Review follow-up: the auth path called pool.try_refresh_current() before
the hinted rotation, so a stale current() pointer could force-refresh a
different, healthy entry — consuming its single-use refresh token, or
(for non-OAuth entries, where a forced refresh marks the entry exhausted
outright) killing it entirely before api_key_hint was ever consulted.
Use try_refresh_matching(api_key_hint=...) to resolve and refresh the
entry that supplied the failing key under the pool lock, falling back to
the previous behavior when no key is known.
Adds a regression test with current() deliberately pointed at the
healthy entry: on the old code the healthy entry is exhausted by the
forced refresh and the pool ends up fully offline; with the fix the
failing entry is exhausted and recovery rotates to the healthy one.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
recover_with_credential_pool identified "which credential failed" via
pool.current(), a shared mutable pointer that is advanced by every
select() (round-robin rotation, concurrent turns, and other processes
reloading the pool reset it to None). By the time recovery ran, it
routinely pointed at a different, healthy entry — mark_exhausted_and_rotate
then stamped the failing request's error message and reset time onto that
innocent entry. With round_robin and one hard-capped key this
deterministically exhausted the healthy key too and took the entire pool
offline ("no available entries") from a single rate-limited credential.
mark_exhausted_and_rotate already supports api_key_hint for exactly this
(the auxiliary-client path passes it); the main conversation-loop path
never did. Pass agent.api_key — kept in sync with the entry in use by
_swap_credential — as the hint on all four rotation call sites, and make
the "already exhausted → rotate immediately" pre-check look up the failing
entry by key with the same fallback to current().
Adds regression tests that fail on the old attribution logic: a fresh
pool (current() is None) failing on key B must mark entry B, never
entry A.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The hermes console entry point is hermes_cli.main:main, and main.py imports
dotenv (via env_loader) and yaml (via config) at module level. In the #57828
failure state — a failed lazy backend refresh wiping a core package's import
files while metadata survives — a normal launch crashed while importing
main.py, before _recover_from_interrupted_install() and the recovery markers
from PR #58004 could act.
- hermes_cli/_early_recovery.py: stdlib-only bootstrap repair invoked at the
very top of main.py, before any third-party import. Probes the fragile
core packages via real imports, force-reinstalls broken ones using the
pyproject.toml pins, shares main.py's single-flight recovery lock, and
never clears markers (the confirmed lifecycle stays with the full recovery
path in main.py).
- Probe/repair tables now have one canonical home in _early_recovery, reused
by main.py so the two layers cannot drift.
- Manual --force-reinstall fallback commands now print pinned specs via
_lazy_refresh_repair_specs() instead of bare package names.
- tests: entry-point lifecycle coverage proving a broken dotenv import
crashes main.py without repair and imports cleanly with it, a stdlib-only
import guard for _early_recovery, and unit coverage for marker gating,
lock single-flight, pinned specs, and marker preservation.
Keep .update-incomplete for full .[all] recovery only. Lazy refresh uses
.lazy-refresh-incomplete and clears only after confirmed import probes;
unavailable probes are indeterminate, not healthy (#58004 review).
Keep .update-incomplete across normal hermes.exe launches, heal via
package-only import probes first, and only clear the marker after repair
succeeds (#57828 / #58004 review).
The startup pruner only considered directories named hermes-* (the
hermes -w scratch trees), so salvage/review/port lanes created with raw
'git worktree add' accumulated forever — a real checkout reached 117
directories / 26 GB with trees dating back months. Two further leaks:
squash-merged branches' local commits stay unreachable from
refs/remotes/* forever, so the unpushed-commits guard preserved fully
merged scratch trees indefinitely; and preserved trees rotted silently
with no visibility.
- Pruner now covers every directory under .worktrees/ except kanban
task trees (t_<hex>, owned by 'hermes kanban gc'). Named (non
hermes-*) trees get a 3x timeline (72h soft / 9d hard) since they
were created deliberately.
- New _worktree_commits_all_merged_upstream(): git-cherry
patch-equivalence check against origin/HEAD|main|master, bounded at
20 commits ahead, fails safe toward preserve. Lets the pruner reap
trees whose every local-only commit already landed upstream via
squash-merge/cherry-pick.
- Dirty guard now applies at every tier (previously the 24-72h tier
skipped it — it only survived because the unpushed check usually
caught the same trees).
- Trees preserved for unpushed/dirty reasons older than 7 days are
listed in a single WARNING so in-flight work can't rot silently.
- tips.py text updated; 13 new behavior-contract tests.
- gateway/run.py: use _adapter_for_source(source) instead of the raw
adapters.get(source.platform) map so the compression-timeout warning
respects transport provenance, relay ingress, and multiplexed profiles
(matches every other user-facing send in the hygiene block).
- tests: add a lock-release verification regression — a fence-cancelled
hygiene compression must leave the per-session compression lock free so
the next attempt (manual /compress retry) acquires it and commits
normally.
Follow-up for the salvaged #35191: the mid-turn pre-API pressure emit in
conversation_loop.py and the idle-resume emit in turn_context.py were not
routed through automatic_compaction_status_message, so an engine with
emit_automatic_compaction_status=False still leaked those lines. Both now
resolve through the hook (phases "pre_api" and "idle") while keeping the
#69550 template constants as the default wording. Suppression also skips the
#69546 structured 'compacted' terminal edge for compress-phase events that
opened no visible phase; failure warnings (_emit_warning) remain never
suppressible, pinned by test.
Covers the follow-up hardening: continuation-marker and summary-as-user
tails keep the flagged standalone snapshot (zero-user provenance #69292
verified via _transcript_has_real_user_turn on the projected rows),
stale snapshot rows are refreshed in place, a previously merged snapshot
is stripped before re-injection, and an all-completed todo store injects
nothing (#26981).
After context compression, the preserved todo list was unconditionally
appended as a standalone user message. When the compressed transcript
already ends with a user message (common case), this creates consecutive
user/user turns — a role-alternation violation some providers reject.
Fix: fold the snapshot into the trailing user message (blank-line separated)
when one exists with plain-string content. Falls back to append when the
tail is non-user, empty, or has structured (list) content.
Rebased on current upstream/main.
Closes#53890
Regression test for the exact issue #61932 report: head + an 8-message
protected tail made exclusively of oversized tool pairs. Pre-fix,
compress_start >= compress_end made compress() a pure no-op and the
retry loop ended in 'Cannot compress further'; post-fix the Phase-1
pressure demotion reclaims the tail in one pass while preserving
tool_call/tool_result pairing.
After multiple in-place compactions, short tool-heavy sessions can leave
nearly every remaining message inside protect_last_n while those messages
are huge completed file/tool outputs. The middle compress window then
makes no material token progress and the turn dies with
"Cannot compress further" (#61932).
Cap the prune message floor at the same bound as tail-cut, and under
pressure demote bulky protected-tail tool bodies (keeping a short recent
floor) so preflight can reclaim headroom without wiping the active ask.
Follow-up for salvaged #58629: thread the previous flush baseline through
the idle-compaction caller in turn_context.py (the one caller the original
PR predates), and add regression coverage that every early-return path in
compress_context (breaker skip, no-progress, plus the completed-rotation
boundary) resets the per-attempt in-place outcome so a stale
_last_compaction_in_place from an earlier successful in-place compaction
can never baseline unflushed turns as persisted.
Relocates the #20424 wiring: the preflight region moved out of
run_agent.py into agent/turn_context.py (and through the
compression.max_attempts unification, #69315), so the contributor's
elif branch is reapplied at its current home as the else arm of the
threshold dispatch chain.
Integration contracts:
- Byte-identical default: the built-in ContextCompressor inherits
ContextEngine.should_compress_preflight() -> False, so the default
path performs no compression and touches no turn bookkeeping
(pinned by test_builtin_compressor_default_sub_threshold_path_unchanged).
- Attempt-cap: the engine gets exactly ONE compress() pass per turn,
mutually exclusive with the cap-bounded threshold multi-pass loop,
so turn-start passes stay within the resolved
compression.max_attempts budget in every case.
- No-op blocking (#64382 / 377244f7c): an engine pass that no-ops
(_compress_context returns the input list object) neither sets nor
clears preflight_compression_blocked and does not re-baseline the
flush history — a sub-threshold maintenance no-op proves nothing
about over-threshold compressibility.
- Engine exceptions are swallowed at debug level; cooldown/defer/
codex-native gates run before the hook is ever consulted.
Salvaged from #20424 by @Beandon13. Fixes#20316.
Context engines that override ``should_compress_preflight()`` (e.g. the
hermes-lcm plugin's incremental leaf-chunk compaction) never had their
hook fired by ``run_conversation`` because the preflight block exited
early once the hardcoded ``>= threshold_tokens`` check failed. As a
result, ``LCM_DEFERRED_MAINTENANCE_ENABLED=1`` and friends were inert
and accumulated raw_backlog debt indefinitely.
Add an ``elif`` branch that delegates to the engine's preflight hook
when the legacy threshold check does not fire. The default
``ContextEngine.should_compress_preflight()`` returns ``False`` so the
built-in ``ContextCompressor`` is unaffected; engines opting in get a
chance to ingest messages and request a single ``compress()`` pass for
deferred maintenance. Exceptions are swallowed at debug level so a
buggy engine cannot break an otherwise-healthy turn.
Closes#20316
- tests/cli/test_compress_type_ahead.py: end-to-end proof of the PR #68284
docstring claim — a prompt queued into _pending_input while /compress runs
survives compaction untouched and is the next item process_loop drains,
i.e. it is processed against the compacted history. Plus a structural
guard that handle_enter never gates on _command_running /
_command_blocks_input (read-only enforcement belongs solely to the
TextArea Condition; a busy-gate in handle_enter would silently drop
type-ahead input).
- tests/test_cli_manual_compress.py: update the one remaining _busy_command
stub (added on main after the PR branched) to accept the new
blocks_input kwarg.
- contributors/emails: map lucas@policastromd.com -> enzo2.
Follow-up to the salvaged #35533 fix: instead of parsing 'here [N]' only in
_mirror_slash_side_effects, parse it inside _compress_session_history — the
single helper all three manual-compress routes converge on (session.compress
RPC, command.dispatch /compress|/compact, slash-exec mirror). Every route now
honors the boundary-aware forms (here [N], up to here, --keep N) with the
same head/tail split + rejoin as cli.py and gateway/slash_commands.py, and
keeps the choke point's existing guards (lock-free LLM call, history_version
race check, deferred context-engine notification).
Tests: choke-point unit tests for 'here N', degenerate-split fallback, and
focus-topic passthrough, plus endpoint-level tests on each of the three
routes.
_mirror_slash_side_effects() passed the raw argument after /compress
directly as focus_topic to _compress_session_history. So /compress here
3 silently used "here 3" as a summary focus topic and did a full
compress instead of preserving the last 3 exchanges verbatim.
Fix: call parse_partial_compress_args() on the argument first. When it
detects a boundary-aware form (here, here N, up to here, --keep N),
split the history into head/tail using split_history_for_partial_compress,
compress only the head via agent._compress_context, and rejoin with
rejoin_compressed_head_and_tail — exactly mirroring cli.py and
gateway/run.py's /compress here implementation (PR #35252).
Non-boundary forms (plain /compress, /compress <focus>) fall through
to _compress_session_history unchanged.
Photon (managed iMessage) shipped without a _PLATFORM_DEFAULTS entry, so it
inherited the noisy global ('all') display defaults and narrated tool
progress / heartbeats / busy-ack detail into a permanent-message iMessage
thread. Register it as TIER_LOW alongside BlueBubbles and Signal, plus a
regression test guarding the tier.
Salvaged from #50511 (Photon tier piece only).
The PR's inline toolset resolution (checking 'memory' in cli_toolsets
list) produced wrong results for composite toolsets like 'hermes-cli'
which expand to include the memory tool. Replace with the canonical
_get_platform_tools() from tools_config.py which correctly handles
composite toolsets and all edge cases.
Update tests to mock _get_platform_tools instead of raw config.
Invalid-tool exhaustion and truncated-tool early returns skipped finalize_turn, leaving role=tool transcripts that become tool→user on the next turn for strict providers. Call close_interrupted_tool_sequence before persist on those paths (same as interrupt aborts).
Assert the version matrix, fresh-DB DELETE fallback, already-WAL left
alone (no checkpoint/DELETE), fixed-SQLite WAL path, and warn-only
doctor output for vulnerable builds (#69784).
Two fixes for desktop hands-free voice:
- The live speech session bound to the first assistant bubble with text, so
a tool-calling turn spoke only the opening narration and silently dropped
every later interim AND the final answer. The conversation selector now
aggregates all unspoken assistant bubbles in order (turn-scoped speech);
auto-speak keeps its latest-reply-only behavior.
- The speak-stream WS producer blocked forever on the text queue, so a
narration line with no trailing whitespace ("Let me check.") sat in the
sentence chunker until end-of-turn — spoken long after the tool finished,
with the UI stuck on "Preparing audio…". Mirror the CLI speaker's idle
flush: sentence-terminated buffers flush after 0.5s of producer silence,
anything else after ~2s; open <think> blocks are never flushed.
The usage gauge is Nous subscription-cap-only (used_fraction requires a cap;
non-Nous providers emit no headers, so no notice fires). A bare percentage
implied a universal unit that doesn't exist, so report the absolute dollars
used of the cap instead: used = cap - remaining, from micros (money-safe),
clamped to [0, cap]. Still a snapshot at band-crossing (re-emits on band
change, not every turn) to keep the single escalating line and stay quiet on
append-only surfaces (messaging pushes one message per crossing).
cua_driver_update_check now short-circuits to None when no driver
resolves; CI has none installed, so pin resolve_cua_driver_cmd in the
update-check and env-sanitization tests.
Cut steady-state Computer Use latency without changing default behavior
or waiting on cua-driver:
- Cap screenshots via set_config(max_image_dimension) on session start
(config: computer_use.max_image_dimension, default 1456)
- Cache aux-vision routing per (provider, model) so captures skip
repeated load_config()
- Add computer_use.capture_after_mode (default som) so users can opt
follow-ups down to ax (elements only) for speed
Auto-detect now disables the cursor overlay on darwin as well as
headless/WSL2 Linux. After start_session, also call
set_agent_cursor_enabled(false) when the policy is on so older drivers
without --no-overlay still tear the overlay down.
Co-authored-by: David Metcalfe <80915+DavidMetcalfe@users.noreply.github.com>
Print gh-image stdout and stderr on attachment failures, then replace the
pending inline-evidence marker in the PR review comment with an escaped
failure notice before preserving the failing workflow result.
The hermes-sweeper review #4701565902 (2026-07-15) flagged two
consistency issues in `_cua_driver_supports_no_overlay` and one
additive-config concern:
1. `cua-backend.py:260` — the `cua-driver --help` support probe
inherited the full parent environment. cua-driver is a third-party
binary; every other spawn site in this file (manifest probe at
`:214`, MCP spawn at `:697`, install probe at `:997`) uses
`_sanitize_subprocess_env(cua_driver_child_env())`. The `--help`
probe should match. This was a low-impact leak (only help output
exits), but inconsistency is the wrong default for a third-party
subprocess.
2. `cua_backend.py:238` — when the manifest returned a `command`
different from the input `driver_cmd` parameter (e.g. a relocated
executable at `/opt/relocated/cua-driver` while the system binary
is at `/usr/bin/cua-driver`), the support probe ran against
`_CUA_DRIVER_CMD` (the default) instead of the manifest-discovered
`command`. Two failure modes:
- The wrapper binary supports `--no-overlay` but the system binary
doesn't → probe returns False → overlay kept despite capability.
- The system binary supports `--no-overlay` but the wrapper doesn't
→ probe returns True → MCP spawn crashes on the unknown flag.
3. The original commit bumped `_config_version` 31→32 for an additive
default (`computer_use.no_overlay: None`). AGENTS.md specifies that
additive defaults in existing sections are handled by deep merge
and should NOT trigger a version bump. After cherry-picking onto
current `origin/main` (which is already at 33), the bump is
effectively dropped — resolved to main's 33.
Changes:
- Add `env=_sanitize_subprocess_env(cua_driver_child_env())` to the
`--help` subprocess (with the same import + rationale comment as
the manifest probe).
- Pass `driver_cmd=command` (or `driver_cmd=driver_cmd` for the
fallback path) into `_mcp_args_with_overlay_flag`, so the support
probe runs against the binary that will actually be launched.
Tests (3 new):
- `test_help_probe_passes_sanitized_env` — verifies `subprocess.run`
is called with an `env=` kwarg.
- `test_manifest_command_drives_support_probe` — verifies the probe
runs against the manifest command when it differs from the input
driver_cmd.
- `test_fallback_uses_input_driver_cmd_for_support_probe` — verifies
the fallback path (no command in manifest) uses the input
driver_cmd.
- `test_probe_distinguishes_support_between_binaries` — sanity check
that the lru_cache key on `driver_cmd` prevents cross-binary
cache leakage.
File-revert negative test confirmed all three of the new
"manifest/probe" tests are load-bearing: with the pre-fix code, they
fail (probe runs against the default binary instead of the resolved
one); with the fix, they pass. 20/20 tests in
`tests/computer_use/test_cua_no_overlay.py` green.
`TestMcpInvocationResolution` (8/8) still green.
Refs: sweeper review #4701565902
Address review feedback from cross-vendor review (Flash + GPT-OSS):
1. Auto-detect now checks for headless Linux (no DISPLAY), WSL2
(/proc/version contains 'microsoft'), instead of all Linux.
Desktop Linux with a compositor keeps the overlay.
2. Add _cua_driver_supports_no_overlay() that probes cua-driver --help
to check if the flag is supported. Older drivers (< 0.6.x) reject
unknown flags, so passing --no-overlay would crash the MCP spawn.
3. Update tests to cover headless vs desktop Linux, WSL2 detection,
version probe, and the unsupported-driver fallback path.
- Add autouse fixture to TestMcpInvocationResolution to disable
--no-overlay flag so existing tests assert baseline args
- Make test_config_load_failure_fails_safe and test_missing_section_enables
platform-aware (Linux auto-detect returns True, macOS/Windows False)
cua-driver's cursor overlay rendering loop can consume CPU indefinitely
when idle (#28152, #47032). On Linux/WSL2, the overlay serves no visual
purpose and the rendering path is the primary source of idle CPU usage.
Add computer_use.no_overlay config option (default: auto-detect) that
passes --no-overlay to cua-driver when enabled. Auto-detection disables
the overlay on Linux (covers WSL2, headless, containers) where it has no
benefit, and keeps it enabled on macOS/Windows where it is visually
useful.
Refs: #28152, #47032
Keep thin-PATH resolution while asserting the absolute binary on the
already-installed version check. Map contributor emails for the salvage.
Co-authored-by: Tianqing Yun <yuntianqing@yahoo.com>
Co-authored-by: Trevor Gordon <trevorbgordon@gmail.com>
Co-authored-by: Adrian Soto Mora <adrian.soto6@gmail.com>
