When streaming fails after partial content delivery (e.g. OpenRouter
timeout kills connection mid-response), the stub response now carries
the accumulated streamed text instead of content=None.
Two fixes:
1. The partial-stream stub response includes recovered content from
_current_streamed_assistant_text — the text that was already
delivered to the user via stream callbacks before the connection
died.
2. The empty response recovery chain now checks for partial stream
content BEFORE falling back to _last_content_with_tools (prior
turn content) or wasting API calls on retries. This prevents:
- Showing wrong content from a prior turn
- Burning 3+ unnecessary retry API calls
- Falling through to '(empty)' when the user already saw content
The root cause: OpenRouter has a ~125s inactivity timeout. When
Anthropic's SSE stream goes silent during extended reasoning, the
proxy kills the connection. The model's text was already partially
streamed but the stub discarded it, triggering the empty recovery
chain which would show stale prior-turn content or waste retries.
OpenCode Zen was in _DOT_TO_HYPHEN_PROVIDERS, causing all dotted model
names (minimax-m2.5-free, gpt-5.4, glm-5.1) to be mangled. The fix:
Layer 1 (model_normalize.py): Remove opencode-zen from the blanket
dot-to-hyphen set. Add an explicit block that preserves dots for
non-Claude models while keeping Claude hyphenated (Zen's Claude
endpoint uses anthropic_messages mode which expects hyphens).
Layer 2 (run_agent.py _anthropic_preserve_dots): Add opencode-zen and
zai to the provider allowlist. Broaden URL check from opencode.ai/zen/go
to opencode.ai/zen/ to cover both Go and Zen endpoints. Add bigmodel.cn
for ZAI URL detection.
Also adds glm-5.1 to ZAI model lists in models.py and setup.py.
Closes#7710
Salvaged from contributions by:
- konsisumer (PR #7739, #7719)
- DomGrieco (PR #8708)
- Esashiero (PR #7296)
- sharziki (PR #7497)
- XiaoYingGee (PR #8750)
- APTX4869-maker (PR #8752)
- kagura-agent (PR #7157)
_check_compression_model_feasibility() called get_model_context_length()
without passing config_context_length, so custom endpoints that do not
support /models API queries always fell through to the 128K default,
ignoring auxiliary.compression.context_length in config.yaml.
Fix: read auxiliary.compression.context_length from config and pass it
as config_context_length (highest-priority hint) so the user-configured
value is always respected regardless of API availability.
Fixes#8499
Three fixes for the (empty) response bug affecting open reasoning models:
1. Allow retries after prefill exhaustion — models like mimo-v2-pro always
populate reasoning fields via OpenRouter, so the old 'not _has_structured'
guard on the retry path blocked retries for EVERY reasoning model after
the 2 prefill attempts. Now: 2 prefills + 3 retries = 6 total attempts
before (empty).
2. Reset prefill/retry counters on tool-call recovery — the counters
accumulated across the entire conversation, never resetting during
tool-calling turns. A model cycling empty→prefill→tools→empty burned
both prefill attempts and the third empty got zero recovery. Now
counters reset when prefill succeeds with tool calls.
3. Strip think blocks before _truly_empty check — inline <think> content
made the string non-empty, skipping both retry paths.
Reported by users on Telegram with xiaomi/mimo-v2-pro and qwen3.5 models.
Reproduced: qwen3.5-9b emits tool calls as XML in reasoning field instead
of proper function calls, causing content=None + tool_calls=None + reasoning
with embedded <tool_call> XML. Prefill recovery works but counter
accumulation caused permanent (empty) in long sessions.
Previously, all invalid API responses (choices=None) were diagnosed
as 'fast response often indicates rate limiting' regardless of actual
response time or error code. A 738s Cloudflare 524 timeout was labeled
as 'fast response' and 'possible rate limit'.
Now extracts the error code from response.error and classifies:
- 524: upstream provider timed out (Cloudflare)
- 504: upstream gateway timeout
- 429: rate limited by upstream provider
- 500/502: upstream server error
- 503/529: upstream provider overloaded
- Other codes: shown with code number
- No code + <10s: likely rate limited (timing heuristic)
- No code + >60s: likely upstream timeout
- No code + 10-60s: neutral response time
All downstream messages (retry status, final error, interrupt message)
now use the classified hint instead of generic rate-limit language.
Reported by community member Lumen Radley (MiMo provider timeouts).
Gemma 4 (26B/31B) uses <thought>...</thought> to wrap its reasoning
output. This tag was not included in the existing list of reasoning tag
variants stripped by _strip_think_blocks(), causing raw thinking blocks
to leak into the visible response.
Added a new re.sub() line for <thought> and extended the cleanup regex
to include 'thought' alongside the existing variants.
Fixes#6148
When running inside WSL (Windows Subsystem for Linux), inject a hint into
the system prompt explaining that the Windows host filesystem is mounted
at /mnt/c/, /mnt/d/, etc. This lets the agent naturally translate Windows
paths (Desktop, Documents) to their /mnt/ equivalents without the user
needing to configure anything.
Uses the existing is_wsl() detection from hermes_constants (cached,
checks /proc/version for 'microsoft'). Adds build_environment_hints()
in prompt_builder.py — extensible for Termux, Docker, etc. later.
Closes the UX gap where WSL users had to manually explain path
translation to the agent every session.
Adds an optional focus topic to /compress: `/compress database schema`
guides the summariser to preserve information related to the focus topic
(60-70% of summary budget) while compressing everything else more aggressively.
Inspired by Claude Code's /compact <focus>.
Changes:
- context_compressor.py: focus_topic parameter on _generate_summary() and
compress(); appends FOCUS TOPIC guidance block to the LLM prompt
- run_agent.py: focus_topic parameter on _compress_context(), passed through
to the compressor
- cli.py: _manual_compress() extracts focus topic from command string,
preserves existing manual_compression_feedback integration (no regression)
- gateway/run.py: _handle_compress_command() extracts focus from event args
and passes through — full gateway parity
- commands.py: args_hint="[focus topic]" on /compress CommandDef
Salvaged from PR #7459 (CLI /compress focus only — /context command deferred).
15 new tests across CLI, compressor, and gateway.
* feat: component-separated logging with session context and filtering
Phase 1 — Gateway log isolation:
- gateway.log now only receives records from gateway.* loggers
(platform adapters, session management, slash commands, delivery)
- agent.log remains the catch-all (all components)
- errors.log remains WARNING+ catch-all
- Moved gateway.log handler creation from gateway/run.py into
hermes_logging.setup_logging(mode='gateway') with _ComponentFilter
Phase 2 — Session ID injection:
- Added set_session_context(session_id) / clear_session_context() API
using threading.local() for per-thread session tracking
- _SessionFilter enriches every log record with session_tag attribute
- Log format: '2026-04-11 10:23:45 INFO [session_id] logger.name: msg'
- Session context set at start of run_conversation() in run_agent.py
- Thread-isolated: gateway conversations on different threads don't leak
Phase 3 — Component filtering in hermes logs:
- Added --component flag: hermes logs --component gateway|agent|tools|cli|cron
- COMPONENT_PREFIXES maps component names to logger name prefixes
- Works with all existing filters (--level, --session, --since, -f)
- Logger name extraction handles both old and new log formats
Files changed:
- hermes_logging.py: _SessionFilter, _ComponentFilter, COMPONENT_PREFIXES,
set/clear_session_context(), gateway.log creation in setup_logging()
- gateway/run.py: removed redundant gateway.log handler (now in hermes_logging)
- run_agent.py: set_session_context() at start of run_conversation()
- hermes_cli/logs.py: --component filter, logger name extraction
- hermes_cli/main.py: --component argument on logs subparser
Addresses community request for component-separated, filterable logging.
Zero changes to existing logger names — __name__ already provides hierarchy.
* fix: use LogRecord factory instead of per-handler _SessionFilter
The _SessionFilter approach required attaching a filter to every handler
we create. Any handler created outside our _add_rotating_handler (like
the gateway stderr handler, or third-party handlers) would crash with
KeyError: 'session_tag' if it used our format string.
Replace with logging.setLogRecordFactory() which injects session_tag
into every LogRecord at creation time — process-global, zero per-handler
wiring needed. The factory is installed at import time (before
setup_logging) so session_tag is available from the moment hermes_logging
is imported.
- Idempotent: marker attribute prevents double-wrapping on module reload
- Chains with existing factory: won't break third-party record factories
- Removes _SessionFilter from _add_rotating_handler and setup_verbose_logging
- Adds tests: record factory injection, idempotency, arbitrary handler compat
The _get_budget_warning() method already returned None unconditionally —
the entire budget warning system was disabled. Remove all dead code:
- _BUDGET_WARNING_RE regex
- _strip_budget_warnings_from_history() function and its call site
- Both injection blocks (concurrent + sequential tool execution)
- _get_budget_warning() method
- 7 tests for the removed functions
The budget exhaustion grace call system (_budget_exhausted_injected,
_budget_grace_call) is a separate recovery mechanism and is preserved.
Normalize api_messages before each API call for consistent prefix
matching across turns:
1. Strip leading/trailing whitespace from system prompt parts
2. Strip leading/trailing whitespace from message content strings
3. Normalize tool-call arguments to compact sorted JSON
This enables KV cache reuse on local inference servers (llama.cpp,
vLLM, Ollama) and improves cache hit rates for cloud providers.
All normalization operates on the api_messages copy — the original
conversation history in messages is never mutated. Tool-call JSON
normalization creates new dicts via spread to avoid the shallow-copy
mutation bug in the original PR.
Salvaged from PR #7875 by @waxinz with mutation fix.
Switch estimate_tokens_rough(), estimate_messages_tokens_rough(), and
estimate_request_tokens_rough() from floor division (len // 4) to
ceiling division ((len + 3) // 4). Short texts (1-3 chars) previously
estimated as 0 tokens, causing the compressor and pre-flight checks to
systematically undercount when many short tool results are present.
Also replaced the inline duplicate formula in run_conversation()
(total_chars // 4) with a call to the shared
estimate_messages_tokens_rough() function.
Updated 4 tests that hardcoded floor-division expected values.
Related: issue #6217, PR #6629
Add display.interim_assistant_messages config (enabled by default) that
forwards completed assistant commentary between tool calls to the user
as separate chat messages. Models already emit useful status text like
'I'll inspect the repo first.' — this surfaces it on Telegram, Discord,
and other messaging platforms instead of swallowing it.
Independent from tool_progress and gateway streaming. Disabled for
webhooks. Uses GatewayStreamConsumer when available, falls back to
direct adapter send. Tracks response_previewed to prevent double-delivery
when interim message matches the final response.
Also fixes: cursor not stripped from fallback prefix in stream consumer
(affected continuation calculation on no-edit platforms like Signal).
Cherry-picked from PR #7885 by asheriif, default changed to enabled.
Fixes#5016
Three root causes of the 'agent stops mid-task' gateway bug:
1. Compression threshold floor (64K tokens minimum)
- The 50% threshold on a 100K-context model fired at 50K tokens,
causing premature compression that made models lose track of
multi-step plans. Now threshold_tokens = max(50% * context, 64K).
- Models with <64K context are rejected at startup with a clear error.
2. Budget warning removal — grace call instead
- Removed the 70%/90% iteration budget warnings entirely. These
injected '[BUDGET WARNING: Provide your final response NOW]' into
tool results, causing models to abandon complex tasks prematurely.
- Now: no warnings during normal execution. When the budget is
actually exhausted (90/90), inject a user message asking the model
to summarise, allow one grace API call, and only then fall back
to _handle_max_iterations.
3. Activity touches during long terminal execution
- _wait_for_process polls every 0.2s but never reported activity.
The gateway's inactivity timeout (default 1800s) would fire during
long-running commands that appeared 'idle.'
- Now: thread-local activity callback fires every 10s during the
poll loop, keeping the gateway's activity tracker alive.
- Agent wires _touch_activity into the callback before each tool call.
Also: docs update noting 64K minimum context requirement.
Closes#7915 (root cause was agent-loop termination, not Weixin delivery limits).
Replace the verbose_logging-gated logging.exception() with an
unconditional logger.debug(exc_info=True). The full traceback now
always lands in agent.log when debug logging is enabled, without
requiring the verbose_logging flag or spamming the console.
Previously, production errors in the 700-line response processing
block (normalization, tool dispatch, final response handling) were
logged as one-line messages with the traceback hidden behind
verbose_logging — making post-mortem debugging difficult.
All retry counters (_invalid_tool_retries, _invalid_json_retries,
_empty_content_retries, _incomplete_scratchpad_retries,
_codex_incomplete_retries) are initialized to 0 at the top of
run_conversation() (lines 7566-7570). The hasattr guards added before
the reset block existed are now dead code — the attributes always exist.
Removed 7 redundant hasattr checks (5 original targets + 2 bonus for
_codex_incomplete_retries found during cleanup).
When _try_activate_fallback() switches to a new provider, retry_count was
reset to 0 but compression_attempts and primary_recovery_attempted were
not. This meant a fallback provider that hit context overflow would only
get the leftover compression budget from the failed primary provider,
and transport recovery was blocked because the flag was still True from
the old provider's attempt.
Reset both counters at all 5 fallback activation sites inside the retry
loop so each fallback provider gets a fresh compression budget (3 attempts)
and its own transport recovery opportunity.
When replaying codex_reasoning_items from previous turns,
duplicate item IDs (rs_*) could appear in the input array,
causing HTTP 400 "Duplicate item found" errors from the
OpenAI Responses API.
Add seen_item_ids tracking in both _chat_messages_to_responses_input()
and _preflight_codex_input_items() to skip already-added reasoning
items by their ID.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The interrupt mechanism in tools/interrupt.py used a process-global
threading.Event. In the gateway, multiple agents run concurrently in
the same process via run_in_executor. When any agent was interrupted
(user sends a follow-up message), the global flag killed ALL agents'
running tools — terminal commands, browser ops, web requests — across
all sessions.
Changes:
- tools/interrupt.py: Replace single threading.Event with a set of
interrupted thread IDs. set_interrupt() targets a specific thread;
is_interrupted() checks the current thread. Includes a backward-
compatible _ThreadAwareEventProxy for legacy _interrupt_event usage.
- run_agent.py: Store execution thread ID at start of run_conversation().
interrupt() and clear_interrupt() pass it to set_interrupt() so only
this agent's thread is affected.
- tools/code_execution_tool.py: Use is_interrupted() instead of
directly checking _interrupt_event.is_set().
- tools/process_registry.py: Same — use is_interrupted().
- tests: Update interrupt tests for per-thread semantics. Add new
TestPerThreadInterruptIsolation with two tests verifying cross-thread
isolation.
Models that do not use <think> tags (e.g. GLM-4.7 on NVIDIA Build,
minimax) may return content=None or empty string when truncated. The
previous _thinking_exhausted check treated any None/empty content as
thinking-budget exhaustion, causing these models to always show the
'Thinking Budget Exhausted' error instead of attempting continuation.
Fix: gate the exhaustion check on _has_think_tags — only trigger the
exhaustion path when the model actually produced reasoning blocks
(<think>, <thinking>, <reasoning>, <REASONING_SCRATCHPAD>). Models
without think tags now fall through to the normal continuation retry
logic (up to 3 attempts).
Fixes#7729
When API routers rewrite finish_reason from "length" to "tool_calls",
truncated JSON arguments bypassed the length handler and wasted 3
retry attempts in the generic JSON validation loop. Now detects
truncation patterns in tool call arguments regardless of finish_reason.
Fixes#7680
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Two-phase design so the warning fires before the user's first message
on every platform:
Phase 1 (__init__):
_check_compression_model_feasibility() runs during agent construction.
Resolves the auxiliary compression model (same chain as call_llm with
task='compression'), compares its context length to the main model's
compression threshold. If too small, emits via _emit_status() (prints
for CLI) and stores the warning in _compression_warning.
Phase 2 (run_conversation, first call):
_replay_compression_warning() re-sends the stored warning through
status_callback — which the gateway wires AFTER construction. The
warning is then cleared so it only fires once.
This ensures:
- CLI users see the warning immediately at startup (right after the
context limit line)
- Gateway users (Telegram, Discord, Slack, WhatsApp, Signal, Matrix,
Mattermost, Home Assistant, DingTalk, etc.) receive it via
status_callback('lifecycle', ...) on their first message
- logger.warning() always hits agent.log regardless of platform
Also warns when no auxiliary LLM provider is configured at all.
Entire check wrapped in try/except — never blocks startup.
11 tests covering: core warning logic, boundary conditions, exception
safety, two-phase store+replay, gateway callback wiring, and
single-delivery guarantee.
Matrix gateway: fix sync loop never dispatching events (#5819)
- _sync_loop() called client.sync() but never called handle_sync()
to dispatch events to registered callbacks — _on_room_message was
registered but never fired for new messages
- Store next_batch token from initial sync and pass as since= to
subsequent incremental syncs (was doing full initial sync every time)
- 17 comments, confirmed by multiple users on matrix.org
Feishu docs: add interactive card configuration for approvals (#6893)
- Error 200340 is a Feishu Developer Console configuration issue,
not a code bug — users need to enable Interactive Card capability
and configure Card Request URL
- Added required 3-step setup instructions to feishu.md
- Added troubleshooting entry for error 200340
- 17 comments from Feishu users
Copilot provider drift: detect GPT-5.x Responses API requirement (#3388)
- GPT-5.x models are rejected on /v1/chat/completions by both OpenAI
and OpenRouter (unsupported_api_for_model error)
- Added _model_requires_responses_api() to detect models needing
Responses API regardless of provider
- Applied in __init__ (covers OpenRouter primary users) and in
_try_activate_fallback() (covers Copilot->OpenRouter drift)
- Fixed stale comment claiming gateway creates fresh agents per message
(it caches them via _agent_cache since the caching was added)
- 7 comments, reported on Copilot+Telegram gateway
Based on PR #7285 by @kshitijk4poor.
Two bugs affecting Qwen OAuth users:
1. Wrong context window — qwen3-coder-plus showed 128K instead of 1M.
Added specific entries before the generic qwen catch-all:
- qwen3-coder-plus: 1,000,000 (corrected from PR's 1,048,576 per
official Alibaba Cloud docs and OpenRouter)
- qwen3-coder: 262,144
2. Random stopping — max_tokens was suppressed for Qwen Portal, so the
server applied its own low default. Reasoning models exhaust that on
thinking tokens. Now: honor explicit max_tokens, default to 65536
when unset.
Co-authored-by: kshitijk4poor <82637225+kshitijk4poor@users.noreply.github.com>
Aligns MiniMax provider with official API documentation. Fixes 6 bugs:
transport mismatch (openai_chat -> anthropic_messages), credential leak
in switch_model(), prompt caching sent to non-Anthropic endpoints,
dot-to-hyphen model name corruption, trajectory compressor URL routing,
and stale doctor health check.
Also corrects context window (204,800), thinking support (manual mode),
max output (131,072), and model catalog (M2 family only on /anthropic).
Source: https://platform.minimax.io/docs/api-reference/text-anthropic-api
Co-authored-by: kshitijk4poor <kshitijk4poor@users.noreply.github.com>
The pre_llm_call plugin hook receives session_id, user_message,
conversation_history, is_first_turn, model, and platform — but not
the sender's user_id. This means plugins cannot perform per-user
access control (e.g. restricting knowledge base recall to authorized
users).
The gateway already passes source.user_id as user_id to AIAgent,
which stores it in self._user_id. This change forwards it as
sender_id in the pre_llm_call kwargs so plugins can use it for
ACL decisions.
For CLI sessions where no user_id exists, sender_id defaults to
empty string. Plugins can treat empty sender_id as a trusted local
call (the owner is at the terminal) or deny it depending on their
ACL policy.
_is_oauth_token() returned True for any key not starting with 'sk-ant-api',
which means MiniMax and Alibaba API keys were falsely treated as Anthropic
OAuth tokens. This triggered the Claude Code compatibility path:
- All tool names prefixed with mcp_ (e.g. mcp_terminal, mcp_web_search)
- System prompt injected with 'You are Claude Code' identity
- 'Hermes Agent' replaced with 'Claude Code' throughout
Fix: Make _is_oauth_token() positively identify Anthropic OAuth tokens by
their key format instead of using a broad catch-all:
- sk-ant-* (but not sk-ant-api-*) -> setup tokens, managed keys
- eyJ* -> JWTs from Anthropic OAuth flow
- Everything else -> False (MiniMax, Alibaba, etc.)
Reported by stefan171.
- Remove auto-activation: when context.engine is 'compressor' (default),
plugin-registered engines are NOT used. Users must explicitly set
context.engine to a plugin name to activate it.
- Add curses_radiolist() to curses_ui.py: single-select radio picker
with keyboard nav + text fallback, matching curses_checklist pattern.
- Rewrite cmd_toggle() as composite plugins UI:
Top section: general plugins with checkboxes (existing behavior)
Bottom section: provider plugin categories (Memory Provider, Context Engine)
with current selection shown inline. ENTER/SPACE on a category opens
a radiolist sub-screen for single-select configuration.
- Add provider discovery helpers: _discover_memory_providers(),
_discover_context_engines(), config read/save for memory.provider
and context.engine.
- Add tests: radiolist non-TTY fallback, provider config save/load,
discovery error handling, auto-activation removal verification.
Follow-up fixes for the context engine plugin slot (PR #5700):
- Enhance ContextEngine ABC: add threshold_percent, protect_first_n,
protect_last_n as class attributes; complete update_model() default
with threshold recalculation; clarify on_session_end() lifecycle docs
- Add ContextCompressor.update_model() override for model/provider/
base_url/api_key updates
- Replace all direct compressor internal access in run_agent.py with
ABC interface: switch_model(), fallback restore, context probing
all use update_model() now; _context_probed guarded with getattr/
hasattr for plugin engine compatibility
- Create plugins/context_engine/ directory with discovery module
(mirrors plugins/memory/ pattern) — discover_context_engines(),
load_context_engine()
- Add context.engine config key to DEFAULT_CONFIG (default: compressor)
- Config-driven engine selection in run_agent.__init__: checks config,
then plugins/context_engine/<name>/, then general plugin system,
falls back to built-in ContextCompressor
- Wire on_session_end() in shutdown_memory_provider() at real session
boundaries (CLI exit, /reset, gateway expiry)
- PluginContext.register_context_engine() lets plugins replace the
built-in ContextCompressor with a custom ContextEngine implementation
- PluginManager stores the registered engine; only one allowed
- run_agent.py checks for a plugin engine at init before falling back
to the default ContextCompressor
- reset_session_state() now calls engine.on_session_reset() instead of
poking internal attributes directly
- ContextCompressor.on_session_reset() handles its own internals
(_context_probed, _previous_summary, etc.)
- 19 new tests covering ABC contract, defaults, plugin slot registration,
rejection of duplicates/non-engines, and compressor reset behavior
- All 34 existing compressor tests pass unchanged
When models return empty responses (no content, no tool calls, no
reasoning), Hermes previously retried 3 times silently then fell through
to '(empty)' — without ever trying the fallback provider chain. Users on
GLM-4.5-Air and similar models experienced what appeared to be a
complete hang, especially in gateway (Telegram/Discord) contexts where
the silent retries produced zero feedback.
Changes:
- After exhausting 3 empty retries, attempt _try_activate_fallback()
before giving up with '(empty)'. If fallback succeeds, reset retry
counter and continue the conversation loop with the new provider.
- Replace all _vprint() calls in recovery paths with _emit_status(),
which surfaces messages through both CLI (_vprint with force=True)
and gateway (status_callback -> adapter.send). Users now see:
* '⚠️ Empty response from model — retrying (N/3)' during retries
* '⚠️ Model returning empty responses — switching to fallback...'
* '↻ Switched to fallback: <model> (<provider>)' on success
* '❌ Model returned no content after all retries [and fallback]'
- Add logger.warning() throughout empty response paths for log file
visibility (model name, provider, retry counts).
- Upgrade _last_content_with_tools fallback from logger.debug to
logger.info + _emit_status so recovery is visible.
- Upgrade thinking-only prefill continuation to use _emit_status.
Tests:
- test_empty_response_triggers_fallback_provider: verifies fallback
activation after 3 empty retries produces content from fallback model
- test_empty_response_fallback_also_empty_returns_empty: verifies
graceful degradation when fallback also returns empty
- test_empty_response_emits_status_for_gateway: verifies _emit_status
is called during retries so gateway users see feedback
Addresses #7180.
Add a close() method to AIAgent that acts as a single entry point for
releasing all resources held by an agent instance. This prevents zombie
process accumulation on long-running gateway deployments by explicitly
cleaning up:
- Background processes tracked in ProcessRegistry
- Terminal sandbox environments
- Browser daemon sessions
- Active child agents (subagent delegation)
- OpenAI/httpx client connections
Each cleanup step is independently guarded so a failure in one does not
prevent the rest. The method is idempotent and safe to call multiple
times.
Also simplifies the background review cleanup to use close() instead
of manually closing the OpenAI client.
Ref: #7131
When _build_api_kwargs() throws an exception, the except handler in
the retry loop referenced api_kwargs before it was assigned. This
caused an UnboundLocalError that masked the real error, making
debugging impossible for the user.
Two _dump_api_request_debug() calls in the except block (non-retryable
client error path and max-retries-exhausted path) both accessed
api_kwargs without checking if it was assigned.
Fix: initialize api_kwargs = None before the retry loop and guard both
dump calls. Now the real error surfaces instead of the masking
UnboundLocalError.
Reported by Discord user gruman0.
`delegate_task` silently truncated batch tasks to 3 — the model sends
5 tasks, gets results for 3, never told 2 were dropped. Now returns a
clear tool_error explaining the limit and how to fix it.
The limit is configurable via:
- delegation.max_concurrent_children in config.yaml (priority 1)
- DELEGATION_MAX_CONCURRENT_CHILDREN env var (priority 2)
- default: 3
Uses the same _load_config() path as the rest of delegate_task for
consistent config priority. Clamps to min 1, warns on non-integer
config values.
Also removes the hardcoded maxItems: 3 from the JSON schema — the
schema was blocking the model from even attempting >3 tasks before
the runtime check could fire. The runtime check gives a much more
actionable error message.
Backwards compatible: default remains 3, existing configs unchanged.
When delegation.base_url routes subagents to a different endpoint, the
correct URL was passed through _resolve_delegation_credentials() and
_build_child_agent() into AIAgent.__init__(), but self.base_url could
fall out of sync with client_kwargs["base_url"] — the value the OpenAI
client actually uses.
This caused billing_base_url in session records to show the parent's
endpoint while actual API calls went to the correct delegation target.
Keep self.base_url in sync with client_kwargs after the credential
resolution block, matching the existing pattern for self.api_key.
Fixes#6825
Broaden the UnicodeEncodeError recovery to handle systems with ASCII-only
locale (LANG=C, Chromebooks) where ANY non-ASCII character causes encoding
failure, not just lone surrogates.
Changes:
- Add _strip_non_ascii() and _sanitize_messages_non_ascii() helpers that
strip all non-ASCII characters from message content, name, and tool_calls
- Update the UnicodeEncodeError handler to detect ASCII codec errors and
fall back to non-ASCII sanitization after surrogate check fails
- Sanitize tool_calls arguments and name fields (not just content)
- Fix bare .encode() in cli.py suspend handler to use explicit utf-8
- Add comprehensive test suite (17 tests)
When switching models at runtime, the config_context_length override
was not being passed to the new context compressor instance. This
meant the user-specified context length from config.yaml was lost
after a model switch.
- Store _config_context_length on AIAgent instance during __init__
- Pass _config_context_length when creating new ContextCompressor in switch_model
- Add test to verify config_context_length is preserved across model switches
Fixes: quando estamos alterando o modelo não está alterando o tamanho do contexto
Automated dead code audit using vulture + coverage.py + ast-grep intersection,
confirmed by Opus deep verification pass. Every symbol verified to have zero
production callers (test imports excluded from reachability analysis).
Removes ~1,534 lines of dead production code across 46 files and ~1,382 lines
of stale test code. 3 entire files deleted (agent/builtin_memory_provider.py,
hermes_cli/checklist.py, tests/hermes_cli/test_setup_model_selection.py).
Co-authored-by: alt-glitch <balyan.sid@gmail.com>
The _call_anthropic() streaming path never updated last_chunk_time during
the event loop — only once at stream start. The stale stream detector in
the outer poll loop uses this timer, so any Anthropic stream longer than
180s was killed even when events were actively arriving. This self-inflicted
a RemoteProtocolError that users saw as:
'⚠️ Connection to provider dropped (RemoteProtocolError). Reconnecting…'
The _call_chat_completions() path already updates last_chunk_time on every
chunk (line 4475). This brings _call_anthropic() to parity.
Also adds deltas_were_sent tracking to the Anthropic text_delta path so
the retry loop knows not to retry after partial delivery (prevents
duplicated output on connection drops mid-stream).
Reported-by: Discord users (Castellani, Codename_11)
The hardcoded User-Agent 'KimiCLI/1.3' is outdated — Kimi CLI is now at
v1.30.0. The stale version string causes intermittent 403 errors from
Kimi's coding endpoint ('only available for Coding Agents').
Update all 8 occurrences across run_agent.py, auxiliary_client.py, and
doctor.py to 'KimiCLI/1.30.0' to match the current official Kimi CLI.
Extends the /fast command to support Anthropic's Fast Mode beta in addition
to OpenAI Priority Processing. When enabled on Claude Opus 4.6, adds
speed:"fast" and the fast-mode-2026-02-01 beta header to API requests for
~2.5x faster output token throughput.
Changes:
- hermes_cli/models.py: Add _ANTHROPIC_FAST_MODE_MODELS registry,
model_supports_fast_mode() now recognizes Claude Opus 4.6,
resolve_fast_mode_overrides() returns {speed: fast} for Anthropic
vs {service_tier: priority} for OpenAI
- agent/anthropic_adapter.py: Add _FAST_MODE_BETA constant,
build_anthropic_kwargs() accepts fast_mode=True which injects
speed:fast + beta header via extra_headers (skipped for third-party
Anthropic-compatible endpoints like MiniMax)
- run_agent.py: Pass fast_mode to build_anthropic_kwargs in the
anthropic_messages path of _build_api_kwargs()
- cli.py: Update _handle_fast_command with provider-aware messaging
(shows 'Anthropic Fast Mode' vs 'Priority Processing')
- hermes_cli/commands.py: Update /fast description to mention both
providers
- tests: 13 new tests covering Anthropic model detection, override
resolution, CLI availability, routing, adapter kwargs, and
third-party endpoint safety
After mid-loop compression (triggered by 413, context_overflow, or Anthropic
long-context tier errors), _compress_context() creates a new session in SQLite
and resets _last_flushed_db_idx=0. However, conversation_history was not cleared,
so _flush_messages_to_session_db() computed:
flush_from = max(len(conversation_history=200), _last_flushed_db_idx=0) = 200
messages[200:] → empty (compressed messages < 200)
This resulted in zero messages being written to the new session's SQLite store.
On resume, the user would see 'Session found but has no messages.'
The preflight compression path (line 7311) already had the fix:
conversation_history = None
This commit adds the same clearing to the three mid-loop compression sites:
- Anthropic long-context tier overflow
- HTTP 413 payload too large
- Generic context_overflow error
Reported by Aaryan (Nous community).
Raise the default httpx stream read timeout from 60s to 120s for all
providers. Additionally, auto-detect local LLM endpoints (Ollama,
llama.cpp, vLLM) and raise the read timeout to HERMES_API_TIMEOUT
(1800s) since local models can take minutes for prefill on large
contexts before producing the first token.
The stale stream timeout already had this local auto-detection pattern;
the httpx read timeout was missing it — causing a hard 60s wall that
users couldn't find (HERMES_STREAM_READ_TIMEOUT was undocumented).
Changes:
- Default HERMES_STREAM_READ_TIMEOUT: 60s -> 120s
- Auto-detect local endpoints -> raise to 1800s (user override respected)
- Document HERMES_STREAM_READ_TIMEOUT and HERMES_STREAM_STALE_TIMEOUT
- Add 10 parametrized tests
Reported-by: Pavan Srinivas (@pavanandums)
When the model mentions <think> as literal text in its response (e.g.
"(/think not producing <think> tags)"), the streaming display treated it
as a reasoning block opener and suppressed everything after it. The
response box would close with truncated content and no error — the API
response was complete but the display ate it.
Root cause: _stream_delta() matched <think> anywhere in the text stream
regardless of position. Real reasoning blocks always start at the
beginning of a line; mentions in prose appear mid-sentence.
Fix: track line position across streaming deltas with a
_stream_last_was_newline flag. Only enter reasoning suppression when
the tag appears at a block boundary (start of stream, after a newline,
or after only whitespace on the current line). Add a _flush_stream()
safety net that recovers buffered content if no closing tag is found
by end-of-stream.
Also fixes three related issues discovered during investigation:
- anthropic_adapter: _get_anthropic_max_output() now normalizes dots to
hyphens so 'claude-opus-4.6' matches the 'claude-opus-4-6' table key
(was returning 32K instead of 128K)
- run_agent: send explicit max_tokens for Claude models on Nous Portal,
same as OpenRouter — both proxy to Anthropic's API which requires it.
Without it the backend defaults to a low limit that truncates responses.
- run_agent: reset truncated_tool_call_retries after successful tool
execution so a single truncation doesn't poison the entire conversation.
Previously /fast only supported gpt-5.4 and forced a provider switch to
openai-codex. Now supports all 13 models from OpenAI's Priority Processing
pricing table (gpt-5.4, gpt-5.4-mini, gpt-5.2, gpt-5.1, gpt-5, gpt-5-mini,
gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, gpt-4o, gpt-4o-mini, o3, o4-mini).
Key changes:
- Replaced _FAST_MODE_BACKEND_CONFIG with _PRIORITY_PROCESSING_MODELS frozenset
- Removed provider-forcing logic — service_tier is now injected into whatever
API path the user is already on (Codex Responses, Chat Completions, or
OpenRouter passthrough)
- Added request_overrides support to chat_completions path in run_agent.py
- Updated messaging from 'Codex inference tier' to 'Priority Processing'
- Expanded test coverage for all supported models
Add /fast slash command to toggle OpenAI Codex service_tier between
normal and priority ('fast') inference. Only exposed for models
registered in _FAST_MODE_BACKEND_CONFIG (currently gpt-5.4).
- Registry-based backend config for extensibility
- Dynamic command visibility (hidden from help/autocomplete for
non-supported models) via command_filter on SlashCommandCompleter
- service_tier flows through request_overrides from route resolution
- Omit max_output_tokens for Codex backend (rejects it)
- Persists to config.yaml under agent.service_tier
Salvage cleanup: removed simple_term_menu/input() menu (banned),
bare /fast now shows status like /reasoning. Removed redundant
override resolution in _build_api_kwargs — single source of truth
via request_overrides from route.
Co-authored-by: Hermes Agent <hermes@nousresearch.com>
When OpenRouter returns 'No endpoints found that support tool use'
(HTTP 404), display a hint explaining that provider routing restrictions
may be filtering out tool-capable providers. Links the user directly
to the model's OpenRouter page to check which providers support tools.
The hint fires in the error display block that runs regardless of whether
fallback succeeds — so the user always understands WHY the model failed,
not just that it fell back.
Reported via Discord: GLM-5.1 on OpenRouter with US-based provider
restrictions eliminated all 4 tool-supporting endpoints (DeepInfra,
Z.AI, Friendli, Venice), leaving only 7 non-tool providers.
When a streaming response is cut mid-tool-call (connection drop, timeout),
the accumulated function.arguments is invalid JSON. The mock response
builder defaulted finish_reason to 'stop', so the agent loop treated it
as a valid completed turn and tried to execute tools with broken args.
Fix: validate tool call arguments with json.loads() during mock response
reconstruction. If any are invalid JSON, override finish_reason to
'length'. In the main loop's length handler, if tool calls are present,
refuse to execute and return partial=True with a clear error instead of
silently failing or wasting retries.
Also fixes _thinking_exhausted to not short-circuit when tool calls are
present — truncated tool calls are not thinking exhaustion.
Original cherry-picked from PR #6776 by AIandI0x1.
Closes#6638.
When the API returns "max_tokens too large given prompt" (input tokens
are within the context window, but input + requested output > window),
the old code incorrectly routed through the same handler as "prompt too
long" errors, calling get_next_probe_tier() and permanently halving
context_length. This made things worse: the window was fine, only the
requested output size needed trimming for that one call.
Two distinct error classes now handled separately:
Prompt too long — input itself exceeds context window.
Fix: compress history + halve context_length (existing behaviour,
unchanged).
Output cap too large — input OK, but input + max_tokens > window.
Fix: parse available_tokens from the error message, set a one-shot
_ephemeral_max_output_tokens override for the retry, and leave
context_length completely untouched.
Changes:
- agent/model_metadata.py: add parse_available_output_tokens_from_error()
that detects Anthropic's "available_tokens: N" error format and returns
the available output budget, or None for all other error types.
- run_agent.py: call the new parser first in the is_context_length_error
block; if it fires, set _ephemeral_max_output_tokens (with a 64-token
safety margin) and break to retry without touching context_length.
_build_api_kwargs consumes the ephemeral value exactly once then clears
it so subsequent calls use self.max_tokens normally.
- agent/anthropic_adapter.py: expand build_anthropic_kwargs docstring to
clearly document the max_tokens (output cap) vs context_length (total
window) distinction, which is a persistent source of confusion due to
the OpenAI-inherited "max_tokens" name.
- cli-config.yaml.example: add inline comments explaining both keys side
by side where users are most likely to look.
- website/docs/integrations/providers.md: add a callout box at the top
of "Context Length Detection" and clarify the troubleshooting entry.
- tests/test_ctx_halving_fix.py: 24 tests across four classes covering
the parser, build_anthropic_kwargs clamping, ephemeral one-shot
consumption, and the invariant that context_length is never mutated
on output-cap errors.
When _try_activate_fallback() swaps to a new provider (e.g.
kimi-coding), resolve_provider_client() correctly injects
provider-specific default_headers (like KimiCLI User-Agent) into the
returned OpenAI client. However, _client_kwargs was saved with only
api_key and base_url, dropping those headers.
Every subsequent API call rebuilds the client from _client_kwargs via
_create_request_openai_client(), producing a bare OpenAI client without
the required headers. Kimi Coding rejects this with 403; Copilot would
lose its auth headers similarly.
This patch reads _custom_headers from the fallback client (where the
OpenAI SDK stores the default_headers kwarg) and includes them in
_client_kwargs so any client rebuild preserves provider-specific headers.
Fixes#6075
Every turn now logs WHY the agent loop ended to agent.log with a
structured INFO line capturing: exit reason, model, api_calls/max,
budget usage, tool turn count, last message role, response length,
and session ID.
When the last message is a tool result and the turn was NOT
interrupted, emits WARNING level (visible in errors.log) — this is
the 'just stops' scenario users report where a tool call completes
but no continuation or final response follows.
10 tracked exit reasons: text_response, interrupted_by_user,
interrupted_during_api_call, budget_exhausted, max_iterations_reached,
all_retries_exhausted_no_response, fallback_prior_turn_content,
empty_response_exhausted, error_near_max_iterations, unknown.
Parse x-ratelimit-* headers from inference API responses (Nous Portal,
OpenRouter, OpenAI-compatible) and display them in the /usage command.
- New agent/rate_limit_tracker.py: parse 12 rate limit headers (RPM/RPH/
TPM/TPH limits, remaining, reset timers), format as progress bars (CLI)
or compact one-liner (gateway)
- Hook into streaming path in run_agent.py: stream.response.headers is
available on the OpenAI SDK Stream object before chunks are consumed
- CLI /usage: appends rate limit section with progress bars + warnings
when any bucket exceeds 80%
- Gateway /usage: appends compact rate limit summary
- 24 unit tests covering parsing, formatting, edge cases
Headers captured per response:
x-ratelimit-{limit,remaining,reset}-{requests,tokens}{,-1h}
Example CLI display:
Nous Rate Limits (captured just now):
Requests/min [░░░░░░░░░░░░░░░░░░░░] 0.1% 1/800 used (799 left, resets in 59s)
Tokens/hr [░░░░░░░░░░░░░░░░░░░░] 0.0% 49/336.0M (336.0M left, resets in 52m)
When a model returns no content, no structured reasoning, and no tool
calls (common with open models), the agent now silently retries up to
3 times before falling through to (empty).
Silent retry (no synthetic messages) keeps the conversation history
clean, preserves prompt caching, and respects the no-synthetic-user-
injection invariant. Most empty responses from open models are
transient (provider hiccups, rate limits, sampling flukes) so a
simple retry is sufficient.
This fills the last gap in the empty-response recovery chain:
1. _last_content_with_tools fallback (prior tool turn had content)
2. Thinking-only prefill continuation (#5931 — structured reasoning)
3. Empty response silent retry (NEW — truly empty, no reasoning)
4. (empty) terminal (last resort after all retries exhausted)
Inline <think> blocks are excluded — the model chose to reason, it
just produced no visible text. That differs from truly empty.
Tests:
- Updated test_truly_empty to expect 4 API calls (1 + 3 retries)
- Added test_truly_empty_response_succeeds_on_nudge
`_cleanup_task_resources` was unconditionally calling `cleanup_vm()` at
the end of every `run_conversation` (i.e. every user turn), tearing down
the docker/daytona/modal sandbox container regardless of its
`persistent_filesystem` setting. This contradicted the documented intent
of `terminal.lifetime_seconds` (idle reaper) and `container_persistent`,
and caused per-turn loss of `/workspace`, `~/.config`, agent CLI auth
state, and any other content living inside the sandbox.
The unconditional teardown was introduced in fbd3a2fd ("prevent leakage
of morph instances between tasks", 2025-11-04) to plug a Morph backend
leak, two days after `lifetime_seconds` shipped in faecbddd. It was
later refactored into `_cleanup_task_resources` in 70dd3a16 without
changing semantics. Code and docs have disagreed since.
Fix: introduce `terminal_tool.is_persistent_env(task_id)` and skip the
per-turn `cleanup_vm` when the active env is persistent. The idle reaper
(`_cleanup_inactive_envs`) still tears persistent envs down once
`terminal.lifetime_seconds` is exceeded. Non-persistent backends (Morph)
are unchanged — still torn down per turn, preserving the original
leak-prevention intent.
Combines the approaches from PR #6309 (duan78) and PR #5963 (KUSH42):
Tiered warnings (from #5963):
- Replaces boolean _context_pressure_warned with float _context_pressure_warned_at
- Fires at 85% (orange) and re-fires at 95% (red/critical)
- Adds 'compacting context...' status message before compression
Gateway dedup (from #6309):
- Class-level dict _context_pressure_last_warned survives across AIAgent
instances (gateway creates a new instance per message)
- 5-minute cooldown per session prevents warning spam
- Higher-tier warnings bypass the cooldown (85% → 95% always fires)
- Compression reset clears the dedup entry for the session
- Stale entries evicted (older than 2x cooldown) to prevent memory leak
Does NOT inject into messages — purely user-facing via _safe_print (CLI)
and status_callback (gateway). Zero prompt cache impact.
Fixes#6309. Fixes#5963.
Three targeted improvements to the compression system:
1. Replace hardcoded truncation limits with named class constants
(_CONTENT_MAX=6000, _CONTENT_HEAD=4000, _CONTENT_TAIL=1500,
_TOOL_ARGS_MAX=1500, _TOOL_ARGS_HEAD=1200). Previous limits
(3000/500) heavily truncated the summarizer's input — a 200-line
edit got cut to 3000 chars before the summarizer ever saw it.
2. Add '## Tools & Patterns' section to both compression prompt
templates (first-pass and iterative). Preserves working tool
invocations, preferred flags, and tool-specific discoveries
across compaction boundaries.
3. Warn users on 2nd+ compression: 'Session compressed N times —
accuracy may degrade. Consider /new to start fresh.'
Ref #499
Local inference providers (Ollama, oMLX, llama-cpp) can take 300+ seconds
for prefill on large contexts. The 180s stale stream detector was killing
these connections while the provider was still processing.
Uses the existing is_local_endpoint() (proper URL parsing with RFC-1918,
localhost, WSL detection) instead of ad-hoc substring matching. The stale
timeout is only disabled when the user hasn't explicitly set
HERMES_STREAM_STALE_TIMEOUT — explicit user config is always honored.
Fixes#5889
The _call_llm() and direct OpenAI fallback paths in flush_memories() both
hardcoded timeout=30.0, ignoring the user-configurable value at
auxiliary.flush_memories.timeout in config.yaml.
Remove the explicit timeout from the auxiliary _call_llm() call so that
_get_task_timeout('flush_memories') reads from config. For the direct
OpenAI fallback, import and use _get_task_timeout() instead of the
hardcoded value.
Add two regression tests verifying both code paths respect the config.
Fixes#6154
Two linked fixes for MiniMax Anthropic-compatible fallback:
1. Normalize httpx.URL to str before calling .rstrip() in auth/provider
detection helpers. Some client objects expose base_url as httpx.URL,
not str — crashed with AttributeError in _requires_bearer_auth() and
_is_third_party_anthropic_endpoint(). Also fixes _try_activate_fallback()
to use the already-stringified fb_base_url instead of raw httpx.URL.
2. Strip Anthropic-proprietary thinking block signatures when targeting
third-party Anthropic-compatible endpoints (MiniMax, Azure AI Foundry,
self-hosted proxies). These endpoints cannot validate Anthropic's
signatures and reject them with HTTP 400 'Invalid signature in
thinking block'. Now threads base_url through convert_messages_to_anthropic()
→ build_anthropic_kwargs() so signature management is endpoint-aware.
Based on PR #4945 by kshitijk4poor (rstrip fix).
Fixes#4944.
Based on #6079 by @tunamitom with critical fixes and comprehensive tests.
Changes from #6079:
- Fix: sanitization overwrite bug — Qwen message prep now runs AFTER codex
field sanitization, not before (was silently discarding Qwen transforms)
- Fix: missing try/except AuthError in runtime_provider.py — stale Qwen
credentials now fall through to next provider on auto-detect
- Fix: 'qwen' alias conflict — bare 'qwen' stays mapped to 'alibaba'
(DashScope); use 'qwen-portal' or 'qwen-cli' for the OAuth provider
- Fix: hardcoded ['coder-model'] replaced with live API fetch + curated
fallback list (qwen3-coder-plus, qwen3-coder)
- Fix: extract _is_qwen_portal() helper + _qwen_portal_headers() to replace
5 inline 'portal.qwen.ai' string checks and share headers between init
and credential swap
- Fix: add Qwen branch to _apply_client_headers_for_base_url for mid-session
credential swaps
- Fix: remove suspicious TypeError catch blocks around _prompt_provider_choice
- Fix: handle bare string items in content lists (were silently dropped)
- Fix: remove redundant dict() copies after deepcopy in message prep
- Revert: unrelated ai-gateway test mock removal and model_switch.py comment deletion
New tests (30 test functions):
- _qwen_cli_auth_path, _read_qwen_cli_tokens (success + 3 error paths)
- _save_qwen_cli_tokens (roundtrip, parent creation, permissions)
- _qwen_access_token_is_expiring (5 edge cases: fresh, expired, within skew,
None, non-numeric)
- _refresh_qwen_cli_tokens (success, preserve old refresh, 4 error paths,
default expires_in, disk persistence)
- resolve_qwen_runtime_credentials (fresh, auto-refresh, force-refresh,
missing token, env override)
- get_qwen_auth_status (logged in, not logged in)
- Runtime provider resolution (direct, pool entry, alias)
- _build_api_kwargs (metadata, vl_high_resolution_images, message formatting,
max_tokens suppression)
Anthropic signs thinking blocks against the full turn content. Any
upstream mutation (context compression, session truncation, orphan
stripping, message merging) invalidates the signature, causing HTTP 400
'Invalid signature in thinking block' — especially in long-lived
gateway sessions.
Strategy (following clawdbot/OpenClaw pattern):
1. Strip thinking/redacted_thinking from all assistant messages EXCEPT
the last one — preserves reasoning continuity on the current
tool-use chain while avoiding stale signature errors on older turns.
2. Downgrade unsigned thinking blocks to plain text — Anthropic can't
validate them, but the reasoning content is preserved.
3. Strip cache_control from thinking/redacted_thinking blocks to
prevent cache markers from interfering with signature validation.
4. Drop thinking blocks from the second message when merging
consecutive assistant messages (role alternation enforcement).
5. Error recovery: on HTTP 400 mentioning 'signature' and 'thinking',
strip all reasoning_details from the conversation and retry once.
This is the safety net for edge cases the proactive stripping
misses.
Addresses the issue reported in PR #6086 by @mingginwan while
preserving reasoning continuity (their PR stripped ALL thinking
blocks unconditionally).
Files changed:
- agent/anthropic_adapter.py: thinking block management in
convert_messages_to_anthropic (strip old turns, downgrade unsigned,
strip cache_control, merge-time strip)
- run_agent.py: one-shot signature error recovery in retry loop
- tests/test_anthropic_adapter.py: 10 new tests covering all cases
Salvaged fixes from community PRs:
- fix(model_switch): _read_auth_store → _load_auth_store + fix auth store
key lookup (was checking top-level dict instead of store['providers']).
OAuth providers now correctly detected in /model picker.
Cherry-picked from PR #5911 by Xule Lin (linxule).
- fix(ollama): pass num_ctx to override 2048 default context window.
Ollama defaults to 2048 context regardless of model capabilities. Now
auto-detects from /api/show metadata and injects num_ctx into every
request. Config override via model.ollama_num_ctx. Fixes#2708.
Cherry-picked from PR #5929 by kshitij (kshitijk4poor).
- fix(aux): normalize provider aliases for vision/auxiliary routing.
Adds _normalize_aux_provider() with 17 aliases (google→gemini,
claude→anthropic, glm→zai, etc). Fixes vision routing failure when
provider is set to 'google' instead of 'gemini'.
Cherry-picked from PR #5793 by e11i (Elizabeth1979).
- fix(aux): rewrite MiniMax /anthropic base URLs to /v1 for OpenAI SDK.
MiniMax's inference_base_url ends in /anthropic (Anthropic Messages API),
but auxiliary client uses OpenAI SDK which appends /chat/completions →
404 at /anthropic/chat/completions. Generic _to_openai_base_url() helper
rewrites terminal /anthropic to /v1 for OpenAI-compatible endpoint.
Inspired by PR #5786 by Lempkey.
Added debug logging to silent exception blocks across all fixes.
Co-authored-by: Hermes Agent <hermes@nousresearch.com>
The response validation stage unconditionally marked Codex Responses API
replies as invalid when response.output was empty, triggering unnecessary
retries and fallback chains. However, _normalize_codex_response can
recover from this state by synthesizing output from response.output_text.
Now the validation stage checks for output_text before marking the
response invalid, matching the normalization logic. Also fixes
logging.warning → logger.warning for consistency with the rest of the
file.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When the model produces structured reasoning (via API fields like .reasoning,
.reasoning_content, .reasoning_details) but no visible text content, append
the assistant message as prefill and continue the loop. The model sees its own
reasoning context on the next turn and produces the text portion.
Inspired by clawdbot's 'incomplete-text' recovery pattern. Up to 2 prefill
attempts before falling through to the existing '(empty)' terminal.
Key design decisions:
- Only triggers for structured reasoning (API fields), NOT inline <think> tags
- Prefill messages are popped on success to maintain strict role alternation
- _thinking_prefill marker stripped from all API message building paths
- Works across all providers: OpenAI (continuation), Anthropic (native prefill)
Verified with E2E tests: simulated thinking-only → real OpenRouter continuation
produces correct content. Also confirmed Qwen models consistently produce
structured-reasoning-only responses under token pressure.
Memory plugins (Mem0, Honcho) used static identifiers ('hermes-user',
config peerName) meaning all gateway users shared the same memory bucket.
Changes:
- AIAgent.__init__: add user_id parameter, store as self._user_id
- run_agent.py: include user_id in _init_kwargs passed to memory providers
- gateway/run.py: pass source.user_id to AIAgent in primary + background paths
- Mem0 plugin: prefer kwargs user_id over config default
- Honcho plugin: override cfg.peer_name with gateway user_id when present
CLI sessions (user_id=None) preserve existing defaults. Only gateway
sessions with a real platform user_id get per-user memory scoping.
Reported by plev333.
Comprehensive cleanup across 80 files based on automated (ruff, pyflakes, vulture)
and manual analysis of the entire codebase.
Changes by category:
Unused imports removed (~95 across 55 files):
- Removed genuinely unused imports from all major subsystems
- agent/, hermes_cli/, tools/, gateway/, plugins/, cron/
- Includes imports in try/except blocks that were truly unused
(vs availability checks which were left alone)
Unused variables removed (~25):
- Removed dead variables: connected, inner, channels, last_exc,
source, new_server_names, verify, pconfig, default_terminal,
result, pending_handled, temperature, loop
- Dropped unused argparse subparser assignments in hermes_cli/main.py
(12 instances of add_parser() where result was never used)
Dead code removed:
- run_agent.py: Removed dead ternary (None if False else None) and
surrounding unreachable branch in identity fallback
- run_agent.py: Removed write-only attribute _last_reported_tool
- hermes_cli/providers.py: Removed dead @property decorator on
module-level function (decorator has no effect outside a class)
- gateway/run.py: Removed unused MCP config load before reconnect
- gateway/platforms/slack.py: Removed dead SessionSource construction
Undefined name bugs fixed (would cause NameError at runtime):
- batch_runner.py: Added missing logger = logging.getLogger(__name__)
- tools/environments/daytona.py: Added missing Dict and Path imports
Unnecessary global statements removed (14):
- tools/terminal_tool.py: 5 functions declared global for dicts
they only mutated via .pop()/[key]=value (no rebinding)
- tools/browser_tool.py: cleanup thread loop only reads flag
- tools/rl_training_tool.py: 4 functions only do dict mutations
- tools/mcp_oauth.py: only reads the global
- hermes_time.py: only reads cached values
Inefficient patterns fixed:
- startswith/endswith tuple form: 15 instances of
x.startswith('a') or x.startswith('b') consolidated to
x.startswith(('a', 'b'))
- len(x)==0 / len(x)>0: 13 instances replaced with pythonic
truthiness checks (not x / bool(x))
- in dict.keys(): 5 instances simplified to in dict
- Redefined unused name: removed duplicate _strip_mdv2 import in
send_message_tool.py
Other fixes:
- hermes_cli/doctor.py: Replaced undefined logger.debug() with pass
- hermes_cli/config.py: Consolidated chained .endswith() calls
Test results: 3934 passed, 17 failed (all pre-existing on main),
19 skipped. Zero regressions.
Two gaps in the codex empty-output handling:
1. _run_codex_create_stream_fallback() skipped all non-terminal events,
so when the fallback path was used and the terminal response had
empty output, there was no recovery. Now collects output_item.done
and text deltas during the fallback stream, backfills on empty output.
2. _normalize_codex_response() hard-crashed with RuntimeError when
output was empty, even when the response had output_text set. The
function already had fallback logic at line 3562 to use output_text,
but the guard at line 3446 killed it first. Now checks output_text
before raising and synthesizes a minimal output item.
Salvages the core fix from PR #5673 (egerev) onto current main.
The chatgpt.com/backend-api/codex endpoint streams valid output items
via response.output_item.done events, but the OpenAI SDK's
get_final_response() returns an empty output list. This caused every
Codex response to be rejected as invalid.
Fix: collect output_item.done events during streaming and backfill
response.output when get_final_response() returns empty. Falls back
to synthesizing from text deltas when no done events were received.
Also moves the synthesis logic from the validation loop (too late, from
#5681) into _run_codex_stream() (before the response leaves the
streaming function), and simplifies the validation to just log
diagnostics since recovery now happens upstream.
Co-authored-by: Egor <egerev@users.noreply.github.com>
Three bugs causing OpenAI Codex sessions to fail silently:
1. Credential pool vs legacy store disconnect: hermes auth and hermes
model store device_code tokens in the credential pool, but
get_codex_auth_status(), resolve_codex_runtime_credentials(), and
_model_flow_openai_codex() only read from the legacy provider state.
Fresh pool tokens were invisible to the auth status checks and model
selection flow.
2. _import_codex_cli_tokens() imported expired tokens from ~/.codex/
without checking JWT expiry. Combined with _login_openai_codex()
saying 'Login successful!' for expired credentials, users got stuck
in a loop of dead tokens being recycled.
3. _login_openai_codex() accepted expired tokens from
resolve_codex_runtime_credentials() without validating expiry before
telling the user login succeeded.
Fixes:
- get_codex_auth_status() now checks credential pool first, falls back
to legacy provider state
- _model_flow_openai_codex() uses pool-aware auth status for token
retrieval when fetching model lists
- _import_codex_cli_tokens() validates JWT exp claim, rejects expired
- _login_openai_codex() verifies resolved token isn't expiring before
accepting existing credentials
- _run_codex_stream() logs response.incomplete/failed terminal events
with status and incomplete_details for diagnostics
- Codex empty output recovery: captures streamed text during streaming
and synthesizes a response when get_final_response() returns empty
output (handles chatgpt.com backend-api edge cases)
Two fixes:
1. Replace all stale 'hermes login' references with 'hermes auth' across
auth.py, auxiliary_client.py, delegate_tool.py, config.py, run_agent.py,
and documentation. The 'hermes login' command was deprecated; 'hermes auth'
now handles OAuth credential management.
2. Fix credential removal not persisting for singleton-sourced credentials
(device_code for openai-codex/nous, hermes_pkce for anthropic).
auth_remove_command already cleared env vars for env-sourced credentials,
but singleton credentials stored in the auth store were re-seeded by
_seed_from_singletons() on the next load_pool() call. Now clears the
underlying auth store entry when removing singleton-sourced credentials.
When the Codex CLI (or VS Code extension) consumes a refresh token before
Hermes can use it, Hermes previously surfaced a generic 401 error with no
actionable guidance.
- In `refresh_codex_oauth_pure`: detect `refresh_token_reused` from the
OAuth endpoint and raise an AuthError explaining the cause and the exact
steps to recover (run `codex` to refresh, then `hermes login`).
- In `run_agent.py`: when provider is `openai-codex` and HTTP 401 is
received, show Codex-specific recovery steps instead of the generic
"check your API key" message.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When using xAI's API directly (base_url contains x.ai), send the
x-grok-conv-id header set to the Hermes session_id. This routes
consecutive requests to the same server, maximizing automatic
prompt cache hits.
Ref: https://docs.x.ai/developers/advanced-api-usage/prompt-caching
OpenCode Go model names with dots (minimax-m2.7, glm-4.5, kimi-k2.5)
were being mangled to hyphens (minimax-m2-7), causing HTTP 401 errors.
Two code paths were affected:
1. model_normalize.py: opencode-go was incorrectly in DOT_TO_HYPHEN_PROVIDERS
2. run_agent.py: _anthropic_preserve_dots() did not check for opencode-go
Fix:
- Remove opencode-go from _DOT_TO_HYPHEN_PROVIDERS (dots are correct for Go)
- Add opencode-go to _anthropic_preserve_dots() provider check
- Add opencode.ai/zen/go to base_url fallback check
- Add regression tests in tests/test_model_normalize.py
Co-authored-by: jacob3712 <jacob3712@users.noreply.github.com>
- Rename per-LLM-call hooks from pre_llm_request/post_llm_request for clarity vs pre_llm_call
- Emit summary kwargs only (counts, usage dict from normalize_usage); keep env_var_enabled for HERMES_DUMP_REQUESTS
- Add is_truthy_value/env_var_enabled to utils; wire hermes_cli.plugins._env_enabled through it
- Update Langfuse local setup doc; add scripts/langfuse_smoketest.py and optional ~/.hermes plugin tests
Made-with: Cursor
Consolidated salvage from PRs #5301 (qaqcvc), #5339 (lance0),
#5058 and #5098 (maymuneth).
Mem0 API v2 compatibility (#5301):
- All reads use filters={user_id: ...} instead of bare user_id= kwarg
- All writes use filters with user_id + agent_id for attribution
- Response unwrapping for v2 dict format {results: [...]}
- Split _read_filters() vs _write_filters() — reads are user-scoped
only for cross-session recall, writes include agent_id
- Preserved 'hermes-user' default (no breaking change for existing users)
- Omitted run_id scoping from #5301 — cross-session memory is Mem0's
core value, session-scoping reads would defeat that purpose
Memory prefetch context fencing (#5339):
- Wraps prefetched memory in <memory-context> fenced blocks with system
note marking content as recalled context, NOT user input
- Sanitizes provider output to strip fence-escape sequences, preventing
injection where memory content breaks out of the fence
- API-call-time only — never persisted to session history
Secret redaction (#5058, #5098):
- Added prefix patterns for Groq (gsk_), Matrix (syt_), RetainDB
(retaindb_), Hindsight (hsk-), Mem0 (mem0_), ByteRover (brv_)
Adds OPENAI_MODEL_EXECUTION_GUIDANCE — XML-tagged behavioral guidance
injected for GPT and Codex models alongside the existing tool-use
enforcement. Targets four specific failure modes:
- <tool_persistence>: retry on empty/partial results instead of giving up
- <prerequisite_checks>: do discovery/lookup before jumping to final action
- <verification>: check correctness/grounding/formatting before finalizing
- <missing_context>: use lookup tools instead of hallucinating
Follows the same injection pattern as GOOGLE_MODEL_OPERATIONAL_GUIDANCE
for Gemini/Gemma models. Inspired by OpenClaw PR #38953 and OpenAI's
GPT-5.4 prompting guide patterns.
Agent activity tracking:
- Add _last_activity_ts, _last_activity_desc, _current_tool to AIAgent
- Touch activity on: API call start/complete, tool start/complete,
first stream chunk, streaming request start
- Public get_activity_summary() method for external consumers
Gateway timeout diagnostics:
- Timeout message now includes what the agent was doing when killed:
actively working vs stuck on a tool vs waiting on API response
- Includes iteration count, last activity description, seconds since
last activity — users can distinguish legitimate long tasks from
genuine hangs
- 'Still working' notifications now show iteration count and current
tool instead of just elapsed time
- Stale lock eviction logs include agent activity state for debugging
Stream stale timeout:
- _emit_status when stale stream is detected (was log-only) — gateway
users now see 'No response from provider for Ns' with model and
context size
- Improved logger.warning with model name and estimated context size
Error path notifications (gateway-visible via _emit_status):
- Context compression attempts now use _emit_status (was _vprint only)
- Non-retryable client errors emit summary before aborting
- Max retry exhaustion emits error summary (was _vprint only)
- Rate limit exhaustion emits specific rate-limit message
These were all CLI-visible but silent to gateway users, which is why
people on Telegram/Discord saw generic 'request failed' messages
without explanation.
Subagent sessions spawned by delegate_task were created with
parent_session_id=NULL and source=cli, making them indistinguishable
from user sessions in hermes sessions list and /resume.
Changes:
- delegate_tool.py: pass parent_agent.session_id to child agent
- run_agent.py: accept parent_session_id param, pass to create_session
- hermes_state.py list_sessions_rich: filter parent_session_id IS NULL
by default (opt-in include_children=True for callers that need them)
- hermes_state.py delete_session: delete child sessions first (FK)
- hermes_state.py prune_sessions: delete children before parents (FK)
session_search already handles parent_session_id correctly — child
sessions are filtered from recent list and resolved to parent root
in full-text search results.
Fixes#5122
As the agent navigates into subdirectories via tool calls (read_file,
terminal, search_files, etc.), automatically discover and load project
context files (AGENTS.md, CLAUDE.md, .cursorrules) from those directories.
Previously, context files were only loaded from the CWD at session start.
If the agent moved into backend/, frontend/, or any subdirectory with its
own AGENTS.md, those instructions were never seen.
Now, SubdirectoryHintTracker watches tool call arguments for file paths
and shell commands, resolves directories, and loads hint files on first
access. Discovered hints are appended to the tool result so the model
gets relevant context at the moment it starts working in a new area —
without modifying the system prompt (preserving prompt caching).
Features:
- Extracts paths from tool args (path, workdir) and shell commands
- Loads AGENTS.md, CLAUDE.md, .cursorrules (first match per directory)
- Deduplicates — each directory loaded at most once per session
- Ignores paths outside the working directory
- Truncates large hint files at 8K chars
- Works on both sequential and concurrent tool execution paths
Inspired by Block/goose SubdirectoryHintTracker.
Add POST /v1/runs to start async agent runs and GET /v1/runs/{run_id}/events
for SSE streaming of typed lifecycle events (tool.started, tool.completed,
message.delta, reasoning.available, run.completed, run.failed).
Changes the internal tool_progress_callback signature from positional
(tool_name, preview, args) to event-type-first
(event_type, tool_name, preview, args, **kwargs). Existing consumers
filter on event_type and remain backward-compatible.
Adds concurrency limit (_MAX_CONCURRENT_RUNS=10) and orphaned run sweep.
Fixes logic inversion in cli.py _on_tool_progress where the original PR
would have displayed internal tools instead of non-internal ones.
Co-authored-by: Mibayy <mibayy@users.noreply.github.com>
Route AIAgent print output to stderr via _print_fn for ACP stdio sessions.
Gate quiet-mode spinner startup on _should_start_quiet_spinner() so JSON-RPC
on stdout isn't corrupted. Child agents inherit the redirect.
Co-authored-by: Git-on-my-level <Git-on-my-level@users.noreply.github.com>
* feat: coerce tool call arguments to match JSON Schema types
LLMs frequently return numbers as strings ("42" instead of 42) and
booleans as strings ("true" instead of true). This causes silent
failures with MCP tools and any tool with strictly-typed parameters.
Added coerce_tool_args() in model_tools.py that runs before every tool
dispatch. For each argument, it checks the tool registry schema and
attempts safe coercion:
- "42" → 42 when schema says "type": "integer"
- "3.14" → 3.14 when schema says "type": "number"
- "true"/"false" → True/False when schema says "type": "boolean"
- Union types tried in order
- Original values preserved when coercion fails or is not applicable
Inspired by Block/goose tool argument coercion system.
* fix: accept reasoning-only responses without retries — set content to "(empty)"
Previously, when a model returned reasoning/thinking but no visible
content, we entered a 120-line retry/classify/compress/salvage cascade
that wasted 3+ API calls trying to "fix" the response. The model was
done thinking — retrying with the same input just burned money.
Now reasoning-only responses are accepted immediately:
- Reasoning stays in the `reasoning` field (semantically correct)
- Content set to "(empty)" — valid non-empty string every provider accepts
- No retries, no compression triggers, no salvage logic
- Session history contains "(empty)" not "" — prevents #2128 session
poisoning where empty assistant content caused prefill rejections
Removes ~120 lines, adds ~15. Saves 2-3 API calls per reasoning-only
response. Fixes#2128.
- Add OLLAMA_API_KEY to credential resolution chain for ollama.com endpoints
- Update requested_provider/_explicit_api_key/_explicit_base_url after /model
switch so _ensure_runtime_credentials() doesn't revert the switch
- Pass base_url/api_key from fallback config to resolve_provider_client()
- Add DirectAlias system: user-configurable model_aliases in config.yaml
checked before catalog resolution, with reverse lookup by model ID
- Add /model tab completion showing aliases with provider metadata
Co-authored-by: LucidPaths <LucidPaths@users.noreply.github.com>
Previously, tool results exceeding 100K characters were silently chopped
with only a '[Truncated]' notice — the rest of the content was lost
permanently. The model had no way to access the truncated portion.
Now, oversized results are written to HERMES_HOME/cache/tool_responses/
and the model receives:
- A 1,500-char head preview for immediate context
- The file path so it can use read_file/search_files on the full output
This preserves the context window protection (inline content stays small)
while making the full data recoverable. Falls back to the old destructive
truncation if the file write fails.
Inspired by Block/goose's large response handler pattern.
Persist structured exhaustion metadata from provider errors, use explicit reset timestamps when available, and expose label-based credential targeting in the auth CLI. This keeps long-lived Codex cooldowns from being misreported as one-hour waits and avoids forcing operators to manage entries by list position alone.
Constraint: Existing credential pool JSON needs to remain backward compatible with stored entries that only record status code and timestamp
Constraint: Runtime recovery must keep the existing retry-then-rotate semantics for 429s while enriching pool state with provider metadata
Rejected: Add a separate credential scheduler subsystem | too large for the Hermes pool architecture and unnecessary for this fix
Rejected: Only change CLI formatting | would leave runtime rotation blind to resets_at and preserve the serial-failure behavior
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Preserve structured rate-limit metadata when new providers expose reset hints; do not collapse back to status-code-only exhaustion tracking
Tested: Focused pytest slice for auth commands, credential pool recovery, and routing (272 passed); py_compile on changed Python files; hermes -w auth list/remove smoke test with temporary HERMES_HOME
Not-tested: Full repository pytest suite, broader gateway/integration flows outside the touched auth and pool paths
Updates _sanitize_tool_calls_for_strict_api docstring to explicitly
mention Fireworks alongside Mistral as strict APIs requiring sanitization.
Also documents the specific fields that are stripped (call_id, response_item_id).
Replaces hardcoded Mistral check with the new _should_sanitize_tool_calls()
method. Updates comment to mention Fireworks alongside Mistral as strict
APIs requiring tool_call field sanitization.
Replaces hardcoded Mistral check with the new _should_sanitize_tool_calls()
method. Ensures summary generation works correctly with Fireworks and
other strict APIs that reject unknown tool_call fields.
Replaces hardcoded Mistral check with the new _should_sanitize_tool_calls()
method. This ensures tool_calls are sanitized for all strict APIs, not
just Mistral. Prevents 400 errors from Fireworks and other providers.
Adds a centralized method to determine when tool_calls need sanitization
for strict APIs. Returns True for all APIs except codex_responses mode.
This prevents 400 errors from providers like Fireworks that reject unknown
fields (call_id, response_item_id) in tool_calls.
Plugin context from pre_llm_call hooks was injected into the system
prompt, breaking the prompt cache prefix every turn when content
changed (typical for memory plugins). Now all plugin context goes
into the current turn's user message — the system prompt stays
identical across turns, preserving cached tokens.
The system prompt is reserved for Hermes internals. Plugins
contribute context alongside the user's input.
Also adds comprehensive documentation for all 6 plugin hooks:
pre_tool_call, post_tool_call, pre_llm_call, post_llm_call,
on_session_start, on_session_end — each with full callback
signatures, parameter tables, firing conditions, and examples.
Supersedes #5138 which identified the same cache-busting bug
and proposed an uncached system suffix approach. This fix goes
further by removing system prompt injection entirely.
Co-identified-by: OutThisLife (PR #5138)
When a skill is active, user_message contains the full SKILL.md content
injected by the skill system. This bloated string was being passed to
memory provider sync_all(), queue_prefetch_all(), and prefetch_all(),
causing providers with query size limits (e.g. Honcho's 10K char limit)
to fail.
Both call sites now use original_user_message (the clean user input,
already defined at line 6516) instead of the skill-inflated user_message:
- Pre-turn prefetch (line ~6695): prefetch_all() query
- Post-turn sync (line ~8672): sync_all() + queue_prefetch_all()
Fixes#4889
* fix(gateway): add message deduplication to Discord and Slack adapters (#4777)
Discord RESUME replays events after reconnects (~7/day observed),
and Slack Socket Mode can redeliver events if the ack was lost.
Neither adapter tracked which messages were already processed,
causing duplicate bot responses.
Add _seen_messages dedup cache (message ID → timestamp) with 5-min
TTL and 2000-entry cap to both adapters, matching the pattern already
used by Mattermost, Matrix, WeCom, Feishu, DingTalk, and Email.
The check goes at the very top of the message handler, before any
other logic, so replayed events are silently dropped.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: prevent duplicate messages on partial stream delivery
When streaming fails after tokens are already delivered to the platform,
_interruptible_streaming_api_call re-raised the error into the outer
retry loop, which would make a new API call — creating a duplicate
message.
Now checks deltas_were_sent before re-raising: if partial content was
already streamed, returns a stub response instead. The outer loop treats
the turn as complete (no retry, no fallback, no duplicate).
Inspired by PR #4871 (@trevorgordon981) which identified the bug.
This implementation avoids monkey-patching exception objects and keeps
the fix within the streaming call boundary.
---------
Co-authored-by: Mibayy <mibayy@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add a provider-agnostic role allowlist guard to _sanitize_api_messages()
that drops messages with roles not accepted by the chat-completions API
(e.g. session_meta). This prevents CLI resume/session restore from
leaking transcript-only metadata into the outgoing messages payload.
Two layers of defense:
1. API-boundary guard: _sanitize_api_messages() now filters messages by
role allowlist (system/user/assistant/tool/function/developer) before
the existing orphaned tool-call repair logic. This protects all
current and future call paths.
2. CLI restore defense-in-depth: Both session restore paths in cli.py
now strip session_meta entries before loading history into
conversation_history, matching the existing gateway behavior.
Closes#4715
Memory provider tools (hindsight_retain, honcho_search, etc.) were
advertised to the model via tool schemas but failed with 'Unknown tool'
at execution time. The concurrent path (_invoke_tool) correctly checks
self._memory_manager.has_tool() before falling through to the registry,
but the sequential path (_execute_tool_calls_sequential) was never
updated with this check. Since sequential is the default for single
tool calls, memory provider tools always hit the registry dispatcher
which returns 'Unknown tool' because they're not registered there.
Add the memory_manager dispatch check between the delegate_task handler
and the quiet_mode fallthrough in the sequential path, with proper
spinner/display handling to match the existing pattern.
Reported by KiBenderOP — all memory providers affected (Honcho,
Hindsight, Holographic, etc.).
Three fixes for long-running gateway sessions that enter a death spiral
when API disconnects prevent token data collection, which prevents
compression, which causes more disconnects:
Layer 1 — Stale token counter fallback (run_agent.py in-loop):
When last_prompt_tokens is 0 (stale after API disconnect or provider
returned no usage data), fall back to estimate_messages_tokens_rough()
instead of passing 0 to should_compress(), which would never fire.
Layer 2 — Server disconnect heuristic (run_agent.py error handler):
When ReadError/RemoteProtocolError hits a large session (>60% context
or >200 messages), treat it as a context-length error and trigger
compression rather than burning through retries that all fail the
same way.
Layer 3 — Hard message count limit (gateway/run.py hygiene):
Force compression when a session exceeds 400 messages, regardless of
token estimates. This catches runaway growth even when all token-based
checks fail due to missing API data.
Based on the analysis from PR #2157 by ygd58 — the gateway threshold
direction fix (1.4x multiplier) was already resolved on main.
Anthropic returns HTTP 429 'Extra usage is required for long context
requests' when a Claude Max subscription doesn't include the 1M context
tier. This is NOT a transient rate limit — retrying won't help.
Only applies to Sonnet models (Opus 1M is general access). Detects
this specific error before the generic rate-limit handler and:
1. Reduces context_length from 1M to 200k (the standard tier)
2. Triggers context compression to fit
3. Retries with the reduced context
The reduction is session-scoped (not persisted) so it auto-recovers
if the user later enables extra usage on their subscription.
Fixes: Sonnet 4.6 instant rate limits on Claude Max without extra usage
Anthropic returns HTTP 429 'Extra usage is required for long context
requests' when a Claude Max subscription doesn't include the 1M context
tier. This is NOT a transient rate limit — retrying won't help.
Detect this specific error before the generic rate-limit handler and:
1. Reduce context_length from 1M to 200k (the standard tier)
2. Trigger context compression to fit
3. Retry with the reduced context
The reduction is session-scoped (not persisted) so it auto-recovers
if the user later enables extra usage on their subscription.
Fixes: Sonnet 4.6 instant rate limits on Claude Max without extra usage
The step_callback previously only forwarded tool names as strings,
so build_tool_complete received result=None and ACP tool_call_update
events had empty content/rawOutput. Now prev_tools carries dicts with
both name and result by pairing each tool_call with its matching
tool-role message via tool_call_id.
* feat(memory): add pluggable memory provider interface with profile isolation
Introduces a pluggable MemoryProvider ABC so external memory backends can
integrate with Hermes without modifying core files. Each backend becomes a
plugin implementing a standard interface, orchestrated by MemoryManager.
Key architecture:
- agent/memory_provider.py — ABC with core + optional lifecycle hooks
- agent/memory_manager.py — single integration point in the agent loop
- agent/builtin_memory_provider.py — wraps existing MEMORY.md/USER.md
Profile isolation fixes applied to all 6 shipped plugins:
- Cognitive Memory: use get_hermes_home() instead of raw env var
- Hindsight Memory: check $HERMES_HOME/hindsight/config.json first,
fall back to legacy ~/.hindsight/ for backward compat
- Hermes Memory Store: replace hardcoded ~/.hermes paths with
get_hermes_home() for config loading and DB path defaults
- Mem0 Memory: use get_hermes_home() instead of raw env var
- RetainDB Memory: auto-derive profile-scoped project name from
hermes_home path (hermes-<profile>), explicit env var overrides
- OpenViking Memory: read-only, no local state, isolation via .env
MemoryManager.initialize_all() now injects hermes_home into kwargs so
every provider can resolve profile-scoped storage without importing
get_hermes_home() themselves.
Plugin system: adds register_memory_provider() to PluginContext and
get_plugin_memory_providers() accessor.
Based on PR #3825. 46 tests (37 unit + 5 E2E + 4 plugin registration).
* refactor(memory): drop cognitive plugin, rewrite OpenViking as full provider
Remove cognitive-memory plugin (#727) — core mechanics are broken:
decay runs 24x too fast (hourly not daily), prefetch uses row ID as
timestamp, search limited by importance not similarity.
Rewrite openviking-memory plugin from a read-only search wrapper into
a full bidirectional memory provider using the complete OpenViking
session lifecycle API:
- sync_turn: records user/assistant messages to OpenViking session
(threaded, non-blocking)
- on_session_end: commits session to trigger automatic memory extraction
into 6 categories (profile, preferences, entities, events, cases,
patterns)
- prefetch: background semantic search via find() endpoint
- on_memory_write: mirrors built-in memory writes to the session
- is_available: checks env var only, no network calls (ABC compliance)
Tools expanded from 3 to 5:
- viking_search: semantic search with mode/scope/limit
- viking_read: tiered content (abstract ~100tok / overview ~2k / full)
- viking_browse: filesystem-style navigation (list/tree/stat)
- viking_remember: explicit memory storage via session
- viking_add_resource: ingest URLs/docs into knowledge base
Uses direct HTTP via httpx (no openviking SDK dependency needed).
Response truncation on viking_read to prevent context flooding.
* fix(memory): harden Mem0 plugin — thread safety, non-blocking sync, circuit breaker
- Remove redundant mem0_context tool (identical to mem0_search with
rerank=true, top_k=5 — wastes a tool slot and confuses the model)
- Thread sync_turn so it's non-blocking — Mem0's server-side LLM
extraction can take 5-10s, was stalling the agent after every turn
- Add threading.Lock around _get_client() for thread-safe lazy init
(prefetch and sync threads could race on first client creation)
- Add circuit breaker: after 5 consecutive API failures, pause calls
for 120s instead of hammering a down server every turn. Auto-resets
after cooldown. Logs a warning when tripped.
- Track success/failure in prefetch, sync_turn, and all tool calls
- Wait for previous sync to finish before starting a new one (prevents
unbounded thread accumulation on rapid turns)
- Clean up shutdown to join both prefetch and sync threads
* fix(memory): enforce single external memory provider limit
MemoryManager now rejects a second non-builtin provider with a warning.
Built-in memory (MEMORY.md/USER.md) is always accepted. Only ONE
external plugin provider is allowed at a time. This prevents tool
schema bloat (some providers add 3-5 tools each) and conflicting
memory backends.
The warning message directs users to configure memory.provider in
config.yaml to select which provider to activate.
Updated all 47 tests to use builtin + one external pattern instead
of multiple externals. Added test_second_external_rejected to verify
the enforcement.
* feat(memory): add ByteRover memory provider plugin
Implements the ByteRover integration (from PR #3499 by hieuntg81) as a
MemoryProvider plugin instead of direct run_agent.py modifications.
ByteRover provides persistent memory via the brv CLI — a hierarchical
knowledge tree with tiered retrieval (fuzzy text then LLM-driven search).
Local-first with optional cloud sync.
Plugin capabilities:
- prefetch: background brv query for relevant context
- sync_turn: curate conversation turns (threaded, non-blocking)
- on_memory_write: mirror built-in memory writes to brv
- on_pre_compress: extract insights before context compression
Tools (3):
- brv_query: search the knowledge tree
- brv_curate: store facts/decisions/patterns
- brv_status: check CLI version and context tree state
Profile isolation: working directory at $HERMES_HOME/byterover/ (scoped
per profile). Binary resolution cached with thread-safe double-checked
locking. All write operations threaded to avoid blocking the agent
(curate can take 120s with LLM processing).
* fix(memory): thread remaining sync_turns, fix holographic, add config key
Plugin fixes:
- Hindsight: thread sync_turn (was blocking up to 30s via _run_in_thread)
- RetainDB: thread sync_turn (was blocking on HTTP POST)
- Both: shutdown now joins sync threads alongside prefetch threads
Holographic retrieval fixes:
- reason(): removed dead intersection_key computation (bundled but never
used in scoring). Now reuses pre-computed entity_residuals directly,
moved role_content encoding outside the inner loop.
- contradict(): added _MAX_CONTRADICT_FACTS=500 scaling guard. Above
500 facts, only checks the most recently updated ones to avoid O(n^2)
explosion (~125K comparisons at 500 is acceptable).
Config:
- Added memory.provider key to DEFAULT_CONFIG ("" = builtin only).
No version bump needed (deep_merge handles new keys automatically).
* feat(memory): extract Honcho as a MemoryProvider plugin
Creates plugins/honcho-memory/ as a thin adapter over the existing
honcho_integration/ package. All 4 Honcho tools (profile, search,
context, conclude) move from the normal tool registry to the
MemoryProvider interface.
The plugin delegates all work to HonchoSessionManager — no Honcho
logic is reimplemented. It uses the existing config chain:
$HERMES_HOME/honcho.json -> ~/.honcho/config.json -> env vars.
Lifecycle hooks:
- initialize: creates HonchoSessionManager via existing client factory
- prefetch: background dialectic query
- sync_turn: records messages + flushes to API (threaded)
- on_memory_write: mirrors user profile writes as conclusions
- on_session_end: flushes all pending messages
This is a prerequisite for the MemoryManager wiring in run_agent.py.
Once wired, Honcho goes through the same provider interface as all
other memory plugins, and the scattered Honcho code in run_agent.py
can be consolidated into the single MemoryManager integration point.
* feat(memory): wire MemoryManager into run_agent.py
Adds 8 integration points for the external memory provider plugin,
all purely additive (zero existing code modified):
1. Init (~L1130): Create MemoryManager, find matching plugin provider
from memory.provider config, initialize with session context
2. Tool injection (~L1160): Append provider tool schemas to self.tools
and self.valid_tool_names after memory_manager init
3. System prompt (~L2705): Add external provider's system_prompt_block
alongside existing MEMORY.md/USER.md blocks
4. Tool routing (~L5362): Route provider tool calls through
memory_manager.handle_tool_call() before the catchall handler
5. Memory write bridge (~L5353): Notify external provider via
on_memory_write() when the built-in memory tool writes
6. Pre-compress (~L5233): Call on_pre_compress() before context
compression discards messages
7. Prefetch (~L6421): Inject provider prefetch results into the
current-turn user message (same pattern as Honcho turn context)
8. Turn sync + session end (~L8161, ~L8172): sync_all() after each
completed turn, queue_prefetch_all() for next turn, on_session_end()
+ shutdown_all() at conversation end
All hooks are wrapped in try/except — a failing provider never breaks
the agent. The existing memory system, Honcho integration, and all
other code paths are completely untouched.
Full suite: 7222 passed, 4 pre-existing failures.
* refactor(memory): remove legacy Honcho integration from core
Extracts all Honcho-specific code from run_agent.py, model_tools.py,
toolsets.py, and gateway/run.py. Honcho is now exclusively available
as a memory provider plugin (plugins/honcho-memory/).
Removed from run_agent.py (-457 lines):
- Honcho init block (session manager creation, activation, config)
- 8 Honcho methods: _honcho_should_activate, _strip_honcho_tools,
_activate_honcho, _register_honcho_exit_hook, _queue_honcho_prefetch,
_honcho_prefetch, _honcho_save_user_observation, _honcho_sync
- _inject_honcho_turn_context module-level function
- Honcho system prompt block (tool descriptions, CLI commands)
- Honcho context injection in api_messages building
- Honcho params from __init__ (honcho_session_key, honcho_manager,
honcho_config)
- HONCHO_TOOL_NAMES constant
- All honcho-specific tool dispatch forwarding
Removed from other files:
- model_tools.py: honcho_tools import, honcho params from handle_function_call
- toolsets.py: honcho toolset definition, honcho tools from core tools list
- gateway/run.py: honcho params from AIAgent constructor calls
Removed tests (-339 lines):
- 9 Honcho-specific test methods from test_run_agent.py
- TestHonchoAtexitFlush class from test_exit_cleanup_interrupt.py
Restored two regex constants (_SURROGATE_RE, _BUDGET_WARNING_RE) that
were accidentally removed during the honcho function extraction.
The honcho_integration/ package is kept intact — the plugin delegates
to it. tools/honcho_tools.py registry entries are now dead code (import
commented out in model_tools.py) but the file is preserved for reference.
Full suite: 7207 passed, 4 pre-existing failures. Zero regressions.
* refactor(memory): restructure plugins, add CLI, clean gateway, migration notice
Plugin restructure:
- Move all memory plugins from plugins/<name>-memory/ to plugins/memory/<name>/
(byterover, hindsight, holographic, honcho, mem0, openviking, retaindb)
- New plugins/memory/__init__.py discovery module that scans the directory
directly, loading providers by name without the general plugin system
- run_agent.py uses load_memory_provider() instead of get_plugin_memory_providers()
CLI wiring:
- hermes memory setup — interactive curses picker + config wizard
- hermes memory status — show active provider, config, availability
- hermes memory off — disable external provider (built-in only)
- hermes honcho — now shows migration notice pointing to hermes memory setup
Gateway cleanup:
- Remove _get_or_create_gateway_honcho (already removed in prev commit)
- Remove _shutdown_gateway_honcho and _shutdown_all_gateway_honcho methods
- Remove all calls to shutdown methods (4 call sites)
- Remove _honcho_managers/_honcho_configs dict references
Dead code removal:
- Delete tools/honcho_tools.py (279 lines, import was already commented out)
- Delete tests/gateway/test_honcho_lifecycle.py (131 lines, tested removed methods)
- Remove if False placeholder from run_agent.py
Migration:
- Honcho migration notice on startup: detects existing honcho.json or
~/.honcho/config.json, prints guidance to run hermes memory setup.
Only fires when memory.provider is not set and not in quiet mode.
Full suite: 7203 passed, 4 pre-existing failures. Zero regressions.
* feat(memory): standardize plugin config + add per-plugin documentation
Config architecture:
- Add save_config(values, hermes_home) to MemoryProvider ABC
- Honcho: writes to $HERMES_HOME/honcho.json (SDK native)
- Mem0: writes to $HERMES_HOME/mem0.json
- Hindsight: writes to $HERMES_HOME/hindsight/config.json
- Holographic: writes to config.yaml under plugins.hermes-memory-store
- OpenViking/RetainDB/ByteRover: env-var only (default no-op)
Setup wizard (hermes memory setup):
- Now calls provider.save_config() for non-secret config
- Secrets still go to .env via env vars
- Only memory.provider activation key goes to config.yaml
Documentation:
- README.md for each of the 7 providers in plugins/memory/<name>/
- Requirements, setup (wizard + manual), config reference, tools table
- Consistent format across all providers
The contract for new memory plugins:
- get_config_schema() declares all fields (REQUIRED)
- save_config() writes native config (REQUIRED if not env-var-only)
- Secrets use env_var field in schema, written to .env by wizard
- README.md in the plugin directory
* docs: add memory providers user guide + developer guide
New pages:
- user-guide/features/memory-providers.md — comprehensive guide covering
all 7 shipped providers (Honcho, OpenViking, Mem0, Hindsight,
Holographic, RetainDB, ByteRover). Each with setup, config, tools,
cost, and unique features. Includes comparison table and profile
isolation notes.
- developer-guide/memory-provider-plugin.md — how to build a new memory
provider plugin. Covers ABC, required methods, config schema,
save_config, threading contract, profile isolation, testing.
Updated pages:
- user-guide/features/memory.md — replaced Honcho section with link to
new Memory Providers page
- user-guide/features/honcho.md — replaced with migration redirect to
the new Memory Providers page
- sidebars.ts — added both new pages to navigation
* fix(memory): auto-migrate Honcho users to memory provider plugin
When honcho.json or ~/.honcho/config.json exists but memory.provider
is not set, automatically set memory.provider: honcho in config.yaml
and activate the plugin. The plugin reads the same config files, so
all data and credentials are preserved. Zero user action needed.
Persists the migration to config.yaml so it only fires once. Prints
a one-line confirmation in non-quiet mode.
* fix(memory): only auto-migrate Honcho when enabled + credentialed
Check HonchoClientConfig.enabled AND (api_key OR base_url) before
auto-migrating — not just file existence. Prevents false activation
for users who disabled Honcho, stopped using it (config lingers),
or have ~/.honcho/ from a different tool.
* feat(memory): auto-install pip dependencies during hermes memory setup
Reads pip_dependencies from plugin.yaml, checks which are missing,
installs them via pip before config walkthrough. Also shows install
guidance for external_dependencies (e.g. brv CLI for ByteRover).
Updated all 7 plugin.yaml files with pip_dependencies:
- honcho: honcho-ai
- mem0: mem0ai
- openviking: httpx
- hindsight: hindsight-client
- holographic: (none)
- retaindb: requests
- byterover: (external_dependencies for brv CLI)
* fix: remove remaining Honcho crash risks from cli.py and gateway
cli.py: removed Honcho session re-mapping block (would crash importing
deleted tools/honcho_tools.py), Honcho flush on compress, Honcho
session display on startup, Honcho shutdown on exit, honcho_session_key
AIAgent param.
gateway/run.py: removed honcho_session_key params from helper methods,
sync_honcho param, _honcho.shutdown() block.
tests: fixed test_cron_session_with_honcho_key_skipped (was passing
removed honcho_key param to _flush_memories_for_session).
* fix: include plugins/ in pyproject.toml package list
Without this, plugins/memory/ wouldn't be included in non-editable
installs. Hermes always runs from the repo checkout so this is belt-
and-suspenders, but prevents breakage if the install method changes.
* fix(memory): correct pip-to-import name mapping for dep checks
The heuristic dep.replace('-', '_') fails for packages where the pip
name differs from the import name: honcho-ai→honcho, mem0ai→mem0,
hindsight-client→hindsight_client. Added explicit mapping table so
hermes memory setup doesn't try to reinstall already-installed packages.
* chore: remove dead code from old plugin memory registration path
- hermes_cli/plugins.py: removed register_memory_provider(),
_memory_providers list, get_plugin_memory_providers() — memory
providers now use plugins/memory/ discovery, not the general plugin system
- hermes_cli/main.py: stripped 74 lines of dead honcho argparse
subparsers (setup, status, sessions, map, peer, mode, tokens,
identity, migrate) — kept only the migration redirect
- agent/memory_provider.py: updated docstring to reflect new
registration path
- tests: replaced TestPluginMemoryProviderRegistration with
TestPluginMemoryDiscovery that tests the actual plugins/memory/
discovery system. Added 3 new tests (discover, load, nonexistent).
* chore: delete dead honcho_integration/cli.py and its tests
cli.py (794 lines) was the old 'hermes honcho' command handler — nobody
calls it since cmd_honcho was replaced with a migration redirect.
Deleted tests that imported from removed code:
- tests/honcho_integration/test_cli.py (tested _resolve_api_key)
- tests/honcho_integration/test_config_isolation.py (tested CLI config paths)
- tests/tools/test_honcho_tools.py (tested the deleted tools/honcho_tools.py)
Remaining honcho_integration/ files (actively used by the plugin):
- client.py (445 lines) — config loading, SDK client creation
- session.py (991 lines) — session management, queries, flush
* refactor: move honcho_integration/ into the honcho plugin
Moves client.py (445 lines) and session.py (991 lines) from the
top-level honcho_integration/ package into plugins/memory/honcho/.
No Honcho code remains in the main codebase.
- plugins/memory/honcho/client.py — config loading, SDK client creation
- plugins/memory/honcho/session.py — session management, queries, flush
- Updated all imports: run_agent.py (auto-migration), hermes_cli/doctor.py,
plugin __init__.py, session.py cross-import, all tests
- Removed honcho_integration/ package and pyproject.toml entry
- Renamed tests/honcho_integration/ → tests/honcho_plugin/
* docs: update architecture + gateway-internals for memory provider system
- architecture.md: replaced honcho_integration/ with plugins/memory/
- gateway-internals.md: replaced Honcho-specific session routing and
flush lifecycle docs with generic memory provider interface docs
* fix: update stale mock path for resolve_active_host after honcho plugin migration
* fix(memory): address review feedback — P0 lifecycle, ABC contract, honcho CLI restore
Review feedback from Honcho devs (erosika):
P0 — Provider lifecycle:
- Remove on_session_end() + shutdown_all() from run_conversation() tail
(was killing providers after every turn in multi-turn sessions)
- Add shutdown_memory_provider() method on AIAgent for callers
- Wire shutdown into CLI atexit, reset_conversation, gateway stop/expiry
Bug fixes:
- Remove sync_honcho=False kwarg from /btw callsites (TypeError crash)
- Fix doctor.py references to dead 'hermes honcho setup' command
- Cache prefetch_all() before tool loop (was re-calling every iteration)
ABC contract hardening (all backwards-compatible):
- Add session_id kwarg to prefetch/sync_turn/queue_prefetch
- Make on_pre_compress() return str (provider insights in compression)
- Add **kwargs to on_turn_start() for runtime context
- Add on_delegation() hook for parent-side subagent observation
- Document agent_context/agent_identity/agent_workspace kwargs on
initialize() (prevents cron corruption, enables profile scoping)
- Fix docstring: single external provider, not multiple
Honcho CLI restoration:
- Add plugins/memory/honcho/cli.py (from main's honcho_integration/cli.py
with imports adapted to plugin path)
- Restore full hermes honcho command with all subcommands (status, peer,
mode, tokens, identity, enable/disable, sync, peers, --target-profile)
- Restore auto-clone on profile creation + sync on hermes update
- hermes honcho setup now redirects to hermes memory setup
* fix(memory): wire on_delegation, skip_memory for cron/flush, fix ByteRover return type
- Wire on_delegation() in delegate_tool.py — parent's memory provider
is notified with task+result after each subagent completes
- Add skip_memory=True to cron scheduler (prevents cron system prompts
from corrupting user representations — closes#4052)
- Add skip_memory=True to gateway flush agent (throwaway agent shouldn't
activate memory provider)
- Fix ByteRover on_pre_compress() return type: None -> str
* fix(honcho): port profile isolation fixes from PR #4632
Ports 5 bug fixes found during profile testing (erosika's PR #4632):
1. 3-tier config resolution — resolve_config_path() now checks
$HERMES_HOME/honcho.json → ~/.hermes/honcho.json → ~/.honcho/config.json
(non-default profiles couldn't find shared host blocks)
2. Thread host=_host_key() through from_global_config() in cmd_setup,
cmd_status, cmd_identity (--target-profile was being ignored)
3. Use bare profile name as aiPeer (not host key with dots) — Honcho's
peer ID pattern is ^[a-zA-Z0-9_-]+$, dots are invalid
4. Wrap add_peers() in try/except — was fatal on new AI peers, killed
all message uploads for the session
5. Gate Honcho clone behind --clone/--clone-all on profile create
(bare create should be blank-slate)
Also: sanitize assistant_peer_id via _sanitize_id()
* fix(tests): add module cleanup fixture to test_cli_provider_resolution
test_cli_provider_resolution._import_cli() wipes tools.*, cli, and
run_agent from sys.modules to force fresh imports, but had no cleanup.
This poisoned all subsequent tests on the same xdist worker — mocks
targeting tools.file_tools, tools.send_message_tool, etc. patched the
NEW module object while already-imported functions still referenced
the OLD one. Caused ~25 cascade failures: send_message KeyError,
process_registry FileNotFoundError, file_read_guards timeouts,
read_loop_detection file-not-found, mcp_oauth None port, and
provider_parity/codex_execution stale tool lists.
Fix: autouse fixture saves all affected modules before each test and
restores them after, matching the pattern in
test_managed_browserbase_and_modal.py.
Makes provider fallback turn-scoped in long-lived CLI sessions. Previously, a single transient failure pinned the session to the fallback provider for every subsequent turn.
- _primary_runtime dict snapshot at __init__ (model, provider, base_url, api_mode, client_kwargs, compressor state)
- _restore_primary_runtime() at top of run_conversation() — restores all state, resets fallback chain index
- _try_recover_primary_transport() — one extra recovery cycle (client rebuild + cooldown) for transient transport errors on direct endpoints before fallback
- Skipped for aggregator providers (OpenRouter, Nous)
- 25 tests
Inspired by #4612 (@betamod). Closes#4612.
Anthropic extended thinking blocks include an opaque 'signature' field
required for thinking chain continuity across multi-turn tool-use
conversations. Previously, normalize_anthropic_response() extracted
only the thinking text and set reasoning_details=None, discarding the
signature. On subsequent turns the API could not verify the chain.
Changes:
- _to_plain_data(): new recursive SDK-to-dict converter with depth cap
(20 levels) and path-based cycle detection for safety
- _extract_preserved_thinking_blocks(): rehydrates preserved thinking
blocks (including signature) from reasoning_details on assistant
messages, placing them before tool_use blocks as Anthropic requires
- normalize_anthropic_response(): stores full thinking blocks in
reasoning_details via _to_plain_data()
- _extract_reasoning(): adds 'thinking' key to the detail lookup chain
so Anthropic-format details are found alongside OpenRouter format
Salvaged from PR #4503 by @priveperfumes — focused on the thinking
block continuity fix only (cache strategy and other changes excluded).
No model, base_url, or provider is assumed when the user hasn't
configured one. Previously the defaults dict in cli.py, AIAgent
constructor args, and several fallback paths all hardcoded
anthropic/claude-opus-4.6 + openrouter.ai/api/v1 — silently routing
unconfigured users to OpenRouter, which 404s for anyone using a
different provider.
Now empty defaults force the setup wizard to run, and existing users
who already completed setup are unaffected (their config.yaml has
the model they chose).
Files changed:
- cli.py: defaults dict, _DEFAULT_CONFIG_MODEL
- run_agent.py: AIAgent.__init__ defaults, main() defaults
- hermes_cli/config.py: DEFAULT_CONFIG
- hermes_cli/runtime_provider.py: is_fallback sentinel
- acp_adapter/session.py: default_model
- tests: updated to reflect empty defaults
OpenAI's newer models (GPT-5, Codex) give stronger instruction-following
weight to the 'developer' role vs 'system'. Swap the role at the API
boundary in _build_api_kwargs() for the chat_completions path so internal
message representation stays consistent ('system' everywhere).
Applies regardless of provider — OpenRouter, Nous portal, direct, etc.
The codex_responses path (direct OpenAI) uses 'instructions' instead of
message roles, so it's unaffected.
DEVELOPER_ROLE_MODELS constant in prompt_builder.py defines the matching
model name substrings: ('gpt-5', 'codex').
* fix: force-close TCP sockets on client cleanup, detect and recover dead connections
When a provider drops connections mid-stream (e.g. OpenRouter outage),
httpx's graceful close leaves sockets in CLOSE-WAIT indefinitely. These
zombie connections accumulate and can prevent recovery without restarting.
Changes:
- _force_close_tcp_sockets: walks the httpx connection pool and issues
socket.shutdown(SHUT_RDWR) + close() to force TCP RST on every socket
when a client is closed, preventing CLOSE-WAIT accumulation
- _cleanup_dead_connections: probes the primary client's pool for dead
sockets (recv MSG_PEEK), rebuilds the client if any are found
- Pre-turn health check at the start of each run_conversation call that
auto-recovers with a user-facing status message
- Primary client rebuild after stale stream detection to purge pool
- User-facing messages on streaming connection failures:
"Connection to provider dropped — Reconnecting (attempt 2/3)"
"Connection failed after 3 attempts — try again in a moment"
Made-with: Cursor
* fix: pool entry missing base_url for openrouter, clean error messages
- _resolve_runtime_from_pool_entry: add OPENROUTER_BASE_URL fallback
when pool entry has no runtime_base_url (pool entries from auth.json
credential_pool often omit base_url)
- Replace Rich console.print for auth errors with plain print() to
prevent ANSI escape code mangling through prompt_toolkit's stdout patch
- Force-close TCP sockets on client cleanup to prevent CLOSE-WAIT
accumulation after provider outages
- Pre-turn dead connection detection with auto-recovery and user message
- Primary client rebuild after stale stream detection
- User-facing status messages on streaming connection failures/retries
Made-with: Cursor
* fix(gateway): persist memory flush state to prevent redundant re-flushes on restart
The _session_expiry_watcher tracked flushed sessions in an in-memory set
(_pre_flushed_sessions) that was lost on gateway restart. Expired sessions
remained in sessions.json and were re-discovered every restart, causing
redundant AIAgent runs that burned API credits and blocked the event loop.
Fix: Add a memory_flushed boolean field to SessionEntry, persisted in
sessions.json. The watcher sets it after a successful flush. On restart,
the flag survives and the watcher skips already-flushed sessions.
- Add memory_flushed field to SessionEntry with to_dict/from_dict support
- Old sessions.json entries without the field default to False (backward compat)
- Remove the ephemeral _pre_flushed_sessions set from SessionStore
- Update tests: save/load roundtrip, legacy entry compat, auto-reset behavior
Show inline diffs in the CLI transcript when write_file, patch, or
skill_manage modifies files. Captures a filesystem snapshot before the
tool runs, computes a unified diff after, and renders it with ANSI
coloring in the activity feed.
Adds tool_start_callback and tool_complete_callback hooks to AIAgent
for pre/post tool execution notifications.
Also fixes _extract_parallel_scope_path to normalize relative paths
to absolute, preventing the parallel overlap detection from missing
conflicts when the same file is referenced with different path styles.
Gated by display.inline_diffs config option (default: true).
Based on PR #3774 by @kshitijk4poor.
The openai SDK's SyncAPIClient.is_closed is a method, not a property.
getattr(client, 'is_closed', False) returned the bound method object,
which is always truthy — causing _is_openai_client_closed() to report
all clients as closed and triggering unnecessary client recreation
(~100-200ms TCP+TLS overhead per API call).
Fix: check if is_closed is callable and call it, otherwise treat as bool.
Fixes#4377
Co-authored-by: Bartok9 <Bartok9@users.noreply.github.com>
Three bugs prevented credential pool rotation from working when multiple
Codex OAuth tokens were configured:
1. credential_pool was dropped during smart model turn routing.
resolve_turn_route() constructed runtime dicts without it, so the
AIAgent was created without pool access. Fixed in smart_model_routing.py
(no-route and fallback paths), cli.py, and gateway/run.py.
2. Eager fallback fired before pool rotation on 429. The rate-limit
handler at line ~7180 switched to a fallback provider immediately,
before _recover_with_credential_pool got a chance to rotate to the
next credential. Now deferred when the pool still has credentials.
3. (Non-issue) Retry budget was reported as too small, but successful
pool rotations already skip retry_count increment — no change needed.
Reported by community member Schinsly who identified all three root
causes and verified the fix locally with multiple Codex accounts.
* feat(file_tools): harden read_file with size guard, dedup, and device blocking
Three improvements to read_file_tool to reduce wasted context tokens and
prevent process hangs:
1. Character-count guard: reads that produce more than 100K characters
(≈25-35K tokens across tokenisers) are rejected with an error that
tells the model to use offset+limit for a smaller range. The
effective cap is min(file_size, 100K) so small files that happen to
have long lines aren't over-penalised. Large truncated files also
get a hint nudging toward targeted reads.
2. File-read deduplication: when the same (path, offset, limit) is read
a second time and the file hasn't been modified (mtime unchanged),
return a lightweight stub instead of re-sending the full content.
Writes and patches naturally change mtime, so post-edit reads always
return fresh content. The dedup cache is cleared on context
compression — after compression the original read content is
summarised away, so the model needs the full content again.
3. Device path blocking: paths like /dev/zero, /dev/random, /dev/stdin
etc. are rejected before any I/O to prevent process hangs from
infinite-output or blocking-input devices.
Tests: 17 new tests covering all three features plus the dedup-reset-
on-compression integration. All 52 file-read tests pass (35 existing +
17 new). Full tool suite (2124 tests) passes with 0 failures.
* feat: make file_read_max_chars configurable, add docs
Add file_read_max_chars to DEFAULT_CONFIG (default 100K). read_file_tool
reads this on first call and caches for the process lifetime. Users on
large-context models can raise it; users on small local models can lower it.
Also adds a 'File Read Safety' section to the configuration docs
explaining the char limit, dedup behavior, and example values.
When an OAuth token refresh fails on a 401 error, the pool recovery
would return 'not recovered' without trying the next credential in the
pool. This meant users who added a second valid credential via
'hermes auth add' would never see it used when the primary credential
was dead.
Now: try refresh first (handles expired tokens quickly), and if that
fails, rotate to the next available credential — same as 429/402
already did.
Adds three tests covering 401 refresh success, refresh-fail-then-rotate,
and refresh-fail-with-no-remaining-credentials.
Some models (e.g. Kimi K2.5 on Alibaba OpenAI-compatible endpoint)
emit reasoning text followed by a closing </think> without a matching
opening <think> tag. The existing paired-tag regexes in
_strip_think_blocks() cannot match these orphaned tags, so </think>
leaks into user-facing responses on all platforms.
Add a catch-all regex that strips any remaining opening or closing
think/thinking/reasoning/REASONING_SCRATCHPAD tags after the existing
paired-block removal pass.
Closes#4285
* feat(auth): add same-provider credential pools and rotation UX
Add same-provider credential pooling so Hermes can rotate across
multiple credentials for a single provider, recover from exhausted
credentials without jumping providers immediately, and configure
that behavior directly in hermes setup.
- agent/credential_pool.py: persisted per-provider credential pools
- hermes auth add/list/remove/reset CLI commands
- 429/402/401 recovery with pool rotation in run_agent.py
- Setup wizard integration for pool strategy configuration
- Auto-seeding from env vars and existing OAuth state
Co-authored-by: kshitijk4poor <82637225+kshitijk4poor@users.noreply.github.com>
Salvaged from PR #2647
* fix(tests): prevent pool auto-seeding from host env in credential pool tests
Tests for non-pool Anthropic paths and auth remove were failing when
host env vars (ANTHROPIC_API_KEY) or file-backed OAuth credentials
were present. The pool auto-seeding picked these up, causing unexpected
pool entries in tests.
- Mock _select_pool_entry in auxiliary_client OAuth flag tests
- Clear Anthropic env vars and mock _seed_from_singletons in auth remove test
* feat(auth): add thread safety, least_used strategy, and request counting
- Add threading.Lock to CredentialPool for gateway thread safety
(concurrent requests from multiple gateway sessions could race on
pool state mutations without this)
- Add 'least_used' rotation strategy that selects the credential
with the lowest request_count, distributing load more evenly
- Add request_count field to PooledCredential for usage tracking
- Add mark_used() method to increment per-credential request counts
- Wrap select(), mark_exhausted_and_rotate(), and try_refresh_current()
with lock acquisition
- Add tests: least_used selection, mark_used counting, concurrent
thread safety (4 threads × 20 selects with no corruption)
* feat(auth): add interactive mode for bare 'hermes auth' command
When 'hermes auth' is called without a subcommand, it now launches an
interactive wizard that:
1. Shows full credential pool status across all providers
2. Offers a menu: add, remove, reset cooldowns, set strategy
3. For OAuth-capable providers (anthropic, nous, openai-codex), the
add flow explicitly asks 'API key or OAuth login?' — making it
clear that both auth types are supported for the same provider
4. Strategy picker shows all 4 options (fill_first, round_robin,
least_used, random) with the current selection marked
5. Remove flow shows entries with indices for easy selection
The subcommand paths (hermes auth add/list/remove/reset) still work
exactly as before for scripted/non-interactive use.
* fix(tests): update runtime_provider tests for config.yaml source of truth (#4165)
Tests were using OPENAI_BASE_URL env var which is no longer consulted
after #4165. Updated to use model config (provider, base_url, api_key)
which is the new single source of truth for custom endpoint URLs.
* feat(auth): support custom endpoint credential pools keyed by provider name
Custom OpenAI-compatible endpoints all share provider='custom', making
the provider-keyed pool useless. Now pools for custom endpoints are
keyed by 'custom:<normalized_name>' where the name comes from the
custom_providers config list (auto-generated from URL hostname).
- Pool key format: 'custom:together.ai', 'custom:local-(localhost:8080)'
- load_pool('custom:name') seeds from custom_providers api_key AND
model.api_key when base_url matches
- hermes auth add/list now shows custom endpoints alongside registry
providers
- _resolve_openrouter_runtime and _resolve_named_custom_runtime check
pool before falling back to single config key
- 6 new tests covering custom pool keying, seeding, and listing
* docs: add Excalidraw diagram of full credential pool flow
Comprehensive architecture diagram showing:
- Credential sources (env vars, auth.json OAuth, config.yaml, CLI)
- Pool storage and auto-seeding
- Runtime resolution paths (registry, custom, OpenRouter)
- Error recovery (429 retry-then-rotate, 402 immediate, 401 refresh)
- CLI management commands and strategy configuration
Open at: https://excalidraw.com/#json=2Ycqhqpi6f12E_3ITyiwh,c7u9jSt5BwrmiVzHGbm87g
* fix(tests): update setup wizard pool tests for unified select_provider_and_model flow
The setup wizard now delegates to select_provider_and_model() instead
of using its own prompt_choice-based provider picker. Tests needed:
- Mock select_provider_and_model as no-op (provider pre-written to config)
- Call _stub_tts BEFORE custom prompt_choice mock (it overwrites it)
- Pre-write model.provider to config so the pool step is reached
* docs: add comprehensive credential pool documentation
- New page: website/docs/user-guide/features/credential-pools.md
Full guide covering quick start, CLI commands, rotation strategies,
error recovery, custom endpoint pools, auto-discovery, thread safety,
architecture, and storage format.
- Updated fallback-providers.md to reference credential pools as the
first layer of resilience (same-provider rotation before cross-provider)
- Added hermes auth to CLI commands reference with usage examples
- Added credential_pool_strategies to configuration guide
* chore: remove excalidraw diagram from repo (external link only)
* refactor: simplify credential pool code — extract helpers, collapse extras, dedup patterns
- _load_config_safe(): replace 4 identical try/except/import blocks
- _iter_custom_providers(): shared generator for custom provider iteration
- PooledCredential.extra dict: collapse 11 round-trip-only fields
(token_type, scope, client_id, portal_base_url, obtained_at,
expires_in, agent_key_id, agent_key_expires_in, agent_key_reused,
agent_key_obtained_at, tls) into a single extra dict with
__getattr__ for backward-compatible access
- _available_entries(): shared exhaustion-check between select and peek
- Dedup anthropic OAuth seeding (hermes_pkce + claude_code identical)
- SimpleNamespace replaces class _Args boilerplate in auth_commands
- _try_resolve_from_custom_pool(): shared pool-check in runtime_provider
Net -17 lines. All 383 targeted tests pass.
---------
Co-authored-by: kshitijk4poor <82637225+kshitijk4poor@users.noreply.github.com>
Adds /btw <question> — ask a quick follow-up using the current
session context without interrupting the main conversation.
- Snapshots conversation history, answers with a no-tools agent
- Response is not persisted to session history or DB
- Runs in a background thread (CLI) / async task (gateway)
- Per-session guard prevents concurrent /btw in gateway
Implementation:
- model_tools.py: enabled_toolsets=[] now correctly means "no tools"
(was falsy, fell through to default "all tools")
- run_agent.py: persist_session=False gates _persist_session()
- cli.py: _handle_btw_command (background thread, Rich panel output)
- gateway/run.py: _handle_btw_command + _run_btw_task (async task)
- hermes_cli/commands.py: CommandDef for "btw"
Inspired by PR #3504 by areu01or00, reimplemented cleanly on current
main with the enabled_toolsets=[] fix and without the __btw_no_tools__
hack.
When context compression fails, users now see hints suggesting /new
or /compress instead of a dead-end error. Covers all 4 error paths:
payload-too-large, max compression attempts (2 paths), and context
length exceeded.
Closes#4061
Salvaged from PR #4076 by SHL0MS.
Co-authored-by: SHL0MS <SHL0MS@users.noreply.github.com>
When context compression fires during run_conversation() in the gateway,
the compressed messages were silently lost on the next turn. Two bugs:
1. Agent-side: _flush_messages_to_session_db() calculated
flush_from = max(len(conversation_history), _last_flushed_db_idx).
After compression, _last_flushed_db_idx was correctly reset to 0,
but conversation_history still had its original pre-compression
length (e.g. 200). Since compressed messages are shorter (~30),
messages[200:] was empty — nothing written to the new session's
SQLite.
Fix: Set conversation_history = None after each _compress_context()
call so start_idx = 0 and all compressed messages are flushed.
2. Gateway-side: history_offset was always len(agent_history) — the
original pre-compression length. After compression shortened the
message list, agent_messages[200:] was empty, causing the gateway
to fall back to writing only a user/assistant pair, losing the
compressed summary and tail context.
Fix: Detect session splits (agent.session_id != original) and set
history_offset = 0 so all compressed messages are written to JSONL.
* feat: add /yolo slash command to toggle dangerous command approvals
Adds a /yolo command that toggles HERMES_YOLO_MODE at runtime, skipping
all dangerous command approval prompts for the current session. Works in
both CLI and gateway (Telegram, Discord, etc.).
- /yolo -> ON: all commands auto-approved, no confirmation prompts
- /yolo -> OFF: normal approval flow restored
The --yolo CLI flag already existed for launch-time opt-in. This adds
the ability to toggle mid-session without restarting.
Session-scoped — resets when the process ends. Uses the existing
HERMES_YOLO_MODE env var that check_all_command_guards() already
respects.
* fix: prevent context pressure warning spam (agent loop + gateway rate-limit)
Two complementary fixes for repeated context pressure warnings spamming
gateway users (Telegram, Discord, etc.):
1. Agent-level loop fix (run_agent.py):
After compression, only reset _context_pressure_warned if the
post-compression estimate is actually below the 85% warning level.
Previously the flag was unconditionally reset, causing the warning
to re-fire every loop iteration when compression couldn't reduce
below 85% of the threshold (e.g. very low threshold like 15%,
or system prompt alone exceeds the warning level).
2. Gateway-level rate-limit (gateway/run.py, salvaged from PR #3786):
Per-chat_id cooldown of 1 hour on compression warning messages.
Both warning paths ('still large after compression' and 'compression
failed') are gated. Defense-in-depth — even if the agent-level fix
has edge cases, users won't see more than one warning per hour.
Co-authored-by: dlkakbs <dlkakbs@users.noreply.github.com>
---------
Co-authored-by: dlkakbs <dlkakbs@users.noreply.github.com>
When the API doesn't provide a call_id for tool calls, the fallback
generated a random uuid4 hex. This made every API call's input unique
when replayed, preventing OpenAI's prompt cache from matching the
prefix across turns.
Replaced all four uuid4 fallback sites with a deterministic hash of
(function_name, arguments, position_index). The same tool call now
always produces the same fallback call_id, preserving cache-friendly
input stability.
Affected code paths:
- _chat_messages_to_responses_input() — Codex input reconstruction
- _normalize_codex_response() — function_call and custom_tool_call
- _build_assistant_message() — assistant message construction
When stdout is closed (piped to a dead process, broken terminal),
Python raises ValueError('I/O operation on closed file'), not OSError.
_safe_print and the API error printer only caught OSError, letting the
ValueError propagate and crash the agent.
Salvaged from PR #3760 by @apexscaleai. Fixes#3534.
Co-authored-by: apexscaleai <apexscaleai@users.noreply.github.com>
Tool call previews (paths, commands, queries) were hardcoded to truncate
at 35-40 chars across CLI spinners, completion lines, and gateway progress
messages. Users could not see full file paths in tool output.
New config option: display.tool_preview_length (default 0 = no limit).
Set a positive number to truncate at that length.
Changes:
- display.py: module-level _tool_preview_max_len with getter/setter;
build_tool_preview() and get_cute_tool_message() _trunc/_path respect it
- cli.py: reads config at startup, spinner widget respects config
- gateway/run.py: reads config per-message, progress callback respects config
- run_agent.py: removed redundant 30-char quiet-mode spinner truncation
- config.py: added display.tool_preview_length to DEFAULT_CONFIG
Reported by kriskaminski
When compression creates a child session with a new session_id,
session_log_file was still pointing to the old session's JSON file.
This caused _save_session_log() to write new data to the wrong file.
Closes#3731.
Co-authored-by: kelsia14 <kelsia14@users.noreply.github.com>
Some providers (Fireworks AI) reject tools=null, and others (Anthropic)
reject tools=[]. The safest approach is to not include the key at all
when there are no tools — the OpenAI SDK treats a missing parameter as
NOT_GIVEN and omits it from the request entirely.
Inspired by PR #3736 (@kelsia14).
Extends the single fallback_model mechanism into an ordered chain.
When the primary model fails, Hermes tries each fallback provider in
sequence until one succeeds or the chain is exhausted.
Config format (new):
fallback_providers:
- provider: openrouter
model: anthropic/claude-sonnet-4
- provider: openai
model: gpt-4o
Legacy single-dict fallback_model format still works unchanged.
Key fix vs original PR: the call sites in the retry loop now use
_fallback_index < len(_fallback_chain) instead of the old one-shot
_fallback_activated guard, so the chain actually advances through
all configured providers.
Changes:
- run_agent.py: _fallback_chain list + _fallback_index replaces
one-shot _fallback_model; _try_activate_fallback() advances
through chain; failed provider resolution skips to next entry;
call sites updated to allow chain advancement
- cli.py: reads fallback_providers with legacy fallback_model compat
- gateway/run.py: same
- hermes_cli/config.py: fallback_providers: [] in DEFAULT_CONFIG
- tests: 12 new chain tests + 6 existing test fixtures updated
Co-authored-by: uzaylisak <uzaylisak@users.noreply.github.com>
When hitting rate limits (429), the agent now:
- Extracts the Retry-After header from the provider response and uses it
as the wait time instead of blind exponential backoff (capped at 120s)
- Shows rate-limit-specific messaging: 'Rate limit reached. Waiting Xs
before retry (attempt N/M)...'
- Shows a distinct exhaustion message: 'Rate limit persisted after N
retries. Please try again later.'
Non-429 errors keep the existing exponential backoff and generic messaging.
Co-authored-by: ygd58 <ygd58@users.noreply.github.com>
When a user runs 'hermes update', the Python process caches old modules
in sys.modules. After git pull updates files on disk, lazy imports of
newly-updated modules fail because they try to import display_hermes_home
from the cached (old) hermes_constants which doesn't have the function.
This specifically broke the gateway auto-restart in cmd_update — importing
hermes_cli/gateway.py triggered the top-level 'from hermes_constants
import display_hermes_home' against the cached old module. The ImportError
was silently caught, so the gateway was never restarted after update.
Users with a running gateway then hit the ImportError on their next
Telegram/Discord message when the stale gateway process lazily loaded
run_agent.py (new version) which also had the top-level import.
Fixes:
- hermes_cli/gateway.py: lazy import at call site (line 940)
- run_agent.py: lazy import at call site (line 6927)
- tools/terminal_tool.py: lazy imports at 3 call sites
- tools/tts_tool.py: static schema string (no module-level call)
- hermes_cli/auth.py: lazy import at call site (line 2024)
- hermes_cli/main.py: reload hermes_constants after git pull in cmd_update
Also fixes 4 pre-existing test failures in test_parse_env_var caused by
NameError on display_hermes_home in terminal_tool.py.
Prep for profiles: user-facing messages now use display_hermes_home() so
diagnostic output shows the correct path for each profile.
New helper: display_hermes_home() in hermes_constants.py
12 files swept, ~30 user-facing string replacements.
Includes dynamic TTS schema description.
Self-hosted Honcho on localhost doesn't require authentication, but
both the activation gates and the SDK client required an API key.
Combined fix from three contributor PRs:
- Relax all 8 activation gates to accept (api_key OR base_url) as
valid credentials (#3482 by @cameronbergh)
- Use 'local' placeholder for the SDK client when base_url points to
localhost/127.0.0.1/::1 (#3570 by @ygd58)
Files changed: run_agent.py (2 gates), cli.py (1 gate),
gateway/run.py (1 gate), honcho_integration/cli.py (2 gates),
hermes_cli/doctor.py (2 gates), honcho_integration/client.py (SDK).
Co-authored-by: cameronbergh <cameronbergh@users.noreply.github.com>
Co-authored-by: ygd58 <ygd58@users.noreply.github.com>
Co-authored-by: devorun <devorun@users.noreply.github.com>
Pasting text from rich-text editors (Google Docs, Word, etc.) can inject
lone surrogate characters (U+D800..U+DFFF) that are invalid UTF-8.
The OpenAI SDK serializes messages with ensure_ascii=False, then encodes
to UTF-8 for the HTTP body — surrogates crash this with:
UnicodeEncodeError: 'utf-8' codec can't encode character '\udce2'
Three-layer fix:
1. Primary: sanitize user_message at the top of run_conversation()
2. CLI: sanitize in chat() before appending to conversation_history
3. Safety net: catch UnicodeEncodeError in the API error handler,
sanitize the entire messages list in-place, and retry once.
Also exclude UnicodeEncodeError from is_local_validation_error
so it doesn't get classified as non-retryable.
Includes 14 new tests covering the sanitization helpers and the
integration with run_conversation().
Ollama reuses index 0 for every tool call in a parallel batch,
distinguishing them only by id. The streaming accumulator now
detects a new non-empty id at an already-active index and redirects
it to a fresh slot, preventing names and arguments from being
concatenated into a single tool call.
No-op for normal providers that use incrementing indices.
Co-authored-by: dmater01 <dmater01@users.noreply.github.com>
* fix(provider): remove MiniMax /v1→/anthropic auto-correction to allow user override
The minimax-specific auto-correction in runtime_provider.py was
preventing users from overriding to the OpenAI-compatible endpoint
via MINIMAX_BASE_URL. Users in certain regions get nginx 404 on
api.minimax.io/anthropic and need to switch to api.minimax.chat/v1.
The generic URL-suffix detection already handles /anthropic →
anthropic_messages, so the minimax-specific code was redundant for
the default path and harmful for the override path.
Now: default /anthropic URL works via generic detection, user
override to /v1 gets chat_completions mode naturally.
Closes#3546 (different approach — respects user overrides instead
of changing the default endpoint).
* fix(display): show reasoning during streaming even when tool calls suppress content
When a model generates content (containing <REASONING_SCRATCHPAD> tags)
alongside tool calls in the same API response, content deltas were
suppressed from streaming once any tool call chunk arrived. This
prevented the CLI's tag extraction from running, so reasoning was
never shown during streaming. The post-response fallback then
displayed reasoning AFTER the already-visible streamed response,
creating a confusing reversed order.
Fix: route suppressed content to stream_delta_callback even when tool
calls are present. The CLI's _stream_delta handles tag extraction —
reasoning tags are routed to the reasoning display box, while
non-reasoning text is handled by the existing stream display logic.
This ensures reasoning appears before tool execution and the final
response, matching the expected visual order.
The TOOL_USE_ENFORCEMENT_GUIDANCE injection (added in #3528) was
hardcoded to only match gpt/codex model names. This makes it a
config option so users can turn it on for any model family.
New config key: agent.tool_use_enforcement
- "auto" (default): matches gpt/codex (existing behavior)
- true: inject for all models
- false: never inject
- list of strings: custom model-name substrings to match
e.g. ["gpt", "codex", "deepseek", "qwen"]
No version bump needed — deep merge provides the default
automatically for existing installs.
12 new tests covering all config modes.
The plugin system defined six lifecycle hooks but only pre_tool_call and
post_tool_call were invoked. This activates the remaining four so that
external plugins (e.g. memory systems) can hook into the conversation
loop without touching core code.
Hook semantics:
- on_session_start: fires once when a new session is created
- pre_llm_call: fires once per turn before the tool-calling loop;
plugins can return {"context": "..."} to inject into the ephemeral
system prompt (not cached, not persisted)
- post_llm_call: fires once per turn after the loop completes, with
user_message and assistant_response for sync/storage
- on_session_end: fires at the end of every run_conversation call
invoke_hook() now returns a list of non-None callback return values,
enabling pre_llm_call context injection while remaining backward
compatible (existing hooks that return None are unaffected).
Salvaged from PR #2823.
Co-authored-by: Nicolò Boschi <boschi1997@gmail.com>
Root cause: Anthropic buffers entire tool call arguments and goes silent
for minutes while thinking (verified: 167s gap with zero SSE events on
direct API). OpenRouter's upstream proxy times out after ~125s of
inactivity and drops the connection with 'Network connection lost'.
Fix: Send the x-anthropic-beta: fine-grained-tool-streaming-2025-05-14
header for Claude models on OpenRouter. This makes Anthropic stream
tool call arguments token-by-token instead of buffering them, keeping
the connection alive through OpenRouter's proxy.
Live-tested: the exact prompt that consistently failed at ~128s now
completes successfully — 2,972 lines written, 49K tokens, 8 minutes.
Additional improvements:
1. Send explicit max_tokens for Claude through OpenRouter. Without it,
OpenRouter defaults to 65,536 (confirmed via echo_upstream_body) —
only half of Opus 4.6's 128K limit.
2. Classify SSE 'Network connection lost' as retryable in the streaming
inner retry loop. The OpenAI SDK raises APIError from SSE error
events, which was bypassing our transient error retry logic.
3. Actionable diagnostic guidance when stream-drop retries exhaust.
Cherry-pick of feat/gpt-tool-steering with modifications:
1. Tool-use enforcement prompt (refactored from GPT-specific):
- Renamed GPT_TOOL_USE_GUIDANCE -> TOOL_USE_ENFORCEMENT_GUIDANCE
- Added TOOL_USE_ENFORCEMENT_MODELS tuple: ('gpt', 'codex')
- Injection logic now checks against the tuple instead of hardcoding
'gpt' — adding new model families is a one-line change
- Addresses models describing actions instead of making tool calls
2. Budget warning history stripping:
- _strip_budget_warnings_from_history() strips _budget_warning JSON
keys and [BUDGET WARNING: ...] text from tool results at the start
of run_conversation()
- Prevents old budget warnings from poisoning subsequent turns
Based on PR #3479 by teknium1.
* fix: cap context pressure percentage at 100% in display
The forward-looking token estimate can overshoot the compaction threshold
(e.g. a large tool result pushes it from 70% to 109% in one step). The
progress bar was already capped via min(), but pct_int was not — causing
the user to see '109% to compaction' which is confusing.
Cap pct_int at 100 in both CLI and gateway display functions.
Reported by @JoshExile82.
* refactor: use real API token counts for compression decisions
Replace the rough chars/3 estimation with actual prompt_tokens +
completion_tokens from the API response. The estimation was needed to
predict whether tool results would push context past the threshold, but
the default 50% threshold leaves ample headroom — if tool results push
past it, the next API call reports real usage and triggers compression
then.
This removes all estimation from the compression and context pressure
paths, making both 100% data-driven from provider-reported token counts.
Also removes the dead _msg_count_before_tools variable.
When finish_reason='length' and the response contains only reasoning
(think blocks or empty content), the model exhausted its output token
budget on thinking with nothing left for the actual response.
Previously, this fell into either:
- chat_completions: 3 useless continuation retries (model hits same limit)
- anthropic/codex: generic 'Response truncated' error with rollback
Now: detect the think-only + length condition early and return immediately
with a targeted error message: 'Model used all output tokens on reasoning
with none left for the response. Try lowering reasoning effort or
increasing max_tokens.'
This saves 2 wasted API calls on the chat_completions path and gives
users actionable guidance instead of a cryptic error.
The existing think-only retry logic (finish_reason='stop') is unchanged —
that's a genuine model glitch where retrying can help.
The Anthropic adapter defaulted to max_tokens=16384 when no explicit value
was configured. This severely limits thinking-enabled models where thinking
tokens count toward max_tokens:
- Claude Opus 4.6 supports 128K output but was capped at 16K
- Claude Sonnet 4.6 supports 64K output but was capped at 16K
With extended thinking (adaptive or budget-based), the model could exhaust
the entire 16K on reasoning, leaving zero tokens for the actual response.
This caused two user-visible errors:
- 'Response truncated (finish_reason=length)' — thinking consumed most tokens
- 'Response only contains think block with no content' — thinking consumed all
Fix: add _ANTHROPIC_OUTPUT_LIMITS lookup table (sourced from Anthropic docs
and Cline's model catalog) and use the model's actual output limit as the
default. Unknown future models default to 128K (the current maximum).
Also adds context_length clamping: if the user configured a smaller context
window (e.g. custom endpoint), max_tokens is clamped to context_length - 1
to avoid exceeding the window.
Closes#2706
Models like GLM-5/5.1 can think for 15+ minutes. The previous 900s
(15 min) default for HERMES_API_TIMEOUT killed legitimate requests.
Raised to 1800s (30 min) in both places that read the env var:
- _build_api_kwargs() timeout (non-streaming total timeout)
- _call_chat_completions() write timeout (streaming connection)
The streaming per-chunk read timeout (60s) and stale stream detector
(180-300s) are unchanged — those are appropriate for inter-chunk timing.
Two independent bugs caused the reasoning box to appear three times when
the model produced reasoning + tool_calls:
Bug A: _build_assistant_message() re-fired reasoning_callback with the full
reasoning text even when streaming had already displayed it. The original
guard only checked structured reasoning_content deltas, but reasoning also
arrives via content tag extraction (<REASONING_SCRATCHPAD>/<think> tags
in delta.content), which went through _fire_stream_delta not
_fire_reasoning_delta. Fix: skip the callback entirely when streaming is
active — both paths display reasoning during the stream. Any reasoning not
shown during streaming is caught by the CLI post-response fallback.
Bug B: The post-response reasoning display checked _reasoning_stream_started,
but that flag was reset by _reset_stream_state() during intermediate turn
boundaries (when stream_delta_callback(None) fires between tool calls).
Introduced _reasoning_shown_this_turn flag that persists across the tool
loop and is only reset at the start of each user turn.
Live-tested in PTY: reasoning now shows exactly once per API call, no
duplicates across tool-calling loops.
The stale stream detector (90s timeout) was killing healthy connections
during the model's thinking phase, producing self-inflicted
RemoteProtocolError ("peer closed connection without sending complete
message body"). Three issues:
1. last_chunk_time was never reset between inner stream retries, so
subsequent attempts inherited the previous attempt's stale budget
2. The non-streaming fallback path didn't reset the timer either
3. 90s base timeout was too aggressive for large-context Opus sessions
where thinking time before first token routinely exceeds 90s
Fix: reset last_chunk_time at the start of each streaming attempt and
before the non-streaming fallback. Increase base timeout to 180s and
scale to 300s for >100K token contexts.
Made-with: Cursor
Three improvements salvaged from PR #3225 by Mibayy:
1. Add /resume slash command handler in CLI process_command(). The
command was registered in the commands registry but had no handler,
so typing /resume produced 'Unknown command'. The handler resolves
by title or session ID, ends the current session cleanly, loads
conversation history from SQLite, re-opens the target session, and
syncs the AIAgent instance. Follows the same pattern as new_session().
2. Add truncation guard in _save_session_log(). When resuming a session
whose messages weren't fully written to SQLite, the agent starts with
partial history and the first save would overwrite the full JSON log
on disk. The guard reads the existing file and skips the write if it
already has more messages than the current batch.
3. Add reopen_session() method to SessionDB. Proper API for clearing
ended_at/end_reason instead of reaching into _conn directly.
Note: Bug 1 from the original PR (INSERT OR IGNORE + _session_db = None)
is already fixed on main — skipped as redundant.
Closes#3123.
When _try_activate_fallback() switches to the fallback model, it
updates the agent's model/provider/client but never touches
self.context_compressor. The compressor keeps the primary model's
context_length and threshold_tokens, so compression decisions use
wrong limits — a 200K primary → 32K fallback still uses 200K-based
thresholds, causing oversized sessions to overflow the fallback.
Update the compressor's model, credentials, context_length, and
threshold_tokens after fallback activation using get_model_context_length()
for the new model.
Cherry-picked from PR #3202 by binhnt92.
Co-authored-by: binhnt92 <binhnt.ht.92@gmail.com>
* fix(gateway): silence flush agent terminal output
quiet_mode=True only suppresses AIAgent init messages.
Tool call output still leaks to the terminal through
_safe_print → _print_fn during session reset/expiry.
Since #2670 injected live memory state into the flush prompt,
the flush agent now reliably calls memory tools — making the
output leak noticeable for the first time.
Set _print_fn to a no-op so the background flush is fully silent.
* test(gateway): add test for flush agent terminal silence + fix dotenv mock
- Add TestFlushAgentSilenced: verifies _print_fn is set to a no-op on
the flush agent so tool output never leaks to the terminal
- Fix pre-existing test failures: replace patch('run_agent.AIAgent')
with sys.modules mock to avoid importing run_agent (requires openai)
- Add autouse _mock_dotenv fixture so all tests in this file run
without the dotenv package installed
* fix(display): route KawaiiSpinner output through print_fn to fully silence flush agent
The previous fix set tmp_agent._print_fn = no-op on the flush agent but
spinner output and quiet-mode cute messages bypassed _print_fn entirely:
- KawaiiSpinner captured sys.stdout at __init__ and wrote directly to it
- quiet-mode tool results used builtin print() instead of _safe_print()
Add optional print_fn parameter to KawaiiSpinner.__init__; _write routes
through it when set. Pass self._print_fn to all spinner construction sites
in run_agent.py and change the quiet-mode cute message print to _safe_print.
The existing gateway fix (tmp_agent._print_fn = lambda) now propagates
correctly through both paths.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(gateway): silence hygiene and compression background agents
Two more background AIAgent instances in the gateway were created with
quiet_mode=True but without _print_fn = no-op, causing tool output to
leak to the terminal:
- _hyg_agent (in-turn hygiene memory agent)
- tmp_agent (_compress_context path)
Apply the same _print_fn no-op pattern used for the flush agent.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* chore(display): remove unused _last_flush_time from KawaiiSpinner
Attribute was set but never read; upstream already removed it.
Leftover from conflict resolution during rebase onto upstream/main.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Dilee <uzmpsk.dilekakbas@gmail.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
The background memory/skill review (_spawn_background_review) runs
after the agent response when turn/iteration counters exceed their
thresholds. It saves memories and skills, then prints a summary like
'💾 Memory updated · User profile updated'. In CLI mode this goes to
the terminal via _safe_print. In gateway mode, _safe_print routes to
print() which goes to stdout — invisible to the user.
Add a background_review_callback attribute to AIAgent. When set, the
background review thread calls it with the summary string after saves
complete. The gateway wires this to adapter.send() via the same
run_coroutine_threadsafe bridge used by status_callback, delivering
the notification to the user's chat.
- add managed modal and gateway-backed tool integrations\n- improve CLI setup, auth, and configuration for subscriber flows\n- expand tests and docs for managed tool support
When third-party tools (Paperclip orchestrator, etc.) spawn hermes chat
as a subprocess, their sessions pollute user session history and search.
- hermes chat --source <tag> (also HERMES_SESSION_SOURCE env var)
- exclude_sources parameter on list_sessions_rich() and search_messages()
- Sessions with source=tool hidden from sessions list/browse/search
- Third-party adapters pass --source tool to isolate agent sessions
Cherry-picked from PR #3208 by HenkDz.
Co-authored-by: Henkey <noonou7@gmail.com>
except Exception does not catch KeyboardInterrupt (inherits from
BaseException). A second Ctrl+C during exit cleanup aborts pending
writes — Honcho observations dropped, SQLite sessions left unclosed,
cron job sessions never marked ended.
Changed to except (Exception, KeyboardInterrupt) at all five sites:
- cli.py: honcho.shutdown() and end_session() in finally exit block
- run_agent.py: _flush_honcho_on_exit atexit handler
- cron/scheduler.py: end_session() and close() in job finally block
Tests exercise the actual production code paths and confirm
KeyboardInterrupt propagates without the fix.
Co-authored-by: dieutx <dangtc94@gmail.com>
* fix(session-db): survive CLI/gateway concurrent write contention
Closes#3139
Three layered fixes for the scenario where CLI and gateway write to
state.db concurrently, causing create_session() to fail with
'database is locked' and permanently disabling session_search on the
gateway side.
1. Increase SQLite connection timeout: 10s -> 30s
hermes_state.py: longer window for the WAL writer to finish a batch
flush before the other process gives up entirely.
2. INSERT OR IGNORE in create_session
hermes_state.py: prevents IntegrityError on duplicate session IDs
(e.g. gateway restarts while CLI session is still alive).
3. Don't null out _session_db on create_session failure (main fix)
run_agent.py: a transient lock at agent startup must not permanently
disable session_search for the lifetime of that agent instance.
_session_db now stays alive so subsequent flushes and searches work
once the lock clears.
4. New ensure_session() helper + call it during flush
hermes_state.py: INSERT OR IGNORE for a minimal session row.
run_agent.py _flush_messages_to_session_db: calls ensure_session()
before appending messages, so the FK constraint is satisfied even
when create_session() failed at startup. No-op when the row exists.
* fix(state): release lock between context queries in search_messages
The context-window queries (one per FTS5 match) were running inside
the same lock acquisition as the primary FTS5 query, holding the lock
for O(N) sequential SQLite round-trips. Move per-match context fetches
outside the outer lock block so each acquires the lock independently,
keeping critical sections short and allowing other threads to interleave.
* fix(session): prefer longer source in load_transcript to prevent legacy truncation
When a long-lived session pre-dates SQLite storage (e.g. sessions
created before the DB layer was introduced, or after a clean
deployment that reset the DB), _flush_messages_to_session_db only
writes the *new* messages from the current turn to SQLite — it skips
messages already present in conversation_history, assuming they are
already persisted.
That assumption fails for legacy JSONL-only sessions:
Turn N (first after DB migration):
load_transcript(id) → SQLite: 0 → falls back to JSONL: 994 ✓
_flush_messages_to_session_db: skip first 994, write 2 new → SQLite: 2
Turn N+1:
load_transcript(id) → SQLite: 2 → returns immediately ✗
Agent sees 2 messages of history instead of 996
The same pattern causes the reported symptom: session JSON truncated
to 4 messages (_save_session_log writes agent.messages which only has
2 history + 2 new = 4).
Fix: always load both sources and return whichever is longer. For a
fully-migrated session SQLite will always be ≥ JSONL, so there is no
regression. For a legacy session that hasn't been bootstrapped yet,
JSONL wins and the full history is restored.
Closes#3212
* test: add load_transcript source preference tests for #3212
Covers: JSONL longer returns JSONL, SQLite longer returns SQLite,
SQLite empty falls back to JSONL, both empty returns empty, equal
length prefers SQLite (richer reasoning fields).
---------
Co-authored-by: Mibayy <mibayy@hermes.ai>
Co-authored-by: kewe63 <kewe.3217@gmail.com>
Co-authored-by: Mibayy <mibayy@users.noreply.github.com>
Two improvements salvaged from PR #2600 (paraddox):
1. Preflight compression now counts tool schema tokens alongside system
prompt and messages. With 50+ tools enabled, schemas can add 20-30K
tokens that were previously invisible to the estimator, delaying
compression until the API rejected the request.
2. Context probe persistence guard: when the agent steps down context
tiers after a context-length error, only provider-confirmed numeric
limits (parsed from the error message) are cached to disk. Guessed
fallback tiers from get_next_probe_tier() stay in-memory only,
preventing wrong values from polluting the persistent cache.
Co-authored-by: paraddox <paraddox@users.noreply.github.com>
_mute_post_response was set True whenever a turn had both content
and tool_calls, suppressing ALL subsequent _vprint output including
tool completion messages. This meant users only saw "preparing
search_files..." but never the result.
Now only mutes output when every tool in the batch is housekeeping
(memory, todo, skill_manage, session_search). Substantive tools
like search_files, read_file, write_file, terminal etc. keep their
completion messages visible.
Also fixes: run_conversation no longer raises on max retries
(returns graceful error dict instead), and cli.py wraps the agent
thread in try/except as a safety net.
Made-with: Cursor
The non-streaming API call path (_interruptible_api_call) had no
wall-clock timeout. When providers keep connections alive with SSE
keep-alive pings but never deliver a response, httpx's inactivity
timeout never fires and the call hangs indefinitely.
Subagents always used the non-streaming path because they have no
stream consumers (quiet_mode=True). This caused delegate_task to
hang for 40+ minutes in production.
The streaming path has two layers of protection:
- httpx read timeout (60s, HERMES_STREAM_READ_TIMEOUT)
- Stale stream detection (90s, HERMES_STREAM_STALE_TIMEOUT)
Both work because streaming sends chunks continuously — a 90-second
gap between chunks genuinely means the connection is broken, even for
reasoning models that take minutes to complete.
Now run_conversation() always prefers the streaming path. The streaming
method falls back to non-streaming automatically if the provider
doesn't support it. Stream delta callbacks are no-ops when no
consumers are registered, so there's no overhead for subagents.
Add _emit_status() helper that sends lifecycle notifications to both
CLI (via _vprint force=True) and gateway (via status_callback). No
retry, fallback, or compression path is silent anymore.
Pathways surfaced:
- General retry backoff: was logger-only, now shows countdown
- Provider fallback: changed raw print() to _emit_status for gateway
- Rate limit eager fallback: new notification before switching
- Empty/malformed response fallback: new notification
- Client error fallback: new notification with HTTP status
- Max retries fallback: new notification before attempting
- Max retries giving up: upgraded from _vprint to _emit_status
- Compression retry (413 + context overflow): upgraded to _emit_status
- Compression success + retry: upgraded to _emit_status (2 instances)
Three categories of cleanup, all zero-behavioral-change:
1. F-strings without placeholders (154 fixes across 29 files)
- Converted f'...' to '...' where no {expression} was present
- Heaviest files: run_agent.py (24), cli.py (20), honcho_integration/cli.py (34)
2. Simplify defensive patterns in run_agent.py
- Added explicit self._is_anthropic_oauth = False in __init__ (before
the api_mode branch that conditionally sets it)
- Replaced 7x getattr(self, '_is_anthropic_oauth', False) with direct
self._is_anthropic_oauth (attribute always initialized now)
- Added _is_openrouter_url() and _is_anthropic_url() helper methods
- Replaced 3 inline 'openrouter' in self._base_url_lower checks
3. Remove dead code in small files
- hermes_cli/claw.py: removed unused 'total' computation
- tools/fuzzy_match.py: removed unused strip_indent() function and
pattern_stripped variable
Full test suite: 6184 passed, 0 failures
E2E PTY: banner clean, tool calls work, zero garbled ANSI
run_conversation raised the raw exception after exhausting retries,
which crashed the background thread in cli.py (unhandled exception
in Thread). Now returns a proper error result dict with failed=True
and persists the session, matching the pattern used by other error
paths (invalid responses, empty content, etc.).
Also wraps cli.py's run_agent thread function in try/except as a
safety net against any future unhandled exceptions from
run_conversation.
Made-with: Cursor
Local models (Ollama, LM Studio) embed reasoning in <think> tags in
delta.content. During streaming, _stream_delta() already displays these
blocks. Then _build_assistant_message() extracts them again and fires
reasoning_callback, causing duplicate display.
Track whether reasoning came from structured fields (reasoning_content)
vs <think> tag extraction. Only fire the callback for <think>-extracted
reasoning when stream_delta_callback is NOT active. Structured reasoning
always fires regardless.
Salvaged from PR #2076 by dusterbloom (Fix A only — Fix B was already
covered by PR #3013's _current_reasoning_callback centralization).
Closes#2069.
Three problems with API error debugging:
1. Terminal showed str(error)[:200] — raw HTML gibberish for Cloudflare
502/503 pages instead of "502 Bad Gateway"
2. errors.log dumped the entire HTML page as unstructured text
3. _dump_api_request_debug was never called when retries exhausted,
only for non-retryable 4xx errors
Adds _summarize_api_error() that extracts <title> and Cloudflare Ray ID
from HTML error pages, and falls back to SDK error body messages. Now
the terminal shows clean one-liners like:
📝 Error: HTTP 502 — openrouter.ai | 502: Bad gateway — Ray 9e226...
Also calls _dump_api_request_debug on max_retries_exhausted so the full
request context is written to ~/.hermes/sessions/ for post-mortem.
Made-with: Cursor
When fallback activates (e.g. minimax → OpenRouter), self.provider,
self.base_url, self.api_mode, and self._client_kwargs were all updated
but self.api_key was not. delegate_tool.py reads parent_agent.api_key
to pass credentials to child agents, so subagents inherited the stale
pre-fallback key (e.g. a minimax key sent to OpenRouter), causing 401
Missing Authentication errors.
Add self.api_key = ... in both the anthropic_messages and
chat_completions branches of _try_activate_fallback().
reset_session_state() was missing two fields added after it was written:
- _user_turn_count: kept accumulating across sessions, affecting
flush_min_turns guard behavior
- context_compressor._previous_summary: old session's compression
summary leaked into new session's iterative compression
Cherry-picked from PR #2640 by dusterbloom. Closes#2635.
When an API call fails, the terminal output now includes the HTTP status
code in the header line and, for 400 errors, the response body from the
provider (truncated to 300 chars). Makes it much easier to diagnose
issues like invalid model names or malformed requests that were
previously hidden behind generic error messages.
Salvaged from PR #2646 by Mibayy. Fixes#2644.
When API calls fail with HTML error pages (e.g., CloudFlare errors), the CLI
was dumping raw HTML content to users like:
📝 Error: <!DOCTYPE html><!--[if lt IE 7]> <html class="no-js ie6...
This commit adds a _clean_error_message() utility method that:
- Detects HTML content and replaces with user-friendly message
- Collapses multiline errors to single line
- Truncates overly long errors (>150 chars)
- Preserves meaningful error text for regular errors
Applied to all user-facing error displays:
- API call failure messages (line 6314)
- Interrupt error responses (line 6324)
- Invalid response error messages (line 6000)
Before: 📝 Error: <!DOCTYPE html><!--[if lt IE 7]>...
After: 📝 Error: Service temporarily unavailable (HTML error page returned)
When context overflow triggers compression, the outer retry loop
restarts via continue without incrementing retry_count. If compression
reduces messages but not enough to fit the context window, this creates
an infinite loop burning API credits: API call → overflow → compress →
retry → overflow → compress → ...
Increment retry_count on compression restarts so the loop exits after
max_retries total attempts.
Cherry-picked from PR #2766 by dieutx.
Adds a wall-clock stale stream detector (HERMES_STREAM_STALE_TIMEOUT,
default 90s) that force-closes the httpx client when no real chunks
arrive, even if SSE keep-alive pings keep the socket alive. Works
with the existing streaming retry loop to recover via fresh connection.
Made-with: Cursor
Centralizes two widely-duplicated patterns into hermes_constants.py:
1. get_hermes_home() — Path resolution for ~/.hermes (HERMES_HOME env var)
- Was copy-pasted inline across 30+ files as:
Path(os.getenv("HERMES_HOME", Path.home() / ".hermes"))
- Now defined once in hermes_constants.py (zero-dependency module)
- hermes_cli/config.py re-exports it for backward compatibility
- Removed local wrapper functions in honcho_integration/client.py,
tools/website_policy.py, tools/tirith_security.py, hermes_cli/uninstall.py
2. parse_reasoning_effort() — Reasoning effort string validation
- Was copy-pasted in cli.py, gateway/run.py, cron/scheduler.py
- Same validation logic: check against (xhigh, high, medium, low, minimal, none)
- Now defined once in hermes_constants.py, called from all 3 locations
- Warning log for unknown values kept at call sites (context-specific)
31 files changed, net +31 lines (125 insertions, 94 deletions)
Full test suite: 6179 passed, 0 failed
After streaming retries are exhausted on transient errors, fall back to
non-streaming instead of propagating the error. Also fall back for any
other pre-delivery stream error (not just 'streaming not supported').
Added user-facing message when streaming is not supported by a model/
provider, directing users to set display.streaming: false in config.yaml
to avoid the fallback delay.
Cherry-picked from PR #3008 by kshitijk4poor. Added UX message for
streaming-not-supported detection.
Co-authored-by: kshitijk4poor <kshitijk4poor@users.noreply.github.com>
- Add 'prompt exceeds max length' to context overflow detection for
Z.AI/GLM 400 errors
- Extract inline reasoning blocks from assistant content as fallback
when no structured reasoning fields are present
- Guard inline extraction so structured API reasoning takes priority
- Update test for reasoning-only response salvage behavior
Cherry-picked from PR #2993 by kshitijk4poor. Added priority guard
to fix test_structured_reasoning_takes_priority failure.
Co-authored-by: kshitijk4poor <kshitijk4poor@users.noreply.github.com>
Each subagent now gets its own IterationBudget instead of sharing the
parent's. The per-subagent cap is controlled by delegation.max_iterations
in config.yaml (default 50). Total iterations across parent + subagents
can exceed the parent's max_iterations, but the user retains control via
the config setting.
Previously, subagents shared the parent's budget, so three parallel
subagents configured for max_iterations=50 racing against a parent that
already used 60 of 90 would each only get ~10 iterations.
Inspired by PR #2928 (Bartok9) which identified the issue (#2873).
* fix(session): surface silent SessionDB failures that cause session data loss
SessionDB initialization and operation failures were logged at debug level
or silently swallowed, causing sessions to never be indexed in the FTS5
database. This made session_search unable to find affected conversations.
In practice, ~48% of sessions can be lost without any visible indication.
The JSON session files are still written (separate code path), but the
SQLite/FTS5 index gets nothing — making session_search return empty results
for affected sessions.
Changes:
- cli.py: Log warnings (not debug) when SessionDB init fails at both
__init__ and _start_session entry points
- run_agent.py: Log warnings on create_session, append_message, and
compression split failures
- run_agent.py: Set _session_db = None after create_session failure to
fail fast instead of silently dropping every message for the session
Root cause: When gateway restarts or DB lock contention occurs during
SessionDB() init, the exception is caught and swallowed. The agent
continues running normally — JSON session logs are written to disk —
but no messages reach the FTS5 index.
* fix: use module logger instead of root logging for SessionDB warnings
Follow-up to cherry-picked PR #2939 — the original used logging.warning()
(root logger) instead of logger.warning() (module logger) in the 5 new
warning calls. Module logger preserves the logger hierarchy and shows the
correct module name in log output.
---------
Co-authored-by: LucidPaths <lc77@outlook.de>
feat: persist reasoning across gateway session turns (schema v6)
Tested against OpenAI Codex (direct), Anthropic (direct + OAI-compat), and OpenRouter → 6 backends. All reasoning field types (reasoning, reasoning_details, codex_reasoning_items) round-trip through the DB correctly.
* fix(run_agent): ensure _fire_first_delta() is called for tool generation events
Added calls to _fire_first_delta() in the AIAgent class to improve the handling of tool generation events, ensuring timely notifications during the processing of function calls and tool usage.
* fix(run_agent): improve timeout handling for chat completions
Enhanced the timeout configuration for chat completions in the AIAgent class by introducing customizable connection, read, and write timeouts using environment variables. This ensures more robust handling of API requests during streaming operations.
* fix(run_agent): reduce default stream read timeout for chat completions
Updated the default stream read timeout from 120 seconds to 60 seconds in the AIAgent class, enhancing the timeout configuration for chat completions. This change aims to improve responsiveness during streaming operations.
* fix(run_agent): enhance streaming error handling and retry logic
Improved the error handling and retry mechanism for streaming requests in the AIAgent class. Introduced a configurable maximum number of stream retries and refined the handling of transient network errors, allowing for retries with fresh connections. Non-transient errors now trigger a fallback to non-streaming only when appropriate, ensuring better resilience during API interactions.
* fix: skills-sh install fails for deeply nested repo structures
Skills in repos with deep directory nesting (e.g.
cli-tool/components/skills/development/senior-backend/) could not be
installed because the candidate path generation and shallow root-dir
scan never reached them.
Added GitHubSource._find_skill_in_repo_tree() which uses the GitHub
Trees API to recursively search the entire repo tree in a single API
call. This is used as a final fallback in
SkillsShSource._discover_identifier() when the standard candidate
paths and shallow scan both fail.
Fixes installation of skills from repos like davila7/claude-code-templates
where skills are nested 4+ levels deep.
Reported by user Samuraixheart.
* docs: unify hooks documentation — add plugin hooks to hooks page, add session:end event
The hooks page only documented gateway event hooks (HOOK.yaml system).
The plugins page listed plugin hooks (pre_tool_call, etc.) that weren't
referenced from the hooks page, which was confusing.
Changes:
- hooks.md: Add overview table showing both hook systems
- hooks.md: Add Plugin Hooks section with available hooks, callback
signatures, and example
- hooks.md: Add missing session:end gateway event (emitted but undocumented)
- hooks.md: Mark pre_llm_call, post_llm_call, on_session_start,
on_session_end as planned (defined in VALID_HOOKS but not yet invoked)
- hooks.md: Update info box to cross-reference plugin hooks
- hooks.md: Fix heading hierarchy (gateway content as subsections)
- plugins.md: Add cross-reference to hooks page for full details
- plugins.md: Mark planned hooks as (planned)
* feat(session_search): add recent sessions mode when query is omitted
When session_search is called without a query (or with an empty query),
it now returns metadata for the most recent sessions instead of erroring.
This lets the agent quickly see what was worked on recently without
needing specific keywords.
Returns for each session: session_id, title, source, started_at,
last_active, message_count, preview (first user message).
Zero LLM cost — pure DB query. Current session lineage and child
delegation sessions are excluded.
The agent can then keyword-search specific sessions if it needs
deeper context from any of them.
* docs: clarify two-mode behavior in session_search schema description
* fix(compression): restore sane defaults and cap summary at 12K tokens
- threshold: 0.80 → 0.50 (compress at 50%, not 80%)
- target_ratio: 0.40 → 0.20, now relative to threshold not total context
(20% of 50% = 10% of context as tail budget)
- summary ceiling: 32K → 12K (Gemini can't output more than ~12K)
- Updated DEFAULT_CONFIG, config display, example config, and tests
* fix: browser_vision ignores auxiliary.vision.timeout config (#2901)
* docs: unify hooks documentation — add plugin hooks to hooks page, add session:end event
The hooks page only documented gateway event hooks (HOOK.yaml system).
The plugins page listed plugin hooks (pre_tool_call, etc.) that weren't
referenced from the hooks page, which was confusing.
Changes:
- hooks.md: Add overview table showing both hook systems
- hooks.md: Add Plugin Hooks section with available hooks, callback
signatures, and example
- hooks.md: Add missing session:end gateway event (emitted but undocumented)
- hooks.md: Mark pre_llm_call, post_llm_call, on_session_start,
on_session_end as planned (defined in VALID_HOOKS but not yet invoked)
- hooks.md: Update info box to cross-reference plugin hooks
- hooks.md: Fix heading hierarchy (gateway content as subsections)
- plugins.md: Add cross-reference to hooks page for full details
- plugins.md: Mark planned hooks as (planned)
* fix: browser_vision ignores auxiliary.vision.timeout config
browser_vision called call_llm() without passing a timeout parameter,
so it always used the 30-second default in auxiliary_client.py. This
made vision analysis with local models (llama.cpp, ollama) impossible
since they typically need more than 30s for screenshot analysis.
Now browser_vision reads auxiliary.vision.timeout from config.yaml
(same config key that vision_analyze already uses) and passes it
through to call_llm().
Also bumped the default vision timeout from 30s to 120s in both
browser_vision and vision_analyze — 30s is too aggressive for local
models and the previous default silently failed for anyone running
vision locally.
Fixes user report from GamerGB1988.
* fix(skills): agent-created skills were incorrectly treated as untrusted community content
_resolve_trust_level() didn't handle 'agent-created' source, so it
fell through to 'community' trust level. Community policy blocks on
any caution or dangerous findings, which meant common patterns like
curl with env vars, systemctl, crontab, cloudflared references etc.
would block skill creation/patching.
The agent-created policy row already existed in INSTALL_POLICY with
permissive settings (allow caution, ask on dangerous) but was never
reached. Now it is.
Fixes reports of skill_manage being blocked by security scanner.
* fix(cli): enhance real-time reasoning output by forcing flush of long partial lines
Updated the reasoning output mechanism to emit complete lines and force-flush long partial lines, ensuring reasoning is visible in real-time even without newlines. This improves user experience during reasoning sessions.
* fix: skip KawaiiSpinner when TUI handles tool progress
In the interactive CLI, the agent runs with quiet_mode=True and
tool_progress_callback set. The quiet_mode condition triggered
KawaiiSpinner for every tool call, but the TUI was already handling
progress display via the spinner widget.
The KawaiiSpinner writes carriage-return animation through StdoutProxy,
triggering run_in_terminal() erase/redraw cycles on every flush. These
redundant cycles cause the status bar to ghost into terminal scrollback.
The thinking spinner already had this guard (checks thinking_callback).
This extends the same pattern to the three tool spinner creation sites:
concurrent tools, delegate_task, and single tool execution.
- threshold: 0.80 → 0.50 (compress at 50%, not 80%)
- target_ratio: 0.40 → 0.20, now relative to threshold not total context
(20% of 50% = 10% of context as tail budget)
- summary ceiling: 32K → 12K (Gemini can't output more than ~12K)
- Updated DEFAULT_CONFIG, config display, example config, and tests
The summary_target_tokens parameter was accepted in the constructor,
stored on the instance, and never used — the summary budget was always
computed from hardcoded module constants (_SUMMARY_RATIO=0.20,
_MAX_SUMMARY_TOKENS=8000). This caused two compounding problems:
1. The config value was silently ignored, giving users no control
over post-compression size.
2. Fixed budgets (20K tail, 8K summary cap) didn't scale with
context window size. Switching from a 1M-context model to a
200K model would trigger compression that nuked 350K tokens
of conversation history down to ~30K.
Changes:
- Replace summary_target_tokens with summary_target_ratio (default 0.40)
which sets the post-compression target as a fraction of context_length.
Tail token budget and summary cap now scale proportionally:
MiniMax 200K → ~80K post-compression
GPT-5 1M → ~400K post-compression
- Change threshold_percent default: 0.50 → 0.80 (don't fire until
80% of context is consumed)
- Change protect_last_n default: 4 → 20 (preserve ~10 full turns)
- Summary token cap scales to 5% of context (was fixed 8K), capped
at 32K ceiling
- Read target_ratio and protect_last_n from config.yaml compression
section (both are now configurable)
- Remove hardcoded summary_target_tokens=500 from run_agent.py
- Add 5 new tests for ratio scaling, clamping, and new defaults
The gateway process runs from the hermes-agent install directory, so
os.getcwd() picks up the repo's AGENTS.md (16k chars) and other dev
context files — inflating input tokens by ~10k on every gateway message.
Fix: use TERMINAL_CWD (which the gateway sets to MESSAGING_CWD or
$HOME) as the cwd for build_context_files_prompt(). In CLI mode,
TERMINAL_CWD is the user's actual project directory, so behavior
is unchanged.
Before: gateway 15-20k input tokens, CLI 6-8k
After: gateway ~6-8k input tokens (same as CLI)
Reported by keri on Discord.
Move OpenRouter to position 1 in the setup wizard's provider list
to match hermes model ordering. Update default selection index and
fix test expectations for the new ordering.
Setup order: OpenRouter → Nous Portal → Codex → Custom → ...
Reset context pressure warnings and update last_prompt_tokens and last_completion_tokens in the context compressor to prevent stale values from causing excessive warnings and re-triggering compression. This change ensures accurate pressure calculations following the compaction process.
Complete cleanup after dropping the mini-swe-agent submodule (PR #2804):
- Remove MSWEA_SILENT_STARTUP and MSWEA_GLOBAL_CONFIG_DIR env var
settings from cli.py, run_agent.py, hermes_cli/main.py, doctor.py
- Remove mini-swe-agent health check from hermes doctor
- Remove 'minisweagent' from logger suppression lists
- Remove litellm/typer/platformdirs from requirements.txt
- Remove mini-swe-agent install steps from install.ps1 (Windows)
- Remove mini-swe-agent install steps from website docs
- Update all stale comments/docstrings referencing mini-swe-agent
in terminal_tool.py, tools/__init__.py, code_execution_tool.py,
environments/README.md, environments/agent_loop.py
- Remove mini_swe_runner from pyproject.toml py-modules
(still exists as standalone script for RL training use)
- Shrink test_minisweagent_path.py to empty stub
The orphaned mini-swe-agent/ directory on disk needs manual removal:
rm -rf mini-swe-agent/
Drop the mini-swe-agent git submodule. All terminal backends now use
hermes-agent's own environment implementations directly.
Docker backend:
- Inline the `docker run -d` container startup (was 15 lines in
minisweagent's DockerEnvironment). Our wrapper already handled
execute(), cleanup(), security hardening, volumes, and resource limits.
Modal backend:
- Import swe-rex's ModalDeployment directly instead of going through
minisweagent's 90-line passthrough wrapper.
- Bake the _AsyncWorker pattern (from environments/patches.py) directly
into ModalEnvironment for Atropos compatibility without monkey-patching.
Cleanup:
- Remove minisweagent_path.py (submodule path resolution helper)
- Remove submodule init/install from install.sh and setup-hermes.sh
- Remove mini-swe-agent from .gitmodules
- environments/patches.py is now a no-op (kept for backward compat)
- terminal_tool.py no longer does sys.path hacking for minisweagent
- mini_swe_runner.py guards imports (optional, for RL training only)
- Update all affected tests to mock the new direct subprocess calls
- Update README.md, CONTRIBUTING.md
No functionality change — all Docker, Modal, local, SSH, Singularity,
and Daytona backends behave identically. 6093 tests pass.
- Added calls to `_fire_first_delta()` in the `AIAgent` class to ensure that the first delta is triggered for both reasoning and thinking updates. This change improves the handling of delta events during streaming, enhancing the responsiveness of the agent's reasoning capabilities.
- Introduced `_on_tool_gen_start` in `HermesCLI` to indicate when tool-call arguments are being generated, enhancing user feedback during streaming.
- Updated `AIAgent` to support a new `tool_gen_callback`, notifying the display layer when tool generation starts, allowing for better user experience during large payloads.
- Ensured that the callback is triggered appropriately during streaming events to prevent user interface freezing.
Added explicit closing of the OpenAI/httpx client in the background review process to prevent "Event loop is closed" errors. This change ensures that the client is properly cleaned up when the review agent is no longer needed, enhancing stability and resource management.
Follow-up to 669c60a6 (cherry-pick of PR #2187, fixes#2177).
The original fix emits a "\n\n" delta immediately after every
_execute_tool_calls() invocation. When the model runs multiple
consecutive tool iterations before producing text (common with
search → read → analyze flows), each iteration appends its own
paragraph break, resulting in 4-6+ blank lines before the actual
response.
Replace the immediate delta with a deferred flag
(_stream_needs_break). _fire_stream_delta() checks the flag and
prepends a single "\n\n" only when the first real text delta
arrives, so multiple back-to-back tool iterations still produce
exactly one paragraph break.
Reverts the sanitizer addition from PR #2466 (originally #2129).
We already have _empty_content_retries handling for reasoning-only
responses. The trailing strip risks silently eating valid messages
and is redundant with existing empty-content handling.
When subagents run in ThreadPoolExecutor threads, the shared stdout handle
can close between thread teardown and KawaiiSpinner cleanup. Python raises
ValueError (not OSError) for I/O operations on closed files:
ValueError: I/O operation on closed file
The _SafeWriter class was only catching OSError, missing this case.
Changes:
- Add ValueError to exception handling in write(), flush(), and isatty()
- Update docstring to document the ThreadPoolExecutor teardown scenario
Fixes#2428
Two related root causes for the '?[33mTool progress: NEW?[0m' garbling
reported on kitty, alacritty, ghostty and gnome-console:
1. /verbose label printing used self.console.print() with Rich markup
([yellow]...[/]). self.console is a plain Rich Console() whose output
goes directly to sys.stdout, which patch_stdout's StdoutProxy
intercepts and mangles raw ANSI sequences.
2. Context pressure status lines (e.g. 'approaching compaction') from
AIAgent._safe_print() had the same problem -- _safe_print() was a
@staticmethod that always called builtin print(), bypassing the
prompt_toolkit renderer entirely.
Fix:
- Convert AIAgent._safe_print() from @staticmethod to an instance method
that delegates to self._print_fn (defaults to builtin print, preserving
all non-CLI behaviour).
- After the CLI creates its AIAgent instance, wire self.agent._print_fn to
the existing _cprint() helper which routes through
prompt_toolkit.print_formatted_text(ANSI(text)).
- Rewrite the /verbose feedback labels to use hermes_cli.colors.Colors
ANSI constants in f-strings and emit them via _cprint() directly,
removing the Rich-markup-inside-patch_stdout anti-pattern.
Fixes#2262
Co-authored-by: Animesh Mishra <animesh.m.7523@gmail.com>
When a non-OpenRouter provider (e.g. minimax, anthropic) is set in
config.yaml but its API key is missing, Hermes silently fell back to
OpenRouter, causing confusing 404 errors.
Now checks if the user explicitly configured a provider before falling
back. Explicit providers raise RuntimeError with a clear message naming
the missing env var. Auto/openrouter/custom providers still fall through
to OpenRouter as before.
Three code paths fixed:
- run_agent.py AIAgent.__init__ — main client initialization
- auxiliary_client.py call_llm — sync auxiliary calls
- auxiliary_client.py call_llm_streaming — async auxiliary calls
Based on PR #2272 by @StefanIsMe. Applied manually to fix a
pconfig NameError in the original and extend to call_llm_streaming.
Co-authored-by: StefanIsMe <StefanIsMe@users.noreply.github.com>
Cherry-picked from PR #2187 by @devorun. Fixes#2177.
When streaming is enabled, text before and after tool calls gets
concatenated without separation. Adds a paragraph break delta after
_execute_tool_calls() so stream consumers insert proper whitespace
between iteration boundaries.
On the native Anthropic Messages API path, convert_messages_to_anthropic()
moves top-level cache_control on role:tool messages inside the tool_result
block. On OpenRouter (chat_completions), no such conversion happens — the
unexpected top-level field causes a silent hang on the second tool call.
Add native_anthropic parameter to _apply_cache_marker() and
apply_anthropic_cache_control(). When False (OpenRouter), role:tool messages
are skipped entirely. When True (native Anthropic), existing behaviour is
preserved.
Fixes#2362
* fix: prevent Anthropic token fallback leaking to third-party anthropic_messages providers
When provider is minimax/alibaba/etc and MINIMAX_API_KEY is not set,
the code fell back to resolve_anthropic_token() sending Anthropic OAuth
credentials to third-party endpoints, causing 401 errors.
Now only provider=="anthropic" triggers the fallback. Generalizes the
Alibaba-specific guard from #1739 to all non-Anthropic providers.
* fix: set provider='anthropic' in credential refresh tests
Follow-up for cherry-picked PR #2383 — existing tests didn't set
agent.provider, which the new guard requires to allow Anthropic
token refresh.
---------
Co-authored-by: 0xbyt4 <35742124+0xbyt4@users.noreply.github.com>
PR #2314 checked for provider names 'alibaba-coding-plan' and
'alibaba-coding-plan-anthropic' which don't exist in the provider
registry. The provider is always 'alibaba' — the condition was dead
code. Fixed to check self.provider == 'alibaba'.
When using Alibaba (DashScope) with an anthropic-compatible endpoint,
model names like qwen3.5-plus were being normalized to qwen3-5-plus.
Alibaba's API expects the dot. Added preserve_dots parameter to
normalize_model_name() and build_anthropic_kwargs().
Also fixed 401 auth: when provider is alibaba or base_url contains
dashscope/aliyuncs, use only the resolved API key (DASHSCOPE_API_KEY).
Never fall back to resolve_anthropic_token(), and skip Anthropic
credential refresh for DashScope endpoints.
Cherry-picked from PR #1748 by crazywriter1. Fixes#1739.
- Add resolve_config_path(): checks $HERMES_HOME/honcho.json first,
falls back to ~/.honcho/config.json. Enables isolated Hermes instances
with independent Honcho credentials and settings.
- Update CLI and doctor to use resolved path instead of hardcoded global.
- Change default session_strategy from per-session to per-directory.
Part 1 of #1962 by @erosika.
Two changes to the error handler in the agent loop:
1. Remove the 'if not pending_handled' block that injected fake
[System error during processing: ...] messages into conversation
history. These polluted history, burned tokens on retries, and
could violate role alternation by injecting as role=user.
The tool_calls error-result path (role=tool) is preserved.
2. Append the error final_response as an assistant message when
hitting the iteration limit, so session resume doesn't produce
consecutive user messages.
Enhanced the review agent to scan and summarize successful tool actions, providing users with a compact overview of updates made during the review process. This includes actions related to memory and user profiles, improving user feedback and interaction clarity.
- Introduced a mechanism to mute output after the main response is delivered, ensuring that subsequent tool calls run without cluttering the CLI.
- Redirected stdout to devnull during the review agent's execution to prevent any print statements from interfering with the main CLI display.
- Added a new attribute `_mute_post_response` to manage output suppression effectively.
When an API call fails, the error output now shows the provider name,
model, and endpoint URL so users can immediately identify which service
rejected their request. Auth errors (401/403) get actionable guidance:
check key validity, model access, and OpenRouter credits link.
Before: 'API call failed (attempt 1/3): PermissionDeniedError'
After: 'API call failed (attempt 1/3): PermissionDeniedError
Provider: openrouter Model: anthropic/claude-sonnet-4
Endpoint: https://openrouter.ai/api/v1
Your API key was rejected by the provider. Check:
• Is the key valid? Run: hermes setup
• Does your account have access to anthropic/claude-sonnet-4?
• Check credits: https://openrouter.ai/settings/credits'
Remove the memory and skill nudges that were appended directly to user
messages, causing backward-looking system instructions to compete with
forward-looking user tasks. Found in 43% of user messages across 15
sessions, with confirmed cases of the agent spending tool calls on
nudge responses before starting the user's actual request.
Replace with a background review agent that runs AFTER the main agent
finishes responding:
- Spawns a background thread with a snapshot of the conversation
- Uses the main model (not auxiliary) for high-precision memory/skill work
- Only has memory + skill_manage tools (5 iteration budget)
- Shares the memory store for direct writes
- Never modifies the main conversation history
- Never competes with the user's task for model attention
- Zero latency impact (runs after response is delivered)
- Same token cost (processes the same context, just on a separate track)
The trigger conditions are unchanged (every 10 user turns for memory,
after 10+ tool iterations for skills). Only the execution path changes:
from inline injection to background fork.
Closes#2227.
Co-authored-by: Test <test@test.com>
Remove the [Files already read — do NOT re-read these] user message
that was injected into the conversation after context compression.
This message used role='user' for system-generated content, creating
a fake user turn that confused models about conversation state and
could contribute to task-redo behavior.
The file_tools.py read tracker (warn on 3rd consecutive read, block
on 4th+) already handles re-read prevention inline without injecting
synthetic messages.
Closes#2224.
Co-authored-by: Test <test@test.com>
- Updated _stream_delta method in HermesCLI to handle None values, flushing the stream and resetting state for clean tool execution.
- Enhanced quiet mode handling in AIAgent to ensure proper display closure before tool execution, preventing display issues with intermediate streamed content.
These changes improve the robustness of the streaming functionality and ensure a smoother user experience during tool interactions.
Cherry-picked from PR #2169 by @0xbyt4.
1. _strip_provider_prefix: skip Ollama model:tag names (qwen:0.5b)
2. Fuzzy match: remove reverse direction that made claude-sonnet-4
resolve to 1M instead of 200K
3. _has_content_after_think_block: reuse _strip_think_blocks() to
handle all tag variants (thinking, reasoning, REASONING_SCRATCHPAD)
4. models.dev lookup: elif→if so nous provider also queries models.dev
5. Disk cache fallback: use 5-min TTL instead of full hour so network
is retried soon
6. Delegate build: wrap child construction in try/finally so
_last_resolved_tool_names is always restored on exception
* feat: context pressure warnings for CLI and gateway
User-facing notifications as context approaches the compaction threshold.
Warnings fire at 60% and 85% of the way to compaction — relative to
the configured compression threshold, not the raw context window.
CLI: Formatted line with a progress bar showing distance to compaction.
Cyan at 60% (approaching), bold yellow at 85% (imminent).
◐ context ▰▰▰▰▰▰▰▰▰▰▰▰▱▱▱▱▱▱▱▱ 60% to compaction 100k threshold (50%) · approaching compaction
⚠ context ▰▰▰▰▰▰▰▰▰▰▰▰▰▰▰▰▰▱▱▱ 85% to compaction 100k threshold (50%) · compaction imminent
Gateway: Plain-text notification sent to the user's chat via the new
status_callback mechanism (asyncio.run_coroutine_threadsafe bridge,
same pattern as step_callback).
Does NOT inject into the message stream. The LLM never sees these
warnings. Flags reset after each compaction cycle.
Files changed:
- agent/display.py — format_context_pressure(), format_context_pressure_gateway()
- run_agent.py — status_callback param, _context_50/70_warned flags,
_emit_context_pressure(), flag reset in _compress_context()
- gateway/run.py — _status_callback_sync bridge, wired to AIAgent
- tests/test_context_pressure.py — 23 tests
* Merge remote-tracking branch 'origin/main' into hermes/hermes-7ea545bf
---------
Co-authored-by: Test <test@test.com>
