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).
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).
feat(gateway): refine Platform._missing_ and platform-connected dispatch
Restricts plugin-name acceptance to bundled plugin scan + registry
(no arbitrary string -> enum-pollution), pulls per-platform connectivity
checks into a _PLATFORM_CONNECTED_CHECKERS lambda map with a clean
_is_platform_connected method, and adds tests covering the checker map,
plugin platform interface, and IRC setup wizard.
Closes remaining functional gaps and adds documentation.
webhook.py: Cross-platform delivery now checks the plugin registry
for unknown platform names instead of hardcoding 15 names in a tuple.
Plugin platforms can receive webhook-routed deliveries.
prompt_builder: Platform hints (system prompt LLM guidance) now fall
back to the plugin registry's platform_hint field. Plugin platforms
can tell the LLM 'you're on IRC, no markdown.'
PlatformEntry: Added platform_hint field for LLM guidance injection.
IRC adapter: Added acquire_scoped_lock/release_scoped_lock in
connect/disconnect to prevent two profiles from using the same IRC
identity. Added platform_hint for IRC-specific LLM guidance.
Removed dead token-empty-warning extension for plugin platforms
(plugin adapters handle their own env vars via check_fn).
website/docs/developer-guide/adding-platform-adapters.md:
- Added 'Plugin Path (Recommended)' section with full code examples,
PLUGIN.yaml template, config.yaml examples, and a table showing all
18 integration points the plugin system handles automatically
- Renamed built-in checklist to clarify it's for core contributors
gateway/platforms/ADDING_A_PLATFORM.md:
- Added Plugin Path section pointing to the reference implementation
and full docs guide
- Clarified built-in path is for core contributors only
Each AIAgent.__init__() was unconditionally starting a daemon thread to
pre-warm the OpenRouter model metadata cache. In gateway mode a new
AIAgent is created for every incoming message, so one OS thread leaked
per request. After ~1 000 messages the process hit the Linux thread
limit and raised RuntimeError: can't start new thread for all subsequent
requests.
Add a module-level threading.Event (_openrouter_prewarm_done) that is
set before the thread is started. Subsequent AIAgent instantiations
skip the spawn entirely; fetch_model_metadata() is cached for 1 hour so
the single background call is sufficient.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
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.
MiniMax's /anthropic endpoint documents cache_control support (0.1x read
pricing, 5-min TTL) for MiniMax-M2.7, M2.5, M2.1, M2. PR #12846 gated
third-party Anthropic-wire caching on 'claude' in model name, which left
MiniMax's own model family re-paying full input tokens every turn.
Opt in explicitly via provider id (minimax / minimax-cn) or host match
(api.minimax.io / api.minimaxi.com). Narrow allowlist mirroring the
existing Qwen/Alibaba branch below; leaves room for a capability-based
surface (ProviderConfig.supports_anthropic_cache) if a third provider
needs it.
Closes#17332
Completes the cfg_get migration started in PR #17304. Covers the
remaining hermes_cli/ and plugins/ config-access sites that the first
PR intentionally left opportunistic.
Migrated (33 sites across 14 files):
hermes_cli/setup.py 13 sites (terminal.*, agent.*, display.*, compression.*, tts.*)
hermes_cli/tools_config.py 7 sites (tts.*, browser.*, web.*, platform_toolsets.*)
hermes_cli/plugins_cmd.py 3 sites (plugins.*, memory.*, context.*)
plugins/memory/honcho/cli.py 3 sites (hosts.*)
hermes_cli/web_server.py 1 site (dashboard.*)
hermes_cli/skills_config.py 1 site (platform_disabled)
hermes_cli/plugins.py 1 site (plugins.disabled)
hermes_cli/status.py 1 site (terminal.backend)
hermes_cli/mcp_config.py 1 site (mcp_servers.*)
hermes_cli/webhook.py 1 site (platforms.webhook)
plugins/memory/__init__.py 1 site (memory.provider)
plugins/memory/hindsight/ 1 site (banks.hermes)
plugins/memory/holographic/ 1 site (plugins.hermes-memory-store)
run_agent.py 1 site (auxiliary.compression)
The helper supports non-literal keys too, so e.g.
cfg.get('hosts', {}).get(HOST, {})
becomes
cfg_get(cfg, 'hosts', HOST, default={})
Migration bugs caught and fixed during this PR:
1. An AST-based batch rewrite naïvely captured the first word token in
a chain, which corrupted 'self._config.get(...).get(...)' into
'self.cfg_get(_config, ...)' (dropping 'self.', creating a broken
method call). Plugins/memory/hindsight caught it via its test suite.
Fixed manually to 'cfg_get(self._config, ...)'.
2. Import-extension heuristic rewrote multi-line parenthesized imports
('from X import (\n A,\n B,\n)') as
'from X import cfg_get, (' — syntactically broken. Fixed by inserting
cfg_get as the first name inside the parentheses.
Combined with PR #17304, the cfg_get migration now covers:
PR #17304 (first batch): 20 sites in tools/ + gateway/
PR #17317 (this one): 33 sites in hermes_cli/ + plugins/ + run_agent.py
Total: 53 sites migrated. Remaining ~8 sites are either:
- Function-call chains (e.g. '_load_stt_config().get(...).get(...)')
that would need double-evaluation or a local binding to migrate
cleanly — intentionally deferred.
- JSON response-navigation (e.g. 'response_data.get('data',{}).get('web'))
which is unrelated to config access and shouldn't use cfg_get.
Verified:
- 412/412 tests/plugins/ pass (including the hindsight test that caught
the self.X regex bug before commit)
- 3181/3189 tests/hermes_cli/ pass (8 pre-existing failures on main,
verified by git-stash comparison)
- Live 'hermes status' and 'hermes config' render correctly (exercise
the migrated terminal.backend, tts.provider, browser.cloud_provider,
compression.threshold, display.tool_progress sites)
- Live 'hermes chat': 1 turn + /quit, zero errors in 11-line log window
No semantic changes — cfg_get was already proven to be a 1:1 match for
the original .get("X",{}).get("Y",default) pattern in PR #17304.
The background skill-review prompts (_SKILL_REVIEW_PROMPT and the **Skills**
half of _COMBINED_REVIEW_PROMPT) steered the reviewer toward passive
behavior — most passes concluded 'Nothing to save.' even when the session
produced real lessons. User-preference corrections (style, format,
legibility, verbosity) were especially lost: they were read as memory
signals only, so skills never carried the fix.
This rewrite changes the stance:
- **Active-update bias.** The reviewer now treats inaction as a missed
learning opportunity. 'Nothing to save.' remains an explicit escape
but is no longer framed as the most-common outcome.
- **User-preference corrections are first-class skill signals.** Style,
tone, format, legibility, verbosity complaints — and the actual
phrasings users use ('stop doing X', 'this is too verbose', 'I hate
when you Y', 'remember this') — now warrant patching the skill that
governs the task, not just writing to memory.
- **Loaded-skill-first preference order.** When a skill was loaded via
/skill-name or skill_view during the session, the reviewer patches
THAT one first. It was in play; it's the right place.
- **Four-step ladder: patch-loaded → patch-umbrella → support-file →
create.** Support files are explicitly enumerated as three kinds:
* references/<topic>.md — session-specific detail OR condensed
knowledge banks (quoted research, API docs excerpts, domain notes)
* templates/<name>.<ext> — starter files to copy and modify
* scripts/<name>.<ext> — statically re-runnable actions
- **Name-veto for CREATE.** New skill names MUST be class-level — no PR
numbers, error strings, codenames, library-alone names, or session
artifacts ('fix-X / debug-Y / audit-Z-today'). If the proposed name
only fits today's task, fall back to one of the patch/support-file
options.
- **Memory scope clarified.** 'who the user is and what the current
situation and state of your operations are' — MEMORY.md is
situational/state, USER.md is identity/preferences.
- **Curator handoff.** Reviewer flags overlap; the background curator
handles consolidation at scale. Single-session reviewer doesn't
attempt umbrella-rebalancing.
Tests: tests/run_agent/test_review_prompt_class_first.py upgraded to
assert the new behavioral contracts (active bias, user-correction
signals, loaded-skill-first, support-file kinds, name-veto, memory
framing, curator handoff). 17 tests, all pass.
Co-authored-by: teknium1 <teknium@users.noreply.github.com>
- Remove dead _lmstudio_loaded_context attribute from run_agent.py (set
but never read — the loaded context is pushed to context_compressor.update_model
which is the actual consumer)
- Cache empty reasoning options with 60s TTL to avoid per-turn HTTP probe
for non-reasoning LM Studio models. Non-empty results cached permanently.
- Extract _lmstudio_server_root(), _lmstudio_request_headers(), and
_lmstudio_fetch_raw_models() shared helpers in models.py — eliminates
URL-strip + auth-header + HTTP-call duplication across probe_lmstudio_models,
ensure_lmstudio_model_loaded, and lmstudio_model_reasoning_options
- Revert runtime_provider.py base_url precedence change: preserve the
established contract (saved config.base_url > env var > default) for all
api_key providers
- Remove unnecessary config version bump 22→23
- Fix TUI test: relax target_model assertion to avoid module-cache flake
- AUTHOR_MAP: added rugved@lmstudio.ai → rugvedS07
CopilotACPClient communicates via subprocess stdio and returns a plain
SimpleNamespace from _create_chat_completion(). The streaming path tries
to iterate this as a stream, crashing with:
TypeError: 'types.SimpleNamespace' object is not iterable
Mirror the existing ACP exclusion pattern (used for Responses API upgrade)
to disable streaming when provider is copilot-acp or base_url starts with
acp:// or acp+tcp://.
Based on PR #9428 by @ningfangbin and issue #16271 by @Joseph19820124.
Fixes#16271
* 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>
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>
* 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>
Adds a pre-call sanitizer that detects assistant messages containing only
reasoning (reasoning / reasoning_content, no visible content, no
tool_calls) and drops them from the API copy. Adjacent user messages
left behind are merged so role alternation is preserved for the
provider.
Mirrors Claude Code's approach in src/utils/messages.ts
(filterOrphanedThinkingOnlyMessages + mergeAdjacentUserMessages). We
drop the whole turn rather than fabricate stub text (the '.' /
'(continued)' pattern from contributor PRs #11098, #13010, #16842 that
were rejected because they put words in the model's mouth).
The stored conversation history (self.messages) is never mutated — only
the per-call api_messages copy. Users still see the reasoning block in
the CLI/gateway transcript; only the wire copy is cleaned. Session
persistence keeps the full trace.
Two call sites covered:
- Main agent loop, after _sanitize_api_messages (catches every turn).
- Iteration-limit-summary fallback path.
Tests: tests/run_agent/test_thinking_only_sanitizer.py — 25 cases
covering detection (string/list content, whitespace-only, tool_calls,
reasoning_details list form), drop behavior, adjacent-user merge
(string+string, list+list, mixed), non-mutation of input dicts, and
system-message handling.
E2E live-tested against 5 providers with a poisoned history (empty
assistant message + reasoning_content): OpenRouter→Anthropic/OpenAI/
DeepSeek-R1/Qwen, native Gemini. All 5 accepted the cleaned request.
Happy-path regression (5/5) confirms the sanitizer is a noop when no
thinking-only turn exists.
Related: #16823 (wontfix — stub-text approach rejected).
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 #15328's vision-unsupported retry branch in run_agent.py.
_strip_images_from_messages() previously deleted any message whose content
was entirely images. That's fine for synthetic user messages injected for
attachment delivery, but it breaks providers for tool-role messages — the
paired tool_call_id on the preceding assistant message ends up unmatched,
which OpenAI-compatible APIs reject with HTTP 400.
Fix: tool-role messages whose content becomes empty are replaced with a
plaintext placeholder that preserves the tool_call_id linkage. Only
non-tool messages are dropped. Added 10 tests covering the role-alternation
invariants + image-type coverage.
Image-rejection detector: expanded phrase list (image content not
supported / multimodal input / vision input / model does not support
image) and gated on 4xx status so transient 5xx errors never get
misinterpreted as 'server said no to images'. Detection is documented as
best-effort English phrase matching.
AUTHOR_MAP: mapped 3820588+ddupont808@users.noreply.github.com to
ddupont808 so release notes attribute the salvage correctly.
Tool handlers (e.g. computer_use capture) return a _multimodal envelope
dict when a screenshot is attached. The tool-message builder was passing
this raw dict as the `content` field of role:tool messages, which is an
illegal format — OpenAI-compatible APIs expect a string or a content-parts
list, not a plain Python dict, and would reject it with a 400/422 error.
Fix: unwrap _multimodal results to their `content` list
([{type:text,...},{type:image_url,...}]) in both the parallel and
sequential tool-call paths. The Anthropic adapter already handles content
lists natively; vision-capable OpenAI-compatible servers (mlx-vlm,
GPT-4o, etc.) accept image_url parts in tool messages directly.
Also add a _vision_supported adaptive fallback: on first image-rejection
error ("Only 'text' content type is supported." etc.) the agent strips all
image parts from the message history and retries with text only, so
text-only endpoints degrade gracefully without crashing the session.
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.
Streaming-only providers (glm, MiniMax, gpt-5.x via aigw, Anthropic via
openai-compat shims) emit reasoning through delta.reasoning_content
chunks that get accumulated into the local reasoning_text string — but
never land on the assistant message object as a top-level attribute. The
prior guard at _build_assistant_message only wrote reasoning_content
when the SDK exposed hasattr(msg, 'reasoning_content'), so these
providers persisted the chain-of-thought under the internal 'reasoning'
key and omitted the protocol-standard field.
The poison was silent until the user later switched to a DeepSeek-v4 or
Kimi thinking model, at which point replay failed with HTTP 400:
'The reasoning_content in the thinking mode must be passed back to the
API.' One reported session store accumulated 4,031 poisoned messages
across 1,101 files (#16844).
Fix: add an additive fallback that promotes the already-sanitized
reasoning_text to reasoning_content when no earlier branch wrote it AND
reasoning text was actually captured. Layered on top of the existing
SDK-attr branch and DeepSeek ''-pad (#15250) rather than replacing them,
so every existing behavior is preserved:
- SDK-exposed reasoning_content (OpenAI/Moonshot/DeepSeek SDK) still
wins.
- DeepSeek tool-call ''-pad still fires when the SDK exposes the attr
but the value is None.
- Non-thinking turns with no reasoning leave the field absent, so
_copy_reasoning_content_for_api's cross-provider leak guard (#15748),
promote-from-'reasoning' tier, and thinking-pad tier remain live at
replay time.
- No empty '' gets eagerly written on every assistant turn (which would
have bypassed the read-side ladder and triggered empty thinking-block
insertion in the Anthropic adapter).
Tests: three new TestBuildAssistantMessage cases covering the streaming
promotion path, SDK precedence, and field-absent-when-no-reasoning
invariant.
Credit @Sanjays2402 for the original diagnosis and patch in #16884;
this is a scoped rework that preserves the existing read-side
compensation code as defense in depth.
Refs #16844, #16884, #15250, #15353, #15748.
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.
Same layering concern as the persisted-assistant scrub already removed:
_emit_interim_assistant_message and the final_response return path were
mutating model output broadly. Streaming scrubber covers real leaks
delta-by-delta; these post-stream scrubs were redundant.
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).
The auto-lowered-threshold warning only named the compression model,
making it confusing when the main and aux models are configured with
the same slug but end up with different resolved context lengths (e.g.
OpenRouter's stepfun/step-3.5-flash catalog value vs. a main-model
context_length override). Users couldn't tell whether the warning
reflected two different models or a context-resolution mismatch.
Now includes both 'model (provider)' labels. The aux provider falls
back to the client's base_url hostname when the configured provider
is 'auto', so users see where compression is actually being called.
PR #13734 fixed the concurrent-tool-executor vector (ThreadPoolExecutor
workers didn't inherit the CLI's TLS approval callback). Two vectors
remained that could still land in the deadlocking input() fallback:
1. _spawn_background_review spawns a raw threading.Thread with no
approval callback installed, so any dangerous-command guard the
review agent trips falls back to input() -> deadlock against the
parent's prompt_toolkit TUI (same class as delegate_task subagents,
fixed in 023b1bff1 / #15491). Install a _bg_review_auto_deny
callback at thread start, clear on finally.
2. prompt_dangerous_approval's fallback unconditionally spawned a
daemon thread calling input() when approval_callback was None.
That fallback can never succeed under prompt_toolkit because the
user's Enter goes to pt's raw-mode stdin capture. Detect an active
pt Application via get_app_or_none() and fail closed (deny + log)
instead, so future threads that forget to install a callback
degrade gracefully instead of hanging 60s invisibly.
Regression guards:
- tests/run_agent/test_background_review.py verifies the review
worker thread sees a callable auto-deny callback mid-run and that
the slot is cleared in the finally block.
- tests/tools/test_approval.py TestFailClosedUnderPromptToolkit
verifies prompt_dangerous_approval returns 'deny' fast under a
mocked pt Application, and that a real callback still wins over
the guard.
When tools execute concurrently via ThreadPoolExecutor, worker threads
could not see the thread-local approval/sudo callbacks registered by
the CLI. This caused dangerous-command prompts to fall back to plain
input(), which deadlocks against prompt_toolkit's raw terminal mode.
Capture parent-thread callbacks before launching workers, register
them locally in each _run_tool thread, and clear them on exit.
Mirrors the existing fix pattern from cli.py run_agent() for the
main agent worker thread (GHSA-qg5c-hvr5-hjgr / #13617).
The background skill/memory review agent was created without toolset
restrictions, inheriting the full default tool set. This allowed it to
use terminal, send_message, delegate_task, and other tools outside its
intended scope, potentially performing unrelated side effects after
skill creation.
Restrict the review agent to only memory and skills toolsets by passing
enabled_toolsets=['memory', 'skills'] during AIAgent construction.
Fixes#15204
* 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.
On provider switches mid-session (e.g. MiniMax -> DeepSeek), the source
assistant turn carries a 'reasoning' field written by the prior provider
but no 'reasoning_content' key. _copy_reasoning_content_for_api would
promote that foreign 'reasoning' to 'reasoning_content' on the outbound
DeepSeek request, leaking a cross-provider chain of thought and in
practice causing HTTP 400.
DeepSeek's own _build_assistant_message always pins reasoning_content=''
at creation time for tool-call turns, so the shape (reasoning set,
reasoning_content absent, tool_calls present) is unreachable from
same-provider DeepSeek history — it can only come from a prior provider.
Pad with '' in that case instead of promoting.
Healthy same-provider 'reasoning' promotion (no tool_calls, or on
providers that do not require the empty-string pin) is unchanged.
When _compress_context rotates session_id (compression split), fire
on_session_start(new_sid, boundary_reason="compression",
old_session_id=<old>) on the active context engine. Plugin engines
(e.g. hermes-lcm) use this to preserve DAG lineage across the rollover
instead of re-initializing fresh per-session state.
Built-in ContextCompressor.on_session_start accepts **kwargs and ignores
them — no behavior change for default users.
Closes hermes-lcm#68 symptom: after Hermes compressed and minted a new
physical session, LCM was treating the split as a fresh /new and losing
continuity (compression_count: 1, store_messages: 0, dag_nodes: 0).
Credit: @Tosko4 (PR #13370) — minimized scope to the boundary_reason
signal only; the broader session-lifecycle refactor will be taken in
separate PRs if justified by concrete plugin need.
Temporary background review agents can initialize Hindsight-backed memory clients, but close() alone skips provider teardown. Shut the memory provider down before closing so aiohttp sessions do not leak at process exit.
Made-with: Cursor
The background memory/skill review (_spawn_background_review) has always
forked a new AIAgent passing only model and provider, then relied on
AIAgent.__init__ to re-resolve credentials from env vars. This works for
users with keys in ~/.hermes/.env but silently falls back to env-var
auto-resolution in all cases, which fails for OAuth-only providers,
session-scoped creds, and credential-pool setups where auth can't be
reconstructed from env.
This used to be invisible -- failures were swallowed via logger.debug().
PR 8a2506af4 (Apr 24) surfaced auxiliary failures to the user, which
made the stale bug visible as:
"Auxiliary background review failed: No LLM provider configured"
Fix: pass api_key, base_url, api_mode, and credential_pool from the
parent's live runtime into the fork -- matching how every other
auxiliary path (compression, memory flush, vision, session search)
already inherits the parent's credentials via _current_main_runtime().
Follow-up to PR #16053 (/btw as /background alias). Cleans up the
plumbing added exclusively for the old ephemeral /btw handler and
repairs a broken btw bypass that landed between my refactor and this
follow-up.
run_agent.py:
- Remove persist_session kwarg, instance attr, and _persist_session
short-circuit. Only /btw ever passed persist_session=False; with
/btw gone the default (always persist) is the only behavior anyone
ever wanted.
gateway/run.py:
- Remove the unreachable 'if _cmd_def_inner.name == "btw"' block
(PR #16059). Canonical name for a /btw message is 'background' after
alias resolution — the comparison could never be true, and it called
_handle_btw_command which no longer exists. The /background branch
above it already dispatches /btw correctly.
tests/gateway/test_running_agent_session_toggles.py:
- Fix test_btw_dispatches_mid_run to mock _handle_background_command
(the real dispatch target for /btw) instead of the deleted
_handle_btw_command.
The background skill-review prompt (spawned after N user turns) now instructs
the reviewer to SURVEY existing skills first, identify the CLASS of task, and
PREFER updating/generalizing an existing skill over creating a new narrow one.
This reduces near-duplicate skill accumulation at the source. Catches the
common failure mode where repeated tasks of the same class each spawn their
own specific skill ("fix-my-tauri-error", "fix-my-electron-error") instead
of a single class-level skill ("desktop-app-build-troubleshooting").
Applied to both _SKILL_REVIEW_PROMPT and the **Skills** half of
_COMBINED_REVIEW_PROMPT. Memory-only review prompt unchanged.
Groundwork for the Curator feature (issue #7816) — the creation-side fix.
Curator handles the retirement/consolidation side in a follow-up PR.
Tests assert the behavioral instructions are present (survey, class, update-
over-create, overlap-flagging, opt-out clause) rather than snapshotting the
full prompt text.
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 exposes an OpenAI-compatible endpoint at
`{resource}.openai.azure.com/openai/v1` that accepts the standard
`openai` Python client. Two issues prevented gpt-5.x models from working:
1. `_max_tokens_param()` only sent `max_completion_tokens` for
`api.openai.com` URLs. Azure also requires `max_completion_tokens`
for gpt-5.x models.
2. The `codex_responses` upgrade gate unconditionally upgraded gpt-5.x
to Responses API. Azure does NOT support the Responses API — it serves
gpt-5.x on the regular `/chat/completions` path, causing a 404.
Fix: add `_is_azure_openai_url()` that matches `openai.azure.com` URLs.
- `_max_tokens_param()` now returns `max_completion_tokens` for Azure.
- The `codex_responses` upgrade gate skips Azure so gpt-5.x stays on
`chat_completions` where Azure actually serves it.
- The fallback-provider api_mode picker also recognises Azure and stays
on chat_completions.
- Tests cover max_tokens routing, api_mode behaviour, and URL detection.
gpt-4.x models on Azure are unaffected (already used chat_completions +
max_tokens, which Azure accepts for those models).
Salvage of PR #10086 — rewritten against current main where the
codex_responses upgrade gate gained copilot-acp / explicit-api_mode
exclusions.
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
Previously _copy_reasoning_content_for_api only padded reasoning_content
when the assistant message had tool_calls. DeepSeek V4 thinking mode
requires the field on every assistant turn, including plain text replies
without tool_calls.
- Remove the 'source_msg.get("tool_calls") and' guard
- Update test: plain assistant turns now get padded for DeepSeek/Kimi
Fixes#15213
When a user sends /stop during a streaming API call, the outer poll loop
detects _interrupt_requested and closes the HTTP connection. However, the
inner _call() thread catches the connection error and enters its retry
loop — opening a FRESH connection without checking the interrupt flag.
On slow providers like ollama-cloud, each retry attempt blocks for the
full stream-read timeout (120s+). With 3 retry attempts this caused
510+ second delays between /stop and actual response — the agent appeared
completely unresponsive despite the stop being acknowledged.
Fix: add an _interrupt_requested check at the top of the streaming retry
loop so the agent exits immediately instead of retrying.
Also fix log truncation: all session key logging in gateway/run.py used
[:20] or [:30] slices, which truncated 'agent:main:telegram:dm:5690190437'
(33 chars) to 'agent:main:telegram:' — losing the identifying chat type
and user ID. Replace with full keys to make logs debuggable.
Reported by user Sidharth Pulipaka via Telegram on ollama-cloud provider.
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.
_check_compression_model_feasibility calls get_model_context_length
without provider=, so Codex OAuth users get 1,050,000 (from models.dev
for 'openai') instead of the actual 272,000 limit. This happens because
_infer_provider_from_url maps chatgpt.com → 'openai' (not 'openai-codex'),
skipping the Codex-specific resolution branch entirely.
Result: compression threshold set at 85% of 1.05M = 892K — conversations
never trigger compression, the context grows unbounded, and when gateway
hygiene eventually forces compression, the Codex endpoint drops the
oversized streaming request ('peer closed connection without sending
complete message body').
Fix: forward self.provider to get_model_context_length so provider-
specific resolution branches (Codex OAuth 272K, Copilot live /models,
Nous suffix-match) fire correctly.
Reported by user on GPT 5.5 via Codex OAuth Pro (paste.rs/vsra3).
When the auxiliary compression model's context is smaller than the main
model's compression threshold, _check_compression_model_feasibility
auto-lowers the session threshold. Previously it set:
new_threshold = aux_context
This let the raw message list grow to exactly aux_context tokens. But
compression and flush_memories actually send system_prompt + tool_schemas
+ messages to the aux model. With 50+ tools that overhead is 25-30K
tokens, so the full request overflowed aux with HTTP 400.
Subtract a headroom estimate from aux_context before setting the new
threshold: the actual tool-schema token count (from
estimate_request_tokens_rough) plus a 12K allowance for the system
prompt (not yet built at __init__ time) and flush-instruction overhead.
Clamp to MINIMUM_CONTEXT_LENGTH so the session still starts even with
an unusually heavy tool schema.
This fixes the 'flush_memories overflow on busy toolsets' path that
Teknium flagged — where main and aux can be nominally the same model
but still 400 because the threshold left no room for the request
overhead. Same fix also protects the normal compression summarisation
request on the same binding aux.
Tests: two new regression tests cover the headroom reservation and the
MINIMUM_CONTEXT_LENGTH floor. Two existing tests updated for the new
(lower) threshold values now that empty-tools still produces a 12K
static headroom deduction.
The memory-flush fallback for api_mode='codex_responses' was unconditionally
adding `temperature` to codex_kwargs before calling _run_codex_stream. The
Responses API does not accept temperature on any supported backend:
- chatgpt.com/backend-api/codex rejects it outright
- api.openai.com + gpt-5/o-series reasoning models reject it
- Copilot Responses rejects it on reasoning models
The CodexAuxiliaryClient adapter and the codex_responses transport both
correctly omit temperature — the flush fallback was the only path putting
it back. On errors from the primary aux path (e.g. expired OAuth token),
users saw `⚠ Auxiliary memory flush failed: HTTP 400: Unsupported parameter:
temperature`.
Reported by Garik [NOUS] on GPT-5.5 via Codex OAuth Pro.
Extracts _needs_kimi_tool_reasoning() for symmetry with the existing
_needs_deepseek_tool_reasoning() helper, so _copy_reasoning_content_for_api
uses the same detection logic as _build_assistant_message. Future changes
to either provider's signals now only touch one function.
Adds tests/run_agent/test_deepseek_reasoning_content_echo.py covering:
- All 3 DeepSeek detection signals (provider, model, host)
- Poisoned history replay (empty string fallback)
- Plain assistant turns NOT padded
- Explicit reasoning_content preserved
- Reasoning field promoted to reasoning_content
- Existing Kimi/Moonshot detection intact
- Non-thinking providers left alone
21 tests, all pass.
DeepSeek V4 thinking mode requires reasoning_content on every
assistant message that includes tool_calls. When this field is
missing from persisted history, replaying the session causes
HTTP 400: 'The reasoning_content in the thinking mode must be
passed back to the API.'
Two-part fix (refs #15250):
1. _copy_reasoning_content_for_api: Merge the Kimi-only and
DeepSeek detection into a single needs_tool_reasoning_echo
check. This handles already-poisoned persisted sessions by
injecting an empty reasoning_content on replay.
2. _build_assistant_message: Store reasoning_content='' on new
DeepSeek tool-call messages at creation time, preventing
future session poisoning at the source.
Additional fix:
3. _handle_max_iterations: Add missing call to
_copy_reasoning_content_for_api in the max-iterations flush
path (previously only main loop and flush_memories had it).
Detection covers:
- provider == 'deepseek'
- model name containing 'deepseek' (case-insensitive)
- base URL matching api.deepseek.com (for custom provider)
``run_conversation`` was calling ``memory_manager.sync_all(
original_user_message, final_response)`` at the end of every turn
where both args were present. That gate didn't consider the
``interrupted`` local flag, so an external memory backend received
partial assistant output, aborted tool chains, or mid-stream resets as
durable conversational truth. Downstream recall then treated the
not-yet-real state as if the user had seen it complete, poisoning the
trust boundary between "what the user took away from the turn" and
"what Hermes was in the middle of producing when the interrupt hit".
Extracted the inline sync block into a new private method
``AIAgent._sync_external_memory_for_turn(original_user_message,
final_response, interrupted)`` so the interrupt guard is a single
visible check at the top of the method instead of hidden in a
boolean-and at the call site. That also gives tests a clean seam to
assert on — the pre-fix layout buried the logic inside the 3,000-line
``run_conversation`` function where no focused test could reach it.
The new method encodes three independent skip conditions:
1. ``interrupted`` → skip entirely (the #15218 fix). Applies even
when ``final_response`` and ``original_user_message`` happen to
be populated — an interrupt may have landed between a streamed
reply and the next tool call, so the strings on disk are not
actually the turn the user took away.
2. No memory manager / no final_response / no user message →
preserve existing skip behaviour (nothing new for providerless
sessions, system-initiated refreshes, tool-only turns that never
resolved, etc.).
3. Sync_all / queue_prefetch_all exceptions → swallow. External
memory providers are strictly best-effort; a misconfigured or
offline backend must never block the user from seeing their
response.
The prefetch side-effect is gated on the same interrupt flag: the
user's next message is almost certainly a retry of the same intent,
and a prefetch keyed on the interrupted turn would fire against stale
context.
### Tests (16 new, all passing on py3.11 venv)
``tests/run_agent/test_memory_sync_interrupted.py`` exercises the
helper directly on a bare ``AIAgent`` (``__new__`` pattern that the
interrupt-propagation tests already use). Coverage:
- Interrupted turn with full-looking response → no sync (the fix)
- Interrupted turn with long assistant output → no sync (the interrupt
could have landed mid-stream; strings-on-disk lie)
- Normal completed turn → sync_all + queue_prefetch_all both called
with the right args (regression guard for the positive path)
- No final_response / no user_message / no memory manager → existing
pre-fix skip paths still apply
- sync_all raises → exception swallowed, prefetch still attempted
- queue_prefetch_all raises → exception swallowed after sync succeeded
- 8-case parametrised matrix across (interrupted × final_response ×
original_user_message) asserts sync fires iff interrupted=False AND
both strings are non-empty
Closes#15218
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Extends _repair_tool_call_arguments() to cover the most common local-model
JSON corruption pattern: llama.cpp/Ollama backends emit literal tabs and
newlines inside JSON string values (memory save summaries, file contents,
etc.). Previously fell through to '{}' replacement, losing the call.
Adds two repair passes:
- Pass 0: json.loads(strict=False) + re-serialise to canonical wire form
- Pass 4: escape 0x00-0x1F control chars inside string values, then retry
Ports the core utility from #12068 / PR #12093 without the larger plumbing
change (that PR also replaced json.loads at 8 call sites; current main's
_repair_tool_call_arguments is already the single chokepoint, so the
upgrade happens transparently for every existing caller).
Credit: @truenorth-lj for the original utility design.
4 new regression tests covering literal newlines, tabs, re-serialisation
to strict=True-valid output, and the trailing-comma + control-char
combination case.
When the streaming path (chat completions) assembled tool call deltas and
detected malformed JSON arguments, it set has_truncated_tool_args=True but
passed the broken args through unchanged. This triggered the truncation
handler which returned a partial result and killed the session (/new required).
_many_ malformations are repairable: trailing commas, unclosed brackets,
Python None, empty strings. _repair_tool_call_arguments() already existed
for the pre-API-request path but wasn't called during streaming assembly.
Now when JSON parsing fails during streaming assembly, we attempt repair
via _repair_tool_call_arguments() before flagging as truncated. If repair
succeeds (returns valid JSON), the tool call proceeds normally. Only truly
unrepairable args fall through to the truncation handler.
This prevents the most common session-killing failure mode for models like
GLM-5.1 that produce trailing commas or unclosed brackets.
Tests: 12 new streaming assembly repair tests, all 29 existing repair
tests still passing.
When a session is split by context compression mid-tool-call, an assistant
message may end up with truncated/invalid JSON in tool_calls[*].function.arguments.
On the next turn this is replayed verbatim and providers reject the entire request
with HTTP 400 invalid_tool_call_format, bricking the conversation in a loop that
cannot recover without manual session quarantine.
This patch adds a defensive sanitizer that runs immediately before
client.chat.completions.create() in AIAgent.run_conversation():
- Validates each assistant tool_calls[*].function.arguments via json.loads
- Replaces invalid/empty arguments with '{}'
- Injects a synthetic tool response (or prepends a marker to the existing one)
so downstream messages keep valid tool_call_id pairing
- Logs each repair with session_id / message_index / preview for observability
Defense in depth: corruption can originate from compression splits, manual edits,
or plugin bugs. Sanitizing at the send chokepoint catches all sources.
Adds 7 unit tests covering: truncated JSON, empty string, None, non-string args,
existing matching tool response (no duplicate injection), non-assistant messages
ignored, multiple repairs.
Fixes#15236
Three interrupt-recovery sites in run_agent.py rebuilt self._anthropic_client
with build_anthropic_client(self._anthropic_api_key, ...) unconditionally.
When provider=bedrock + api_mode=anthropic_messages (AnthropicBedrock SDK
path), self._anthropic_api_key is the sentinel 'aws-sdk' — build_anthropic_client
doesn't accept that and the rebuild either crashed or produced a non-functional
client.
Extract a _rebuild_anthropic_client() helper that dispatches to
build_anthropic_bedrock_client(region) when provider='bedrock', falling back
to build_anthropic_client() for native Anthropic and other anthropic_messages
providers (MiniMax, Kimi, Alibaba, etc.). Three inline rebuild sites now call
the helper.
Partial salvage of #14680 by @bsgdigital — only the _rebuild_anthropic_client
helper. The normalize_model_name Bedrock-prefix piece was subsumed by #14664,
and the aux client aws_sdk branch was subsumed by #14770 (both in the same
salvage PR as this commit).
## 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>
Try to activate fallback model after errors was calling get_model_context_length()
without the config_context_length parameter, causing it to fall through to
DEFAULT_FALLBACK_CONTEXT (128K) even when config.yaml has an explicit
model.context_length value (e.g. 204800 for MiniMax-M2.7).
This mirrors the fix already present in switch_model() at line 1988, which
correctly passes config_context_length. The fallback path was missed.
Fixes: context_length forced to 128K on fallback activation
ssl.SSLError (and its subclass ssl.SSLCertVerificationError) inherits from
OSError *and* ValueError via Python's MRO. The is_local_validation_error
check used isinstance(api_error, (ValueError, TypeError)) to detect
programming bugs that should abort immediately — but this inadvertently
caught ssl.SSLError, treating a TLS transport failure as a non-retryable
client error.
The error classifier already maps SSLCertVerificationError to
FailoverReason.timeout with retryable=True (its type name is in
_TRANSPORT_ERROR_TYPES), but the inline isinstance guard was overriding
that classification and triggering an unnecessary abort.
Fix: add ssl.SSLError to the exclusion list alongside the existing
UnicodeEncodeError carve-out so TLS errors fall through to the
classifier's retryable path.
Closes#14367
Claude-style and some Anthropic-tuned models occasionally emit tool
names as class-like identifiers: TodoTool_tool, Patch_tool,
BrowserClick_tool, PatchTool. These failed strict-dict lookup in
valid_tool_names and triggered the 'Unknown tool' self-correction
loop, wasting a full turn of iteration and tokens.
_repair_tool_call already handled lowercase / separator / fuzzy
matches but couldn't bridge the CamelCase-to-snake_case gap or the
trailing '_tool' suffix that Claude sometimes tacks on. Extend it
with two bounded normalization passes:
1. CamelCase -> snake_case (via regex lookbehind).
2. Strip trailing _tool / -tool / tool suffix (case-insensitive,
applied twice so TodoTool_tool reduces all the way: strip
_tool -> TodoTool, snake -> todo_tool, strip 'tool' -> todo).
Cheap fast-paths (lowercase / separator-normalized) still run first
so the common case stays zero-cost. Fuzzy match remains the last
resort unchanged.
Tests: tests/run_agent/test_repair_tool_call_name.py covers the
three original reports (TodoTool_tool, Patch_tool, BrowserClick_tool),
plus PatchTool, WriteFileTool, ReadFile_tool, write-file_Tool,
patch-tool, and edge cases (empty, None, '_tool' alone, genuinely
unknown names).
18 new tests + 17 existing arg-repair tests = 35/35 pass.
Closes#14784
Extracts pool-rotation-room logic into `_pool_may_recover_from_rate_limit`
so single-credential pools no longer block the eager-fallback path on 429.
The existing check `pool is not None and pool.has_available()` lets
fallback fire only after the pool marks every entry as exhausted. With
exactly one credential in the pool (the common shape for Gemini OAuth,
Vertex service accounts, and any personal-key setup), `has_available()`
flips back to True as soon as the cooldown expires — Hermes retries
against the same entry, hits the same daily-quota 429, and burns the
retry budget in a tight loop before ever reaching the configured
`fallback_model`. Observed in the wild as 4+ hours of 429 noise on a
single Gemini key instead of falling through to Vertex as configured.
Rotation is only meaningful with more than one credential — gate on
`len(pool.entries()) > 1`. Multi-credential pools keep the current
wait-for-rotation behaviour unchanged.
Fixes#11314. Related to #8947, #10210, #7230. Narrower scope than
open PRs #8023 (classifier change) and #11492 (503/529 credential-pool
bypass) — this addresses the single-credential 429 case specifically
and does not conflict with either.
Tests: 6 new unit tests in tests/run_agent/test_provider_fallback.py
covering (a) None pool, (b) single-cred available, (c) single-cred in
cooldown, (d) 2-cred available rotates, (e) multi-cred all cooling-down
falls back, (f) many-cred available rotates. All 18 tests in the file
pass.
Previously _handle_credential_pool_error handled 401, 402, and 429
but silently ignored 403. When a provider returns 403 for a revoked or
unauthorised credential (e.g. Nous agent_key invalidated by a newer
login), the pool was never rotated and every subsequent request
continued to use the same failing credential.
Treat 403 the same as 402: immediately mark the current credential
exhausted and rotate to the next pool entry, since a Forbidden response
will not resolve itself with a retry.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When using GitHub Copilot as provider, HTTP 401 errors could cause
Hermes to silently fall back to the next model in the chain instead
of recovering. This adds a one-shot retry mechanism that:
1. Re-resolves the Copilot token via the standard priority chain
(COPILOT_GITHUB_TOKEN -> GH_TOKEN -> GITHUB_TOKEN -> gh auth token)
2. Rebuilds the OpenAI client with fresh credentials and Copilot headers
3. Retries the failed request before falling back
The fix handles the common case where the gho_* OAuth token remains
valid but the httpx client state becomes stale (e.g. after startup
race conditions or long-lived sessions).
Key design decisions:
- Always rebuild client even if token string unchanged (recovers stale state)
- Uses _apply_client_headers_for_base_url() for canonical header management
- One-shot flag guard prevents infinite 401 loops (matches existing pattern
used by Codex/Nous/Anthropic providers)
- No token exchange via /copilot_internal/v2/token (returns 404 for some
account types; direct gho_* auth works reliably)
Tests: 3 new test cases covering end-to-end 401->refresh->retry,
client rebuild verification, and same-token rebuild scenarios.
Docs: Updated providers.md with Copilot auth behavior section.
json.JSONDecodeError inherits from ValueError. The agent loop's
non-retryable classifier at run_agent.py ~L10782 treated any
ValueError/TypeError as a local programming bug and short-circuited
retry. Without a carve-out, a transient JSONDecodeError from a
provider that returned a malformed response body, a truncated stream,
or a router-layer corruption would fail the turn immediately.
Add JSONDecodeError to the existing UnicodeEncodeError exclusion
tuple so the classified-retry logic (which already handles 429/529/
context-overflow/etc.) gets to run on bad-JSON errors.
Tests (tests/run_agent/test_jsondecodeerror_retryable.py):
- JSONDecodeError: NOT local validation
- UnicodeEncodeError: NOT local validation (existing carve-out)
- bare ValueError: IS local validation (programming bug)
- bare TypeError: IS local validation (programming bug)
- source-level assertion that run_agent.py still carries the carve-out
(guards against accidental revert)
Closes#14782
Two related paths where Codex auth failures silently swallowed the
fallback chain instead of switching to the next provider:
1. cli.py — _ensure_runtime_credentials() calls resolve_runtime_provider()
before each turn. When provider is explicitly configured (not "auto"),
an AuthError from token refresh is re-raised and printed as a bold-red
error, returning False before the agent ever starts. The fallback chain
was never tried. Fix: on AuthError, iterate fallback_providers and
switch to the first one that resolves successfully.
2. run_agent.py — inside the codex_responses validity gate (inner retry
loop), response.status in {"failed","cancelled"} with non-empty output
items was treated as a valid response and broke out of the retry loop,
reaching _normalize_codex_response() outside the fallback machinery.
That function raises RuntimeError on status="failed", which propagates
to the outer except with no fallback logic. Fix: detect terminal status
codes before the output_items check and set response_invalid=True so
the existing fallback chain fires normally.
- Load prompt_caching.cache_ttl in AIAgent (5m default, 1h opt-in)
- Document DEFAULT_CONFIG and developer guide example
- Add unit tests for default, 1h, and invalid TTL fallback
Made-with: Cursor
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).
Closes#11616.
The agent's API retry loop hardcoded max_retries = 3, so users with
fallback providers on flaky primaries burned through ~3 × provider
timeout (e.g. 3 × 180s = 9 minutes) before their fallback chain got a
chance to kick in.
Expose a new config key:
agent:
api_max_retries: 3 # default unchanged
Set it to 1 for fast failover when you have fallback providers, or
raise it if you prefer longer tolerance on a single provider. Values
< 1 are clamped to 1 (single attempt, no retry); non-integer values
fall back to the default.
This wraps the Hermes-level retry loop only — the OpenAI SDK's own
low-level retries (max_retries=2 default) still run beneath this for
transient network errors.
Changes:
- hermes_cli/config.py: add agent.api_max_retries default 3 with comment.
- run_agent.py: read self._api_max_retries in AIAgent.__init__; replace
hardcoded max_retries = 3 in the retry loop with self._api_max_retries.
- cli-config.yaml.example: documented example entry.
- hermes_cli/tips.py: discoverable tip line.
- tests/run_agent/test_api_max_retries_config.py: 4 tests covering
default, override, clamp-to-one, and invalid-value fallback.
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).
Follow-up for #13862 — the post-init api_mode upgrade at __init__ (direct OpenAI /
gpt-5-requires-responses path) runs AFTER the eager transport warm. Clear the cache
so the stale chat_completions entry is evicted.
Cosmetic: correctness was already fine since _get_transport() keys by current
api_mode, but this avoids leaving unused cache state behind.
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#67318. Some open models (notably Gemma
variants served via OpenRouter) emit tool calls as XML blocks inside
assistant content instead of via the structured tool_calls field:
<function name="read_file"><parameter name="path">/tmp/x</parameter></function>
<tool_call>{"name":"x"}</tool_call>
<function_calls>[{...}]</function_calls>
Left unstripped, this raw XML leaked to gateway users (Discord, Telegram,
Matrix, Feishu, Signal, WhatsApp, etc.) and the CLI, since hermes-agent's
existing reasoning-tag stripper handled only <think>/<thinking>/<thought>
variants.
Extend _strip_think_blocks (run_agent.py) and _strip_reasoning_tags
(cli.py) to cover:
* <tool_call>, <tool_calls>, <tool_result>
* <function_call>, <function_calls>
* <function name="..."> ... </function> (Gemma-style)
The <function> variant is boundary-gated (only strips when the tag sits
at start-of-line or after sentence punctuation AND carries a name="..."
attribute) so prose mentions like 'Use <function> declarations in JS'
are preserved. Dangling <function name="..."> with no close is
intentionally left visible — matches OpenClaw's asymmetry so a truncated
streaming tail still reaches the user.
Tests: 9 new cases in TestStripThinkBlocks (run_agent) + 9 in new file
tests/run_agent/test_strip_reasoning_tags_cli.py. Covers Qwen-style
<tool_call>, Gemma-style <function name="...">, multi-line payloads,
prose preservation, stray close tags, dangling open tags, and mixed
reasoning+tool_call content.
Note: this port covers the post-streaming final-text path, which is what
gateway adapters and CLI display consume. Extending the per-delta stream
filter in gateway/stream_consumer.py to hide these tags live as they
stream is a separate follow-up; for now users may see raw XML briefly
during a stream before the final cleaned text replaces it.
Refs: openclaw/openclaw#67318
