Whole-bug-class follow-up to the profiles/ gate: cron/, platforms/
pairing, and legacy pairing/ ran chown_hermes_tree unconditionally on
every boot with the identical warm-boot cost profile. Same
tree_has_non_hermes_owner gate; find evaluates the top directory first
and -quits on the first mismatch, so a mis-owned tree short-circuits in
O(1) while a clean tree pays one read-only walk instead of a full
chown -R inode rewrite.
_reset_fts_index_to_empty used a no-WHERE DELETE, whose docstring
claimed FTS5 treats it as an efficient drop-all. That's true only for
ordinary rowid tables — on external-content FTS5 each deleted row's
tokens are regenerated from the content table, making it O(rows)
(measured ~12us/row: 0.22s @100K, 5.2s @400K, ~25s projected @2M) while
holding the write lock. It also corrupts the index when indexed rows
have diverged from messages — exactly the broken-bookkeeping shape this
repair path handles. The FTS5 'delete-all' special command is the
documented O(1) truncate for external-content tables (measured 1.6ms
@100K) and truncates unconditionally regardless of divergence.
_fts_external_index_empty_with_messages runs on every writable open via
the _init_schema fts_storage_version stamp condition. COUNT(*) is a full
b-tree scan on both messages and messages_fts_docsize (~100ms per open
on a 2M-row DB, measured); the function only ever compares against
zero, so EXISTS(SELECT 1 ...) gives the identical boolean in O(1).
Demote wrote the empty v23 schema via executescript inside BEGIN IMMEDIATE,
which commits early and can leave trash + empty indexes without rebuild
markers. Re-run then tore down trash and stamped fts_storage_version with
docsize=0, permanently losing historical session search.
Stage markers with the demote, create schema only after they are durable,
heal empty-index bookkeeping on resume, and refuse settle until the base
index is populated. Settle refusal returns ok=False instead of raising,
and resume fails fast if the base v23 table cannot be re-created.
Orphan-marker repair only resets a missing fts_rebuild_progress to 0 once
the index is known empty: the chunk worker replays its whole selected id
range without an anti-join, so a partially indexed DB that lost only its
progress key is first reset to a known-empty surface, then rebuilt.
Ported onto the SessionDB mixin split (hermes_state_search.py /
hermes_state_schema.py).
The RO branch's new FTS capability probe raises sqlite3.DatabaseError on
a malformed store (the probe itself only catches OperationalError). The
outer __init__ handler re-raises without closing self._conn, leaking a
tracked connection for the process lifetime — which makes
_backup_db_file refuse its raw-copy, so the writable heal that follows
(web_server's stale-schema/malformed reopen) repairs the store WITHOUT
the forensic backup repair_state_db_schema promises. Close-then-reraise
on any probe failure, mirroring _open_probed's cleanup discipline.
Regression test: corrupt sqlite_master (duplicate messages_fts row),
assert the failed RO open leaves no live tracked connection and the
subsequent writable heal creates its malformed-backup file. Mutation-
checked: no-oping the cleanup handler makes the test fail.
Review follow-up on the #76194 salvage: the readonly swap makes this
function leak the live config-cache 'approvals' sub-dict to callers.
All current callers are read-only (audited); the docstring now carries
the do-not-mutate contract for future ones.
The terminal-command guard path loaded config 2-3x per invocation via
load_config(), which pays a defensive deepcopy of the entire config on
every call (~356us of the ~376us warm-cache cost measured on a real
config.yaml). All six swapped call sites were audited read-only — every
caller takes scalar reads or iterates the returned structures; none
mutate (the save path at save_permanent_allowlist keeps load_config) —
so they now use load_config_readonly(), the API built for exactly this
(precedent: #74211, #74322; the one unsafe-site lesson from #56085's
salvage is covered by the mutation audit and a cache-integrity test).
Measured (real config.yaml, warm cache): load_config 376.0us ->
load_config_readonly 19.9us (18.9x); full guard pass
check_all_command_guards('ls -la','local') 930.7us -> 241.8us (3.85x).
Tests: new test_approval_config_readonly.py drives the real functions
against a temp HERMES_HOME — readonly call counts per function, a
no-deepcopy pin for the full guard pass, and cache-identity/integrity
checks. Existing test mocks retargeted from load_config to
load_config_readonly (same injection intent). Note: 6
test_approval_mode_parity failures are pre-existing ordering flakes —
identical with the change stashed on clean main.
Review follow-up on the #76098 salvage: the 4096-entry count bound alone
doesn't bound MEMORY — write_file/patch argument strings run 100KB+, so
a long-lived gateway process under sustained heavy write workloads could
pin ~800MB of evicted-session strings. A 32MB byte budget extends the
existing FIFO eviction; common-case args (0.5-2KB) never hit it. New
guard test mutation-checked (fails with the byte leg disabled).
The pre-send normalization pass re-canonicalized every historical tool
call's argument JSON on every API-call iteration — quadratic in session
tool-call count. Route it through a bounded value-keyed memo (the
_MSG_TOKENS_CACHE idiom from agent/model_metadata.py): per-iteration
cost is now proportional to new tool calls, not all of them.
Measured (simulated growing session, repo venv): session-total
canonicalization cost 1056 ms -> 58 ms at 500 tool calls x 2 KB args
(18.3x), 5744 ms -> 79 ms at 500 x 16 KB (72.5x). Byte-parity with the
pre-fix logic is asserted at every iteration (unicode, nested,
malformed, empty, non-string args), and a call-count test proves
json.loads invocations went from K(K+1)/2 to K per session.
Review follow-up on the #76142 salvage: MemoryStore path-resolves and
shares one process-wide connection per file, so MemoryStore(":memory:")
creates a literal ./:memory: FILE whose state leaks across test runs —
the second run of the file failed all three spy tests because the
NULL-vector test had permanently wiped hrr_vector in the leaked db.
tmp_path isolates each run; verified two consecutive runs green + full
tests/plugins/memory/ green.
FactRetriever.search() re-encoded the query vector once per candidate,
related() re-encoded both role atoms once per fact row, and probe()
re-encoded the role-content atom once per row. All three encoders are
deterministic (SHA-256 counter blocks), so the hoisted vectors are
bit-identical to the per-iteration values they replace.
Measured (300-fact store, dim=1024, median of 30 calls): search()
11.62 -> 1.46 ms/call (8.0x; encode_text 30 -> 1 per call), related()
63.08 -> 16.17 ms/call (3.9x; encode_atom 601 -> 3 per call), probe()
431.93 -> 389.36 ms/call (1.1x; dominated by per-fact content encoding,
which is inherent to the algorithm and unchanged).
Tests: call-count regression tests for each hoist plus a bit-exact
parity test of search() against the pre-fix per-candidate loop.
Every ORDER BY id query on the messages table sorted or scanned the
whole session: get_messages_around's window seek, latest_message_row_id
(LIMIT 1), and get_messages' full-load ordering all paid O(session
history) per call — hot mid-turn via session_search and reactions.
messages.id is an original column (INTEGER PRIMARY KEY AUTOINCREMENT),
so the index lives in SCHEMA_SQL next to idx_messages_session — no
legacy-column migration hazard (the kanban lesson from #28776 does not
apply).
Measured (real schema, one 20k-message session, median of 30):
get_messages_around 7.08 -> 0.22 ms (32x), latest_message_row_id
3.37 -> 0.011 ms (307x), get_messages full load 111.6 -> 98.6 ms
(1.13x — remaining cost is row deserialization, not the sort).
Window results byte-identical at probe points across the session.
Tests: VM-step pin (get_messages_around bounded work, calibrated
~12 vs ~855 handler calls, threshold 300 — fails without the index)
and window parity with/without the index. No EXPLAIN/plan text
(behavior contracts, AGENTS.md).
Review follow-ups on the #76287 salvage:
- scheduler.py no longer calls advance_next_run after the batch switch;
keeping the import invites a future test to patch the wrong seam.
- advance_next_run returns >= 1 instead of == 1 so a corrupted jobs file
with duplicate ids (both records advanced by the batch) still reports
True after advancing and saving.
The scheduler's pre-dispatch loop called advance_next_run per due job —
one full load_jobs() + one full save_jobs() of the jobs file each — so
N due jobs cost N reads + N writes of the whole file (gateway-restart
catch-up or co-scheduled bursts). advance_next_runs() does one load +
at most one save for the whole due set with identical per-job semantics;
advance_next_run() is now a thin wrapper over it.
Measured (50 due recurring jobs, real jobs file): 107.9 ms -> 2.5 ms
(45x; 50 loads + 50 saves -> 1 + 1).
Tests: batch advances recurring and skips one-shots, single load + save
I/O pin (fails pre-fix — no such function), no save when nothing
advances, and per-job wrapper semantics unchanged. Related: #60946 and
#75833 both restructure this loop's call site for correctness — neither
addresses the I/O cost, and this batch primitive composes with either
dispatch design; happy to rebase onto whichever lands first.
Review follow-ups on the #76357 salvage:
- _message_reference_from_ids + _reply_reference_for_send collapse the
3x duplicated MessageReference construction (naming mirrors telegram's
_reply_to_message_id_for_send).
- The overflow elif's comment claimed PartialMessage has no to_reference;
discord.py 2.7.1's PartialMessage does (message.py L1901) — the branch
is belt-and-suspenders for duck-typed priors, now labeled as such.
Every reply paid one extra Discord API round trip: the text send path,
the voice send path, and the edit path each called fetch_message() just
to obtain a reference or an editable handle. Discord resolves
message_reference payloads from ids alone, and PartialMessage.edit()
works without a fetch — so build MessageReference directly (with
fail_if_not_exists=False, preserving the deleted-target behavior the
existing send-side 10008 retry already covered) and use
channel.get_partial_message() for edits. Overflow continuations keep
threading via an ids-built reference fallback for PartialMessage.
Measured by call count (deterministic): reply sends and edits now make
ZERO fetch_message calls where they made 1 per reply and 1 per edit
(including every streaming edit tick).
Tests: pin that first-mode replies construct the reference without any
fetch, deleted-target retry test updated to assert fetch await_count==0
(retry now happens purely send-side), overflow/edit mocks retargeted
from fetch_message to get_partial_message (any fetch regression breaks
all five). Note: 4 discord-suite failures are pre-existing ordering
flakes — identical with the change stashed on clean main.
Architecture fix for the bug class behind the Termux --version NameError
(live on main since eb4040242): version-printing kept being reimplemented
as *_fast() copies at the top of hermes_cli/main.py, each duplicating
canonical logic (project-root resolution, container detection, profile
detection). The copies drift silently — eb4040242 edited the canonical
output and referenced the PROJECT_ROOT module constant inside the fast
function, which doesn't exist yet at the fast exit point.
- hermes_cli/_startup_fast.py: THE implementations, stdlib-only. main.py's
*_fast() names become thin delegates (kept for test/back-compat), and
PROJECT_ROOT itself derives from the same helper — the constant and the
fast path can no longer disagree.
- Fast output now includes the .install_method stamp (one cheap file read)
and a 'Run hermes version for update status' pointer, so globalizing the
fast path doesn't silently drop slow-path info.
- Guard tests: (1) import-weight — subprocess-imports _startup_fast and
fails if any heavy module (config/yaml/argparse/cli/run_agent/httpx)
lands in sys.modules; (2) subprocess parity on+off Termux — the test
that would have caught eb4040242 the day it landed; (3) install-method
stamp surfacing.
hermes --version: ~3.8s cold / 0.2-0.4s warm -> 0.01-0.02s everywhere.
- _print_fast_version_info referenced the PROJECT_ROOT module constant,
which is defined AFTER the ultrafast exit point. On current main this
is a LIVE latent bug: the Termux fast path NameErrors on --version
(eb4040242 changed the print to use PROJECT_ROOT without noticing the
constant doesn't exist yet on that path). Compute the root locally.
- PR tests asserted the old 'Project:' label; main renamed it to
'Install directory:' (eb4040242). Expectations updated.
Exact parity with utils.atomic_replace: its target fsync is wrapped in
try/except OSError. A failed fsync after a successful copy must not
surface the already-completed write as an error (Windows can raise on
fsync of a read-only handle).
- agent_import.dump_yaml_file now calls utils.atomic_yaml_write instead
of hand-rolling safe_dump + atomic_write_text — same temp+fsync+atomic
rename and symlink preservation, plus mode/owner preservation a
0600-secured config.yaml needs
- openclaw script: the EXDEV/EBUSY copy fallback gains copystat + target
fsync so the docstring's 'mirrors utils.atomic_replace' durability
claim is true on cross-device deployments
- trim load_yaml_file's docstring to the behavior contract
The inlined temp-file + os.replace in openclaw_to_hermes.dump_yaml_file
replaced a symlinked config.yaml with a regular file, silently detaching
managed deployments that symlink ~/.hermes/config.yaml into a dotfiles repo or
profile package. The bare path.write_text it replaced followed the link, and
utils.atomic_replace -- which the hermes_cli twin reaches through
atomic_write_text -- resolves the link for exactly this reason (#16743).
Mirror that here: resolve the symlink before creating the temp file so the
rename lands on the real file, and fall back to copyfile on EXDEV/EBUSY now
that the target can live on another device. Covered by a regression test that
fails when the resolution is removed.
Also guard the permission-denied test for Windows: os.geteuid does not exist
there and chmod-based denial is unreliable, so skip on non-POSIX.
agent_import.py carries a private load_yaml_file/dump_yaml_file pair that
returned {} for an absent file AND for a present file it could not read or
parse. Three importers -- import_permission_allowlist, import_permission_denylist
and import_mcp_servers -- read config.yaml through it, merge one section into
the result, and write the whole mapping straight back. So a YAML syntax error,
a permission problem or a broken mount meant the importer replaced every
setting the user had with only the one to three keys it merged, and still
reported the item as "imported". The write was a bare path.write_text(), so an
interrupted import truncated the file instead.
Distinguish the two cases at the read. Absent, or present but empty, still
yields {} so first-time creation works. Present but unreadable, unparseable, or
not a mapping raises ConfigReadError; the three sites funnel through a new
load_target_config() that records the refusal as a per-item error and leaves the
file byte-identical. Dry-run refuses too, rather than previewing an "imported"
that would destroy the config. dump_yaml_file now writes through
utils.atomic_write_text, which the module already imports and uses for the
memory store.
This is the invariant hermes_cli/config.py already enforces for its own writers
via require_readable_config_before_write / atomic_config_write, whose docstring
names this exact root cause and calls itself "the single chokepoint every
config-update path should use". agent_import.py has its own helper pair and so
was never covered; it was the last config.yaml writer without the guard.
The identical helper pair lives in openclaw_to_hermes.py, the script this module
was ported from, where twelve config.yaml read-modify-write sites share the same
defect; fixed there too. Its refusal is recorded at the run_if_selected dispatch
point, which flips the existing _config_apply_blocked flag so the remaining
config-mutating options short-circuit instead of each rediscovering the same
unreadable file. The atomic write is inlined with tempfile + os.replace because
that script runs standalone with only the stdlib on its path.
kill_started_since duplicated kill_all's collect-under-lock/kill-outside-lock
loop line for line; it is now a thin delegate through new kill_all kwargs
(exclude_ids, source, consume_output). Public signatures unchanged — existing
callers and test monkeypatch seams keep working. kill_process's docstring now
names the deliberate consume_output=True exception for abandoned-turn reaping
so the deviation isn't 'fixed' later.
The API server intentionally lets concurrent runs share a client-provided
session_id (= process task_id), so the SSE-disconnect reap could kill a
process a still-live concurrent run spawned after the disconnecting run's
baseline — the same stale-reaper bug class the gateway path gates via
run_generation.
- Per-task-id run epochs (monotonic counter): each run claims the epoch at
publish; a reaper holding a superseded epoch declines to kill. A missing
entry (the run's own clear pruned it) still reaps, so the leak fix isn't
silently disabled.
- _publish_turn_process_ownership / _clear_turn_process_ownership helpers
replace the copy-pasted marker set/clear blocks, so attribute names and
epoch bookkeeping can't drift between surfaces.
- /v1/runs — the third own-lifecycle surface — now records ownership and
reaps on POST /v1/runs/{id}/stop and on server-side SSE cancellation,
closing the remaining sibling paths of #76115.
Two follow-ups from review of the salvaged fix:
- _reap_gateway_turn_processes now returns 0 for a blank task_id.
ProcessSession.task_id defaults to empty for sessionless callers, so a
blank turn id would match (and kill) every unrelated empty-task process
instead of the turn's own.
- The asyncio poll loop checks executor completion BEFORE the watchdog's
timeout flag. When both race in the same window, the completed run has
already persisted its real reply to session history; surfacing the
'agent inactive' diagnostic would contradict the stored transcript.
This matches _abandon_timed_out_gateway_turn's own worker-done-wins
tiebreak.
dbbb10d39 shipped without direct test coverage for its own new logic
— the same gap teknium's review flagged on the competing PR. Close it:
- _reap_gateway_turn_processes: skips when is_still_current() is
False, proceeds when True, fails open (reaps) if the check itself
raises rather than silently disabling the leak fix.
- _abandon_timed_out_gateway_turn: still marks the turn abandoned
(interrupt fires) even when the reap itself is skipped.
- api_server._reap_disconnected_agent_processes: reaps the
baseline-diff for an owned turn, no-ops when the agent never
recorded ownership markers.
- APIServerAdapter._run_agent: markers are populated with the right
task_id/baseline during the turn and cleared once it completes,
closing the same race window fixed in gateway/run.py for this
separate agent-lifecycle surface.
Addresses the hermes-sweeper review on #76188:
1. task_id is session-scoped (task_id == session_id), not turn-scoped,
and the reap runs on a detached thread. A replacement turn could
claim the same session and spawn a legitimate process before the
previous turn's reaper thread actually enumerates its targets,
killing that new process by mistake.
Fixed by gating the reap on the existing run_generation mechanism
(_is_session_run_current) instead of inventing a new ownership
token: the timeout path captures its own run_generation at turn
start, the interrupt path captures the generation immediately after
invalidating it. If a newer turn has since claimed the session, the
reap is skipped — that newer turn owns its own baseline, so nothing
is left permanently unreaped.
2. gateway/platforms/api_server.py's SSE handlers for chat-completions
and the /api/sessions responses endpoint run their own agent
lifecycle via _run_agent() and never passed through TurnRunner, so
client-disconnect abandonment there had no baseline and no reap —
contradicting the PR's stated disconnect coverage. Both disconnect
handlers now snapshot/reap through the same
tools.process_registry primitives, via a small
_reap_disconnected_agent_processes() helper shared by both call
sites.
An agent turn can spawn a long-running background subprocess (e.g.
`next build`) and later be abandoned via inactivity timeout, /stop,
/new, or a client disconnect. Before this fix the gateway interrupted
the agent loop but never touched the subprocess: it kept running
inside the gateway's cgroup, unbounded, until memory pressure starved
the event loop and made every platform/cron look hung (#76115).
The process registry already knew how to kill a process tree — the
missing piece was per-turn ownership: nothing distinguished a process
that predates the turn (must survive), a process the turn started and
finished successfully (must survive), and a process an abandoned turn
left running (must be reaped).
- tools/process_registry.py: snapshot_running_ids() captures a turn's
starting baseline; kill_started_since() reaps only IDs created after
it, scoped to one task_id.
- gateway/turn_context.py: TurnContext carries process_task_id +
process_baseline so the timeout/interrupt paths can reach them.
- gateway/run.py: baseline is snapshotted right before the turn's
executor task starts; the inactivity-timeout path and the explicit
/stop|/new|disconnect interrupt path both reap via the same helper.
A daemon-thread watchdog backs up the asyncio-based timeout poll,
since a starved event loop is exactly the failure mode this bug
causes. The turn's own worker clears its ownership markers the
instant it finishes, closing a race where a /stop landing right
after normal completion could reap a background process the turn
deliberately left running.
Related but insufficient on their own: #37454 (cgroup ExecStopPost
reaper only fires on service restart) and #68915 (orphaned-pipe
grandchild detection, a registry bug not a turn-lifecycle gap).
Neither ties process cleanup to turn abandonment.
- logger.warning when the 6h ceiling stops the heartbeat (matches the
delegate_task stale-stop precedent) so the eventual watchdog reap is
explainable from logs instead of silent
- new mutation-checked test: past the ceiling the heartbeat stops while
the job still completes
- clearer assertion messages (surface res on failure)
- heartbeat loop continues past a raising activity callback instead of
silently stopping (matches delegate_task / touch_activity_if_due
swallow-and-continue semantics) — one transient error must not drop
watchdog protection for the rest of a long job
- hard 6h elapsed ceiling so a wedged job under HERMES_CRON_TIMEOUT=0
(unlimited child watchdog) cannot mask the gateway watchdog forever
- public get_activity_callback() accessor in tools/environments/base.py
instead of importing the private _get_activity_callback cross-module
- tests: deterministic heartbeat test (event-gated, no timing sleep),
no-callback test now asserts the thread is truly never created, new
exception-survival guard; dead started event removed
- fix comment: delegate_task heartbeat cadence is 30s, not 10s
/simplify-code review found _poll_for_token has a second caller:
web_server._nous_poller (dashboard/desktop device login), which surfaces
str(e) as the UI error_message — so wrapping only in
_nous_device_code_login left the dashboard showing the bare timeout.
Move the enrichment into _poll_for_token's deadline raise so every
caller inherits the guidance, and drop the now-redundant try/except
wrap in the CLI login. Add a source-level regression test driving the
real poll loop (authorization_pending stub client) to the deadline.
A bare 'Timed out waiting for device authorization' gives the user
nothing to act on. The most common cause is Portal sign-in failing in
the opened browser tab (including the server-side CAPTCHA loop from
issue #20605), so point at the Portal login page and the hermes portal
retry command.
Salvaged from PR #75290 by @HexLab98 (timeout-guidance kernel only).
The URL-rewrite portion of that PR was dropped: the live Portal has no
/device route (verified 404 with a real user_code), so rewriting the
manage-subscription verification URL would break login entirely.
Guidance text reworded to reference only real URLs.
/simplify-code findings on the salvage stack:
- extract _category_skill_dirs() as the single category detector; the
install guard and hermes_cli._existing_categories() now share it
(third copy of the heuristic eliminated)
- filter rglob hits through is_excluded_skill_path so vendored /
support-dir SKILL.md files (node_modules, references/pkg) no longer
misclassify a plain directory as a category and block install
- fix inaccurate WHY comment (lock-file check only guards hub-installed
skills, not hand-authored dirs), drop underscore prefixes on locals,
fold the file-collision guard under the single exists() check
kshitij