`hermes update` runs in the pre-pull interpreter. The auto-restart phase
imports freshly-pulled gateway source, which resolves sibling imports
against the OLD sys.modules cache — so any update where an already-cached
module gained a new export ImportErrored the whole phase and left the
gateway serving pre-update code (2026-08-20 field failure: new gateway.py
needs cli_output.line_input, cached cli_output predates it).
Class fix replacing the per-symptom _UPDATE_RUNTIME_RELOAD_MODULES
approach: _purge_stale_hermes_modules() evicts every cached module under
the Hermes package prefixes (hermes_cli/gateway/tools/tui_gateway/agent)
right before the restart phase, so later lazy imports rebuild a
self-consistent module graph from the updated checkout. The updater's own
executing modules are exempt (purging them buys nothing; reload-in-place
is the unsafe op, and we never reload). Root-segment check spares
prefix-lookalike packages. Best-effort, never raises.
5 new tests incl. an end-to-end repro of the field failure shape
(stale module missing symbol -> ImportError -> purge -> import resolves).
tests/tools/test_website_policy.py and tests/cli/test_surrogate_sanitization.py
repeatedly failed under the parallel CI runner (process-teardown timeouts /
async-timeout flakes) across multiple unrelated PRs this cycle, while passing
locally. Removed per maintainer direction to stop the flake taxing every PR.
posix.sh now probes `update --help` before the real update call; the fake
counted the probe as call #1, shifting the exits.N mapping so the retry
gate never fired. Answer the probe out-of-band so counted calls remain
actual update attempts.
The desktop updater ran `hermes update --yes`, which auto-restored any
uncommitted source-tree edits onto the freshly updated checkout. On dirty
from-source installs this silently carried local modifications across every
update and could break the rebuilt app (field report: Windows update handoff
leaving the app 'crashed').
New `hermes update --keep-stash`: local changes are still autostashed so the
update can proceed, but are never re-applied — they stay parked in git stash
with printed recovery guidance. Both desktop handoff scripts (windows.ps1,
posix.sh) now pass it, probing `update --help` first so older installed
backends without the flag keep working. Failure paths are unchanged (stash
preserved, no restore); updates.non_interactive_local_changes: discard still
wins.
Tests: park/restore/failure-path coverage incl. a sabotage-verified
regression test; docs updated.
Builds on @fattchris resolve_turn_limit salvage (#67696): flips the default
from a numeric cap to unlimited across all construction paths (CLI, agent_init,
run_agent subagents), adds inf/infinity/null to the unlimited spellings, and
sets DEFAULT_CONFIG agent.max_turns to null. The turn cap caused more problems
than it solved (silent mid-task truncation).
Addresses teknium1 sweeper review on PR #67696:
1. Gateway bridge: Skip str(None) bridging when YAML value is Python None
(from or bare ). Previously str(None) → None → unlimited
instead of default 90. Now clears stale env var so resolver applies default.
2. TUI: Route _cfg_max_turns through resolve_turn_limit instead of bare
int(). Old code crashed on none/unlimited and swallowed 0 via
. HERMES_TUI_MAX_TURNS env var also routed through
resolver.
3. Docs: Document unlimited spellings (none/unlimited/infinite/0/-1) in
configuration.md.
4. Tests: Add TestGatewayBridgeNullHandling (4 tests) and TestTUIResolver
(8 tests) covering null handling, string spellings, env var override,
and legacy root-level config.
All 50 tests pass.
Previously agent.max_turns only accepted positive integers. Setting it to
'none', 'unlimited', or 0 — all natural ways to say 'no limit' — either
crashed int() or was silently skipped by `or` checks, falling back to 90.
This adds resolve_turn_limit() in hermes_cli/config.py as the single
normalization point. It accepts:
- int/float → int(raw) (floats truncated)
- numeric string ('120') → int(raw)
- 'none'/'unlimited'/'infinite'/'∞'/'-1'/'0' (case-insensitive,
whitespace-tolerant) → sys.maxsize sentinel
- YAML None/null → default (90)
- bool/list/dict/garbage → default (with debug log)
All config-reading sites (cli.py, gateway/run.py, cron/scheduler.py) now
call this instead of bare int(), so agent.max_turns: none in config.yaml
becomes a first-class supported spelling of 'unlimited'.
The sentinel (sys.maxsize) survives the str()→int() round-trip through
the HERMES_MAX_ITERATIONS env-var bridge in gateway/run.py and works in
every <, >=, remaining = max - used comparison without requiring call
sites to learn about a special value.
Includes 38 tests covering the full spelling table, the str→int env-var
round-trip, and sentinel properties.
Every long-lived Hermes process is now positively identifiable so reapers
never have to guess lineage from PPID archaeology or cmdline shape:
- hermes_cli/process_identity.py (new): HERMES_SPAWN tag build/parse,
spawn-ledger.json self-registration keyed on (pid, create_time) — PID
reuse cannot forge the pair — with #89298-style corrupt-file quarantine,
and a kill-on-close job-object self-attach (BREAKAWAY_OK preserved for
the existing CREATE_BREAKAWAY_FROM_JOB escape hatches).
- serve/dashboard (web_server.py) and the gateway entry point register
themselves at startup and attach to the job; Desktop legacy
HERMES_PARENT_PID/winms marker reused as spawner identity so lineage
works with every Desktop version.
- Desktop stamps HERMES_SPAWN on backend spawns (parent-process-identity.ts).
- hermes update gets a positive-identity rung ahead of the heuristic ones:
_ledger_reapable_backend_pids reaps holders the ledger PROVES are orphaned
backends (purpose reapable + recorded spawner provably dead) in ANY update
context. Ledger-unknown holders fall through to the existing rungs.
22 new tests, sabotage-verified.
The one-shot keyless extract rescue (d1eefe6ac) treated ANY whole-batch
failure as a backend outage. A website-policy refusal also arrives as a
failed batch, so blocked URLs were routed through the free-tier ring:
in CI the ring's live fetch attempt returned a result for the wrong URL
or a bare None error, turning test_website_policy reds on main (slices
8/12 and 12/12) — and in production it would fetch content the user
explicitly blocked.
_rescue_extract now partitions policy blocks (blocked_by_policy flag or
policy error text) out of the rescue set: they are preserved verbatim,
only genuine failures ride the ring, and order/merge parity is kept.
Two sabotage-verified regression tests pin the class.
Field incident (2026-08-20): a Windows Desktop update hand-off
(update --yes --gateway --force) left a swarm of per-profile serve
backends (mr-tester, probe-inherit, turqoise, clippy, maroon, …) holding
cryptography/_rust.pyd. Some still had a live parent (the tearing-down
Electron process, or the venv launcher->worker two-hop chain mid-exit),
so the strict orphan-only reap (_orphaned_desktop_backend_pids, which
bails the instant ANY holder has a live parent) disqualified the whole
set and the venv-holder guard dead-ended. The user saw a ~12-minute hang,
force-closed, and the half-done state stranded bot sessions.
New rung: _handoff_reapable_backend_pids reaps surviving Hermes
serve/dashboard backends from this venv — live parent or not — but ONLY
in the hand-off context the caller gates on: args.gateway AND the
update-incomplete marker present AND no live hermes.exe shim. In that
window nothing legitimate supervises or respawns a serve backend (the
Desktop tree-kills its backends and parks any relaunch behind the marker,
#50238), so a surviving backend is a leak, not a race. A non-backend
holder (operator REPL, stray script) still disqualifies the whole set;
psutil-unavailable returns None (keep refusing). Wired as the final rung
before the existing dead-end, after the orphan-only reap.
One conflict, gateway/relay/adapter.py send_for_platform: main added the
turn-final draft-seal interception (_sfp_metadata with the _interim_send
marker stripped, seal-or-fall-through); this branch added format-hint
stamping on the same frame. COMPOSED: the plain-send frame now stamps
_with_format_hints_for_platform over _sfp_metadata (the stripped copy),
so both the seal fall-through contract and the cron-lane block hints
hold. Note: the seal frame itself (op:draft final) does not stamp hints
— cron sends are never open drafts, so the flagship path is unaffected;
noted as a connector-PR follow-up for streamed interactive finals.
seed_mock.assert_not_called() alone could pass for the wrong reason — a
harness failure before delivery also leaves the seed uncalled. Assert
the real adapter recorded exactly one live send to the origin chat, so
the test pins the D6 thread-fallback decision, not an accidental
no-delivery.
Unspecced MagicMock/AsyncMock adapters fabricate
supports_inchannel_continuable_for_platform as a truthy callable, so the
scheduler's duck-typed D6 gate silently took the relay accessor branch in
every in-channel test — the native scalar fallback the fixtures describe
was never exercised, and setting supports_inchannel_continuable=False on
a mock could not force thread mode. Pin the accessor to None on both
mock fixtures (matching a real native adapter, which never defines the
method), and add a fallback-boundary test with a real minimal adapter
class: scalar False -> in_channel fails safe to thread, flat seed never
fires.
`hermes update` on Windows detached on every run, including the
`Already up to date!` no-op that never touches the venv. emozilla hit
the visible half: the shim exits, PowerShell takes the console back,
and a child prints the result under a fresh prompt — it reads as a
frozen update. The invisible half is worse: the hand-off sat ahead of
the fetch, so it also carried off the stash and branch-switch
questions, which #90205 then had to answer by closing stdin. Nobody
who mods Hermes got asked about their local changes again.
The shim lock is real and the child is still required — a launcher
holds venv\Scripts\hermes.exe open without FILE_SHARE_DELETE for the
whole command, so the quarantine rename is refused and uv fails with
os error 32. A parent that waits deadlocks against the handle it is
itself holding, and Windows has no exec to escape with.
But that lock only binds one step. Move the hand-off to the dependency
sync boundary, beside the native-module deferral that solves the same
"this process holds a file the sync must replace" problem — and for
the reason that placement already exists (#86735: a preflight ahead of
the fetch re-bricked the flow it was meant to protect). Everything
before the sync now runs foreground in the user's console: the
preflight, the stash question, the branch switch, git pull. An
up-to-date run never hands off at all.
Deferring to the next launch cannot substitute here the way it does
for a mapped .pyd: every future `hermes` launch is also the shim, so
the marker would defer forever. The child re-runs the update to keep
the node/web/lazy-refresh tail, and takes the sync it was spawned for
rather than the up-to-date early return.
The seed-key fix made the SESSION scoped, but the delivery leg still
dropped the scope: cron route_metadata carried only job_id (+thread),
DeliveryRouter stamps scope_id only for the configured HOME channel, and
the RelayAdapter's per-chat scope cache is cold after a gateway restart
(learned from inbound only). A scoped Slack origin that is not the home
chat therefore egressed with NO tenant discriminator, and the connector's
fail-closed guard could reject the brief before delivery — the
delivery-leg sibling of the seed-key scope gap.
Copy origin.scope_id into the live text and media routing metadata for
ORIGIN-MATCHING targets only (setdefault — never overrides router/home
stamping). Fan-out/broadcast targets are excluded by the origin gate: a
fan-out target's tenant is not the origin's, and a wrong scope is worse
than none.
Tests: restart-shaped positive (scoped non-home origin -> scope_id on
routed metadata, RED before this fix) and legacy negative (scope-less
origin stamps nothing).
Every drive_preview action answered with the entire inventory — around 120
elements of ref, role, label, and an up-to-eight-rung `:nth-child` selector
chain. On a real app shell that was ~24.5k characters, re-sent after every
click, so a ten-step task paid for ten copies of a page that had barely moved.
Handles are now durable and legible. An element is named after what it is and
what it says — `btn-sign-in`, `inp-email`, `srch-search-projects` — minted once
per page and never reused, with duplicates disambiguated as `btn-edit`,
`btn-edit-1`. Each one remembers a stable attribute, its role, its accessible
name, and the nearest landmark it sits in, so when a framework destroys the
node and builds a new one the handle moves across and the agent is told
`rebound` rather than being handed a removal it has to react to and an addition
it has to re-read. The re-bind ladder is anchortree's (Apache-2.0), minus its
geometry rung, which can never clear the threshold on its own.
Because the handles hold, the first look at a page returns the inventory and
every look after it returns only what moved. `changed` carries the ref and
whichever of label/value/disabled actually shifted — role and selector are
absent by construction, since a change in either would mean the re-bind ladder
was looking at a different element. A delta gives way to a full re-read when
half the page is new, where there is nothing left to reuse.
The selector column is gone with it. It was 74% of the inventory on an
85-element page, nothing downstream ever read it, and a positional chain is
wrong the moment a sibling appears. An `#id` or `[data-testid]` survives when
the page offers one; everything else is addressed by handle.
Legibility is what makes the delta work rather than a nicety. `+ btn-sign-in`
on turn nine reads on its own, where `+ @e42` sends the model back to an
inventory twenty thousand tokens ago.
Measured on an 85-element app shell: 18,693 -> 4,930 characters for a baseline,
and a steady turn that moved two things costs ~200.
The in-app browser was a one-way mirror. open_preview put a page in the pane
and read_preview read its text back, but nothing could touch it. A click meant
falling back to the browser_* tools, which drive a separate Chromium the user
cannot see — so "log into this and pull my invoices" happened in a different
browser from the one on screen, with none of the sessions the user is already
signed into.
Four pieces, and they only make sense together:
· an in-page engine that inventories what is interactable and performs the
verb, injected as source because it has to run inside the guest page;
· the preview.act.request bridge from the gateway into the pane;
· drive_preview, for acting: elements, click, type, scroll, press, and the
pane's own back/forward/reload;
· annotate_preview, for marking without acting.
Those last two started as one tool doing two unrelated jobs. Leaving a mark is
not an action — it outlives the turn that drew it — so it gets its own verb,
and the interaction verb gets a name that says what it does.
Gating is the existing surface rule: desktop_ui folds in on session
source: 'desktop', and the bridge refuses to act for a background session, so a
turn running behind the user's back cannot reach into the page they are working
in.
Two details worth a reviewer's attention. Typing assigns through the
prototype's value setter, because React shadows value with its own accessor and
ignores an input event whose value it believes it already wrote — a plain
el.value = … types into a field that snaps back on the next render. And
clicking replays the pointer/mouse pair before activation, because frameworks
bind to mousedown as often as to click.
#84327 excluded _spawn_supervised's permanent watchers (session-expiry,
kanban, reconnect, the scale-to-zero watcher itself, ...) from
_scale_to_zero_has_live_background_work() via a _hermes_supervised_watcher
tag, because counting them made an armed gateway consider itself busy
forever and never go dormant.
Two more permanent, infinite-loop tasks are added to _background_tasks
OUTSIDE _spawn_supervised and were untagged:
- _loop_heartbeat_task (loop_heartbeat_forever, #66892): a `while True`
loop started unconditionally in start() on every gateway boot. Extracted
the inline spawn block into _start_loop_heartbeat_task() so it's
independently testable, matching the existing _start_heartbeat_poller()
pattern.
- _heartbeat_poll_task (_poll_loop in _start_heartbeat_poller): also a
`while True` loop, started the first time a session registers a
heartbeat watch, and then permanent for the rest of the process.
Because _loop_heartbeat_task starts on every boot, it alone made
_scale_to_zero_has_live_background_work() return True forever on every
armed instance, regardless of the #84327 fix -- confirmed empirically
against the real method with the exact untagged-task shape this task has.
Two new regression tests spawn each task through its real production
entry point and assert the busy check returns False; both fail against
the unfixed code (missing method / real assertion failure).
Co-authored-by: pierrenode <298902573+pierrenode@users.noreply.github.com>
Co-authored-by: Ben Barclay <ben@nousresearch.com>
test_flat_key_wins_over_subblock asserted the OPPOSITE of its name (the
sub-block wins, matching _relay_slack_extra). Rename to what it proves.
Also note in _resolve_cron_surface_mode why its fallback differs from
_relay_slack_extra's all-or-nothing sub-dict: the flat key is the legacy
staging shape, and a flat knob applies to every fronted platform, gated
only by the per-platform D6 capability check.
RelayAdapter.supports_inchannel_continuable is a scalar adopted from the
PRIMARY identity's handshake descriptor, but one RelayAdapter fronts N
platforms and the connector advertises the bit per platform. Reading the
scalar for every logical platform both leaked a Slack-primary True onto
other fronted platforms (activating the flat surface their descriptor
never advertised) and suppressed a non-primary platform's advertised
True (forcing thread mode on capable Slack behind a Discord primary).
Add supports_inchannel_continuable_for_platform(platform): resolves the
platform's own negotiated descriptor via descriptor_for_platform (the
same Phase 1.5 seam max_message_length uses), scalar fallback only when
the per-platform descriptor is unavailable. The scheduler's D6 gate
prefers the query when the adapter provides it; native adapters keep
the class-attribute path byte-identically.
Tests: two-platform descriptor matrix (primary-True no-leak,
non-primary-True honored, unknown-platform scalar fallback).
Two gaps in the block-formatting hint stamping:
1. Wrong descriptor: _format_hints gated on self.descriptor — the PRIMARY
identity's scalar — while one RelayAdapter fronts N platforms. A
Slack-primary adapter stamped Slack hints onto known Discord chats; a
Discord-primary adapter suppressed hints for Slack chats whose own
negotiated descriptor advertised the bit. Resolve per destination:
send/edit use _descriptor_for_chat (the same seam max_message_length
already uses) plus the chat's logical platform for the config
sub-block; the knob lookup is now per-logical-platform
(platforms.relay.extra.<platform>.*) instead of hardwired to slack.
2. Missing lane: send_for_platform — the scheduled/persisted-home lane
(gateway/delivery.py), i.e. the CRON delivery path, the flagship
consumer of the in_channel brief — never stamped hints at all. Stamp
there too, resolving descriptor_for_platform(logical) off the
transport; the scalar descriptor is used only when it belongs to that
exact platform (fail closed).
Tests: Slack-primary/Discord-chat no-leak, Discord-primary/Slack-chat
still-stamps, send_for_platform stamps for capable platform and stays
clean for incapable — all against a two-platform negotiated-descriptor
transport. Existing single-platform suite unchanged and green.
The seed was decoupled from the mirror opt-in (in_channel is the
continuation surface regardless of attach_to_session), but the
thread-id-clearing gate above it still read mirror_this_target. With the
advertised default config (attach_to_session=false, cron.mirror_delivery
unset) and an origin carrying a real thread_id, the brief kept delivering
INTO the origin thread while the flat (thread_id=None) session got
seeded — brief and continuation surface in different places, so a plain
reply never saw it.
Flatten on the same gate as the seed: origin_target (with the existing
live_adapter_ready guard). Fan-out/broadcast targets are unaffected.
Test drives _deliver_result with a thread-carrying origin and default
knobs, asserting on the routed DeliveryTarget.thread_id — RED on the old
gate, GREEN now.
'hermes --version' (and -V) now prints the full version report — banner
version line with upstream SHA, install directory, authoritative install
method, Python and OpenAI SDK versions, and update status — making the
separate 'hermes version' subcommand redundant. The subcommand is removed.
- _startup_fast.print_fast_version_info() is now THE canonical version
printer: static lines print instantly from stdlib probes, then the
banner label, install-method resolver, and update check lazy-import
after the first line is on screen (each degrades gracefully).
- main.py _print_version_info() delegates to it (used by /version in the
CLI chat surface and the --version flag path); the old duplicate
implementation is deleted.
- hermes_cli/subcommands/version.py removed; parser wiring, subcommand
sets, console-engine extraction entry, and tests updated. Hermes
Console keeps a 'version' command wired to the shared printer.
- Termux fast paths now include update status too (previously
check_updates=False).
- Docs/i18n, CONTRIBUTING, SECURITY, and nix checks updated to
'hermes --version'.
build_session_key embeds the workspace segment (scope_id) in every Slack
dm/group/thread key, but both cron seed helpers built their SessionSource
without it: the seeded row keyed agent:main:slack:dm:<chat>:<thread> while
a real scoped reply keys agent:main:slack:dm:<team>:<chat>:<thread> — a
row no reply ever resolves to. DMs were rescued only incidentally by the
legacy-key claim-once migration; scoped channels/threads got continuation
amnesia, and identical channel ids in two workspaces could collide.
Capture HERMES_SESSION_SCOPE_ID into the cron origin (_origin_from_env —
the session-context var async_delegation already snapshots), add scope_id
to _seed_cron_thread_session/_seed_cron_channel_session, and pass the
origin's scope at all three seed call sites.
Tests: scoped dm-thread / channel-thread / flat-channel seed-vs-reply key
equality through the real build_session_key, plus a two-workspace
non-collision guard.
A GitHub-side HTTP 429 during 'hermes update' printed only
'Failed to fetch updates from origin.' — and the curl
'unable to access ... returned error: 429' shape even matched the
network-error branch, blaming the user's connection for a GitHub
outage.
- new _classify_fetch_failure(): 429/rate-limit -> 'GitHub is rate
limiting requests or having an outage — try again in 5 minutes';
5xx -> outage message with githubstatus.com; ordered BEFORE the
generic 'unable to access' network check
- both fetch-failure sites (update apply + --check) now share the
classifier via _print_fetch_failure(), and both always print the
first raw stderr line so the wire error stays diagnosable
- tests: classifier matrix + E2E against a live local HTTP server
returning 429 through real git
Fixes#89287
When the chosen/keyed backend fails a web_search or web_extract call
(bad key, upstream outage, 5xx, raised exception), that single call
retries on the keyless free-tier ring instead of erroring. The next
call attempts the chosen backend again — no sticky failover, no state.
Resolves the keyed half of #78984/#32159 (keyless half landed in the
ring PR).
- tools/web_tools.py: _rescue_eligible (keyed ring vendors + non-ring
backends eligible; keyless-mode calls excluded — they already walked
the ring), _rescue_search/_rescue_extract (search annotates
rescued_from + backend_error naming the original failure and the
retry-next-call semantics; extract rescues only whole-batch failures,
partial failures pass through untouched; rescue failure preserves the
ORIGINAL backend error with the rescue note appended)
- both dispatchers wrap the provider call: failure-results AND raised
exceptions rescue; ineligible paths re-raise unchanged
- web.keyless_rescue config key (default true; implicitly off when
keyless_fallback is off); docs updated
Live E2E: keyed Tavily with an invalid key 401'd and the call was
served by the real ring with the rescue annotation; a second call
re-attempted Tavily first (statelessness proven); whole-batch extract
rescue returned real page content. 13 new tests; 67 green across the
keyless suites.
Fresh installs with zero web credentials now rotate web_search/
web_extract across FIVE vendors' public free tiers — Exa, Parallel,
Tavily, Firecrawl, Keenable — instead of a 2-vendor 50/50 split, with
next-in-line ring failover on rate limits (multi-hop until a vendor
serves or the ring is exhausted; served_by marks the actual vendor).
- plugins/web/keenable/: new bundled provider (search via /v1/search,
fetch via /v1/fetch; keyed Bearer or keyless with the mandatory
X-Keenable-Title app header). Credit: integration proposed by
Ilya Gusev (Keenable) in #49758; Free/Paid picker rows included.
- keyless_mcp: tavily/firecrawl/keenable keyless search+extract
wrappers, _KEYLESS_RING + per-process round-robin cursor (seeded by
the random session id, advances per unpinned request), pinned-vendor
entry (pin = start there; rotation off), paid-pinned vendors excluded
from the ring entirely.
- Tavily/Firecrawl providers route keyless traffic through the ring;
both are now default-on ring members (no longer selection-gated).
- web_tools/registry: keenable in backend sets, auto-detect, availability
probes; _keyless_preference() delegates to the ring cursor.
- KEENABLE_API_KEY in OPTIONAL_ENV_VARS; docs updated (ring semantics).
Live E2E: all 10 vendorXcapability paths (5 search + 5 extract) served
real results keyless; rotation cycled all five vendors over 5 dispatch
calls; double-throttle failover walked exa->parallel->tavily.
Review follow-up on the salvaged #89444:
- Warn fires only from the conversation-loop pre-API site, reusing the
unconditionally computed request_pressure_tokens (zero marginal cost,
covers turn-start AND mid-turn growth) — drops the duplicate every-turn
estimate the turn-context block paid.
- Turn-context block now only RE-ARMS the dedup once the session is back
under the window, so warn -> /compress -> regrow warns again (the dedup
was previously never cleared with compression disabled).
- Char pre-check treats non-string (multimodal) content as over-gate —
len() of a part list defeated the 20k char floor (probe: 10 'chars' vs
~70k real tokens) — and compares against the window, not a flat 20k.
- Deletes the unreachable get_model_context_length fallback from both
sites (context_compressor always exists; its context_length property
hard-floors positive; the fallback would have been a synchronous
network probe mid-turn that also bypassed config overrides) and the
undeduped inline _emit_warning fallback (third copy of the message).
- Tests bind the PRODUCTION warn/clear methods (previously a verbatim
fake reimplementation left them uncovered) and add dedup, re-arm,
no-rearm-while-over, and multimodal-gate coverage.
When compression is explicitly disabled (compression.enabled: false), conversations can grow past the model's context window across hundreds of messages (e.g., 824 messages / 460K+ tokens in #89297). Serializing massive JSON payloads repeatedly under memory-constrained environments leads to swap thrashing (STAT=U) and unhandled provider errors.
Add a pre-flight uncompressed context overflow guardrail in build_turn_context and a deduped _warn_uncompressed_context_overflow method on AIAgent to alert users to run /compact or enable compression before unmanageable payloads freeze the process.
Two gaps behind the recurring 'hermes -w timed out after 30 seconds':
1. Rebase-merge leak: git cherry only catches patch-identical commits.
Salvage flows routinely change the diff (conflict resolution, follow-up
commits), so 12 of 22 'unpushed' trees on the incident box had MERGED
PRs and were preserved forever. The pruner now falls back to
'gh pr list --head <branch> --state merged' — authoritative, memoized
on (branch, head_sha) with True-only caching, fail-safe to preserve.
2. Creation timeout 30s -> 120s: the ~10k-file checkout measured 113s at
near-zero CPU under multi-agent disk contention vs 1.2s idle. 30s
killed legitimate creates and threw away completed work.
The startup pruner is deliberately conservative (unattended, pre-banner),
so real installs accumulate what it can never touch: trees preserved for
untracked-only scratch, and orphaned local branches beyond the two
auto-generated prefixes it deletes. A measured multi-agent box: 35 trees /
15GB / 244 local branches, 120 of them fully merged.
New attended surface (hermes_cli/worktree_gc.py + worktree_cmd.py):
- hermes worktree list — audit every tree: age, size, verdict, reason,
plus deletable-branch count
- hermes worktree prune [--dry-run|--trees-only|--branches-only]
- /worktree prune [--dry-run] — same engine in-session; never touches the
session's own active tree
- startup escalation: one WARNING when .worktrees/ exceeds 10 trees or
5GB, naming the reclaim commands (silence is how boxes hit 15GB)
Safety invariants (shared with the startup pruner via cli.py primitives):
tracked modifications and unique unpushed commits never deleted at any
age; live-locked trees untouched; branch deletion gated on worktree
removal success; untracked-only scratch ARCHIVED to
~/.hermes/archive/worktree-prune/ before its tree is reaped.
Branch GC is content-gated, not name-gated: any local branch fully merged
or git-cherry patch-equivalent upstream is safe to delete (rebase merges
rewrite SHAs, so --merged alone misses the dominant leak); unique-commit,
checked-out, protected, and stale-base (>50 ahead) branches are kept.
Classification is parallel (8 workers) — 244 branches audit in ~64s live.
git timeouts degrade to keep (returncode 124) instead of crashing the
audit — live-verified failure on a 746MB .git repo.
16 behavior-contract tests against real git fixtures; live dry-run on the
production repo: 12 trees reclaimable, 120 branches deletable, 0 false
positives among kept trees.