* 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>
Replace the fragile hardcoded context length system with a multi-source
resolution chain that correctly identifies context windows per provider.
Key changes:
- New agent/models_dev.py: Fetches and caches the models.dev registry
(3800+ models across 100+ providers with per-provider context windows).
In-memory cache (1hr TTL) + disk cache for cold starts.
- Rewritten get_model_context_length() resolution chain:
0. Config override (model.context_length)
1. Custom providers per-model context_length
2. Persistent disk cache
3. Endpoint /models (local servers)
4. Anthropic /v1/models API (max_input_tokens, API-key only)
5. OpenRouter live API (existing, unchanged)
6. Nous suffix-match via OpenRouter (dot/dash normalization)
7. models.dev registry lookup (provider-aware)
8. Thin hardcoded defaults (broad family patterns)
9. 128K fallback (was 2M)
- Provider-aware context: same model now correctly resolves to different
context windows per provider (e.g. claude-opus-4.6: 1M on Anthropic,
128K on GitHub Copilot). Provider name flows through ContextCompressor.
- DEFAULT_CONTEXT_LENGTHS shrunk from 80+ entries to ~16 broad patterns.
models.dev replaces the per-model hardcoding.
- CONTEXT_PROBE_TIERS changed from [2M, 1M, 512K, 200K, 128K, 64K, 32K]
to [128K, 64K, 32K, 16K, 8K]. Unknown models no longer start at 2M.
- hermes model: prompts for context_length when configuring custom
endpoints. Supports shorthand (32k, 128K). Saved to custom_providers
per-model config.
- custom_providers schema extended with optional models dict for
per-model context_length (backward compatible).
- Nous Portal: suffix-matches bare IDs (claude-opus-4-6) against
OpenRouter's prefixed IDs (anthropic/claude-opus-4.6) with dot/dash
normalization. Handles all 15 current Nous models.
- Anthropic direct: queries /v1/models for max_input_tokens. Only works
with regular API keys (sk-ant-api*), not OAuth tokens. Falls through
to models.dev for OAuth users.
Tests: 5574 passed (18 new tests for models_dev + updated probe tiers)
Docs: Updated configuration.md context length section, AGENTS.md
Co-authored-by: Test <test@test.com>
When streaming was enabled, two visual feedback mechanisms were
completely suppressed:
1. The thinking spinner (TUI toolbar) was skipped because the entire
spinner block was gated on 'not self._has_stream_consumers()'.
Now the thinking_callback fires in streaming mode too — the
raw KawaiiSpinner is still skipped (would conflict with streamed
tokens) but the TUI toolbar widget works fine alongside streaming.
2. Tool progress lines (the ┊ feed) were invisible because _vprint
was blanket-suppressed when stream consumers existed. But during
tool execution, no tokens are actively streaming, so printing is
safe. Added an _executing_tools flag that _vprint respects to
allow output during tool execution even with stream consumers
registered.
Based on PR #1859 by @magi-morph (too stale to cherry-pick, reimplemented).
GPT-5.x models reject tool calls + reasoning_effort on
/v1/chat/completions with a 400 error directing to /v1/responses.
This auto-detects api.openai.com in the base URL and switches to
codex_responses mode in three places:
- AIAgent.__init__: upgrades chat_completions → codex_responses
- _try_activate_fallback(): same routing for fallback model
- runtime_provider.py: _detect_api_mode_for_url() for both custom
provider and openrouter runtime resolution paths
Also extracts _is_direct_openai_url() helper to replace the inline
check in _max_tokens_param().
Follow-up to PR #2101 (InB4DevOps). Adds three missing context compressor
resets in reset_session_state():
- compression_count (displayed in status bar)
- last_total_tokens
- _context_probed (stale context-error flag)
Also fixes the test_cli_new_session.py prompt_toolkit mock (missing
auto_suggest stub) and adds a regression test for #2099 that verifies
all token counters and compressor state are zeroed on /new.
- Add <thinking> tag to streaming filter's tag list
- When show_reasoning is on, route XML reasoning content to the
reasoning display box instead of silently discarding it
- Expand _strip_think_blocks to handle all tag variants:
<think>, <thinking>, <THINKING>, <reasoning>, <REASONING_SCRATCHPAD>
Local models (especially Qwen 3.5) sometimes wrap their entire response
inside <think> tags, leaving actual content empty. Previously this caused
3 retries and then an error, wasting tokens and failing the request.
Now when retries are exhausted and reasoning_text contains the response,
it is used as final_response instead of returning an error. The user
sees the actual answer instead of "Model generated only think blocks."
* fix(codex): treat reasoning-only responses as incomplete, not stop
When a Codex Responses API response contains only reasoning items
(encrypted thinking state) with no message text or tool calls, the
_normalize_codex_response method was setting finish_reason='stop'.
This sent the response into the empty-content retry loop, which
burned 3 retries and then failed — exactly the pattern Nester
reported in Discord.
Two fixes:
1. _normalize_codex_response: reasoning-only responses (reasoning_items_raw
non-empty but no final_text) now get finish_reason='incomplete', routing
them to the Codex continuation path instead of the retry loop.
2. Incomplete handling: also checks for codex_reasoning_items when deciding
whether to preserve an interim message, so encrypted reasoning state is
not silently dropped when there is no visible reasoning text.
Adds 4 regression tests covering:
- Unit: reasoning-only → incomplete, reasoning+content → stop
- E2E: reasoning-only → continuation → final answer succeeds
- E2E: encrypted reasoning items preserved in interim messages
* fix(codex): ensure reasoning items have required following item in API input
Follow-up to the reasoning-only response fix. Three additional issues
found by tracing the full replay path:
1. _chat_messages_to_responses_input: when a reasoning-only interim
message was converted to Responses API input, the reasoning items
were emitted as the last items with no following item. The Responses
API requires a following item after each reasoning item (otherwise:
'missing_following_item' error, as seen in OpenHands #11406). Now
emits an empty assistant message as the required following item when
content is empty but reasoning items were added.
2. Duplicate detection: two consecutive reasoning-only incomplete
messages with identical empty content/reasoning but different
encrypted codex_reasoning_items were incorrectly treated as
duplicates, silently dropping the second response's reasoning state.
Now includes codex_reasoning_items in the duplicate comparison.
3. Added tests for both the API input conversion path and the duplicate
detection edge case.
Research context: verified against OpenCode (uses Vercel AI SDK, no
retry loop so avoids the issue), Clawdbot (drops orphaned reasoning
blocks entirely), and OpenHands (hit the missing_following_item error).
Our approach preserves reasoning continuity while satisfying the API
constraint.
---------
Co-authored-by: Test <test@test.com>
* fix: detect context length for custom model endpoints via fuzzy matching + config override
Custom model endpoints (non-OpenRouter, non-known-provider) were silently
falling back to 2M tokens when the model name didn't exactly match what the
endpoint's /v1/models reported. This happened because:
1. Endpoint metadata lookup used exact match only — model name mismatches
(e.g. 'qwen3.5:9b' vs 'Qwen3.5-9B-Q4_K_M.gguf') caused a miss
2. Single-model servers (common for local inference) required exact name
match even though only one model was loaded
3. No user escape hatch to manually set context length
Changes:
- Add fuzzy matching for endpoint model metadata: single-model servers
use the only available model regardless of name; multi-model servers
try substring matching in both directions
- Add model.context_length config override (highest priority) so users
can explicitly set their model's context length in config.yaml
- Log an informative message when falling back to 2M probe, telling
users about the config override option
- Thread config_context_length through ContextCompressor and AIAgent init
Tests: 6 new tests covering fuzzy match, single-model fallback, config
override (including zero/None edge cases).
* fix: auto-detect local model name and context length for local servers
Cherry-picked from PR #2043 by sudoingX.
- Auto-detect model name from local server's /v1/models when only one
model is loaded (no manual model name config needed)
- Add n_ctx_train and n_ctx to context length detection keys for llama.cpp
- Query llama.cpp /props endpoint for actual allocated context (not just
training context from GGUF metadata)
- Strip .gguf suffix from display in banner and status bar
- _auto_detect_local_model() in runtime_provider.py for CLI init
Co-authored-by: sudo <sudoingx@users.noreply.github.com>
* fix: revert accidental summary_target_tokens change + add docs for context_length config
- Revert summary_target_tokens from 2500 back to 500 (accidental change
during patching)
- Add 'Context Length Detection' section to Custom & Self-Hosted docs
explaining model.context_length config override
---------
Co-authored-by: Test <test@test.com>
Co-authored-by: sudo <sudoingx@users.noreply.github.com>
Three bugs prevented providers like MiniMax from using their
Anthropic-compatible endpoints (e.g. api.minimax.io/anthropic):
1. _VALID_API_MODES was missing 'anthropic_messages', so explicit
api_mode config was silently rejected and defaulted to
chat_completions.
2. API-key provider resolution hardcoded api_mode to 'chat_completions'
without checking model config or detecting Anthropic-compatible URLs.
3. run_agent.py auto-detection only recognized api.anthropic.com, not
third-party endpoints using the /anthropic URL convention.
Fixes:
- Add 'anthropic_messages' to _VALID_API_MODES
- API-key providers now check model config api_mode and auto-detect
URLs ending in /anthropic
- run_agent.py and fallback logic detect /anthropic URL convention
- 5 new tests covering all scenarios
Users can now either:
- Set MINIMAX_BASE_URL=https://api.minimax.io/anthropic (auto-detected)
- Set api_mode: anthropic_messages in model config (explicit)
- Use custom_providers with api_mode: anthropic_messages
Co-authored-by: Test <test@test.com>
Adds model name and provider to the system prompt metadata block,
alongside the existing session ID and timestamp. These are frozen
at session start and don't change mid-conversation, so they won't
break prompt caching.
SOUL.md now loads in slot #1 of the system prompt, replacing the
hardcoded DEFAULT_AGENT_IDENTITY. This lets users fully customize
the agent's identity and personality by editing ~/.hermes/SOUL.md
without it conflicting with the built-in identity text.
When SOUL.md is loaded as identity, it's excluded from the context
files section to avoid appearing twice. When SOUL.md is missing,
empty, unreadable, or skip_context_files is set, the hardcoded
DEFAULT_AGENT_IDENTITY is used as a fallback.
The default SOUL.md (seeded on first run) already contains the full
Hermes personality, so existing installs are unaffected.
Co-authored-by: Test <test@test.com>
* Improve tool batching independence checks
* fix: address review feedback on path-aware batching
- Log malformed/non-dict tool arguments at debug level before
falling back to sequential, instead of silently swallowing
the error into an empty dict
- Guard empty paths in _paths_overlap (unreachable in practice
due to upstream filtering, but makes the invariant explicit)
- Add tests: malformed JSON args, non-dict args, _paths_overlap
unit tests including empty path edge cases
- web_crawl is not a registered tool (only web_search/web_extract
are); no addition needed to _PARALLEL_SAFE_TOOLS
---------
Co-authored-by: kshitij <82637225+kshitijk4poor@users.noreply.github.com>
* perf: cache base_url.lower() via property, consolidate triple load_config(), hoist set constant
run_agent.py:
- Add base_url property that auto-caches _base_url_lower on every
assignment, eliminating 12+ redundant .lower() calls per API cycle
across __init__, _build_api_kwargs, _supports_reasoning_extra_body,
and the main conversation loop
- Consolidate three separate load_config() disk reads in __init__
(memory, skills, compression) into a single call, reusing the
result dict for all three config sections
model_tools.py:
- Hoist _READ_SEARCH_TOOLS set to module level (was rebuilt inside
handle_function_call on every tool invocation)
* Use endpoint metadata for custom model context and pricing
---------
Co-authored-by: kshitij <82637225+kshitijk4poor@users.noreply.github.com>
- Update _is_anthropic_oauth in _try_refresh_anthropic_client_credentials()
when token type changes during credential refresh
- Set _is_anthropic_oauth in _try_activate_fallback() Anthropic path
- Move _turns_since_memory and _iters_since_skill init to __init__ so
nudge counters accumulate across run_conversation() calls in CLI mode
- Remove unreachable retry_count >= max_retries block after raise
Adds 7 regression tests. Salvaged from PR #1797 by @0xbyt4.
Add first-class GitHub Copilot and Copilot ACP provider support across
model selection, runtime provider resolution, CLI sessions, delegated
subagents, cron jobs, and the Telegram gateway.
This also normalizes Copilot model catalogs and API modes, introduces a
Copilot ACP OpenAI-compatible shim, and fixes service-mode auth by
resolving Homebrew-installed gh binaries under launchd.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
- Add summary_base_url config option to compression block for custom
OpenAI-compatible endpoints (e.g. zai, DeepSeek, Ollama)
- Remove compression env var bridges from cli.py and gateway/run.py
(CONTEXT_COMPRESSION_* env vars no longer set from config)
- Switch run_agent.py to read compression config directly from
config.yaml instead of env vars
- Fix backwards-compat block in _resolve_task_provider_model to also
fire when auxiliary.compression.provider is 'auto' (DEFAULT_CONFIG
sets this, which was silently preventing the compression section's
summary_* keys from being read)
- Add test for summary_base_url config-to-client flow
- Update docs to show compression as config.yaml-only
Closes#1591
Based on PR #1702 by @uzaylisak
Salvage of PR #1321 by @alireza78a (cherry-picked concept, reimplemented
against current main).
Phase 1 — Pre-call message sanitization:
_sanitize_api_messages() now runs unconditionally before every LLM call.
Previously gated on context_compressor being present, so sessions loaded
from disk or running without compression could accumulate dangling
tool_call/tool_result pairs causing API errors.
Phase 2a — Delegate task cap:
_cap_delegate_task_calls() truncates excess delegate_task calls per turn
to MAX_CONCURRENT_CHILDREN. The existing cap in delegate_tool.py only
limits the task array within a single call; this catches multiple
separate delegate_task tool_calls in one turn.
Phase 2b — Tool call deduplication:
_deduplicate_tool_calls() drops duplicate (tool_name, arguments) pairs
within a single turn when models stutter.
All three are static methods on AIAgent, independently testable.
29 tests covering happy paths and edge cases.
When a fallback model is configured, switch to it immediately upon
detecting rate-limit conditions (429, quota exhaustion, empty/malformed
responses) instead of exhausting all retries with exponential backoff.
Two eager-fallback checks:
1. Invalid/empty API responses — fallback attempted before retry loop
2. HTTP 429 / rate-limit keyword detection — fallback before backoff
Both guarded by _fallback_activated for one-shot semantics.
Cherry-picked from PR #1413 by usvimal.
Co-authored-by: usvimal <usvimal@users.noreply.github.com>
compression_attempts was initialized inside the outer while loop,
resetting to 0 on every iteration. Since compression triggers a
'continue' back to the top of the loop, the counter never accumulated
past 1 — effectively allowing unlimited compression attempts.
Move initialization before the outer while loop so the cap of 3
applies across the entire run_conversation() call.
Two edge cases could inject messages that violate role alternation:
1. Invalid JSON recovery (line ~5985): After 3 retries of invalid JSON
tool args, a user-role recovery message was injected. But the
assistant's tool_calls were never appended, so the sequence could
become user → user. Fix: append the assistant message with its
tool_calls, then respond with proper tool-role error results.
2. System error handler (line ~6238): Always injected a user-role
error message, which creates consecutive user messages if the last
message was already user. Fix: dynamically choose the role based on
the last message to maintain alternation.
length_continue_retries and truncated_response_prefix were initialized
once before the outer loop and never reset after a successful
continuation. If a conversation hit length truncation once (counter=1),
succeeded on continuation, did more tool calls, then hit length again,
the counter started at 1 instead of 0 — reducing available retries
from 3 to 2. The stale truncated_response_prefix would also leak
into the next response.
Reset both after the prefix is consumed on a successful final response.
RedactingFormatter was imported inside 'if not has_errors_log_handler:'
(line 461) but also used unconditionally in the verbose_logging block
(line 479). When the error log handler already exists (e.g. second
AIAgent in the same process) AND verbose_logging=True, the import was
skipped and line 479 raised NameError.
Fix: Move the import one level up so it's always available regardless
of whether the error log handler already exists.
* fix: thread safety for concurrent subagent delegation
Four thread-safety fixes that prevent crashes and data races when
running multiple subagents concurrently via delegate_task:
1. Remove redirect_stdout/stderr from delegate_tool — mutating global
sys.stdout races with the spinner thread when multiple children start
concurrently, causing segfaults. Children already run with
quiet_mode=True so the redirect was redundant.
2. Split _run_single_child into _build_child_agent (main thread) +
_run_single_child (worker thread). AIAgent construction creates
httpx/SSL clients which are not thread-safe to initialize
concurrently.
3. Add threading.Lock to SessionDB — subagents share the parent's
SessionDB and call create_session/append_message from worker threads
with no synchronization.
4. Add _active_children_lock to AIAgent — interrupt() iterates
_active_children while worker threads append/remove children.
5. Add _client_cache_lock to auxiliary_client — multiple subagent
threads may resolve clients concurrently via call_llm().
Based on PR #1471 by peteromallet.
* feat: Honcho base_url override via config.yaml + quick command alias type
Two features salvaged from PR #1576:
1. Honcho base_url override: allows pointing Hermes at a remote
self-hosted Honcho deployment via config.yaml:
honcho:
base_url: "http://192.168.x.x:8000"
When set, this overrides the Honcho SDK's environment mapping
(production/local), enabling LAN/VPN Honcho deployments without
requiring the server to live on localhost. Uses config.yaml instead
of env var (HONCHO_URL) per project convention.
2. Quick command alias type: adds a new 'alias' quick command type
that rewrites to another slash command before normal dispatch:
quick_commands:
sc:
type: alias
target: /context
Supports both CLI and gateway. Arguments are forwarded to the
target command.
Based on PR #1576 by redhelix.
---------
Co-authored-by: peteromallet <peteromallet@users.noreply.github.com>
Co-authored-by: redhelix <redhelix@users.noreply.github.com>
In headless environments (systemd, Docker, nohup) stdout can become
unavailable mid-session. Raw print() raises OSError which crashes
cron jobs — agent finishes work but delivery never happens because
the error handler's own print() also raises OSError.
Fix:
- Add _safe_print() static method that wraps print() with try/except
OSError — silently drops output when stdout is broken
- Make _vprint() use _safe_print() — protects all calls through the
verbose print path
- Convert raw print() calls in run_conversation() hot path to use
_safe_print(): starting conversation, interrupt, budget exhausted,
preflight compression, context cache, conversation completed
- Error handler print (the cascading crash point) gets explicit
try/except with logger.error() fallback so diagnostics aren't lost
Fixes#845Closes#1358 (superseded — PR was 323 commits stale with a bug)
Add HERMES_API_MODE env var and model.api_mode config field to let
custom OpenAI-compatible endpoints opt into codex_responses mode
without requiring the OpenAI Codex OAuth provider path.
- _get_configured_api_mode() reads HERMES_API_MODE env (precedence)
then model.api_mode from config.yaml; validates against whitelist
- Applied in both _resolve_openrouter_runtime() and
_resolve_named_custom_runtime() (original PR only covered openrouter)
- Fix _dump_api_request_debug() to show /responses URL when in
codex_responses mode instead of always showing /chat/completions
- Tests for config override, env override, invalid values, named
custom providers, and debug dump URL for both API modes
Inspired by PR #1041 by @mxyhi.
Co-authored-by: mxyhi <mxyhi@users.noreply.github.com>
* fix: prevent infinite 400 failure loop on context overflow (#1630)
When a gateway session exceeds the model's context window, Anthropic may
return a generic 400 invalid_request_error with just 'Error' as the
message. This bypassed the phrase-based context-length detection,
causing the agent to treat it as a non-retryable client error. Worse,
the failed user message was still persisted to the transcript, making
the session even larger on each attempt — creating an infinite loop.
Three-layer fix:
1. run_agent.py — Fallback heuristic: when a 400 error has a very short
generic message AND the session is large (>40% of context or >80
messages), treat it as a probable context overflow and trigger
compression instead of aborting.
2. run_agent.py + gateway/run.py — Don't persist failed messages:
when the agent returns failed=True before generating any response,
skip writing the user's message to the transcript/DB. This prevents
the session from growing on each failure.
3. gateway/run.py — Smarter error messages: detect context-overflow
failures and suggest /compact or /reset specifically, instead of a
generic 'try again' that will fail identically.
* fix(skills): detect prompt injection patterns and block cache file reads
Adds two security layers to prevent prompt injection via skills hub
cache files (#1558):
1. read_file: blocks direct reads of ~/.hermes/skills/.hub/ directory
(index-cache, catalog files). The 3.5MB clawhub_catalog_v1.json
was the original injection vector — untrusted skill descriptions
in the catalog contained adversarial text that the model executed.
2. skill_view: warns when skills are loaded from outside the trusted
~/.hermes/skills/ directory, and detects common injection patterns
in skill content ("ignore previous instructions", "<system>", etc.).
Cherry-picked from PR #1562 by ygd58.
---------
Co-authored-by: buray <ygd58@users.noreply.github.com>
* feat: add Vercel AI Gateway as a first-class provider
Adds AI Gateway (ai-gateway.vercel.sh) as a new inference provider
with AI_GATEWAY_API_KEY authentication, live model discovery, and
reasoning support via extra_body.reasoning.
Based on PR #1492 by jerilynzheng.
* feat: add AI Gateway to setup wizard, doctor, and fallback providers
* test: add AI Gateway to api_key_providers test suite
* feat: add AI Gateway to hermes model CLI and model metadata
Wire AI Gateway into the interactive model selection menu and add
context lengths for AI Gateway model IDs in model_metadata.py.
* feat: use claude-haiku-4.5 as AI Gateway auxiliary model
* revert: use gemini-3-flash as AI Gateway auxiliary model
* fix: move AI Gateway below established providers in selection order
---------
Co-authored-by: jerilynzheng <jerilynzheng@users.noreply.github.com>
Co-authored-by: jerilynzheng <zheng.jerilyn@gmail.com>
Two tests lacked filesystem isolation causing them to pick up real
~/.claude/.credentials.json tokens on machines with Claude Code installed.
- test_prefers_oauth_token_over_api_key: add tmp_path, mock Path.home,
clear CLAUDE_CODE_OAUTH_TOKEN env
- test_falls_back_to_token: same isolation
Also commit run_agent.py generic-400 retry fix.
Anthropic prompt caching splits input into cache_read_input_tokens,
cache_creation_input_tokens, and non-cached input_tokens. The context
counter only read input_tokens (non-cached portion), showing ~3 tokens
instead of the real ~18K total. Now includes cached portions for
Anthropic native provider only — other providers (OpenAI, OpenRouter,
Codex) already include cached tokens in their prompt_tokens field.
Before: 3/200K | 0%
After: 17.7K/200K | 9%
- 429 rate limit and 529 overloaded were incorrectly treated as
non-retryable client errors, causing immediate failure instead of
exponential backoff retry. Users hitting Anthropic rate limits got
silent failures or no response at all.
- Generic "Sorry, I encountered an unexpected error" now includes
error type, details, and status-specific hints (auth, rate limit,
overloaded).
- Failed agent with final_response=None now surfaces the actual
error message instead of returning an empty response.
* feat: improve memory prioritization — user preferences over procedural knowledge
Inspired by OpenAI Codex's memory prompt improvements (openai/codex#14493)
which focus memory writes on user preferences and recurring patterns
rather than procedural task details.
Key insight: 'Optimize for reducing future user steering — the most
valuable memory prevents the user from having to repeat themselves.'
Changes:
- MEMORY_GUIDANCE (prompt_builder.py): added prioritization hierarchy
and the core principle about reducing user steering
- MEMORY_SCHEMA (memory_tool.py): reordered WHEN TO SAVE list to put
corrections first, added explicit PRIORITY guidance
- Memory nudge (run_agent.py): now asks specifically about preferences,
corrections, and workflow patterns instead of generic 'anything'
- Memory flush (run_agent.py): now instructs to prioritize user
preferences and corrections over task-specific details
* feat: more aggressive skill creation and update prompting
Press harder on skill updates — the agent should proactively patch
skills when it encounters issues during use, not wait to be asked.
Changes:
- SKILLS_GUIDANCE: 'consider saving' → 'save'; added explicit instruction
to patch skills immediately when found outdated/wrong
- Skills header: added instruction to update loaded skills before finishing
if they had missing steps or wrong commands
- Skill nudge: more assertive ('save the approach' not 'consider saving'),
now also prompts for updating existing skills used in the task
- Skill nudge interval: lowered default from 15 to 10 iterations
- skill_manage schema: added 'patch it immediately' to update triggers
Thorough code review found 5 issues across run_agent.py, cli.py, and gateway/:
1. CRITICAL — Gateway stream consumer task never started: stream_consumer_holder
was checked BEFORE run_sync populated it. Fixed with async polling pattern
(same as track_agent).
2. MEDIUM-HIGH — Streaming fallback after partial delivery caused double-response:
if streaming failed after some tokens were delivered, the fallback would
re-deliver the full response. Now tracks deltas_were_sent and only falls
back when no tokens reached consumers yet.
3. MEDIUM — Codex mode lost on_first_delta spinner callback: _run_codex_stream
now accepts on_first_delta parameter, fires it on first text delta. Passed
through from _interruptible_streaming_api_call via _codex_on_first_delta
instance attribute.
4. MEDIUM — CLI close-tag after-text bypassed tag filtering: text after a
reasoning close tag was sent directly to _emit_stream_text, skipping
open-tag detection. Now routes through _stream_delta for full filtering.
5. LOW — Removed 140 lines of dead code: old _streaming_api_call method
(superseded by _interruptible_streaming_api_call). Updated 13 tests in
test_run_agent.py and test_openai_client_lifecycle.py to use the new
method name and signature.
4573 tests passing.
Previously the fallback only triggered on specific error keywords like
'streaming is not supported'. Many third-party providers have partial
or broken streaming — rejecting stream=True, crashing on stream_options,
dropping connections mid-stream, returning malformed chunks, etc.
Now: any exception during the streaming API call triggers an automatic
fallback to the standard non-streaming request path. The error is logged
at INFO level for diagnostics but never surfaces to the user. If the
fallback also fails, THAT error propagates normally.
This ensures streaming is additive — it improves UX when it works but
never breaks providers that don't support it.
Tests: 2 new (any-error fallback, double-failure propagation), 15 total.
