
Co-authored-by: Chris Danis <cdanis@gmail.com>
Per https://platform.claude.com/docs/en/build-with-claude/fast-mode:
"Fast mode is currently supported on Opus 4.6 only. Sending speed: fast
with an unsupported model returns an error."
Pre-fix, _is_anthropic_fast_model() returned True for any claude-* model,
so /fast on Opus 4.7 (or Sonnet/Haiku) would persist agent.service_tier=fast
in config.yaml and the adapter would inject extra_body["speed"] = "fast"
on every subsequent request. Opus 4.7 returns:
HTTP 400: 'claude-opus-4-7' does not support the `speed` parameter.
This wedged sessions across model upgrades (a user who ran /fast on Opus 4.6
and later switched the default model to 4.7 hit a hard 400 on every turn
until they manually edited config.yaml).
Changes:
- _is_anthropic_fast_model: gate on "opus-4-6" / "opus-4.6" only
- anthropic_adapter: add _supports_fast_mode predicate as defensive guard
so stale request_overrides on an unsupported model are dropped silently
instead of 400'ing
- Tests: flip the assertions that mirrored the bug (Sonnet/Haiku/Opus 4.7
asserting fast-mode support) to match the documented API contract
Commit 408dd8aa added a non-string guard for Pass 1 (dedup), but the same
pattern exists in Pass 2 (summarization/pruning) where content.startswith()
and len() are called on potentially non-string tool content.
When a provider returns tool results with non-string content (e.g. dict or
int from llama.cpp or similar), the pruning pass crashes with AttributeError.
Add the same isinstance(content, str) guard to Pass 2 for consistency.
Keep the configured vision provider when base_url is overridden so credential-pool lookup still resolves provider-specific API keys (e.g. ZAI_API_KEY), and add a regression test for this path.
Generic 400 and server-disconnect heuristics used absolute token/message-count fallbacks that are too aggressive for 1M context sessions. Gate those absolute fallbacks to smaller context windows while preserving relative pressure checks.
Fixes#16351
ENV-assignment and JSON-field regex patterns in redact_sensitive_text()
cause false positives when reading source code files:
- MAX_TOKENS=*** triggers the ENV assignment pattern
- "apiKey": "test" in test fixtures triggers the JSON field pattern
Add code_file=False parameter. When code_file=True, skip only the
ENV-assignment and JSON-field regex passes; all other patterns (prefixes,
auth headers, private keys, DB connstrings, JWTs, URL secrets) are
still applied.
Update file_tools.py (read_file and search_files) to pass code_file=True
so agent code analysis is not polluted by false-positive redactions.
Closes#15934
Extends the existing _normalize_tool_input_schema to also drop top-level
union keywords that Anthropic's tool schema validator rejects with HTTP 400.
Several upstream and plugin tools ship schemas with a top-level oneOf/
allOf/anyOf (common for Pydantic discriminated unions). The existing
strip_nullable_unions pass only handles anyOf-with-null patterns; a
non-null top-level union keyword sails through and hits the API.
Salvage of #16471 — approach folded into the existing normalize helper
rather than introducing a parallel _sanitize_input_schema function, to
avoid two schema-munging code paths running against the same input.
Co-authored-by: Grey0202 <grey0202@users.noreply.github.com>
Previously only HTTP 404/503 and specific error strings triggered a fallback
to the main model when the summary model was unavailable. Timeout errors
(HTTP 408/429/502/504, or error strings containing 'timeout') entered a
short cooldown instead, leaving context to grow unbounded for the rest of
the session.
Add _is_timeout detection alongside _is_model_not_found so that transient
timeout errors on the summary model also trigger immediate fallback to the
main model, preventing compression failure from cascading.
Closes#15935
DashScope's Anthropic-compatible endpoint enforces max_tokens ∈ [1, 65536].
Adding "qwen3" to _ANTHROPIC_OUTPUT_LIMITS prevents 400 errors that were
misclassified as context overflow, triggering premature compression.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
on_session_reset() cleared _previous_summary, _last_summary_error, and
_ineffective_compression_count but left _summary_failure_cooldown_until
intact. When a transient summary error sets a 60 s cooldown (or 600 s
for a missing-provider RuntimeError) and the user immediately runs /reset
or /new, the cooldown carries into the new session. If the new session
reaches the compression threshold before the cooldown expires,
_generate_summary() returns None early, middle turns are silently dropped
without a summary, and the agent continues with no indication that
compaction was skipped.
Fix: set _summary_failure_cooldown_until = 0.0 in on_session_reset(),
matching the value assigned in __init__ and symmetric with the other
per-session fields already cleared there.
Fixes#15547
_classify_removed_skills used naive 'in' substring matching to detect
whether a removed skill's name appeared in skill_manage arguments.
Short/common skill names (api, git, test, foo, etc.) matched
incorrectly when they appeared as substrings of longer words in file
paths (references/api-design.md) or content (latest, testing).
Replace with field-aware matching:
- file_path: needle must match a complete filename stem or directory
name, with -/_ normalised for variant tolerance
- content fields: word-boundary regex (\b) prevents embedding in
longer words
Also add 3 regression tests covering the false-positive scenarios.
_try_anthropic() lacked the explicit_api_key parameter added to
_try_openrouter() in #18768. When resolve_provider_client() is called
with provider="anthropic" and an explicit key (e.g. from a fallback_model
entry with api_key set), the key was silently ignored — _try_anthropic()
always fell back to resolve_anthropic_token(), so the fallback returned
None,None for users without a default Anthropic credential configured.
Fix: add explicit_api_key: str = None to _try_anthropic() and use
explicit_api_key or <pool/env fallback> in both the pool-present and
no-pool paths. Pass explicit_api_key=explicit_api_key at the call site
in resolve_provider_client(). Symmetric with the _try_openrouter() fix.
No behavior change when explicit_api_key is None.
KANBAN_GUIDANCE layer 3 of the system prompt started with 'You are a
Kanban worker', overriding the profile's SOUL.md identity at layer 1.
Profiles with strict role boundaries (e.g. a reviewer profile that
never writes code) still executed implementation tasks because the
kanban identity claim diluted SOUL's.
Drop the identity line. Layer 3 now describes the task-execution
protocol only; SOUL.md remains the sole identity slot.
Fixes#19351
Curator review fork now forwards per-slot credentials from auxiliary.curator
and legacy curator.auxiliary to resolve_runtime_provider, matching the
canonical aux task schema. Add regression tests for binding and main fallback.
Enable OpenRouter's response caching feature (beta) via X-OpenRouter-Cache
headers. When enabled, identical API requests return cached responses for
free (zero billing), reducing both latency and cost.
Configuration via config.yaml:
openrouter:
response_cache: true # default: on
response_cache_ttl: 300 # 1-86400 seconds
Changes:
- Add openrouter config section to DEFAULT_CONFIG (response_cache + TTL)
- Add build_or_headers() in auxiliary_client.py that builds attribution
headers plus optional cache headers based on config
- Replace inline _OR_HEADERS dicts with build_or_headers() at all 5 sites:
run_agent.py __init__, _apply_client_headers_for_base_url(), and
auxiliary_client.py _try_openrouter() + _to_async_client()
- Add _check_openrouter_cache_status() method to AIAgent that reads
X-OpenRouter-Cache-Status from streaming response headers and logs
HIT/MISS status
- Document in cli-config.yaml.example
- Add 28 tests (22 unit + 6 integration)
Ref: https://openrouter.ai/docs/guides/features/response-caching
When resolve_provider_client() passes explicit_api_key for OpenRouter auxiliary
tasks, _try_openrouter() now accepts and honors this parameter instead of
silently ignoring it and falling back to OPENROUTER_API_KEY env var.
Root cause: _try_openrouter() had no explicit_api_key parameter, so even
when callers wanted to pass a runtime credential pool key, it could not be used.
Fix:
- Add explicit_api_key: str = None parameter to _try_openrouter()
- Prioritize explicit_api_key over pool key and env var
- Update resolve_provider_client() call site to pass explicit_api_key
Regression coverage:
- Test that explicit_api_key is passed to OpenAI client when provided
- Test that fallback to OPENROUTER_API_KEY still works when explicit_api_key is None
Closes#18338
When _seed_from_env() reads API keys to populate the credential pool, it
should treat ~/.hermes/.env as the authoritative source — not os.environ.
Stale env vars inherited from parent shell processes (Codex CLI, test
scripts, etc.) can shadow deliberate changes to the .env file, causing
auth.json to cache an outdated key that leads to silent 401 errors.
This is especially visible with OpenRouter: if a parent process exported
OPENROUTER_API_KEY=test-key-fresh and the user later updates .env with a
valid key, restarting Hermes still picks up the stale os.environ value,
writes it back to auth.json, and all API calls fail with 401.
Fixes#18254
Providers like Google Vertex, Azure, and Amazon Bedrock reject API
requests with duplicate tool names (HTTP 400: 'Tool names must be
unique'). The upstream injection paths in run_agent.py already dedup
after PR #17335, but two API-boundary functions pass tools through
without checking:
- agent/auxiliary_client.py: _build_call_kwargs() (all non-Anthropic
providers in chat_completions mode)
- agent/anthropic_adapter.py: convert_tools_to_anthropic() (Anthropic
Messages API path)
Add defensive dedup guards at both sites. Duplicates are dropped with
a warning log, converting a hard 400 failure into a recoverable
condition. This is intentionally conservative — the root-cause dedup
in run_agent.py is the primary defense; these guards add resilience
against future injection-path regressions.
Includes 8 new tests covering unique passthrough, duplicate removal,
empty/None edge cases.
Closes#18478
The process-global `_skill_commands` dict in agent/skill_commands.py
was seeded by whichever platform scanned first, and
`get_skill_commands()` only rescanned when the cache was empty. In a
long-lived gateway process serving multiple platforms (Telegram +
Discord + Slack), the first platform's
`skills.platform_disabled` view was silently inherited by the
others — so a skill disabled for Telegram would also disappear from
Discord's slash menu, and vice versa.
Track the platform scope the cache was populated for
(`_skill_commands_platform`) and rescan in `get_skill_commands()`
when the currently-active platform no longer matches. Platform
resolution uses the same precedence as `_is_skill_disabled`:
`HERMES_PLATFORM` env var then `HERMES_SESSION_PLATFORM` from the
gateway session context.
Fixes#14536
Salvages #14570 by LeonSGP43.
Co-authored-by: LeonSGP <leon@sgp43.com>
* fix(curator): authoritative absorbed_into declarations on skill delete
Closes#18671. The classification pipeline that feeds cron-ref rewriting
used to infer consolidation vs pruning from two brittle signals: the
curator model's post-hoc YAML summary block, and a substring heuristic
scanning other tool calls for the removed skill's name. Both miss in
real consolidations — the model forgets the YAML under reasoning
pressure, and the heuristic misses when the umbrella's patch content
describes the absorbed behavior abstractly instead of naming the old
slug. When both miss, the skill falls through to 'no-evidence fallback'
pruned, and #18253's cron rewriter drops the cron ref entirely instead
of mapping it to the umbrella. Same observable symptom as pre-#18253:
'Skill(s) not found and skipped' at the next cron run.
The fix makes the model declare intent at the moment of deletion.
skill_manage(action='delete') now accepts absorbed_into:
- absorbed_into='<umbrella>' -> consolidated, target must exist on disk
- absorbed_into='' -> explicit prune, no forwarding target
- missing -> legacy path, falls through to heuristic/YAML
The curator reconciler reads these declarations off llm_meta.tool_calls
BEFORE either the YAML block or the substring heuristic. Declaration
wins. Fallback logic stays intact for backward compat with any caller
(human or older curator conversation) that doesn't populate the arg.
Changes
- tools/skill_manager_tool.py: add absorbed_into param to skill_manage
+ _delete_skill. Validate target exists when non-empty. Reject
absorbed_into=<self>. Wire through dispatcher + registry + schema.
- agent/curator.py: new _extract_absorbed_into_declarations() walks
tool calls for skill_manage(delete) with the arg. _reconcile_classification
accepts absorbed_declarations= and treats them as authoritative. Curator
prompt updated to require the arg on every delete.
- Tests: 7 new skill_manager tests covering the tool contract (valid
target, empty string, nonexistent target, self-reference, whitespace,
backward compat, dispatcher plumbing). 11 new curator tests covering
the extractor + authoritative reconciler path + mixed-legacy-and-
declared runs.
Validation
- 307/307 targeted tests pass (curator + cron + skill_manager suites).
- E2E #18671 repro: 3 narrow skills, 1 umbrella, cron job referencing
all 3. Model emits NO YAML block. Heuristic misses (patch prose
doesn't name old slugs). Delete calls carry absorbed_into. Result:
both PR skills correctly classified 'consolidated' + cron rewritten
['pr-review-format', 'pr-review-checklist', 'stale-junk'] ->
['hermes-agent-dev']; stale-junk pruned via absorbed_into=''.
- E2E backward-compat: delete without absorbed_into, model emits YAML
-> routed via existing 'model' source, cron still rewritten correctly.
* feat(curator): capture + restore cron skill links across snapshot/rollback
Before this, rolling back a curator run restored the skills tree but cron
jobs still pointed at the umbrella skills the curator had rewritten them
to. The user would see their old narrow skills back on disk but their
cron jobs still configured with the merged umbrella — not actually 'back
to how it was'.
Snapshot side: snapshot_skills() now captures ~/.hermes/cron/jobs.json
alongside the skills tarball, as cron-jobs.json. The manifest gets a new
'cron_jobs' block with {backed_up, jobs_count} so rollback (and the CLI
confirm dialog) can surface what's in the snapshot. If jobs.json is
missing/unreadable/malformed, snapshot proceeds without cron data — the
skills backup is the core guarantee; cron is additive.
Rollback side: after the skills extract succeeds, the new
_restore_cron_skill_links() reconciles the backed-up jobs into the live
jobs.json SURGICALLY. Only 'skills' and 'skill' fields are restored, and
only on jobs matched by id. Everything else about a cron job — schedule,
last_run_at, next_run_at, enabled, prompt, workdir, hooks — is live
state the user or scheduler has modified since the snapshot; overwriting
it would regress unrelated activity.
Reconciliation rules:
- Job in backup AND live, skills differ → skills restored.
- Job in backup AND live, skills match → no-op.
- Job in backup, NOT in live → skipped (user deleted it
after snapshot; their choice
is later than the snapshot).
- Job in live, NOT in backup → untouched (user created it
after snapshot).
- Snapshot missing cron-jobs.json at all → rollback still succeeds,
reports 'not captured'
(older pre-feature snapshots
keep working).
Writes go through cron.jobs.save_jobs under the same _jobs_file_lock the
scheduler uses, so rollback doesn't race tick().
Also:
- hermes_cli/curator.py: rollback confirm dialog now shows
'cron jobs: N (will be restored for skill-link fields only)' when the
snapshot has cron data, or 'not in snapshot (<reason>)' otherwise.
- rollback()'s message string includes a 'cron links: ...' clause
summarizing the reconciliation outcome.
Tests
- 9 new cases: snapshot-with-cron, snapshot-without-cron, malformed-json
captured-as-raw, full rollback-restores-skills-and-cron, rollback
touches only skill fields, rollback skips user-deleted jobs, rollback
leaves user-created jobs untouched, rollback still works with
pre-feature snapshot that has no cron-jobs.json, standalone unit test
on _restore_cron_skill_links exercising the full report shape.
Validation
- 484/484 targeted tests pass (curator + cron + skill_manager suites).
- E2E: real snapshot_skills, real cron rewrite, real rollback. Before:
['pr-review-format', 'pr-review-checklist', 'pr-triage-salvage'].
After curator: ['hermes-agent-dev']. After rollback: ['pr-review-format',
'pr-review-checklist', 'pr-triage-salvage']. Non-skill fields (id,
name, prompt) preserved across the round trip.
* fix(curator): defer first run and add --dry-run preview (#18373)
Curator was meant to run 7 days after install, not on the very first
gateway tick. On a fresh install (no .curator_state), should_run_now()
returned True immediately because last_run_at was None — so the gateway
cron ticker fired Curator against a fresh skill library moments after
'hermes update'. Combined with the binary 'agent-created' provenance
model (anything not bundled and not hub-installed), this consolidated
hand-authored user workflow skills without consent.
Changes:
- should_run_now(): first observation seeds last_run_at='now' and returns
False. The next real pass fires one full interval_hours later (7 days
by default), matching the original design intent.
- hermes curator run --dry-run: produces the same review report without
applying automatic transitions OR permitting the LLM to call
skill_manage / terminal mv. A DRY-RUN banner is prepended to the
prompt and the caller skips apply_automatic_transitions. State is
NOT advanced so a preview doesn't defer the next scheduled real pass.
- hermes update: prints a one-liner on fresh installs pointing at
--dry-run, pause, and the docs. Silent on steady state.
- Docs: curator.md and cli-commands.md explain the deferred first-run
behavior and warn that hand-written SKILL.md files share the
'agent-created' bucket, with guidance to pin or preview before the
first pass.
Tests:
- test_first_run_defers replaces the old 'first run always eligible'
assertion — same fixture, inverted expectation.
- test_maybe_run_curator_defers_on_fresh_install covers the gateway tick
path end-to-end.
- Three new dry-run tests cover state-advance suppression, prompt
banner injection, and apply_automatic_transitions skipping.
Fixes#18373.
* feat(curator): pre-run backup + rollback (#18373)
Every real curator pass now snapshots ~/.hermes/skills/ into
~/.hermes/skills/.curator_backups/<utc-iso>/skills.tar.gz before calling
apply_automatic_transitions or the LLM review. If a run consolidates or
archives something the user didn't want touched, 'hermes curator
rollback' restores the tree in one command. Dry-run is skipped — no
mutation means no snapshot needed.
Changes:
- agent/curator_backup.py (new): tar.gz snapshot + safe rollback. The
snapshot excludes .curator_backups/ (would recurse) and .hub/ (managed
by the skills hub). Extract refuses absolute paths and .. components,
and uses tarfile's filter='data' on Python 3.12+. Rollback takes a
pre-rollback safety snapshot FIRST, stages the current tree into
.rollback-staging-<ts>/ so the extract lands in an empty dir, and
cleans the staging dir on success. A failed extract restores the
staged contents.
- agent/curator.py: run_curator_review() calls curator_backup.
snapshot_skills(reason='pre-curator-run') before apply_automatic_
transitions. Best-effort — a failed snapshot logs at debug and the
run continues (a transient disk issue shouldn't silently disable
curator forever).
- hermes_cli/curator.py: new 'hermes curator backup' and 'hermes curator
rollback' subcommands. rollback supports --list, --id <ts>, -y.
- hermes_cli/config.py: curator.backup.{enabled, keep} config block
with sane defaults (enabled=true, keep=5).
- Docs: curator.md gets a 'Backups and rollback' section; cli-commands
.md table gets the new rows.
Tests (new file tests/agent/test_curator_backup.py, 16 cases):
- snapshot creates tarball + manifest with correct counts
- snapshot excludes .curator_backups/ (recursion guard) and .hub/
- snapshot disabled via config returns None without creating anything
- snapshot uniquifies ids within the same second (-01 suffix)
- prune honors keep count, newest-first
- list_backups + _resolve_backup cover newest-default and unknown-id
- rollback restores a deleted skill with content intact
- rollback is itself undoable — safety snapshot shows up in list_backups
- rollback with no snapshots returns an error
- rollback refuses tarballs with absolute paths or .. components
- real curator runs take a 'pre-curator-run' snapshot; dry-runs do not
All curator tests: 210 passing locally.
The anyOf collapse in _repair_schema returned early, skipping the
nullable-strip and enum-cleanup steps. When a schema had anyOf
[{enum: [..., null, '']}, {type: null}] alongside a parent-level
'nullable: true', collapsing to the single non-null branch produced a
merged node that still had both 'nullable' and the bad enum values —
Moonshot would still 400 on it.
Fix: fall through to Rules 1/3 when the collapse produces a single
merged node; only return early for the multi-branch case (pure
anyOf preservation) or when there was no null branch to remove.
Adds a test that locks in the combined-case expectation.
When a schema node inside anyOf has enum values but no explicit 'type',
Rule 3 (enum cleanup) ran before _fill_missing_type, so node_type was
None and the enum was never cleaned. Moonshot then rejected the schema
with 'enum value (<nil>) does not match any type in [string]'.
Fix: reorder operations — fill missing type first, strip nullable,
then clean enum. This ensures enum cleanup always has a type to check.
Also fixes test expectation: empty string in enum is now correctly
stripped (Moonshot rejects it too).
Closes#16875
When the curator consolidates skill X into umbrella Y, any cron job
that listed X in its skills field would fail to load X at run time —
the scheduler logs a warning and skips it, so the scheduled job runs
without the instructions it was scheduled to follow.
cron.jobs.rewrite_skill_refs(consolidated, pruned) now updates jobs
in-place: consolidated names route to the umbrella target (dedup
when umbrella is already present), pruned names are dropped.
agent.curator._write_run_report calls it after classification,
best-effort so a cron-side failure never breaks the curator itself.
Results are recorded in run.json (counts.cron_jobs_rewritten + full
cron_rewrites payload), a separate cron_rewrites.json for convenience
when jobs were touched, and a section in REPORT.md.
Reported by @tombielecki.
The user-visible /compress banner and the post-compression last_prompt_tokens
writeback both counted only the raw message transcript (chars/4). With a 15KB
system prompt and 30 tool schemas (~26KB), a 4-message transcript that looks
like ~45 tokens to the transcript-only estimator is really ~10.5K tokens of
request pressure — a 234x gap.
Two user-facing consequences:
- Banner shows 'Compressing … (~45 tokens)…' while compression is actually
firing on 10K+ tokens of real pressure, confusing users about why
compression triggered (reported by @codecovenant on X; #6217).
- Post-compression last_prompt_tokens writeback omits tool schemas, so the
next should_compress() check compares real usage against a stale
underestimate — compression triggers late, potentially past the model's
context limit on small-context models (#14695).
Swap estimate_messages_tokens_rough() for estimate_request_tokens_rough()
at every user-visible banner and at the post-compression writeback.
estimate_request_tokens_rough() already existed for exactly this purpose
and includes system prompt + tool schemas.
Touched call sites:
- run_agent.py: post-compression last_prompt_tokens writeback, post-tool
call should_compress() fallback when provider usage is missing
- cli.py: /compress banner + summary
- gateway/run.py: gateway /compress banner + summary
- tui_gateway/server.py: TUI /compress status + summary
- acp_adapter/server.py: ACP /compact before/after
Left intentionally alone:
- Session-hygiene fallback and the 'no agent' /status path in gateway/run.py
— no agent instance is in scope to query for system prompt/tools, and the
existing 30-50% overestimate wobble on hygiene is safety-accepted.
- Verbose-mode 'Request size' logging — informational only, already counts
system prompt via api_messages[0].
Also relabels the feedback line from 'Rough transcript estimate' to
'Approx request size' so the metric label matches what it actually measures.
Credits: diagnoses from @devilardis (#14695) and @Jackten (#6217);
user report @codecovenant on X (2026-04-30).
Closes#14695Closes#6217
The initial guardrail PR consolidated failure classification by pointing
display._detect_tool_failure at the new classify_tool_failure helper,
which was strictly broader: it flagged any JSON result with
"success": false / "failed": true / non-empty "error", plus plain-text
"traceback" and "error:" prefixes. That would uptick the user-visible
[error] tag on tools that return {"success": false} as a benign signal
(memory fullness, todo state, etc.) and feed the failure-streak counter
at the same time.
Restore display._detect_tool_failure to its pre-PR semantics verbatim.
Tighten classify_tool_failure (the guardrail's internal safety-fallback
used only when callers don't pass failed=) to match _detect_tool_failure
exactly, so the two never disagree. Production callers in run_agent.py
already pass an explicit failed= derived from _detect_tool_failure, so
the guardrail counter is driven by the same signal the CLI shows.
When a user defines `custom_providers: [{name: kimi, ...}]` and references
`provider: kimi` from fallback_model or the main config, the built-in alias
rewriting (`kimi` → `kimi-coding`) was hijacking the request before the
named-custom lookup ran. `_get_named_custom_provider` also refused to
return a match when the raw name resolved to any built-in (including aliases),
so the custom endpoint was unreachable.
Fix at both layers of the resolution chain so every caller benefits, not
just `_try_activate_fallback`:
- hermes_cli/runtime_provider.py: narrow `_get_named_custom_provider`'s
built-in-wins guard to canonical provider names only. An alias like
`kimi` that resolves to a different canonical (`kimi-coding`) no longer
blocks the custom lookup; a canonical name like `nous` still does.
- agent/auxiliary_client.py: in `resolve_provider_client`, try the named-
custom lookup with the original (pre-alias-normalization) name before the
alias-normalized one, so aliased requests reach the user's custom entry.
Also honour `explicit_base_url` and `explicit_api_key` in the API-key
provider branch so callers that pass explicit hints (e.g. fallback
activation) can override the registered defaults.
Tests added for:
- custom `kimi` shadowing built-in alias (regression for #15743)
- custom `nous` NOT shadowing canonical built-in (behaviour preserved)
- bare `kimi` without any custom entry still routing to built-in
- explicit base_url/api_key override on the API-key provider branch
Original PR #17827 by @Feranmi10 identified the same bug class and
implemented a narrower fix in `_try_activate_fallback`; this reshapes the
fix to live in the shared resolution layer so all callers benefit.
Fixes#15743
Co-authored-by: Feranmi10 <89228157+Feranmi10@users.noreply.github.com>
Treat skill views and edits as activity when curator reports and applies lifecycle transitions, so recently loaded or patched skills are not displayed or transitioned as never used.\n\nAdds regression tests for activity derivation, automatic transitions, and CLI status output.
* fix(curator): split 'archived' into consolidated vs pruned in run reports
Users who watched a curator run saw skills like 'anthropic-api' listed
under 'Skills archived' and interpreted that as pruning — but the curator
had actually absorbed those skills into a new umbrella (e.g. 'llm-providers')
during the same run. The directory gets archived for safety (all removals
are recoverable), but the content still lives under a different name.
Users then 'restored' what they thought were deleted skills and ended up
with confusingly duplicated skillsets (old-name + absorbed-inside-umbrella).
Classify removed skills using this run's skill_manage tool calls:
- consolidated: content absorbed into a surviving/newly-created skill
(evidenced by a skill_manage write_file/patch/create/edit whose target
is a different skill AND whose file_path/content references the
removed skill's name)
- pruned: archived without consolidation evidence (truly stale)
REPORT.md now shows two distinct sections:
- 'Consolidated into umbrella skills' — with `removed → merged into umbrella`
- 'Pruned — archived for staleness' — pure staleness archives
run.json schema additions (backward compatible):
- counts.consolidated_this_run, counts.pruned_this_run
- consolidated: [{name, into, evidence}, ...]
- pruned: [names]
- archived: retained as the union for backward compat
Also: relabel the auto-transitions 'archived' counter to 'archived (no
LLM, pure time-based staleness)' so it's clearly distinct from LLM-pass
archives.
Tests: 9 new tests in test_curator_classification.py covering consolidation
evidence parsing (write_file/patch/create), hyphen/underscore name variants,
self-reference rejection, destination-must-exist, mixed runs, and
malformed-JSON fallback safety. Existing test_report_md_is_human_readable
updated to cover the new section names.
E2E: isolated HERMES_HOME, realistic 3-skill run, REPORT.md verified
end-to-end.
* feat(curator): hybrid model-declared + heuristic classification
Extend the consolidated-vs-pruned split with LLM-authored intent:
1. Curator prompt now requires a structured YAML block at the end of the
final response (consolidations / prunings with short rationale).
2. _parse_structured_summary() extracts it tolerantly — missing block,
malformed YAML, partial lists all fall back to heuristic cleanly.
3. _reconcile_classification() merges model intent with the tool-call
heuristic:
- Model wins on rationale when its umbrella exists post-run
- Model hallucination (umbrella doesn't exist) is downgraded to the
heuristic's finding, or pruned if there's no evidence either
- Heuristic catches model omission — consolidations the model
enumerated tools for but forgot to list get surfaced with a
'(detected via tool-call audit)' tag
4. REPORT.md now shows per-row rationale alongside 'removed → umbrella'
and flags audit-only rows so the user knows why no reason is shown.
Backward compat: run.json's 'archived' field (union) is preserved.
'pruned' is now a list of dicts with {name, source, reason};
'pruned_names' is the flat-name list for legacy consumers.
Tests: 15 new covering YAML parse edge cases (malformed, empty lists,
bare-string entries, missing fields), reconciler rules (model wins,
hallucination fallback, heuristic catches omission, prune with reason),
and an end-to-end report-render test with all four paths exercised.
Fixes HTTP 404 errors when using Anthropic-compatible providers (Kimi Coding, MiniMax, MiniMax-CN) for auxiliary tasks.
Root cause: `_to_openai_base_url()` rewrites `/anthropic` → `/v1` so the OpenAI SDK hits the right endpoint. But the rewritten URL was then passed to `_maybe_wrap_anthropic`, whose `_endpoint_speaks_anthropic_messages` detector only fires on `/anthropic` or `api.kimi.com/coding`. Detector saw `/v1` → returned False → no Anthropic wrap → 404 on every aux call.
Fix: preserve the raw base_url before rewriting and pass it to `_maybe_wrap_anthropic` for transport detection, while still giving the rewritten URL to the OpenAI client constructor.
Closes#17705, #17413, #17086, #10469.
Co-authored-by: oak <chengoak@users.noreply.github.com>
bump_use() existed and was tested but had zero production call sites —
use_count stayed 0 for all skills, breaking Curator's stale-detection
logic which relies on last_used_at.
Wire bump_use() into:
1. build_skill_invocation_message() — when a user invokes /skill-name
2. build_preloaded_skills_prompt() — when a skill is preloaded at session start
Both are the canonical 'a skill is actively being used' moments, distinct
from 'browsing' (bump_view in skill_view tool call).
Closes#17782
Archived skills (moved to ~/.hermes/skills/.archive/ by the curator)
were still surfaced in the <available_skills> system prompt under a
fake '.archive' category, causing the agent to load and try to use
deprecated skills. The os.walk in iter_skill_index_files() only
excluded .git/.github/.hub.
Add '.archive' to EXCLUDED_SKILL_DIRS, and to the two other places
that hardcode the same exclusion tuple (gateway/run.py and
agent/skill_commands.py).
Three fixes bundled for curator reliability on existing installs and
broken/partial installs:
1. run_agent.py: defer `import fire` into the __main__ block. `fire` is
only used by `fire.Fire(main)` when running run_agent.py directly as
a CLI — it is NOT needed for library usage. Importing it at module
top made `from run_agent import AIAgent` from a daemon thread (e.g.
the curator's forked review agent) crash with ModuleNotFoundError
on broken/partial installs where `fire` isn't present.
2. hermes_cli/config.py: add version 22 → 23 migration that writes the
`curator` + `auxiliary.curator` sections to config.yaml with their
defaults, only filling keys the user hasn't overridden. Existing
configs from before PR #16049 / the April 2026 `auxiliary.curator`
unification had neither section on disk, so users couldn't see or
edit the settings in their config.yaml (runtime deep-merge papered
over it at read time, but the file never reflected reality).
3. hermes_cli/config.py: `ensure_hermes_home()` now pre-creates
`~/.hermes/logs/curator/` alongside cron/sessions/logs/memories on
every CLI launch. Managed-mode (NixOS) variant mkdir's it
defensively after the activation-script existence checks, since the
activation script may not know about this subpath.
4. agent/curator.py: `_reports_root()` mkdir's the dir at call time as
belt-and-suspenders for entry paths that bypass both
ensure_hermes_home() and the v23 migration (gateway-only installs,
bare library use).
E2E validated in isolated HERMES_HOME: fresh install gets full defaults
seeded; partial-override config keeps user's `enabled: false` and
custom `interval_hours` while filling the missing keys; re-running the
migration is a no-op.
When a user sets model.context_length in config.yaml, the value was only
used for Hermes' internal compression decisions (context_compressor) but
NOT for Ollama's num_ctx parameter. Ollama auto-detects context from GGUF
metadata (often 256K+) and allocates that much VRAM regardless of the
user's config — causing OOM on smaller GPUs like the P100 (16GB).
Root cause: two separate context values existed independently:
- context_compressor.context_length = config value (e.g. 65536) ✓
- _ollama_num_ctx = GGUF metadata value (e.g. 256000) ✗ ignored config
Changes:
1. Cap Ollama num_ctx to config context_length (run_agent.py)
When model.context_length is explicitly set and no explicit
ollama_num_ctx override exists, cap the auto-detected GGUF value
to the user's context_length. This is the core fix — it prevents
Ollama from allocating more VRAM than the user budgeted.
2. Pass config_context_length through all secondary call sites
Several paths called get_model_context_length() without the config
override, falling through to the 256K default fallback:
- cli.py: @-reference expansion and /model switch display
- gateway/run.py: @-reference expansion and /model switch display
- tui_gateway/server.py: @-reference expansion
- hermes_cli/model_switch.py: resolve_display_context_length()
3. Normalize root-level context_length in config (hermes_cli/config.py)
_normalize_root_model_keys() now migrates root-level context_length
into the model section, matching existing behavior for provider and
base_url. Users who wrote `context_length: 65536` at the YAML root
instead of under `model:` had it silently ignored.
4. Fix misleading comments (agent/model_metadata.py)
DEFAULT_FALLBACK_CONTEXT is 256K (CONTEXT_PROBE_TIERS[0]), not 128K
as two comments stated.
Tests: 3 new tests for root-level context_length normalization.
All existing context_length tests pass (96 tests).
The `gemini` provider also serves Gemma (e.g. `gemma-4-31b-it`) and
historically other Google models like PaLM. Those reject
`extra_body.thinking_config` with HTTP 400:
Unknown name "thinking_config": Cannot find field
`_build_gemini_thinking_config()` was unconditionally producing a
config dict for any model on the `gemini` / `google-gemini-cli`
provider, which `ChatCompletionsTransport.build_kwargs` then dropped
into `extra_body["thinking_config"]`. The result: every chat turn for
Gemma users on the gemini provider blew up at the API edge.
The fix is the same shape Hermes already uses for the Gemini-2.5 vs
Gemini-3 family clamping: normalise the model id, strip an
`OpenRouter`-style `google/` prefix, and short-circuit early when the
result doesn't start with `gemini`. We return `None` rather than
`{"includeThoughts": False}`, because the API rejects the field name
itself — even the polite "off" form trips the same 400.
Three regression tests cover Gemma with reasoning enabled, Gemma with
reasoning disabled, and the `google/gemma-…` OpenRouter-style id; the
existing Gemini-2.5 / Gemini-3 / `google/gemini-…` cases keep passing
because the Gemini guard fires after the prefix strip.
Fixes#17426
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Voscko reported curator.auxiliary.provider/model was advertised in the
docs but ignored — the review fork read only model.provider/default. The
narrow fix would wire the one-off key through, but that leaves curator
as a parallel system: not in `hermes model` → auxiliary picker, not in
the dashboard Models tab, missing per-task base_url/api_key/timeout/
extra_body.
Unify curator with the rest of the aux task system so `hermes model`
and the dashboard configure it like every other aux task.
Four sources of truth updated:
- hermes_cli/config.py — add 'curator' slot to DEFAULT_CONFIG.auxiliary
(timeout=600 since reviews run long), drop the one-off curator.auxiliary
block from DEFAULT_CONFIG.curator.
- hermes_cli/main.py — add ('curator', 'Curator', 'skill-usage review pass')
to _AUX_TASKS so the CLI picker offers it.
- hermes_cli/web_server.py — add 'curator' to _AUX_TASK_SLOTS so the
dashboard REST endpoint accepts it.
- web/src/pages/ModelsPage.tsx — add Curator entry so the dashboard
Models tab renders the task.
agent/curator.py _resolve_review_model() now reads auxiliary.curator
first (canonical), falls back to legacy curator.auxiliary (with an info
log asking users to migrate), then falls back to the main chat model.
Pre-unification users keep working.
Docs updated: docs/user-guide/features/curator.md now points at
`hermes model` → auxiliary → Curator and the dashboard Models tab.
Tests: 6 unit tests on _resolve_review_model (auto default, canonical
slot honored, partial override fallback, legacy fallback with
deprecation log assertion, new-wins-over-legacy, empty-config safety)
plus a cross-registry test that curator is wired into all four sources
of truth. test_aux_tasks_keys_all_exist_in_default_config already
covers the DEFAULT_CONFIG ↔ _AUX_TASKS invariant.
Reported by Voscko on Discord.
The _CODEX_AUX_MODEL constant had already rotated twice in 6 weeks
(gpt-5.3-codex -> gpt-5.2-codex -> now broken again at gpt-5.2-codex)
because ChatGPT-account Codex gates which models it accepts via an
undocumented, shifting allow-list that OpenAI publishes no changelog
for. Any pinned default will keep going stale. Issue #17533 reports
the current breakage: every ChatGPT-account auxiliary fallback fails
with HTTP 400 "model is not supported" and the 60s pause loop degrades
long sessions.
Rather than reset the clock with another stale pin (PR #17544 proposes
gpt-5.2-codex -> gpt-5.4), remove the hardcoded second-order Codex
fallback entirely:
- Delete `_CODEX_AUX_MODEL`.
- Drop `_try_codex` from `_get_provider_chain()` (the auto chain now
ends at api-key providers; 4 rungs instead of 5).
- Rename `_try_codex() -> _build_codex_client(model)` and require an
explicit model from the caller. No more guessing.
- `resolve_provider_client("openai-codex", model=None)` now warns and
returns (None, None) instead of silently guessing a stale model ID.
- Remove `_try_codex` from the `provider="custom"` fallback ladder
(same stale-constant trap).
- `_resolve_strict_vision_backend("openai-codex")` routes through
`resolve_provider_client` so the caller's explicit model is honored.
Codex-main users are unaffected: Step 1 of `_resolve_auto` already
uses `main_provider` + `main_model` directly and passes the user's
configured Codex model through `resolve_provider_client`, which never
touched `_CODEX_AUX_MODEL`. Per-task overrides (`auxiliary.<task>.provider/model`)
continue to work and are the supported way to route specific aux tasks
through Codex.
Users whose main provider fails with a payment/connection error and
who have ONLY ChatGPT-account Codex auth will now see the 60s pause
without a stale-model-rejection noise line in between -- same outcome,
cleaner failure.
Closes#17533. Supersedes #17544 (which resets the clock on the
same stale-constant problem).
Keep context-1m-2025-08-07 in OAuth requests by default so 1M-capable
subscriptions retain full context. When Anthropic rejects a request with
400 'long context beta is not yet available for this subscription',
disable the beta for the rest of the session, rebuild the client, and
retry once.
Addresses #17680 (thanks @JayGwod for the clean reproduction) without
forcing every OAuth user off the 1M context window.
Changes:
- agent/error_classifier.py: new FailoverReason.oauth_long_context_beta_forbidden;
pattern matches 400 + 'long context beta' + 'not yet available'. Narrow
enough that the existing 429 tier-gate pattern keeps its own reason.
- agent/anthropic_adapter.py: _common_betas_for_base_url,
build_anthropic_client, build_anthropic_kwargs gain drop_context_1m_beta
kwarg. Default=False (1M stays). OAuth OAUTH_ONLY_BETAS unchanged.
- agent/transports/anthropic.py: build_kwargs forwards the flag.
- run_agent.py: self._oauth_1m_beta_disabled flag, retry-once guard,
recovery branch next to the image-shrink path. _rebuild_anthropic_client
honors the flag. The main build_kwargs call site threads it through for
fast-mode extra_headers.
- hermes_cli/doctor.py, hermes_cli/models.py: sibling OAuth /v1/models
probes get the same reactive retry — previously they'd falsely report
the Anthropic API as unreachable for affected subscriptions.
Tests: 2190 tests/agent/ + 94 adjacent integration tests pass. New unit
tests cover the classifier pattern (including the collision guard against
the 429 tier-gate) and the drop_context_1m_beta adapter behavior (default
keeps 1M, flag strips only 1M while preserving every other beta).
Salvage-follow-up to @shannonsands's /reload-skills PR. Trims the feature to
match the design: user-initiated rescan, no prompt-cache reset, no new
schema surface, no phantom user turn, and the next-turn note carries each
added/removed skill's 60-char description (not just its name).
Changes vs the original PR:
* Drop the in-process skills prompt-cache clear in reload_skills(). Skills
are invoked at runtime via /skill-name, skills_list, or skill_view —
they don't need to live in the system prompt for the model to use them.
Keeping the cache intact preserves prefix caching across the reload so
/reload-skills pays no cache-reset cost. (MCP has to break the cache
because tool schemas must be known at conversation start; skills do not.)
* Drop the skills_reload agent tool and SKILLS_RELOAD_SCHEMA from
tools/skills_tool.py, plus the four skills_reload enumerations in
toolsets.py. No new schema surface — agents can already see a freshly-
installed skill via skill_view / skills_list the moment it's on disk.
* Replace the phantom 'role: user' turn injection with a one-shot queued
note. CLI uses self._pending_skills_reload_note (same pattern as
_pending_model_switch_note, prepended to the next API call and cleared).
Gateway uses self._pending_skills_reload_notes[session_key]. The note
is prepended to the NEXT real user message in this session, so message
alternation stays intact and nothing out-of-band is persisted to the
transcript.
* reload_skills() now returns added/removed as
[{'name': str, 'description': str}, ...] (description truncated to 60
chars — matches the curator / gateway adapter budget). The injected
next-turn note formats each entry as 'name — description' so the model
can actually reason about which new skills to call without running
skills_list first.
* Only emit the note when the diff is non-empty. On empty diff, print
'No new skills detected' and do nothing else.
* Tests rewritten to cover the queue semantics, the description payload,
and a regression guard that the prompt-cache snapshot is preserved.
Adds a public reload path for the in-process skill caches so newly
installed (or removed) skills become visible mid-session without a
gateway restart. Mirrors the shape of /reload-mcp.
Three surfaces:
* /reload-skills slash command — CLI (cli.py) and gateway (gateway/run.py),
with /reload_skills alias for Telegram autocomplete and an explicit
Discord registration.
* skills_reload agent tool (tools/skills_tool.py) — lets agents/subagents
pick up freshly-installed skills via tool call.
* agent.skill_commands.reload_skills() — shared helper that clears
_skill_commands, _SKILLS_PROMPT_CACHE (in-process LRU), and the
on-disk .skills_prompt_snapshot.json, then returns an added/removed
diff plus the new total count.
Tested:
* tests/agent/test_skill_commands_reload.py (9 cases)
* tests/cli/test_cli_reload_skills.py (3 cases)
* tests/gateway/test_reload_skills_command.py (4 cases)
Use case: NemoClaw / OpenShell-style sandboxed orchestrators that drop
skills into ~/.hermes/skills mid-session, plus agentic flows where the
agent itself installs a skill via the shell tool and needs it bound
without a gateway restart. The Python helper
clear_skills_system_prompt_cache(clear_snapshot=True) already exists
internally — this PR just exposes it via slash command and tool.
Close integration gaps discovered by auditing qwen-oauth's file coverage.
These are surfaces the original salvage missed — they all existed on
main and were added in the 747 commits since PR #15203 was opened.
Coverage added:
- agent/credential_pool.py: seed pool from auth.json providers.minimax-oauth
so `hermes auth list` reflects logged-in state and
`hermes auth remove minimax-oauth <N>` works through the standard flow.
- agent/credential_sources.py: register RemovalStep for minimax-oauth
with suppression-aware `_clear_auth_store_provider`.
- agent/models_dev.py: PROVIDER_TO_MODELS_DEV mapping (-> 'minimax' family).
- hermes_cli/providers.py: HermesOverlay entry (anthropic_messages transport,
oauth_external auth_type, api.minimax.io/anthropic base).
- hermes_cli/model_normalize.py: add to _MATCHING_PREFIX_STRIP_PROVIDERS so
`minimax-oauth/MiniMax-M2.7` in config.yaml gets correctly repaired.
- hermes_cli/status.py: render MiniMax OAuth block in `hermes doctor`
(logged-in / region / expires_at / error).
- hermes_cli/web_server.py: register in OAUTH_PROVIDER_REGISTRY + dispatch
branch in _resolve_provider_status so the dashboard auth page shows it.
- website/docs/integrations/providers.md: full 'MiniMax (OAuth)' section.
- website/docs/reference/cli-commands.md: --provider enum.
- website/docs/user-guide/features/fallback-providers.md: fallback table row.
- scripts/release.py AUTHOR_MAP: amanning3390 mapping (CI gate).
Wire MiniMax-M2.7 and MiniMax-M2.7-highspeed into the model catalog,
CLI model picker, and agent auxiliary/metadata subsystems.
Changes:
- hermes_cli/models.py:
- Add 'minimax-oauth' to _PROVIDER_MODELS with MiniMax-M2.7 and
MiniMax-M2.7-highspeed
- Add ProviderEntry('minimax-oauth', 'MiniMax (OAuth)', ...) to
CANONICAL_PROVIDERS near existing minimax entries
- Add aliases: minimax-portal, minimax-global, minimax_oauth in
_PROVIDER_ALIASES
- hermes_cli/main.py:
- Add 'minimax-oauth' to provider_labels dict
- Insert 'minimax-oauth' into providers list in
select_provider_and_model() near the other minimax entries
- Add 'minimax-oauth' to --provider argparse choices
- Add _model_flow_minimax_oauth() function: ensures login via
_login_minimax_oauth(), resolves runtime credentials, prompts for
model selection, saves model choice and config
- Add dispatch elif branch for selected_provider == 'minimax-oauth'
- agent/auxiliary_client.py:
- Add 'minimax-oauth': 'MiniMax-M2.7-highspeed' to
_API_KEY_PROVIDER_AUX_MODELS
- Add 'minimax-oauth' to _ANTHROPIC_COMPAT_PROVIDERS set
- agent/model_metadata.py:
- Add 'minimax-oauth' to _PROVIDER_PREFIXES frozenset
- MiniMax-M2.7 context length (200_000) already covered by the
existing 'minimax' substring match in DEFAULT_CONTEXT_LENGTHS
DeepSeek's /anthropic endpoint requires thinking blocks to be replayed
in multi-turn conversations for reasoning continuity. The existing code
classified api.deepseek.com as a generic third-party endpoint and stripped
ALL thinking blocks, causing HTTP 400 from DeepSeek.
Fix: add _is_deepseek_anthropic_endpoint() detector (following the Kimi
precedent) and a dedicated branch that strips only signed Anthropic blocks
while preserving unsigned ones synthesised from reasoning_content.
This follows the exact same pattern as the Kimi exemption (issue #13848)
and does not change behavior for any other third-party endpoint (Azure,
Bedrock, MiniMax, etc.).
FixesNousResearch/hermes-agent#16748
The ~/.openclaw/ detection banner (#16327) had two problems flagged in #16629:
1. It only pitched 'hermes claw cleanup' (destructive archive) and never
mentioned 'hermes claw migrate' — the actual non-destructive path that
ports config/memory/skills into Hermes.
2. The copy anthropomorphized the bug ('the agent can still get confused',
'dutifully reads') and framed OpenClaw as a competitor to eliminate
('instead of Hermes's').
Rewrite so migrate leads, cleanup is a clearly-labelled follow-up with a
warning that archiving breaks OpenClaw for users still running it.
Closes#16629
The guard that drops Anthropic's `thinking` kwarg for Kimi endpoints was
matched on `https://api.kimi.com/coding` only. Users configuring a
custom Kimi-compatible gateway (or an official Moonshot host) with
`api_mode: anthropic_messages` fall through to the generic third-party
path, which strips thinking blocks AND still sends
`thinking={enabled,...}` → upstream rejects with HTTP 400
"reasoning_content is missing in assistant tool call message at index N"
on the next request after a tool call.
Replace `_is_kimi_coding_endpoint` callers (history replay + thinking
kwarg gate) with `_is_kimi_family_endpoint(base_url, model)` that also
matches the `api.kimi.com` / `moonshot.ai` / `moonshot.cn` hosts and
Kimi/Moonshot family model names (`kimi-`, `moonshot-`, `k1.`, `k2.`,
…) for custom / proxied endpoints. Keeps the UA-header check in
`build_anthropic_client` URL-only — the `claude-code/0.1.0` header is
an official-Kimi contract.
Plumbs optional `model` through `convert_messages_to_anthropic` so
the unsigned reasoning_content→thinking block synthesised for Kimi's
history validation survives the third-party signature-stripping pass
on custom hosts too.
Closes#17057.
The normalize_model_name() function unconditionally converted dots to
hyphens in all model names. This caused non-Anthropic models (e.g.
gpt-5.4) to be mangled to gpt-5-4 when routed through the Anthropic
adapter path, resulting in HTTP 404 from the backend.
Now only applies dot-to-hyphen conversion for models starting with
"claude-" or "anthropic/", which are the actual Anthropic model IDs.
Fixes NousResearch/hermes-agent#17171
Related: #7421, #13061, #16417
* docs(anthropic): correct OAuth scope to Max plan + extra usage credits only
The previous docs pass (#17399) overstated what Anthropic OAuth works
with. In practice Hermes can only route against a Claude Max plan that
has purchased extra usage credits — the base Max allowance is not
consumed, and Claude Pro is not supported at all. Without Max + extra
credits, users must fall back to an ANTHROPIC_API_KEY (pay-per-token).
Updates the four pages touched in #17399:
- integrations/providers.md
- user-guide/features/credential-pools.md
- reference/environment-variables.md
- getting-started/quickstart.md
* fix(aux): skip kimi-coding in vision auto-detect (closes#17076)
Kimi Coding Plan's /coding endpoint (Anthropic Messages wire) has no
image_in capability — Kimi's own docs confirm and suggest switching to
a vision-capable model. Vision lives on the separate Kimi Platform
(api.moonshot.ai, OpenAI-wire, pay-as-you-go). When the user has
kimi-coding as main provider and auxiliary.vision.provider=auto,
resolve_vision_provider_client was handing back an AnthropicAuxiliaryClient
wrapped around /coding which 404'd on every vision request.
Add a _PROVIDERS_WITHOUT_VISION frozenset ({kimi-coding, kimi-coding-cn})
and gate the main-provider vision branch on membership. On a skip the
auto-detect falls through to OpenRouter → Nous like any other
main-provider-unavailable case.
Explicit per-task overrides (auxiliary.vision.provider=kimi-coding) are
unaffected — the skip only applies when the caller is in auto mode.
Tests: 4 new targeted tests in TestVisionAutoSkipsKimiCoding covering
the skip path, CN variant, explicit-override passthrough, and a guard
against accidental skip-list widening.
Fixes#6672
Memory providers now receive on_session_switch() whenever AIAgent.session_id
rotates mid-process — /resume, /branch, /reset, /new, and context
compression. Before this, providers that cached per-session state in
initialize() (Hindsight's _session_id, _document_id, accumulated
_session_turns, _turn_counter) kept writing into the old session's
record after the agent had moved on.
MemoryProvider ABC
------------------
- New optional hook on_session_switch(new_session_id, *,
parent_session_id='', reset=False, **kwargs) with no-op default for
backward compat. reset=True signals /reset or /new — providers should
flush accumulated per-session buffers. reset=False for /resume,
/branch, compression where the logical conversation continues.
MemoryManager
-------------
- on_session_switch() fans the hook out to every registered provider.
Isolated try/except per provider — one bad provider can't block others.
- Empty/None new_session_id is a no-op to avoid corrupting provider state
during shutdown paths.
run_agent.py
------------
- _sync_external_memory_for_turn now passes session_id=self.session_id
into sync_all() and queue_prefetch_all(). Providers with defensive
session_id updates in sync_turn (Hindsight already had this at
plugins/memory/hindsight/__init__.py:1199) now actually receive the
current id.
- Compression block at ~L8884 already notified the context engine of
the rollover; now also calls
_memory_manager.on_session_switch(reason='compression').
cli.py
------
- new_session() fires reset=True, reason='new_session' so providers
flush buffers.
- _handle_resume_command fires reset=False, reason='resume' with the
previous session as parent_session_id.
- _handle_branch_command fires reset=False, reason='branch' with the
parent session_id already captured for the DB parent link.
gateway/run.py
--------------
- _handle_resume_command now evicts the cached AIAgent, mirroring
/branch and /reset. The next message rebuilds a fresh agent whose
memory provider initialize() runs with the correct session_id —
matches the pattern the gateway already uses for provider state
cross-session transitions.
Hindsight reference implementation
----------------------------------
- plugins/memory/hindsight/__init__.py adds on_session_switch that:
updates _session_id, mints a fresh _document_id (prevents
vectorize-io/hindsight#1303 overwrite), and clears _session_turns /
_turn_counter / _turn_index so in-flight batches don't flush under
the new document id. parent_session_id only overwritten when provided
(avoids clobbering on a bare switch).
Tests
-----
- tests/agent/test_memory_session_switch.py: new dedicated file. ABC
default no-op, manager fan-out, failure isolation, empty-id no-op,
session_id propagation through sync_all/queue_prefetch_all, Hindsight
state transitions for every reset/non-reset case, parent preservation.
- tests/cli/test_branch_command.py: new test verifying /branch fires
the hook with correct parent_session_id + reset=False + reason.
- tests/gateway/test_resume_command.py: new test verifying /resume
evicts the cached agent.
- tests/run_agent/test_memory_sync_interrupted.py: updated existing
assertions to account for the session_id kwarg on sync_all and
queue_prefetch_all.
E2E verified (real imports, tmp HERMES_HOME):
- /resume: session_id updates, doc_id fresh, buffers cleared, parent set
- /branch: session_id forks, parent links to original
- /new: reset=True clears accumulated state
- compression: reason='compression' propagated, lineage preserved
- Empty id: no-op, state preserved
- Legacy provider without on_session_switch: no crash
Reported by @nicoloboschi (Hindsight maintainer); related scope-widening
comment by @kidonng extending coverage to compression.
Fixes#16825. Sessions using MiniMax-M2.7 via minimax-cn showed
estimated_cost_usd=0.0 and cost_status='unknown' because neither
provider had a billing route or pricing entry. Adds official_docs_snapshot
entries ($0.30/M input, $1.20/M output) for both minimax and minimax-cn,
and adds explicit routing in resolve_billing_route so both resolve to
billing_mode='official_docs_snapshot' instead of falling through to 'unknown'.
Every curator pass now emits a dated report directory under
`~/.hermes/logs/curator/{YYYYMMDD-HHMMSS}/` with two files:
- `run.json` — machine-readable full record (before/after snapshot,
state transitions, all tool calls, model/provider, timing, full LLM
final response untruncated, error if any)
- `REPORT.md` — human-readable markdown: model + duration header,
auto-transition counts, LLM consolidation stats, archived-this-run
list, new-skills-this-run list, state transitions, the full LLM
final summary, and a recovery footer pointing at the archive + the
`hermes curator restore` command
Reports live under `logs/curator/`, not inside `skills/` — they're
operational telemetry, not user-authored skill data, and belong
alongside `agent.log` / `gateway.log`.
Internals:
- `_run_llm_review()` now returns a dict (final, summary, model,
provider, tool_calls, error) instead of a bare truncated string so
the reporter has full fidelity
- Report writer is fully best-effort — any failure logs at DEBUG and
never breaks the curator itself. Same-second rerun gets a numeric
suffix so reports can't clobber each other
- Report path stamped into `.curator_state` as `last_report_path`
- `hermes curator status` surfaces a "last report:" line so users
can immediately open the latest run
Tests (all green):
- 7 new tests in tests/agent/test_curator_reports.py covering: report
location (logs not skills), both files written, run.json shape and
diff accuracy, markdown structure, error path still writes, state
transitions captured, same-second runs get unique dirs
- Existing test_run_review_synchronous_invokes_llm_stub updated to
stub the new dict-returning _run_llm_review signature
Live E2E: ran a synchronous pass against a 1-skill test collection
with a stubbed LLM; report written correctly, state stamped with
last_report_path, markdown human-readable, run.json machine-parseable.
Based on three live test runs against 346 agent-created skills on the
author's own setup (~6.5 min, opus-4.7, 86 API calls), the curator
prompt needed three sharpenings before it consistently produced real
umbrella consolidation instead of passive audit output:
**Umbrella-first framing.** The original 'decide keep/patch/archive/
consolidate' framing lets opus default to 'keep' whenever two skills
aren't byte-identical. The new prompt explicitly tells the reviewer
that pairwise distinctness is the wrong bar — the right question is
'would a human maintainer write this as N separate skills, or one
skill with N labeled subsections?' Expect 10-25 prefix clusters; merge
each into an umbrella via one of three methods.
**Three concrete consolidation methods.** (a) Merge into an existing
umbrella (patch the broadest skill, archive siblings); (b) Create a
new umbrella SKILL.md (skill_manage action=create); (c) Demote
session-specific detail into references/, templates/, or scripts/
under the umbrella via skill_manage action=write_file, then archive
the narrow sibling. This matches the support-file vocabulary the
review-prompt side already uses (PR #17213).
**Two observed bailouts pre-empted:** 'usage counters are zero so I
can't judge' (rule 4: judge on content, not use_count) and 'each has
a distinct trigger' (rule 5: pairwise distinctness is the wrong bar).
**Config-aware parent inheritance.** _run_llm_review() was building
AIAgent() without explicit provider/model, hitting an auto-resolve
path that returned empty credentials → HTTP 400 'No models provided'
against OpenRouter. Fork now inherits the user's main provider and
model (via load_config + resolve_runtime_provider) before spawning —
runs on whatever the user is currently on, OAuth-backed or
pool-backed included.
**Unbounded iteration ceiling.** max_iterations=8 was way too low for
an umbrella-build pass over hundreds of skills. A live pass takes
50-100 API calls (scanning, clustering, skill_view'ing candidates,
patching umbrellas, mv'ing siblings). Raised to 9999 — the natural
stopping criterion is 'no more clusters worth processing', not an
arbitrary tool-call budget.
**Tests updated:** test_curator_review_prompt_has_invariants accepts
DO NOT / MUST NOT and drops 'keep' from the required-verb set (the
umbrella-first prompt correctly deemphasizes 'keep' as a first-class
decision label since passive keep-everything is the failure mode
being prevented). Added test_curator_review_prompt_is_umbrella_first
asserting the umbrella framing, class-level thinking, references/
+ templates/ + scripts/ support-file mentions, and the 'use_count
is not evidence of value' pre-emption. Added
test_curator_review_prompt_offers_support_file_actions asserting
skill_manage action=create and action=write_file are both named.
**Live validation on author's setup:**
- Run 1 (old prompt): 3 archives, stopped after surveying — typical passive outcome
- Run 2 (consolidation prompt): 44 archives, 3 patches, surfaced the 50-skill mlops reorg duplicate bug but didn't umbrella
- Run 3 (this prompt): 249 archives + 18 new class-level umbrellas created, reducing agent-created skills from 346 → 118 with every archived skill's content preserved as references/ under its umbrella. Pinned skill untouched. Full report in PR description.
Weekly is closer to how skill churn actually works — most agent-created
skills don't change multiple times per day, so a daily review is pure
cost without benefit. Bumping the default to 7 days reduces aux-model
spend while still catching drift and staleness on the timescales that
matter (30d stale, 90d archive).
Changes:
- DEFAULT_INTERVAL_HOURS: 24 -> 168 (7 days)
- config.yaml default: interval_hours: 24 -> 24 * 7
- CLI status line renders as '7d' when interval is a whole-day multiple
- Test `test_old_run_eligible` decoupled from the exact default: it now
uses 2 * get_interval_hours() so future tweaks don't break it
The LLM review prompt mentioned bespoke `archive_skill` and `pin_skill`
tools that are not registered as model tools. Swap the prompt to rely
on the real surface:
- skill_manage action=patch — for patching and consolidation
- terminal — to `mv` skill dirs into .archive/
Also drop `pin` from the model's decision list — pinning is a user
opt-out for `hermes curator pin <skill>`, not something the model
should do autonomously.
Decision list is now: keep / patch / consolidate / archive.
Tests updated: prompt-invariant test now asserts the existing tools
are referenced and that bespoke tool names do NOT appear. New test
prevents `pin` from being re-added as a model decision.
Adds the Curator — an auxiliary-model background task that periodically
reviews AGENT-CREATED skills and keeps the collection tidy: tracks usage,
transitions unused skills through active → stale → archived, and spawns
a forked AIAgent to consolidate overlaps and patch drift.
Default: enabled, inactivity-triggered (no cron daemon). Runs on CLI
startup and gateway boot when the last run is older than interval_hours
(default 24) AND the agent has been idle for min_idle_hours (default 2).
Invariants (all load-bearing):
- Never touches bundled or hub-installed skills (.bundled_manifest +
.hub/lock.json double-filter)
- Never auto-deletes — archive only. Archives are recoverable
via `hermes curator restore <skill>`
- Pinned skills bypass all auto-transitions
- Uses the aux client; never touches the main session's prompt cache
New files:
- tools/skill_usage.py — sidecar .usage.json telemetry, atomic writes,
provenance filter
- agent/curator.py — orchestrator: config, idle gating, state-machine
transitions (pure, no LLM), forked-agent review prompt
- hermes_cli/curator.py — `hermes curator {status,run,pause,resume,
pin,unpin,restore}` subcommand
- tests/tools/test_skill_usage.py — 29 tests
- tests/agent/test_curator.py — 25 tests
Modified files (surgical patches):
- tools/skills_tool.py — bump view_count on successful skill_view
- tools/skill_manager_tool.py — bump patch_count on skill_manage
patch/edit/write_file/remove_file; forget record on delete
- hermes_cli/config.py — add curator: section to DEFAULT_CONFIG
- hermes_cli/commands.py — add /curator CommandDef with subcommands
- hermes_cli/main.py — register `hermes curator` subparser via
register_cli() from hermes_cli.curator
- cli.py — /curator slash-command dispatch + startup hook
- gateway/run.py — gateway-boot hook (mirrors CLI)
Validation:
- 54 new tests across skill_usage + curator, all passing in 3s
- 346 tests across all touched files' neighbors green
- 2783 tests across hermes_cli/ + gateway/test_run_progress_topics.py green
- CLI smoke: `hermes curator status/pause/resume` work end-to-end
Companion to PR #16026 (class-first skill review prompt) — together
they form a loop: the review prompt stops near-duplicate skill creation
at the source, and the curator prunes/consolidates what still accumulates.
Refs #7816.
Relative entries in skills.external_dirs were resolved against the
process cwd via Path.resolve(), making them silently fail when Hermes
was launched from a different directory.
Resolve relative paths against get_hermes_home() for consistent
behavior across CLI, gateway, and cron contexts. Absolute paths
and env-var/tilde expansion are unchanged.
Three modules independently implemented the same "preserve head+tail of
a secret, mask the middle" logic with slightly different behaviors that
had started to drift:
hermes_cli/config.py redact_key — 12-char floor, 4+4, DIM '(not set)'
hermes_cli/status.py redact_key — 12-char floor, 4+4, plain '(not set)' ← drift
hermes_cli/dump.py _redact — 12-char floor, 4+4, empty string
The visible bug: 'hermes status' displayed the '(not set)' placeholder
in plain text while 'hermes config' showed it in dim text. Same concept,
inconsistent UI.
Introduces mask_secret() in agent/redact.py as the canonical helper,
with head/tail/floor/placeholder/empty kwargs. The three call sites
become one-line wrappers that differ only in the 'empty' handling:
config.redact_key → mask_secret(k, empty=color('(not set)', Colors.DIM))
status.redact_key → mask_secret(k, empty=color('(not set)', Colors.DIM))
dump._redact → mask_secret(v) # empty → ''
agent.redact._mask_token (log redactor, different policy: 18-char floor,
6+4 visible, '***' on empty) also ports to mask_secret but retains its
own empty-case handling to preserve the historical '***' return.
Net: the three display-time redactors now agree on formatting, the
canonical helper lives in one place, and future tweaks (e.g. adding
bullet-point masking, changing the head/tail widths) happen once.
Verified:
- 3/3 tests/hermes_cli/test_web_server.py::TestRedactKey pass
- 89/89 agent/tests/test_redact.py + tests/tools/test_browser_secret_exfil.py
+ tests/hermes_cli/test_redact_config_bridge.py pass
- Live 'hermes status', 'hermes config', 'hermes dump' all render the
same way they did before (verified against actual env with real
keys: OpenRouter, Firecrawl, Browserbase, FAL, Tinker all show
'prefix...suffix'; Kimi shows '***' at <12 chars; unset shows
'(not set)' uniformly).
Co-authored-by: teknium1 <teknium@users.noreply.github.com>
* perf(startup): lazy-import OpenAI, Anthropic, Firecrawl, account_usage
Four heavy SDK/module imports are now deferred off the hot startup path.
Net savings on cold module imports:
cli 1200 → 958 ms (-242)
run_agent 1220 → 901 ms (-319)
tools.web_tools 711 → 423 ms (-288)
agent.anthropic_adapter 230 → 15 ms (-215)
agent.auxiliary_client 253 → 68 ms (-185)
Four independent changes in one PR since they all use the same pattern
and share the same risk profile (heavy SDK import → lazy proxy or
function-local import):
1. tools/web_tools.py:
'from firecrawl import Firecrawl' moved into _get_firecrawl_client(),
which is only called when backend='firecrawl'. Users on Exa/Tavily/
Parallel pay zero firecrawl cost.
2. cli.py + gateway/run.py:
'from agent.account_usage import ...' moved into the /limits handlers.
account_usage transitively pulls the OpenAI SDK chain; only needed
when the user runs /limits.
3. agent/anthropic_adapter.py:
'try: import anthropic as _anthropic_sdk' replaced with a cached
'_get_anthropic_sdk()' accessor. The three usage sites
(build_anthropic_client, build_anthropic_bedrock_client,
read_claude_code_credentials_from_keychain) now resolve via the
accessor. All pre-existing test patches of
'agent.anthropic_adapter._anthropic_sdk' keep working because the
accessor respects any value already in module globals.
4. agent/auxiliary_client.py AND run_agent.py:
'from openai import OpenAI' replaced with an '_OpenAIProxy()' module-
level object that looks like the OpenAI class but imports the SDK on
first call/isinstance check. This preserves:
- 15+ in-module OpenAI(...) construction sites in auxiliary_client
and the single site in run_agent's _create_openai_client (Python's
function-scope name lookup finds the proxy, forwards the call);
- 'patch("agent.auxiliary_client.OpenAI", ...)' and
'patch("run_agent.OpenAI", ...)' test patterns used by 28+ test
files (patch replaces the module attribute as usual).
Tried two alternatives first:
- 'from openai._client import OpenAI' — doesn't skip openai/__init__.py
(the audit's hypothesis here was wrong).
- Module-level __getattr__ — works for external access but Python
function-scope name resolution skips __getattr__, so in-module
OpenAI(...) calls NameError.
Note: 'openai' still loads on 'import cli' because
cli.py -> neuter_async_httpx_del() -> openai._base_client, and
run_agent.py -> code_execution_tool.py (module-level
build_execute_code_schema) -> _load_config() -> 'from cli import
CLI_CONFIG'. Deferring those is a separate, larger change — out of scope
for this PR. The savings above all come from avoiding the openai/*,
anthropic/*, and firecrawl/* top-level type-tree imports on paths that
don't need them.
Verified:
- 302/302 tests in tests/agent/{test_anthropic_adapter,
test_bedrock_1m_context, test_minimax_provider, test_anthropic_keychain}
pass. Two pre-existing failures on main unchanged.
- 106/106 tests/agent/test_auxiliary_client.py pass (1 pre-existing fail).
- 97/97 tests/run_agent/test_create_openai_client_kwargs_isolation.py,
test_plugin_context_engine_init.py, test_invalid_context_length_warning.py,
test_api_max_retries_config.py,
tests/hermes_cli/test_gemini_provider.py, test_ollama_cloud_provider.py
pass (1 pre-existing fail).
- Live hermes chat smoke: 2 turns + /model switch + tool calls, zero
errors in the 57-line agent.log window.
- Module-level import of run_agent + auxiliary_client + anthropic_adapter
no longer pulls 'anthropic' or 'firecrawl' at all.
* fix(gateway): restore top-level account_usage import for test-patch surface
CI caught two failures in tests/gateway/test_usage_command.py that I
missed locally:
AttributeError: 'module' object at gateway.run has no attribute 'fetch_account_usage'
The test uses monkeypatch.setattr('gateway.run.fetch_account_usage', ...)
to inject a fake account-fetch call. Moving the import inside the
handler deleted that module-level attribute, breaking the patch surface.
Restoring the top-level import in gateway/run.py gives up the ~230 ms
gateway-boot savings from that one lazy, but:
1. the gateway is a long-running daemon — boot cost is paid once per
install, not per turn;
2. the other four lazy-imports (firecrawl, openai, anthropic, cli's
account_usage) remain in place and still account for the bulk of
the savings reported in the PR body;
3. preserving the patch surface keeps the established
'gateway.run.fetch_account_usage' monkeypatch pattern working
without touching tests.
Verified: tests/gateway/test_usage_command.py — 8 passed, 0 failed.
Full targeted sweep (2336 tests across agent/gateway/hermes_cli/run_agent):
2332 passed, 4 failed — all 4 pre-existing on main.
---------
Co-authored-by: teknium1 <teknium@users.noreply.github.com>
Auxiliary tasks (title_generation, vision, compression, web_extract,
session_search) now pick the correct wire protocol based on the
endpoint, not just on which resolve_provider_client branch built the
client. Fixes 404s on Kimi Coding Plan and any other named provider
whose endpoint speaks Anthropic Messages.
Root cause: the 'api_key' branch of resolve_provider_client (and the
Step 2 fallback chain inside _resolve_auto) always built a plain
OpenAI client regardless of what the endpoint actually spoke. For
provider=kimi-coding + model=kimi-for-coding, that meant:
POST https://api.kimi.com/coding/v1/chat/completions
{ "model": "kimi-for-coding", ... }
→ 404 resource_not_found_error
The /coding route only accepts the Anthropic Messages shape (the main
agent already uses api_mode=anthropic_messages for it). Earlier fixes
(#16819, #22ddac4b1) patched the anonymous-custom, named-custom, and
external-process branches — but the named api_key branch (kimi-coding,
minimax, zai, future /anthropic providers) was the fourth sibling and
never got the same treatment.
Fix: one module-level helper _maybe_wrap_anthropic() that rewraps a
plain OpenAI client in AnthropicAuxiliaryClient when:
- api_mode is explicitly 'anthropic_messages', OR
- the URL ends in '/anthropic', OR
- the host is api.kimi.com + path contains '/coding', OR
- the host is api.anthropic.com.
Wired into _wrap_if_needed (covers all resolve_provider_client
branches that already go through it) and into the Step 2 api_key
fallback chain inside _resolve_auto. Explicit api_mode still wins:
passing api_mode='chat_completions' forces OpenAI wire, and already-
wrapped specialized adapters (Codex, Gemini native, CopilotACP) pass
through unchanged.
E2E verified:
- resolve_provider_client('kimi-coding', 'kimi-for-coding')
→ AnthropicAuxiliaryClient (was plain OpenAI, which 404'd)
- _resolve_auto Step 1 for kimi-coding runtime → AnthropicAuxiliaryClient
- resolve_provider_client('openrouter', ...) → plain OpenAI (no regression)
- api_mode='chat_completions' override → plain OpenAI (explicit wins)
Tests:
- tests/agent/test_auxiliary_transport_autodetect.py (new): 21 tests
covering URL detection, wrap decisions, and integration.
- 204/205 existing auxiliary tests pass (1 pre-existing failure on
main, unrelated to this change).
Co-authored-by: teknium1 <teknium@users.noreply.github.com>
Mechanical cleanup across 43 files — removes 46 unused imports
(F401) and 14 unused local variables (F841) detected by
`ruff check --select F401,F841`. Net: -49 lines.
Also fixes a latent NameError in rl_cli.py where `get_hermes_home()`
was called at module line 32 before its import at line 65 — the
module never imported successfully on main. The ruff audit surfaced
this because it correctly saw the symbol as imported-but-unused
(the call happened before the import ran); the fix moves the import
to the top of the file alongside other stdlib imports.
One `# noqa: F401` kept in hermes_cli/status.py for `subprocess`:
tests monkeypatch `hermes_cli.status.subprocess` as a regression
guard that systemctl isn't called on Termux, so the name must
exist at module scope even though the module body doesn't reference
it. Docstring explains the reason.
Also fixes an invalid `# noqa:` directive in
gateway/platforms/discord.py:308 that lacked a rule code.
Co-authored-by: teknium1 <teknium@users.noreply.github.com>
Auxiliary callers that configure reasoning via
auxiliary.<task>.extra_body.reasoning were having that config silently
dropped by the Codex Responses adapter — it only forwarded
messages/model/tools through to responses.stream(), never translating
chat.completions-shaped reasoning hints into the Responses API's
top-level reasoning + include fields.
Mirror the main-agent translation from agent/transports/codex.py:
- extra_body.reasoning.effort → resp_kwargs.reasoning.{effort, summary:"auto"}
- 'minimal' → 'low' clamp (Codex backend rejects 'minimal')
- Always include ['reasoning.encrypted_content'] when reasoning is enabled
- {'enabled': False} → omit reasoning and include entirely
- Non-dict reasoning values are ignored defensively
Reported by @OP (Apr 26 feedback bundle).
## Changes
- agent/auxiliary_client.py: _CodexCompletionsAdapter.create() now reads
and translates extra_body.reasoning before calling responses.stream()
- tests/agent/test_auxiliary_client.py: 9 new tests covering all effort
levels, the minimal→low clamp, the disabled path, the no-op paths,
and defensive handling of wrong-shape inputs
Co-authored-by: teknium1 <teknium@users.noreply.github.com>
When openai-codex tokens expire or the ChatGPT account hits a 429
window, the pool entry gets marked STATUS_EXHAUSTED with
last_error_reset_at many hours in the future. If the user then runs
`hermes model` / `hermes auth openai-codex` to reauth, fresh tokens
land in ~/.hermes/auth.json but the pool entry stayed frozen behind
its reset_at — every request kept failing with 'credential pool: no
available entries (all exhausted or empty)' until the original window
elapsed.
_available_entries() already had auth.json/credentials-file resync
branches for anthropic/claude_code and nous/device_code; openai-codex
was missing. Added _sync_codex_entry_from_auth_store() mirroring the
nous version (reads state["tokens"][{access,refresh}_token] +
state["last_refresh"]) and wired it into the exhausted-entry resync
loop.
Also softens the 'codex CLI not found' doctor warning — native
device-code OAuth does not require the Codex binary, only
importing existing Codex CLI tokens does. Downgraded to an info line.
Reported on Discord by p1aceho1der: Codex stalled indefinitely after
a rate-limit reset, reauth didn't help, and doctor falsely warned
that the codex CLI was required.
Co-authored-by: teknium1 <teknium@users.noreply.github.com>
Gemini 3 Flash documents low/medium/high as the accepted thinkingLevel
values. The salvaged bridge was forwarding Hermes' "minimal" effort to
Flash verbatim, which is not a documented Gemini level and risks a 400
from the native adapter.
Clamp minimal->low on Flash (matching how Pro already clamps minimal+low
down), and funnel anything outside {low, medium, high} into medium to
keep the request valid by construction. No behaviour change for the
documented effort levels.
Telegram has no native table syntax. The gateway auto-rewrites pipe
tables into row-group bullets (see previous commit), but letting models
know up front means they emit the clean form directly instead of
relying on post-processing to synthesize headings.
Also helps users whose MEMORY.md formatting policies were being
overridden — the platform hint now carries the guidance.
Extract the islink/realpath guard from the 16743 fix into a single
atomic_replace() helper in utils.py, then migrate every os.replace()
call site in the codebase to use it.
The original PR #16777 correctly identified and fixed the bug, but
only patched 9 of ~24 call sites. The same bug class (managed
deployments that symlink state files silently losing the link on
every write) still existed at auth.json, sessions file, gateway
config, env_loader, webhook subscriptions, debug store, model
catalog, pairing, google OAuth, nous rate guard, and more.
Rather than add another 10+ copies of the same three-line guard,
consolidate into atomic_replace(tmp, target) which:
- resolves symlinks via os.path.realpath before os.replace
- returns the resolved real path so callers can re-apply permissions
- is a drop-in replacement for os.replace at the use sites
Changes:
- utils.py: new atomic_replace() helper + atomic_json_write /
atomic_yaml_write now call it instead of inlining the guard
- 16 files: all os.replace() call sites migrated to atomic_replace()
- agent/{google_oauth, nous_rate_guard, shell_hooks}.py
- cron/jobs.py
- gateway/{pairing, session, platforms/telegram}.py
- hermes_cli/{auth, config, debug, env_loader, model_catalog, webhook}.py
- tools/{memory_tool, skill_manager_tool, skills_sync}.py
Tests: tests/test_atomic_replace_symlinks.py pins the invariant for
atomic_replace + atomic_json_write + atomic_yaml_write, covers plain
files, first-time creates, broken symlinks, and permission preservation.
Refs #16743
Builds on #16777 by @vominh1919.
os.replace(tmp, path) replaces the symlink itself with a regular file,
breaking users who symlink config.yaml, SOUL.md, or .env from ~/.hermes/
to a dotfiles repo or managed profile package.
Fix: resolve symlinks via os.path.realpath() before os.replace(), so the
real file is overwritten in-place while the symlink survives.
Fixed in 7 files covering all os.replace call sites:
- utils.py (atomic_json_write, atomic_yaml_write — fixes save_config)
- hermes_cli/config.py (env sanitizer, save_env_value, remove_env_value)
- tools/skill_manager_tool.py (_atomic_write_text — SOUL.md writes)
- tools/memory_tool.py (memory file writes)
- tools/skills_sync.py (manifest writes)
- cron/jobs.py (job state + output file writes)
- agent/shell_hooks.py (hook file writes)
FixesNousResearch/hermes-agent#16743
DeepSeek API returns HTTP 400 with 'Insufficient Balance' message when
account funds are depleted. This pattern was not in _BILLING_PATTERNS,
causing the error to be misclassified instead of triggering billing
exhaustion handling (e.g., fallback to alternate provider).
Suggested by teknium1 in PR review of #15586.
Adds tools.schema_sanitizer.strip_nullable_unions as the single
implementation for collapsing anyOf/oneOf nullable unions. Both the
MCP input-schema normalizer and the Anthropic tool-schema guard now
delegate to it instead of re-implementing the same walk three times.
The global sanitizer also gains a final pass so any tool that slips
past the two earlier hooks (plugin tools, non-MCP custom tools with
Pydantic-shaped schemas) still gets safe input_schemas on Anthropic.
- tools/schema_sanitizer.py:
* New public strip_nullable_unions(schema, keep_nullable_hint=True).
* _sanitize_single_tool() calls it as a final pass (hint preserved
so coerce_tool_args can still map string "null" to None).
- tools/mcp_tool.py: _normalize_mcp_input_schema delegates.
- agent/anthropic_adapter.py: _normalize_tool_input_schema delegates
with keep_nullable_hint=False (Anthropic does not recognize nullable).
No behavioral change for the fix itself; tests (73/73 targeted +
E2E across MCP→sanitizer→Anthropic paths) pass.
provider_model_ids("bedrock") fell through to a static _PROVIDER_MODELS
table containing only hardcoded us.* model IDs. Users configured for
non-US AWS regions (eu-central-1, ap-northeast-1, etc.) saw wrong or no
models in /model and autocomplete.
Root causes fixed:
1. models.py: provider_model_ids() now calls discover_bedrock_models()
keyed by the resolved region before falling back to the static table.
A new bedrock_model_ids_or_none() helper in bedrock_adapter.py
consolidates the discover -> extract IDs -> fallback pattern used by
all three call sites.
2. providers.py: registers bedrock in HERMES_OVERLAYS with
transport=bedrock_converse and auth_type=aws_sdk so
get_provider("bedrock") and resolve_provider_full("bedrock") work.
3. model_switch.py: list_authenticated_providers() sections 2 and 3
detect AWS credentials via has_aws_credentials() for aws_sdk
overlays and use live discovery for the model list.
4. bedrock_adapter.py: resolve_bedrock_region() reads the configured
region from botocore.session before falling back to us-east-1,
covering users who set their region in ~/.aws/config via a named
profile rather than env vars.
5. tui_gateway/server.py: passes provider= to get_model_context_length()
so context window lookups work correctly for the Bedrock provider.
* fix(anthropic): remove Claude Code fingerprinting from OAuth Messages API path
OAuth requests now identify as Hermes on the wire. Removed:
- "You are Claude Code, Anthropic's official CLI for Claude." system
prompt prepend
- Hermes Agent → Claude Code / Nous Research → Anthropic
system-prompt substitutions
- mcp_ tool-name prefix on outgoing tool schemas + message history
- Matching mcp_ strip on inbound tool_use blocks (strip_tool_prefix path
removed from AnthropicTransport.normalize_response, + all 5 call
sites in run_agent.py and auxiliary_client.py)
- user-agent: claude-cli/<v> (external, cli) and x-app: cli headers on
the Messages API client
Added:
- OAuth path strips context-1m-2025-08-07 — Anthropic rejects OAuth
requests carrying it with HTTP 400 'This authentication style is
incompatible with the long context beta header.'
Kept (auth plumbing, not identity spoofing):
- _is_oauth_token classifier and is_oauth flag threading
- Bearer vs x-api-key auth routing
- _OAUTH_ONLY_BETAS (claude-code-20250219, oauth-2025-04-20) — backend
requires these on the OAuth-gated Messages endpoint
- _OAUTH_CLIENT_ID (Claude Code's) — Anthropic doesn't issue OAuth
creds to third parties; this is the only way the login flow works
- claude-cli/<v> User-Agent on the OAuth token exchange + refresh
endpoints at platform.claude.com/v1/oauth/token — bare requests get
Cloudflare 1010 blocked
Verified live against api.anthropic.com with a fresh sk-ant-oat01-*
token:
- claude-haiku-4-5 simple message: HTTP 200, 'OK' response
- claude-haiku-4-5 tool call: HTTP 200, stop_reason=tool_use, tool
named 'terminal' (no mcp_ prefix) round-tripped correctly
- Outgoing wire: no user-agent, no x-app, real Hermes identity in
system prompt, real tool name in schema
Closes/supersedes #16820 (mcp_ PascalCase normalization patch — no longer
needed since the mcp_ round-trip is gone).
* fix(anthropic): resolve_anthropic_token() reads credential pool first
Close the gap where ~/.hermes/auth.json → credential_pool.anthropic
(where hermes login + dashboard PKCE flow write OAuth tokens) was not
in resolve_anthropic_token()'s source list.
Before: users who authed via hermes login got the token written into
the pool, but legacy fallback code paths (auxiliary_client, models
catalog fetch, explicit-runtime path) that call resolve_anthropic_token()
saw None and raised 'No Anthropic credentials found' — even though the
token was sitting in auth.json.
New priority 1: pool.select() with env-sourced entries skipped. Skipping
env:* entries preserves the existing env-var priority logic further
down the chain (static env OAuth → refreshable Claude Code upgrade via
_prefer_refreshable_claude_code_token).
Surfaced while writing the hermes-agent-dev skill playbook for
'finding a live OAuth token for an E2E test'.
---------
Co-authored-by: teknium1 <teknium@users.noreply.github.com>
Registers tencent-tokenhub (https://tokenhub.tencentmaas.com/v1) as a
new API-key provider with model tencent/hy3-preview (256K context).
- PROVIDER_REGISTRY entry + TOKENHUB_API_KEY / TOKENHUB_BASE_URL env vars
- Aliases: tencent, tokenhub, tencent-cloud, tencentmaas
- openai_chat transport with is_tokenhub branch for top-level
reasoning_effort (Hy3 is a reasoning model)
- tencent/hy3-preview:free added to OpenRouter curated list
- 60+ tests (provider registry, aliases, runtime resolution,
credentials, model catalog, URL mapping, context length)
- Docs: integrations/providers.md, environment-variables.md,
model-catalog.json
Author: simonweng <simonweng@tencent.com>
Salvaged from PR #16860 onto current main (resolved conflicts with
#16935 Azure Anthropic env-var hint tests and the --provider choices=
list removal in chat_parser).
Follow-up to PR #16819 applying the same treatment to the two sibling
fallback sites in resolve_provider_client() that carry the identical bug
class as the anonymous-custom branch:
- Named custom provider (providers: / custom_providers: config entries):
apply _to_openai_base_url() on the OpenAI-wire path (chat_completions /
codex_responses), leave custom_base untouched on the anthropic_messages
path where the /anthropic surface is intentional. Prefer
main_runtime.get('model') over _read_main_model() so the entry model
still wins first. The ImportError fallback for anthropic_messages now
redoes query-param extraction against the rewritten URL so the final
OpenAI client hits /v1.
- external_process branch (copilot-acp): same main_runtime.get('model')
fallback before _read_main_model() so auxiliary tasks on this provider
track live /model switches instead of stale config.yaml.
Keeps the fix consistent across all three custom-endpoint fallback sites
in resolve_provider_client().
Extends the cua-driver computer-use backend to drive backgrounded macOS
windows without stealing keyboard or mouse focus from the foreground app.
All changes target the cua-driver MCP backend and the shared dispatcher.
## cua_backend.py
**Window-aware capture**: capture() now calls list_windows + get_window_state
instead of the removed capture tool. Prefers structuredContent.windows
(MCP 2024-11-05+ cua-driver) for zero-parse window enumeration; falls back
to regex-parsed text for older builds. Stores the selected (pid, window_id)
as sticky context so subsequent action calls do not need a redundant round-trip.
**Action routing**: click/scroll/type_text/key all carry the sticky pid
(and window_id for element-indexed clicks). type_text routes through
type_text_chars (individual key events) rather than AX attribute write --
WebKit AXTextFields reject attribute writes from backgrounded processes.
**Key parsing**: _parse_key_combo splits cmd+s-style strings into
(key, [modifiers]) and routes to hotkey (modifier present) or
press_key (bare key) -- cua-driver actual tool names.
**set_value method**: new set_value(value, element) calls the cua-driver
set_value MCP tool. For AXPopUpButton / HTML select in a backgrounded Safari,
AXPress opens the native macOS popup which closes immediately when the app is
non-frontmost; set_value AX-presses the matching child option directly
(no menu required, no focus steal).
**focus_app**: reimplemented as a pure window-selector (enumerates
list_windows, sets sticky pid/window_id) without ever raising the window
or stealing focus.
**list_apps**: fixed tool name from listApps to list_apps; handles plain-text
response via regex when structured data is absent.
**Structured-content extraction**: _extract_tool_result now surfaces
structuredContent from MCP results, enabling the list_windows window array
without text parsing.
**Helpers**: _parse_windows_from_text, _parse_elements_from_tree,
_split_tree_text, _parse_key_combo extracted as module-level functions.
## schema.py
Added set_value to the action enum with a description explaining when to
prefer it over click (select/popup elements, sliders, no focus steal).
Added value field for set_value payloads.
## tool.py
Routed set_value action through _dispatch to backend.set_value.
Added set_value to _DESTRUCTIVE_ACTIONS (approval-gated).
Fixed MIME-type detection in _capture_response: cua-driver may return
JPEG; detect from base64 magic bytes (/9j/ -> image/jpeg, else image/png)
rather than hardcoding image/png.
## agent/display.py + run_agent.py
Guard _detect_tool_failure and result-preview logic against non-string
function_result values: multimodal tool results (dicts with _multimodal=True)
are not string-sliceable; treat them as successes and fall back to str()
for length/preview.
Background macOS desktop control via cua-driver MCP — does NOT steal the
user's cursor or keyboard focus, works with any tool-capable model.
Replaces the Anthropic-native `computer_20251124` approach from the
abandoned #4562 with a generic OpenAI function-calling schema plus SOM
(set-of-mark) captures so Claude, GPT, Gemini, and open models can all
drive the desktop via numbered element indices.
- `tools/computer_use/` package — swappable ComputerUseBackend ABC +
CuaDriverBackend (stdio MCP client to trycua/cua's cua-driver binary).
- Universal `computer_use` tool with one schema for all providers.
Actions: capture (som/vision/ax), click, double_click, right_click,
middle_click, drag, scroll, type, key, wait, list_apps, focus_app.
- Multimodal tool-result envelope (`_multimodal=True`, OpenAI-style
`content: [text, image_url]` parts) that flows through
handle_function_call into the tool message. Anthropic adapter converts
into native `tool_result` image blocks; OpenAI-compatible providers
get the parts list directly.
- Image eviction in convert_messages_to_anthropic: only the 3 most
recent screenshots carry real image data; older ones become text
placeholders to cap per-turn token cost.
- Context compressor image pruning: old multimodal tool results have
their image parts stripped instead of being skipped.
- Image-aware token estimation: each image counts as a flat 1500 tokens
instead of its base64 char length (~1MB would have registered as
~250K tokens before).
- COMPUTER_USE_GUIDANCE system-prompt block — injected when the toolset
is active.
- Session DB persistence strips base64 from multimodal tool messages.
- Trajectory saver normalises multimodal messages to text-only.
- `hermes tools` post-setup installs cua-driver via the upstream script
and prints permission-grant instructions.
- CLI approval callback wired so destructive computer_use actions go
through the same prompt_toolkit approval dialog as terminal commands.
- Hard safety guards at the tool level: blocked type patterns
(curl|bash, sudo rm -rf, fork bomb), blocked key combos (empty trash,
force delete, lock screen, log out).
- Skill `apple/macos-computer-use/SKILL.md` — universal (model-agnostic)
workflow guide.
- Docs: `user-guide/features/computer-use.md` plus reference catalog
entries.
44 new tests in tests/tools/test_computer_use.py covering schema
shape (universal, not Anthropic-native), dispatch routing, safety
guards, multimodal envelope, Anthropic adapter conversion, screenshot
eviction, context compressor pruning, image-aware token estimation,
run_agent helpers, and universality guarantees.
469/469 pass across tests/tools/test_computer_use.py + the affected
agent/ test suites.
- `model_tools.py` provider-gating: the tool is available to every
provider. Providers without multi-part tool message support will see
text-only tool results (graceful degradation via `text_summary`).
- Anthropic server-side `clear_tool_uses_20250919` — deferred;
client-side eviction + compressor pruning cover the same cost ceiling
without a beta header.
- macOS only. cua-driver uses private SkyLight SPIs
(SLEventPostToPid, SLPSPostEventRecordTo,
_AXObserverAddNotificationAndCheckRemote) that can break on any macOS
update. Pin with HERMES_CUA_DRIVER_VERSION.
- Requires Accessibility + Screen Recording permissions — the post-setup
prints the Settings path.
Supersedes PR #4562 (pyautogui/Quartz foreground backend, Anthropic-
native schema). Credit @0xbyt4 for the original #3816 groundwork whose
context/eviction/token design is preserved here in generic form.
Flips security.redact_secrets from true to false in DEFAULT_CONFIG, and
the HERMES_REDACT_SECRETS env-var fallback in agent/redact.py now
requires explicit opt-in ("1"/"true"/"yes"/"on") to enable.
New installs and users without a security.redact_secrets key get pass-
through tool output. Existing users whose config.yaml explicitly sets
redact_secrets: true keep redaction on — the config-yaml -> env-var
bridges in hermes_cli/main.py and gateway/run.py still honor their
setting.
Also updates the inline config comments, website docs, and the
hermes-agent skill so /hermes config set security.redact_secrets true
is now the documented way to turn it on.
On AWS Bedrock (and Azure AI Foundry), Claude Opus 4.6/4.7 and Sonnet 4.6
are capped at 200K context unless the request carries the
`context-1m-2025-08-07` beta header. On native Anthropic (api.anthropic.com)
1M went GA so the header is a harmless no-op, but Bedrock/Azure still gate
it as beta as of 2026-04.
Hermes was advertising 1M in model_metadata.py (`claude-opus-4-7: 1000000`)
while silently sending a request without the beta — so Bedrock users saw
a 200K ceiling with no error message, and no config knob unblocked it.
Claude Code sends this header by default, which is why the same Bedrock
credentials worked there.
- Add `context-1m-2025-08-07` to `_COMMON_BETAS` (alongside interleaved
thinking and fine-grained tool streaming).
- Strip it in `_common_betas_for_base_url` for MiniMax bearer-auth
endpoints — they host their own models, not Claude, so Anthropic beta
headers are irrelevant and could risk rejection.
- Attach `_COMMON_BETAS` as `default_headers` on the AnthropicBedrock
client. Previously that constructor passed no betas at all, so native
Anthropic had the 1M unlock via default_headers but Bedrock didn't.
- Fast-mode per-request `extra_headers` already rebuilds from
`_common_betas_for_base_url`, so it picks up the 1M beta automatically.
Reported by user 'Rodmar' on Discord: Bedrock Opus 4.7 stuck at 200K while
same credentials worked in Claude Code.
A misconfigured auxiliary.compression.model is a user-fixable problem that silent recovery would hide. The previous retry-on-main logic transparently swallowed aux-model failures whenever the fallback succeeded, leaving the user's broken config in place and racking up future failures.
Track the aux-model failure on the compressor alongside the existing fallback-placeholder fields:
- _last_aux_model_failure_model: str | None
- _last_aux_model_failure_error: str | None
Both are set at the moment the aux model errors (captured before summary_model is cleared for retry), regardless of whether the retry succeeds. Cleared at compress() start and on on_session_reset() so a clean run doesn't leak stale warnings.
Surface at three places:
- gateway hygiene auto-compress: ℹ note to the platform adapter (thread_id preserved)
- gateway /compress command: ℹ line appended to the reply
- CLI via _emit_warning: deduped on (model, error) so repeat compactions don't spam
Distinct from the existing ⚠️ dropped-turns warning — different severity, different emoji, explicit 'context is intact' reassurance.
The existing retry-on-main path in _generate_summary only fires for errors that match the _is_model_not_found heuristic (404/503, 'model_not_found', 'does not exist', 'no available channel'). Other misconfiguration errors — 400s from aggregators, provider-specific 'no route' strings, opaque rejections — fall straight through to the transient-cooldown branch, which drops N turns of context and inserts a static placeholder.
Losing context is almost always worse than one extra summary attempt. Add a best-effort retry-on-main for the unknown-error branch, guarded by the same invariants as the existing fast-path retry: only when summary_model differs from main, and only once per compressor (_summary_model_fallen_back).
Tests cover: 404 fast-path fallback still works, unknown 400 now falls back, same-model aux skips retry (no infinite loop), and a double-failure (aux + main) stops at 2 calls.
The per-call reset block at the top of compress() cleared
_last_summary_dropped_count and _last_summary_fallback_used but
not _last_summary_error. Functionally this didn't break the
gateway warning path (callers gate on _last_summary_fallback_used
first, and _last_summary_error is overwritten on the next failure),
but it left the three tracking fields inconsistent — anyone
reading _last_summary_error standalone after a successful compress
would see a stale value from a previous failed compress.
Reset all three together so the per-call contract is uniform.
The fallback placeholder said "N conversation turns were removed" while the
gateway warning said "N historical message(s) were removed". Use "messages"
in both so users don't wonder if the two counters refer to different things.
When auxiliary compression's summary LLM call fails (e.g. model 404,
auxiliary model misconfigured), the compressor still drops the selected
turns and inserts a static fallback placeholder — the dropped context
is unrecoverable.
Previously the only signal of this was a WARNING in agent.log. Gateway
users (Telegram/Discord/etc.) had no way to know context was lost
because the existing _emit_warning path requires a status_callback,
and the gateway hygiene path uses a temporary _hyg_agent with
quiet_mode=True and no callback wired up.
Changes:
- ContextCompressor: track _last_summary_fallback_used and
_last_summary_dropped_count on each compress() call. Cleared at the
start of compress() and on session reset.
- gateway/run.py hygiene: after auto-compress, inspect the temp
agent's compressor; if fallback was used, send a visible ⚠️ warning
to the user via the platform adapter (TG/Discord/etc.) including
dropped count and the underlying error.
- gateway/run.py /compress: append the same warning to the manual
compress reply so users running /compress see the failure too.
Acceptance:
- Summary success: no user-visible warning (unchanged).
- Summary failure on gateway hygiene: user receives a TG/Discord
message with dropped count + error + remediation hint.
- Summary failure on /compress: warning appended to the command reply.
- CLI status_callback / _emit_warning path is untouched.
- Test coverage: two new tests verify the tracking fields are set on
failure and cleared on subsequent success.
Reviewer pushback on the original boundary-hardening commits — three
overreach points pulled plugin-specific policy into shared core paths:
1. gateway/run.py hardcoded a '## Honcho Context' literal split for
vision-LLM output. Plugin-format heading in framework code; could
truncate legitimate output naturally containing that header.
Drop the literal split; keep generic sanitize_context (the wrapper
strip is plugin-agnostic). Plugin-specific cleanup belongs at the
provider boundary, not the shared gateway path.
2. run_agent.run_conversation scrubbed user_message and
persist_user_message before the conversation loop. User text is
sacred — if a user types a literal <memory-context> tag we must
not silently delete it. The producer (build_memory_context_block)
is the only legitimate emitter; user input should never need the
reverse op.
3. _build_assistant_message scrubbed model output before persistence.
Same hazard: would silently mutate legitimate documentation/code
the model emits containing the literal markers. The streaming
scrubber catches real leaks delta-by-delta before content is
concatenated; persist-time scrub was redundant belt-and-suspenders.
4. _fire_stream_delta stripped leading newlines from every delta unless
a paragraph break flag was set. Mid-stream '\n' is legitimate
markdown — lists, code fences, paragraph breaks — and chunk
boundaries are arbitrary. Narrow lstrip to the very first delta
of the stream only (so stale provider preamble still gets cleaned
on turn start, but mid-stream formatting survives).
Plus: build_memory_context_block now logs a warning when its defensive
sanitize_context strips something — surfaces buggy providers returning
pre-wrapped text instead of silently double-fencing.
Net architectural change: scrub surface collapses from 8 sites to 3
(StreamingContextScrubber on output deltas, plugin→backend send,
build_memory_context_block input-validation). Plugin-specific strings
stay out of shared runtime paths. User input and persisted assistant
output are no longer mutated.
Tests: rescoped TestMemoryContextSanitization (helper-correctness only,
no source-inspection of removed call sites), updated vision tests to
drop '## Honcho Context' literal-split assertions, updated
_build_assistant_message persistence test to assert preservation.
Added: cross-turn scrubber reset, build_memory_context_block warn-on-
violation, mid-stream newline preservation (plain + code fence).
sanitize_context() uses a non-greedy block regex that needs both
<memory-context> open and close tags present in a single string. When a
provider streams the fenced memory block across multiple deltas (typical
for recalled-context leaks — the payload often arrives in 10+ 1-80 char
chunks), the per-delta sanitize stripped the lone open/close tags via
_FENCE_TAG_RE but let the payload in between flow straight to the UI.
Adds StreamingContextScrubber: a small stateful scrubber that tracks
open/close tag pairs across deltas, holds back partial-tag tails at
chunk boundaries, and discards span contents wholesale (including the
system-note line that fragments across deltas).
Wired into _fire_stream_delta; reset per user turn; benign trailing
partial-tag tails are flushed at the end of each model call. Mid-span
interruption (provider drops closing tag) drops the orphaned content
rather than leaking it — truncated answer > leaked memory.
Follow-up to #13672 (@dontcallmejames).
- config.py: remove dead ENV_VARS_BY_VERSION[17] entry (current _config_version
is 22, so all users are past version 17 and would never be prompted for
GMI_API_KEY on upgrade — consistent with how arcee was added)
- auxiliary_client.py: use google/gemini-3.1-flash-lite-preview as GMI aux
model instead of anthropic/claude-opus-4.6 (matches cheap fast-model pattern
used by all other providers: zai→glm-4.5-flash, kimi→kimi-k2-turbo-preview,
stepfun→step-3.5-flash, kilocode→google/gemini-3-flash-preview)
- test_gmi_provider.py: fix malformed write_text() call in doctor test
(was: write_text("GMI_API_KEY=*** encoding="utf-8") → missing closing quote,
wrote literal string 'GMI_API_KEY=*** encoding=' to .env file)
- test_gmi_provider.py + test_auxiliary_client.py: update aux model assertions
to match new cheaper default
- docs/integrations/providers.md: add 'gmi' to inline 'Supported providers'
fallback list (was only in the table, not the inline list at line ~1181)
- docs/reference/cli-commands.md: add 'gmi' to --provider choices list
Thread a vision-request flag through auxiliary provider resolution so Copilot clients can include Copilot-Vision-Request only for vision tasks. This preserves normal text requests while ensuring Copilot vision payloads reach the vision-capable route.
Add regression coverage for Copilot vision routing and keep cached text and vision clients separate so a text client without the header is not reused for vision.
Co-authored-by: dhabibi <9087935+dhabibi@users.noreply.github.com>
* feat(image-input): native multimodal routing based on model vision capability
Attach user-sent images as OpenAI-style content parts on the user turn when
the active model supports native vision, so vision-capable models see real
pixels instead of a lossy text description from vision_analyze.
Routing decision (agent/image_routing.py::decide_image_input_mode):
agent.image_input_mode = auto | native | text (default: auto)
In auto mode:
- If auxiliary.vision.provider/model is explicitly configured, keep the
text pipeline (user paid for a dedicated vision backend).
- Else if models.dev reports supports_vision=True for the active
provider/model, attach natively.
- Else fall back to text (current behaviour).
Call sites updated: gateway/run.py (all messaging platforms), tui_gateway
(dashboard/Ink), cli.py (interactive /attach + drag-drop).
run_agent.py changes:
- _prepare_anthropic_messages_for_api now passes image parts through
unchanged when the model supports vision — the Anthropic adapter
translates them to native image blocks. Previous behaviour
(vision_analyze → text) only runs for non-vision Anthropic models.
- New _prepare_messages_for_non_vision_model mirrors the same contract
for chat.completions and codex_responses paths, so non-vision models
on any provider get text-fallback instead of failing at the provider.
- New _model_supports_vision() helper reads models.dev caps.
vision_analyze description rewritten: positions it as a tool for images
NOT already visible in the conversation (URLs, tool output, deeper
inspection). Prevents the model from redundantly calling it on images
already attached natively.
Config default: agent.image_input_mode = auto.
Tests: 35 new (test_image_routing.py + test_vision_aware_preprocessing.py),
all existing tests that reference _prepare_anthropic_messages_for_api
still pass (198 targeted + new tests green).
* feat(image-input): size-cap + resize oversized images, charge image tokens in compressor
Two follow-ups that make the native image routing safer for long / heavy
sessions:
1) Oversize handling in build_native_content_parts:
- 20 MB ceiling per image (matches vision_tools._MAX_BASE64_BYTES,
the most restrictive provider — Gemini inline data).
- Delegates to vision_tools._resize_image_for_vision (Pillow-based,
already battle-tested) to downscale to 5 MB first-try.
- If Pillow is missing or resize still overshoots, the image is
dropped and reported back in skipped[]; caller falls back to text
enrichment for that image.
2) Image-token accounting in context_compressor:
- New _IMAGE_TOKEN_ESTIMATE = 1600 (matches Claude Code's constant;
within the realistic range for Anthropic/GPT-4o/Gemini billing).
- _content_length_for_budget() helper: sums text-part lengths and
charges _IMAGE_CHAR_EQUIVALENT (1600 * 4 chars) per image/image_url/
input_image part. Base64 payload inside image_url is NOT counted
as chars — dimensions don't matter, only image-presence.
- Both tail-cut sites (_prune_old_tool_results L527 and
_find_tail_cut_by_tokens L1126) now call the helper so multi-image
conversations don't slip past compression budget.
Tests: 9 new in test_image_routing.py (oversize triggers resize,
resize-fails-returns-None, oversize-skipped-reported), 11 new in
test_compressor_image_tokens.py (flat charge per image, multiple images,
Responses-API / Anthropic-native / OpenAI-chat shapes, no-inflation on
raw base64, bounds-check on the constant, integration test that an
image-heavy tail actually gets trimmed).
* fix(image-input): replace blanket 20MB ceiling with empirically-verified per-provider limits
The previous commit imposed a hardcoded 20 MB base64 ceiling on all
providers, triggering auto-resize on anything larger. This was wrong in
both directions:
* Too loose for Anthropic — actual limit is 5 MB (returns HTTP 400
'image exceeds 5 MB maximum' above that).
* Too strict for OpenAI / Codex / OpenRouter — accept 49 MB+ without
complaint (empirically verified April 2026 with progressive PNG
sizes).
New behaviour:
* _PROVIDER_BASE64_CEILING table: only anthropic and bedrock have a
ceiling (5 MB, since bedrock-on-Claude shares Anthropic's decoder).
* Providers NOT in the table get no ceiling — images attach at native
size and we trust the provider to return its own error if it
disagrees. A provider-specific 400 message is clearer than us
guessing wrong and silently degrading image quality.
* build_native_content_parts() gains a keyword-only provider arg;
gateway/CLI/TUI pass the active provider so Anthropic users get
auto-resize protection while OpenAI users don't pay it.
* Resize target dropped from 5 MB to 4 MB to slide safely under
Anthropic's boundary with header overhead.
Empirical measurements (direct API, no Hermes in the loop):
image b64 anthropic openrouter/gpt5.5 codex-oauth/gpt5.5
0.19 MB ✓ ✓ ✓
12.37 MB ✗ 400 5MB ✓ ✓
23.85 MB ✗ 400 5MB ✓ ✓
49.46 MB ✗ 413 ✓ ✓
Tests: rewrote TestOversizeHandling (5 tests): no-ceiling pass-through,
Anthropic resize fires, Anthropic skip on resize-fail, build_native_parts
routes ceiling by provider, unknown provider gets no ceiling. All 52
targeted tests pass.
* refactor(image-input): attempt native, shrink-and-retry on provider reject
Replace proactive per-provider size ceilings with a reactive shrink path
on the provider's actual rejection. All providers now attempt native
full-size attachment first; if the provider returns an image-too-large
error, the agent silently shrinks and retries once.
Why the previous design was wrong: hardcoding provider ceilings
(anthropic=5MB, others=unlimited) meant OpenAI users on a 10MB image
paid no tax, but Anthropic users lost quality on anything >5MB even
though the empirical behaviour at provider-reject time is the same
(shrink + retry). Baking the table into the routing layer also
requires updating Hermes every time a provider's limit changes.
Reactive design:
- image_routing.py: _file_to_data_url encodes native size, no ceiling.
build_native_content_parts drops its provider kwarg.
- error_classifier.py: new FailoverReason.image_too_large + pattern
match ("image exceeds", "image too large", etc.) checked BEFORE
context_overflow so Anthropic's 5MB rejection lands in the right
bucket.
- run_agent.py: new _try_shrink_image_parts_in_messages walks api
messages in-place, re-encodes oversized data: URL image parts
through vision_tools._resize_image_for_vision to fit under 4MB,
handles both chat.completions (dict image_url) and Responses
(string image_url) shapes, ignores http URLs (provider-fetched).
New image_shrink_retry_attempted flag in the retry loop fires the
shrink exactly once per turn after credential-pool recovery but
before auth retries.
E2E verified live against Anthropic claude-sonnet-4-6:
- 17.9MB PNG (23.9MB b64) attached at native size
- Anthropic returns 400 "image exceeds 5 MB maximum"
- Agent logs '📐 Image(s) exceeded provider size limit — shrank and
retrying...'
- Retry succeeds, correct response delivered in 6.8s total.
Tests: 12 new (8 shrink-helper shapes + 4 classifier signals),
replaces 5 proactive-ceiling tests with 3 simpler 'native attach works'
tests. 181 targeted tests pass. test_enum_members_exist in
test_error_classifier.py updated for the new enum value.
Adds a short always-on pointer to the system prompt: when the user asks
about configuring, setting up, troubleshooting, or using Hermes Agent
itself, load the hermes-agent skill via skill_view(name='hermes-agent')
and fall back to https://hermes-agent.nousresearch.com/docs via
web_extract. Keeps sessions without skill_view loaded useful too — the
docs URL + web_extract is enough to answer most questions.
The guidance is appended right after DEFAULT_AGENT_IDENTITY (or SOUL.md)
so it ships regardless of which toolset profile is active. Footprint is
~560 chars, behind the existing prompt cache.
Closes#15775.
Title generation swallowed exceptions at debug level and returned None,
so a depleted auxiliary provider (e.g. OpenRouter 402) silently left
sessions with NULL titles. Reporter observed 45 untitled sessions
accumulated over 19 days with no user-visible indication.
- agent/title_generator.py: accept optional failure_callback, bump log
to WARNING, invoke callback on call_llm exception (swallowing callback
errors so nothing can crash the fire-and-forget worker thread).
- cli.py, gateway/run.py: pass agent._emit_auxiliary_failure as the
callback so failures route through the existing user-visible warning
channel.
- tests: cover callback fires / errors are swallowed / no-callback
legacy behavior / maybe_auto_title forwards kwarg to worker.
The bare-string isinstance guard added in 80ae2621 covered _find_tail_cut_by_tokens
(line 1084) but missed the identical pattern in _calculate_protect_tail_boundary
(line 487, the protect-tail scan loop). Both loops call .get("text", "") on every
list item in message["content"]; both crash with AttributeError when that list
contains a bare string.
Apply the same dict/str/fallback isinstance guard to the protect-tail path.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
raw_content from message["content"] can be a list that contains bare
strings, not only dicts. The previous `p.get("text", "")` call raised
AttributeError on string items, crashing context compression for any
session that had a message with mixed content.
Guard with isinstance checks: dict → .get("text"), str → len(p),
fallback → len(str(p)). Adds a regression test covering the bare-string
case that would have AttributeError'd on the pre-fix code.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
_find_tail_cut_by_tokens called len(content) to estimate message tokens.
When content is a list of blocks (multimodal: text + image_url), len()
returns block count (e.g. 2) rather than character count, so a message
with 500 chars of text was counted as ~10 tokens instead of ~135.
This caused the backward walk to exhaust all messages before hitting the
budget ceiling; the head_end safeguard then forced cut = n - min_tail,
shrinking the protected tail to the bare minimum and preventing effective
compression of long multimodal conversations.
Fix mirrors the existing pattern in _prune_old_tool_results (line 487):
sum(len(p.get("text", "")) for p in raw_content)
if isinstance(raw_content, list) else len(raw_content)
Tests: 3 new cases in TestTokenBudgetTailProtection — regression guard
(confirms the test fails with the bug), plain-string regression guard,
and image-only block edge case.
Fixes#16087.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two related fixes for OpenClaw-residue problems after an OpenClaw→Hermes
migration (especially migrations done via OpenClaw's own tool, which
doesn't archive the source directory).
1. optional-skills/migration/openclaw-migration/scripts/openclaw_to_hermes.py:
rebrand_text() was rewriting ~/.openclaw/config.yaml → ~/.Hermes/config.yaml
(capital H — a directory that doesn't exist). Now case-preserving:
"OpenClaw" → "Hermes" (prose), but "openclaw" → "hermes" (so filesystem
paths land on the real Hermes home). Regex logic unchanged — replacement
function now checks if the matched text was all-lowercase and emits the
replacement in the matching case.
2. agent/onboarding.py + cli.py: one-time startup banner the first time
Hermes launches and finds ~/.openclaw/. Tells the user to run
`hermes claw cleanup` to archive it, gated on the existing onboarding
seen-flag framework (onboarding.seen.openclaw_residue_cleanup in
config.yaml). Fires once per install; re-running requires wiping that
flag or running cleanup directly.
Tests:
- 4 new TestDetectOpenclawResidue tests (present / absent / file-instead-
of-dir / default-home smoke)
- 2 TestOpenclawResidueHint tests (content check)
- 2 TestOpenclawResidueSeenFlag tests (flag isolation + round-trip)
- test_rebrand_text_preserves_filesystem_path_casing regression test
with 4 scenarios including the exact ~/.openclaw/config.yaml case
- Existing test_rebrand_text_* tests updated to the new case-preserving
contract (lowercase input → lowercase output)
Co-authored-by: teknium1 <teknium@noreply.github.com>
Four small tool-description / skill-content tweaks addressing recurring
model mistakes seen in @versun's docx feedback (Kimi 2.6, but the patterns
apply to every model):
1. browser_navigate description: call out .md/.txt/.json/.yaml/.csv/.xml,
raw.githubusercontent.com, and API endpoints as specifically preferring
curl or web_extract. The generic "prefer web_search or web_extract" was
too weak; models kept firing up the browser for plain-text URLs.
2. delegate_task description: two additions.
(a) Pass user language / output-style preferences in 'context' when they
differ from English — otherwise subagents default to English and their
summaries contaminate the final reply (caused the bilingual digest bug).
(b) Subagent summaries are self-reports, not verified facts. For
operations with external side-effects (HTTP uploads, remote writes,
file creation at shared paths), require a verifiable handle (URL, ID,
path) and verify it yourself before claiming success.
3. agent/prompt_builder.py Skills-mandatory block: new explicit line
"Whenever the user asks to configure / set up / modify / install /
enable / disable / troubleshoot Hermes Agent itself, load the
`hermes-agent` skill first." The generic "load what's relevant" didn't
route Hermes-meta questions (like "how do I turn off redaction?") to
the one skill that has the answer.
4. skills/autonomous-ai-agents/hermes-agent/SKILL.md: new "Security &
Privacy Toggles" section covering security.redact_secrets (with the
import-time-snapshot restart-required caveat), privacy.redact_pii,
approvals.mode (manual/smart/off) + --yolo + HERMES_YOLO_MODE, shell
hooks allowlist, and how to disable network/media tools entirely.
Every command verified against the actual config keys — no invented
knobs.
Co-authored-by: teknium1 <teknium@noreply.github.com>
`_resolve_effective_accept()` used `return bool(cfg_val)` for the
`hooks_auto_accept` config key. In Python, `bool("false")` is `True`,
so a user setting `hooks_auto_accept: "false"` (quoted YAML string)
in `config.yaml` would silently enable auto-approval of every shell
hook, bypassing the consent prompt entirely.
Replace the coercion with the same type-aware parsing already used for
the HERMES_ACCEPT_HOOKS env var three lines above: bool passthrough,
strings checked against {1,true,yes,on} case-insensitively, everything
else (including "false", None, 0, ints) rejected.
Add TestHooksAutoAcceptParsing guarding the regression across all four
value shapes (bool, string-truthy, string-falsy, missing/None).
Reported by @sprmn24 in #16244.
Enter while the agent is busy can now inject the typed text via /steer —
arriving at the agent after the next tool call — instead of interrupting
(current default) or queueing for the next turn.
Changes:
- cli.py: keybinding honors busy_input_mode='steer' by calling
agent.steer(text) on the UI thread (thread-safe), with automatic
fallback to 'queue' when the agent is missing, steer() is unavailable,
images are attached, or steer() rejects the payload. /busy accepts
'steer' as a fourth argument alongside queue/interrupt/status.
- gateway/run.py: busy-message handler and the PRIORITY running-agent
path both route through running_agent.steer() when the mode is 'steer',
with the same fallback-to-queue safety net. Ack wording tells users
their message was steered into the current run. Restart-drain queueing
now also activates for 'steer' so messages aren't lost across restarts.
- agent/onboarding.py: first-touch hint has a steer branch for both
CLI and gateway.
- hermes_cli/commands.py: /busy args_hint updated to include steer,
and 'steer' is registered as a subcommand (completions).
- hermes_cli/web_server.py: dashboard select widget offers steer.
- hermes_cli/config.py, cli-config.yaml.example, hermes_cli/tips.py:
inline docs updated.
- website/docs/user-guide/cli.md + messaging/index.md: documented.
- Tests: steer set/status path for /busy; onboarding hints;
_load_busy_input_mode accepts steer; busy-session ack exercises
steer success + two fallback-to-queue branches.
Requested on X by @CodingAcct.
Default is unchanged (interrupt).
Azure OpenAI content filters (Default/DefaultV2) treat bracketed
[SYSTEM: ...] meta-instructions as prompt-injection attempts and
reject requests with HTTP 400.
Replacing [SYSTEM: with [IMPORTANT: preserves the same semantic
meaning for the model while bypassing the Azure heuristic.
Fixes#6576
Follow-up to cherry-picked PR #15920:
- agent/credential_pool.py: hoist 'from hermes_cli.config import get_env_value'
to module top instead of inline try/except in each seed site (3 sites).
No import cycle — hermes_cli/config.py doesn't depend on agent.credential_pool.
- hermes_cli/auth.py: same hoist for the _resolve_api_key_provider_secret loop.
- tests/tools/test_credential_pool_env_fallback.py: replace smoke-only tests
with real .env file I/O. Each test writes a temp ~/.hermes/.env, verifies
_seed_from_env / _resolve_api_key_provider_secret read from it, and asserts
the full priority chain: os.environ > .env > credential_pool. Uses
'deepseek' as the test provider since 'openai' isn't in PROVIDER_REGISTRY
and _seed_from_env's generic path requires a real pconfig lookup.
_resolve_api_key_provider_secret() and _seed_from_env() only checked
os.environ for provider API keys. When keys exist in ~/.hermes/.env but
are not loaded into the process environment (e.g. ACP adapter entry
point, post-session-start .env edits, or non-CLI entry points), the
resolution returns an empty string, causing HTTP 401 failures.
Changes:
- credential_pool._seed_from_env: use get_env_value() which checks both
os.environ and ~/.hermes/.env file, preventing _prune_stale_seeded_entries
from removing valid entries whose env var isn't in os.environ
- credential_pool._seed_from_env: same fix for openrouter and
base_url_env_var resolution
- auth._resolve_api_key_provider_secret: use get_env_value() instead of
os.getenv(), and add credential_pool fallback when env resolution fails
Fixes#15914
PR #16046 added /busy and /verbose hints to the classic CLI and the
gateway runner but skipped the Ink TUI (and therefore the dashboard
/chat page, which embeds the TUI via PTY). This extends the same
latch to the TUI with TUI-native wording.
The TUI's busy-input model is not the /busy knob from the CLI —
single Enter while busy auto-queues, double Enter on an empty line
interrupts. The new busy-input hint teaches THAT gesture instead of
telling the user to flip a config that does not apply.
Changes:
- agent/onboarding.py — add busy_input_hint_tui() + tool_progress_hint_tui()
- tui_gateway/server.py — onboarding.claim JSON-RPC (Ink triggers busy
hint on enqueue) + _maybe_emit_onboarding_hint helper hooked into
_on_tool_complete for the 30s/tool_progress=all path. Same
config.yaml latch so each hint fires at most once per install across
CLI, gateway, and TUI combined.
- ui-tui/src/gatewayTypes.ts — OnboardingClaimResponse + onboarding.hint event
- ui-tui/src/app/createGatewayEventHandler.ts — render the hint event as sys()
- ui-tui/src/app/useSubmission.ts — claim busy_input_prompt on first
busy enqueue
- tests/agent/test_onboarding.py — +3 cases for TUI hint shape
- tests/tui_gateway/test_protocol.py — +4 cases for onboarding.claim
- website/docs/user-guide/tui.md — new 'Interrupting and queueing'
section explaining the TUI's double-Enter model and the hints
Validation:
scripts/run_tests.sh tests/agent/test_onboarding.py \
tests/tui_gateway/test_protocol.py \
tests/gateway/test_busy_session_ack.py
-> 66 passed
npm --prefix ui-tui run type-check -> clean
npm --prefix ui-tui run lint -> clean
npm --prefix ui-tui run build -> clean
Instead of a blocking first-run questionnaire, show a one-time hint the first
time the user hits each behavior fork:
1. First message while the agent is working — appends a hint to the busy-ack
explaining the /busy queue vs /busy interrupt knob, phrased to match the
mode that was just applied (don't tell a queue-mode user to switch to
queue).
2. First tool that runs for >= 30s in the noisiest progress mode
(tool_progress: all) — prints a hint about /verbose to cycle display
modes (all -> new -> off -> verbose). Gated on /verbose actually being
usable on the surface: always shown on CLI; on gateway only shown when
display.tool_progress_command is enabled.
Each hint is latched in config.yaml under onboarding.seen.<flag>, so it
fires exactly once per install across CLI, gateway, and cron, then never
again. Users can wipe the section to re-see hints.
New:
- agent/onboarding.py — is_seen / mark_seen / hint strings, shared by
both CLI and gateway.
- onboarding.seen in DEFAULT_CONFIG (hermes_cli/config.py) and in
load_cli_config defaults (cli.py). No _config_version bump — deep
merge handles new keys.
Wired:
- gateway/run.py: _handle_active_session_busy_message appends the hint
after building the ack. progress_callback tracks tool.completed
duration and queues the tool-progress hint into the progress bubble.
- cli.py: CLI input loop appends the busy-input hint on the first busy
Enter; _on_tool_progress appends the tool-progress hint on the first
>=30s tool completion. In-memory CLI_CONFIG is also updated so
subsequent fires in the same process are suppressed immediately.
All writes go through atomic_yaml_write and are wrapped in try/except
so onboarding can never break the input/busy-ack paths.
`_apply_model_switch_result` (the interactive `/model` picker's
confirmation path) printed `ModelInfo.context_window` straight from
models.dev, which reports the vendor-wide value (1.05M for gpt-5.5 on
openai). ChatGPT Codex OAuth caps the same slug at 272K, so the picker
showed 1M while the runtime (compressor, gateway `/model`, typed
`/model <name>`) correctly used 272K — the classic 'sometimes 1M,
sometimes 272K' mismatch on a single model.
Both display paths now go through `resolve_display_context_length()`,
matching the fix that `_handle_model_switch` received earlier.
Also bump the stale last-resort fallback in DEFAULT_CONTEXT_LENGTHS
(`gpt-5.5: 400000 -> 1050000`) to match the real OpenAI API value; the
272K Codex cap is already enforced via the Codex-OAuth branch, so the
fallback now reflects what every non-Codex probe-miss should see.
Tests: adds `test_apply_model_switch_result_context.py` with three
scenarios (Codex cap wins, OpenRouter shows 1.05M, resolver-empty falls
back to ModelInfo). Updates the existing non-Codex fallback test to
assert 1.05M (the correct value).
## Validation
| path | before | after |
|-------------------------------|-----------|-----------|
| picker -> gpt-5.5 on Codex | 1,050,000 | 272,000 |
| picker -> gpt-5.5 on OpenAI | 1,050,000 | 1,050,000 |
| picker -> gpt-5.5 on OpenRouter | 1,050,000 | 1,050,000 |
| typed /model gpt-5.5 on Codex | 272,000 | 272,000 |
#14934 added deepseek-v4-pro / deepseek-v4-flash to the DeepSeek native
provider but the context-window lookup still falls back to the existing
"deepseek" substring entry (128K). DeepSeek V4 ships with a 1M context
window, so any caller relying on get_model_context_length() for
pre-flight token budgeting (compression, context warnings) under-counts
by ~8x.
Add explicit lowercase entries for the four DeepSeek model ids that
ship 1M context:
- deepseek-v4-pro
- deepseek-v4-flash
- deepseek-chat (legacy alias, server-side maps to v4-flash non-thinking)
- deepseek-reasoner (legacy alias, server-side maps to v4-flash thinking)
Longest-key-first substring matching means these explicit entries also
cover the vendor-prefixed forms (deepseek/deepseek-v4-pro on OpenRouter
and Nous Portal) without regressing the existing 128K fallback for
older / unknown DeepSeek model ids on custom endpoints.
Source: https://api-docs.deepseek.com/zh-cn/quick_start/pricing
Nous Portal multiplexes multiple upstream providers (DeepSeek, Kimi,
MiMo, Hermes) behind one endpoint. Before this fix, any 429 on any of
those models recorded a cross-session file breaker that blocked EVERY
model on Nous for the cooldown window -- even though the caller's
own RPM/RPH/TPM/TPH buckets were healthy. Users hit a DeepSeek V4 Pro
capacity error, restarted, switched to Kimi 2.6, and still got
'Nous Portal rate limit active -- resets in 46m 53s'.
Nous already emits the full x-ratelimit-* header suite on every
response (captured by rate_limit_tracker into agent._rate_limit_state).
We now gate the breaker on that data: trip it only when either the
429's own headers or the last-known-good state show a bucket with
remaining == 0 AND a reset window >= 60s. Upstream-capacity 429s
(healthy buckets everywhere, but upstream out of capacity) fall
through to normal retry/fallback and the breaker is never written.
Note: the in-memory 'restart TUI/gateway to clear' workaround
circulated in Discord does NOT work -- the breaker is file-backed at
~/.hermes/rate_limits/nous.json. The workaround for users still
affected by a bad state file is to delete it.
Reported in Discord by CrazyDok1 and KYSIV (Apr 2026).
Azure OpenAI requires an `api-version` query parameter on every request.
When users include it in the base_url (e.g. `?api-version=2025-04-01-preview`),
the OpenAI SDK silently drops it during URL construction, causing 404 errors.
Extract query params from base_url and pass them via `default_query` so the
SDK appends them to every request. This is a generic solution that works for
any custom endpoint requiring query parameters, not just Azure.
No-op for URLs without query params — fully backward compatible.
Fixes#15779. Custom-provider per-model context_length (`custom_providers[].models.<id>.context_length`) is now honored across every resolution path, not just agent startup. Also adds 256K as the top probe tier and default fallback.
## What changed
New helper `hermes_cli.config.get_custom_provider_context_length()` — single source of truth for the per-model override lookup, with trailing-slash-insensitive base-url matching.
`agent.model_metadata.get_model_context_length()` gains an optional `custom_providers=` kwarg (step 0b — runs after explicit `config_context_length` but before every other probe).
Wired through five call sites that previously either duplicated the lookup or ignored it entirely:
- `run_agent.py` startup — refactored to use the new helper (dedups legacy inline loop, keeps invalid-value warning)
- `AIAgent.switch_model()` — re-reads custom_providers from live config on every /model switch
- `hermes_cli.model_switch.resolve_display_context_length()` — new `custom_providers=` kwarg
- `gateway/run.py` /model confirmation (picker callback + text path)
- `gateway/run.py` `_format_session_info` (/info)
## Context probe tiers
`CONTEXT_PROBE_TIERS = [256_000, 128_000, 64_000, 32_000, 16_000, 8_000]` — was `[128_000, ...]`. `DEFAULT_FALLBACK_CONTEXT` follows tier[0], so unknown models now default to 256K. The stale `128000` literal in the OpenRouter metadata-miss path is replaced with `DEFAULT_FALLBACK_CONTEXT` for consistency.
## Repro (from #15779)
```yaml
custom_providers:
- name: my-custom-endpoint
base_url: https://example.invalid/v1
model: gpt-5.5
models:
gpt-5.5:
context_length: 1050000
```
`/model gpt-5.5 --provider custom:my-custom-endpoint` → previously "Context: 128,000", now "Context: 1,050,000".
## Tests
- `tests/hermes_cli/test_custom_provider_context_length.py` — new file, 19 tests covering the helper, step-0b integration, and the 256K tier invariants
- `tests/hermes_cli/test_model_switch_context_display.py` — added regression tests for #15779 through the display resolver
- `tests/gateway/test_session_info.py` — updated default-fallback assertion (128K → 256K)
- `tests/agent/test_model_metadata.py` — updated tier assertions for the new top tier
The Codex Responses API rejects input_text inside assistant messages —
only output_text and refusal are valid content types for assistant role.
_chat_content_to_responses_parts() previously hardcoded all text content
to input_text regardless of the message role. When an assistant message
had list-format content (multimodal or structured), this produced invalid
input_text parts that the API rejected with:
Invalid value: 'input_text'. Supported values are: 'output_text' and 'refusal'.
Fix: add a role parameter to _chat_content_to_responses_parts() that
selects output_text for assistant messages and input_text for user
messages. Thread this through _chat_messages_to_responses_input() and
_preflight_codex_input_items().
Fixes#15687
The AIAgent.flush_memories pre-compression save, the gateway
_flush_memories_for_session, and everything feeding them are
obsolete now that the background memory/skill review handles
persistent memory extraction.
Problems with flush_memories:
- Pre-dates the background review loop. It was the only memory-save
path when introduced; the background review now fires every 10 user
turns on CLI and gateway alike, which is far more frequent than
compression or session reset ever triggered flush.
- Blocking and synchronous. Pre-compression flush ran on the live agent
before compression, blocking the user-visible response.
- Cache-breaking. Flush built a temporary conversation prefix
(system prompt + memory-only tool list) that diverged from the live
conversation's cached prefix, invalidating prompt caching. The
gateway variant spawned a fresh AIAgent with its own clean prompt
for each finalized session — still cache-breaking, just in a
different process.
- Redundant. Background review runs in the live conversation's
session context, gets the same content, writes to the same memory
store, and doesn't break the cache. Everything flush_memories
claimed to preserve is already covered.
What this removes:
- AIAgent.flush_memories() method (~248 LOC in run_agent.py)
- Pre-compression flush call in _compress_context
- flush_memories call sites in cli.py (/new + exit)
- GatewayRunner._flush_memories_for_session + _async_flush_memories
(and the 3 call sites: session expiry watcher, /new, /resume)
- 'flush_memories' entry from DEFAULT_CONFIG auxiliary tasks,
hermes tools UI task list, auxiliary_client docstrings
- _memory_flush_min_turns config + init
- #15631's headroom-deduction math in
_check_compression_model_feasibility (headroom was only needed
because flush dragged the full main-agent system prompt along;
the compression summariser sends a single user-role prompt so
new_threshold = aux_context is safe again)
- The dedicated test files and assertions that exercised
flush-specific paths
What this renames (with read-time backcompat on sessions.json):
- SessionEntry.memory_flushed -> SessionEntry.expiry_finalized.
The session-expiry watcher still uses the flag to avoid re-running
finalize/eviction on the same expired session; the new name
reflects what it now actually gates. from_dict() reads
'expiry_finalized' first, falls back to the legacy 'memory_flushed'
key so existing sessions.json files upgrade seamlessly.
Supersedes #15631 and #15638.
Tested: 383 targeted tests pass across run_agent/, agent/, cli/,
and gateway/ session-boundary suites. No behavior regressions —
background memory review continues to handle persistent memory
extraction on both CLI and gateway.
Generalize the temperature-specific 400 retry that shipped in PR #15621 so
the same reactive strategy covers any provider that rejects an arbitrary
request parameter — — not just temperature.
- agent/auxiliary_client.py:
* New _is_unsupported_parameter_error(exc, param): matches the same six
phrasings the old temperature detector did plus 'unrecognized parameter'
and 'invalid parameter', against any named param.
* _is_unsupported_temperature_error is now a thin back-compat wrapper so
existing imports and tests keep working.
* The max_tokens → max_completion_tokens retry branch in call_llm and
async_call_llm now (a) gates on 'max_tokens is not None' so we do not
pop a key that was never set and silently substitute a None value on
the retry, and (b) also matches the generic helper in addition to the
legacy 'max_tokens' / 'unsupported_parameter' substring checks — picking
up phrasings like 'Unknown parameter: max_tokens' that previously slipped
through.
- tests/agent/test_unsupported_parameter_retry.py: 18 new tests covering
the generic detector across params, the back-compat wrapper, and the two
hardenings to the max_tokens retry branch (None gate + generic phrasing).
Credit: retry-generalization pattern from @nicholasrae's PR #15416. That PR
also proposed the reactive temperature retry which landed independently via
PR #15621 + #15623 (co-authored with @BlueBirdBack). This commit salvages
the remaining hardening ideas onto current main.
Universal reactive fix for 'HTTP 400: Unsupported parameter: temperature'
across all providers/models — not just Codex Responses.
The same backend can accept temperature for some models and reject it for
others (e.g. gpt-5.4 accepts but gpt-5.5 rejects on the same OpenAI
endpoint; similar patterns on Copilot, OpenRouter reasoning routes, and
Anthropic Opus 4.7+ via OAI-compat). An allow/deny-list by model name does
not scale.
call_llm / async_call_llm now detect the concrete 'unsupported parameter:
temperature' 400 and transparently retry once without temperature. Kimi's
server-managed omission and Opus 4.7+'s proactive strip stay in place —
this is the safety net for everything else.
Changes:
- agent/auxiliary_client.py: add _is_unsupported_temperature_error helper;
wire into both sync and async call_llm paths before the existing
max_tokens/payment/auth retry ladder
- tests/agent/test_unsupported_temperature_retry.py: 19 tests covering
detector phrasings, sync + async retry, no-retry-without-temperature,
and non-temperature 400s not triggering the retry
Builds on PR #15620 (codex_responses fallback) which stripped temperature
up front for that one api_mode. This PR closes the gap for every other
provider/model combo via reactive retry.
Credit: retry approach and detector originate from @BlueBirdBack's PR #15578.
Co-authored-by: BlueBirdBack <BlueBirdBack@users.noreply.github.com>
update_model() recalculated threshold_tokens but left tail_token_budget
and max_summary_tokens at their __init__ values. When switching from a
200K model to 32K, the tail budget stayed at ~20K tokens (62% of 32K)
instead of the intended ~10%.
Adds budget recalculation in update_model() and 2 regression tests.
gpt-5.x on the Codex Responses API sometimes degenerates and emits
Harmony-style `to=functions.<name> {json}` serialization as plain
assistant-message text instead of a structured `function_call` item.
The intent never makes it into `response.output` as a function_call,
so `tool_calls` is empty and `_normalize_codex_response()` returns
the leaked text as the final content. Downstream (e.g. delegate_task),
this surfaces as a confident-looking summary with `tool_trace: []`
because no tools actually ran — the Taiwan-embassy-email bug report.
Detect the pattern, scrub the content, and return finish_reason=
'incomplete' so the existing Codex-incomplete continuation path
(run_agent.py:11331, 3 retries) gets a chance to re-elicit a proper
function_call item. Encrypted reasoning items are preserved so the
model keeps its chain-of-thought on the retry.
Regression tests: leaked text triggers incomplete, real tool calls
alongside leak-looking text are preserved, clean responses pass
through unchanged.
Reported on Discord (gpt-5.4 / openai-codex).
## Problem
When a pooled HTTPS connection to the Bedrock runtime goes stale (NAT
timeout, VPN flap, server-side TCP RST, proxy idle cull), the next
Converse call surfaces as one of:
* botocore.exceptions.ConnectionClosedError / ReadTimeoutError /
EndpointConnectionError / ConnectTimeoutError
* urllib3.exceptions.ProtocolError
* A bare AssertionError raised from inside urllib3 or botocore
(internal connection-pool invariant check)
The agent loop retries the request 3x, but the cached boto3 client in
_bedrock_runtime_client_cache is reused across retries — so every
attempt hits the same dead connection pool and fails identically.
Only a process restart clears the cache and lets the user keep working.
The bare-AssertionError variant is particularly user-hostile because
str(AssertionError()) is an empty string, so the retry banner shows:
⚠️ API call failed: AssertionError
📝 Error:
with no hint of what went wrong.
## Fix
Add two helpers to agent/bedrock_adapter.py:
* is_stale_connection_error(exc) — classifies exceptions that
indicate dead-client/dead-socket state. Matches botocore
ConnectionError + HTTPClientError subtrees, urllib3
ProtocolError / NewConnectionError, and AssertionError
raised from a frame whose module name starts with urllib3.,
botocore., or boto3.. Application-level AssertionErrors are
intentionally excluded.
* invalidate_runtime_client(region) — per-region counterpart to
the existing reset_client_cache(). Evicts a single cached
client so the next call rebuilds it (and its connection pool).
Wire both into the Converse call sites:
* call_converse() / call_converse_stream() in
bedrock_adapter.py (defense-in-depth for any future caller)
* The two direct client.converse(**kwargs) /
client.converse_stream(**kwargs) call sites in run_agent.py
(the paths the agent loop actually uses)
On a stale-connection exception, the client is evicted and the
exception re-raised unchanged. The agent's existing retry loop then
builds a fresh client on the next attempt and recovers without
requiring a process restart.
## Tests
tests/agent/test_bedrock_adapter.py gets three new classes (14 tests):
* TestInvalidateRuntimeClient — per-region eviction correctness;
non-cached region returns False.
* TestIsStaleConnectionError — classifies botocore
ConnectionClosedError / EndpointConnectionError /
ReadTimeoutError, urllib3 ProtocolError, library-internal
AssertionError (both urllib3.* and botocore.* frames), and
correctly ignores application-level AssertionError and
unrelated exceptions (ValueError, KeyError).
* TestCallConverseInvalidatesOnStaleError — end-to-end: stale
error evicts the cached client, non-stale error (validation)
leaves it alone, successful call leaves it cached.
All 116 tests in test_bedrock_adapter.py pass.
Signed-off-by: Andre Kurait <andrekurait@gmail.com>
Bedrock's aws_sdk auth_type had no matching branch in
resolve_provider_client(), causing it to fall through to the
"unhandled auth_type" warning and return (None, None). This broke
all auxiliary tasks (compression, memory, summarization) for Bedrock
users — the main conversation loop worked fine, but background
context management silently failed.
Add an aws_sdk branch that creates an AnthropicAuxiliaryClient via
build_anthropic_bedrock_client(), using boto3's default credential
chain (IAM roles, SSO, env vars, instance metadata). Default
auxiliary model is Haiku for cost efficiency.
Closes#13919
## Problem
`get_model_context_length()` in `agent/model_metadata.py` had a resolution
order bug that caused every Bedrock model to fall back to the 128K default
context length instead of reaching the static Bedrock table (200K for
Claude, etc.).
The root cause: `bedrock-runtime.<region>.amazonaws.com` is not listed in
`_URL_TO_PROVIDER`, so `_is_known_provider_base_url()` returned False.
The resolution order then ran the custom-endpoint probe (step 2) *before*
the Bedrock branch (step 4b), which:
1. Treated Bedrock as a custom endpoint (via `_is_custom_endpoint`).
2. Called `fetch_endpoint_model_metadata()` → `GET /models` on the
bedrock-runtime URL (Bedrock doesn't serve this shape).
3. Fell through to `return DEFAULT_FALLBACK_CONTEXT` (128K) at the
"probe-down" branch — never reaching the Bedrock static table.
Result: users on Bedrock saw 128K context for Claude models that
actually support 200K on Bedrock, causing premature auto-compression.
## Fix
Promote the Bedrock branch from step 4b to step 1b, so it runs *before*
the custom-endpoint probe at step 2. The static table in
`bedrock_adapter.py::get_bedrock_context_length()` is the authoritative
source for Bedrock (the ListFoundationModels API doesn't expose context
window sizes), so there's no reason to probe `/models` first.
The original step 4b is replaced with a one-line breadcrumb comment
pointing to the new location, to make the resolution-order docstring
accurate.
## Changes
- `agent/model_metadata.py`
- Add step 1b: Bedrock static-table branch (unchanged predicate, moved).
- Remove dead step 4b block, replace with breadcrumb comment.
- Update resolution-order docstring to include step 1b.
- `tests/agent/test_model_metadata.py`
- New `TestBedrockContextResolution` class (3 tests):
- `test_bedrock_provider_returns_static_table_before_probe`:
confirms `provider="bedrock"` hits the static table and does NOT
call `fetch_endpoint_model_metadata` (regression guard).
- `test_bedrock_url_without_provider_hint`: confirms the
`bedrock-runtime.*.amazonaws.com` host match works without an
explicit `provider=` hint.
- `test_non_bedrock_url_still_probes`: confirms the probe still
fires for genuinely-custom endpoints (no over-reach).
## Testing
pytest tests/agent/test_model_metadata.py -q
# 83 passed in 1.95s (3 new + 80 existing)
## Risk
Very low.
- Predicate is identical to the original step 4b — no behaviour change
for non-Bedrock paths.
- Original step 4b was dead code for the user-facing case (always hit
the 128K fallback first), so removing it cannot regress behaviour.
- Bedrock path now short-circuits before any network I/O — faster too.
- `ImportError` fall-through preserved so users without `boto3`
installed are unaffected.
## Related
- This is a prerequisite for accurate context-window accounting on
Bedrock — the fix for #14710 (stale-connection client eviction)
depends on correct context sizing to know when to compress.
Signed-off-by: Andre Kurait <andrekurait@gmail.com>
Bedrock model IDs use dots as namespace separators (anthropic.claude-opus-4-7,
us.anthropic.claude-sonnet-4-5-v1:0), not version separators.
normalize_model_name() was unconditionally converting all dots to hyphens,
producing invalid IDs that Bedrock rejects with HTTP 400/404.
This affected both the main agent loop (partially mitigated by
_anthropic_preserve_dots in run_agent.py) and all auxiliary client calls
(compression, session_search, vision, etc.) which go through
_AnthropicCompletionsAdapter and never pass preserve_dots=True.
Fix: add _is_bedrock_model_id() to detect Bedrock namespace prefixes
(anthropic., us., eu., ap., jp., global.) and skip dot-to-hyphen
conversion for these IDs regardless of the preserve_dots flag.
Bug 3 — Stale OAuth token not detected in 'hermes model':
- _model_flow_anthropic used 'has_creds = bool(existing_key)' which treats
any non-empty token (including expired OAuth tokens) as valid.
- Added existing_is_stale_oauth check: if the only credential is an OAuth
token (sk-ant- prefix) with no valid cc_creds fallback, mark it stale
and force the re-auth menu instead of silently accepting a broken token.
Bug 4 — macOS Keychain credentials never read:
- Claude Code >=2.1.114 migrated from ~/.claude/.credentials.json to the
macOS Keychain under service 'Claude Code-credentials'.
- Added _read_claude_code_credentials_from_keychain() using the 'security'
CLI tool; read_claude_code_credentials() now tries Keychain first then
falls back to JSON file.
- Non-Darwin platforms return None from Keychain read immediately.
Tests:
- tests/agent/test_anthropic_keychain.py: 11 cases covering Darwin-only
guard, security command failures, JSON parsing, fallback priority.
- tests/hermes_cli/test_anthropic_model_flow_stale_oauth.py: 8 cases
covering stale OAuth detection, API key passthrough, cc_creds fallback.
Refs: #12905
FixesNousResearch/hermes-agent#9813
Root cause: _is_oauth_token() only recognized sk-ant-* and eyJ* patterns,
but Claude Code OAuth tokens from CLAUDE_CODE_OAUTH_TOKEN use cc- prefix
Fix: Add cc- prefix detection so these tokens route through Bearer auth
Two small fixes triggered by a support report where the user saw a
cryptic 'HTTP 400 - Error 400 (Bad Request)!!1' (Google's GFE HTML
error page, not a real API error) on every gemini-2.5-pro request.
The underlying cause was an empty GOOGLE_API_KEY / GEMINI_API_KEY, but
nothing in our output made that diagnosable:
1. hermes_cli/dump.py: the api_keys section enumerated 23 providers but
omitted Google entirely, so users had no way to verify from 'hermes
dump' whether the key was set. Added GOOGLE_API_KEY and GEMINI_API_KEY
rows.
2. agent/gemini_native_adapter.py: GeminiNativeClient.__init__ accepted
an empty/whitespace api_key and stamped it into the x-goog-api-key
header, which made Google's frontend return a generic HTML 400 long
before the request reached the Generative Language backend. Now we
raise RuntimeError at construction with an actionable message
pointing at GOOGLE_API_KEY/GEMINI_API_KEY and aistudio.google.com.
Added a regression test that covers '', ' ', and None.
Concurrent Hermes processes (e.g. cron jobs) refreshing a Nous OAuth token
via resolve_nous_runtime_credentials() write the rotated tokens to auth.json.
The calling process's pool entry becomes stale, and the next refresh against
the already-rotated token triggers a 'refresh token reuse' revocation on
the Nous Portal.
_sync_nous_entry_from_auth_store() reads auth.json under the same lock used
by resolve_nous_runtime_credentials, and adopts the newer token pair before
refreshing the pool entry. This complements #15111 (which preserved the
obtained_at timestamps through seeding).
Partial salvage of #10160 by @konsisumer — only the agent/credential_pool.py
changes + the 3 Nous-specific regression tests. The PR also touched 10
unrelated files (Dockerfile, tips.py, various tool tests) which were
dropped as scope creep.
Regression tests:
- test_sync_nous_entry_from_auth_store_adopts_newer_tokens
- test_sync_nous_entry_noop_when_tokens_match
- test_nous_exhausted_entry_recovers_via_auth_store_sync
The least_used strategy selected entries via min(request_count) but
never incremented the counter. All entries stayed at count=0, so the
strategy degenerated to fill_first behavior with no actual load balancing.
Now increments request_count after each selection and persists the update.
The Copilot provider resolved context windows via models.dev static data,
which does not include account-specific models (e.g. claude-opus-4.6-1m
with 1M context). This adds the live Copilot /models API as a higher-
priority source for copilot/copilot-acp/github-copilot providers.
New helper get_copilot_model_context() in hermes_cli/models.py extracts
capabilities.limits.max_prompt_tokens from the cached catalog. Results
are cached in-process for 1 hour.
In agent/model_metadata.py, step 5a queries the live API before falling
through to models.dev (step 5b). This ensures account-specific models
get correct context windows while standard models still have a fallback.
Part 1 of #7731.
Refs: #7272
Raw GitHub tokens (gho_/github_pat_/ghu_) are now exchanged for
short-lived Copilot API tokens via /copilot_internal/v2/token before
being used as Bearer credentials. This is required to access
internal-only models (e.g. claude-opus-4.6-1m with 1M context).
Implementation:
- exchange_copilot_token(): calls the token exchange endpoint with
in-process caching (dict keyed by SHA-256 fingerprint), refreshed
2 minutes before expiry. No disk persistence — gateway is long-running
so in-memory cache is sufficient.
- get_copilot_api_token(): convenience wrapper with graceful fallback —
returns exchanged token on success, raw token on failure.
- Both callers (hermes_cli/auth.py and agent/credential_pool.py) now
pipe the raw token through get_copilot_api_token() before use.
12 new tests covering exchange, caching, expiry, error handling,
fingerprinting, and caller integration. All 185 existing copilot/auth
tests pass.
Part 2 of #7731.
Pass an explicit HOME into Copilot ACP child processes so delegated ACP runs do not fail when the ambient environment is missing HOME.
Prefer the per-profile subprocess home when available, then fall back to HOME, expanduser('~'), pwd.getpwuid(...), and /home/openclaw. Add regression tests for both profile-home preference and clean HOME fallback.
Refs #11068.
Two narrow fixes motivated by #15099.
1. _seed_from_singletons() was dropping obtained_at, agent_key_obtained_at,
expires_in, and friends when seeding device_code pool entries from the
providers.nous singleton. Fresh credentials showed up with
obtained_at=None, which broke downstream freshness-sensitive consumers
(self-heal hooks, pool pruning by age) — they treated just-minted
credentials as older than they actually were and evicted them.
2. When the Nous Portal OAuth 2.1 server returns invalid_grant with
'Refresh token reuse detected' in the error_description, rewrite the
message to explain the likely cause (an external process consumed the
rotated RT without persisting it back) and the mitigation. The generic
reuse message led users to report this as a Hermes persistence bug when
the actual trigger was typically a third-party monitoring script calling
/api/oauth/token directly. Non-reuse errors keep their original server
description untouched.
Closes#15099.
Regression tests:
- tests/agent/test_credential_pool.py::test_nous_seed_from_singletons_preserves_obtained_at_timestamps
- tests/hermes_cli/test_auth_nous_provider.py::test_refresh_token_reuse_detection_surfaces_actionable_message
- tests/hermes_cli/test_auth_nous_provider.py::test_refresh_non_reuse_error_keeps_original_description
Google AI Studio's free tier (<= 250 req/day for gemini-2.5-flash) is
exhausted in a handful of agent turns, so the setup wizard now refuses
to wire up Gemini when the supplied key is on the free tier, and the
runtime 429 handler appends actionable billing guidance.
Setup-time probe (hermes_cli/main.py):
- `_model_flow_api_key_provider` fires one minimal generateContent call
when provider_id == 'gemini' and classifies the response as
free/paid/unknown via x-ratelimit-limit-requests-per-day header or
429 body containing 'free_tier'.
- Free -> print block message, refuse to save the provider, return.
- Paid -> 'Tier check: paid' and proceed.
- Unknown (network/auth error) -> 'could not verify', proceed anyway.
Runtime 429 handler (agent/gemini_native_adapter.py):
- `gemini_http_error` appends billing guidance when the 429 error body
mentions 'free_tier', catching users who bypass setup by putting
GOOGLE_API_KEY directly in .env.
Tests: 21 unit tests for the probe + error path, 4 tests for the
setup-flow block. All 67 existing gemini tests still pass.
PR #14935 added a Codex-aware context resolver but only new lookups
hit the live /models probe. Users who had run Hermes on gpt-5.5 / 5.4
BEFORE that PR already had the wrong value (e.g. 1,050,000 from
models.dev) persisted in ~/.hermes/context_length_cache.yaml, and the
cache-first lookup in get_model_context_length() returns it forever.
Symptom (reported in the wild by Ludwig, min heo, Gaoge on current
main at 6051fba9d, which is AFTER #14935):
* Startup banner shows context usage against 1M
* Compression fires late and then OpenAI hard-rejects with
'context length will be reduced from 1,050,000 to 128,000'
around the real 272k boundary.
Fix: when the step-1 cache returns a value for an openai-codex lookup,
check whether it's >= 400k. Codex OAuth caps every slug at 272k (live
probe values) so anything at or above 400k is definitionally a
pre-#14935 leftover. Drop that entry from the on-disk cache and fall
through to step 5, which runs the live /models probe and repersists
the correct value (or 272k from the hardcoded fallback if the probe
fails). Non-Codex providers and legitimately-cached Codex entries at
272k are untouched.
Changes:
- agent/model_metadata.py:
* _invalidate_cached_context_length() — drop a single entry from
context_length_cache.yaml and rewrite the file.
* Step-1 cache check in get_model_context_length() now gates
provider=='openai-codex' entries >= 400k through invalidation
instead of returning them.
Tests (3 new in TestCodexOAuthContextLength):
- stale 1.05M Codex entry is dropped from disk AND re-resolved
through the live probe to 272k; unrelated cache entries survive.
- fresh 272k Codex entry is respected (no probe call, no invalidation).
- non-Codex 1M entries (e.g. anthropic/claude-opus-4.6 on OpenRouter)
are unaffected — the guard is strictly scoped to openai-codex.
Full tests/agent/test_model_metadata.py: 88 passed.
Gemini's Schema validator requires every `enum` entry to be a string,
even when the parent `type` is integer/number/boolean. Discord's
`auto_archive_duration` parameter (`type: integer, enum: [60, 1440,
4320, 10080]`) tripped this on every request that shipped the full
tool catalog to generativelanguage.googleapis.com, surfacing as
`Gateway: Non-retryable client error: Gemini HTTP 400 (INVALID_ARGUMENT)
Invalid value ... (TYPE_STRING), 60` and aborting the turn.
Sanitize by dropping the `enum` key when the declared type is numeric
or boolean and any entry is non-string. The `type` and `description`
survive, so the model still knows the allowed values; the tool handler
keeps its own runtime validation. Other providers (OpenAI,
OpenRouter, Anthropic) are unaffected — the sanitizer only runs for
native Gemini / cloudcode adapters.
Reported by @selfhostedsoul on Discord with hermes debug share.
Keep auxiliary provider resolution aligned with the switch and persisted main-provider paths when models.dev returns github-copilot slugs.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Auxiliary tasks (session_search, flush_memories, approvals, compression,
vision, etc.) that route to a named custom provider declared under
config.yaml 'providers:' with 'api_mode: anthropic_messages' were
silently building a plain OpenAI client and POSTing to
{base_url}/chat/completions, which returns 404 on Anthropic-compatible
gateways that only expose /v1/messages.
Two gaps caused this:
1. hermes_cli/runtime_provider.py::_get_named_custom_provider — the
providers-dict branch (new-style) returned only name/base_url/api_key/
model and dropped api_mode. The legacy custom_providers-list branch
already propagated it correctly. The dict branch now parses and
returns api_mode via _parse_api_mode() in both match paths.
2. agent/auxiliary_client.py::resolve_provider_client — the named
custom provider block at ~L1740 ignored custom_entry['api_mode']
and unconditionally built an OpenAI client (only wrapping for
Codex/Responses). It now mirrors _try_custom_endpoint()'s three-way
dispatch: anthropic_messages → AnthropicAuxiliaryClient (async wrapped
in AsyncAnthropicAuxiliaryClient), codex_responses → CodexAuxiliaryClient,
otherwise plain OpenAI. An explicit task-level api_mode override
still wins over the provider entry's declared api_mode.
Fixes#15033
Tests: tests/agent/test_auxiliary_named_custom_providers.py gains a
TestProvidersDictApiModeAnthropicMessages class covering
- providers-dict preserves valid api_mode
- invalid api_mode values are dropped
- missing api_mode leaves the entry unchanged (no regression)
- resolve_provider_client returns (Async)AnthropicAuxiliaryClient for
api_mode=anthropic_messages
- full chain via get_text_auxiliary_client / get_async_text_auxiliary_client
with an auxiliary.<task> override
- providers without api_mode still use the OpenAI-wire path
Follow-up to PR #14533 — applies the same _resolve_requests_verify()
treatment to the one requests.get() site the PR missed (Codex OAuth
chatgpt.com /models probe). Keeps all seven requests.get() callsites
in model_metadata.py consistent so HERMES_CA_BUNDLE / REQUESTS_CA_BUNDLE /
SSL_CERT_FILE are honored everywhere.
Co-authored-by: teknium1 <teknium@hermes-agent>
- hermes_cli/auth.py: add _default_verify() with macOS Homebrew certifi
fallback (mirrors weixin 3a0ec1d93). Extend env var chain to include
REQUESTS_CA_BUNDLE so one env var works across httpx + requests paths.
- agent/model_metadata.py: add _resolve_requests_verify() reading
HERMES_CA_BUNDLE / REQUESTS_CA_BUNDLE / SSL_CERT_FILE in priority
order. Apply explicit verify= to all 6 requests.get callsites.
- Tests: 18 new unit tests + autouse platform pin on existing
TestResolveVerifyFallback to keep its "returns True" assertions
platform-independent.
Empirically verified against self-signed HTTPS server: requests honors
REQUESTS_CA_BUNDLE only; httpx honors SSL_CERT_FILE only. Hermes now
honors all three everywhere.
Triggered by Discord reports — Nous OAuth SSL failure on macOS
Homebrew Python; custom provider self-signed cert ignored despite
REQUESTS_CA_BUNDLE set in env.
Manual /compress crashed with 'LCMEngine' object has no attribute
'_align_boundary_forward' when any context-engine plugin was active.
The gateway handler reached into _align_boundary_forward and
_find_tail_cut_by_tokens on tmp_agent.context_compressor, but those
are ContextCompressor-specific — not part of the generic ContextEngine
ABC — so every plugin engine (LCM, etc.) raised AttributeError.
- Add optional has_content_to_compress(messages) to ContextEngine ABC
with a safe default of True (always attempt).
- Override it in the built-in ContextCompressor using the existing
private helpers — preserves exact prior behavior for 'compressor'.
- Rewrite gateway /compress preflight to call the ABC method, deleting
the private-helper reach-in.
- Add focus_topic to the ABC compress() signature. Make _compress_context
retry without focus_topic on TypeError so older strict-sig plugins
don't crash on manual /compress <focus>.
- Regression test with a fake ContextEngine subclass that only
implements the ABC (mirrors LCM's surface).
Reported by @selfhostedsoul (Discord, Apr 22).
OpenRouter returns a 404 with the specific message
'No endpoints available matching your guardrail restrictions and data
policy. Configure: https://openrouter.ai/settings/privacy'
when a user's account-level privacy setting excludes the only endpoint
serving a model (e.g. DeepSeek V4 Pro, which today is hosted only by
DeepSeek's own endpoint that may log inputs).
Before this change we classified it as model_not_found, which was
misleading (the model exists) and triggered provider fallback (useless —
the same account setting applies to every OpenRouter call).
Now it classifies as a new FailoverReason.provider_policy_blocked with
retryable=False, should_fallback=False. The error body already contains
the fix URL, so the user still gets actionable guidance.
On ChatGPT Codex OAuth every gpt-5.x slug actually caps at 272,000 tokens,
but Hermes was resolving gpt-5.5 / gpt-5.4 to 1,050,000 (from models.dev)
because openai-codex aliases to the openai entry there. At 1.05M the
compressor never fires and requests hard-fail with 'context window
exceeded' around the real 272k boundary.
Verified live against chatgpt.com/backend-api/codex/models:
gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.3-codex, gpt-5.2-codex,
gpt-5.2, gpt-5.1-codex-max → context_window = 272000
Changes:
- agent/model_metadata.py:
* _fetch_codex_oauth_context_lengths() — probe the Codex /models
endpoint with the OAuth bearer token and read context_window per
slug (1h in-memory TTL).
* _resolve_codex_oauth_context_length() — prefer the live probe,
fall back to hardcoded _CODEX_OAUTH_CONTEXT_FALLBACK (all 272k).
* Wire into get_model_context_length() when provider=='openai-codex',
running BEFORE the models.dev lookup (which returns 1.05M). Result
persists via save_context_length() so subsequent lookups skip the
probe entirely.
* Fixed the now-wrong comment on the DEFAULT_CONTEXT_LENGTHS gpt-5.5
entry (400k was never right for Codex; it's the catch-all for
providers we can't probe live).
Tests (4 new in TestCodexOAuthContextLength):
- fallback table used when no token is available (no models.dev leakage)
- live probe overrides the fallback
- probe failure (non-200) falls back to hardcoded 272k
- non-codex providers (openrouter, direct openai) unaffected
Non-codex context resolution is unchanged — the Codex branch only fires
when provider=='openai-codex'.
Fixes a broader class of 'tools.function.parameters is not a valid
moonshot flavored json schema' errors on Nous / OpenRouter aggregators
routing to moonshotai/kimi-k2.6 with MCP tools loaded.
## Moonshot sanitizer (agent/moonshot_schema.py, new)
Model-name-routed (not base-URL-routed) so Nous / OpenRouter users are
covered alongside api.moonshot.ai. Applied in
ChatCompletionsTransport.build_kwargs when is_moonshot_model(model).
Two repairs:
1. Fill missing 'type' on every property / items / anyOf-child schema
node (structural walk — only schema-position dicts are touched, not
container maps like properties/$defs).
2. Strip 'type' at anyOf parents; Moonshot rejects it.
## MCP normalizer hardened (tools/mcp_tool.py)
Draft-07 $ref rewrite from PR #14802 now also does:
- coerce missing / null 'type' on object-shaped nodes (salvages #4897)
- prune 'required' arrays to names that exist in 'properties'
(salvages #4651; Gemini 400s on dangling required)
- apply recursively, not just top-level
These repairs are provider-agnostic so the same MCP schema is valid on
OpenAI, Anthropic, Gemini, and Moonshot in one pass.
## Crash fix: safe getattr for Tool.inputSchema
_convert_mcp_schema now uses getattr(t, 'inputSchema', None) so MCP
servers whose Tool objects omit the attribute entirely no longer abort
registration (salvages #3882).
## Validation
- tests/agent/test_moonshot_schema.py: 27 new tests (model detection,
missing-type fill, anyOf-parent strip, non-mutation, real-world MCP
shape)
- tests/tools/test_mcp_tool.py: 7 new tests (missing / null type,
required pruning, nested repair, safe getattr)
- tests/agent/transports/test_chat_completions.py: 2 new integration
tests (Moonshot route sanitizes, non-Moonshot route doesn't)
- Targeted suite: 49 passed
- E2E via execute_code with a realistic MCP tool carrying all three
Moonshot rejection modes + dangling required + draft-07 refs:
sanitizer produces a schema valid on Moonshot and Gemini
OpenAI launched GPT-5.5 on Codex today (Apr 23 2026). Adds it to the static
catalog and pipes the user's OAuth access token into the openai-codex path of
provider_model_ids() so /model mid-session and the gateway picker hit the
live ChatGPT codex/models endpoint — new models appear for each user
according to what ChatGPT actually lists for their account, without a Hermes
release.
Verified live: 'gpt-5.5' returns priority 0 (featured) from the endpoint,
400k context per OpenAI's launch article. 'hermes chat --provider
openai-codex --model gpt-5.5' completes end-to-end.
Changes:
- hermes_cli/codex_models.py: add gpt-5.5 to DEFAULT_CODEX_MODELS + forward-compat
- agent/model_metadata.py: 400k context length entry
- hermes_cli/models.py: resolve codex OAuth token before calling
get_codex_model_ids() in provider_model_ids('openai-codex')
Commit 43de1ca8 removed the _nr_to_assistant_message shim in favor of
duck-typed properties on the ToolCall dataclass. However, the
extra_content property (which carries the Gemini thought_signature) was
omitted from the ToolCall definition. This caused _build_assistant_message
to silently drop the signature via getattr(tc, 'extra_content', None)
returning None, leading to HTTP 400 errors on subsequent turns for all
Gemini 3 thinking models.
Add the extra_content property to ToolCall (matching the existing
call_id and response_item_id pattern) so the thought_signature round-trips
correctly through the transport → agent loop → API replay path.
Credit to @celttechie for identifying the root cause and providing the fix.
Closes#14488
## Merged
Adds MiMo v2.5-pro and v2.5 support to Xiaomi native provider, OpenCode Go, and setup wizard.
### Changes
- Context lengths: added v2.5-pro (1M) and v2.5 (1M), corrected existing MiMo entries to exact values (262144)
- Provider lists: xiaomi, opencode-go, setup wizard
- Vision: upgraded from mimo-v2-omni to mimo-v2.5 (omnimodal)
- Config description updated for XIAOMI_API_KEY
- Tests updated for new vision model preference
### Verification
- 4322 tests passed, 0 new regressions
- Live API tested on Xiaomi portal: basic, reasoning, tool calling, multi-tool, file ops, system prompt, vision — all pass
- Self-review found and fixed 2 issues (redundant vision check, stale HuggingFace context length)
NormalizedResponse and ToolCall now have backward-compat properties
so the agent loop can read them directly without the shim:
ToolCall: .type, .function (returns self), .call_id, .response_item_id
NormalizedResponse: .reasoning_content, .reasoning_details,
.codex_reasoning_items
This eliminates the 35-line shim and its 4 call sites in run_agent.py.
Also changes flush_memories guard from hasattr(response, 'choices')
to self.api_mode in ('chat_completions', 'bedrock_converse') so it
works with raw boto3 dicts too.
WS1 items 3+4 of Cycle 2 (#14418).
3-layer chain (transport → v2 → v1) was collapsed to 2-layer in PR 7.
This collapses the remaining 2-layer (transport → v1 → NR mapping in
transport) to 1-layer: v1 now returns NormalizedResponse directly.
Before: adapter returns (SimpleNamespace, finish_reason) tuple,
transport unpacks and maps to NormalizedResponse (22 lines).
After: adapter returns NormalizedResponse, transport is a
1-line passthrough.
Also updates ToolCall construction — adapter now creates ToolCall
dataclass directly instead of SimpleNamespace(id, type, function).
WS1 item 1 of Cycle 2 (#14418).
Replace direct normalize_anthropic_response() call in
_AnthropicCompletionsAdapter.create() with
AnthropicTransport.normalize_response() via get_transport().
Before: auxiliary_client called adapter v1 directly, bypassing
the transport layer entirely.
After: auxiliary_client → get_transport('anthropic_messages') →
transport.normalize_response() → adapter v1 → NormalizedResponse.
The adapter v1 function (normalize_anthropic_response) now has
zero callers outside agent/anthropic_adapter.py and the transport.
This unblocks collapsing v1 to return NormalizedResponse directly
in a follow-up (the remaining 2-layer chain becomes 1-layer).
WS1 item 2 of Cycle 2 (#14418).
* feat(agent): add PLATFORM_HINTS for matrix, mattermost, and feishu
These platform adapters fully support media delivery (send_image,
send_document, send_voice, send_video) but were missing from
PLATFORM_HINTS, leaving agents unaware of their platform context,
markdown rendering, and MEDIA: tag support.
Salvaged from PR #7370 by Rutimka — wecom excluded since main already
has a more detailed version.
Co-Authored-By: Marco Rutsch <marco@rutimka.de>
* test: add missing Markdown assertion for feishu platform hint
---------
Co-authored-by: Marco Rutsch <marco@rutimka.de>
Consolidate 4 per-transport lazy singleton helpers (_get_anthropic_transport,
_get_codex_transport, _get_chat_completions_transport, _get_bedrock_transport)
into one generic _get_transport(api_mode) with a shared dict cache.
Collapse the 65-line main normalize block (3 api_mode branches, each with
its own SimpleNamespace shim) into 7 lines: one _get_transport() call +
one _nr_to_assistant_message() shared shim. The shim extracts provider_data
fields (codex_reasoning_items, reasoning_details, call_id, response_item_id)
into the SimpleNamespace shape downstream code expects.
Wire chat_completions and bedrock_converse normalize through their transports
for the first time — these were previously falling into the raw
response.choices[0].message else branch.
Remove 8 dead codex adapter imports that have zero callers after PRs 1-6.
Transport lifecycle improvements:
- Eagerly warm transport cache at __init__ (surfaces import errors early)
- Invalidate transport cache on api_mode change (switch_model, fallback
activation, fallback restore, transport recovery) — prevents stale
transport after mid-session provider switch
run_agent.py: -32 net lines (11,988 -> 11,956).
PR 7 of the provider transport refactor.
Port from openclaw/openclaw#66664. The build_anthropic_kwargs call site
used 'max_tokens or _get_anthropic_max_output(model)', which correctly
falls back when max_tokens is 0 or None (falsy) but lets negative ints
(-1, -500), fractional floats (0.5, 8192.7), NaN, and infinity leak
through to the Anthropic API. Anthropic rejects these with HTTP 400
('max_tokens: must be greater than or equal to 1'), turning a local
config error into a surprise mid-conversation failure.
Add two resolver helpers matching OpenClaw's:
_resolve_positive_anthropic_max_tokens — returns int(value) only if
value is a finite positive number; excludes bools, strings, NaN,
infinity, sub-one positives (floor to 0).
_resolve_anthropic_messages_max_tokens — prefers a positive requested
value, else falls back to the model's output ceiling; raises
ValueError only if no positive budget can be resolved.
The context-window clamp at the call site (max_tokens > context_length)
is preserved unchanged — it handles oversized values; the new resolver
handles non-positive values. These concerns are now cleanly separated.
Tests: 17 new cases covering positive/zero/negative ints, fractional
floats (both >1 and <1), NaN, infinity, booleans, strings, None, and
integration via build_anthropic_kwargs.
Refs: openclaw/openclaw#66664
_generate_summary() takes (turns_to_summarize, focus_topic) but the
summary model fallback path passed (messages, summary_budget) — where
'messages' is not even in scope, causing a NameError.
Fix the recursive call to pass the correct variables so the fallback
to the main model actually works when the summary model is unavailable.
Fixes: #10721
Mid-stream SSL alerts (bad_record_mac, tls_alert_internal_error, handshake
failures) previously fell through the classifier pipeline to the 'unknown'
bucket because:
- ssl.SSLError type names weren't in _TRANSPORT_ERROR_TYPES (the
isinstance(OSError) catch picks up some but not all SDK-wrapped forms)
- the message-pattern list had no SSL alert substrings
The 'unknown' bucket is still retryable, but: (a) logs tell the user
'unknown' instead of identifying the cause, (b) it bypasses the
transport-specific backoff/fallback logic, and (c) if the SSL error
happens on a large session with a generic 'connection closed' wrapper,
the existing disconnect-on-large-session heuristic would incorrectly
trigger context compression — expensive, and never fixes a transport
hiccup.
Changes:
- Add ssl.SSLError and its subclass type names to _TRANSPORT_ERROR_TYPES
- New _SSL_TRANSIENT_PATTERNS list (separate from _SERVER_DISCONNECT_PATTERNS
so SSL alerts route to timeout, not context_overflow+compress)
- New step 5 in the classifier pipeline: SSL pattern check runs BEFORE
the disconnect check to pre-empt the large-session-compress path
Patterns cover both space-separated ('ssl alert', 'bad record mac')
and underscore-separated ('ERR_SSL_SSL/TLS_ALERT_BAD_RECORD_MAC')
forms. This is load-bearing because OpenSSL 3.x changed the error-code
separator from underscore to slash (e.g. SSLV3_ALERT_BAD_RECORD_MAC →
SSL/TLS_ALERT_BAD_RECORD_MAC) and will likely churn again — matching on
stable alert reason substrings survives future format changes.
Tests (8 new):
- BAD_RECORD_MAC in Python ssl.c format
- OpenSSL 3.x underscore format
- TLSV1_ALERT_INTERNAL_ERROR
- ssl handshake failure
- [SSL: ...] prefix fallback
- Real ssl.SSLError instance
- REGRESSION GUARD: SSL on large session does NOT compress
- REGRESSION GUARD: plain disconnect on large session STILL compresses
os.walk() by default does not follow symlinks, causing skills
linked via symlinks to be invisible to the skill discovery system.
Add followlinks=True so that symlinked skill directories are scanned.
Port from cline/cline#10266.
When OpenAI-compatible proxies (OpenRouter, Vercel AI Gateway, Cline)
route Claude models, they sometimes surface the Anthropic-native cache
counters (`cache_read_input_tokens`, `cache_creation_input_tokens`) at
the top level of the `usage` object instead of nesting them inside
`prompt_tokens_details`. Our chat-completions branch of
`normalize_usage()` only read the nested `prompt_tokens_details` fields,
so those responses:
- reported `cache_write_tokens = 0` even when the model actually did a
prompt-cache write,
- reported only some of the cache-read tokens when the proxy exposed them
top-level only,
- overstated `input_tokens` by the missed cache-write amount, which in
turn made cost estimation and the status-bar cache-hit percentage wrong
for Claude traffic going through these gateways.
Now the chat-completions branch tries the OpenAI-standard
`prompt_tokens_details` first and falls back to the top-level
Anthropic-shape fields only if the nested values are absent/zero. The
Anthropic and Codex Responses branches are unchanged.
Regression guards added for three shapes: top-level write + nested read,
top-level-only, and both-present (nested wins).
Zhipu AI (智谱) serves both international users via api.z.ai and
China-based users via open.bigmodel.cn. The domestic endpoint was not
mapped in _URL_TO_PROVIDER, causing Hermes to treat it as an unknown
custom endpoint and fall back to the default 128K context length
instead of resolving the correct 200K+ context via models.dev or the
hardcoded GLM defaults.
This affects users of both the standard API
(https://open.bigmodel.cn/api/paas/v4) and the Coding Plan
(https://open.bigmodel.cn/api/coding/paas/v4).
- Adds 'ctx_size' field to _CONTEXT_LENGTH_KEYS tuple
- Enables hermes agent to correctly detect context size from custom LLMs
running on Lemonade server that use this field name instead of the
standard keys (max_seq_len, n_ctx_train, n_ctx)
Fixes#12976
The generic "gemma": 8192 fallback was incorrectly matching gemma4:31b-cloud
before the more specific Gemma 4 entries could match, causing Hermes to assign
only 8K context instead of 262K. Added "gemma-4" and "gemma4" entries before
the fallback to correctly handle Gemma 4 model naming conventions.
Replace xiaomi/mimo-v2-pro with xiaomi/mimo-v2.5-pro and xiaomi/mimo-v2.5
in the OpenRouter fallback catalog and the nous provider model list.
Add matching DEFAULT_CONTEXT_LENGTHS entries (1M tokens each).
`is_local_endpoint()` leaned on `ipaddress.is_private`, which classifies
RFC-1918 ranges and link-local as private but deliberately excludes the
RFC 6598 CGNAT block (100.64.0.0/10) — the range Tailscale uses for its
mesh IPs. As a result, Ollama reached over Tailscale (e.g.
`http://100.77.243.5:11434`) was treated as remote and missed the
automatic stream-read / stale-stream timeout bumps, so cold model load
plus long prefill would trip the 300 s watchdog before the first token.
Add a module-level `_TAILSCALE_CGNAT = ipaddress.IPv4Network("100.64.0.0/10")`
(built once) and extend `is_local_endpoint()` to match the block both
via the parsed-`IPv4Address` path and the existing bare-string fallback
(for symmetry with the 10/172/192 checks). Also hoist the previously
function-local `import ipaddress` to module scope now that it's used by
the constant.
Extend `TestIsLocalEndpoint` with a CGNAT positive set (lower bound,
representative host, MagicDNS anchor, upper bound) and a near-miss
negative set (just below 100.64.0.0, just above 100.127.255.255, well
outside the block, and first-octet-wrong).
Anthropic's API can legitimately return content=[] with stop_reason="end_turn"
when the model has nothing more to add after a turn that already delivered the
user-facing text alongside a trivial tool call (e.g. memory write). The transport
validator was treating that as an invalid response, triggering 3 retries that
each returned the same valid-but-empty response, then failing the run with
"Invalid API response after 3 retries."
The downstream normalizer already handles empty content correctly (empty loop
over response.content, content=None, finish_reason="stop"), so the only fix
needed is at the validator boundary.
Tests:
- Empty content + stop_reason="end_turn" → valid (the fix)
- Empty content + stop_reason="tool_use" → still invalid (regression guard)
- Empty content without stop_reason → still invalid (existing behavior preserved)
Follow-up to the cherry-picked PR #13897 fix. Three issues found:
1. CRITICAL: The thinking block synthesised from reasoning_content was
immediately stripped by the third-party signature management code
(Kimi is classified as _is_third_party_anthropic_endpoint). Added a
Kimi-specific carve-out that preserves unsigned thinking blocks while
still stripping Anthropic-signed blocks Kimi can't validate.
2. Empty-string reasoning_content was silently dropped because the
truthiness check ('if reasoning_content and ...') evaluates to False
for ''. Changed to 'isinstance(reasoning_content, str)' so the
tier-3 fallback from _copy_reasoning_content_for_api (which injects
'' for Kimi tool-call messages with no reasoning) actually produces
a thinking block.
3. The thinking block was appended AFTER tool_use blocks. Anthropic
protocol requires thinking -> text -> tool_use ordering. Changed to
blocks.insert(0, ...) to prepend.
FixesNousResearch/hermes-agent#13848
Kimi's /coding endpoint speaks the Anthropic Messages protocol but has its
own thinking semantics: when thinking is enabled, Kimi validates message
history and requires every prior assistant tool-call message to carry
OpenAI-style reasoning_content.
The Anthropic path never populated that field, and
convert_messages_to_anthropic strips all Anthropic thinking blocks on
third-party endpoints — so the request failed with HTTP 400:
"thinking is enabled but reasoning_content is missing in assistant
tool call message at index N"
Now, when an assistant message contains tool_calls and a
reasoning_content string, we append a {"type": "thinking", ...} block
to the Anthropic content so Kimi can validate the history. This only
affects assistant messages with tool_calls + reasoning_content; plain
text assistant messages are unchanged.
The 404 branch in _classify_by_status had dead code: the generic
fallback below the _MODEL_NOT_FOUND_PATTERNS check returned the
exact same classification (model_not_found + should_fallback=True),
so every 404 — regardless of message — was treated as a missing model.
This bites local-endpoint users (llama.cpp, Ollama, vLLM) whose 404s
usually mean a wrong endpoint path, proxy routing glitch, or transient
backend issue — not a missing model. Claiming 'model not found' misleads
the next turn and silently falls back to another provider when the real
problem was a URL typo the user should see.
Fix: only classify 404 as model_not_found when the message actually
matches _MODEL_NOT_FOUND_PATTERNS ("invalid model", "model not found",
etc.). Otherwise fall through as unknown (retryable) so the real error
surfaces in the retry loop.
Test updated to match the new behavior. 103 error_classifier tests pass.
Adds a first-class 'stepfun' API-key provider surfaced as Step Plan:
- Support Step Plan setup for both International and China regions
- Discover Step Plan models live from /step_plan/v1/models, with a
small coding-focused fallback catalog when discovery is unavailable
- Thread StepFun through provider metadata, setup persistence, status
and doctor output, auxiliary routing, and model normalization
- Add tests for provider resolution, model validation, metadata
mapping, and StepFun region/model persistence
Based on #6005 by @hengm3467.
Co-authored-by: hengm3467 <100685635+hengm3467@users.noreply.github.com>
* feat(plugins): pluggable image_gen backends + OpenAI provider
Adds a ImageGenProvider ABC so image generation backends register as
bundled plugins under `plugins/image_gen/<name>/`. The plugin scanner
gains three primitives to make this work generically:
- `kind:` manifest field (`standalone` | `backend` | `exclusive`).
Bundled `kind: backend` plugins auto-load — no `plugins.enabled`
incantation. User-installed backends stay opt-in.
- Path-derived keys: `plugins/image_gen/openai/` gets key
`image_gen/openai`, so a future `tts/openai` cannot collide.
- Depth-2 recursion into category namespaces (parent dirs without a
`plugin.yaml` of their own).
Includes `OpenAIImageGenProvider` as the first consumer (gpt-image-1.5
default, plus gpt-image-1, gpt-image-1-mini, DALL-E 3/2). Base64
responses save to `$HERMES_HOME/cache/images/`; URL responses pass
through.
FAL stays in-tree for this PR — a follow-up ports it into
`plugins/image_gen/fal/` so the in-tree `image_generation_tool.py`
slims down. The dispatch shim in `_handle_image_generate` only fires
when `image_gen.provider` is explicitly set to a non-FAL value, so
existing FAL setups are untouched.
- 41 unit tests (scanner recursion, kind parsing, gate logic,
registry, OpenAI payload shapes)
- E2E smoke verified: bundled plugin autoloads, registers, and
`_handle_image_generate` routes to OpenAI when configured
* fix(image_gen/openai): don't send response_format to gpt-image-*
The live API rejects it: 'Unknown parameter: response_format'
(verified 2026-04-21 with gpt-image-1.5). gpt-image-* models return
b64_json unconditionally, so the parameter was both unnecessary and
actively broken.
* feat(image_gen/openai): gpt-image-2 only, drop legacy catalog
gpt-image-2 is the latest/best OpenAI image model (released 2026-04-21)
and there's no reason to expose the older gpt-image-1.5 / gpt-image-1 /
dall-e-3 / dall-e-2 alongside it — slower, lower quality, or awkward
(dall-e-2 squares only). Trim the catalog down to a single model.
Live-verified end-to-end: landscape 1536x1024 render of a Moog-style
synth matches prompt exactly, 2.4MB PNG saved to cache.
* feat(image_gen/openai): expose gpt-image-2 as three quality tiers
Users pick speed/fidelity via the normal model picker instead of a
hidden quality knob. All three tier IDs resolve to the single underlying
gpt-image-2 API model with a different quality parameter:
gpt-image-2-low ~15s fast iteration
gpt-image-2-medium ~40s default
gpt-image-2-high ~2min highest fidelity
Live-measured on OpenAI's API today: 15.4s / 40.8s / 116.9s for the
same 1024x1024 prompt.
Config:
image_gen.openai.model: gpt-image-2-high
# or
image_gen.model: gpt-image-2-low
# or env var for scripts/tests
OPENAI_IMAGE_MODEL=gpt-image-2-medium
Live-verified end-to-end with the low tier: 18.8s landscape render of a
golden retriever in wildflowers, vision-confirmed exact match.
* feat(tools_config): plugin image_gen providers inject themselves into picker
'hermes tools' → Image Generation now shows plugin-registered backends
alongside Nous Subscription and FAL.ai without tools_config.py needing
to know about them. OpenAI appears as a third option today; future
backends appear automatically as they're added.
Mechanism:
- ImageGenProvider gains an optional get_setup_schema() hook
(name, badge, tag, env_vars). Default derived from display_name.
- tools_config._plugin_image_gen_providers() pulls the schemas from
every registered non-FAL plugin provider.
- _visible_providers() appends those rows when rendering the Image
Generation category.
- _configure_provider() handles the new image_gen_plugin_name marker:
writes image_gen.provider and routes to the plugin's list_models()
catalog for the model picker.
- _toolset_needs_configuration_prompt('image_gen') stops demanding a
FAL key when any plugin provider reports is_available().
FAL is skipped in the plugin path because it already has hardcoded
TOOL_CATEGORIES rows — when it gets ported to a plugin in a follow-up
PR the hardcoded rows go away and it surfaces through the same path
as OpenAI.
Verified live: picker shows Nous Subscription / FAL.ai / OpenAI.
Picking OpenAI prompts for OPENAI_API_KEY, then shows the
gpt-image-2-low/medium/high model picker sourced from the plugin.
397 tests pass across plugins/, tools_config, registry, and picker.
* fix(image_gen): close final gaps for plugin-backend parity with FAL
Two small places that still hardcoded FAL:
- hermes_cli/setup.py status line: an OpenAI-only setup showed
'Image Generation: missing FAL_KEY'. Now probes plugin providers
and reports '(OpenAI)' when one is_available() — or falls back to
'missing FAL_KEY or OPENAI_API_KEY' if nothing is configured.
- image_generate tool schema description: said 'using FAL.ai, default
FLUX 2 Klein 9B'. Rewrote provider-neutral — 'backend and model are
user-configured' — and notes the 'image' field can be a URL or an
absolute path, which the gateway delivers either way via
extract_local_files().
