hermes send "MEDIA:/x.png This Caption" now arrives as one native captioned
bubble instead of a separate text message followed by an uncaptioned bubble.
Root cause: the standalone senders (hermes send / cron / send_message tool)
stripped the MEDIA: tag, sent the remaining text as its own message, and
called the media send with no caption -- even though hermes send's help
advertises the captioned form and the bridges/adapters already support a
caption. Signal already captioned correctly.
- tools/send_message_tool.py: new _media_caption_split() chokepoint decides
caption-vs-separate-body (single captionable non-voice file within the
platform's message-length cap). Wired into the Telegram, WhatsApp and
Discord dispatch paths.
- Telegram/WhatsApp/Discord: when the single captioned file is missing, the
caption text is delivered as a plain message so it is never silently lost.
- Telegram caption send gets a MarkdownV2->plain parse fallback.
- Tests: _media_caption_split unit tests + per-platform caption tests
(ride, multi-file fallback, voice exclusion, over-limit fallback,
missing-file text fallback); updated the 3 tests that asserted the old
text-then-media split.
Closes the gap reported against #58911 (the MEDIA_CAPTION directive PR);
credit to @ferreiraesilva for surfacing the caption behavior.
Review finding: callers mutate the returned dicts in place —
hermes_cli/web_server.py annotates s['enabled']/s['usage'] on the skills
list — so handing out the cached objects poisons the cache for every
subsequent caller (and is a cross-thread shared-mutable hazard in the
gateway). Return [dict(s) for s in cached] on both hit and miss paths;
warm-path cost is negligible (241x speedup retained on a 300-skill
fixture). Regression test mutates a returned list/dict and asserts the
next cached call is clean.
Review findings on the cherry-picked cache (follow-up to #58985):
- The cache key was the max mtime of only the TOP-LEVEL scan dirs.
Adding/removing a skill inside a category subdir bumps the category
dir's mtime, NOT the root's, so the cache served a stale list
indefinitely. Replace with a per-dir signature covering roots +
immediate children (one scandir per dir; mirrors
hermes_cli/profiles.py::_count_skills from d5eee133e).
- The disabled-set is config-driven and changes with no filesystem
mtime bump; fold it into the signature so /skills disable takes
effect without a restart.
- Platform is part of the signature (gateway processes serve multiple
platform scopes; scan results are platform-filtered).
- Add a 30s TTL to bound staleness from in-place SKILL.md edits (file
mtime is invisible to any directory signature).
- The original also keyed dirs off the module-level SKILLS_DIR constant;
the scan itself uses _skills_dir() (live profile HERMES_HOME) — use
the same resolution for the signature.
Mutation-verified: nested-add, disabled-set, and TTL tests fail against
the pre-fix cache and pass with it.
_find_all_skills() re-reads every SKILL.md on every call, which is
wasteful when nothing changed between turns. Cache results keyed by
the max mtime across all scanned skill directories — a skill write
touches the directory, bumping mtime past the cached value and
triggering an automatic re-scan.
skip_disabled True/False are cached separately.
This commit is unstacked from #58984; it carries only the skill
discovery cache change.
(cherry picked from commit cd65673a8f)
Completes the session-binding class on the gateway surface (#55578),
matching the TUI rules:
1. Fail-closed pinning: switch_session() re-opens ended sessions, so
pinning a completion to a spawning session that has since ENDED
(user /new, closed rotation) would resurrect a conversation the user
explicitly ended and inject into it. The injection path now checks
the pinned row's ended_at first and drops the injection with a
WARNING when the spawning session is dead or unknown - the result
stays in the delegation records.
2. /new ends the old conversation's delegations: _handle_reset_command
calls interrupt_for_session() with the expiring durable session id
(matching the parent_session_id pin stamped at dispatch) plus the
routing key as fallback, so a reset can't leave dangling subagents
whose completions have no live owner.
interrupt_for_session() gains the parent_session_id selector because a
gateway chat's session_key (the platform conversation key) survives a
reset while the session id rotates - key-based matching alone could
never sever a gateway conversation's delegations.
Background delegate_task completions only carried session_key. When multiple
active sessions shared a routing peer, get_or_create_session could recover the
latest ended_at IS NULL row and inject the subagent result into the wrong
session.
Capture parent_agent.session_id at dispatch time, include it on async-delegation
completion events, and pin gateway routing via switch_session when the
synthetic completion message is handled.
Fixes#57498
Extends the salvaged session_key filter with the same fail-closed,
compression-chain-aware ownership gate the poller uses (#55578):
- drain_notifications() accepts an owns_event callback; when provided,
an async-delegation event is consumed ONLY on positive proof of
ownership, and a broken callback re-queues (never leaks). Bare key
equality remains for single-session callers (CLI); no filter remains
legacy behavior.
- The TUI post-turn drain passes _session_owns_notification_event, so
it can't adopt another session's (or an orphan's) delegation payload,
while a post-compression session still claims its own pre-compression
dispatches - the gap bare key equality left open.
The completion event already carries the dispatching session's session_key
(captured at dispatch time in delegate_tool.py:2798), but the delivery
router ignored it — results landed in whatever session was active at
completion time instead of the session that dispatched the subagent.
Changes:
- drain_notifications() in process_registry.py: optional session_key
filter. Non-matching async_delegation events are re-queued instead of
consumed, so they remain available for the correct session's drain.
- cli.py process_loop: passes active session_key to drain_notifications()
- tui_gateway/server.py post-turn drain: passes session_key from the
TUI session dict
- gateway/run.py _build_process_event_source: logs warning when routing
metadata is unresolvable (previously silent drop)
- Regression tests verifying session-scoped drain filtering
Fixes#58684
Two invariants layered on the origin-routing commit (#55578):
1. Fail closed on orphaned async-delegation payloads. The poller's
belongs-elsewhere check handles events owned by another LIVE session,
but an event whose owner is gone previously fell through and was
adopted by whichever poller saw it - injecting one chat's delegation
output into another chat. Delegation completions are now injected
only into a session that PROVABLY owns them (origin UI id, or
session-key/lineage match via the compression chain); unowned
payloads are dropped from injection with a WARNING (the subagent's
output is already persisted in the delegation records, so nothing is
lost). The shutdown drain applies the same rule. Non-delegation
events keep the historical adopt-orphans behavior.
2. A session's in-flight async delegations end with the session.
_finalize_session now calls interrupt_for_session(): delegations
commissioned by the closing UI session are interrupted always;
key-matched delegations only when the TUI owns the session lifecycle,
so closing a viewer tab on a live gateway session never kills the
gateway's own background work.
Carry the live TUI session id with async delegation completion events and prefer the commissioning UI session when desktop pollers share the completion queue. Resolve compressed session keys to their continuation before treating events as orphaned, and capture the live parent agent session id for TUI/ACP dispatch.
Three fixes for the silent post-restart ticker stall:
1. _jobs_lock() bounds its cross-process flock: LOCK_NB polled against a
30s deadline instead of an unbounded LOCK_EX taken while holding the
process-wide RLock. On timeout it logs at ERROR and degrades to
in-process-only locking (the existing fallback path), so a sibling
process wedged while holding .jobs.lock can no longer freeze every
cron function - including the ticker's get_due_jobs() and thus the
heartbeat - forever with zero logging.
2. fire_claim/run_claim freshness checks are bounded on both sides
(0 <= age < ttl): a claim stamped in the future (clock/TZ skew across
a restart) was previously fresh forever, making the job permanently
unfireable and every manual run report 'already being fired'.
3. _execute_job_now distinguishes paused/disabled/missing jobs from a
genuinely held claim instead of mislabeling them all as 'already
being fired'.
safe_load() raises ComposerError on multi-document streams (k8s manifests)
and ConstructorError on application-defined tags (CloudFormation !Sub,
Ansible !vault) — both valid YAML syntax. Now that the linter's verdict is
a fail-closed write gate, those false positives would refuse legitimate
writes outright. Switch to yaml.parse() (scanner+parser only), which still
catches real syntax failures.
write_file() previously called _atomic_write() first and only ran the
JSON/YAML/TOML/Python syntax check afterward as an informational lint
delta -- a parse failure never set the top-level `error` key, so a
corrupt structured-data write still landed on disk (and file_tools.py's
files_modified gating, which keys off `error`, silently reported it as
a successful modification).
Move the in-process syntax check for JSON/YAML/TOML ahead of
_atomic_write() and refuse the write outright on a parse failure: no
temp file, no rename, nothing touches disk, and the result carries a
top-level `error` so callers correctly see it as unmodified.
Deliberately scoped to _FAIL_CLOSED_INPROC_EXTS (JSON/YAML/TOML), not
all of LINTERS_INPROC -- .py is excluded because this codebase's own
test fixtures (TestPatchReplacePostWriteVerification et al.) write
arbitrary non-Python text through *.py paths purely to exercise
write-mechanics; a hard block there broke 3 previously-passing tests
during development. Python keeps its pre-existing non-blocking
lint-delta report.
Adds tests/tools/test_write_file_syntax_gate.py: invalid JSON/YAML/YML/
TOML refused with nothing written (new file) and nothing modified
(existing file); valid JSON/YAML still written byte-for-byte; a
non-linted extension with garbage content is unaffected; invalid Python
is confirmed NOT hard-refused (still just reported).
signal.SIGKILL / os.killpg don't exist on Windows. The watchdog is only
spawned on POSIX (wrap site gates on os.name), but guard via getattr with
a plain terminate/kill fallback so an accidental Windows import can't
AttributeError.
Two fixes on top of the salvaged parent-death watchdog:
- Apply the watchdog wrap AFTER the OSV malware preflight so the check
inspects the real npx/uvx package instead of the python wrapper
(the wrap previously made the preflight a silent no-op for every
stdio server).
- The real server runs in its own process group under the watchdog, so
the graceful-shutdown killpg no longer reached it; the watchdog now
forwards SIGTERM/SIGINT to the child's group, keeping wedged servers
killable on clean shutdown.
A stdio MCP server (e.g. `npx -y mcp-remote <url>`) is spawned as a direct
child of the Hermes process. Existing teardown (MCPServerTask.shutdown() /
_kill_orphaned_mcp_children()) reaps it correctly on a clean exit, but a
kill -9 / crash / force-quit of the Hermes process skips that path entirely
-- the child (and its own descendants, e.g. mcp-remote's spawned node
process) is orphaned and keeps running. Repeated ungraceful restarts pile up
N orphaned processes racing to hold the same upstream SSE session, producing
errors like 'Invalid request parameters' on legitimate reconnects.
macOS/Linux have no portable equivalent of prctl(PR_SET_PDEATHSIG) at the
Python subprocess level, so this adds a thin supervisor
(tools/mcp_stdio_watchdog.py) that:
- execs the real command as its own child in its own process group
- passes stdin/stdout/stderr through untouched (MCP stdio protocol
talks directly over those streams)
- polls the original spawning PID with the same orphan-detection
algorithm already proven in tui_gateway/slash_worker.py (ppid
comparison + psutil creation-time guard against PID reuse)
- SIGTERM-then-SIGKILL's the child's process group the moment the
original parent is gone
Wired into _run_stdio via a new _wrap_command_with_watchdog() helper,
POSIX-only (matches the existing killpg-based cleanup's platform scope),
fails open (any error resolving pid/create-time falls back to the
unwrapped command) so this can never be the reason a working MCP server
stops starting.
Verified: reproduced the exact orphan scenario standalone (fake parent
process spawns watchdog + fake long-running MCP child, kill -9 the fake
parent, confirm the watchdog reaps the child within its poll window with
zero leaked processes). Updated test_mcp_tool_issue_948.py's resolved-path
assertion to check the watchdog-wrapped command instead of the raw
resolved binary. Full test_mcp_tool.py + test_mcp_stability.py +
test_mcp_tool_issue_948.py suite: 232 passed. Full -k mcp sweep across the
whole test tree: 1003 passed, 2 skipped, 0 failed.
Sibling sites of the same bug class as the salvaged stdio fix:
- SSE, streamable-HTTP (new + deprecated API) initialize() calls are now
bounded by the same connect_timeout, so an endpoint that accepts the
connection but never answers the handshake cannot park the run() task
forever.
- start() now cancels its ensure_future'd run() task when the caller's
connect timeout cancels start() itself — the orphaned-task leak was
the root mechanism behind #59349, and this closes the class for any
future pre-ready hang.
A stdio MCP server that never completes `initialize` (e.g. emits a
non-JSON-RPC frame and then blocks on stdin) leaks a child process plus its
stdio pipes/pidfd on every discovery-retry cycle — unbounded, until the
gateway hits EMFILE and every new open()/spawn fails (#59349).
Root cause (confirmed by instrumenting the live repro, and different from the
issue's own hypothesis): the spawned child IS captured in `new_pids`, so the
report's "new_pids empty at finally" guess is not it. The real cause is that
`session.initialize()` hangs forever on the garbage stream. `connect_timeout`
only bounds the caller's `.result()` wait on the foreground thread — it does
NOT cancel the `_run_stdio` coroutine on the background MCP loop. So the
coroutine is stuck at `await session.initialize()` permanently, its cleanup
`finally` never runs, the child is never reaped, and it stays invisible to the
orphan-reaper (whose `_orphan_stdio_pids` set never gets populated).
Fix: wrap `session.initialize()` in `asyncio.wait_for(..., connect_timeout)`
so a stalled handshake fails instead of hanging. The TimeoutError unwinds
through the SDK context managers (closing the child's stdin -> EOF -> exit)
and lets the existing `finally` reap any straggler. Cross-platform — no
signals/pgid/proc.
Scope: stdio only. The HTTP path has the same `await session.initialize()`
shape but spawns no subprocess (so it can't cause this leak) and already has
httpx transport timeouts.
Verified: the reporter's repro goes from unbounded growth to draining to zero;
added a hermetic regression test (fake transport whose `initialize()` hangs,
asserts the connect is bounded by connect_timeout) that fails on the pre-fix
code and passes on the fix; 566 existing MCP tests pass; ruff clean.
Repro confirmed on macOS (pipe FDs); the Linux-specific pidfd growth in the
report should be equivalent — the reporter offered to validate on Linux.
Closes#59349
Merge the two cherry-picked reap call sites into one unscoped sweep at
the top of _run_stdio (the unscoped sweep is a superset of the
per-server one), and run it via asyncio.to_thread so the 2s
SIGTERM->SIGKILL escalation cannot stall the shared MCP event loop.
When an MCP stdio subprocess fails to connect (token expiry, port
contention, timeout), the run() reconnect loop retries with backoff.
Each retry calls _run_stdio() which spawns a new process pair, but the
previous failed pair was only detected as orphaned (added to
_orphan_stdio_pids) — never actually killed. This caused rapid zombie
accumulation: 5 failed attempts × 2 procs each = 10 orphans competing
for the same port.
Add a _kill_orphaned_mcp_children() call at the top of _run_stdio(),
before the _snapshot_child_pids() baseline, so any orphans from prior
failed attempts are reaped before a new subprocess is spawned.
Fixes#57355
_run_approval_gate's gateway branch only queued via submit_pending, so
plugin-escalated approvals never sent the interactive embed+buttons on
Discord/Telegram/Slack (#59413) - the user was never notified and the
action stayed silently blocked. Mirror check_dangerous_command's path:
when a session notify callback is registered, run the blocking
_await_gateway_decision round-trip (redacted payload, once/session/
always persistence, deny/timeout produce definitive BLOCKED outcomes);
fall back to submit_pending only when no callback exists.
Fixes#59413.
* feat(trace): upload sessions to HF Agent Trace Viewer
Salvage trace upload as a smaller CLI-first feature: deterministic Claude Code JSONL export, fail-closed redaction, lazy Hugging Face dependency, and no gateway slash-command wiring.
* chore(trace): drop external porting references from docstrings
Describe the trace-upload design in Hermes' own terms.
* feat(sessions): fold trace upload into 'sessions export --format trace'
Integrates the HF Agent Trace Viewer exporter (PR #36145) onto the
unified export surface instead of a separate 'hermes trace' subcommand:
- --format trace: Claude Code JSONL to stdout/file, or one
<id>.trace.jsonl per session for filtered bulk export; defaults to
the most recent session when no --session-id/filters given.
- --upload pushes to the user's private HF traces dataset (--public to
opt out of private); reads HF_TOKEN with guided setup when missing.
- traces are secret-redacted by default (force mode); --no-redact opts
out after review; redaction failure blocks export (fail closed).
- hermes_cli/trace.py + subcommands/trace.py removed; agent/trace_upload.py
is the single engine. Docs EN + zh-Hans; 4 new CLI tests.
BaseEnvironment writes shell snapshots and cwd metadata through the process
umask. With a common 022 umask, snapshot files containing exported environment
state landed at mode 0644 even though they can include env-carried credentials
from the parent process.
Set umask 077 only around Hermes metadata writes: the initial snapshot
bootstrap and the post-command snapshot/cwd refresh. User commands still run
under the caller's original umask, while Hermes-owned snapshot and cwd files
are created owner-only.
This intentionally does not copy the source PR's global orphan sweep; deleting
all matching /tmp snapshot files could interfere with concurrent Hermes
processes. The security-critical local disclosure fix is the file mode clamp.
This is salvageable because the source report still identifies a concrete
credential-disclosure path, but the safe subset is smaller than the original
proposal: clamp only the Hermes-owned snapshot writes and leave process-wide
cleanup, user command umask, and concurrent sessions alone.
Salvages source PR: https://github.com/NousResearch/hermes-agent/pull/20056
Related issue: https://github.com/NousResearch/hermes-agent/issues/48441
Co-authored-by: Andrew Homeyer <andrew@hndl.app>
Same bug class as skills_tool: module-level SKILLS_DIR pinned at import
under the launch HERMES_HOME makes skill_manage() write/edit against the
wrong profile in long-lived multi-profile runtimes. Apply the same
_skills_dir() call-time resolution (honoring explicit test patches of
SKILLS_DIR) to _containing_skills_root, _resolve_skill_dir,
_find_skill_in_other_profiles, and create-result path reporting.
Refs #40677
Port from anomalyco/opencode#31877: JSON Schema type arrays like
["number","string"] (common in MCP tool schemas) were collapsed to the
first non-null type, silently dropping every other branch. Several
tool-call backends reject the array form outright — llama.cpp's grammar
generator and Gemini via OpenAI-compatible transports (e.g. GitHub
Copilot proxying to Gemini) 400 on it.
_sanitize_node now mirrors @ai-sdk/google: a single non-null type stays
type:X (+nullable if null was present), multiple non-null types become
an anyOf of single-type schemas so no branch is lost, and an all-null
array becomes type:null. Single-null collapse is unchanged.
Verified nested (object props, array items) survive the full sanitize
pipeline — combinator stripping is top-level-only and nullable-union
collapse only fires on single-survivor unions, so multi-type anyOf is
left intact.
Port from openclaw/openclaw#91950: normalize LLM-generated URLs like 'https:// docs.example' before web tool safety checks while preserving path and query encoding semantics.
Port from nearai/ironclaw#4547: treat a JSON null memory target as omitted so strict providers that fill optional fields with null use the documented default target instead of failing validation.
A user-approved terminal/execute_code command could be SIGINT-killed
(exit 130 + "[Command interrupted]") by a stale interrupt bit that landed
on the execution thread during the blocking approval-wait, while the
result still carried the "...approved by the user." note. The terminal
tool runs sequentially inline on the execution thread, and nothing
cleared or re-checked the bit between approval-grant and env.execute.
Clear the current thread's interrupt bit once before an approved command
spawns its child (terminal foreground; execute_code local + remote), and
enrich the note to "...approved by the user, then interrupted." on a
genuine post-start interrupt instead of implying success. A genuine
interrupt arriving after execution starts (or during a retry backoff)
still SIGINTs the command; non-approved commands keep current behavior.
Adds regression tests covering stale-bit-clears, genuine-interrupt-still-
kills, the retry-backoff window, natural-exit-130 (not mislabeled), and
execute_code local + remote.
When a bundled web provider (firecrawl, tavily, exa, ...) is listed in
plugins.disabled, its provider never registers and the web_search/
web_extract dispatchers emitted the misleading "No web extract provider
configured. Set web.extract_backend to ..." — even though the backend was
configured correctly. The real fix is to re-enable the plugin.
- web_tools.py + web_search_registry.py: when the configured backend names
a disabled bundled web plugin, both dispatchers now point the user at the
actual cause (re-enable the plugin) instead of a wrong config hint.
- plugins_cmd.py cmd_enable: enabling by canonical key now also clears the
manifest-name alias (web-firecrawl) from plugins.disabled, so the
suggested command actually re-enables the plugin ('explicit disable wins'
matches on the name too).
- plugins_cmd.py cmd_toggle / _run_composite_ui / _run_composite_fallback:
the interactive 'hermes plugins' menu now persists the canonical key
(web/firecrawl), never the bare manifest name — the drift that put the
offending entry in plugins.disabled in the first place.
Follow-up to #59518 (which fixed web credential resolution, a different
cause). Fixes the disabled-plugin symptom reported after that PR.
The Firecrawl provider used os.getenv() to read FIRECRAWL_API_KEY and
FIRECRAWL_API_URL, which only checks the process environment. When
values are supplied through Hermes's ~/.hermes/.env config mechanism
(via hermes_cli.config.get_env_value), they are not guaranteed to be
present in os.environ for every gateway/tool execution path.
Switch to get_env_value() which checks both os.environ and the .env
file, matching the pattern used by other providers (nous_subscription,
setup, discord adapter).
Fixes#40190
_sync_back_once defers a SIGINT that lands mid-sync, then re-delivers it once the
sync completes so the user's Ctrl+C isn't lost. It did so with
os.kill(os.getpid(), signal.SIGINT). That is not graceful on Windows: os.kill
only treats CTRL_C_EVENT(0)/CTRL_BREAK_EVENT(1) as console events; any other
value (SIGINT == 2) routes to TerminateProcess(sig), so a Ctrl+C during a
remote-backend (ssh/daytona/modal) sync-back hard-kills the whole CLI session
(exit code 2) on Windows instead of raising KeyboardInterrupt.
Use signal.raise_signal(signal.SIGINT) (3.8+), which invokes the restored
handler through C raise() on every platform. Verified on Windows: raise_signal
runs the handler (graceful) while os.kill(getpid, SIGINT) TerminateProcess-es
the process. Adds a cross-platform regression test that runs on Windows too (it
stubs the locked sync body, so unlike test_file_sync_back.py it needs no fcntl).
The Skills Hub 'Browse Hub' landing page and index-backed search render
empty on fresh deployments (e.g. Fly.io VPS agents) with no stale cache.
Root cause: the centralized index at /docs/api/skills-index.json is a
large body (~34MB, tens of MB compressed) served with Content-Encoding:
br. httpx's streaming Brotli decoder — backed by brotlicffi 1.2.0.1,
which is pinned so aiohttp can decode Discord attachments — trips over
its own output_buffer_limit on a payload this size and raises:
DecodingError("brotli: decoder process called with data when
'can_accept_more_data()' is False")
_load_hermes_index() catches that (DecodingError is an httpx.HTTPError
subclass) and silently falls back to the on-disk cache. On a fresh box
that cache never existed, so HermesIndexSource.is_available is False,
the index contributes 0 skills, and the hub landing page — which is
built solely from an empty-query index search — is blank. Existing
installs only appear to work because they serve a (possibly weeks-)stale
cached index instead.
Fix: request 'gzip, deflate' on the index fetch so httpx never
negotiates the broken Brotli path, and retry once with 'identity' if a
DecodingError still occurs (defends against a proxy that ignores the
header). Falls through to the stale cache only when both attempts fail.
Verified on a live staging VPS agent: index_available flips False->True
and the featured landing list repopulates from 0 to 12.
Also un-freezes already-deployed images: skills added after an image was
built (e.g. the 'unbroker' optional skill) become reachable again via
the index, which is the whole point of the centralized catalog.
register_mcp_servers now nudges cached entries whose session is None
via _signal_reconnect, so a new agent session recovers a parked server
immediately instead of waiting up to _PARKED_RETRY_INTERVAL for the
next self-probe (#50170). Gate-check idea credit: @izumi0uu (#50184),
@LeonSGP43 (#37772), @Tranquil-Flow (#37899).
_wait_for_server_session_ready used a time.monotonic deadline; the
circuit-breaker tests freeze monotonic, turning the loop into an
infinite spin (300s SIGKILL in CI-parity runs). Bound by iteration
count instead.
Four independent pre-request stalls sat on the critical path between
prompt submission and the first streamed token, measured with cProfile
against a live process:
1. Discord capability detection (~2.0s, worst 5s): get_tool_definitions
-> _get_dynamic_schema made a BLOCKING https call to discord.com
inside AIAgent.__init__ for any user with DISCORD_BOT_TOKEN set, on
every platform, every cold process. Now non-blocking: memory cache ->
24h disk cache -> permissive default + one background detection that
seeds the disk cache for the next process. The permissive default is
pinned per-process so tool schemas never flip mid-conversation
(prompt-cache safety); it mirrors the existing detection-failure
fallback (all actions exposed, 403s enriched at call time).
2. Ollama /api/show probe (~0.3s): get_model_context_length step 5e
POSTed to <base_url>/api/show for KNOWN providers (openrouter etc.),
got a 404, and never cached the miss - so every fresh process paid a
full HTTP round-trip. Known non-Ollama providers now skip the probe;
local/custom/unknown endpoints keep the exact previous behavior.
3. env_probe subprocess sweep (~0.5s): the Python-toolchain probe ran
4-8 subprocess calls inside the FIRST system prompt build. Now warmed
off-thread during agent init; the prompt build hits the cache (same
lock, so a mid-flight warm just joins instead of recomputing).
4. tools.mcp_tool import (~0.4s): the between-turns MCP refresh in
build_turn_context imported the whole mcp package even with zero MCP
servers configured. MCP tools can only exist if tools.mcp_tool was
already imported (discovery/reload paths), so gate the import on
sys.modules membership - no behavior change for MCP users.
CLI additionally pre-imports run_agent + openai off-thread during the
idle banner window (same pattern as the /model picker prewarm), hiding
the remaining ~1.5s of module imports while the user types. Fixes 1-4
apply to every interaction layer (CLI, gateway, TUI, desktop, cron).
Measured cold first turn (submit -> request dispatched, openrouter,
discord token set): 4.3s before -> 0.9s after CLI prewarm (~80%); the
agent-side non-import cost drops 2.9s -> 0.36s (init) + 0.27s (turn
prologue).
Some MCP servers (e.g. Spring Boot apps with a React SPA) serve their
frontend on any unmatched GET route. The MCP endpoint works perfectly
via POST (JSON-RPC), but a GET to /mcp falls through to the SPA
controller and returns text/html. Hermes's preflight content-type probe
sees HTML instead of application/json or text/event-stream and refuses
to connect.
This adds a per-server config option that
bypasses the content-type probe, letting the SDK connect directly via
POST where it works fine.
```yaml
mcp_servers:
stirling-pdf:
url: http://localhost:8090/mcp
headers:
X-API-KEY: <key>
skip_preflight: true
```
Related: #52460 (OAuth redirect preflight), #51600 (skip probe on mcp add),
#40366 (skip probe on reconnect — already merged).
Some MCP servers (e.g. DocuSeal) serve their web UI on HEAD/GET but
speak Streamable HTTP only via POST. The preflight probe now tries a
lightweight JSON-RPC `initialize` POST before rejecting endpoints
whose HEAD/GET returns a non-MCP content type (e.g. `text/html`).
If the POST returns `application/json` or `text/event-stream` with a
2xx status, the endpoint is accepted. Otherwise the original rejection
behaviour is preserved.
Adds 5 new test cases covering the POST probe path:
- POST rescues HTML HEAD with JSON response
- POST rescues HTML HEAD with event-stream response
- POST still rejects when it also returns HTML
- POST still rejects on non-2xx status
- POST not attempted when HEAD already returns valid MCP content type
