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
The BlueBubbles adapter was merged but missing setup wizard support:
- Add _setup_bluebubbles() guided setup (server URL, password, allowlist,
home channel, webhook port)
- Add to _GATEWAY_PLATFORMS registry so it appears in 'hermes setup gateway'
- Add to any_messaging check and home channel missing warning
- Add to gateway status display in 'hermes setup'
- Add BLUEBUBBLES_SERVER_URL, BLUEBUBBLES_PASSWORD, BLUEBUBBLES_ALLOWED_USERS
to OPTIONAL_ENV_VARS with descriptions and categories
Previously the only way to configure BlueBubbles was manually editing .env.
Two issues resolved:
1. Add opencode.ai to _URL_TO_PROVIDER mapping so base_url routes through
models.dev lookup (which has mimo-v2-pro at 1M context) instead of
falling back to probing /models (404) and defaulting to 128K.
2. Fix _format_context_length to round cleanly: 1048576 → '1M' instead
of '1.048576M'. Applies same rounding logic to K values.
Add QEMU cross-compilation and multi-arch manifest support so Apple
Silicon (M1/M2/M3) and other ARM-based systems get native images.
- Add docker/setup-qemu-action for arm64 emulation on amd64 runners
- Smoke test stays amd64-only (load:true can't export multi-arch)
- Both push steps (main + release) now build linux/amd64,linux/arm64
- Bump timeout 30->60min for QEMU cross-compilation overhead
- Add permissions: contents: read (least-privilege hardening)
Salvaged from PR #3998 by Mibayy. Also addresses #5005 and #3913.
Co-authored-by: Mibayy <Mibayy@users.noreply.github.com>
Tests for the new behavior paths:
- Large tool outputs no longer block compaction (motivating scenario)
- Hard minimum of 3 tail messages always protected
- 1.5x soft ceiling for oversized messages
- Small conversations still compress (min 8 messages)
- Token-budget prune path in _prune_old_tool_results
- Fallback to message-count when no token budget
PR #6240 changed tail protection from protect_last_n to min(3, ...)
which increased the minimum compressible message count and shifted
tail boundaries. Three tests broke:
- test_summary_role_avoids_consecutive_user_messages: 6→8 msgs
- test_double_collision_user_head_assistant_tail: 7→8 msgs
- test_no_collision_scenarios_still_work: 6→8 msgs
All tests now exceed the new min_for_compress threshold (6) and
maintain proper role alternation in both head and tail sections.
Tail protection was effectively message-count based despite having a
token budget, because protect_last_n=20 acted as a hard floor. A single
50K-token tool output would cause all 20 recent messages to be
preserved regardless of budget, leaving little room for summarization.
Changes:
- _find_tail_cut_by_tokens: min_tail reduced from protect_last_n (20)
to 3; token budget is now the primary criterion
- Soft ceiling at 1.5x budget to avoid cutting mid-oversized-message
- _prune_old_tool_results: accepts optional protect_tail_tokens so
pruning also respects the token budget instead of a fixed count
- compress() minimum message check relaxed from protect_first_n +
protect_last_n + 1 to protect_first_n + 3 + 1
- Tool group alignment (no splitting tool_call/result) preserved
Port missing features from the hindsight-hermes external integration
package into the native plugin. Only touches plugin files — no core
changes.
Features:
- Tags on retain/recall (tags, recall_tags, recall_tags_match)
- Recall config (recall_max_tokens, recall_max_input_chars, recall_types,
recall_prompt_preamble)
- Retain controls (retain_every_n_turns, auto_retain, auto_recall,
retain_async via aretain_batch, retain_context)
- Bank config via Banks API (bank_mission, bank_retain_mission)
- Structured JSON retain with per-message timestamps
- Full session accumulation with document_id for dedup
- Custom post_setup() wizard with curses picker
- Mode-aware dep install (hindsight-client for cloud, hindsight-all for local)
- local_external mode and openai_compatible LLM provider
- OpenRouter support with auto base URL
- Auto-upgrade of hindsight-client to >=0.4.22 on session start
- Comprehensive debug logging across all operations
- 46 unit tests
- Updated README and website docs
Prevents unbounded memory growth in _assistant_threads dict.
Evicts oldest entries when exceeding _ASSISTANT_THREADS_MAX (5000),
matching the pattern used by _mentioned_threads and _seen_messages.
The test_non_internal_event_without_user_triggers_pairing test relied on
no Discord auth env vars being set, but gateway/run.py loads dotenv at
module level. In environments with DISCORD_ALLOW_ALL_USERS=True in .env,
the auth check passed instead of triggering the pairing flow.
Clear DISCORD_ALLOW_ALL_USERS, DISCORD_ALLOWED_USERS, GATEWAY_ALLOW_ALL_USERS,
and GATEWAY_ALLOWED_USERS via monkeypatch to ensure test isolation.
When a background process with notify_on_complete=True finishes, the
gateway injects a synthetic MessageEvent to notify the session. This
event was constructed without user_id, causing _is_user_authorized()
to reject it and — for DM-origin sessions — trigger the pairing flow,
sending "Hi~ I don't recognize you yet!" with a pairing code to the
chat owner.
Add an `internal` flag to MessageEvent that bypasses authorization
checks for system-generated synthetic events. Only the process watcher
sets this flag; no external/adapter code path can produce it.
Includes 4 regression tests covering the fix and the normal pairing path.
`_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.
ORIGINAL INCIDENT:
Discord forum descriptions (the topic field on ForumChannel) were invisible
to the agent. When a user set project instructions in a forum's description
(e.g. tool-evaluations), threads created in that forum had no Channel Topic
in their session context. Discovered while evaluating per-forum auto-context
injection for web-tap-terminal development threads.
ISSUE IN THE CODE:
In gateway/platforms/discord.py, all three session entry points
(_handle_message, _build_slash_event, _dispatch_thread_session) read
chat_topic via getattr(channel, 'topic', None). Discord Thread objects
don't carry a topic — only the parent ForumChannel does. So chat_topic
was always None for forum threads, and the Channel Topic line was never
injected into build_session_context_prompt output. The infrastructure to
handle this was already in place — _is_forum_parent() detects forum
channels, _format_thread_chat_name() traverses to the parent, and
build_session_context_prompt() renders Channel Topic when present. The
forum parent was being identified; its topic just wasn't being read.
HOW THIS COMMIT FIXES IT:
Adds _get_effective_topic(channel, is_thread) helper that reads
channel.topic first, then falls back to the parent forum's topic when
the channel is a thread inside a forum. All three session entry points
now call this helper instead of inlining getattr(channel, 'topic', None).
Existing tests pass unchanged.
Co-authored-by: dhabibi <9087935+dhabibi@users.noreply.github.com>
The new degradation warning reads compression_count as an int,
but the existing test's MagicMock returns a MagicMock object
for that attribute, causing '>=' comparison to fail.
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
- Bind exception in warning send handler (was using stale _ne from outer scope)
- Calculate remaining time until timeout correctly: (timeout - warning) // 60
instead of warning // 60 (which equals elapsed time, not remaining)
Introduce gateway_timeout_warning (default 900s) as a pre-timeout alert
layer. When inactivity reaches the warning threshold, a single
notification is sent to the user offering to wait or reset. If
inactivity continues to the gateway_timeout (default 1800s), the full
timeout fires as before.
This gives users a chance to intervene before work is lost on slow
API providers without disabling the safety timeout entirely.
Config: agent.gateway_timeout_warning in config.yaml, or
HERMES_AGENT_TIMEOUT_WARNING env var (0 = disable warning).
Step c in switch_model() blindly converted the first colon to a slash for
aggregator providers, even when the model name already contained a slash
(vendor/model format). This mangled variant tags like :free into /free,
causing 400 Bad Request from the API.
Fix: skip the colon→slash conversion when the model already has a slash,
since the colon is a variant tag, not a vendor separator. The module
docstring already documented this intent (line 17-18) but the
implementation didn't enforce it.
Reported via Discord. Related to PR #6088 (which identified the same bug
but placed the fix in model_normalize.py instead of model_switch.py where
the actual mangling occurs).
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
Documents the per-platform tool_progress_overrides config key added in
PR #6348. Shows example YAML with Signal set to 'off' while Telegram
stays on 'verbose'. Lists all valid platform keys.
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
When managed_persistence is enabled, cleanup_browser() was calling
camofox_close() which destroys the server-side browser context via
DELETE /sessions/{userId}, killing login sessions across cron runs.
Add camofox_soft_cleanup() — a public wrapper that drops only the
in-memory session entry when managed persistence is on, returning True.
When persistence is off it returns False so the caller falls back to
the full camofox_close(). The inactivity reaper still handles idle
resource cleanup.
Also surface a logger.warning() when _managed_persistence_enabled()
fails to load config, replacing a silent except-and-return-False.
Salvaged from #6182 by el-analista (Eduardo Perea Fernandez).
Added public API wrapper to avoid cross-module private imports,
and test coverage for both persistence paths.
Co-authored-by: Eduardo Perea Fernandez <el-analista@users.noreply.github.com>
Fixes#4647 — Signal replies duplicated when gateway streaming is enabled.
Root cause: stream_consumer.py did not handle the case where send() returns
success=True but no message_id (Signal behavior). Every stream delta produced
a separate send() call (7+ messages instead of 2), plus the gateway sent
another full duplicate since already_sent was never set.
Changes:
- stream_consumer.py: Add elif branch for success-without-message_id — enters
fallback mode (sets already_sent, disables editing, sends only continuation)
- signal.py send(): Extract timestamp from signal-cli RPC result as message_id
so stream consumer follows normal edit→fallback path
- signal.py: Add public stop_typing() delegating to _stop_typing_indicator()
so base adapter's _keep_typing finally block can clean up typing tasks
- gateway/run.py: Per-platform tool_progress_overrides (#6164) — lets users
set e.g. signal: off while keeping telegram: all
- hermes_cli/config.py: Add tool_progress_overrides to DEFAULT_CONFIG
Refs: #4647, #6164
- Docker env tests: verify _build_init_env_args() instead of per-execute
Popen flags (env forwarding is now init-time only)
- Docker: preserve explicit forward_env bypass of blocklist from main
- Daytona tests: adapt to SDK-native timeout, _ThreadedProcessHandle,
base.py interrupt handling, HERMES_STDIN_ heredoc prefix
- Modal tests: fix _load_module to include _ThreadedProcessHandle stub,
check ensurepip in _resolve_modal_image instead of __init__
- SSH tests: mock time.sleep on base module instead of removed ssh import
- Add missing BaseEnvironment attributes to __new__()-based test fixtures
Add configurable reply-reference behavior for Discord, matching the
existing Telegram (TELEGRAM_REPLY_TO_MODE) and Mattermost
(MATTERMOST_REPLY_MODE) implementations.
Modes:
- 'off': never reply-reference the original message
- 'first': reply-reference on first chunk only (default, current behavior)
- 'all': reply-reference on every chunk
Set DISCORD_REPLY_TO_MODE=off in .env to disable reply-to messages.
Changes:
- gateway/config.py: parse DISCORD_REPLY_TO_MODE env var
- gateway/platforms/discord.py: read reply_to_mode from config, respect
it in send() — skip fetch_message entirely when 'off'
- hermes_cli/config.py: add to OPTIONAL_ENV_VARS for hermes setup
- 23 tests covering config, send behavior, env var override
- docs: discord.md env var table + environment-variables.md reference
Closes community request from Stuart on Discord.
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.
Users were setting model.provider to "main" after reading the auxiliary
provider docs, causing "Unknown provider" errors. The "main" alias is
only valid inside auxiliary:, compression:, and fallback_model: configs
where it means "use the same provider as my main agent chat."
Added warning admonitions and inline clarifications to:
- configuration.md: Auxiliary Models provider list and Provider Options table
- fallback-providers.md: Provider Options for Auxiliary Tasks table
Reported by community member cn on Discord.
Fixes 9 test failures on current main, incorporating ideas from PR stack
#6219-#6222 by xinbenlv with corrections:
- model_metadata: sync HF context length key casing
(minimaxai/minimax-m2.5 → MiniMaxAI/MiniMax-M2.5)
- cli.py: route quick command error output through self.console
instead of creating a new ChatConsole() instance
- docker.py: explicit docker_forward_env entries now bypass the
Hermes secret blocklist (intentional opt-in wins over generic filter)
- auxiliary_client: revert _read_main_provider() to simple
provider.strip().lower() — the _normalize_aux_provider() call
introduced in 5c03f2e7 stripped the custom: prefix, breaking
named custom provider resolution
- auxiliary_client: flip vision auto-detection order to
active provider → OpenRouter → Nous → stop (was OR → Nous → active)
- test: update vision priority test to match new order
Based on PR #6219-#6222 by xinbenlv.
* fix(nix): export HERMES_HOME system-wide when addToSystemPackages is true
The `addToSystemPackages` option's documentation (and the `:::tip` block in
`website/docs/getting-started/nix-setup.md`) promises that enabling it both
puts the `hermes` CLI on PATH and sets `HERMES_HOME` system-wide so interactive
shells share state with the gateway service. The module only did the former,
so running `hermes` in a user shell silently created a separate `~/.hermes/`
directory instead of the managed `${stateDir}/.hermes`.
Implement the documented behavior by also setting
`environment.variables.HERMES_HOME = "${cfg.stateDir}/.hermes"` in the same
mkIf block, and update the option description to match.
Fixes#6044
* fix(nix): preserve group-readable permissions in managed mode
The NixOS module sets HERMES_HOME directories to 0750 and files to 0640
so interactive users in the hermes group can share state with the gateway
service. Two issues prevented this from working:
1. hermes_cli/config.py: _secure_dir() unconditionally chmod'd HERMES_HOME
to 0700 on every startup, overwriting the NixOS module's 0750. Similarly,
_secure_file() forced 0600 on config files. Both now skip in managed mode
(detected via .managed marker or HERMES_MANAGED env var).
2. nix/nixosModules.nix: the .env file was created with 0600 (owner-only),
while config.yaml was already 0640 (group-readable). Changed to 0640 for
consistency — users granted hermes group membership should be able to read
the managed .env.
Verified with a NixOS VM integration test: a normal user in the hermes group
can now run `hermes version` and `hermes config` against the managed
HERMES_HOME without PermissionError.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: zerone0x <zerone0x@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
mistralai v2.x is a namespace package — `Mistral` class lives at
`mistralai.client`, not at the top-level `mistralai` module. The
previous `from mistralai import Mistral` raises ImportError at runtime.
Update both production code and test fixture to use the correct path.
- Add HERMES_QWEN_BASE_URL to OPTIONAL_ENV_VARS in config.py (was missing
despite being referenced in code)
- Remove redundant qwen-oauth entry from _API_KEY_PROVIDER_AUX_MODELS
(non-aggregator providers use their main model for aux tasks automatically)
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)
Fixes#6259. Three bugs fixed:
1. Config key mismatch: _get_client() and _start_daemon() read
'llmApiKey' (camelCase) but save_config() stores 'llm_api_key'
(snake_case). The config value was never read — only the env var
fallback worked.
2. Missing base URL support: users on OpenRouter or custom endpoints
had no way to configure HINDSIGHT_API_LLM_BASE_URL through setup.
Added llm_base_url to config schema with empty default, passed
conditionally to HindsightEmbedded constructor.
3. Daemon config change detection: config_changed now also checks
HINDSIGHT_API_LLM_BASE_URL, and the daemon profile .env includes
the base URL when set.
Keeps HINDSIGHT_API_LLM_API_KEY (with double API) in the daemon
profile .env — this matches the upstream hindsight .env.example
convention.
Documents the proxy env var support added in PR #3591 (salvage of #3411
by @kufufu9). Covers HTTPS_PROXY/HTTP_PROXY/ALL_PROXY precedence,
configuration methods, and scope.
