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.
