Builds the escalation layer on top of HexLab98's heal-log windowing
(salvaged from PR #96916):
- Per-session heal counters (heal events + messages healed) tracked by the
repair path in agent_runtime_helpers.py, session totals preserved across
10-minute log windows.
- Threshold escalation: after N heals in a session window (default 3,
configurable via agent.sanitizer_heal_escalation_threshold in
config.yaml, 0 = off) log ONE ERROR carrying session id + heal pattern
(events/messages/window/threshold), then stay quiet.
- ONE-TIME out-of-band user notice queued at the threshold and delivered by
the conversation loop through _emit_warning (status callback -> gateway
status message / CLI print). Never injected into conversation context or
the wire copy: prompt caching, role alternation, and durable history are
untouched. Never re-arms on a new window; scoped per session.
- Counters visible in diagnostics: get_sanitizer_heal_stats() rendered in
the /debug share // hermes debug report, and the config key surfaced in
hermes dump overrides. errors.log carries the ERROR line for `hermes logs
errors`.
Fill empty non-final user/assistant turns on the wire copy during send-time projection so the sanitizer does not re-heal the same poisoned row every call. When a caller still hits the owner, log WARNING then one ERROR per session window instead of flooding errors.log.
Second leg of the masoria 20-minute 'Summarizing' stall (Aug 31 2026
bundle): after the Codex timeout, compression fell back to OpenRouter,
which defaulted the omitted output cap to the model's full 65,536-token
window and rejected with '402 ... can only afford 7117' — on an account
whose balance easily covered a summary. Three fallbacks, three 402s,
zero summaries.
- _create_with_progress: when a 402 names an affordable budget, retry
ONCE with that cap (minus 64-token margin, 512-token floor); plain
exhaustion 402s and within-budget 402s re-raise unchanged. Single
funnel covers primary, fallback, and retry call sites.
- extends the #41055 OpenRouter max_tokens preservation onto the current
_build_call_kwargs gate (explicit caps survive; None still omitted).
Live A/B against a mock OpenRouter enforcing a 7117-token budget with
the real SDK + real adapter: main fails with the exact bundle 402;
fixed branch retries clamped and returns the summary.
_build_call_kwargs strips max_tokens for non-Anthropic providers to
avoid wire-format issues (Copilot, ZAI, GPT-5). However, OpenRouter
free/limited-credit tiers need max_tokens because the model's full
output window exceeds the credit budget, causing HTTP 402.
Without max_tokens, the 402 triggers fallback to a text-only model
which then fails with 'unknown variant image_url, expected text'.
Include max_tokens when provider is 'openrouter' or base_url contains
'openrouter.ai'.
Fixes#41035
Socket shutdown() releases readers blocked on a real transport, but the
owner can be blocked inside the SDK event stream (or a transportless
double). Close the attempt-owned stream from the Timer — the same
attempt-scoped wake the hard-cancel branch uses — so the owner unwinds
and performs the real FD release in its finally. Fixes the
test_codex_timeout_and_explicit_cancel_have_one_linearized_outcome red.
Adapt the #99660 blocked-before-first-event watchdog test to the FD-ownership
contract: the stranger-thread Timer marks the timeout but never close()s;
the owning thread surfaces the TimeoutError and releases the FDs on unwind.
_CodexCompletionsAdapter.create arms a daemon threading.Timer that calls
client.close() when the aux Responses stream exceeds its timeout. On a
stalled stream -- the failure the timeout exists for -- the Timer is the
only thing that fires, so the close runs on a thread that does not own
the in-flight httpx connection.
That is the FD-ownership violation the repo already fixed twice on the
main transport (#29507, #67142, #70773): close() releases the raw TLS fd
while the owner's OpenSSL BIO still caches that integer, the kernel
recycles it into the next open() in the process -- a SessionDB or
kanban.db handle -- and the owner's unwinding TLS flush writes an
application-data record into that database file.
agent/auxiliary_client.py had no thread-ownership machinery at all: the
guarded twins (_retire_shared_openai_client, _abort_request_openai_client)
live in run_agent.py and are unreachable from this adapter, which holds no
AIAgent reference.
Dispatch on ownership the way chat_completion_helpers already does: from a
stranger thread only force_close_tcp_sockets() (shutdown(SHUT_RDWR), which
is FD-safe from any thread), and let the owning thread release the FDs when
it unwinds. The owner-thread caller (_check_cancelled) keeps closing
directly. Cache eviction (#23432) is unchanged.
The #95294/#91277 fleet contract requires the pending-restart check to
execute on every already-up-to-date pass; the verification exit(1) now
fires only after the catchup runs, so a vulnerable runtime demotes the
outcome to partial without stranding the fleet on stale code.
Verification-features-must-not-false-positive-on-rollout rule: only a
POSITIVE vulnerable SQLite probe demotes success to partial. A dev
checkout without a venv (or a failed probe subprocess) keeps the
success banner and the fleet-restart catchup path — the CI-red
test_update_fleet_restart_pending already-up-to-date trio pinned this.
The previous form was `mock_install.assert_not_called(), ("msg")` — a bare
tuple expression whose parenthetical never surfaces as a failure message.
Switch to `assert mock_install.call_count == 0, "msg"` so the diagnostic
actually appears when the guard regresses (review feedback on #93163).
On Windows, hermes update launched from the install's own venv can never
complete the managed-runtime repair: Windows keeps the image of the
updater's venv\Scripts\python.exe (and the waiting hermes.exe launcher
ancestor) mapped until exit, so the park rename in _cut_over_candidate
always fails with ERROR_ACCESS_DENIED. The pre-flight holder scan
deliberately excludes the calling process and its ancestors, so the guard
passes and the repair burns its retries on a structurally unwinnable
rename - forever, since the failure is non-fatal and the 'next update
will retry' message is misleading for this case.
Detect the self-lock (sys.executable or a launcher ancestor inside the
live venv) before provisioning and defer with actionable guidance instead
of walking into the doomed cutover. The deferral happens pre-provisioning,
so the incomplete generation-* leftovers the reporter observed are no
longer produced. No-op off Windows: POSIX renames work while the updater
maps the tree. Mirrors the existing _defer_update_for_self_lock pattern.
Regression tests prove the fix bites: neutralized guard -> repair proceeds
to provisioning (red); restored guard -> deferred before provisioning
(green). Verified on Windows 11 against a real venv.
Closes#93032
- Thread event.message_id (raw inbound id) as TurnContext.inbound_message_id
instead of reusing event_message_id, which is the reply/thread anchor and
can be the replied-to message on Slack/Mattermost/Buzz or None for
Telegram topics.
- Add platform_message_id to the schema-foreign strip sets in
ChatCompletionsTransport.convert_messages and the summary path so strict
providers never see the persistence-only key.
Follow-ups on top of the #94595 cherry-pick, implementing the maintainer
review's two blockers:
Blocker 1 (turn-admission chokepoint): _run_prompt_submit itself now runs
the ownership admission, so synthesized turns that never pass through the
prompt.submit RPC handler (crash auto-continue from cold session.resume,
wake-ups) are fenced too. Auto-continue additionally checks ownership
BEFORE emitting message.start and leaves the marker in place, closing the
#94778 shape where backend B resumed a session backend A was actively
running and auto-continued A's fresh interrupted-turn marker into a
duplicate concurrent turn.
Blocker 2 (fail-closed registry semantics): try_acquire_active_session no
longer converts an unreadable/corrupt registry into an untracked go-ahead.
Ownership uncertainty is a distinct typed refusal —
SESSION_COORDINATION_UNAVAILABLE — because "could not prove ownership" must
never be collapsed into "no owner exists". The TUI gateway claim helper
fails closed on claim exceptions for every surface, not just desktop.
Also: empty session ids short-circuit to a no-op lease (nothing to fence,
and the strict registry schema rejects empty ids), and the existing
fail-open tests were updated to assert the new fail-closed contract.
The unit tests share one interpreter, so they cannot exercise the failure the
fence exists to prevent: two SEPARATE gateway processes, each holding its own
snapshot of a conversation, both writing to it. That is how the defect was found
and it is the only way to show it is closed.
This drives two real `python -m tui_gateway.entry` processes over stdio and
checks the whole sequence, including the parts that are easy to get wrong:
session.create claims nothing an idle composer must not hold a session
the lease keys on the STORED id a lease keyed on the runtime handle would
fence nothing, since two processes
resuming one conversation have different
runtime ids by construction
B may still RESUME reading is never fenced; only writing is
B's submit -> SESSION_NOT_OWNED typed, and the registry is unchanged
A killed, B retries -> accepted a dead owner is pruned, not permanent
No provider is needed. The fence is checked before the agent is built, so a
submit that later fails for want of a model still proves who owns the session --
which keeps the probe free of credentials and of inference cost.
Against the parent commit it stops at the second check with an empty registry,
which is the defect stated exactly: with no cap configured, nothing was recorded
and therefore nothing could be refused.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Cherry-picked from PR #94595 (author: Futahua) onto current main, with the
maintainer-review revision points folded in during the rebase:
- the lease engages UNCONDITIONALLY: try_acquire_active_session no longer
returns a disabled no-op lease when max_concurrent_sessions is unset;
the concurrency cap stays an orthogonal, optional policy checked second
- ownership uncertainty fails CLOSED (SESSION_COORDINATION_UNAVAILABLE)
instead of degrading to an untracked go-ahead: a corrupt/unreadable
registry must not be collapsed into 'no owner exists' (review blocker 2)
- the ownership admission sits at the _run_prompt_submit chokepoint that
EVERY fresh turn source crosses, and crash auto-continue acquires (or
bails) BEFORE emitting message.start — closing the #94778 bypass where
backend B's auto-continue ran a duplicate turn while backend A was live
(review blocker 1)
- the TUI gateway claim helper fails closed on claim exceptions for every
surface, not just desktop
- CLI and messaging-gateway call sites pass live_session_id metadata so
the (pid, live id) re-entrancy identity protects them from self-fencing
on a leaked lease
Co-authored-by: teknium1 <teknium1@users.noreply.github.com>
Salvaged from #98741, composed on top of the merged #98424 preflight
fail-closed boundary. A host-ceiling compression timeout is now a typed,
thread-safe outcome consumed by every automatic caller:
- conversation_compression.py: threading.local + per-agent lock timeout
state (mark/reset/read helpers) upgrading #98424's simple attribute
where overlapping automatic/manual compression entrypoints matter;
the _last_compression_timed_out attribute stays as compat mirror.
- conversation_loop.py: the mid-turn pre-API pass and the provider
overflow (413/400 context_length_exceeded) recovery path end the turn
with the typed compression_exhausted recovery contract instead of
re-sending the unchanged oversized request and re-entering compression
in the same turn.
- run_agent.py/turn_context.py: forwarder resets the typed state per
attempt; the #98424 turn-start check reads it through the typed helper.
Tests: thread-safety/atomicity of the state helpers, overflow-recovery
non-re-entry, and typed terminal result.
When the default/global state.db is corrupt at gateway startup,
SessionStore degrades (_db=None) and record_gateway_session_peer never
self-heals. Under multiplexed profile routes the AIAgent's lazy
_ensure_db_session was then the ONLY durable write for the session, and
it created the row identity-less (session_key/chat_id/chat_type/
thread_id/user_id/origin_json all NULL) — unrecoverable by
find_latest_gateway_session_for_peer, so Telegram chats forgot prior
turns.
Persist the routing identity the agent already carries into
create_session; plain CLI sessions keep the old identity-less shape.
Salvaged leg 1 of #88804; the transcript-recovery leg is covered by the
scope-aware session DB resolution already on main.
Follow-ups on top of the salvaged commits from PR #87111 (@HexLab98) and
PR #87265 (@JoaoMarcos44):
- keep main's #92991 stall watchdog (150s progress-based) as the single
steady-state liveness probe instead of adding a second overlapping one
- orphaned-client aclose() cleanup uses the wall-clock thread deadline and
is tracked in _background_tasks so a wedged close can neither hang nor
leak one task per reconnect attempt (from #87265's review findings)
- merge #87265's no-keepalive getUpdates pool (max_keepalive_connections=0)
with #87111's TCP-keepalive socket options on all transports
- add tests/gateway/test_telegram_closewait_windows_live.py: live probes
against a real half-closing HTTP server, skipif non-win32, wired into
the on-demand windows-venv-e2e lane (wine2e/**)
Prevent the dedicated getUpdates pool from reusing server-closed connections and replace a polling HTTP client left open after a timed-out CLOSE-WAIT drain. Keep the general Bot API pool reusable so concurrent sends and edits are unaffected. Add regression coverage for both transport limits and stale-client replacement. Fixes#87057
After updater.stop() times out, HTTPXRequest.initialize() is a no-op unless
the client is already closed, so start_polling reused the wedged socket and
the gateway stayed alive but deaf. Rebuild the polling client after a hung
drain, watch getUpdates I/O independently of get_me(), and enable TCP
keepalive on the fallback transport.
The native generateContent adapter never runs uncapped: when
model.max_tokens is unset it sends maxOutputTokens=65,535
(GEMINI_DEFAULT_MAX_OUTPUT_TOKENS) because Gemini treats an omitted cap
as a low internal default. The context compressor's trigger is
pct×(window − max_tokens), and constructing it with max_tokens=None
reserved 0 — so on a 128K Gemma window the trigger landed at 98,304
while the real safe input budget was 65,537, and the provider 400'd
before compaction fired.
Live repro (real imports, temp HERMES_HOME, native Gemini base_url,
window=131072, max_tokens unset):
before: compressor.max_tokens=None, threshold_tokens=98304,
wire maxOutputTokens=65535 → trigger ABOVE the safe budget
after: compressor.max_tokens=65535, threshold_tokens=64000 → below it
Scoped to the native Gemini wiring (provider names + native base_url via
is_native_gemini_base_url; the /openai compat endpoint is excluded). The
generic provider-default reservation gap remains tracked in #63839.
Reported by @Artemonim in #57275 (residual claim 4).
model.ollama_num_ctx is resolved AFTER the context compressor is
constructed, so a config that sets only ollama_num_ctx (without
model.context_length) ran every request at the smaller served num_ctx
while the compressor still targeted the probed GGUF window (e.g. 256K
Gemma metadata). The compaction trigger then sat several times above the
window the server actually serves and never fired — reproducing the
original #57275 'blows past the limit' symptom on current main.
Live repro (real imports, temp HERMES_HOME, config = {model:
{ollama_num_ctx: 65536}}, probed window 262144):
before: _ollama_num_ctx=65536, compressor.context_length=262144,
threshold_tokens=196608 (300% of the served window)
after: compressor.context_length=65536, threshold below the window
The clamp is one-directional (a num_ctx larger than the resolved window
never inflates the compressor) and reuses update_model() so every
threshold-derived budget recalibrates. Overlaps #60103 (silent-clamp
dead zone) — this is the init-order half.
Reported by @Artemonim in #57275 (residual claim 3).
Real windows-latest coverage for the #88136 salvage: probes drive the
actual _git_is_trampoline/_locate_real_git/_ensure_non_trampoline_git
helpers against the runner's genuine Git-for-Windows install plus a real
fork-bomb-guard trampoline stand-in. Wired into the on-demand
windows-venv-e2e lane (wine2e/** pushes only).
Review feedback on #88136 (monerostar): a profile-scoped `hermes update`
sets HERMES_HOME to <root>/profiles/<name>, but the Hermes-managed
PortableGit tree lives under the SHARED root (<root>/git/...). The locator
checked get_hermes_home() only, so a broken trampoline during a
profile-scoped update was not swapped and fell through to ZIP.
Extract _portable_git_candidates() (shared root first, profile home as
fallback) and add a regression test for the profile layout.
A Git-for-Windows trampoline launcher (bin\git.exe / cmd\git.exe shim,
~46KB) that fails to re-exec the real git-core binary refuses every git
call with a "BUG (fork bomb)" guard instead of running it (#87876).
Detect the trampoline up front via `git --version`, locate a real git
binary (Git for Windows or Hermes-managed PortableGit locations), and
rebuild the git command with it so fetch/pull/checkout keep working with
a real git instead of degrading to the ZIP fallback. When no real binary
can be found, leave the command untouched so the existing fetch-failure
handler still falls back to the ZIP path on Windows (#88046).
Google Gemini/Gemma overflow errors read 'Unable to submit request because
the input token count is 32825 but model only supports up to 32768'.
parse_context_limit_from_error had no pattern for the 'supports up to N'
phrasing, so overflow recovery kept the wrong window and burned its retry
attempts instead of recalibrating to the provider-reported limit.
Add the anchored pattern (limit follows 'supports up to'; the larger input
count before it is never captured) plus regression tests covering the exact
message and the get_context_length_from_provider_error recalibration path.
Reported by @Artemonim in #57275 (residual claim 5).
The MINIMUM_CONTEXT_LENGTH floor in _compute_threshold_tokens only
degraded to the 85% trigger when it met or exceeded the effective
window exactly (#14690). Near-minimum windows slipped through: at
context_length=65536 the threshold passed through at 64,000 — 97.7%
of the window, ~1.5K tokens of output room — so pre-API compaction
effectively could not fire.
Providers that silently truncate over-window prompts instead of
rejecting them (e.g. ollama's OpenAI-compatible /v1 endpoint) never
deliver the reactive context-overflow backstop either. Observed live
on a 65,536-token local model: the session rode into the window
ceiling and each length-continuation retry re-sent a window-filling
prompt (65,120 -> 65,273 prompt tokens, 263 output tokens of room)
until the turn died with "Response remained truncated after 4
continuation attempts" — every retry paying a full multi-minute
prefill.
Cap the floored threshold at _MIN_CTX_TRIGGER_RATIO (85%) of the
effective input budget whenever the floor is the binding term. An
explicit threshold_percent above 85% is user intent and stays
uncapped; windows where the floor lands at/below the cap are
unchanged.
Builds on webtecnica's escape-aware _split_key_path (#84152, cherry-picked
with authorship preserved; earliest fix in the family was RelaxJonh's #80253
greedy-match approach — both behaviors now ship together):
- _greedy_literal_match: when navigating an EXISTING mapping, prefer an
existing literal key equal to the dot-join of the next N path segments
(longest match wins). Dotted model IDs are the norm, so the common
unescaped command (config set providers.p.models.grok-4.6.supports_vision
true) now hits the real key across set/get/unset instead of creating a
phantom sibling. Plain dotted paths with no dotted-key collision split
exactly as before.
- _phantom_sibling + ValueError in _set_nested: refuse to CREATE a new
intermediate mapping that would shadow an existing dotted literal sibling
(Soju06's fail-loudly suggestion on #84064); set_config_value surfaces it
as a clean CLI error with the escaped spelling to use.
- utils.py::atomic_roundtrip_yaml_update (the second split site, #91607 —
/model + TUI persistence) now uses the same escape-aware split + greedy
literal matching.
- CFG-04 empty-segment guard now splits escape-aware so escaped keys are
not misclassified.
- Tests for every repro shape in the family: #84064 provider model keys,
#80006 Matrix room IDs, #91095 dotted models under custom_providers list
index (incl. escaped creation-when-absent), #91607 model_overrides via
atomic_roundtrip_yaml_update, #99124 dotted leaf keys; plus
backward-compat coverage. Also fixed the carrier's one stale assertion
(structured-value coercion landed on main after #84152 branched) and
removed its dead _MCP_SECRETS_CONFIG fixture flagged in review.
- Docs: 'Dots inside key names' section in website/docs/reference/cli-commands.md.
Fixes#84064, fixes#80006, fixes#91095, fixes#91607, fixes#99124
Follow-up to the salvaged #98571: forward the interrupt to the compute
host whenever the parent 'running' mirror is stale, but only for
sessions that actually have hosted activity — HostSupervisor.interrupt()
calls start(), so an unconditional forward would spawn a compute-host
child just to deliver an interrupt for an idle lazy session.
Adds a regression test asserting the idle-lazy-session no-spawn path.
Refs #92916
reconcileRegistryDrift only healed remote/cloud v1 routes. A v1 global
mode:'ssh' route (host, no url) written by Settings after the one-shot
migration had no registry identity: resolvedConnectionId returned null,
primary stayed 'local', and every launch re-homed the window onto a
fresh local backend. Because the heal skipped SSH entirely, the two
config files re-drifted after every update relaunch instead of
converging once.
Normalize the v1 SSH descriptor into a v2 kind:'ssh' entry (via the
same validated normalizeConnectionInput path the editor uses) and align
primary/lastUsed, with the same narrow-heal rules as remote: already-
registered targets and deliberate primary picks are left alone, and
unusable hosts never touch the registry.
Diagnosis credit: mgallmur-glitch (root cause) and jakewvincent
(re-drift after update relaunch) on #93888.
A no-mux tunnel is a single persistent `ssh -N -L` child. On main, ANY
death of that child after readiness immediately set tunnel.alive=false,
which poisons SshConnection.isAlive() forever; upstream lifecycle probes
then treat the whole SSH connection as dead, tear down the scope, and
SIGTERM a perfectly healthy backend (~10s after HERMES_BACKEND_READY in
the #96266 logs: '[ssh] connection closed (no-mux tunnels killed)' ->
'Ignoring stale Hermes backend exit (SIGTERM)' -> 90s port-announcement
timeout, with retry/repair looping the same failure).
Now a post-readiness child death is a tunnel FLAP: the child is
restarted with a bounded budget (5 attempts, 1s delay by default,
injectable for tests) and only an exhausted budget marks the tunnel —
and thus the connection — unhealthy. Deliberate teardown (cancelForward
/ close) sets tunnel.stopping, cancels any pending restart timer, and
never restarts. Pre-readiness deaths keep failing fast with classified
stderr (auth/bind errors unchanged).
Fixes the kill chain of #96266.
The isinstance(primary, Path) branch in _candidate_sentinel_paths was dead
weight: the surrounding except Exception already covers non-Path test
doubles, and .resolve() failing on them falls through to the plain
inequality comparison. Verified the pre-existing fail-safe stat fixture
(test_is_engaged_fails_safe_on_stat_error) still passes without it.
Module docstring still claimed 'a single os.stat'; the fleet-root check
makes it one or two stats. Updated.
Profile processes launch with HERMES_HOME=~/.hermes/profiles/<name>, so
`hermes pause` at the fleet root did not bind fleet-analyst dispatch
(t_7b65ff88). Check/resume both the process home and the fleet root.
Generalize the HERMES_S6_SUPERVISED_CHILD supervisor-marker mechanism so
ANY supervised gateway launch (systemd, launchd, Windows Scheduled Task,
external supervisor) skips the active_profile redirect in
_apply_profile_override(). Previously only the s6 container marker was
honored, so a systemd-launched default-profile gateway with
HERMES_HOME=<root> followed the sticky active_profile file and silently
assumed another profile's identity — logging under that profile's tree
and connecting with its Telegram bot token (double-polling a token owned
by that profile's own live gateway).
- hermes_cli/main.py: honor HERMES_SUPERVISED_CHILD (new generalized
marker), HERMES_S6_SUPERVISED_CHILD (back-compat), INVOCATION_ID
(systemd; gateway commands only, since it leaks into every descendant
of systemd-launched processes), and HERMES_GATEWAY_EXTERNAL_SUPERVISOR.
- hermes_cli/gateway.py: export HERMES_SUPERVISED_CHILD=1 in generated
systemd units (user + system) and the launchd plist.
- hermes_cli/gateway_windows.py: export it from the Scheduled-Task cmd/vbs
launchers and the windowless respawn env overlay.
- hermes_cli/service_manager.py: export it alongside the s6 sentinel.
- tests: regression coverage for all markers + non-gateway INVOCATION_ID
neutrality + generated-unit marker presence.
Fixes#74872
Add _pid_record_belongs_to_current_profile() helper that verifies a
PID record's persisted hermes_home matches the current process. Use
it in get_running_pid() and get_runtime_status_running_pid() so the
default-profile gateway never mistakes another profile's gateway PID
as its own.
In _apply_profile_override(), clear HERMES_HOME instead of returning
early when it points to a profile directory but no --profile flag was
given, letting the sticky active_profile logic resolve the right one.
In _guard_existing_gateway_process_conflict(), detect stale PID files
from other profiles and emit a warning.