Multiple @-references in one message (esp. @url: refs, each a full
web_extract round-trip) were expanded in a serial `for ref in refs: await`
loop. Switch to asyncio.gather over the independent _expand_reference calls,
reassembling warnings/blocks in original positional order so output is
byte-identical to the serial path; the token-budget check is unchanged.
Generic + provider-agnostic: helps every web backend equally (exa/tavily/
firecrawl/parallel) since it's above the provider layer. RED/GREEN test:
3 url refs @ 0.2s each = 0.60s serial -> ~0.20s concurrent.
#53552 flipped verify_on_stop to default OFF because the guard fired on
doc/markdown/skill edits and felt like noise. That doc/markdown/skill
suppression already shipped in the same change (_filter_verifiable_paths in
agent/verification_stop.py), so the original noise rationale no longer holds:
the guard already skips prose-only turns.
Restore the surface-aware "auto" default — ON for interactive coding surfaces
(CLI, TUI, desktop) and programmatic callers, OFF for conversational messaging
surfaces (Telegram, Discord, etc.) where the verification narrative would reach
a human as chat noise. The missing/unrecognized fallback in
verify_on_stop_enabled now resolves to the same surface-aware default instead of
hard OFF, so both the DEFAULT_CONFIG value and the resolver agree.
Scope: this changes the shipped default for fresh installs and configs without
an explicit verify_on_stop key. Existing configs that #53552/#54740 migrated to
an explicit `false` are respected and unchanged — this PR does not add a
force-migration of those values back to auto.
agent.coding_instructions (a string or list) is appended to the coding brief as
its own stable system block, so users can pin project-wide workflow rules
without editing the shipped brief. Coding-posture only and cache-safe (resolved
once per session; takes effect next session). Empty by default.
Add a `pre_verify` user/plugin/shell hook fired once per turn when the agent
edited code and is about to finish, after the existing verify-on-stop guard. A
hook can keep the agent going one more turn (run a check, defer it, tidy the
diff) by returning {"action":"continue","message":...} (the Claude-Code Stop
shape {"decision":"block","reason":...} is accepted too). Hooks receive coding,
attempt, final_response, and sorted changed_paths so they can self-scope and
self-throttle; the path is bounded by agent.max_verify_nudges and preserves
message-role alternation.
Hermes still ships its default coding guidance (agent.verify_guidance, on by
default), but it now rides the evidence-based verify-on-stop missing-evidence
nudge instead of a separate default pre_verify continuation, so it costs no
extra model turn of its own. Guidance reuses the shared utils.is_truthy_value
parser rather than a local copy.
Assemble a per-profile graph of memories and learned skills over time
(agent/learning_graph.py) and serve it at GET /api/learning/graph
(hermes_cli/web_server.py), with tests. The radial time axis the desktop
renders is derived from this payload; the REST path stays under /learning
for backend compatibility.
* feat(display): friendly human-phrased tool labels for built-in tools
Built-in tools now render ChatGPT-style status verbs ('Searching the web
for ...', 'Reading <file>', 'Browsing <url>') on the CLI spinner and
gateway/desktop tool-progress instead of the raw tool name.
- agent/display.py: _TOOL_VERBS map + build_tool_label() + set/get
friendly-labels flag (default on). Custom/plugin/MCP tools fall back to
the raw preview; verbose gateway mode left untouched (debug surface).
- tool_executor.py / tui_gateway / gateway: route the three spinner sites,
the TUI _tool_ctx, and the gateway all/new progress line through the label.
- config: display.friendly_tool_labels (default True, per-platform aware).
Zero new core tool / schema footprint — pure display layer.
* docs: add PR infographic for friendly tool labels
* fix(display): preserve arg preview in gateway friendly labels + update tests
The first gateway pass re-derived the label from the callback's `args`, which
is empty ({}) at the gateway tool.started callsite — the command/query lives in
the `preview` string, so terminal rendered as a bare '💻 Running' and dedup
collapsed consecutive commands. Now the gateway prefixes the verb onto the
already-computed preview via get_tool_verb/tool_verb_connector/verb_drops_preview,
preserving the command/url/query. CLI spinner path (real args) keeps build_tool_label.
Tests: update test_run_progress_topics exact-format assertions to the friendly
form ('💻 Running pwd'), add a format-agnostic preview extractor for the
truncation tests (works for both quoted-legacy and verb-prefixed output).
* test(tui): update resume-display context to friendly tool label
_tool_ctx now uses build_tool_label, so the desktop resume-view context for a
search_files turn reads 'Searching files for resume' instead of the bare
'resume' preview — consistent with live tool-progress. Update the assertion.
* test(tui): harden no-race worker test against sibling shard leakage
test_session_create_no_race_keeps_worker_alive flaked under -j 8: a daemon
build thread leaked from a prior session.create test in the same shard process
fires close/unregister against its own (foreign) session_key after this test
patches the global approval hooks, polluting the captured lists. Scope the
assertions to this session's own session_key so the regression intent
(this session's worker/notify must survive) is preserved while the test
becomes immune to shard composition. Not related to friendly-tool-labels.
Let users click the status bar context indicator to see how tokens are
split across system prompt, tools, rules, skills, MCP, and conversation.
Co-authored-by: Cursor <cursoragent@cursor.com>
The five _resolved_api_call_stale_timeout_base integration tests reloaded
hermes_cli.config + hermes_cli.timeouts via importlib.reload to clear cached
config. Under xdist that mutates module-global state shared across the worker
process, so a sibling test could leave the config cache in a state that made
get_provider_stale_timeout return a leaked value — intermittently failing
test_reasoning_floor_applies_to_opus_4_thinking (shard 6 flake, #52217 area).
Patch run_agent.get_provider_stale_timeout per-test instead: floor-path tests
get None (resolver falls through to the reasoning floor / env var / default),
the explicit-config test gets 60.0 (priority-1 short-circuit). Same assertions,
no shared-module mutation, deterministic under parallel execution.
A single agent turn can fan out N vision_analyze calls at once — the
classic trigger is "analyze every frame of this video", where ffmpeg
explodes a clip into dozens of frames and the model calls vision_analyze
on each. Every call does a CPU-heavy base64-encode/resize burst AND holds
a long-lived LLM stream open. The tool executor runs concurrent tool calls
on a per-session ThreadPoolExecutor (_MAX_TOOL_WORKERS=8), and multiple
agent sessions share one process (the dashboard runs the agent in-process),
so there was no global ceiling. In prod (June 2026) a video-frame fan-out
pinned a worker thread at ~100% CPU and starved the shared asyncio event
loop that also serves the dashboard's /api/status liveness probe, flapping
the instance to UNHEALTHY even though nothing had crashed.
Add a process-global threading.BoundedSemaphore that bounds how many vision
analyses run concurrently across the whole process, held across the entire
analysis (image load + encode + LLM call) in the single _handle_vision_analyze
chokepoint (covers both the native fast path and the legacy aux-LLM path).
It is a threading semaphore, NOT asyncio: each vision call is dispatched
through model_tools._run_async on a per-thread event loop, so an asyncio
primitive bound to one loop cannot coordinate across them. The acquire is
offloaded via run_in_executor so waiting for a slot never blocks the calling
loop.
Default: min(host CPUs, 4), floored at 1 — respect the host's concurrency,
or lower. Override via auxiliary.vision.max_concurrency (config.yaml) or
HERMES_VISION_MAX_CONCURRENCY (env). Values < 1 are ignored so the cap can
never be disabled into an unbounded fan-out.
Tests: bounded-fan-out regression guard + a control proving it would fail
without the cap; resolver tests for host-cpu default, ceiling clamp, low-cpu
host, env override, and sub-1 rejection. Pre-existing handler tests updated
for the now-async _handle_vision_analyze. Verified via the real
registry.dispatch -> _run_async per-thread-loop path (16 concurrent calls,
peak bounded to cap).
Regression tests for the injection fix: outside a git repo only cwd is
checked (planted ancestor .hermes.md is ignored), a cwd-local .hermes.md
is still found, and inside a git repo the parent walk to the git root
still works.
git remote set-url with an embedded password (https://PASSWORD@github.com)
leaked the credential into agent output — the redaction engine only masked
user:pass@ DB connection strings, never the colon-less bare-token userinfo
form a git remote uses.
Add _URL_BARE_TOKEN_RE: scheme://TOKEN@host for web/transport schemes
(http/https/wss/git/ssh/ftp), 8+ char floor to skip short usernames, token
class forbidding /:@ so an @ in a path/query is never treated as userinfo.
Deliberately scoped to the bare-token form only. The user:pass@ colon form
and query-string tokens stay passing through (#34029, 'pass web URLs through
unchanged') so magic-link / OAuth round-trip skills keep working — a bare
credential in userinfo is never a workflow token (those live in the query
string), so masking it can't break a skill.
The curator's inactivity prune archived any non-pinned agent-created
skill whose activity was older than archive_after_days (90d). A skill
loaded only by a cron job had its usage bumped solely when the job
fired, so paused jobs, infrequent (quarterly/annual) schedules, and
far-future one-shots aged their skills out from under them — the next
run then failed to load the now-archived skill.
- cron/jobs.py: add referenced_skill_names() returning skills used by
ANY job (incl. paused/disabled).
- curator.apply_automatic_transitions(): skip cron-referenced skills
like pinned; add a use=0 grace floor so a never-used skill is not
marked stale/archived until it is at least stale_after_days old.
- LLM review pass: candidate list marks cron=yes; prompt forbids
pruning cron-referenced skills and never-used skills under 30 days.
Tested E2E against a real cron job + real usage records and with 4 new
unit tests.
The salvaged #35519 regression guard asserted that default (non-file_read)
mode keeps a head/tail `ghp_S1...Pn2T` mask for a `token: <key>` line. On
current main the YAML config pass (`_YAML_ASSIGN_RE`, key `token`) re-masks
the already-prefix-masked value to `***`, so the assertion was stale. Switch
to a bare-token context so the guard isolates what it claims (prefix-mask
head/tail shape in default mode) without depending on the YAML collapse.
When security.redact_secrets is on (default), read_file/search_files/cat
applied redact_sensitive_text(code_file=True) to file content, which still
ran prefix masking. An API key in config.yaml (ghp_..., sk-..., xai-..., etc.)
came back as a head/tail mask like `ghp_S1...Pn2T` — a plausible-looking
truncated key. When an agent read that and wrote it back to config, the masked
value replaced the real credential, silently breaking auth (401). Production
evidence: a config.yaml found containing the exact 13-char masked GitHub PAT.
The two community PRs (#35529, #35534) fixed the corruption by NOT redacting
prefixes for config reads — but that exposes the user's real keys to the agent
context, model, and logs (a security regression). This takes the safer route:
keep redacting, but for file content emit a NON-REUSABLE sentinel.
- New `_mask_token_nonreusable`: prefix secrets -> `«redacted:ghp_…»` (vendor
label preserved for debuggability; zero secret bytes; angle-bracket/ellipsis
wrapper is syntactically invalid as a token so it can't be mistaken for or
written back as a usable key).
- New `redact_sensitive_text(file_read=True)` routes prefix matches through it
(implies code_file=True). Default/log/display mode is UNCHANGED — `_mask_token`
still keeps head/tail (fine for logs, never written back).
- Wired the 3 file_tools.py call sites (read_file / search_files / cat) to
file_read=True.
Fixes both the corruption AND avoids the secret-exposure of the un-redact
approach. 6 new tests (sentinel shape, no-leak, not-a-plausible-key, default
mode unchanged, file_read implies code_file, sk- prefix); 88 redact tests pass;
mutation-verified (reverting to the old mask fails the sentinel/leak tests).
Co-authored-by: liuhao1024 <sunsky.lau@gmail.com>
Co-authored-by: adammatski1972 <289282750+adammatski1972@users.noreply.github.com>
Closes#35519. Supersedes #35529, #35534.
* fix(security): redact secrets in background process + foreground env-dump output
Terminal-output redaction was incomplete (#43025):
- Gap 1: process(action=poll/log/wait) returned background stdout verbatim —
no redaction at all. A background printenv/server/test emitting a key leaked
raw to the model, session.db, and CLI display. Same for the gateway
background-process watcher's completion/progress notifications.
- Gap 2: the foreground terminal path hardcoded code_file=True, which skips the
ENV-assignment pass, so an opaque token (no vendor prefix) from env/printenv
leaked even there.
Adds agent.redact.redact_terminal_output(output, command) as the single policy
for ALL terminal-output surfaces: env-dump commands (env/printenv/set/export/
declare) get the ENV-assignment pass (code_file=False) to mask opaque tokens;
other commands stay on code_file=True to avoid false positives on source dumps.
Wired into terminal_tool, process_registry (_handle_process boundary), and the
gateway watcher. Respects security.redact_secrets (no force) — opt-out preserved.
* docs: add infographic for #43025 terminal-output redaction fix
Two related redaction bugs from #43083:
1. build_assistant_message redacted tool-call arguments in-memory. That dict
feeds both the replayed conversation history and state.db (which is itself
replayed verbatim on session resume), so the model read back its own
PGPASSWORD='***' psql call and copied the placeholder, breaking every
credential-dependent command on the second turn. The masking gave no real
protection either — the same secret still leaks through tool OUTPUT. Remove
it. Keeping secrets out of the replayable store is a separate
tokenization/vault concern (security.redact_secrets still governs
storage-time redaction elsewhere).
2. _AUTH_HEADER_RE's greedy \S+ credential class ate a closing quote when the
token sat flush against it (Authorization: Bearer sk-.."), turning value
corruption into syntax corruption (unterminated quote -> shell EOF /
SyntaxError). Exclude " and ' from the token class; real credentials never
contain them.
Closes#43083.
The system-prompt backend probe imported a nonexistent symbol —
`from tools.environments import get_environment` — which always raised
ImportError: cannot import name 'get_environment'. The exception is caught
and only drops the live backend description to a static fallback, so it is
cosmetic, but it broke the live OS/user/cwd probe for every non-local
backend (docker/singularity/modal/daytona/ssh).
The real factory is `_create_environment` in tools.terminal_tool. Build the
environment the same way the live terminal path does (select backend image,
assemble ssh/container config from _get_env_config()), then run the probe.
Note: this does NOT affect tool loading — tool selection runs each tool's
check_fn and never consults this probe. Regression from #52147 (2026-06-25).
Closes#53667 (probe import); the 'cronjob-only' tool-collapse symptom is
not reproducible — tool selection has no probe dependency and memory's
check_fn is unconditionally True.
Secret redaction is display/output-scoped on main — write_file writes
content verbatim, terminal/execute_code redact only output not the
command/source. The real bug is in displayed tool OUTPUT (read_file,
terminal, execute_code):
_DB_CONNSTR_RE's password group [^@]+ was greedy across newlines, so on a
multi-line block it scanned past the DSN line to the next stray '@' (a
Python @decorator), replacing every intervening character — including line
breaks — with ***. That dropped lines and concatenated the next line onto
the f-string line, making read_file output look corrupted (the file on disk
was always correct). Reported in #33801.
Fix:
- Forbid whitespace in the userinfo/password groups ([^:\s]+ / [^@\s]+) so
the match can never span a line break. A real DSN password never contains
whitespace. This alone kills the catastrophic line-dropping.
- Under code_file=True, preserve a password group that is a pure {...} brace
expression — f"postgresql://{user}:{pass}@{host}" is an f-string template,
not a live credential. Literal passwords are still masked.
- Pass code_file=True at the terminal and execute_code output redaction call
sites (file_tools already did) so code-execution output isn't corrupted by
ENV/JSON/template false positives. Real prefixes, auth headers, JWTs, and
private keys are still redacted.
Verified E2E against the reporter's exact pydantic-settings module: file
written verbatim, read_file shows the DSN f-string + @model_validator intact
with zero *** corruption, while a literal postgresql://admin:pw@host DSN and
a real sk- key are still masked.
Reported-by: koishi70
Reported-by: pfrenssen
When a provider's output-layer safety filter (MiniMax "output new_sensitive
(1027)", Azure content_filter, etc.) kills a streaming response after deltas
were already sent, interruptible_streaming_api_call swallows the raw error
into a finish_reason=length partial-stream stub. The conversation loop then
burned 3 continuation retries against the SAME primary — re-hitting the
content-deterministic filter every time — and gave up with "Response remained
truncated after 3 continuation attempts", never consulting fallback_providers.
Builds on @595650661's classifier change (cherry-picked) so error_classifier
recognizes the filter; then:
- chat_completion_helpers: run the swallowed error through error_classifier at
the stub-creation point and stamp _content_filter_terminated on the stub
(single source of truth — no parallel pattern list).
- conversation_loop: read the tag and activate the fallback chain BEFORE
burning any continuation retries; roll partial content back to the last
clean turn and re-issue against the new provider (restart_with_rebuilt_messages).
Plain network stalls are unaffected (only content_policy_blocked is tagged).
Credits #32479 (@sweetcornna) and #33845 (@Tranquil-Flow) which fixed the
same issue via the stub-tag and loop-escalation approaches respectively.
Live E2E confirmed: before, _try_activate_fallback called 0x; after, fallback
fires on the first stub and the fallback provider completes the turn.
A WebUI/TUI session whose last turn died mid-tool-loop (stale-timeout kill,
interrupt, or process restart before the tool result was written) persists a
dangling assistant(tool_calls) or interrupted assistant->tool tail. The
messaging gateway already strips these tails before replay (the #49201 fix),
but the TUI/WebUI resume path fed db.get_messages_as_conversation() straight
in as the agent's conversation_history with no cleanup. The model re-issued
the unanswered call on every resume -- including after a full WebUI + Gateway
restart, since the poison lives in the SessionDB, not memory -- leaving the
session permanently 'thinking'. Only deleting the session recovered it.
- Extract the two strippers + helper from gateway/run.py into a shared
agent/replay_cleanup.py (sanitize_replay_history wraps both).
- gateway/run.py re-exports under the historical private names; messaging
behavior unchanged.
- Both TUI cold-resume sites now sanitize the model-fed history while leaving
the display transcript untouched, so the user still sees their full history.
Verified E2E against a real SessionDB: dangling and interrupted tails are
stripped from the model feed, healthy mid-progress tool sequences are
preserved, and the display transcript is always the full raw history.
* fix(agent): config-driven intent-ack continuation for all api_modes (#27881)
The agent could end a turn after only stating intent ('I will run a health
check...') without executing the announced tool call, forcing the user to
re-prompt. A continuation guard that catches this and nudges the model to
proceed already existed but was hard-gated to the codex_responses api_mode,
so Gemini/Claude/OpenRouter turns never benefited.
- New agent.intent_ack_continuation config (default 'auto' = codex-only,
byte-stable for existing conversations). 'true'/model-list opts every
api_mode in; 'false' disables. Mirrors agent.tool_use_enforcement's shape.
- looks_like_codex_intermediate_ack gains require_workspace (default True).
The opted-in path drops the codebase/filesystem requirement so general
autonomous workflows (server ops, deploys, API calls) are caught, not just
coding tasks. Future-ack + action-verb + short-content + no-prior-tool
guards still apply; the 2-nudge-per-turn cap is unchanged.
- Resolution centralized in intent_ack_continuation_mode (off/codex_only/all).
* docs(infographic): intent-ack continuation (#27881)
_restore_primary_runtime restored the construction-time api_key snapshot and
never consulted the credential pool. After the pool rotated away from a
revoked/exhausted entry mid-session, every new turn restored the dead key,
re-failed instantly, burned the remaining entries, and fell through to
cross-provider fallback.
After restoring the snapshot, re-select the pool's current best entry and
swap the live credential in via _swap_credential (which already rebuilds the
OpenAI/Anthropic client, reapplies base-url headers, and carries the #33163
base_url / OAuth-detection fixes). Falls back to the snapshot key when the
pool is absent, empty, or the entry has no usable key.
Salvaged from #25206 onto current main: the original targeted the pre-refactor
monolithic method in run_agent.py; the logic now lives in
agent/agent_runtime_helpers.py and is collapsed onto _swap_credential instead
of re-inlining the client rebuild.
Fixes#25205
A document attached alongside an image in the same Discord message was
swept into the vision pipeline and 400'd the whole turn ("Could not
process image"), and was simultaneously never surfaced to the agent as a
readable file. Restores the "any file type works" contract for mixed
messages and fixes the HTTP 400.
Bug 1 — mixed attachments: the inbound routing loop keyed image/audio/video
classification off the message-level type (PHOTO/VOICE/AUDIO), so a doc in
a PHOTO message landed in image_paths and poisoned the vision call. The
document context-note path was gated on message_type == DOCUMENT, so that
same doc never reached the agent at all. Now classification is
per-attachment (trust each attachment's own MIME; fall back to the
message-level type only when MIME is unknown), via shared _event_media_is_*
helpers used by both _build_media_placeholder and the main inbound loop.
The document note now fires for any non-image/audio/video attachment
regardless of message-level type.
Bug 2 — uncommon formats: AVIF/HEIC/BMP/TIFF/ICO produced the same generic
400 because providers only accept PNG/JPEG/GIF/WEBP. image_routing now
transcodes those to PNG via Pillow before declaring media_type, skipping
cleanly (logged) if Pillow/plugins are missing. SVG is vector — Pillow
can't rasterize it — so it's skipped rather than transcoded.
Closes#25935.
Co-authored-by: LeonSGP43 <cine.dreamer.one@gmail.com>
Co-authored-by: cypres0099 <74935762+cypres0099@users.noreply.github.com>
The Anthropic SDK clients were built without max_retries, so the SDK
default (max_retries=2) retried 429/5xx with its own backoff that ignores
Retry-After — double-retrying inside hermes's outer loop and burning
request slots against a bucket that won't refill for minutes. Set
max_retries=0 on all Anthropic/AnthropicBedrock client constructions so
the outer conversation loop (which already honors Retry-After) owns retry.
Also raise the Retry-After cap in the conversation loop from 120s to 600s.
Anthropic Tier 1 input-token buckets reset in ~171s, so the 120s cap made
hermes retry before the reset window and re-trip the limit.
Refs #26293
Claude Code OAuth refresh tokens are single-use; Claude Code refreshes on
its own schedule, so by the time Hermes notices an expired token Claude
Code may have already rotated it. Re-read live credential sources first and
adopt a valid token rather than POSTing a possibly-stale refresh token.
Ports the _refresh_oauth_token hardening from PR #40107 (chazmaniandinkle)
on top of the keychain/file reconciliation from PR #21112 (nodejun).
Adds AUTHOR_MAP entry for nodejun.
read_claude_code_credentials() previously returned the macOS Keychain
entry as soon as one existed, even if its OAuth token was already
expired. Callers then ran is_claude_code_token_valid() on the result
and got False, so resolve_anthropic_token() returned None — surfacing
the misleading 'No Anthropic credentials found' error even when
~/.claude/.credentials.json held a perfectly valid token.
Now reads both sources and prefers the non-expired one. When both are
valid (or both expired), prefers the later expiresAt so any subsequent
refresh uses the freshest refresh_token.
Adds TestReadClaudeCodeCredentialsDesync covering the four reconciliation
cases. The existing 'keychain wins' priority test still passes because
both fixtures share the same expiresAt and the tiebreaker is >=.
_try_openrouter() returned (None, None) whenever an OpenRouter credential
pool existed but was exhausted (_select_pool_entry -> (True, None)), making
the OPENROUTER_API_KEY env-var fallback unreachable. Auxiliary tasks
(compression, vision, web_extract) silently failed even with a valid env key.
Now the pool-present branch only returns early when it successfully builds a
client; an exhausted pool falls through to the env-var path. The final
failure (pool exhausted AND no env var) still marks the provider unhealthy.
Fixes#23452.
Co-authored-by: ambition0802 <noreply@github.com>
When the primary provider returns 401 and the auth-refresh path is
unavailable or fails, both call_llm() and async_call_llm() reached the
should_fallback gate without _is_auth_error in the condition, so the
auxiliary task (e.g. compression) was dropped silently — losing message
history. Add _is_auth_error to should_fallback (NOT is_capacity_error) in
both sync and async paths, plus an 'auth error' reason branch.
Auth stays a non-capacity error: it falls back in auto mode via the
is_auto gate, but on an explicitly-configured provider it still respects
the user's choice and raises rather than silently switching providers.
On a MoA session, auxiliary tasks (title generation, compression, vision, …)
ran through _resolve_auto with provider='moa' / model='<preset>', which sent
the preset name (e.g. 'opus-gpt') as the model id to resolve_provider_client —
producing 'HTTP 400: opus-gpt is not a valid model ID' on every turn (visible
as the title-generation warning).
MoA is a virtual provider with no real HTTP endpoint; aux tasks don't need the
reference fan-out. _resolve_auto now resolves a 'moa' main provider to the
preset's aggregator slot (its acting model) and continues Step 1 with that real
provider+model, dropping the virtual moa://local base_url + placeholder key so
the aggregator resolves via its own provider credentials. Mirrors the MoA
context-length resolution.
Verified live: a MoA turn no longer emits the 'not a valid model ID' warning.
Test: tests/agent/test_auxiliary_main_first.py (19 pass).
A MoA session's model is the preset name (e.g. 'opus-gpt') and its base_url is
the virtual local endpoint, so get_model_context_length() missed every probe
and fell through to the 256K fallback — even when the aggregator is a 1M-context
model. The acting model in MoA IS the aggregator, so resolve the context window
from the aggregator slot's real provider+model.
- model_metadata.get_model_context_length: when provider=='moa', resolve the
preset's aggregator slot through resolve_runtime_provider and recurse with the
aggregator's real provider/model/base_url. Explicit model.context_length still
wins (checked first); falls through to the generic default if resolution fails.
Tests: opus-gpt preset now reports 1M (the aggregator window), config override
still honored.
The secret redactor only matched uppercase env-style keys ([A-Z0-9_]),
so config-file assignments like spring.datasource.password=secret,
app.api.key=xyz, and YAML password: secret leaked verbatim when the
agent ran cat/grep on application.properties or .env files (issue #16413).
Adds three case-insensitive config-key matchers that run only in a
config-file context, preserving the existing #4367 (lowercase code/prose)
and web-URL-passthrough carve-outs:
- _CFG_DOTTED_RE: namespaced keys (contain a dot) — unambiguously config
- _CFG_ANCHORED_RE: bare secret-word keys at line start (incl. export)
- _YAML_ASSIGN_RE: unquoted colon config (password: value)
Value capture stops at whitespace and '&' so form bodies stay pair-wise;
the '://' guard keeps intentional web-URL query-param passthrough intact.
Reported-by: Murtaza1211
Z.AI / Zhipu reuse HTTP 429 for server-wide overload. The 429 status
path classified these unconditionally as rate_limit with
should_rotate_credential=True, so an overloaded provider exhausted the
credential pool after two errors — fatal for a single-key user, who has
nothing to rotate to.
The credential is valid; the server is just busy. Disambiguate the 429
body against a shared _OVERLOADED_PATTERNS list and route overload
language to FailoverReason.overloaded (retryable, no rotation), matching
the existing 503/529 path and the message-only path (#52890). Genuine
rate limits (no overload language) still rotate.
Extracted the inline overloaded tuple #52890 added into the shared
_OVERLOADED_PATTERNS constant so the status-code and message paths use
one list.
Closes#14038.
The curator_env fixture left async review threads (synchronous=False spawns
a daemon 'curator-review' thread that calls save_state() on completion)
running past test teardown. save_state() resolves the state path from
HERMES_HOME at write time, so a straggler could write into the next test's
tmp home, corrupting test_state_file_survives_corrupt_read (and others)
under CI load. Join the thread on teardown while HERMES_HOME is still
pinned to this test's home.
The 'whatsapp' and 'signal' PLATFORM_HINTS told the agent 'Please do not
use markdown as it does not render' — factually wrong. Both adapters
actively convert markdown to native formatting:
- whatsapp_common.format_message(): **bold**, ~~strike~~, # headers,
links, code blocks -> WhatsApp native syntax
- signal_format.markdown_to_signal(): same conversions via bodyRanges,
plus '- item' / '* item' bullets -> '• ' Unicode bullets
The wrong hint made the agent strip bullets and bold the adapter would
have rendered (#12224). Rewrote both hints to mirror whatsapp_cloud:
markdown is auto-converted, bullet lists work, tables are not supported.
Added a contract test asserting markdown-converting platforms never
forbid markdown in their hint.
The verify-on-stop guard fired too eagerly — including on doc/markdown/skill
edits with nothing to verify, where it pushed a pointless /tmp verification
script. Three changes:
1. Default OFF for new installs: agent.verify_on_stop defaults to false
(was the "auto" surface-aware sentinel). _config_version bumped 30 -> 31.
2. One-time migration (v30 -> v31): existing installs are switched off once,
but only when the value is missing or still the "auto" sentinel — an
explicit true/false the user set is preserved.
3. Path filter: build_verify_on_stop_nudge() now drops documentation/prose
paths (.md/.mdx/.rst/.txt/LICENSE/CHANGELOG/...) so even when explicitly
enabled, a doc-only turn never nudges. Mixed doc+code turns still nudge on
the code paths.
The legacy "auto" sentinel is still honored when set explicitly (ON for
interactive coding surfaces, OFF for messaging). HERMES_VERIFY_ON_STOP env
override unchanged.
When operator config has provider=anthropic with model.base_url pointing
at a non-Anthropic host (e.g. https://openrouter.ai/api/v1 with provider=anthropic),
the auxiliary Anthropic path was unconditionally applying that override.
Main-session traffic routed correctly because the main path attaches the
right credential for the actual destination, but every side-channel call
(memory extractors, reflection, vision, title generation, janus
extractor/promise) sent ANTHROPIC_API_KEY to the foreign host and 401'd.
Gate the override on hostname == api.anthropic.com. Operators routing main
through a non-Anthropic provider must use that provider's own auxiliary
client; the Anthropic aux path now stays pointed at api.anthropic.com.
Regression tests cover openrouter, openai, anthropic-with-path, empty, and
anthropic-default-base_url cases.
Wire get_reasoning_stale_timeout_floor() into both stale detectors so known
reasoning models (Nemotron 3 Ultra, OpenAI o1/o3, Opus 4.x thinking, DeepSeek
R1, Qwen QwQ, Grok reasoning) tolerate multi-minute thinking phases instead of
the upstream gateway idle-killing the socket (BrokenPipeError) before first
token. Applied as max(default, floor) — never overrides explicit user config,
never lowers an existing threshold.
The reasoning_timeouts.py allowlist module already landed on main via #52795,
so this salvage carries only the wiring + tests (the duplicate module and the
stale-base MoA reverts from the original PR branch are dropped).
Salvaged from #52238. Fixes#52217.
The salvaged #51875 added a background-review write guard in skill_manage
that refuses mutations to skills.external_dirs skills — but it only fires
when is_background_review() is true. The curator's LLM review fork ran with
the default _memory_write_origin='assistant_tool', so the guard never
triggered during the exact curation pass it exists to protect against
(GH-47688).
- Set _memory_write_origin='background_review' on the curator review fork so
turn_context binds it onto the write-origin ContextVar and the guard fires.
- Add a regression test asserting the fork runs under the background_review
origin (the invariant linking the fork to the guard).
- AUTHOR_MAP: map yu-xin-c for the salvaged commit.
Force redact_sensitive_text(force=True) on the browser_type text arg so
recognized credentials (API keys, tokens, JWTs) are masked in tool
progress, previews, callbacks, and return payloads even when the global
security.redact_secrets opt-out is set — a typed credential reaching chat
history is a security boundary, not log hygiene. Normal typed text matches
no pattern and stays fully readable for debuggability.
Tests assert the API-key-shaped secret is masked across every surface and
that normal text passes through unchanged.
CredentialPool._sync_device_code_entry_to_auth_store rotated single-use
OAuth refresh tokens but wrote the new chain only into the active profile
store. When a profile resolves a grant from the global-root fallback
(read_credential_pool, #18594) and the pool then refreshes it, root was
left holding a now-revoked refresh token — every other profile reading the
stale root grant subsequently died with refresh_token_reused / invalid_grant
once its access token expired.
This is the credential-pool analog of #43589 (which fixed the non-pool xAI
refresh path in _save_xai_oauth_tokens). Detect the read-from-root case
(profile lacks its own providers.<id> block) BEFORE the profile save and,
after it, write the rotated chain back to the global root via a best-effort,
seat-belted write-through. A profile that genuinely shadows root (owns the
block) is untouched; classic mode (profile == root) is a no-op; a failed root
write never breaks the profile's own save. Covers openai-codex (reported),
xai-oauth, and nous through the shared sync path.
Two-part fix:
Part 1 (classifier override at agent/error_classifier.py:720-738):
A transport disconnect on a reasoning model — even on a large session —
now routes to FailoverReason.timeout instead of context_overflow. Without
this, large-session reasoning-model disconnects route to the compression
branch and silently delete conversation history on a phantom
context-length error. The override is strictly targeted: non-reasoning
models (gpt-4o, claude-3-5-sonnet, llama-3.3-70b, etc.) still route to
context_overflow on large sessions — the existing intentional behavior
for chat models whose proxy doesn't idle-kill during prefill/generation.
Part 2 (new agent/thinking_timeout_guidance.py + integration at
agent/conversation_loop.py:3488-3567):
New is_thinking_timeout() and build_thinking_timeout_guidance() helpers.
When a known reasoning model (NVIDIA Nemotron 3 Ultra, OpenAI o1/o3,
Anthropic Opus 4.x thinking, DeepSeek R1, Qwen QwQ, xAI Grok reasoning)
hits a transport-kill on a small session (classifier says timeout
directly) or after Part 1 routes correctly (large session), the user
now sees reasoning-specific guidance with three actionable workarounds
in priority order:
1. Set providers.<provider>.models.<model>.stale_timeout_seconds: 900
in ~/.hermes/config.yaml (Hermes's built-in floor is already 600s
for known reasoning models; raise further if upstream is even
tighter).
2. Lower reasoning_budget or set reasoning_effort: medium on this
model if the provider supports it.
3. Use a smaller / faster reasoning model if the task doesn't
require deep thinking.
The new guidance takes precedence via if/elif over the existing
_is_stream_drop block, so a reasoning-model user with a transport-kill
message sees actionable advice instead of the misleading "try
execute_code with Python's open() for large files" advice (which is
correct for the unrelated large-file-write stream-drop case but
actively wrong for the thinking-timeout case).
Verified:
- 478 tests passing across 9 directly-relevant files (49 new + 429
existing, zero regressions).
- Ruff lint clean on all 4 modified/new files.
- Negative test: 6 parametrized regression guards confirm non-reasoning
models still route to context_overflow on large sessions; 4
parametrized gates confirm non-timeout classifier reasons never
trigger the guidance; 5 parametrized cases confirm non-transport
messages never trigger it.
- Regression guard: new guidance message does NOT contain
"execute_code" or "open()" — the misleading advice is fully
replaced, not appended alongside.
- Cross-vendor dual review via agy -p:
- Gemini 3.5 Flash (Medium) — passed: true, zero blockers, one
SHOULD-FIX (vprint block duplication — fixed by extracting
detection into a helper module).
- GPT-OSS 120B (Medium) — passed: true, zero blockers, two nits
(test placement — adopted at tests/agent/test_thinking_timeout_guidance.py;
primary-model capture — accepted as non-issue per Flash's nit).
Dependency note for maintainers:
This PR includes agent/reasoning_timeouts.py (the reasoning-model
allowlist module from PR #52238) because the Layer 1 override is
load-bearing on get_reasoning_stale_timeout_floor(). After PR #52238
lands on main, this PR's duplicate agent/reasoning_timeouts.py should
be rebased away. Either PR can land first; the other rebase is
mechanical.
Fixes#52271.
The salvaged context-window screen (#52392) skips fallback candidates that
are too small, and the rate-limit/403 fixes skip candidates that are at
capacity. A third hard failure remained uncovered: a fallback that builds a
client fine but returns a 400 because it structurally cannot run the model.
The canonical case is a configured openai-codex / ChatGPT-account fallback
asked to compress a glm-5.2 conversation:
400 - {'detail': "The 'glm-5.2' model is not supported when using
Codex with a ChatGPT account."}
This is a request-validation error, so should_fallback was False and the
explicit-provider gate blocked it — the auxiliary task (compression) aborted
every turn, dropping middle turns without a summary and churning the session,
which is exactly what destroys the prompt cache.
Adds _is_model_incompatible_error() (400 + capability phrasing, excluding
not-found and billing 400s which the sibling predicates own) and treats it as
a fallback-worthy capacity error in both sync and async call_llm, so the chain
skips the incapable route and continues to the next viable candidate.
The runtime auxiliary fallback chain (_try_configured_fallback_chain and
_try_main_fallback_chain) returned the first reachable candidate without
checking whether the candidate's context window was large enough for the
task. For task='compression' this meant a reachable but undersized
fallback (e.g. 32K) could be selected and then fail, even when a later
larger-context fallback was available.
This adds two small helpers:
_task_minimum_context_length(task)
Returns MINIMUM_CONTEXT_LENGTH (64K) for compression, None for
other tasks (vision, web_extract, etc.).
_candidate_context_window(provider, model, ...)
Thin wrapper around get_model_context_length that returns None on
probe failure so unknown/custom endpoints pass through unchanged
(preserves the existing fallback surface).
Both fallback loops now skip reachable candidates whose resolved context
is below the task minimum and continue iterating. The success path
(first viable candidate wins) is unchanged. Return shape and ordering
for healthy candidates are preserved.
Six regression tests cover:
L2 configured chain skips too-small candidate
L2 chain continues after skipping, returns last viable
L3 main chain skips too-small candidate
L4 unknown-context candidate passes through
L5 non-compression task is not filtered
L6 minimum constant matches MINIMUM_CONTEXT_LENGTH (64K)
3/6 fail on upstream/main without the production change (verified); all
6 pass with the fix. Full test_auxiliary_client.py suite (231 tests)
and related compression tests (130 tests) remain green.
When an explicit aux provider cannot build a client before any request is
sent (missing raw env key, exhausted/unavailable OAuth or credential-pool
auth, resolver returning (None, None)), call_llm raised a misleading
"no API key was found" error and bypassed the configured fallback_chain
entirely. A provider authenticated through Hermes auth / the credential
pool (e.g. ollama-cloud) whose pool entry is exhausted hit this path, so
compression failed instead of routing to the configured fallback.
Adds _try_configured_fallback_for_unavailable_client() and wires it into
both sync and async call_llm before the raise, and into the startup
compression feasibility check.
Salvaged from #51835 by @herbalizer404.
Rate-limit (429) errors on explicit-provider auxiliary tasks were
silently failing instead of triggering the fallback chain. The
is_capacity_error gate only checked payment and connection errors,
excluding rate limits — so when a configured provider like
openai-codex hit its rate limit, auxiliary tasks (kanban_decomposer,
vision, web_extract, approval, etc.) had zero resilience.
Add _is_rate_limit_error() to is_capacity_error at both call sites
(sync and async paths) so rate limits trigger fallback regardless
of whether the provider was auto-detected or explicitly configured.
Fixes#52228
Ollama Cloud (and similar) return 403 with bodies like "this model requires
a subscription, upgrade for access" or "you have reached your session usage
limit, upgrade for higher limits". These are capacity/billing conditions
semantically identical to credit exhaustion, but _is_payment_error() did not
recognize them (403 missing from the status set; keywords missing), so the
configured fallback_chain was never tried and compression failed outright.
Adds 403 to the status set and the subscription/session-usage keywords.
Salvaged from #49076 by @herbalizer404.
These 7 test sites assert rotation behavior (fork, child sessions, lock
contention, logging session-context follows id rotation, boundary hooks fire
on rotation). Pin each builder to in_place=False explicitly so they keep
exercising the retained rotation fallback regardless of the global default
(flipped to True in #38763). Rotation stays a working opt-out fallback and
deserves continued coverage — these are NOT deleted.
Pinned sites:
- test_compression_concurrent_fork._build_agent_with_db
- test_compression_logging_session_context._build_agent_with_db
- test_compression_rotation_state._build_agent_with_db
- test_compression_boundary_hook._make_agent (2 helpers: CompressionBoundaryHook + SessionCompressEvent)
- test_compression_concurrent_sessions._build_agent_with_db
build_turn_context() created the DB session row via _ensure_db_session()
before the system prompt was restored/built, so a fresh API/gateway agent
carrying client-managed history inserted a row with system_prompt=NULL. That
tripped the misleading 'stored system prompt is null; rebuilding from scratch
... investigate the previous turn's write path' warning and a guaranteed
first-turn prefix cache miss. Move row creation to after _cached_system_prompt
is populated.
Verified live (OpenRouter + claude-sonnet-4.5): persistent-agent turns show
cache_read jumping to the full prefix on turn 2+ (write 24411 -> read 24411),
and the persisted system_prompt is non-NULL so fresh-agent restore keeps the
prefix cache warm.
Tests: turn-context ordering regression asserting _ensure_db_session runs
after _cached_system_prompt is populated.
/learn told the agent to fill the skill `author` field, and the system
prompt environment probe surfaces the OS login name (user=$(whoami) in
prompt_builder.py), so the model wrote the host username into published
SKILL.md frontmatter — a privacy leak the user never opted into, and
inconsistent run to run as the most-salient identity changed.
The /learn authoring prompt now sets `author` to the literal value
`Hermes` and explicitly forbids deriving it from the host environment
(OS/login user, git config, or any probeable identity). The skill names
itself as the tool that wrote it.
Closes#52368.
#48879 closed the tool-call sequence on interrupt inside finalize_turn so a
/stop after a tool no longer persists a `tool` tail that the next user message
turns into a `tool -> user` role-alternation violation (which strict providers
like Gemini/Claude react to by hallucinating a continuation and ignoring prior
context — what users see as "lost context after stop").
But the retry-wait, error-handling, and post-error retry-wait interrupt aborts
in conversation_loop return early and never reach finalize_turn, so they still
persisted and returned a raw `tool` tail. Interrupting during provider
backoff/rate-limiting (common under heavy work) hit exactly this path.
Extract the close into a shared close_interrupted_tool_sequence helper and apply
it at every interrupt abort (finalize_turn + the three early returns) so the
whole bug class is fixed, not just the one site.
The verify-on-stop guard (PRs #52296, #52297) defaulted ON for every
session, so on gateway messaging surfaces (Telegram, Discord, etc.) the
model complied with the nudge by writing a hermes-verify temp script and
emitting an ad-hoc verification summary, which the gateway delivered to
the end user as chat noise.
Resolve a surface-aware default instead. The DEFAULT_CONFIG value becomes
the sentinel "auto", which verify_on_stop_enabled() resolves to ON for
interactive coding surfaces (CLI, TUI, desktop) and programmatic callers,
and OFF for conversational messaging surfaces. The surface is read from
HERMES_SESSION_PLATFORM (what the gateway actually binds), with
HERMES_SESSION_SOURCE and HERMES_PLATFORM as fallbacks, matching the
sibling resolution in skill_commands.py and prompt_builder.py. An explicit
HERMES_VERIFY_ON_STOP env var or a boolean agent.verify_on_stop config
still overrides in either direction.
The passive evidence ledger and the call site are untouched.
The /learn authoring prompt taught a subset of the HARDLINE skill rules,
and stated the <=60-char description rule without making the model enforce
it — so generated descriptions overshot (up to 202 chars), which the
60-char system-prompt skill index then silently truncates.
- description: add the index-truncation rationale, a count-and-trim
self-check, and a good/bad length example so the model actually hits <=60.
- add platforms-gating rule (OS-bound primitives -> declare platforms:).
- add author-credits-human-first rule.
- round out the Hermes-tool framing with the full wrapped-tool mapping and
references/templates layout.
Closes#52367.
Move terminal/execute_code/read_file preview compaction into agent.display so CLI, gateway, and Ink TUI all inherit the same labels that desktop introduced in #52321.
The shared preview keeps raw args intact while trimming display-only shell plumbing (`cd`, pipe tails, banner/status echoes) and read_file line ranges. Desktop now prefers backend `context` for live rows and keeps its TypeScript fallback only for hydrated history.
The pet generation image-processing suite is deterministic but expensive enough
to blow the per-file CI timeout on Linux (140s), and it is not relevant to the
fast timeout PR's normal signal. Keep it available for manual validation, but do
not run it by default.
Set HERMES_RUN_SLOW_PET_TESTS=1 to enable the suite. The canonical test wrapper
now preserves that opt-in variable through its hermetic env.
Recurring cron jobs were prompt-cache-cold on every fire. session_id is
built as cron_<job_id>_<timestamp>, and the Codex/Responses transport used
session_id directly as prompt_cache_key — so the timestamp changed the cache
key on every run and the static prefix (agent identity + tool schemas) was
re-paid each tick.
Derive prompt_cache_key from a SHA-256 of the static prefix (instructions +
sorted tool schemas) instead. Repeated fires of the same job share one
content-addressed key (pck_<hash>) and reuse the warm prefix within the
provider's cache TTL. The key changes exactly when the prefix changes —
edit the job's prompt or toolset and it re-keys; leave it alone and it stays
stable.
session_id is left untouched for transcript isolation, log correlation, and
the Codex/xAI session-scope routing headers (session_id, x-client-request-id,
x-grok-conv-id) — those are the per-fire identity, not the cache key. Only the
prompt_cache_key body field (standard OpenAI/Codex path and the xAI extra_body
field) is content-addressed.
Closes#51395.
Co-authored-by: spiky02plateau <spiky02plateau@users.noreply.github.com>
Co-authored-by: JoaoMarcos44 <JoaoMarcos44@users.noreply.github.com>
OpenRouter/Nous image gen now runs a quality-first model chain by default:
attempt the highest-fidelity OpenAI image model first, then fall back to
Gemini 3 Pro Image when it's access-gated/unavailable/times out. An explicit
OPENROUTER_IMAGE_MODEL / config model override pins one model with no fallback.
Atlas validation rejects malformed model output instead of shipping it: adds a
per-state collapse guard (a single sliver/fragment row no longer passes because
other rows are healthy), on top of the existing postage-stamp + multi-pose
checks.
Desktop: pet-gen native notifications are now "global" (not tied to a chat
session), so a background generation started from the command center fires an
OS notification when the user is away even with no active session. Adds a
neutral "This can take up to 5 minutes." banner on step 1, and lets the
provider picker auto-size.
Tests updated/added for the OpenRouter fallback chain, the collapse guard, and
the global notification path.
Make verification closure the default coding behavior after landed file edits while keeping bounded retries and config/env switches for users who need to disable it.
Ship the final pet-generation UX polish (provider picker behavior, step-2 cancel flow, banner integration, and visual consistency) and make saturated-chroma background removal C-op driven so hatch processing no longer hammers the machine during long runs.
Closes#47707
Context engines and memory providers expose tool schemas via
get_tool_schemas(). agent_init.py wrapped each as
{"type":"function","function":_schema} without validating that
_schema carries a top-level name. A provider returning an entry already
in OpenAI tool form ({"type":"function","function":{...}}) was then
double-wrapped into a tool whose function has no name. Strict providers
(e.g. DeepSeek) reject the entire request with HTTP 400
'tools[N].function: missing field name', so one malformed schema
silently disables the whole toolset and breaks every turn. The schema
was also never added to valid_tool_names, so even lenient providers
could not call it.
Add a shared normalize_tool_schema() helper that unwraps an
already-wrapped entry and returns None for anything lacking a resolvable
string name. Wire it into the agent_init context-engine loop and all
three memory_manager surfaces (inject_memory_provider_tools,
add_provider routing index, get_all_tool_schemas), so a single bad
plugin schema is skipped with a warning instead of poisoning the
request.
Verification: 209 targeted agent/memory tests pass (incl. 9 new).
New tests assert the unwrap + skip-nameless behavior and fail without
the fix.
When delegate_task spawns a child agent with a different model/provider, the
child's init_agent loaded the plugin context-engine GLOBAL singleton by
reference (`_selected_engine = _candidate`) and then called update_model() on
it with the child's (smaller) context_length. Because parent and child shared
the same object, this mutated the PARENT's compressor: e.g. DeepSeek 1M ctx
silently dropped to 204800 and the compression threshold from 200K to 40K
after any delegate_task with a different model.
Deepcopy the singleton before assigning/mutating it (agent_init.py) so the
child gets its own instance and the parent's compressor is untouched.
Salvaged from #42452 by @liuhao1024 (authorship preserved). Added a
source-pin regression test that fails if the production line reverts to the
bare alias, plus an end-to-end test driving get_plugin_context_engine() and a
StubEngine.update_model() — the original PR's tests exercised copy.deepcopy in
isolation but did not guard the actual agent_init code path.
Closes#42449. Supersedes #42469, #42474 (same one-line fix, no test).
The preflight-compression gate only ran the (expensive) token estimate when
the message COUNT exceeded protect_first_n + protect_last_n + 1. A session
with a handful of very large messages never tripped the count condition, so
compression was never attempted and the turn eventually hit a hard
context-overflow error.
Add _should_run_preflight_estimate() with OR semantics: run the estimate when
either the message count exceeds the protected ranges (the historical gate)
OR a cheap char-based estimate already crosses the configured threshold. The
downstream estimate_request_tokens_rough() stays authoritative — this is only
a hint that decides whether to pay for the full estimate.
Salvaged from #27435 by @texhy (authorship preserved). Re-applied on current
main: the preflight gate moved from conversation_loop.py to turn_context.py
since the PR was opened, so the helper + gate are placed there; the test
imports the real MINIMUM_CONTEXT_LENGTH instead of a hardcoded literal.
Closes#27405.
Turn a text prompt into a petdex-spec spritesheet (8×9 grid of 192×208
cells), grounded so every animation row stays the same creature:
- orchestrate: base drafts (distinct variation nudges) → per-row grounded
generation → atlas compose; one image call per row, rows fan out in parallel.
- atlas: frame-perfect registration in normalize_cells — 1-D cross-correlation
of each frame's column-mass profile locks the body (robust to limbs/cape),
one shared per-state scale, bottom-anchored; plus alpha-hole repair, gutter
severing, and interior-seeded chroma-pocket clearing.
- prompts: pixel-art-by-default style hints + registration constraints.
- store: local pet write (register_local_pet), slugify/unique_slug,
export_pet, slug-realigning rename_pet, createdBy provenance.
When context compaction's summary generation fails, the compressor's default
path (abort_on_summary_failure=False) drops the middle window and inserts a
static 'summary unavailable' marker — destroying the compacted turns. #29559
reported the field impact: a Connection error at the compaction moment dropped
124->15 messages (110 lost) for a long browser-automation task; #25585 is the
same failure mode (failed summary commits a destructive compaction anyway).
compress() already has an EXCEPTION to the historical drop default: auth
failures (401/403) ALWAYS abort and preserve the session, because rotating into
a placeholder-summary child on a broken credential strands the user. A transient
network/connection error is the same situation in reverse: it WILL recover, and
retrying then is strictly better than discarding context for a momentary blip.
Extend the always-abort carve-out to terminal connection/network failures:
- new _last_summary_network_failure flag, set in _generate_summary's terminal
failure branch when _is_connection_error(e) (reached only after any main-model
fallback is exhausted), reset alongside the auth flag;
- compress() aborts when it's set (returns messages unchanged,
_last_compress_aborted=True), independent of abort_on_summary_failure;
- a network-specific operator warning (distinct from the auth + config-flag
messages).
Scoped to connection errors only: a generic 500/400 still takes the historical
fallback-drop path (test_non_auth_failure_still_uses_fallback_path stays green).
Tests: network-failure detection + abort-despite-flag-false, both mutation-checked
(removing the flag-set fails detection; removing the carve-out fails the abort).
Add two regression tests for the salvaged #48706 fix:
- login token exchange targets platform.claude.com first
- falls back to console.anthropic.com when the new host is unreachable
Also map the salvaged contributor's noreply email in release.py
AUTHOR_MAP (CI author-map gate).
Regression coverage for the synthetic-assistant close: interrupt after a
successful tool must persist an assistant tail (placeholder when no
delivered text), real delivered text is preserved, and non-interrupted
or non-tool tails are left untouched.
deliver=origin (or omitted) from a TUI or classic-CLI session produces a
job with origin=null, because those sessions never populate the
HERMES_SESSION_PLATFORM/CHAT_ID context vars that _origin_from_env reads.
The scheduler then resolves no delivery target and skips delivery — the
job runs and saves output to last_output, but nothing reaches the user
and they only find out by polling cronjob(action='list') (#51568).
This is by design (local sessions have no live-delivery channel), so the
fix surfaces it instead of silently dropping the intent:
- cronjob create now appends an informational notice to its result when
a created job resolves to zero delivery targets and the user did not
explicitly ask for deliver='local'. The check uses the scheduler's own
_resolve_delivery_targets so it accounts for origin, home channels,
'all', and explicit platform targets — no false positives.
- PLATFORM_HINTS gains a 'tui' entry (the TUI had none) and the 'cli'
hint now states that cron jobs from these sessions are local-only and
that deliver must target a gateway-connected platform to notify the
user. This stops the agent promising a delivery that never happens.
No scheduler/delivery behavior change; no new env var; cron isolation
invariant untouched.
Open-ended skill learning across every surface. /learn <free text> takes a
description of any source — a directory, a URL, the workflow you just walked
the agent through, or pasted notes — and the live agent gathers it with the
tools it already has (read_file/search_files, web_extract, the conversation,
the pasted text), then authors a SKILL.md via skill_manage following the
house authoring standards (<=60-char description, the standard section order,
Hermes-tool framing, no invented commands).
No engine, no model-tool footprint, works on any terminal backend (local,
Docker, remote): /learn builds a standards-guided prompt and hands it to the
agent as a normal turn.
- agent/learn_prompt.py: shared standards-guided prompt builder
- /learn registry entry (both surfaces) + CLI handler (inject onto input
queue) + gateway handler (rewrite turn, fall through, /blueprint pattern)
- tui_gateway command.dispatch returns a send directive -> TUI + dashboard chat
- dashboard Skills page 'Learn a skill' panel (dir + URL + open-ended text)
composes a /learn request and runs it in chat
- docs (slash-commands ref + skills feature page), 11 targeted tests
Inspired by OpenAI Codex's Record & Replay and the /learn concept from #47234
(dir-distillation engine); reworked to be open-ended and engine-free per
review.
A "one-shot" is a single stateless model call that runs OUTSIDE any conversation:
it never touches session history, never breaks prompt caching, and returns plain
text. UI surfaces need this for small generative chores — a commit message from a
diff, a rename suggestion, a summary — where an agent turn would pollute the
thread and hand-rolling an LLM call at every call site would be worse.
- `agent/oneshot.py`: `run_oneshot(...)` over the existing auxiliary-client
plumbing (same path as title generation). Two call shapes: explicit
instructions/input, or a registered `template` + `variables` (templates own the
prompt engineering so it stays consistent across CLI/TUI/desktop). Ships a
`commit_message` template. Model selection inherits the live session via
`main_runtime`, else the configured aux `task` backend.
- `tui_gateway/server.py`: `llm.oneshot` RPC (long-handler) inheriting the
session's model when `session_id` resolves.
Stateless by construction — no session mutation, cache untouched.
Follow-up to the coding-context posture (#43316): that PR detects each repo's
verify loop (manifests, package manager, exact test/lint/build commands, context
files) and bakes it into the system-prompt snapshot — but only as a string, for
the model. Non-prompt consumers (the desktop verify UI) had no way to read it
without re-sniffing and drifting from the prompt.
Split detection from rendering, keeping one source of truth:
- `detect_project_facts(root) -> ProjectFacts` (frozen) holds the structured
facts; `_project_facts()` now renders it into the same snapshot lines, so the
prompt block stays byte-identical (cache-safe).
- `project_facts_for(cwd)` resolves the workspace root (git, else marker) and
returns the structured facts, or None outside a workspace.
- `project.facts` gateway RPC surfaces it to any client (desktop/TUI/ACP).
Tests assert the structured output and that the UI-facing commands never drift
from what the prompt block renders (one detector feeds both).
The success/staged gating and op-expansion for mirroring built-in memory
writes to external providers lived in a standalone agent/memory_write_bridge.py
helper called inline from two core call sites (tool_executor.py,
agent_runtime_helpers.py). That left the mirror decision-making in the agent
loop, outside the memory-provider interface.
Fold it into a new MemoryManager.notify_memory_tool_write() entry point: the
loop now hands over the raw tool result + args and a metadata callback, and the
manager decides whether/what to mirror. Both core call sites collapse to a
single call; the orphan module is removed. No MemoryProvider ABC change.
Tests rewritten as behavior tests against the manager method.
Mirror built-in memory writes to external providers only after the native memory tool succeeds and is not staged for approval. Keep OpenViking's built-in memory mirroring add-only, since Hermes native memory entries do not yet have stable OpenViking file URIs for replace/remove.
Add a narrow viking_forget tool for exact user memory file deletion and document the current OpenViking write/delete behavior.
The compaction trigger compared estimated input against context_length *
threshold, but the provider reserves max_tokens of OUTPUT out of the same
window. With a large max_tokens (e.g. 65536 on a custom provider) the usable
input budget is materially smaller than the raw window, so sessions hit a
provider 400 before compaction ever fired.
_compute_threshold_tokens now subtracts the output reservation
(context_length - max_tokens) before applying the percentage and the
small-window 85% guard. max_tokens is stored on the compressor (threaded from
agent.max_tokens at construction) and reused across update_model() switches;
None = provider default = no reservation (full-window behavior, unchanged).
Reimplemented on the current _compute_threshold_tokens surface (the inline
threshold calc the original PR targeted was since refactored for the
small-window #14690 fix); composes with that 85% guard on the effective budget.
Credit: @kyssta-exe (#43651) — original design for the output-token
reservation in the compaction threshold.
Closes#43547.
After a compaction, the post-compression path parks last_prompt_tokens=-1 and
sets awaiting_real_usage_after_compression=True, but last_real_prompt_tokens
still holds the stale pre-compression value (above threshold). should_defer_
preflight_to_real_usage() hit the 'last_real_prompt_tokens >= threshold => False'
short-circuit and let preflight fire a SECOND compaction before the provider
reported real post-compaction usage. Add an early-return on the awaiting flag so
deferral holds for exactly one turn; update_from_response() clears it.
The flag-setting half (#36718) already landed on main via the in-place
compaction path (conversation_compression.py); this adds the missing
should_defer guard that consumes it.
Credit:
- @ashishpatel26 (#38133) — diagnosis + the should_defer early-return design
- @Tranquil-Flow (#36769) — same #36718 fix, identical guard placement
Closes#36718.
The tail-protection budget walks estimated an assistant message's tokens from content + function.arguments only, dropping each tool_call's id, type and function.name (plus JSON structure). Assistant turns that fan out into parallel tool calls were undercounted by 2-15x (a 4-tool-call turn measures ~73 vs ~1,090 real tokens), so the protected tail overshot tail_token_budget and compression ran far below its intended ratio — context kept growing.
Consolidate the three duplicated budget walks (_prune_old_tool_results and the two passes in _find_tail_cut_by_tokens) into a single _estimate_msg_budget_tokens() helper that counts the full tool_call envelope via len(str(tc)), consistent with how _estimate_message_chars estimates message size elsewhere.
Tested on Windows: new tests/agent/test_compressor_tool_call_budget.py plus the existing compression suite (test_context_compressor, compressor_image_tokens, cross_session_guard, infinite_compaction_loop) — 209 passed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Follow-up to the salvaged preflight token-progress fix: require a material
(>5%) token reduction to count as progress, matching the overflow-handler
retry path (conversation_loop.py, #39550), so a sub-5% wobble can't keep the
3-pass preflight loop spinning. Adds boundary + zero-token regression tests.
/simplify-code QUALITY finding: the `if callable(_available_entries): ... else:
pool.select()` ladder was dead for the real CredentialPool type (`_available_entries`
is always a bound method) AND the select() fallback violated the helper's read-only
contract — select() -> _select_unlocked() runs _available_entries(clear_expired=True,
refresh=True), which persists to auth.json and triggers a network refresh. Call
_available_entries(clear_expired=False, refresh=False) directly inside the existing
try/except instead.
Also drops the now-dead `select=` stubs from the 6 pool tests (they only existed to
satisfy the removed fallback branch). Behavior unchanged; 6 pool tests pass and the
read-only / null-token contract tests were mutation-checked (flipping the flags /
removing the None-guard fails the respective test).
Rebased onto god-file Phase 1 refactor — preflight compression has moved
from agent/conversation_loop.py to agent/turn_context.py (no semantic
change in the refactor itself; the bug below was carried over verbatim).
The preflight compression loop in ``turn_context.py`` uses
``len(messages) >= _orig_len`` to decide whether a compression pass has
made progress. That conflates two different conditions: a true no-op
(transcript materially unchanged) and effective token compression that
summarises message contents but keeps the same number of rows. The
second case is misread as "Cannot compress further" — the session then
surfaces ``Context length exceeded`` and auto-resets even when the
post-compression estimate is far below the model context window.
Observed example from #39548: a Telegram session on GPT-5.5 with a 1M
context dropped from ~288k → ~183k tokens (a 36% reduction) while
preserving 220 messages. The loop treats that as exhaustion and the
gateway auto-resets the session.
Fix
---
Add ``_compression_made_progress(orig_len, new_len, orig_tokens, new_tokens)``
and call it after the post-pass ``estimate_request_tokens_rough`` (which
is moved up to run *before* the progress check instead of after it).
Either a row-count reduction OR a token-count reduction now counts as
progress; only when neither moves do we break out as "stuck".
Fixes#39548
* feat(providers): remove google-gemini-cli + google-antigravity OAuth providers
Google now actively bans accounts for third-party tools that piggyback on
Gemini CLI / Antigravity / Code Assist OAuth, and because abuse prevention
sits at a backend layer the ban can extend to the entire Google account
(Gmail/Drive), with a second violation being permanent.
Ref: https://github.com/google-gemini/gemini-cli/discussions/20632
Removes both OAuth inference providers entirely (modules, provider profiles,
auth/runtime/config/models wiring, the /gquota Code Assist quota command,
the antigravity-cli optional skill, desktop + docs surface in en + zh-Hans).
The API-key 'gemini' provider (GOOGLE_API_KEY/GEMINI_API_KEY against
generativelanguage.googleapis.com) is unaffected and stays fully supported.
* fix(skills): keep the antigravity-cli skill — only the OAuth provider is removed
The antigravity-cli optional skill orchestrates the external `agy` binary as
a coding-agent tool via the terminal tool — it does NOT wrap Hermes inference
through the banned google-antigravity OAuth provider, so it carries none of
the account-ban risk that motivated removing that provider. Restore the skill,
its docs page, the sidebar entry, and the optional-skills catalog row. The
google-antigravity / google-gemini-cli inference providers stay fully removed.
Salvage follow-up on top of @pmos69's #29474. The PR resolved the
Antigravity OAuth client purely by discovering it from an installed `agy`
binary or HERMES_ANTIGRAVITY_CLIENT_ID/SECRET env vars, so users without
agy installed hit a hard 'client ID not available' error.
Antigravity's desktop OAuth client is a public, non-confidential installed-app
client (PKCE provides the security), baked into every copy of the Antigravity
CLI — same posture as the gemini-cli credentials Hermes already ships in
google_oauth.py. Bake it in as the final fallback (env -> discovery -> public
default) and add the public default Code Assist project as the discovery
fallback, matching the reference Antigravity flow. Now consumers can
authenticate directly without agy installed.
Secret redaction only matched `Authorization: Bearer <token>`. Other auth
headers passed through verbatim into logs, tool output, and transcripts:
- `Authorization: Basic <base64>` — leaks base64(user:password)
- `Authorization: token <pat>` / any non-Bearer scheme
- `Proxy-Authorization: ...`
- `x-api-key: <key>` (Anthropic and many providers) and `api-key`,
`x-goog-api-key`, `x-auth-token`, `x-access-token`, ... — opaque values with
no known vendor prefix were caught by nothing
A logged request or an echoed `curl -H "x-api-key: ..."` command therefore
leaked live credentials.
Generalize the Authorization rule to mask the credential for any scheme (and
Proxy-Authorization) while preserving the header name and scheme word for
debuggability, and add an api-key header rule for the single-opaque-value
headers. Bearer behavior is unchanged; plain prose containing the word
"authorization" (no colon-delimited value) is left untouched.
Adds regression tests for Basic/token/Proxy auth and the x-api-key/api-key
headers, including inside a curl command.
Bedrock Claude routes through the AnthropicBedrock SDK and injects
cache_control, so cached tokens are always reported — but the pricing
table had no cache cost fields for any Bedrock model, so /usage showed
"cost unknown" on every cached session. Also, cross-region inference
profiles (us./global./eu. prefixes) never matched the bare pricing keys.
- Add cache_read/cache_write rates to the four Bedrock Claude rows
(read 0.1x input, write 1.25x input per the Bedrock pricing page).
- Normalize the cross-region prefix in the Bedrock pricing lookup,
mirroring is_anthropic_bedrock_model's prefix list.
Closes#50295.
The 'Session compressed N times — accuracy may degrade' warning went
through _vprint (CLI stdout only), so the Ink TUI / Telegram / Discord
never saw it — unlike the two other compression warnings in the same
module, which route through _emit_status (and store _compression_warning
for late-bound gateway status_callback replay).
Set agent._compression_warning + call agent._emit_status() for this
warning too, matching the sibling pattern. _emit_status still _vprints
for the CLI, so CLI output is unchanged; TUI / gateway surfaces now
receive it via status_callback (and replay_compression_warning can
re-deliver it once a late-bound gateway callback is wired).
Co-authored-by: liuhao1024 <sunsky.lau@gmail.com>
When an OpenAI-compatible proxy (e.g. cmkey.cn, one-api Anthropic channels)
returns a well-formed HTTP 200 whose summary content is null or empty/
whitespace-only, _generate_summary coerced it to "" and stored a prefix-only
summary — silently replacing the compacted turns with nothing. The model then
lost all in-progress context after compression (#11978, #11914).
_validate_llm_response already guards None / empty-choices, so those never
reach the compressor; the gap was a well-formed response with empty *content*.
Now treat empty content as a summary failure: raise so it routes through the
existing main-model fallback then transient cooldown, dropping the turns
without a summary rather than wiping context with an empty one.
Also narrow the bare 'except RuntimeError' so only genuine 'No LLM provider
configured' errors take the 600s no-provider cooldown; empty/invalid-response
RuntimeErrors from a configured provider now correctly get the main-model
fallback instead of being misrouted into the long no-provider cooldown.
Reported by @Hung2124; area identified by @annguyenNous in #39590.
load_pool() is meant to be a read, but it persistently pruned env-seeded
pool entries whenever the calling process's os.environ lacked the seeding
var. A process without MINIMAX_API_KEY would delete the persisted
env:MINIMAX_API_KEY entry from auth.json for every other process, causing
auth.json to oscillate and auxiliary auto-detect to fall through to the
wrong provider.
env:* entries are persisted references re-hydrated from the environment on
each load — a missing var means "cannot re-seed right now", not "source is
gone forever". _prune_stale_seeded_entries now gates env-source removal
behind prune_env_sources (default True for explicit cleanup paths);
load_pool() passes prune_env_sources=False. File-backed singletons
(device-code OAuth, hermes_pkce) still prune when their backing file is
gone, and explicit removal via `hermes auth remove` (source suppression)
is unaffected.
Fixes#9331.
Co-authored-by: houko <suzukaze.haduki@gmail.com>
The compaction threshold is max(context_length * threshold_percent,
MINIMUM_CONTEXT_LENGTH=64000). The floor prevents premature compression on
large models, but degenerates at small windows: a model at exactly 64000
ctx gets max(32000, 64000) = 64000 — a threshold equal to the ENTIRE
window. should_compress() can then never fire, because the provider
rejects the request before usage reaches 100%. Auto-compression silently
never triggers for any model whose context_length <= MINIMUM /
threshold_percent (e.g. 64K-per-slot local models).
Centralize the calc in _compute_threshold_tokens(). When the floor would
meet or exceed the context window, trigger at 85% of the window
(_MIN_CTX_TRIGGER_RATIO) — high enough that a minimum-context model uses
most of its budget before compacting (compacting at the 50% percentage
would waste half the small window), but below 100% so compaction actually
fires before the provider rejects the request. This mirrors the existing
gpt-5.5/Codex 85% autoraise rationale. Large-context behavior (floor at
64000) is unchanged; both call sites (__init__ and update_model) use the
shared helper.
Co-authored-by: soynchux <soynchuux@gmail.com>
Co-authored-by: LeonSGP43 <154585401+LeonSGP43@users.noreply.github.com>
Co-authored-by: Tranquil-Flow <tranquil_flow@protonmail.com>
When a turn hit max_iterations, finalize_turn ran three unguarded cleanup
steps after the model's summary — _save_trajectory (file I/O), _cleanup_task_resources
(remote VM/browser teardown), and _persist_session (SQLite write). Any raise
there propagated out of run_conversation, discarding the partial final_response
the caller was waiting for; subprocess wrappers saw an empty stdout with no
traceback (#8049).
Each step is now guarded independently so one failure can't skip the others.
Failures log at ERROR with a traceback and are surfaced on the result dict via
cleanup_errors; the partial response is always returned.
Closes#8049.
ContextCompressor.update_model() recomputed context_length/threshold/budgets
but kept the cross-call calibration state (last_real_prompt_tokens,
last_rough_tokens_when_real_prompt_fit, last_compression_rough_tokens,
awaiting_real_usage_after_compression, _ineffective_compression_count) from the
PREVIOUS model.
Those fields encode 'the provider proved this prompt fit' / 'preflight can be
deferred' decisions valid only for the model that produced them. Carried across
a switch to a smaller-context model, should_defer_preflight_to_real_usage() used
the old model's 'it fit' history to SKIP a preflight compression the new model
actually needed — sending an oversized prompt the provider rejects (#23767).
update_model() now clears that state; the new model's first response repopulates
it via update_from_response(). Verified E2E: after a 200K->65,536 switch, defer
no longer suppresses and should_compress fires on an over-threshold estimate.
protect_first_n keeps the first N non-system messages verbatim through
compaction so the original task framing survives. But it was applied on
EVERY compression pass: the same early user turns were re-copied into each
child session and never summarized away, so across a long, repeatedly-
compressed session those old messages became immortal and grew the
protected head unboundedly (#11996, P1).
Decay it: protect_first_n applies on the FIRST compaction only. Once the
session has been compressed at least once (compression_count >= 1, or a
handoff summary already exists), the early turns are captured in the
summary, so _effective_protect_first_n() returns 0 and only the system
prompt stays protected. The decay is read at compress_start computation
time, before compression_count/_previous_summary are mutated at the end of
compress(), so the first pass still protects correctly.
Co-authored-by: truenorth-lj <liliangjya@gmail.com>
Co-authored-by: davidvv <david.vv@icloud.com>
When LLM summarization fails, the deterministic fallback summary rendered
the latest user ask (active_task = "User asked: '<ask>'") verbatim under
THREE headings — Historical Task Snapshot, Historical In-Progress State,
and Historical Pending User Asks. Re-presenting an already-handled ask as
unresolved in-progress/pending work made the model re-answer it AND treat
the resurrected ask as the active turn, burying the genuinely-new
post-compaction user message (#49307: answer repetition + new-instruction
loss, P1).
Keep the latest ask once, under Task Snapshot, as historical context only.
The In-Progress and Pending-Asks sections now say 'Unknown / None
recoverable from deterministic fallback' (consistent with the Active
State / Key Decisions / Resolved Questions sections) and explicitly note
the ask is historical, not outstanding. The raw turn text still appears in
the verbatim 'Last Dropped Turns' transcript — that's the dropped-turn
record, not a re-labeled instruction.
Note: the separate role=assistant standalone-summary regurgitation
(#33256) is left as-is — that role choice is constrained by strict message
alternation (user collides with a user-ending head) and is already
mitigated by the summary end-marker; forcing the role would risk the
alternation invariant.
Co-authored-by: r266-tech <r2668940489@gmail.com>
Co-authored-by: kyssta-exe <kyssta-exe@users.noreply.github.com>
Context compression is atomic, but a gateway interrupt (an incoming user
message while the agent is busy) could abort the in-flight summary call.
The Codex Responses aux stream polls the thread interrupt flag and raised
InterruptedError unconditionally — so compression fell back to a degraded
static 'summary unavailable' marker, losing the real handoff (#23975).
Add a thread-local interrupt-protection flag (aux_interrupt_protection
context manager) in auxiliary_client; the Codex stream's cancellation
check honors it. The compressor wraps its summary call_llm in the context
manager. Timeouts still fire (a hung call must die) and all other aux
tasks (vision, web_extract, title_generation, …) stay interruptible.
Re-entrant, so the main-model retry recursion is safe.
Co-authored-by: konsisumer <der@konsi.org>
Three state-loss bugs at the compression rotation boundary, fixed together
because they all live in the same ~80-line rotation block:
- #33618: a persistent /goal did not follow the rotation. load_goal does a
flat per-session lookup with no lineage walk, so a goal silently died when
compression minted a fresh child id. Added migrate_goal_to_session() and
call it after the child session is created (move-not-copy: the parent row
is archived as cleared so exactly one active goal row exists).
- #33906/#33907: if the child create_session raised (FK constraint,
contended write), the outer handler only warned and let the agent continue
on the NEW id — which has no row in state.db — producing an orphan session.
Now the rotation rolls agent.session_id back to the still-indexed parent
(reopening it) instead of stranding the conversation on a phantom id.
- #27633: the compaction-boundary on_session_start notification omitted the
platform kwarg, so context-engine plugins saw source=unknown for every
message after the boundary. Forward platform (matching the initial
session-start call in agent_init.py).
Co-authored-by: denisqq <21260182+denisqq@users.noreply.github.com>
Co-authored-by: zccyman <16263913+zccyman@users.noreply.github.com>
Co-authored-by: liuhao1024 <sunsky.lau@gmail.com>
Add a regression test for #47868 asserting convert_messages strips the
internal per-message timestamp field, plus the identity-return path for
timestamp-free message lists. Map x7peeps for the release attribution gate.
Add the shared pet engine under agent/pet/: spritesheet manifest loading
and in-process caching, six-state animation model, frame rendering, and
the persistent pet store. Register the display.pet config block (pet,
scale, enabled, etc.) that every surface reads from. Covered by
tests/agent/test_pet_engine.py.
When the auxiliary summary call fails with an authentication/permission
error (HTTP 401/403), context compression now ABORTS and preserves the
session unchanged instead of rotating into a child session with a
placeholder summary.
Before: a 401 (invalid/blocked key, or a token pointed at the wrong
inference host) fell through every transient-error check to 'return
None', and because compression.abort_on_summary_failure defaults False,
compress() took the static-fallback path and rotated the session anyway
(messages N->N). The user landed on a fresh-but-broken session that kept
failing the same way — paying for a full-context API call each turn with
no useful compression.
After: _generate_summary classifies 401/403 as a non-recoverable auth
failure (_last_summary_auth_failure) and compress() aborts on it
regardless of abort_on_summary_failure. A distinct auxiliary summary_model
that 401s still retries once on the main model first (its dedicated creds
may be the only broken thing); the abort only sticks when the main model
itself auth-fails or the fallback also auth-fails. The existing
_last_compress_aborted handling in conversation_compression.py already
skips rotation and emits a warning, so no session rotation occurs.
Tests: TestAuthFailureAborts — 401/403 flagging, compress() aborts despite
flag=False, non-auth failures keep the historical fallback path, and
aux-model auth failure recovers on main without aborting.
When the active provider returns a 401/403 that survives its per-provider
credential-refresh attempt (revoked OAuth, blocked/expired key, or an
account pinned to a dead/staging inference endpoint), the conversation
loop now escalates to the configured fallback chain instead of dead-ending.
Before: the generic failover dispatch fired only for {rate_limit, billing};
auth/auth_permanent fell through to 'switch providers manually' advice and
never called _try_activate_fallback(). A user whose primary credential was
broken kept thrashing on the same dead credential every turn — the main
agent appeared 'stuck in fallback mode' while never actually failing over.
This also affected auxiliary tasks (compression, vision, title-gen), since
auto-resolved aux follows the main provider.
After: a persistent auth failure with a configured fallback chain switches
to the next provider (mirroring the rate-limit/billing failover path),
guarded one-shot per attempt by TurnRetryState.auth_failover_attempted.
When no fallback is configured the behavior is unchanged — it falls through
to the existing terminal handling and provider-specific troubleshooting
guidance.
Tests: test_auth_provider_failover.py — 401/403 classify as auth, the
gating condition fires only with a chain present + guard unset, the guard
blocks repeats, and non-auth (500) errors do not trigger auth failover.
The session-stable system prompt embeds Model:/Provider: identity lines,
but mid-turn failover (try_activate_fallback) swaps the runtime without
touching them, so a fallback model misreports itself as the primary when
asked "what model are you?".
rewrite_prompt_model_identity() rewrites the last occurrence of each line
on _cached_system_prompt when a fallback activates (and back on restore,
byte-identical so the primary's prefix cache still hits). The rewrite is
never persisted to the session DB. _sync_failover_system_message() patches
the in-flight api_messages[0] at all 8 failover sites so the current turn
ships the corrected identity. Cache-safe: the fallback's prefix cache is
cold on a model switch anyway.
Co-authored-by: Hermes Agent <noreply@nousresearch.com>
Second review pass (Codex + Hermes subagent). Codex reproduced a real race with
a two-thread harness; both converged on the remaining issues.
- Generation-aware publish (fixes a lost-update race): two refresh callers (the
late-refresh daemon and the between-turns prologue around turn 1) could each
compute a snapshot outside the lock; a SLOWER caller holding an OLDER registry
generation could acquire the publish lock after a newer caller and clobber it,
deleting just-landed tools. refresh_agent_mcp_tools now captures
registry._generation before computing and refuses to publish a stale set;
agent._tool_snapshot_generation tracks the published generation.
- Context-engine routing names (_context_engine_tool_names) are now staged on a
local and published atomically with the snapshot, and only claimed when this
rebuild actually appended the schema — matching agent_init's dedup so a
registry/plugin tool of the same name keeps its own dispatch. (Previously
mutated live, before the publish lock, and on no-change refreshes.)
- CLI /reload-mcp: self.enabled_toolsets is resolved once at startup, so a
server newly ENABLED in config mid-session wasn't picked up (TUI already
re-resolved). Merge now-connected MCP server names into the override (unless
the user pinned all/*), mirroring startup, and keep self.enabled_toolsets in
sync. Closes the CLI/TUI parity hole.
- ACP (acp_adapter/server.py) routed through the shared helper — it was a 5th
sibling rebuild that re-injected memory tools but NOT context-engine tools and
bypassed the atomic/name-diff path (inert today, fragile).
- mcp_startup._resolve_discovery_timeout pulls its default from DEFAULT_CONFIG
(single source of truth) instead of a stale hardcoded 5.0 literal.
- Tests: stale-generation-no-clobber, _skip_mcp_refresh honored, timeout
fallback uses DEFAULT_CONFIG.
A slow MCP server (HTTP/OAuth, 2-6s cold connect) that finishes connecting
after the agent's one-time tool snapshot was uncallable for the rest of the
session. The merged pre-first-turn late-refresh only helps during the dead air
before the user's first keystroke; once a turn starts it bails to protect the
prompt cache, so a user who types before the server connects never gets the
tools without a manual /reload-mcp.
Refresh the snapshot in the per-turn prologue (build_turn_context), before this
turn's first API call assembles tools=. This is cache-safe by construction: the
refresh only ever extends a fresh request prefix at a turn boundary, never
mutates the cached prefix of an in-flight turn. So late tools become callable on
the user's NEXT turn automatically, with no /reload-mcp and no cache cost.
- tools/mcp_tool.py: has_registered_mcp_tools() — cheap guard so sessions with
no MCP servers (the common case) skip the rebuild entirely.
- agent/turn_context.py: call the shared refresh_agent_mcp_tools() helper at the
top of the prologue when MCP servers are registered.
- tests: 3 contract tests through the real build_turn_context (adds late tool;
skipped when no servers; no snapshot churn when unchanged).
.hermes/plans/: SPEC + PLAN documenting the root cause, the cache-safety
constraint, and why the existing fixes (#48403/#41630/#42802) don't close it.
The credential gate. When multiplexing is active, a profile's secrets resolve
from a context-local scope, never the process-global os.environ (which in a
multiplexer may hold another profile's keys, and is inherited by every
subprocess spawned with env=dict(os.environ)).
- agent/secret_scope.py: get_secret() backed by a secret-scope contextvar.
FAIL-CLOSED: when multiplex is active and no scope is installed, an unscoped
read RAISES UnscopedSecretError instead of falling back to os.environ — a
missed/new call site crashes loudly at that line rather than leaking a
cross-profile value. Genuinely-global vars (HERMES_*, PATH, kanban paths,
…) keep reading os.environ via an allowlist. load_env_file/build_profile_
secret_scope parse a profile .env into an isolated dict WITHOUT mutating
os.environ. Off by default => transparent os.getenv behavior.
- hermes_cli/runtime_provider.py: all credential/provider/base-url reads go
through _getenv -> get_secret.
- agent/credential_pool.py: env fallbacks route through get_secret (the
~/.hermes/.env-first preference is preserved and already profile-correct via
the home override).
- tools/mcp_tool.py: MCP config interpolation resolves through
get_secret, so a server's picks up the routed profile's value.
- gateway/run.py: set_multiplex_active() at GatewayRunner init; per-turn .env
reload is a no-op for credentials in multiplex mode (secrets come from the
scope, not global env); _profile_runtime_scope context manager combines the
HERMES_HOME override + secret scope; _run_agent wraps _run_agent_inner in
that scope (resolved via _resolve_profile_home_for_source) when multiplexing.
Propagates into the agent worker thread for free via the existing
copy_context() in _run_in_executor_with_context.
Tests: 13 unit (fail-closed, scope isolation, global allowlist, .env parsing
without environ mutation) + 7 E2E (runtime_provider + MCP interpolation prove
two profiles isolated, unscoped read raises, globals still read environ).
Add platform_hints config so an admin can append to or replace Hermes'
built-in platform hint for a single messaging platform (WhatsApp, Slack,
Telegram, ...) without affecting other platforms. Enables enterprise
managed profiles to steer platform-aware skills (e.g. invoke a custom
table-formatting skill on WhatsApp where Markdown tables don't render)
while leaving Telegram/Slack/CLI behavior unchanged.
- hermes_cli/config.py: document platform_hints in DEFAULT_CONFIG
- agent/agent_init.py: load platform_hints -> agent._platform_hint_overrides
- agent/system_prompt.py: _resolve_platform_hint() applies append/replace
(replace wins; bare string = append shorthand); defensive on bad config
- tests: 16 cases covering append/replace/shorthand/isolation/malformed
Override only affects the platform-hint segment of the system prompt;
SOUL/context/memory tiers and general instructions are unchanged.
* feat(billing): nous_billing http client + BillingState core (phase 2b)
Phase 2b terminal-billing client foundation:
- hermes_cli/nous_billing.py: typed client for the 4 /api/billing/* endpoints
(state/charge/poll/auto-top-up). Raises typed errors (BillingScopeRequired,
BillingRateLimited, BillingAuthError) mapped from the live-verified contract;
fail-open is the caller's job. Idempotency-Key enforced client-side.
- agent/billing_view.py: surface-agnostic BillingState core + Decimal money
parsing (server emits decimal strings, not 2dp), fail-open builder,
idempotency-key gen, custom-amount validation.
- 51 unit tests (decimal parse/format, payload tiering, error->exception
matrix, fail-open, amount validation).
Plan: docs/plans/2026-06-13-001-phase-2b-terminal-billing-tui-plan.md
* feat(billing): billing:manage scope + lazy step-up re-auth (phase 2b)
- NOUS_BILLING_MANAGE_SCOPE constant.
- nous_token_has_billing_scope(): split-based scope check (no false-positive
substring match).
- step_up_nous_billing_scope(): re-runs the device flow requesting
billing:manage, reusing the held credential's portal/inference URLs + client_id
(so a preview stays a preview), persists like _login_nous but WITHOUT the model
picker. Returns True iff the minted token carries the scope (False when NAS
silently downscopes a non-admin / unticked grant).
Lazy step-up (plan D-A): normal login path unchanged; 403 insufficient_scope
from a billing call triggers this. 7 unit tests.
* feat(billing): billing JSON-RPC methods for the TUI (phase 2b)
billing.state / charge / charge_status / auto_reload / step_up in
tui_gateway/server.py. Return STRUCTURED success envelopes (result.ok +
result.error=<code>) rather than JSON-RPC-level errors, so the Ink rpc() promise
always resolves and the TUI branches on the typed billing error code
(insufficient_scope, rate_limited, no_payment_method, …) to render the right
affordance. Money serialized as decimal STRINGS + display strings. charge mints
+ echoes an idempotency_key for retry reuse. 16 unit tests.
* feat(billing): /billing CLI handler + command registry (phase 2b)
- CommandDef("billing", subcommands=buy|auto-reload|limit), added to
_SLACK_VIA_HERMES_ONLY so it routes via /hermes on Slack (keeps the 50-cap
parity test green, same as /credits).
- cli.py::_show_billing + screen helpers: all 5 screens (overview, buy→confirm→
poll, auto-reload, monthly-limit read-only). Reuses _prompt_text_input_modal /
_prompt_text_input (D-C). Non-interactive (_app is None) renders text + portal
deep-link, never prompts (R7). Decimal money end-to-end. 2s/5-min cancellable
poll loop; 429/503 = retry not failure; settled = ledger truth. Lazy step-up on
403 insufficient_scope. no_payment_method treated as mainline funnel-to-portal.
- 6 CLI tests; 156 command tests (incl. Slack/Telegram parity) green.
* feat(billing): /billing Ink TUI screens + tests (phase 2b)
- ui-tui/src/app/slash/commands/billing.ts: /billing TUI command covering all 5
screens — overview (text), buy <amt> → ConfirmReq → charge → non-blocking 2s/
5-min poll loop → settled/failed/timeout branches, auto-reload <below> <to> →
ConfirmReq → PATCH, limit (read-only). Reuses the existing ConfirmReq overlay
(D-C) — no bespoke component. Typed-error envelope branching: insufficient_scope
arms the lazy step-up confirm; no_payment_method/rate_limited/cap funnel to
portal. Client-side amount validation mirrors the server (bounds + 2dp).
- gatewayTypes.ts: Billing* response interfaces.
- registry.ts: register billingCommands.
- billingCommand.test.ts: 12 vitest cases (overview/gating/buy-confirm-poll-
settled/no_payment_method/step-up/limit/auto-reload/validation).
TUI build green; 12/12 vitest pass; slash tests pass once @hermes/ink is built.
* docs(billing): scrub private cross-repo references
NAS is a private repo — remove all references to it from the public PR:
- drop the cross-repo planning doc (planning scaffolding, not a deliverable;
the PR description documents the design)
- replace 'NAS' / 'PR #412 preview' mentions in code + test comments with
generic 'the server' / 'a preview deployment'
* docs(billing): scrub final NAS reference in step-up docstring
* docs(billing): drop dangling plan-doc refs
The phase-2b plan doc was removed in the cross-repo scrub (300afcc0b)
but two module docstrings still pointed at it. Drop the dead refs.
* feat(billing): interactive /billing overlay + step-up UX, portal-URL & token fixes
Adds the interactive /billing TUI overlay and hardens the terminal-billing
client across CLI and TUI.
- TUI: full /billing overlay state machine (overview to buy to confirm,
auto-reload, read-only monthly limit) reusing the existing confirm overlay.
- Step-up: surface the verification link in-transcript and open the browser
via the TUI's own opener (the device flow runs in the headless gateway, so a
printed URL was being dropped); run the step-up handler off the main loop and
emit the link as an out-of-band event so the gateway stays responsive.
- Step-up copy is scope-accurate ("Billing permission granted") and re-checks
/state so it never claims "enabled" when the org kill-switch is still off.
- Portal deep-links resolve to absolute URLs against the active portal base
(the server emits them relative) - fixes a bare "/billing?topup=open" link.
- Billing calls refresh an expired access token via the stored refresh token
instead of reporting a false "not logged in".
- Optimistic funnel: advise "set up a saved card on the portal" up front when
no card is on file (advisory, not a hard gate).
- Token resolution is cached briefly so the 2s charge poll loop stops
re-locking + re-reading the auth store on every tick; 401 re-resolves fresh.
- Remove the temporary demo-mode shims.
Validation: 87 Python billing tests, 88 TS tests (billing command + gateway
event handler), tsc clean, ink + ui-tui builds green.
* docs(billing): add /billing TUI screenshots for PR
* fix(cli): guard _last_invalidate on bare instances; update stale prompt-fallback test
The UI-invalidate throttle read self._last_invalidate unconditionally, which
raised AttributeError on HermesCLI instances built without __init__ (the
thread-safety test's object.__new__ shell). Guard the read with getattr.
The off-main-thread branch of _prompt_text_input was changed (#23185) to cancel
cleanly to None instead of falling back to a bare input() that would hang on the
slash-worker thread; the test still asserted the old direct-input fallback.
Update it to assert the current intended behavior: returns None, calls neither
run_in_terminal nor input(), and does not hang.
The universal PARALLEL_TOOL_CALL_GUIDANCE block already lives on main, but it
shipped with two rough edges this change cleans up:
- It duplicated the batching steer for Google models. The
GOOGLE_MODEL_OPERATIONAL_GUIDANCE block still carried its own
"Parallel tool calls" bullet, so Gemini/Gemma received the instruction
twice in one prompt. Drop the redundant bullet — the universal block is now
the single source.
- Its comment claimed "nothing in the open-source system prompt encouraged
batching," which was wrong: the steer existed for Google models only. Reword
to say the gap was that every *other* model got nothing.
- Tighten the test that asserts the steer (precedence-correct), and add an
invariant guarding against re-introducing the Google duplicate.
* Port from cline/cline#11514: encourage parallel tool calls
Add a universal system-prompt guidance block telling the model to batch
independent tool calls (reads, searches, web fetches, read-only commands)
into a single assistant turn instead of one call per turn. The runtime
already executes independent batches concurrently (read-only tools always;
non-overlapping path-scoped file ops); the open-source system prompt had
nothing steering the model to PRODUCE the batch. Fewer round-trips means
less resent context, which compounds over a long conversation.
- prompt_builder.py: new PARALLEL_TOOL_CALL_GUIDANCE block (short, static,
cache-amortised) modeled on TASK_COMPLETION_GUIDANCE.
- system_prompt.py: inject right after the task-completion block, gated by
agent.valid_tool_names + the new toggle.
- agent_init.py: read agent.parallel_tool_call_guidance (default True).
- config.py: add the default under the agent section.
- test_prompt_builder.py: behavior-contract tests (batching steer, dependent
carve-out, length bound) — invariants, not wording snapshots.
Adapted from Cline's TypeScript tool-surface guidance to hermes-agent's
Python prompt-assembly architecture and config-over-env conventions.
* fix(desktop): never persist or restore a named custom provider as bare "custom"
Custom providers vanish from the Desktop/TUI model picker with
"No LLM provider configured" — repeatedly fixed (#44062, #44109, #45578)
and repeatedly regressed (#44022, #47714) because every fix only recovered
the entry identity from a persisted base_url. When a session is
persisted/restored with the resolved provider "custom" and NO base_url, bare
"custom" leaked through verbatim; resolve_runtime_provider("custom") routes to
the OpenRouter default URL with no api_key, so the next turn/resume dies.
Bare "custom" is the resolved billing class shared by every named providers:/
custom_providers: entry — it is not a routable identity. Centralize the
"never let bare custom escape" invariant in one helper,
runtime_provider.canonical_custom_identity(), and apply it at all four leak
sites in tui_gateway/server.py:
- _ensure_session_db_row — the ORIGIN: first DB write seeds the bad row
- _runtime_model_config — live persist
- _stored_session_runtime_overrides — resume restore (heals old rows; drops
unrecoverable bare custom so resume falls back to config default)
- _make_agent — rebuild / per-turn
The helper recovers custom:<name> from the endpoint URL when present, else
from config.model.provider (the durable identity left when no base_url
survived). Regression tests in test_custom_provider_session_persistence.py
lock the no-base_url vector at every site so it cannot regress again.
Salvage corrections on top of @XVVH's #44341:
- Make native web_search injection a 1:1 swap for an already-present client
web_search function, NOT an additive grant. The original unconditionally
appended {"type":"web_search"} on every is_xai_responses turn with any
tools, force-enabling Grok server-side search even when the user never
enabled the web toolset (bypassing Hermes web-provider config + tool-trace
plumbing). Now gated on a client web_search actually being present.
- Reconcile grok-composer context to 200000 (merged in #47908) rather than
262144; 200k is xAI's published usable context window for Composer 2.5,
262144 is the /v1/responses input+output budget.
- Update tests to match scoped behavior + add a no-web-toolset guard test.
- AUTHOR_MAP entry for #44341 salvage.
Incomplete-guard (server-side *_call items at in_progress no longer flip
has_incomplete_items) and preflight built-in-tool allowlist kept as-is.
- model_metadata: grok-composer-2.5-fast → 262144 (OAuth slug not in /v1/models)
- codex transport: inject native {"type":"web_search"} for is_xai_responses;
drop client web_search to avoid duplicate-name 400s
- codex adapter: do not treat in-progress server-side *_call items as incomplete
- tests: adapter, transport build_kwargs, model_metadata, oauth recovery
Weak open models (mimo, nemotron-class) that see tool-call XML/JSON sitting in
file contents or tool output get primed and emit their own structured tool
calls mimicking the payload — usually with an empty/whitespace name. Those
calls can't be fuzzy-repaired toward a real tool, so the dispatch loop returns
an error and the model retries. Before this fix, every empty-name error dumped
the full tool catalog back to the model, which fed the priming loop more names
to mimic and inflated context 3-4x across the retry budget.
A blank/whitespace-only tool name now gets a terse anti-priming error that
tells the model in-context tool-call syntax is DATA, with no catalog dump. A
genuinely-wrong-but-nonempty name (a real typo) still gets the full catalog so
the model can self-correct.
Not a sandbox/auth boundary issue: Hermes never parses tool-call text from
content into executable calls (structured tool_calls only; the lone text->call
parser is the Copilot ACP transport and it also rejects empty names). The
reporter's own debug dump confirms the injection never executed.
Behavior-contract test added: empty-name -> terse error, no catalog; nonempty
unknown -> catalog preserved. Exercised end-to-end via run_conversation against
an in-process mock provider.
Context files (AGENTS.md, CLAUDE.md, .hermes.md, .cursorrules, SOUL.md) were
hard-capped at a flat 20K chars before head/tail truncation. Among the agent
harnesses we track, only Codex caps project docs at all (32 KiB); Claude Code,
OpenCode, and Cline load them whole. The flat 20K predates large context
windows and silently truncates real-world AGENTS.md files.
B — dynamic cap: when context_file_max_chars is unset (now the shipped
default), the cap scales with the model's context window
(ctx_tokens * 4 * 0.06, floor 20K, ceiling 500K). Small-context models stay at
the historical 20K; a 200K model gets 48K; large models stop truncating real
docs. An explicit context_file_max_chars still wins. Context length is resolved
once per conversation (stable -> prompt cache untouched).
C — when truncation does happen, the marker now names the concrete file path
and tells the agent to read_file it for the full content.
Validation: 154 targeted tests + full agent/ + hermes_cli/ + test_config
(0 failures); E2E against a real 60K AGENTS.md confirms small windows truncate
with the path-bearing marker, large windows load whole, and the system prompt
is byte-stable across rebuilds.
Rolling back to the oldest curator snapshot failed and deleted that
snapshot. rollback() takes a safety snapshot first, and snapshot_skills()
ends by pruning the backups directory down to keep (5 by default). At the
steady keep limit that prune removed the oldest snapshot, which is the very
one being restored, so the extract found no skills.tar.gz and the rollback
stopped with "snapshot extract failed (state restored)".
Thread an optional protect set through snapshot_skills() into _prune_old()
so the pre rollback safety snapshot can never evict the snapshot being
restored. Add two regression tests covering restore of the oldest snapshot
at the keep limit.
Fixes#47612
The curator now defaults to prune-only: the deterministic inactivity pass
(mark stale / archive long-unused skills) still runs whenever the curator is
enabled, but the opinionated LLM umbrella-building consolidation fork is OFF
by default.
- agent/curator.py: add DEFAULT_CONSOLIDATE=False + get_consolidate(); gate
the forked aux-model review in run_curator_review behind it (new consolidate
param, None=read config). When off, the LLM pass is skipped entirely (no
aux-model cost); the run is still recorded and reported.
- config.py: add curator.consolidate (default false); v29->v30 migration seeds
the key for existing installs without clobbering a user-set value.
- hermes_cli/curator.py: 'hermes curator run --consolidate' override; status
shows consolidate state; prune-only notice on run.
- docs + tests.
The double-underscore prefix swap fixed bare native tools but SKIPPED tools
already named mcp_<server>_<tool> (real MCP servers, e.g. mcp_linear_get_issue):
they went on the OAuth wire single-underscore and still tripped Anthropic's
third-party billing classifier -> HTTP 400 'extra usage, not plan limits'.
Verified empirically against a live Max subscription: a single mcp_ tool flips
the whole request to the extra-usage lane; mcp__ is accepted.
- build_anthropic_kwargs: promote ANY leading single-underscore mcp_ to mcp__
(bare names -> mcp__name; mcp_<server>_<tool> -> mcp__<server>_<tool>),
never double-prefixing an already-mcp__ name. Same for tool_use blocks in
history.
- normalize_response: reverse the mcp__ wire name back to whichever original
the registry knows — the single-underscore mcp_<server>_<tool> form for MCP
server tools, or the bare name for native tools — preferring a name that
already resolves natively.
- Tests rewritten to assert the invariant: ZERO single-underscore mcp_ names
reach the OAuth wire, and the mcp__ round-trip resolves back to the
registered name for both native and MCP-server tools.
Builds on liuhao1024's mcp__ prefix commit (cherry-picked). Closes the
MCP-server gap that left any session with an MCP server configured still
billing to extra usage.
Anthropic's Claude-Code request classifier treats tool names with a
single-underscore `mcp_<x>` prefix as non-Claude-Code / third-party,
routing the request to extra-usage billing (HTTP 400). Real Claude Code
uses double underscores: `mcp__<server>__<tool>`.
Change the tool-name prefix from `mcp_` to `mcp__` in both the outgoing
path (build_anthropic_kwargs) and the incoming path
(normalize_response). Update the skip-guard to check for both `mcp_`
and `mcp__` prefixes so native MCP server tools (which use the legacy
single-underscore format) are not double-prefixed.
Fixes#46675
Support files under references/, templates/, assets/, and scripts/ are progressive-disclosure data loaded through skill_view(..., file_path=...). They should not be treated as standalone skills during discovery or collision checks.
This prevents archived skill packages or support markdown files inside a real skill from shadowing active skills with the same name while still allowing top-level categories named scripts/templates/assets/references.
Tests cover:
- pruning nested SKILL.md files inside skill support directories
- preserving support-named top-level categories
- avoiding skill_view collisions from support markdown
- keeping archived package SKILL.md files accessible only through file_path
Follow-up to salvaged PR #41619: replace the module-global
_truncation_warnings list with a contextvars.ContextVar so concurrent
gateway-session prompt builds can't drain or clear each other's pending
warnings (cross-session leak). Adds a context-isolation test.
PROBLEM: Automatic context files such as SOUL.md and AGENTS.md were capped by a hardcoded CONTEXT_FILE_MAX_CHARS value. Amy's local fork had raised that constant from 20K to 25K so a larger SOUL.md would not be silently truncated, but the hardcoded 25K value changed upstream default behavior and made the patch less generally useful.
SOLUTION: Restore the upstream-compatible 20K default, add a context_file_max_chars config setting for users who intentionally keep larger identity/project-context files, keep chat-visible truncation warnings, and document the new setting. Tests cover the default, config override, explicit max_chars precedence, and the warning text.
Generalizes #32663 (@ehz0ah). The slash-skill scaffolding pollution
affected every auto-syncing memory provider — mem0, hindsight, retaindb,
byterover, honcho, supermemory all store/embed the raw user turn, so a
/skill invocation poisoned their stores with the full skill body, not just
openviking.
- Lift the contributor's parser into agent/skill_commands.py as the canonical
extract_user_instruction_from_skill_message(), co-located with the message
builders so the markers can't drift.
- Strip once in MemoryManager.{prefetch_all,queue_prefetch_all,sync_all} —
fixes the whole provider fan-out, bare /skill turns are skipped entirely.
- OpenViking's _derive_openviking_user_text() now delegates to the shared
helper as defense-in-depth (no duplicated marker literals).
- Marker-drift regression now asserts against the canonical skill_commands
constants; add manager-level coverage proving every provider gets clean text.
A live config.set model switch already moved the next API call to the new model,
but the conversation could still restore an old sessions.system_prompt snapshot
whose Model/Provider lines named the previous runtime. That made "what model are
you?" answer from stale metadata even while inference ran on the new model.
After a live switch we now refresh the stored system prompt and append a real
system-history pivot (not a fake user turn) so the transcript itself records the
new model/provider. Restore also rejects already-stale prompt snapshots when
their Model/Provider lines disagree with the runtime, so existing bad sessions
self-heal.
Keep request dump writes on the shared atomic JSON path, add regression coverage for request body/error/stdout redaction, and map the salvaged contributor email for release attribution.
converse() and converse_stream() were added in boto3 1.34.59. When Hermes
is installed editable into system Python (e.g. Ubuntu 24.04 ships 1.34.46),
the system boto3 takes precedence and calls to converse_stream fail with
AttributeError. Add an early version check in _require_boto3() that raises
a clear RuntimeError with upgrade instructions.
Remove the free Parallel Search MCP path and restore the keyed Parallel backend behavior from before it was introduced.
Also drops the keyless fallback registration/display labeling tests and returns the Parallel SDK pin to the prior version.
GLM-5.2 ships with a 1M (1,048,576) token context window. Without this
entry, Hermes falls through to the generic 'glm' key (202,752 tokens),
under-reporting the context bar and prematurely compressing conversations.
The 1M limit was verified empirically via needle-in-a-haystack retrieval
at 789,240 prompt tokens on api.z.ai/api/coding/paas/v4 — zero errors,
zero truncation, correct retrieval at every tested size (25K through 789K).
Changes:
- agent/model_metadata.py: add 'glm-5.2': 1_048_576 before 'glm' fallback
- hermes_cli/models.py: add glm-5.2 to zai curated models
- hermes_cli/setup.py: add glm-5.2 to setup wizard zai list
- hermes_cli/auth.py: add glm-5.2 to coding plan endpoint probes
- plugins/model-providers/zai/__init__.py: add glm-5.2 to fallback_models
- tests/agent/test_model_metadata.py: context resolution + vendor-prefix tests
A chat-completions response that carries real text or tool calls *alongside*
a `message.refusal` note is a normal, usable turn — the model did work. The
prior logic flipped finish_reason to `content_filter` whenever a refusal
string was present, so the conversation loop reframed a content-bearing turn
as a *failed* safety refusal (failed=True) and buried the model's actual
output inside the "model declined" template, or dropped tool calls entirely.
Only promote to a terminal `content_filter` when the refusal is the sole
payload (no visible text AND no tool calls). The refusal explanation is still
recorded in provider_data in every case for observability. Refusal-only
responses (the bug this feature targets) are unaffected and still surface
terminally; the empty+refusal, bare content_filter passthrough, and no-refusal
common cases are byte-identical to before.
Updates the partial-content test to the corrected contract and adds a
tool_calls-alongside-refusal regression guard.
OpenRouter (and every other OpenAI-compatible provider) uses the default
chat_completions transport, so it is already covered by the refusal fix:
an upstream Claude / moderation refusal arrives as
finish_reason="content_filter" (often empty content, no message.refusal).
Add a regression test asserting the transport passes that finish reason
straight through to the loop's content_filter handler.
(cherry picked from commit 60168a513bc9edc508aa8968d0163bd5feb87055)
A Claude refusal (HTTP 200, stop_reason="refusal", empty content) was
laundered into a generic retry loop and surfaced as a misleading
"rate limited / invalid response" or "no content after retries" error,
burning paid attempts reproducing a deterministic refusal.
This hit two distinct paths:
- Direct Anthropic (anthropic_messages): validate_response rejected the
empty-content refusal *before* normalize_response mapped refusal ->
content_filter, so it fell into the invalid-response retry loop.
- Nous Portal / OpenAI-compatible (chat_completions): the portal surfaces
a Claude refusal via message.refusal with empty content, which sailed
past validation and died in the empty-response retry loop.
Fix (one unified content_filter dispatch for all backends):
- AnthropicTransport.validate_response: accept empty content when
stop_reason == "refusal" so it flows to normalize_response.
- ChatCompletionsTransport.normalize_response: promote message.refusal to
content + a content_filter finish reason.
- conversation_loop: handle finish_reason == "content_filter" - fire the
api_request_error hook (content_policy_blocked), try a configured
fallback once, else return a clear terminal refusal message. Never retry
a deterministic refusal.
Supersedes #43084, which fixed only the direct-Anthropic path and could
not reach the chat_completions/portal path.
Tests: transport-level (validate_response refusal, message.refusal
promotion) + end-to-end loop (refusal surfaced, exactly one API call).
(cherry picked from commit 01f546f92cb1629ec1427be270dbd7c504e962ad)
The previous test patched ssl.create_default_context globally with a bare
SSLContext that has zero CA certs. Both verify_ca_bundle() and the macOS
fallback got the same mocked context, so the test verified nothing useful:
both paths produced empty get_ca_certs() and the assertion that no
exception escaped was vacuously satisfied.
Only mock the fallback call (no cafile) — let the certifi call hit the
real SSL stack and fail with SSLError on the broken PEM. The mock
fallback returns a context with load_default_certs() so the test now
verifies the real scenario: broken certifi → SSLConfigurationError,
macOS system trust store → success.
Also pads the broken PEM past the 1 KB size guard so the size check
doesn't short-circuit before ssl.create_default_context(cafile=...) runs.
Reported by @liuhao1024 in PR review.
A stale certifi CA bundle after a partial `hermes update` used to crash
the agent on the first outbound HTTPS call with a raw traceback and
trap the gateway in a retry loop.
This patch:
* Adds `agent/errors.py` with a typed `SSLConfigurationError`
* Adds `agent/ssl_guard.py` with a `verify_ca_bundle()` pre-flight
that asserts the bundle exists, is non-trivial in size, and can build
a working SSLContext. On macOS, it falls back to the system trust
store when the bundle is empty but the system store is healthy
(covers corporate proxies / MDM setups).
* Wires the guard into `run_agent.py` and `gateway/run.py` right
after the `hermes_bootstrap` import, inside a try/except so a bug
in the guard itself can never prevent startup.
* Adds a `SSL / CA Certificates` section to `hermes_cli doctor` so
users can detect the failure with one command.
* Adds unit tests covering the healthy, missing, empty, skip-env, and
macOS-fallback paths.
* Adds an RCA document describing the failure mode and the recovery
path (`pip install -e .`).
When the bundle is broken the user sees:
\u26a0\ufe0f SSL certificate bundle issue detected.
Run: pip install -e .
`HERMES_SKIP_SSL_GUARD=1` disables the check for sandboxed
environments that ship their own trust store.
Bedrock Converse rejects non-default sampling parameters for Opus 4.7 and 4.8 with a ValidationException. Reuse the Anthropic-native sampling-param guard in the Bedrock kwargs builder so those models omit temperature/topP while older Claude and non-Claude models keep existing behavior.
Includes the stop-sequence regression from the parallel fix to ensure stopSequences still pass through for restricted Opus models.
Co-authored-by: Tranquil-Flow <tranquil_flow@protonmail.com>
Remove the rich_messages config toggle entirely so Telegram replies always try the Bot API 10.1 rich-message path first, with the existing MarkdownV2 fallback/latch behavior for unsupported endpoints and per-message failures.
Restore the Telegram platform hint to encourage rich Markdown tables/task lists/math now that the rich path is the default, and remove the config/docs surface for the old toggle.
Custom endpoints carry two naming conventions for the same provider: the
agent's provider attribute is the generic 'custom' label while the pool
is keyed 'custom:<normalized-name>'. The defensive guard in
recover_with_credential_pool compared them literally, logged
'Credential pool provider mismatch: pool=custom:<name>, agent=custom',
and skipped recovery — so 401 refresh and 429 rotation never ran for
ANY custom-provider user (seen in the field on a Fireworks setup whose
dead key burned full retry cycles every turn with the skip warning on
each one).
Accept the pair only when the agent's CURRENT base_url resolves to the
same pool key via get_custom_provider_pool_key, preserving the guard's
original purpose (#33088/#33163): a fallback provider or a different
custom endpoint still skips pool mutation.
21 cases pinning the new ``_ensure_last_assistant_message_in_tail``
anchor and its interaction with the existing tail-cut path:
* ``TestFindLastAssistantMessageIdx`` — helper contract: prefers a
content-bearing assistant message, skips ``tool_calls``-only
stubs, multimodal text-block content counts, falls back to
"any assistant" when no content-bearing reply exists, honours
``head_end``, returns -1 when there's none.
* ``TestEnsureLastAssistantMessageInTail`` — direct: no-op when
already in the tail, walks ``cut_idx`` back when the reply is
in the compressed middle, never crosses into the head region,
re-aligns through a preceding ``tool_call`` / ``tool_result``
group instead of orphaning it.
* ``TestFindTailCutByTokensAnchorsAssistant`` — integration:
reporter repro (long tool-output run after the visible reply)
now preserves the reply; user and assistant anchors compose
in a single tail-cut call; a soft-ceiling-overrunning oversized
tool result no longer strands the prior reply.
* ``TestCompactionRollupReproduction`` — end-to-end through
``compress()`` with a stubbed ``_generate_summary``: the
visible reply text survives either as its own standalone
assistant message (normal path) or concatenated onto the
merged summary tail (double-collision path the WebUI then
re-splits). The standalone-summary case is asserted strictly
(exactly one summary row, exactly one separate assistant
row carrying the reply) — that's the dominant path and any
drift there reintroduces the original bug.
* ``TestSourceGuardrail`` — static asserts on
``agent/context_compressor.py``: the helper exists, the
anchor is wired into ``_find_tail_cut_by_tokens`` AFTER the
user-message anchor (so chaining is monotonic), the
content-bearing preference is preserved, and the issue
number is referenced so future bisects can find this fix.
When the compression summary lands as an assistant-role message (head ends
with user), the end marker was not appended. Models may regurgitate the
summary text as their own visible output when there's no clear boundary
signal (#33256).
The end marker was already appended for user-role summaries (#11475, #14521)
but the assistant-role path was missed in the original fix. This ensures ALL
standalone summary messages carry the boundary marker, preventing summary
text from leaking into user-visible chat output.
- Use reply_parameters per the sendRichMessage spec instead of the
undocumented reply_to_message_id scalar (silently ignored -> reply
anchor quietly dropped).
- Latch rich sends off after an endpoint-capability failure (old PTB /
server without sendRichMessage) so every later reply doesn't pay a
doomed extra roundtrip; per-message BadRequests do NOT latch.
- Default rich_messages to OFF (opt-in) while the day-old Bot API 10.1
endpoint is validated live; revert the prompt-hint table guidance
until the default flips on.
- Tests: reply_parameters shape, send-latch behavior, BadRequest
non-latch; rich tests opt in explicitly via extra.
Introduce opportunistic support for Telegram Bot API 10.1 rich messages by sending raw agent Markdown via sendRichMessage and streaming previews via sendRichMessageDraft. Implements a rich-path fast‑path in gateway/platforms/telegram.py (RICH_MESSAGE_MAX_BYTES=32768, feature gate platforms.telegram.extra.rich_messages, bot capability checks, routing/thread handling, and conservative fallback rules: permanent/capability errors fall back to the legacy MarkdownV2 path, transient/network errors are surfaced without legacy-resend). Also add a latch for draft capability failures (_rich_draft_disabled) and preserve legacy chunking and draft behavior when needed. Update agent prompt hints (telegram encourages rich Markdown/tables), add CLI config example option, update English and Chinese docs to describe rich messages and fallbacks, and add/adjust tests for rich send and draft behavior.
* feat(billing): /usage → portal top-up browser handoff
Add the terminal side of the billing slice (phase 2a): start a top-up by
throwing the user to the portal billing page with the top-up modal open. The
terminal does not confirm, poll, or track payment — checkout completes in the
browser and the next /usage shows the new balance.
- nous_account.py: parse organisation.slug/name from /api/oauth/account into
NousPortalAccountInfo; add nous_portal_topup_url() building the org-pinned
{base}/orgs/{slug}/billing?topup=open with a null-slug fallback to the legacy
{base}/billing?topup=open (never /orgs/None/...).
- portal_cli.py: 'hermes portal topup' — fresh account fetch, identity line
(Topping up as <email> / org <name>), browser open with printed-URL fallback,
no-wait closing copy. No polling/confirmation (deferred to 2b).
- account_usage.py: the shared /usage credits block now links the org-pinned
top-up URL (auto-opens the modal) + points to the command.
Depends on NAS #409 (organisation.slug/name + ?topup=open). Do not merge until
that is live on the target env; until then /api/oauth/account returns
organisation: { id } only and the URL falls back to legacy.
* feat(billing): /credits command for balance + top-up handoff
Replace the standalone `hermes portal topup` subcommand with an in-session
/credits slash command — a focused money surface (balance in, top-up out) that
works in the CLI, TUI, and every messaging platform from one registry entry.
- commands.py: register /credits (Info category). Slack is at its 50-slash cap,
so /credits is routed via /hermes credits on Slack only (new
_SLACK_VIA_HERMES_ONLY set) to avoid clamping a canonical command off the
native list and breaking Telegram parity; native everywhere else.
- account_usage.py: build_credits_view() — one portal fetch → balance lines +
identity line + org-pinned top-up URL + depleted flag, consumed by all
surfaces. Reuses the same snapshot/URL builder as /usage so numbers match.
- cli.py: _show_credits() — balance block + identity line + 3-button panel
(Open top-up / Copy link / Cancel) via the existing prompt_toolkit modal.
ASK, never auto-launch; headless falls back to printing the URL.
- gateway/slash_commands.py: _handle_credits_command() — renders the block +
tappable top-up URL + no-wait copy; works on button and plain-text platforms.
- /usage credits line now points to /credits.
- Retire `hermes portal topup` (portal_cli.py back to baseline); the engine
(slug/name parse + nous_portal_topup_url) stays as the shared core.
No polling, no payment confirmation (billing phase 2a). Depends on NAS #409.
* fix(credits): /credits works in the TUI slash-worker (non-interactive)
In the TUI, /credits runs in the slash-worker subprocess where there is no
live prompt_toolkit app and stdin is the JSON-RPC pipe. _show_credits called
the 3-button modal unconditionally, which fell back to reading stdin →
exception → slash.exec rejected → the command produced no output (only the
pre-existing 'Credit access paused' banner showed).
- _show_credits: when self._app is None (TUI worker / piped / non-interactive),
render the text variant — balance block + tappable top-up URL + no-wait line,
same affordance as the messaging surfaces — and skip the modal entirely. The
3-button panel still renders in the interactive CLI.
- Depleted banner copy: 'run /usage for balance' → 'run /credits to top up'
now that /credits is the dedicated money surface (+ tests).
- Regression tests: _show_credits with self._app=None renders text and never
invokes the modal; logged-out path.
* feat(tui): credits.view RPC for the /credits tappable top-up button
Add a credits.view JSON-RPC method returning the structured CreditsView
(logged_in, balance_lines, identity_line, topup_url, depleted) so the TUI can
render a clickable <Link> top-up button instead of plain text. Account-
independent (portal fetch gated on a logged-in Nous account), fail-open to
{logged_in: false} on any hiccup. Mirrors session.usage's credits-block pattern.
Frontend (TUI-local /credits command + Ink component) lands separately.
* feat(tui): /credits command with keyboard-driven top-up confirm
TUI-local /credits: fetches the structured balance via the credits.view RPC,
prints the balance + identity + top-up URL, then arms the EXISTING confirm
overlay (Enter = open top-up in browser via openExternalUrl, Esc = cancel).
Reuses ConfirmReq — no new overlay component/state/input handler. Headless
(openExternalUrl returns false) falls back to printing the URL.
- gatewayTypes.ts: CreditsViewResponse.
- commands/credits.ts: the command (mirrors /status's rpc+guarded pattern).
- registry.ts: register creditsCommands.
- test: balance+overlay armed, headless fallback, no-url, logged-out (4 cases).
Matches the CLI /credits 'Enter to open' affordance. Phase 2a: no polling.
The subscription-cap usage gauge (50/75/90% bands) ignored purchased
(top-up) credits: a sub user with top-up funds got a sticky warn banner
at 90% of their cap — permanently at >=100%, alongside grant_spent —
despite being fully able to keep inferencing. The cap is the wrong
denominator for an account that can keep spending.
- evaluate_credits_notices: purchased_micros > 0 suppresses the usage
band (grant_spent already covers the cap-reached + top-up case with
the remaining balance). A top-up landing mid-session clears any
showing band; spending top-up down to 0 resumes the gauge.
- New display.credits_notices config (default true): false silences all
credits notices. State capture and /usage are unaffected. Read once
per agent (cached) in _emit_credits_notices, fail-open true.
- Docs: configuration.md display block.
The original fix added agent/memory_manager.py:flatten_message_content, but
that helper was a near-exact duplicate of
agent/codex_responses_adapter.py:_summarize_user_message_for_log — same
None/str/list dispatch, same {text,input_text,output_text}/{image_url,input_image}
part sets, the identical [N image(s)] marker, and the same str() fallback. The
only difference was the join separator (newline for memory vs space for the
log/trajectory previews the existing helper already serves), and that helper is
already imported into agent/turn_finalizer.py — the same file whose call site the
memory fix touches.
Parameterize the existing helper with sep=' ' (default preserves every current
logging/trajectory caller byte-for-byte) and call it with sep='\n' at the memory
boundary; drop the forked flatten_message_content. Repoints the unit tests to the
consolidated helper and adds a case locking the default space-join.
Single source of truth for multimodal-content flattening; no behavior change for
the fix or for existing callers.
Multimodal turns carry message content as a list of typed parts
({type: "text"|"image_url", ...}). _sync_external_memory_for_turn
passed that list straight into MemoryManager.sync_all, and providers
feed it to regexes — Honcho's sync_turn calls sanitize_context, where
re.sub raised 'expected string or bytes-like object, got list'. Every
turn with an attached image silently never synced.
Flatten to plain text at the boundary: text parts joined, images noted
as an [N image(s)] marker so the attachment isn't erased from recall.
Fixing here covers all providers instead of patching each plugin.
(cherry picked from commit 705bdb6ffe9deb60885182fa48f63675d4ba2e35)
Tell coding agents to activate shell setup once per session instead of re-sourcing it before every command, and pin the existing LocalEnvironment env-snapshot behavior with regression tests.
The prompt consolidation above retires the carveout-era prefix. Without a
frozen copy in _HISTORICAL_SUMMARY_PREFIXES, summaries persisted by
pre-upgrade builds would lose detection (_is_context_summary_content) and
renormalization (_strip_summary_prefix) — the exact regression class the
tuple exists to prevent. Adds contract tests covering every frozen prefix.
Refs #41607#38364#42812
The coding-posture brief told GPT/Codex models to use patch mode='patch'
(V4A) for structured/multi-file changes but mode='replace' "for a single
small swap". That second nudge points those models at a format their
first-party harness never taught them.
Verified against openai/codex (current main): apply_patch is the ONLY file
editor in codex-rs — zero occurrences of str_replace/old_string anywhere in
the repo; the grammar (core/src/tools/handlers/apply_patch.lark) is exactly
the V4A dialect our patch_parser implements; the shipped model prompts
(gpt_5_codex, gpt-5.2-codex, gpt-5.1-codex-max + instruction templates)
explicitly say to use apply_patch "for single file edits"; and the tool is
gated per model via ModelInfo.apply_patch_tool_type, i.e. OpenAI ships
V4A-for-everything as model metadata.
The GPT-family line now steers to mode='patch' for all edits, single-file
included. The replace-family line (Claude + open-weight) is unchanged —
Claude Code's FileEdit is old_string/new_string/replace_all exact string
replacement (confirmed from Anthropic's shipped sdk-tools.d.ts, the only
file editor in its tool union), matching our mode='replace'.
The coding posture's names-only demotion of non-coding skill categories
(#44342) applied under the default auto mode, silently changing the skill
index for every user in a git repo. Index changes must be opt-in: demotion
now only fires under agent.coding_context=focus, alongside the toolset
collapse. auto/on leave the skill index untouched; focus semantics are
unchanged (demoted, never hidden; deny-list keeps coding-adjacent and
custom categories at full entries).
Real-world failure with the original index pruning: under the default auto
posture, an agent-created ops skill in a demoted category vanished from the
prompt's skill index mid-project, and the agent silently fell back to a
stale sibling skill instead. The "discovery-only" premise didn't hold —
models do not reach for skills_list to rediscover what the index stops
showing them, and agent-created skills are the model's accumulated project
memory (runbooks, pitfalls, operating rules).
Gating pruning behind the opt-in focus mode was the wrong fix too: users
opening a worktree don't know the config exists, so the index-noise win
would effectively never ship.
Instead, the coding posture now DEMOTES non-coding categories rather than
hiding them: each demoted category renders as a single names-only line
("gaming [names only]: allthemons10-ops, mc-backup") with a footer note
explaining the omitted descriptions. Every skill name stays in the prompt,
so memory-anchored recall ("load <name>") keeps working in every mode,
while the description noise is still cut. Applies in auto/on/focus alike;
the general posture demotes nothing. Deny-list semantics unchanged —
unknown/custom categories and coding-adjacent ones keep full entries.
API renamed to match the honest semantics: hidden_skill_categories →
compact_skill_categories, build_skills_system_prompt(hidden_categories=) →
compact_categories=.
IAM policies scoped to bedrock:InvokeModel only (a common least-privilege
setup) reject converse_stream() with AccessDeniedException. The agent loop
hard-prefers streaming and the denial never matched the 'stream not
supported' auto-fallback, so InvokeModel-only users looped on AccessDenied
forever.
- agent/bedrock_adapter.py: new is_streaming_access_denied_error()
detector (ClientError code check + wrapped-SDK message match);
call_converse_stream() falls back to converse() on denial.
- agent/chat_completion_helpers.py: bedrock_converse streaming branch
retries inline via converse() and sets _disable_streaming so later
turns skip the doomed stream attempt; the chat-completions retry
block also recognizes the denial for the AnthropicBedrock SDK path
(message pre-check avoids importing bedrock_adapter — and its lazy
boto3 install — for unrelated providers).
Both paths print a one-line notice telling the user which IAM action
restores streaming.
* feat(agent): coding-context posture with per-model edit-format tuning
Hermes detects when it's running in a coding context — an interactive
surface (CLI, TUI, ACP, desktop) sitting in a code workspace (git repo or
recognised project root) — and shifts into a coding posture. Outside that
(chat platforms, non-workspaces) nothing changes.
The posture is modelled as a frozen RuntimeMode selected from a small
ContextProfile registry (coding/general). A profile is data: the toolset to
collapse to, the operating brief to inject, and seams for model routing and
memory. Every domain reads the same resolved object instead of re-probing
git/config on its own:
- System prompt — RuntimeMode.system_blocks(): an operating brief (gather
context before editing, edit through tools not chat, verify with terminal,
cap retry loops) plus a live git/workspace snapshot, built once and baked
into the stable prompt tier so per-conversation caching is preserved.
- Per-model edit-format tuning — the brief nudges each model family toward
the patch mode it handles best: OpenAI/Codex toward mode='patch' (V4A
multi-file diffs), Anthropic toward mode='replace' (string replacement).
The model id rides on RuntimeMode; unknown families keep neutral wording.
- Skill index — non-coding skill categories are pruned from the prompt's
skill index (discovery-only; skills_list/skill_view still reach the full
catalog, with a disclosure note).
- Toolset — only under the opt-in 'focus' mode does the posture collapse to
the coding toolset + enabled MCP servers; the default posture is
prompt-only and never overrides configured toolsets.
Activation via agent.coding_context: auto (default), focus, on, off.
Subagents inherit the posture for free via toolset inheritance + the shared
prompt builder. Detection is not memoized so a long-lived gateway/TUI
process can't pin a stale posture across working directories.
* feat(agent): cover new-file authoring in the coding edit-format nudge
The per-model edit-format guidance only addressed editing existing code
(patch mode='patch' vs 'replace'), but authoring a brand-new file —
write_file, not patch — is a large fraction of real coding work and the
nudge was silent on it. Surfaced when building a single-file artifact where
the dominant operation was write_file and the steering offered no guidance.
Both family lines now lead with "author new files with write_file; for
edits to existing code prefer ...". Tests assert write_file appears in each
family's brief; unknown families still get neutral wording.
* docs(agent): correct memoization docstring + clarify TUI config-load asymmetry
* feat(agent): sharpen the coding posture — verify-loop facts, wider edit steering, $HOME guard
Tuning pass on the coding posture from dogfooding it as a harness:
- Workspace snapshot now hands the model its verify loop up front:
detected manifests + package manager (lockfile sniff), the exact
verify commands (package.json scripts, Makefile targets,
scripts/run_tests.sh, pytest config), and which context files
(AGENTS.md / CLAUDE.md / .cursorrules) exist at the root. Marker-only
(non-git) projects get the snapshot too instead of nothing. The
"verify before claiming done" brief line was the highest-value piece
in evals — this turns it from advice into an executable loop instead
of making the model rediscover the test command every session. Still
stat-cheap, size-guarded reads, built once at prompt time.
- Edit-format steering covers the families Hermes actually serves:
Gemini and open-weight coding models (DeepSeek, Qwen, Kimi, GLM,
Grok, Hermes, Llama, Mistral, Devstral, MiniMax) steer to
mode='replace' — their RL scaffolds use str_replace-style editors.
Previously only GPT/Codex and Claude families got steering; the
models Hermes users disproportionately run all fell to neutral.
- Operating brief gains four behaviors elite harnesses encode: batch
independent reads/searches in one turn; fix root causes and the bug
class (sibling call paths), not the reported site; no drive-by
refactors/renames/reformatting; never read, print, or commit secrets.
Plus a patch-failure escalation ladder: after the same region fails
twice, rewrite the enclosing function/file with write_file instead of
a third patch attempt.
- $HOME dotfiles guard: a git repo rooted exactly at the home directory
(or a marker sitting in it, e.g. a global ~/AGENTS.md) is user config,
not a code workspace — without the guard, every session anywhere under
a dotfiles-managed home silently flipped to the coding posture. Real
projects under such a home still detect via their own markers/repos;
'on' mode bypasses the guard.
Drop the hermes_state.py column + persistence plumbing from the salvaged
interleaved-thinking fix. The ordered-block channel covers the failure
window in-memory (turn replayed within the live conversation loop). A
session reloaded from disk after a crash falls back to reconstruction;
if that replay 400s, the thinking-signature recovery (#43667) strips
reasoning_details and retries — one degraded call in a rare resume path
instead of a schema column. Replaces the DB-roundtrip test with a
fallback-shape test.
Two additive hardening changes on the interleaved-thinking replay path
introduced by this PR's anthropic_content_blocks channel. Both are scoped
to that channel's blast radius; neither changes correct behavior.
1. Replay-time tool-input re-sourcing (credential safety).
The ordered-block channel captures each tool_use `input` from the RAW
API response in normalize_response, which is NOT credential-redacted.
The parallel tool_calls[].function.arguments IS redacted at storage
time (build_assistant_message, #19798). The verbatim-replay fast path
in _convert_assistant_message replayed the raw block input, so a secret
a model inlined into a tool call (e.g. an Authorization header value
passed inside a terminal command) would ride back onto the wire even
though it is redacted everywhere else in history. Re-source tool_use
input from the redacted tool_calls map by
sanitized id; interleave order (the reason this channel exists) is
unaffected. Adapted from #36071, which re-sources tool inputs the same
way on its replay path.
2. Broaden the thinking-replay 400 classifier (defense-in-depth).
error_classifier only matched "signature" + "thinking", so the
frozen-block variant — "thinking ... blocks in the latest assistant
message cannot be modified. These blocks must remain as they were in
the original response." — carried no "signature" token and fell through
to a non-retryable abort. The anthropic_content_blocks channel prevents
the reorder that triggers this 400 at the source, but if any future
mutator reintroduces it, the turn now self-heals via the existing
strip-reasoning-and-retry recovery instead of crash-looping. A negative
case ensures an unrelated "cannot be modified" 400 (no "thinking") is
not swept in. Mirrors the classifier broadening in #36087 and #36071.
Tests
- tests/agent/test_anthropic_thinking_block_order.py: a replay test
asserting an inlined secret is redacted on the wire while interleave
order is preserved.
- tests/agent/test_error_classifier.py: three cases — frozen-block 400
native and via OpenRouter route to thinking_signature/retryable; an
unrelated "cannot be modified" 400 does not.
Both grafts verified RED (tests fail with the change reverted) then GREEN.
Full adapter, transport, classifier and output-field-leak suites pass.
Co-authored-by: AlexanderBFoley <92330381+AlexanderBFoley@users.noreply.github.com>
HTTP 400 "messages.N.content.M.text.parsed_output: Extra inputs are not
permitted" on the native Anthropic transport. Anthropic SDK 0.87.0 response
blocks carry output-only attributes the Messages *input* schema forbids: text
blocks get `parsed_output` and `citations=None`, tool_use blocks get `caller`.
normalize_response captured blocks verbatim via _to_plain_data and replayed
them as request input on the next turn, so the forbidden fields leaked back ->
400. Like the earlier thinking-block bug, one poisoned turn wedges every
subsequent request in the session (even the diagnostic turn), recoverable only
by switching models or deleting the session.
This is a defect in the anthropic_content_blocks channel added for the
interleaved-thinking fix: it preserved block ORDER correctly but copied every
SDK attribute, including output-only ones.
Fix — whitelist input-permitted fields per block type at all three leak points:
- agent/transports/anthropic.py normalize_response: sanitize at CAPTURE so the
poison never persists to state.db (defence-in-depth).
- agent/anthropic_adapter.py _sanitize_replay_block (new): whitelist used on the
ordered-blocks replay path; also recovers already-poisoned stored sessions.
- agent/anthropic_adapter.py _convert_content_part_to_anthropic: a stored
`text` part is rebuilt from whitelisted fields instead of dict(part) verbatim
(this was the exact content.N.text.parsed_output failure locus).
Whitelist not blacklist, so future SDK output-only fields can't reintroduce it.
Block order and thinking-block signatures are preserved (the reason the channel
exists). Adds tests/agent/test_anthropic_output_field_leak.py; full adapter
suite green (163 tests). Existing poisoned state.db rows scrubbed out-of-band.
Interleaved-thinking turns (adaptive thinking, Claude 4.6+/Opus 4.8) emit
content blocks like:
thinking_1(signed) tool_use_1 thinking_2(signed) tool_use_2
Anthropic signs each thinking block against the turn content preceding it
at its position. normalize_response split the turn into two parallel lists
(reasoning_details + tool_calls), discarding cross-type order, and
_convert_assistant_message rebuilt it as [all thinking][text][all tool_use].
That moved thinking_2 ahead of tool_use_1, invalidating its signature, so
Anthropic rejected the latest assistant message with HTTP 400:
messages.N.content.M: `thinking` or `redacted_thinking` blocks in the
latest assistant message cannot be modified.
Observed repeatedly in agent.conversation_loop against api.anthropic.com /
claude-opus-4-8, recurring across sessions on multi-thinking-block turns.
Fix: carry a verbatim, order-preserving copy of the turn's content blocks
(anthropic_content_blocks) end-to-end - capture in normalize_response,
persist/restore through state.db, and replay unchanged for the latest
assistant message. Gated to turns that actually interleave signed thinking
with tool_use, so normal turns are unaffected.
Adds 3 regression tests including a SQLite round-trip covering the
crash-recovery reload path.
Make Parallel the web search/extract backend with a zero-setup free tier:
- Keyless (no PARALLEL_API_KEY): web_search/web_extract work out of the box via
Parallel's free hosted Search MCP (search.parallel.ai/mcp), and parallel
becomes the default backend when no other web credentials are configured
(ahead of ddgs, which is search-only). A small hand-rolled Streamable-HTTP
JSON-RPC client speaks the MCP's web_search/web_fetch tools; the existing
web_search/web_extract tools are the only tools registered.
- Keyed (PARALLEL_API_KEY set): uses the Parallel v1 REST endpoints
(client.search / client.extract with advanced_settings.full_content) — no beta.
Bumps parallel-web 0.4.2 -> 0.6.0.
- Attribution: on the free path only, results carry provider/attribution and the
CLI tool line reads "Parallel search" / "Parallel fetch"; the paid path is
unbranded.
- Selection/registration: web tools register unconditionally (free MCP backstop)
while check_web_api_key remains a real usability probe; explicit per-capability
backends are honored (so misconfig surfaces) rather than masked by the fallback.
Tested: live web_search/web_extract against search.parallel.ai in keyless and
keyed modes; unit suites for the MCP client, backend selection, and display
labeling; full agent run shows the "Parallel search" label on the free path.
* fix(streaming): stop socket read timeout from preempting stale-stream detector
The stale-stream detector is deliberately scaled to 180-300s so reasoning
models (e.g. Opus) can pause mid-stream during extended thinking. But the
httpx socket read timeout stayed at a flat 120s for cloud providers and fired
first, tearing down healthy reasoning streams before the detector (which owns
retry + diagnostics) could act. Symptom: every Copilot/Opus turn dies with
ReadTimeout at a consistent ~125s and never completes.
Floor the cloud socket read timeout at the stale-stream timeout so it can no
longer fire before the detector. Local providers and explicit
HERMES_STREAM_READ_TIMEOUT / request_timeout_seconds overrides are unchanged.
* test(streaming): pin read-timeout >= stale-stream invariant for cloud reasoning streams
Cover the contract that the httpx socket read timeout is never shorter than
the stale-stream detector for cloud providers on the default: small contexts
floor to 180s, >=50K to 240s, >=100K to 300s; explicit overrides win; local
providers and the unresolved-value fallback are unaffected.
Anthropic returns a 400 when the thinking/redacted_thinking blocks in the
latest assistant message are mutated upstream: 'thinking or redacted_thinking
blocks in the latest assistant message cannot be modified. These blocks must
remain as they were in the original response.'
The classifier's thinking_signature branch only matched on the substring
'signature', so this variant fell through to a non-retryable client error
and hard-aborted the turn -- even though the existing strip-reasoning_details
-and-retry recovery would have healed it.
Broaden the 400 match to also catch 'cannot be modified' / 'must remain as
they were' (still gated on 'thinking'), routing it to the same recovery.
Adds a negative-case test so unrelated 'cannot be modified' 400s are not
swept in.
Defense-in-depth, orthogonal to the root-cause work in #35975 / #17861
(which prevent the block mutation in the first place). Only changes a
terminal-failure into a one-shot recovery.
Signed-off-by: Ian Culling <ian@culling.ca>
Rebased onto current main and re-ported across the restructured
surfaces: model flows now thread confirm_provider/base_url/api_key
through hermes_cli/model_setup_flows.py, the Discord picker lives in
plugins/platforms/discord/adapter.py, and the web dashboard picker
applies chat-mode switches via config.set so the expensive-model
confirmation can ride the response.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A GPT-5 model rejecting max_tokens returns a 400 whose message contains the
literal substring 'max_tokens' — one of the _CONTEXT_OVERFLOW_PATTERNS. The 400
path in _classify_400 checked overflow patterns before any request-validation
check (which only existed on the 5xx path), so the parameter error was routed
into the compression loop, re-sent with the same bad param, and ended in
'Cannot compress further' on a tiny context.
Hoist a request-validation guard (unsupported/unknown parameter) above the
context-overflow check in _classify_400. Deliberately excludes the generic
invalid_request_error code, which OpenAI also stamps on real overflow 400s, so
genuine overflows still compress. Pairs with the max_completion_tokens param
fix that stops the bad request at the source.
Also adds AUTHOR_MAP entry for the salvaged PR #13902 commit.
Third-party OpenAI-compatible endpoints (self-hosted gateways, OpenRouter,
Azure proxies) fronting gpt-4o / gpt-4.1 / gpt-5+ / o1-o4 models silently
received max_tokens and 400'd with unsupported_parameter, because the three
kwarg-selection sites only checked base_url_hostname(...) == "api.openai.com"
and fell through to max_tokens on every other host. The constraint is
enforced server-side by the model family, not by the URL, so name-based
detection is required as a fallback.
Changes:
- utils.py: new shared helper model_forces_max_completion_tokens(model) that
prefix-matches gpt-4o, gpt-4.1, gpt-5, o1, o3, o4 families on normalized
(lowercased, vendor-prefix-stripped) names.
- run_agent.py: _max_tokens_param ORs the helper into the URL check.
- agent/auxiliary_client.py:
- auxiliary_max_tokens_param gains an optional keyword-only model arg.
- _build_call_kwargs inline branch applies the same check for both
provider == "custom" and non-custom paths.
Tests:
- tests/test_model_forces_max_completion_tokens.py: 31 new cases covering
positive families, negatives (classic gpt-4, claude, llama, mistral, qwen,
deepseek), vendor prefixes, case-insensitivity, whitespace, None/empty,
and substring-not-prefix guards.
- tests/run_agent/test_run_agent.py::TestMaxTokensParam: 5 new model-based
cases (custom + gpt-5.4, openrouter + gpt-4o-mini, custom + o1-preview,
classic gpt-4-turbo keeps max_tokens, llama3 keeps max_tokens).
- tests/agent/test_auxiliary_client.py::TestAuxiliaryMaxTokensParam: new
class, 7 tests covering the URL x model matrix.
A binary @file: ref (PDF, docx, spreadsheet, …) expanded to a bare
"binary files are not supported" warning with no content. The model saw a
failure and gave up — e.g. a dropped PDF came back as a text note claiming the
type was unsupported, even though the file was staged on disk right next to it.
Inject an actionable content block instead: the path, mime type, size, and a
nudge to use its tools to read/convert/view the file (and explicitly not to tell
the user the type is unsupported). General across every binary type — not
PDF-specific. The file already resolves where the agent's tools run (local cwd
or the staged copy in a remote session workspace), so it can act on it directly.
New Anthropic models without a recognized version substring (claude-fable-5
and future named/numbered releases) were classified as legacy and routed down
the manual-thinking path, which made OpenRouter emit thinking.type.disabled —
a form reasoning-mandatory Claude models reject with a non-retryable HTTP 400.
Invert the brittle version-substring allowlists to default-to-modern (mirroring
_get_anthropic_max_output): unknown Claude models get the adaptive/xhigh/
no-sampling contract, with an explicit legacy list for older families. Non-Claude
Anthropic-Messages models (minimax, qwen3, …) keep the manual path.
- anthropic_adapter: _supports_adaptive_thinking / _supports_xhigh_effort /
_forbids_sampling_params now default unknown Claude models to modern; legacy
families enumerated in _LEGACY_MANUAL_THINKING_CLAUDE_SUBSTRINGS.
- openrouter profile: omit reasoning entirely (→ adaptive default) instead of
forwarding {enabled:false} for reasoning-mandatory Anthropic models; legacy
Anthropic + all non-Anthropic models still pass the disable form through.
- model_metadata + output-limit table: register claude-fable-5 (1M ctx, 128K out).
Tests assert the invariant ("unknown Claude model -> modern contract; legacy
stays manual; non-Claude unaffected"), not specific model names.
OpenRouter-routed slugs that are absent from models.dev (e.g. a freshly
shipped anthropic/claude-fable-5) fell through to the generic
DEFAULT_CONTEXT_LENGTHS["claude"]=200K entry and under-reported their real
1M window. The step-6 OpenRouter live-metadata fallback was gated on
`not effective_provider`, but an OpenRouter selection sets
effective_provider="openrouter" (inferred from the base URL), so that
branch was dead code for every OR model.
Add a dedicated step-5 OpenRouter branch that consults the live /models
catalog (authoritative, refreshes as new slugs ship) before models.dev and
the hardcoded family defaults — mirroring the existing Nous/Copilot/GMI
branches. Keeps the Kimi-family 32k underreport guard. Per-model values are
respected (claude-haiku-4.5 stays 200K), so it does not blanket-bump to 1M.
Regression tests cover the fable-5 case, the genuinely-200k case, and the
Kimi guard.
A Responses-API-shaped payload carrying instructions=/input=/store=/
parallel_tool_calls= can reach the native Anthropic messages.stream() /
messages.create() call under a rare api_mode-flip race (e.g. a concurrent
auxiliary vision call mutating a shared agent between the kwargs build and
the stream dispatch). The Anthropic SDK rejects these with a non-retryable
TypeError that kills the whole turn and propagates the entire fallback chain.
Add sanitize_anthropic_kwargs() at both Anthropic dispatch sites: it drops
the Responses-only keys in place and logs a WARNING (with #31673 breadcrumb)
when one is present, so the underlying race stays visible in the wild
instead of being silently papered over.
When agent.interrupt() fires during an active LLM call, the main poll loop
force-closes the worker-local httpx client to stop token generation. That
raises a transport error (RemoteProtocolError) on the worker thread — the
EXPECTED consequence of our own close, not a network bug.
The streaming retry loop misclassified it as a transient connection error
and retried; each doomed retry stalled for the full stream-stale timeout
(up to 300s). Because the gateway caches AIAgent instances per session, the
stale worker outlived the interrupted turn and raced the next turn's request
on shared client state — the root of the multi-minute cascading-interrupt
hang reported in the wild.
Fix: a request-local _request_cancelled token set by the poll loop right
before the force-close, in both interruptible_api_call (non-streaming) and
interruptible_streaming_api_call. The worker's exception handler checks the
token and exits cleanly — no retry, no fallback, no 'reconnecting' status —
instead of treating the forced error as transient. The token is request-
local (not agent._interrupt_requested, which is cleared at turn boundaries)
so a stale worker outliving its turn still recognizes its own forced close.
Original diagnosis and fix by @kristianvast (PR #6600), against the then-
inline methods in run_agent.py. Those were since extracted into
agent/chat_completion_helpers.py, so the fix is reapplied there.
Co-authored-by: Kristian Vastveit <kristianvast@users.noreply.github.com>
A misconfigured/slow external memory provider could hold the agent in
the 'running' state for minutes after the final response was delivered.
MemoryManager.sync_all / queue_prefetch_all looped provider.sync_turn /
queue_prefetch INLINE on the turn-completion path; a provider making a
blocking network/daemon call (a broken Hindsight daemon was observed
blocking ~298s before failing) blocked run_conversation from returning.
Because every interface (CLI, TUI, gateway) marks the agent 'running'
until run_conversation returns, the agent stayed busy for the full block
and any follow-up message triggered an aggressive interrupt that dropped
the message.
Dispatch provider sync/prefetch to a lazily-created single-worker
background executor. sync_all / queue_prefetch_all return immediately;
work completes (or fails, logged) in the background. A single worker
serializes writes so turn N lands before turn N+1. flush_pending()
provides a barrier for session boundaries and deterministic tests.
shutdown_all() drains the executor with a bounded timeout so a wedged
provider can never hang teardown.
Builtin-only / no-provider sessions spawn no executor (zero new threads
in the common case).
A one-off transient transport failure (streaming-close / incomplete
chunked read / 5xx / 408) on an auxiliary LLM call escalated straight to
provider/model fallback (or, for context compression, dropped the summary
and entered cooldown), even when an immediate retry on the same provider
would have succeeded.
Add a single same-target retry at the top of call_llm() and
async_call_llm() — before the existing except-chain — gated on a new
_is_transient_transport_error() that reuses the canonical
_is_connection_error() detector plus a 5xx/408 status check. A second
failure (or any non-transient error: auth, other 4xx, malformed payload)
falls through to first_err and the existing fallback handling unchanged.
This lives in call_llm so every auxiliary task (compression, memory flush,
title generation, session search, vision) shares one transient-retry
surface, rather than each caller re-implementing it. The context
compressor needs no change — it calls call_llm and inherits the retry; its
existing fallback-to-main path (#18458) now composes naturally (retry the
aux model once, then fall back to main only if the retry also fails).
Co-authored-by: ARegalado1 <alberto.regalado@ymail.com>
run_conversation's inner retry loop tracked recovery state in ~15 scattered
bare booleans (per-provider OAuth refresh guards, format-recovery guards,
restart signals). They are now fields on a single TurnRetryState dataclass the
loop mutates in place (_retry.<flag>), giving the recovery bookkeeping a named,
testable home.
Loop-control vars (retry_count, max_retries, max_compression_attempts) stay as
plain locals — they're while-mechanics, not recovery bookkeeping.
Behavior-neutral: pure local→attribute rewrite of 42 references; kwarg NAMES
preserved (e.g. has_retried_429=_retry.has_retried_429). Live simple + tool
turns OK.
Validation: tests/run_agent/ 1615 passed / 0 failed under per-file process
isolation; new test_turn_retry_state.py pins the field contract.
When context compaction rotates agent.session_id, it updates the gateway/tools
session context (set_current_session_id -> HERMES_SESSION_ID env + ContextVar)
but never updates the separate logging session context. The [session_id] tag on
log lines comes from hermes_logging._session_context (set once per turn in
conversation_loop.py), so post-compaction log lines in the same turn carry the
STALE old id while the message/DB/gateway state carry the new one — breaking log
correlation exactly at the compaction boundary.
Call hermes_logging.set_session_context(agent.session_id) alongside the existing
set_current_session_id, guarded so a logging failure can't regress the routing
update. Logs-only; no runtime or caching impact.
Refs #34089
The curator's idle-archival path (apply_automatic_transitions under
prune_builtins) could archive the bundled `plan` skill, killing the
/plan slash command silently — typing /plan then returned 'Unknown
command' with no signal that a skill had vanished. The archived skill's
hash stays in .bundled_manifest, so 'hermes update' wouldn't re-seed it.
Add PROTECTED_BUILTIN_SKILLS ({plan}) enforced at the master gate
is_curation_eligible() (covers archive_skill + the transition walk) and
in the candidate enumerator (so the LLM consolidation pass never sees
them). Immune to prune_builtins, pin state, and LLM judgment.
The auxiliary Codex adapter maintained its own chat->Responses conversion
loop that forwarded every non-system message's role verbatim into
Responses input[]. When flush_memories()/compression replayed session
history containing assistant tool_calls + role=tool results, those tool
messages leaked into the request and the Responses API rejected them with
HTTP 400: Invalid value: 'tool'.
Route _CodexCompletionsAdapter.create() through the same shared converter
the main agent transport uses (_chat_messages_to_responses_input), so tool
calls become function_call items and tool results become function_call_output
items with a valid call_id. Single conversion path means no future drift.
Also remove the now-dead _convert_content_for_responses() helper — its only
caller was the private conversion loop this change deletes.
Co-authored-by: ProgramCaiCai <techxacm@gmail.com>
Phase 1 of the god-file decomposition plan. run_conversation's ~470-line
once-per-turn setup block (stdio guarding, retry-counter resets, user-message
sanitization, todo/nudge hydration, system-prompt restore-or-build,
crash-resilience persistence, preflight compression, the pre_llm_call hook, and
external-memory prefetch) is moved verbatim into build_turn_context(), which
returns a TurnContext dataclass the loop unpacks.
Behavior-neutral move-and-name refactor: the builder mutates `agent` exactly as
the inline code did; only the locals the loop reads back are returned.
- run_conversation: 4602 -> 4217 LOC (-385)
- agent/conversation_loop.py: 4965 -> ~4580 LOC
- new agent/turn_context.py: focused, dependency-injected, unit-tested in isolation
Tests: tests/run_agent/ 1570 passed / 0 failed under per-file process isolation.
Relocation follow-ups: 413_compression mocks now patch both module references;
nudge/on_turn_start source-inspection guards point at the extracted module.
When a cron or background session compacts, it sets _previous_summary for
iterative updates. If that session ends without /new or /reset (which calls
on_session_reset()), the stale summary survives on the ContextCompressor
instance. A subsequent live messaging session's compaction then injects it as
'PREVIOUS SUMMARY:' into the summarizer prompt — contaminating the live
session with unrelated content from the prior session.
Add an else guard in compress(): when no handoff summary is found in the
current messages but _previous_summary is non-empty, discard it so
_generate_summary() starts fresh instead of iteratively updating a stale
cross-session summary.
Fixes#38788
The per-session compression lock prevents same-window concurrent forks but
not cross-turn ones: the background-review fork shares the parent's
session_id, so if it won a compression race its new child session was never
adopted by the gateway (the fork is single-lifecycle). The next foreground
turn then started from the stale parent and compressed it again, leaving the
same parent with two sibling children.
Set review_agent.compression_enabled = False so the fork never triggers
compression. Both trigger sites in conversation_loop.py gate on
compression_enabled before calling _compress_context, so the fork can never
rotate the shared parent. Review needs full context anyway — compressing
would degrade the memory/skill summary.
The per-session lock is kept as defense-in-depth for any future shared-session
path. Adds a regression test that fails without the flag and passes with it.
Closes#38727
Problem: get_model_context_length() had an early return at the end of the
custom-endpoint probe branch (step 3) that returned DEFAULT_FALLBACK_CONTEXT
(256K) without ever consulting the hardcoded DEFAULT_CONTEXT_LENGTHS catalog
(step 8). Models served through a custom/proxied gateway (e.g. corporate
Anthropic proxy) that didn't expose Ollama or local-server endpoints would
hit this path and get capped at 256K, even when the model name clearly
matched a known entry in the catalog (e.g. claude-opus-4-8 → 1M).
Changes:
- agent/model_metadata.py: Before returning DEFAULT_FALLBACK_CONTEXT at the
end of the custom-endpoint branch, consult DEFAULT_CONTEXT_LENGTHS using
the same longest-key-first fuzzy matching as step 8. Only fall through
to 256K if no catalog entry matches.
- tests/agent/test_model_metadata.py: Updated existing test and added new
test covering the custom-endpoint → catalog fallback behavior.
Fixes#38865
_supports_vision_override() in image_routing.py checked model.supports_vision
and providers.<name>.models, but not the legacy list-style custom_providers
config. A custom provider entry like:
custom_providers:
- name: my-provider
models:
my-model:
supports_vision: true
was ignored, causing image_input_mode=auto to route through the auxiliary
vision_analyze path instead of natively attaching images.
Fix: added a lookup step for custom_providers list entries, matching by
provider name (including 'custom:<name>' variants at runtime).
providers.<name>.models still takes precedence over custom_providers.
13 new tests covering: true/false override, custom: prefix matching,
no-match fallback, non-dict entries, empty lists, models key missing.
* feat(onboarding): opt-in structured profile-build path on first contact
On a user's very first gateway message, Hermes now optionally offers to
build a short profile of them — then, only with consent, gathers durable
facts and persists them to the user-profile memory store (memory tool,
target="user") so future sessions start already knowing who they are.
Inspired by Poke's zero-input onboarding, but consent-first by design:
- The agent OFFERS, never assumes. Declining stops it immediately.
- Before ANY external lookup it states what it will look up and asks.
- It never reads connected accounts (email/calendar) silently — the
exact privacy concern that made naive implementations feel invasive.
Wiring reuses existing infrastructure end-to-end:
- gateway/run.py first-message hook (was a plain self-intro) now swaps in
the profile-build directive when enabled and not yet offered.
- agent/onboarding.py gains profile_build_mode()/profile_build_directive()
+ PROFILE_BUILD_FLAG, latched once via the existing onboarding.seen
mechanism so the offer fires at most once per install.
- config default onboarding.profile_build: "ask" (set "off" to disable).
Added to an existing section, so no _config_version bump needed.
No new storage layer, no new injection path, no prompt-cache impact.
* fix(dashboard): fold onboarding into agent tab to avoid 1-field category
onboarding.profile_build is the only schema-surfaced onboarding field
(onboarding.seen is an internal latch dict), so the dashboard CONFIG_SCHEMA
single-field-category invariant rejected it. Merge onboarding -> agent like
the other small categories.
Compaction summaries now receive the current date and instruct the
summarizer to rewrite completed actions as absolute, dated, past-tense
facts (e.g. "email John about the proposal" -> "Sent the proposal email
to John on 2026-06-07"). A resumed conversation no longer re-issues work
that already happened or treats a finished action as still pending.
The date is resolved via hermes_time.now() (date-only, user-configured
timezone) inside _generate_summary. The compaction summary is a
mid-conversation message that is never part of the cached prefix, so the
date does not affect prompt-cache stability. Date resolution is
best-effort: a clock failure omits the rule rather than blocking
compaction. The rule rides the shared template, so both first-compaction
and iterative-update prompts carry it.
Inspired by Poke's summarization (temporal anchoring + semantic
preservation).
The salvaged main-agent fix (sanidhyasin) applies model.default_headers
to the primary OpenAI client, but the auxiliary client (title generation,
context compression, vision routing) builds its own clients and did not
read the override. For a `provider: custom` endpoint behind a gateway/WAF
that rejects the OpenAI SDK's identifying headers, the main turn would
succeed while auxiliary calls to the same endpoint still failed with the
opaque 502/4xx from #40033.
Add agent.auxiliary_client._apply_user_default_headers() (user values win
over provider/SDK defaults; no-op when unconfigured) and apply it at every
OpenAI-wire client construction site:
- _try_custom_endpoint() — config-level `model.provider: custom`
- the named custom-provider branch (custom_providers/providers entries),
including the anthropic-SDK-missing OpenAI-wire fallback
- the api-key-provider, async-conversion, and main resolve_provider_client
fallback branches
To prevent the two clients ever drifting on precedence/value handling,
AIAgent._apply_user_default_headers (run_agent.py) now delegates the config
read + merge to this shared helper (run_agent already imports from
auxiliary_client). Native Anthropic/Bedrock branches are untouched (they
don't use the OpenAI wire).
8 new tests (helper semantics + config-level custom + named custom);
full aux + attribution header suites green (295).
compress_context() sets last_prompt_tokens=-1 right after compression to
mark "no real API usage yet". The preflight display-seed used
`_preflight_tokens > (last_prompt_tokens or 0)`, and `(-1 or 0)` is -1
(truthy), so any positive rough estimate clobbered the sentinel with a
schema-inflated count — re-triggering compression on the next turn.
Treat any negative value as "no real data yet" and skip the seed.
Salvaged from #40246 as the minimal root-cause fix. The original also
added an `_awaiting_suppression_count` bounded-window state machine to
should_compress() across 3 files; left out here to keep blast radius
small — the sentinel guard alone fixes the re-fire. The suppression
window can be added separately if the usage=None-stub edge case warrants it.
Co-authored-by: davidgut1982 <davidgut1982@users.noreply.github.com>
The ChatGPT Codex OAuth backend hard-caps gpt-5.5 at a 272K context window
(verified live: a ~330K-token request to chatgpt.com/backend-api/codex/responses
is rejected with context_length_exceeded while ~250K succeeds; the same slug
exposes 1.05M on the direct OpenAI API / OpenRouter and 400K on Copilot). At the
default 50% trigger, auto-compaction fires at ~136K — half the usable window.
Raise the trigger to 85% (~231K) on this exact route only, gated by a new
compression.codex_gpt55_autoraise config flag (default true). When it fires,
emit a one-time notice (CLI inline print + gateway status_callback replay) with
the exact opt-back-out command. gpt-5.5 on any other provider keeps the user's
global threshold.
- _is_codex_gpt55() matches the 5.5 family only on provider=openai-codex
- _compression_threshold_for_model() now provider-aware + opt-out param
- config key + _config_version bump (27->28) for backfill
- docs + tests (40 cases in test_arcee_trinity_overrides.py)
* feat(tui): HERMES_DEV_CREDITS live-spend dev readout (L0 tracer for usage-aware credits)
L0 of the usage-aware-credits feature: a dev-only, env-gated tracer that
exercises the real header -> CreditsState -> TUI pipe end-to-end behind
HERMES_DEV_CREDITS, de-risking the L1/L5 build before the notice policy exists.
- agent/credits_tracker.py: CreditsState + parse_credits_headers (headers are
strings -> paid_access via == "true", never bool(); retain-last-known; only
subscription_micros may be negative; *_usd kept verbatim).
- run_agent.py: _capture_credits / get_credits_state / get_credits_spent_micros,
session-start baseline latch, + dev-gated "credits" capture log.
- agent/chat_completion_helpers.py: capture on the streaming response.
- agent/agent_init.py: init _credits_state + _credits_session_start_micros.
- tui_gateway/server.py: _get_usage emits dev_credits_spent_micros only when flagged.
- ui-tui appChrome.tsx / types.ts: cents delta status segment + "(dev credits)" banner.
Off by default; silent for normal users. Validated live against staging
(capture log delta matches the TUI segment). Throwaway consumer (readout/log/
banner); credits_tracker + the capture plumbing are the real feature foundation.
* test(credits): lock parser under 9-state matrix + harden validation (L2)
Add tests/agent/test_credits_tracker.py with 92 tests covering the 9-state
matrix (healthy, sub_90pct, grant_exhausted, purchased_only, tool_pool_free,
depleted, debt, missing, no_org) plus validation edge cases: version strict==1
with warn-once latch for v>1, bool-string trap (paid_access/tool_pool_gated_off
== "true"/"false", never bool()), half-pair subscription limit treated as
both-absent while parse succeeds, USD regex ^-?\d+\.\d{2}$, non-int micros
→ None, negative non-subscription micros → None, as_of_ms junk → None, zero
limit ZeroDivision guard.
Harden agent/credits_tracker.py to match the spec:
- Add tool_pool_micros/tool_pool_gated_off/from_header fields to CreditsState
- Add depleted property (== not paid_access, never remaining==0)
- Change used_fraction guard to key off subscription_limit_micros (the actual
denominator) not denominator_kind (metadata)
- Replace fail-soft _safe_int with a sentinel-returning variant; full validation
now returns None on any malformed field rather than silently defaulting
- Add module-level warn-once latch for version > 1
- Add USD regex validation; add denominator_kind allow-list check
- Parse x-nous-tool-pool-* prefix headers (not x-nous-credits-tool-pool-*)
* feat(credits): notice spine — AgentNotice + notice_callback/notice_clear_callback + TUI binding (L1)
L1 of usage-aware credits: the driver-agnostic notice delivery spine that L4's
policy will fire through and L5's TUI render will consume.
- agent/credits_tracker.py: AgentNotice dataclass (text/level/kind/ttl_ms/key/id;
kind defaults "sticky", kept TTL-expressive for a future config seam).
- run_agent.py: AIAgent gains notice_callback + notice_clear_callback slots and
_emit_notice / _emit_notice_clear emitters (swallow all callback errors — a
notice must never break the agent loop; no-op when unbound).
- agent/agent_init.py: thread both callbacks through init_agent.
- tui_gateway/server.py: bind both in _agent_cbs → notification.show / notification.clear
WS events (snake_case payload, matching the existing gateway-event convention).
- ui-tui/src/gatewayTypes.ts: notification.show / notification.clear arms on GatewayEvent.
- tests/run_agent/test_notice_spine.py: 15 tests (emitter fire + fail-open + no-op,
signature threading, TUI binding payload shape).
Messaging push is out of v1 (binds neither callback). CLI binding + the TUI render/
decode land with L4 (firing) and L5 (render) so turn-end flush is wired correctly.
* feat(credits): threshold reconciliation policy + tests (L4.1)
* feat(credits): wire threshold policy into capture + latch (L4.2)
After a fresh header parse, _capture_credits runs evaluate_credits_notices against
the agent's _credits_latch and emits the result — clears first, then shows (so a
recovered depletion clears before the "restored" success lands, and depleted wins
the latest-wins slot). Gated on a bound notice_callback: messaging (no callbacks)
still caches state for /usage but runs no policy. Parse stays fail-open (miss →
keep last-known); the eval/emit path warns on failure rather than swallowing, so a
depletion-notice bug can't vanish silently.
- run_agent.py: _capture_credits split into parse (swallow→miss) + policy (warn);
latch lazy-guarded (object.__new__ safety).
- agent/agent_init.py: init agent._credits_latch = {"active": set(), "seen_below_90": False}.
* feat(tui): render credits notices in the status bar (L5, Strategy B)
The TUI now renders the notification.show / notification.clear gateway events the
agent emits — a level-colored notice overrides the status/verb slot when not busy.
- Notice state machine on turnController (pendingNotice + dedicated noticeTimer +
show/clear/applyNotice/flushPendingNotice/clearNoticeState). createGatewayEventHandler
decodes the events and delegates.
- Render priority busy > notice > status (appChrome StatusRule); notice text rendered
verbatim (its glyph comes from the policy), shrinkable so it never clips model│ctx;
dev-credits banner + Δ segment preserved. UiState.notice is snake_case (matches wire).
- Busy-wins: a notice arriving mid-turn is held and flushed at the THREE turn-end sites
(recordMessageComplete / interruptTurn / recordError) — never idle(), which reset()
also calls (would leak across sessions); reset() clears instead.
- Dedicated noticeTimer (never statusTimer); TTL starts on visibility with an id-guard;
latest-wins cancels the prior timer; clear is key-matched (no-op on mismatch); a sticky
survives a turn (flush no-ops with no pending); session reset clears (no cross-session leak).
- 20 tests (handler/turnController logic incl. R3-C2 timer isolation + render priority).
* feat(credits): cold-start seed for new Nous sessions (L3)
A genuinely-new Nous session has no inference header yet, so seed credits state from
the authoritative GET /api/oauth/account snapshot at session start (in the new-session
branch of _restore_or_build_system_prompt — inline, since the on_session_start plugin
hook gets no agent reference). The seed runs the shared notice policy, so a session that
opens already depleted warns IMMEDIATELY rather than only after the first turn.
- Maps the nested account fields (paid_service_access → paid_access; total_usable /
subscription / purchased on paid_service_access_info; rollover on subscription), each
None-guarded; float dollars → micros via round(d*1e6), *_usd left "" (render formats
from micros — never synthesize a verbatim usd from a float).
- Magnitudes-only: no monthlyCredits on the endpoint → subscription_limit_* unset →
used_fraction None → no warn90 from the seed (% only once a header lands, per D-E).
- Provider-guarded to Nous; fail-open (any error leaves _credits_state None, never
blocks startup); paid_access unknown ⇒ True (never falsely depleted).
- run_agent.py: extracted the warm-path policy/emit block into a shared
_emit_credits_notices() so capture and the seed fire notices identically.
* feat(credits): /usage Nous credits magnitudes view + recovery trigger (L6)
Add Nous credit dollar magnitudes to /usage (subscription / top-up / total
+ rollover + renewal + portal CTA), magnitudes-only per v1 (no % until the
account endpoint exposes a denominator). Reuses the existing account-usage
render machinery via a new pure build_nous_credits_snapshot() that maps a
NousPortalAccountInfo to an AccountUsageSnapshot; no nous branch is added to
fetch_account_usage (keeps the per-provider boundary intact).
CLI /usage also doubles as a depletion-recovery trigger: a force_fresh
account fetch, kept in a SEPARATE local so it never clobbers the
header-sourced agent._credits_state (which alone carries used_fraction). If
paid access recovered while credits.depleted is latched and a notice
consumer is bound, it reuses agent._emit_credits_notices() to clear it.
Gateway /usage displays magnitudes only — messaging binds no notice
consumer, so it performs no recovery emit.
Fail-open throughout: any portal hiccup leaves /usage unaffected.
* refactor(credits): dedupe HERMES_DEV_CREDITS flag parse via shared helpers
The dev-flag truthy check was inlined in three places. Replace with the shared
utils.is_truthy_value (run_agent.py, tui_gateway/server.py — also drops a
redundant inline `import os`) and a hoisted DEV_CREDITS_MODE export in
ui-tui/src/config/env.ts (consumed by appChrome, which also stops recomputing the
env check on every render). Behaviour-preserving; identical truthy set.
* fix(credits): cut dead /usage recovery trigger + bound portal fetches (L6 review)
Adversarial review found the /usage depletion-recovery trigger dead AND broken:
the CLI binds no notice_clear_callback, the TUI runs /usage in a separate
slash-worker subprocess (its own agent/latch), and the no-clobber rule made it
evaluate stale paid_access anyway. Recovery already happens on the next inference
(warm path), so the trigger was redundant — remove it and stop the depleted
notice over-promising.
- cli.py: remove the dead recovery block; bound the /usage portal fetch with a
10s wall-clock timeout (ThreadPoolExecutor) like the per-provider fetch —
urllib's per-socket timeout is not a wall-clock guarantee.
- agent/credits_tracker.py: reword the depleted CTA to "run /usage for balance"
(no false recovery promise; /usage shows fresh magnitudes, sticky clears next turn).
- agent/conversation_loop.py: same wall-clock timeout on the cold-start seed fetch
so a stalled portal can't hang session startup; tidy its time import.
* chore(credits): dev notice-state fixtures (HERMES_DEV_CREDITS_FIXTURE)
Throwaway dev scaffolding to exercise the notice pipeline without real spend or
Redis seeding. Set HERMES_DEV_CREDITS_FIXTURE to a state name (healthy / sub_90pct
/ grant_exhausted / depleted / clear) or a file path whose contents name a state
(re-read each turn → flip states live for recovery testing). _capture_credits
injects the chosen CreditsState instead of parsing real headers and runs the
shared notice policy. Deletable with the rest of the HERMES_DEV_CREDITS scaffolding.
* feat(credits): /usage monthly-grant % gauge
The portal /api/oauth/account subscription block now carries monthly_credits
(the per-period grant allowance, the % denominator). The consumer parsed
monthly_charge but dropped monthly_credits, so /usage stayed magnitudes-only.
Capture monthly_credits into NousPortalSubscriptionInfo + _subscription_from_payload.
build_nous_credits_snapshot emits a Subscription usage window (real % used, routed
through the existing render machinery) when monthly_credits is a finite positive
denominator and credits_remaining is finite and <= cap; otherwise it degrades to
magnitudes-only (older portals, rollover-over-cap, or non-finite payloads).
Guards (adversarial-review-driven): reject non-finite operands (json.loads parses
bare NaN/Infinity by default → would render $nan + a false 100% used), reject
bools, guard div-by-zero (cap>0), and suppress the gauge when remaining > cap
(rollover spanning the period makes the cap a nonsensical denominator → the
$X-of-$Y detail would read as a contradiction). Debt (remaining<0) clamps to 100%.
Money rule preserved: the ratio + magnitudes are computed from numeric float
account fields via display formatting, never by parsing a server *_usd string
(there are none on these dataclasses).
13 gauge tests added (tests/agent/test_nous_credits_gauge.py).
* fix(credits): show /usage Nous block whenever a Nous account is present
/usage runs in a slash-worker subprocess whose resolved inference provider is
often not "nous" even when the user has a Nous account, so gating the Nous
credits block on (provider == "nous") hid it entirely — the account data was
fully available but never rendered.
Gate instead on "a Nous account is logged in": a cheap local auth-state lookup
(get_provider_auth_state('nous') has an access_token) decides whether to attempt
the portal fetch, regardless of which provider inference runs on. In the gateway
the block is also lifted out of the 'if provider:' scope so a Nous-credentialled
user with another (or no) resident inference provider still sees their balance.
Fail-open and the per-fetch wall-clock timeout are preserved.
* fix(credits): show /usage Nous block when there's no live agent (TUI slash-worker)
In the TUI, /usage runs in a slash-worker subprocess that resumes the session
WITHOUT building an agent (self.agent is None), so _show_usage early-returned
"(._.) No active agent" before ever reaching the Nous credits block — which is
agent-independent (a portal fetch gated on Nous auth-state). Extract the block
into _print_nous_credits_block() and run it at the no-agent / no-calls
early-returns too (returns True if it printed, so the fallback message only
shows when there's genuinely nothing).
Verified live against staging: the block + monthly-grant gauge now render in the
slash-worker /usage path (previously hidden). The plain CLI REPL + messaging
paths are unchanged (they have a live agent).
* feat(credits): escalating 50/75/90 usage bands (single status line)
Replace the lone 90%-used warning with three escalating bands (50 info, 75 warn,
90 warn) shown as ONE status-bar line: it displays the highest band the
subscription grant has crossed, replaces the line as usage climbs, steps back
down on recovery, and clears below 50%. No stacking, no per-turn churn.
Bands live in a tunable CREDITS_USAGE_BANDS list; the policy derives everything
from it. Single notice key (credits.usage) with a usage_band latch field so the
notice only re-emits when the band actually changes. The crossing gate
(seen_below_90) is preserved so a fresh live session that opens mid-range stays
quiet until it has been observed below the lowest band (cold-start primes it when
it wants an open-high warning). Denominator math unchanged: % = subscription
grant burn (cap - grant_remaining)/cap, clamped [0,1]; top-up never moves the %.
Migrated test_credits_policy.py to the new key + added TestUsageBands (climb,
step-down, recovery-clear, idempotent, inclusive boundaries).
* feat(credits): hydrate notices at session OPEN via shared seed (TUI + first-turn)
Notices previously only fired inside a conversation turn (first message), so a
session that opened already depleted / past a usage band showed nothing at
'ready'. Extract the cold-start seed into a shared seed_credits_at_session_start()
and call it (a) in the TUI/desktop agent build right after the notice callback is
wired (fires at 'ready', before any message) and (b) as the first-turn fallback in
conversation_loop. Idempotent (skips once _credits_state exists) and fail-open.
The seed now maps monthly_credits -> subscription_limit_micros +
denominator_kind='subscription_cap', so used_fraction is computable at seed time
and usage-band warnings (not just depletion) hydrate on open. Primes the crossing
latch so a session opening already in a band warns immediately. Degrades to
depletion-only when monthly_credits is absent (older portals).
Adds test_credits_cold_start.py covering open-at-band, depletion, debt, no-cap
degradation, and the shared seed (fires/idempotent/skips-non-nous).
* feat(credits): /usage monthly-grant % gauge + fixture support + TUI surfacing
agent/account_usage.py: build_nous_credits_snapshot emits a subscription %% gauge
when the portal supplies a positive, finite monthly_credits denominator with
remaining <= cap (guards reject NaN/Infinity and rollover-over-cap, which would
render $nan or a contradictory $X-of-$Y); degrades to magnitudes-only otherwise.
Adds shared nous_credits_lines() (auth-gated, wall-clock-bounded portal fetch) so
the CLI and TUI /usage render the same block, and _snapshot_from_credits_state()
so HERMES_DEV_CREDITS_FIXTURE drives /usage offline too.
TUI: session.usage RPC carries credits_lines (agent-independent) and the /usage
panel renders them regardless of API-call count or resume state — previously the
TUI's separate /usage implementation only showed token counts.
Money rule preserved: %% and magnitudes come from numeric float account fields via
display formatting, never by parsing a server *_usd string.
* feat(credits): CLI REPL inline notices (parity with TUI)
The plain CLI agent bound no notice callbacks, so credit notices were TUI-only.
Bind notice_callback/notice_clear_callback on the CLI AIAgent; _on_notice renders
a single level-colored line above the prompt (error red / warn yellow / success
green / info dim) via _cprint, and seed credits at session open so a depletion or
usage-band warning shows before the first message — the same hydration the TUI
got. _on_notice_clear is a no-op (the REPL prints lines, no persistent slot).
* test(credits): add sub_50pct + sub_75pct dev fixtures for the new usage bands
The fixture set jumped 10%% -> 90%%; add sub_50pct (uf 0.5 -> band 50 info) and
sub_75pct (uf 0.75 -> band 75 warn) so the new escalating bands are exercisable
via HERMES_DEV_CREDITS_FIXTURE across all three surfaces (notice, session-open
seed, /usage gauge).
* fix(credits): usage-band notice clears on next prompt (not sticky-forever)
A 50/75/90 usage heads-up was sticky and camped the status bar indefinitely. Clear
the visible credits.usage notice when a new turn starts (startMessage), so it shows
until your next prompt then yields. The server latch is unchanged, so it won't
re-nag at the same band — it only re-shows when the band actually changes (climb)
or clears when usage drops below the lowest band. Depletion stays sticky.
* refactor(credits): consolidate the /usage credits block behind nous_credits_lines()
The CLI (_print_nous_credits_block) and the messaging gateway (_handle_usage_command)
each re-implemented the auth-gate + portal fetch + render, and both bypassed the
dev-fixture short-circuit that only the TUI honored — so /usage ignored
HERMES_DEV_CREDITS_FIXTURE on the CLI and in chat. Route both through the shared
agent.account_usage.nous_credits_lines() helper: one fetch/render path, one auth
gate, and the fixture works on every surface (~60 fewer duplicated lines).
The gateway usage test recorded only the last asyncio.to_thread call; /usage now
dispatches both the account fetch and the credits fetch, so it records every call
and matches the account fetch by its provider arg.
* fix(credits): keep the /usage gauge type-safe and log its fail-open path
_is_finite_num is now a TypeGuard[float], so the type checker narrows the gauge
operands (monthly_credits / credits_remaining) and the magnitudes passed to
_fmt_usd through it — no more None-operand warnings on the arithmetic. Add a debug
breadcrumb on the nous_credits_lines portal-fetch fail-open so a dead /usage block
is diagnosable in agent.log without a dev flag.
* fix(credits): harden the header tracker — prod-leak gate, hot-path probe, fire-and-forget seed
- Prod-leak guard: dev fixtures (HERMES_DEV_CREDITS_FIXTURE) now also require
HERMES_DEV_CREDITS, so a stray fixture var can't surface fabricated balances on a
real account. Matches the documented run workflow (both vars set together).
- Hot-path probe: parse_credits_headers checks for the version sentinel header
before allocating a lowercased copy of the response headers — skips that work on
every non-Nous API call. Behaviour-identical and still case-insensitive.
- Fire-and-forget seed: the real portal fetch in seed_credits_at_session_start now
runs in a daemon thread, so a slow/unreachable portal never delays session "ready"
(previously blocked up to 10s). The dev-fixture path stays synchronous; the thread
re-checks idempotency before hydrating (a live header may land first).
- Diagnostics: debug breadcrumbs on the parse and seed fail-open paths so a crashed
parser / dead seed is distinguishable from a legitimate no-headers miss.
Cold-start tests set HERMES_DEV_CREDITS alongside the fixture to match the gate.
* test(tui): fix env-timing in the StatusRule dev-credits assertion
DEV_CREDITS_MODE is read once at module load (config/env), so mutating
process.env.HERMES_DEV_CREDITS inside the test couldn't flip it — the dev-banner
assertion only passed if the env was exported before vitest started, and failed in a
normal run. Move that assertion to a sibling file that mocks config/env with
DEV_CREDITS_MODE: true (scoped, no module-reset / React-identity hazard).
* test(credits): cover the dev-fixture /usage render and usage-band clear-on-prompt
- _snapshot_from_credits_state (the offline /usage renderer) had no direct test:
lock the gauge math, the verbatim *_usd magnitudes, the depletion line and the
fixture marker, plus the no-cap (no gauge) and None-state cases.
- turnController.startMessage had no test for clearing the credits.usage notice on
the next prompt while leaving credits.depleted sticky.
* feat(credits): deliver credit notices over messaging gateways
Bind notice_callback/notice_clear_callback on the per-turn gateway agent
so usage-band / depletion / restored notices reach Telegram/Discord/Slack/
etc. Previously the messaging gateway bound neither callback, so the agent's
_emit_credits_notices early-returned and a chat user crossing a band got
nothing unless they ran /usage manually.
- render_notice_line(): AgentNotice -> single plaintext line (level glyph +
text), plaintext-only so it renders uniformly without per-platform escaping.
Fail-soft on malformed/empty notices.
- Standalone push for every notice (messaging has no persistent status bar):
route through the shared _deliver_platform_notice rail (honors private/
public delivery + thread metadata), scheduled onto the gateway loop via
safe_schedule_threadsafe from the agent's sync worker thread — same pattern
as _status_callback_sync.
- The fired-once latch lives on the cached (reused-in-place) agent and
persists across turns, so a band crosses once -> one push, no per-turn
re-nag. Re-fires only after idle-eviction rebuilds the agent (a reminder).
- Recovery ('Credit access restored') rides the show path (emitted as a
success notice, not a clear). notice_clear_callback is a no-op: a sent
platform message can't be cleanly retracted.
Tests: render glyph/levels/fail-soft + public/private delivery seam through
_deliver_platform_notice + no-adapter no-op.
* fix(credits): don't double the glyph on messaging notices
render_notice_line prepended a per-level glyph, but the notice policy already
bakes the glyph into the text (and the TUI + CLI render it verbatim) — so every
credit notice over messaging came out doubled ("⚠ ⚠ Credits 90% used",
"⛔ ✕ Credit access paused"). Emit the text verbatim instead; drop the now-dead
level→glyph map.
The render tests fed glyph-less text (and the success case only checked
startswith), so the doubling slipped through. Rework them around the verbatim
contract and add an end-to-end regression that runs real evaluate_credits_notices
output through render_notice_line and asserts the line is returned unchanged.
* fix: respect disabled auto-compaction on context overflow
Port from anomalyco/opencode#30749.
When compression.enabled is false, NO automatic compaction trigger may
fire. The proactive token-threshold paths (preflight + post-response
should_compress gate) already honoured the setting, but the three
provider-overflow recovery paths in the agent loop — long-context-tier
429, 413 payload-too-large, and context-overflow — called
_compress_context() unconditionally, silently compressing and rotating
the session against the user's explicit choice.
Add a single guard at the top of the overflow-recovery dispatch: when
compression is disabled and the error is one of those three overflow
classes, surface a terminal error (compaction_disabled: True) telling the
user to /compress manually, /new, switch to a larger-context model, or
reduce attachments. Manual /compress (force=True) is unaffected — it never
enters this loop.
Tests: new TestOverflowWithCompactionDisabled (413 + 400 overflow don't
compress when disabled; control case still compresses when enabled).
Existing overflow-recovery tests updated to enable compaction explicitly
(they verify the recovery fires); fixture defaults flipped to True to
match production (compression.enabled defaults to True).
* fix(gemini): default native maxOutputTokens + strip OpenAI extra_body on Gemini endpoints
Two distinct failures hit users on the gemini provider with only Google
AI Studio keys set.
1. Truncation loop: build_gemini_request() only set maxOutputTokens when
max_tokens was non-None. Hermes passes None to mean "unlimited", but
Gemini's native generateContent does NOT treat an absent maxOutputTokens
as full budget — it applies a low internal default and stops early with
finishReason=MAX_TOKENS, truncating tool calls. The agent then retries
3x and refuses the incomplete call. Now default to the published 65,535
ceiling (shared by all current Gemini text models) when max_tokens=None.
2. HTTP 400 on Gemini endpoint: the chat_completions transport assembles
profile extra_body (Nous portal 'tags', reasoning, provider prefs) and
sends it via the OpenAI client to whatever base_url is resolved. When a
profile that emits extra_body (e.g. Nous) is active but the endpoint is a
native Gemini base_url — typical when only Google creds exist and a
fallback/aux call lands on Gemini — Google rejects the unknown 'tags'
field with a non-retryable 400. Strip all non-thinking_config extra_body
keys when the resolved endpoint is native Gemini.
Verified E2E against real transport code: tags stripped on native Gemini,
preserved on Nous and the /openai compat endpoint; maxOutputTokens=65535
on None, explicit values respected.
Docker Compose service names (e.g. ollama, litellm, hermes-litellm)
are unqualified hostnames with no dots. These are always local — they
resolve via Docker DNS, /etc/hosts, or mDNS. Without this fix, the
stale stream timeout fires on local LLM proxies, causing infinite
reconnect loops.
Closes#7905
Three Copilot inline review comments on #37664, two worth landing
in a polish pass before merge:
1. auxiliary_client.py:270 — Copilot suggested keeping the
minimax-* entries in _API_KEY_PROVIDER_AUX_MODELS_FALLBACK as
a safety net for environments where the profile-based
resolution can't import or run plugin discovery. **Declined.**
The deepseek precedent (commit 773a0faca) explicitly removed
deepseek from the same dict for the same reason — the profile
layer is the source of truth and the dict is a legacy
pre-profiles-system fallback. We do not want to fragment the
codebase by provider: either the profile layer is authoritative
or the dict is. The minimax PR picks profile (matching deepseek)
and the dict stays cleaned up. The risk Copilot raises is
real but theoretical — plugin discovery runs at import time of
the providers module, which is the first thing any modern
Hermes entrypoint imports.
2. tests/agent/test_minimax_provider.py:162 — Copilot flagged
that the test class relies on _get_aux_model_for_provider()
resolving via provider profiles but doesn't explicitly trigger
plugin discovery. **Fixed.** Added 'import model_tools # noqa:
F401' at the top of both test_minimax_aux_is_standard and
test_minimax_aux_not_highspeed. The fixtures in the parallel
test_minimax_profile.py already did this; the legacy test in
test_minimax_provider.py was order-dependent and would silently
break if anyone reorganised the test ordering. Pinned the
dependency explicitly so the test is order-independent.
3. tests/plugins/model_providers/test_minimax_profile.py:46 —
Copilot flagged that the docstring referenced a hard-coded
line number 'hermes_cli/models.py:298' that would go stale.
**Fixed.** Replaced with the symbol reference
'hermes_cli.models._PROVIDER_MODELS[\'minimax\']' which is
stable under file edits and grep-friendly. The new docstring
also reads more naturally — readers don't have to look up
'what's at line 298' to follow the reasoning.
All 221 minimax-related tests still pass.
The minimax / minimax-cn / minimax-oauth profiles still advertised
M2.7 (and M2.7-highspeed for OAuth) as their default_aux_model,
predating the M3 release (2026-06-01). The user-facing
_PROVIDER_MODELS['minimax'] catalog top entry is M3, and the
recommended config for a Token-Plan install now sets
model.default: MiniMax-M3, so the aux default was the only
remaining drift.
Updates:
* minimax default_aux_model: M2.7 -> M3
* minimax-cn default_aux_model: M2.7 -> M3
* minimax-oauth default_aux_model: M2.7-highspeed -> M2.7
(M3 is not on the OAuth / Coding Plan tier per
platform docs as of this PR; the highspeed
variant was the 2x-cost regression from #4082
that PR #6082 collapsed to plain M2.7 for
minimax / minimax-cn but missed OAuth)
* agent/auxiliary_client.py: drop the three legacy
_API_KEY_PROVIDER_AUX_MODELS_FALLBACK entries for the minimax
family. _get_aux_model_for_provider() reads from
ProviderProfile.default_aux_model first (line 250) and only
falls back to the dict when the profile has no aux model or
the profile import fails. With the profile now set, the dict
entries are dead code and a drift hazard. Mirrors the deepseek
cleanup in 773a0faca.
* tests/agent/test_minimax_provider.py: update the existing
TestMinimaxAuxModel assertions from MiniMax-M2.7 to MiniMax-M3
(the intent — 'standard, not highspeed' — is unchanged; the
pin value is).
* tests/plugins/model_providers/test_minimax_profile.py: new
file mirroring tests/plugins/model_providers/test_deepseek_profile.py.
Pins each of the three profiles' default_aux_model and
asserts _get_aux_model_for_provider() returns it. A second
class guards against the highspeed regression coming back.
Refs:
- Closes#36196 in spirit (M3 support — the catalog half of
that issue is #36212; this PR covers the profile half)
- Related: #4082 (M2.7-highspeed 2x-cost), #6082 (previous
M2.7-highspeed -> M2.7 fix that missed OAuth + the
auxiliary_client.py fallback dict)
- Pattern: 773a0faca (same profile-layer fix for deepseek)
The salvaged conversion emitted type:"input_video", which MiniMax M3 rejects
just like the original video_url block. Per MiniMax's Anthropic-compat docs,
the video content block is type:"video" with an image-style source (base64 or
url). Fixes the block type, converts URL-based videos too, and adds 4 video
conversion tests (none shipped with the original PR).
The ``grok-4.3`` (1M context) catalog entry was added on 2026-05-15
(ce0e189d3). Between 2026-04-10 (when ``grok-4`` at 256,000 was first
added by b57769718) and 2026-05-15, grok-4.3 slugs resolved via the
generic ``grok-4`` substring catch-all and that 256,000 value was
persisted to context_length_cache.yaml. Users who first queried
grok-4.3 in that 35-day window are stuck at 256K forever — the cache
is read at step 1 before the hardcoded defaults in step 8, so the
correct 1M entry is never reached.
Mirror the existing Kimi/Codex/MiniMax-M3 stale-cache guards: add
_model_name_suggests_grok_4_3() and an elif branch that drops any
cached value ≤ 256,000 for a grok-4.3 slug so the next lookup falls
through to the 1M hardcoded default.
Adds 4 regression tests: helper unit test, stale-drop-and-re-resolve,
correct-cache-preserved, and no-clobber for plain grok-4 (256K correct).
kshitijk4poor
Brooklyn Nicholson
Teknium
Austin Pickett
HexLab98
Ben Barclay
sgaofen
lkevincc
Jack Maloney
Shashwat Gokhe
konsisumer
Chaz Dinkle
jun
LeonSGP43
qWaitCrypto
LeonSGP43
Bartok
ethernet
helix4u
Moonsong
DavidMetcalfe
yu-xin-c
rebel
Tranquil-Flow
herbalizer404
pyxl-dev
GodsBoy
benbenwyb
liuhao1024
texhy
Hao Zhe
Basil Al Shukaili
JackJin
Shannon Sands
pmos69
devorun
yeyitech
Sancho
alt-glitch
tt-a1i
Victor Kyriazakos
Siddharth Balyan
XVVH
Max Freedom Pollard
kyssta-exe
Wolfram Ravenwolf
mr-r0b0t
SHL0MS
chromalinx
ITheEqualizer
achaljhawar
ashishpatel26
Tranquil-Flow
Henrik Bentel
xxxigm
Erosika
RaumfahrerSpiffy
Matt Harris
emozilla
xxxigm
Ian Culling
rob-maron
Robin Fernandes
Xiangji
JP Lew
JimStenstrom
Basil Al Shukaili
islam666
Fearvox
AhmetArif0