Keep parsing, contracts, gates and persisted goal mutations in one dispatcher. Adapters retain authorization, rendering and scheduling; TUI drafting resolves the target session profile off the RPC reader. Document ACP as unsupported rather than implying a goal loop exists.
Codex OAuth caps gpt-6-astra at the same 272K window as gpt-5.4/5.5/5.6,
so the global 50% trigger compacted at ~136K. Extend the existing
codex_gpt55_autoraise gate to any slug containing "astra" (minus the
opt-in -900k picker variants, which already unlock the wider window).
Other routes (OpenAI direct, OpenRouter) keep the user threshold.
- acquire(): the discard-and-recurse path becomes one more iteration of the existing wait loop;
the two inline 'with lifecycle_lock: _teardown(db)' copies reuse _teardown_generation, and the
type-narrowing asserts go away with the recursion. release() reads generation.path directly.
- Restore the real os.replace inode swaps in test_state_db_file_identity.py and the registry tests:
those files carry no windows_only marker so they never run on Windows, and the monkeypatched
predicate stopped exercising the stat->identity->halt path anywhere.
- Drop the auto-archive change and its 4 tests: on main the sweep gets a bare SessionDB and
db.close() already releases a registry-shared handle, so the described NameError leak only
existed on this branch's earlier head. trace_upload: acquire(None) already defaults.
- Trim the barrier tests to the invariant pair (retired drain must not lift a pending current
teardown; replacement not published before the last close settles) plus the raising-close
settlement; comments say the WHY once.
The #102827 corruption is pure zero holes -- frames lost across a WAL
generation. SessionDB.close() produces exactly that when it runs against a
file another live handle is still writing: PRAGMA wal_checkpoint(PASSIVE),
then the connection close that lets SQLite unlink -wal/-shm. The dangerous
event is a physical close overlapping any other live physical lifetime for
the same path, so both sides of it are closed here.
Late write vs. close: a cron watchdog timeout only stops waiting, and
ThreadPoolExecutor.shutdown(wait=False) cannot interrupt a worker already
inside run_conversation. The agent and its registry reference are now held
until that worker's Future completes, so its last frames land before any
checkpoint.
Close vs. open: the per-path barrier now COUNTS admitted teardowns. A path
can own several closes at once -- the current generation's final release and
a retired generation's drain are admitted independently under the registry
lock, and the per-path mutex only serializes teardowns that already entered
it. With one bare event per path, a releasing thread descheduled between
generation removal and the mutex let the next teardown to settle remove and
signal the shared event: close_all() returned over a pending close and
acquire() published a replacement writer on top of a handle still inside
checkpoint/unlink. _TeardownBarrier tracks event + pending count,
_admit_teardown_locked registers each close in the same lock section that
removes the generation, and only the last settled teardown lifts the
barrier. Physical I/O stays outside the registry lock and unrelated paths
still progress independently.
The auto-archive sweep called release_or_close in its finally while the
import was local to a different function, so every eligible sweep raised
NameError, the outer except Exception swallowed it at debug level, and the
borrowed registry reference was never returned -- a holder leak that pins a
retired generation open. The helper is now bound in the calling scope.
Remaining in-process writable SessionDB() call sites (trace upload, the
API-server profile cache, the web-server writable paths, startup schema
reconcile) go through the canonical registry acquire/release_or_close, and
gateway maintenance borrows pinned handles instead of iterating an unpinned
snapshot.
Regressions: overlapping final releases of the current and retired
generations in both orderings with the first paused before the lifecycle
mutex, teardown-error settlement, an unrelated-path control, and refcount
assertions for the auto-archive sweep on success, on failure, across
repeated sweeps and with auto-archive disabled.
Fixes#102827
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BzxCWw6SuHXhXMdkiEwMa2
_first_user_line indexed splitlines()[0] on a whitespace-only user message
(image-only / tool-only turn) and raised IndexError, which aborted the whole
CLI picker and the desktop session.foreign.list RPC. partition("\n") yields
"" for blank text and the loop moves on to the first real user line.
Reported first in #92290.
- Sidebar: the entry joins SIDEBAR_NAV (same chrome, active state, data-tour
handle as the other rows) instead of a one-off Button below the rail;
i18n moves to sidebar.nav['session-import'].
- Reuse common.retry / common.refresh / common.back instead of duplicating them
under sessionImport in six locales.
- hermes_cli/foreign_session_browser.py -> foreign_sessions_browser.py so it
sorts as a sibling of the foreign_sessions module it extends.
One unreadable or rotated log must not hide the rest of the list; discovery
tolerates OSError from resolve()/stat() per entry and reuses the stat result.
Serve/dashboard runtimes return from match_runtime_outcomes before
_gateway_named_in is reached, and the systemd flavour of this contract is
already pinned by test_serve_reconciles_against_its_own_unit_vocabulary.
The restart phase records macOS LaunchAgent labels (ai.hermes.gateway).
match_runtime_outcomes used a substring check for "hermes-gateway", so a
successful Desktop update on the default profile always tripped
"Planned runtimes the restart phase never touched" and exited 1.
Use the exact systemd/launchd/s6 matcher for both plan reconciliation
and abort-recovery so the two cannot drift.
Portal serves anthropic/* on two routes. The native /v1/messages wire, which
Hermes has used since 02d5e23085, re-writes the previous turn's prompt cache
on 14-20% of consecutive calls in concurrent tool loops; the chat/completions
route does not. Measured 2026-09-06, 20 concurrent sessions x 6 tool calls on
Fable 5.1, same account, same hour, same first-party pin:
Nous /v1/messages 180 pairs, 25 stuck (13.9%); 3 earlier 40-session
runs 15.0 / 17.3 / 20.2%; unchanged by the pin
Nous /v1/chat/completions 320 pairs (2 runs), 0 stuck
OpenRouter direct, pinned 161 pairs, 0 stuck
"stuck" = cache_read on call k+1 equals cache_read on call k instead of
call k's prompt size: the last write was not visible, the turn (~28K) was
written twice. Every stuck pair had byte-identical system, tools and
per-message shas, so it is not client-side mutation. At $10-20/M for
cache writes that is 15-20% of a fan-out's write bill.
The cause is inside the portal's native route (every element of its
outgoing request reproduces clean from outside; NousResearch/api#227
carries the diagnostics). Until it is fixed, anthropic/* rides
chat/completions. nous.anthropic_wire: native opts back in. Cost of
chat: prior-turn thinking travels as OpenAI-style reasoning fields, and
cache_control scopes are translated by the portal's adapter.
Tests: new test_nous_anthropic_wire_default.py reads the knob through the
real config loader in a temp HERMES_HOME and through
resolve_runtime_provider (both settings); the existing native-wire
contract suite selects native via an autouse fixture so the wire keeps
working for the flip-back. Live: a real AIAgent on this branch,
provider=nous + anthropic/claude-fable-5.1, dispatches to
chat/completions with the session_id intact.
Trim the salvaged suite to the behaviour contracts: a picker exception or Ctrl+C leaves
config.yaml's model exactly as snapshotted (parametrized), and an absent active_provider
stays absent through snapshot+restore. The mock-dispatch tests (asserting calls into our own
helpers) are dropped. evals/cli_fallback_add_picker_error.py drives the real
`hermes fallback add` under a Linux PTY into an ordinary picker OSError and checks the
persisted model: stranded on base, restored after.
Parametrized refusal test (bare/fenced/uppercase-fenced truncated objects) asserts
profile.yaml is byte-identical afterwards; one prose test keeps the lenient
fallback contract. Replaces five near-duplicate tests from #104075.
Campaign tracker: https://github.com/NousResearch/hermes-agent/issues/104154
hermes_cli/profile_describer.py's aux-LLM response parser is documented as
"lenient, never raises": when the reply doesn't parse as JSON via
_extract_json_blob, it falls back to treating the WHOLE raw reply as plain
prose and persists it (truncated to 280 chars) as profile.yaml's
description.
That fallback exists for models that ignore the JSON-output instruction and
just answer in prose -- reasonable. But it doesn't distinguish that case
from a reply that DID start out as the requested JSON object and was cut
off mid-object by the aux model or transport. _extract_json_blob requires a
matching closing brace, so a truncated object (no `}` at all, or one cut
off mid-string) returns None just like plain prose does, and the fallback
then persists the raw JSON fragment verbatim: literal leading `{`, `\n`
escapes, a sentence chopped mid-word (#104067).
Fix: when parsing fails, check whether the reply (after the existing
code-fence strip) still looks JSON-shaped (starts with `{`). If so, it's a
malformed/truncated structured reply, not prose -- refuse it and return
ok=False instead of persisting the fragment. Only fall through to the
raw-text-as-prose fallback when the reply never looked like JSON to begin
with, so genuinely prose-only aux replies keep working exactly as before.
Also logs one INFO line on the refusal path, per the issue's note that the
silent fallback made this undiagnosable.
Update (review feedback from @jerrygooch): profile_describer._FENCE_RE was
case-sensitive, so an uppercase ```JSON fence wasn't stripped -- stripping
left "JSON\n{..." which doesn't start with "{", so a truncated response
under that fence variant still fell through to the prose fallback and got
persisted, defeating the fix for that case. Added re.IGNORECASE to
_FENCE_RE, aligning it with kanban_specify._FENCE_RE (which already used
re.IGNORECASE), and added a regression test for the uppercase-fence case.
Added tests/hermes_cli/test_profile_describer.py coverage: a truncated
object cut off mid-sentence (the real-world shape from the issue), one cut
off right after the opening brace, the same truncation under a ```json
fence, the same truncation under an uppercase ```JSON fence, and a
regression guard that a genuine plain-prose reply still hits the existing
lenient fallback unchanged.
Fixes#104067
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LUtGTop5iGKi5vfWnKnwET
The tail clause was correct only by ordering (exit_code != 0 there meant
exit_code is None). Name what it encodes: a stop_reason counts only when
nothing vouched for success. Same truth table. The three literal-dict tests on
the predicate collapse into one parametrized contract; the handoff-exit test
binds to COMMAND_BOUNDARY_STOP_REASON instead of re-spelling it.
update_contract writes {"outcome": "refused", "stop_reason": <code>} with no
exit_code; that is the one production receipt where the stop_reason clause in
_receipt_looks_unfinished is load-bearing. The previous negative control used
exit_code=1, which the exit_code branch already catches. Docstring reworded:
a KeyboardInterrupt never lands on a success receipt (the boundary finalize is
a no-op once the inner path finalized).
docker info needs the /info endpoint, which socket proxies such as tecnativa block by
default, so doctor reported 'docker daemon not running' against a fully working
DOCKER_HOST. docker version hits /version and is what the backend itself probes with.
Independent review found the delegation WAIT was a plain timer: the batch
results arrived with 1,199 s left on a 1,200 s wait and nothing re-judged,
so integration sat unfinished until the timer ran out. The wait now records
how many batches it was set for (GoalState.waiting_on_delegations) and
is_waiting() lifts it the moment fewer are live, so the next idle tick
resumes with the result in hand. A plain timed wait (no delegations) is
unchanged. The TUI/Desktop goal path now passes the live delegation count
to the judge like the CLI and gateway do.
Live with the real auxiliary judge: WAIT 900 s with 4 live batches; one
returns -> is_waiting() False immediately.
Test: parked with 4 live, lifted at 3, plain timed wait unaffected.
JUDGE_SYSTEM_PROMPT allowed WAIT only for "a background process listed
below" or a stated backoff. A fan-out orchestrator's in-flight work is
delegated subagents, which are not registry processes, so when the agent
said "waiting on 4 worker batches, nothing to dispatch" the judge answered
CONTINUE ("no background process is listed to gate on"). In the
1,393-agent run 4 of 5 nudges fired 6-153 s after exactly such a turn and
each bought a status recap: 19 API calls at ~450K context, ~$4.9, and no
work a batch-complete notification would not have produced.
The judge now receives an "Active delegations: N batch(es) still running"
line (count of live async_delegation records whose parent_session_id is
this session, via the existing _session_records selector) and a WAIT
branch for it: wait_for_seconds 600-1800 when the response says it is
waiting on them with nothing dispatchable. Both goal-loop callers (CLI,
gateway) pass the count.
Live A/B with the real auxiliary judge on the run's response shape
("4 worker batches still running ... nothing else is dispatchable until
these return"): main 3/3 CONTINUE, branch 3/3 WAIT 1200 s.
Tests (2): the prompt carries the delegations line only when N > 0 and
a wait verdict parses through; count_active_delegations is scoped to the
spawning session and ignores completed records.
In a multiplex dashboard (one process serving every profile on the host),
a secondary profile's terminal.backend could leak into the shared
os.environ and silently override the launch profile's terminal backend.
A profile configured for terminal.backend: ssh would find itself running
every command locally because a sibling profile with backend: local had
polluted the shared environment.
The tell is an asymmetric env state on the launch session:
TERMINAL_ENV=local # leaked from a sibling profile
TERMINAL_SSH_HOST=10.10.0.103 # still the launch profile's, untouched
A restart does not help: as soon as the sibling profile is touched again
(a turn, a cron tick), the shared env is re-poisoned.
Root cause: env_loader._reapply_terminal_config_bridge() guards "only
re-bridge config into the shared os.environ when this load is for the
process's own profile" via _process_hermes_home(). That helper delegated
to get_hermes_home(), which follows the context-local
set_hermes_home_override a per-request task installs. When a per-turn
handler is scoped to a secondary profile and triggers a dotenv reload,
both sides of the comparison resolve to that secondary profile, the guard
passes, and the bridge writes the wrong profile's terminal.backend into
the shared environment. Because the secondary profile's config carries no
TERMINAL_SSH_* keys, only TERMINAL_ENV is overwritten, producing the
asymmetric state above.
The helper was introduced for the config-cache key (where following the
active home is correct) and later reused for the terminal bridge guard,
where the override-following semantics are wrong.
Fix: delegate _process_hermes_home() to get_process_hermes_home(), the
override-immune resolver built for exactly this. It reflects the scope the
process was launched under, ignoring per-task overrides. Both call sites
(the bridge guard and the secrets-cache reuse check) want the true process
home. Single-profile / CLI behaviour is unchanged: with no active
override the two resolvers are identical.
Complements the terminal_tool._active_environments session-key fix: that
scopes the per-session environment cache; this keeps the shared os.environ
the cache reads TERMINAL_ENV from uncontaminated in the first place.
Adds tests/hermes_cli/test_terminal_bridge_profile_scope.py covering both
the override-immune resolver and the no-leak reload behaviour.
The parametrized heights 5 and 8 exercise the renderer's zero-budget
branch (description too large to fit any row beside the reserved choice
rows), but the test never asserted what that branch draws. Pin it: no
description content and no omission notice, while the choices and key
hint stay visible.
Follow-up polish on #103350 by helix4u.
In a fan-out run the /goal loop parked for 3 h 22 min at the end on a
grandchild's poller (waiting_on_session=proc_21a6fe2369a1, 00:47 -> 04:09)
while the root itself had nothing running. Every one of the root's 7 logged
judge verdicts was WAIT on a child-owned process.
Two causes. gather_background_processes() was called with no task_id from
both goal-loop callers (CLI cli_loops_mixin, gateway run_goals), and the
registry's task_id is the CONTAINER key, which collapses to one value for
every agent in the process, so the judge's process list was every one of
~1,300 subagents' pollers. And a pid/session barrier had no ceiling: once
the judge said WAIT on a session that never exits, nothing resumed judging;
waiting_since was recorded and never read.
Now list_sessions() reports owner_task_id (the RAW spawning id the
registry already keeps for ownership checks), gather_background_processes
takes owner_task_id and both callers pass their own session id (CLI turns
register processes under self.session_id; gateway turns under
turn_ctx.session_id), and is_waiting() ages out a pid/session barrier
after _MAX_BARRIER_WAIT_S (30 min). Timed barriers keep their own deadline.
Tests: only the owning session's running processes are returned when
owner_task_id is given (unfiltered behaviour unchanged); a live barrier
older than the ceiling clears and judging resumes.
A gateway with no profile mapping (or one whose relaunch could not be armed) is
SIGTERMed and listed under "Restart manually" — by design it has no successor and
publishes no fleet-matrix row. It still counted in ``killed_pids`` and the
pre-restart snapshot, so ``_fleet_probe_expected_runtimes`` demanded rows that
could not exist and a fully successful update exited 1 with "Fleet version check
returned no rows even though gateway runtimes were expected", leaving the
fleet_restart_pending marker behind and every later CLI start warning about it.
Track the unmapped stops on the restart outcome and subtract them from the
row-predicting signals (``fleet_probe_signals``); relaunched/systemd gateways
still predict rows exactly as before.
A dashboard-only runtime plan (gateway never started) made
_fleet_probe_expected_runtimes() return True from the unfiltered
'plan.runtimes is non-empty' check. collect_fleet_versions() reports
gateway identities only, so the probe waited for rows that cannot exist,
printed the incomplete-verification warning, and exited 1 after a
successful update (#97332).
Key the plan-derived expectation on kind == 'gateway' records — the same
row-capability rule already applied to the Windows resume token (#93406)
— and update the two tests that pinned the old object() placeholder so
they pin the runtime-kind distinction. Restart-phase, killed-PID, and
pre-restart-PID signals still fail closed unchanged.
Widen the salvaged receipt protection to every updater-owned module: the
pre-update plan's ``RuntimeRecord`` dataclass (``update_inventory``) has the same
failure — the plan is built before the purge and the post-restart reconciliation
filters with ``isinstance``, so an evicted module means a fresh class, zero matches,
and an empty runtime-outcome report. One prefix rule instead of a name list that
has to be re-fixed per symptom.
The runtime-only assertion passed on unfixed main (it exercised
model_tools, not _get_platform_tools). Replace it with the contract the
fix exists for: a toolset is listed iff at least one of its tools survives
_select_tool_names.
hermes tools --summary and CLI /tools used name-level disabled_toolsets subtraction, so disabling debugging still showed terminal/web/file as enabled while runtime stripped those tools. Prune platform toolsets after tool-level subtraction and pass disabled_toolsets into CLI get_tool_definitions calls.
Fixes#97015
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(desktop): treat a macOS ps: parent-marker mismatch as inconclusive, not death
`ps -o lstart=` is a TZ/locale-rendered wall-clock string. Electron caches it
once per app lifetime; the backend re-renders it per spawn. A timezone change
(travel, or the DST boundary) while the app stays open makes the SAME instant
differ byte-for-byte, and _is_serve_orphaned() treated that as proof the parent
died -- os._exit(0) ~5ms after HERMES_BACKEND_READY, silently, on every respawn
until a full quit.
Route mismatches through _parent_start_marker_mismatch_is_conclusive(): linux:/
win:/winms: machine markers stay conclusive (PID-reuse defence intact); any ps:
side degrades to the PID-liveness check the legacy Desktop path already uses.
Log one warning before os._exit(0) so the exit is no longer traceless.
Fixes#95693Fixes#93958
Co-authored-by: Ahmett101 <Ahmett101@users.noreply.github.com>
* fix(desktop): reject truncated ps: parent markers; blank marker env means absent
A ps: marker that got whitespace-split in env plumbing (`ps:Sat`) passed
_valid_parent_start_marker and armed a watchdog that could never match, so
the backend exited 0 right after HERMES_BACKEND_READY. Require a full
lstart value (>=4 tokens, a 4-digit year, a time). Treat empty
HERMES_PARENT_START_MARKER / HERMES_PARENT_NONCE as absent so a blank
inherited value degrades to PID-only tracking instead of disarming or
misfiring.
* chore: map contributor email for drewTuzson
* fix(desktop): log when the parent-death watchdog disarms on an unusable marker
Disarming is the fail-safe branch (the backend keeps serving) but it also
means this backend will never reap itself when the Desktop dies. Leave one
warning naming the rejected marker so that downgrade is not traceless.
* test(desktop): live macOS proof that a TZ-drifted ps: marker no longer kills the backend
Runs the real start_server in a subprocess against the host ps, with the
marker rendered under Europe/Paris and the backend under America/New_York.
Asserts READY + still alive after several watchdog polls, then that the
backend still exits once the stand-in parent is killed. Fails on main with
the reported symptom (exit 0, live parent); macos_only lane.
---------
Co-authored-by: ygd58 <buraysandro9@gmail.com>
Co-authored-by: Ahmett101 <Ahmett101@users.noreply.github.com>
Co-authored-by: Drew Tuzson <drew.tuzson@uqual.com>
* fix(dashboard): prevent PTY input from blocking event loop
* fix(win-pty): don't terminate a healthy ConPTY on write cancellation; log leaked write workers
Review follow-up to the backpressure fix.
CancelledError on WinPtyBridge.write() ran the same path as a timeout and
force-terminated the ConPTY. Cancellation means the owning socket went away
mid-write, which is the keep-alive session's normal reattach case, not a
wedged child; killing the process there defeats the PTY-outlives-socket
design. Give the in-flight write the shutdown grace window and only
terminate if it never lands.
When terminate() fails to unblock pywinpty, the worker stays parked in the
default executor. That was swallowed by a bare except; log it so a slow
thread-pool starvation is diagnosable.
---------
Co-authored-by: Austin Pickett <pickett.austin@gmail.com>
* feat(pty_bridge): mark the dashboard-spawned TUI with HERMES_PTY_HOST
Ink needs to know when its emulator is the dashboard's xterm.js rather
than a native terminal, so it can drop hidden-tab recovery work that only
makes sense for emulators that coalesce output.
Co-authored-by: Raymond <supere989@users.noreply.github.com>
* fix(tui): skip the focus-in erase+repaint under the dashboard PTY
Ink answers a DECSET 1004 focus-in with a full clear+repaint to heal rows
a native emulator may have dropped while the tab was hidden. xterm.js fed
by the dashboard WebSocket never drops frames, and it reports focus on
every OS window blur/focus, so under the dashboard that repaint was a
visible "session reloaded" flash on every alt-tab. Keep the mode
re-assert and keep delivering the focus report to TerminalFocusProvider
(the composer hides its cursor on blur); only the repaint is skipped.
Co-authored-by: Raymond <supere989@users.noreply.github.com>
* fix(web): restore terminal focus after an OS app-switch
Alt-tabbing away and back lands browser focus on <body>, so Ctrl+V never
reached the composer. Pull focus back into xterm on window focus under
the same ownership rule tab activation already uses, extracted into
shouldRestoreTerminalFocus so both paths share it.
Co-authored-by: Raymond <supere989@users.noreply.github.com>
---------
Co-authored-by: Raymond <supere989@users.noreply.github.com>
Follow-up to the salvaged fix from #102994. The quickstart leg tests
stubbed web_deps.late itself, so the resolution path that regressed
(late() -> web_server facade -> getattr) was never exercised; the PR's
proposed tests pinned the module string instead, which fails on any
future move of the symbol even when the call site is updated.
Stub only the leaf (web_server_config._apply_model_assignment_sync) so
the real late() runs, and add one invariant test: _assign_default must
call it with ("main", "llamacpp", model_id, "", "", ""). Red on base
(AttributeError), green with the fix; not tied to where the symbol lives.
Drop the docstring parenthetical that described test mechanics.
The whole-codebase refactor moved _apply_model_assignment_sync from the
web_server facade into web_server_config, but _assign_default in
web_routers/local_models.py still resolved it through the default late()
module (web_server), which no longer carries the symbol. A local-model
quickstart therefore failed its final 'making it your default' step with
AttributeError: module 'hermes_cli.web_server' has no attribute
'_apply_model_assignment_sync'.
Pass the defining module explicitly to web_deps.late() so the call
resolves against hermes_cli.web_server_config, matching how the sibling
models.py router imports the same helper. Update the two quickstart test
stubs to accept the module argument the real late(name, module=...) takes.
The compat sweep dropped the auth-facade re-export of _mark_qwen_oauth_active
while hermes_cli/auth_commands.py still reached it through the auth_mod module
alias, so every 'hermes auth add qwen' died with AttributeError on this branch.
Import it from hermes_cli/auth_qwen where it now lives. Module-alias attribute
reads dodge import-time checks; the sweep in the PR thread found this to be the
only production hit of that class.
Reported-by: yoniebans (PR #102117 review)