Main's 243c9182b1/a16fd675df/7142dc4580 added load_config_readonly
sibling stubs across 38 files; our pruned versions of 11 of those files
kept only the load_config stubs. Re-applied the pairing at every
surviving site (26 patch()/setattr sites) — same return_value/
side_effect as the adjacent load_config stub. 494 tests green across
the 11 files.
The #71637 prune fix cut one stage of -w startup, but the base-ref
resolution right after it still ran an uncapped-in-practice
'git fetch origin main' (timeout=30) on every launch — and on a flaky
smart-HTTP connection that fetch intermittently stalled to the full 30s,
then cascaded into step 2's SECOND 30s fetch. Measured: back-to-back
fetches of 0.9s, 1.0s, 63.5s on the same box with healthy TLS (~185ms).
_resolve_worktree_base now:
- skips the fetch entirely when FETCH_HEAD is < 5 min old and the
tracking ref exists (repeat launches pay zero network cost)
- caps the fetch at 5s and falls back to the locally-known tracking
ref (labelled 'cached') on timeout/failure instead of cascading into
a second fetch — genuine staleness stays backstopped by the pre-push
stale-base gate
- caps 'git remote show origin' the same way
Worst case drops ~60s -> ~5s; warm path is ~0.02s (was up to 30.8s).
sync_base=False and the offline HEAD fallback are unchanged.
Final pass: per_model_threshold_init_ordering, memory_provider_init,
plugin_context_engine_init, api_max_retries_config,
invalid_context_length_warning, tool_call_guardrail_runtime all stub
agent_init config reads that now resolve through load_config_readonly().
Verified by running the complete 59-file suspect list (every test file
that stubs load_config in string or attribute form and intersects the
swapped modules): 3,063 tests, 0 failed.
Second stub shape the first sweep missed: monkeypatch.setattr(config_mod,
"load_config", ...) — attribute-form instead of string-form. Four files
(compression_max_attempts, preflight_compression_cap_e2e,
codex_gpt55_autoraise_notice, proactive_prune_config) stub agent_init
config reads that now go through load_config_readonly(). Swept the whole
tree for the attribute form; remaining hits stub modules that still use
the mutable loader. 1,210 tests green across the 19-file re-sweep.
The readonly swaps mean agent/ modules now read config through
load_config_readonly(); tests that stubbed only load_config stopped
intercepting those reads. Added sibling readonly stubs (same
return_value/side_effect/lambda) at every affected site across 11 test
files, found via a tree-wide sweep of load_config stubs cross-referenced
against the swapped modules. 3,915 tests green across the 38-file
sibling sweep.
agent_init's config reads now go through load_config_readonly(); 8
tests that stubbed only load_config stopped intercepting them. Each of
the 10 patch sites gains a sibling readonly patch with the same
return_value. 461/461 file-local tests green.
Pins the audit findings: normalizer never mutates its input
(api_key_env + camelCase forms), providers-dict round-trips leave a
cached config byte-identical, and the normalized models mapping does
not alias the caller's dict.
Salvaged from #56085 (@Stoltemberg), rebased onto current main: sites
main had already converted (credential_pool, auxiliary_client MoA
paths, model_metadata, moa_loop, agent_runtime_helpers) resolve to
main's versions; the remaining ~29 read-only sites across 16 agent/
files swap to the no-deepcopy readonly loader (~135us saved per call).
Full per-site mutation audit performed (every enclosing function read,
escapes traced): 23 SAFE, 5 ESCAPES with read-only consumers, 1 UNSAFE
path (init_agent -> get_compatible_custom_providers -> normalizer
in-place alias writes) fixed by the preceding no-mutate commits, which
make the normalizer copy-safe for ALL callers.
Companion to the no-mutate fix: normalized['models'] retained a
reference to the caller's (possibly cached) models dict, and the
normalized entry escapes into long-lived runtime state
(agent._custom_providers). Shallow-copy so runtime writes can never
reach the shared config cache.
The normalizer writes alias keys into the entry it is given
(entry['key_env'] = entry['api_key_env'] and entry[snake] =
entry[camel]) while building its normalized copy. Two of its three
callers — get_compatible_custom_providers and
providers_dict_to_custom_providers — pass live sub-dicts straight
from load_config_readonly()'s shared cache (only
_custom_provider_entry_to_provider_config defends with dict(entry)).
A config written with the documented camelCase / api_key_env aliases
therefore gets its cached copy polluted with injected duplicate keys,
violating the cache's explicit no-mutation contract; every later
load_config() deepcopy inherits the duplicates, and any
save_config(load_config()) flow (setup wizard, dashboard writes,
model persist) writes them back to config.yaml. The aux-client TLS
resolution runs this on every auxiliary client build, so the
mutation also happens unlocked on worker threads against a shared
object.
Shallow-copy the entry up front; the function's return value is a
separately-built dict, so behavior is otherwise unchanged.
Buzz Desktop v0.5.1 now renders Hermes' ACP model menu in agent runtime
settings. Add a short note under the Buzz Desktop host section explaining
where the list comes from (the shared authenticated-provider inventory),
the provider:model / custom:<name>:<model> ID shapes, and that a pick is
session-scoped rather than a Hermes-wide default change.
The 1.0s sleep between sequential tool calls has been present since the
initial commit with no documented rationale. It sleeps between local
tool executions — the next LLM request goes out only after the whole
batch — so it rate-limits nothing, and the parallel read-only path
already runs with no delay. Every multi-tool turn pays (N-1) seconds
of dead time. Remove the sleep, the internal tool_delay plumbing, and
dead test assignments. AIAgent.__init__ keeps tool_delay as a
deprecated no-op keyword for one release so existing programmatic
callers construct cleanly; passing it emits a DeprecationWarning.
- Remove tests/-shadowing sys.path.insert(dirname/'..') from 11 test files:
it prepended the tests/ dir itself to sys.path, so 'import agent' /
'import hermes_cli' resolved to the test packages and collection died
with ModuleNotFoundError depending on import order (2 files failed in
every full-suite run; 9 more were latent).
- Patch call_llm in 5 context-compressor tests that called compress()
unmocked: each burned ~50s attempting live LLM traffic through the
relay before falling back (572s file — the slowest in the suite, and
flaky under the 300s per-file timeout). File now runs in ~5s.
- agent/redact.py: fix two catastrophically-backtracking regexes hit by
the compressor's redaction pass on large payloads —
_STRICT_URL_USERINFO_RE anchors on the mandatory '//' (optional-scheme
prefix backtracked O(n^2): ~55s on a 320KB payload, now sub-ms;
output-equivalence fuzz-verified on 20k random strings), and the
_CFG_DOTTED_RE/_CFG_ANCHORED_RE subs gain an exact linear keyword
pre-gate so secret-free text skips the quadratic pattern entirely.
- tests/gateway/test_feishu.py: version-guard the extra_ua_tags SDK
signature check; the repo pins lark-oapi==1.6.8 but stale local
installs (1.5.3) fail the assertion — skip below the pin.
- tests/tools/test_managed_browserbase_and_modal.py: stub
agent.redact + agent.credential_persistence in the fake agent package
(empty __path__ blocks all real agent.* imports added since the fake
was written).
- tests/gateway/test_startup_restart_race.py: raise wait_for timeouts
2s -> 30s; 2s wall-clock on a loaded 40-worker box flaked in the
baseline run (passes instantly when the box is quiet).
Follow-up to the salvaged #62026 ownership fix, folding in #72054's
CancelledError rule by @adurham: start() already cancels/reaps its own
run task when the caller's connect timeout cancels start() itself, so
_connect_server() must propagate cancellation without awaiting a
redundant shutdown() inside a cancelled context. Non-cancellation
failures on the unclaimed (standalone probe) path still reap the parked
task, now with the reap failure logged instead of raising over the real
error.
Also maps mrz@mrzlab630.pw for the attribution check.
Co-authored-by: Adam Durham <amdnative@gmail.com>
_stop_mcp_loop() stopped and closed the background loop without reaping
the tasks still on it. A task left suspended is resumed later by the GC,
whose finalizer drives its cleanup against the now-closed loop:
Exception ignored in: <coroutine object MCPServerTask.run ...>
File "tools/mcp_tool.py", line 2947, in run
parked = await self._wait_for_reconnect_or_shutdown(
File "tools/mcp_tool.py", line 2161, in _wait_for_reconnect_or_shutdown
t.cancel()
RuntimeError: Event loop is closed
shutdown_mcp_servers() only reaps servers held in _servers, so a server
that parked after exhausting its initial-connect budget — never inserted
there, because start() raises _error before the caller registers it — has
no owner to signal it and stays suspended until the loop is gone.
Drain the loop the way asyncio.run() does: cancel the remaining tasks and
gather them while the loop is still open, so each runs its own finally.
Cancel alone is not enough — Task.cancel() only schedules the throw.
This resolves the reported traceback, but not the ownership bug that
strands the task in the first place; that needs a follow-up. Deliberately
not using "Fixes" so #60197 stays open for it.
Addresses #60197
Addresses #66113
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Second, deeper pass over tools/gateway/hermes_cli plus first pass over
the trees wave 1 missed (acp, acp_adapter, skills, computer_use, docker,
dashboard, conformance, monitoring, secret_sources, hermes_state,
providers). Same rubric as wave 1 (AGENTS.md test policy); security,
alternation/caching invariants, issue-number regressions, and E2E kept.
Real test-quality fixes found and rooted out along the way:
- tests/tools/test_command_guards.py made real auxiliary-LLM HTTPS calls
(DEFAULT_CONFIG smart-approval leaked in) — pinned approval
mode=manual via autouse fixture: 17.4s → 0.4s.
- test_model_switch_custom_providers.py / test_user_providers_model_switch.py
silently probed live provider catalogs (~2s/test) — stubbed
cached_provider_model_ids/provider_model_ids/fetch_api_models.
- test_telegram_noise_filter.py: 15-platform copy-paste matrix over
shared gateway.run logic → 3 representative platforms (55s → 3.9s).
- test_gateway_shutdown.py: stop()'s 5s interrupt-deadline loop spun on
MagicMock agents — interrupt.side_effect now clears _running_agents
(22s → 1.0s).
- test_gateway_inactivity_timeout.py poll-harness timings shrunk 3-5x
(24s → 1.1s); test_mcp_stability.py backoff/SIGTERM-grace sleeps
patched (15.4s → 2.5s); test_async_delegation.py negative-drain wait
5s → 0.5s.
- test_telegram_init_deadline.py: loop-block margin restored to 1.0s
with rationale comment — the watchdog-dump assertion needs the loop
blocked well past deadline+grace under parallel load (flaked once in
the 40-worker verification run at a 0.2s margin).
Verification: full hermetic suite via scripts/run_tests.sh —
2,438 files, 21,718 tests passed, 0 failed, 293.9s wall.
Suite totals vs original baseline: 46,820 → 19,757 test functions
(−57.8%), wall 583.5s → 293.9s (−50%), subprocess CPU 13,564s → 11,623s.
Resolves the PR's conflict with main (2252 commits). Two conflicts, both
"each side added an independent block in the same place" — kept both:
- gateway/run.py — the housekeeping loop. This branch adds the Skill Sync
pulls inside the CURATOR_EVERY branch (12-space indent); main adds a
stale-session auto-archive as a sibling `if` at loop level (8-space).
Different scopes, so the naive union would have mis-nested the archive
block into the curator branch; kept each at its own indent level.
- tools/skill_manager_tool.py — the _edit_skill result dict. This branch
appends the org auto-propose note; main appends
_add_description_prompt_preview(). Independent, order-insensitive.
No behaviour dropped from either side.
Verified: 3552 passed / 0 failed across 63 suites (scope regenerated to
include main's new maybe_auto_archive / _add_description_prompt_preview
consumers) via scripts/run_tests.sh. `hermes sync` and `hermes sync status`
still work against a live token, resolving the production plane default.
The Pyright Optional-parameter warnings in skill_manager_tool.py are
pre-existing on main (`content: str = None` etc.), not introduced here.
The voice-interruption fix (5081551f0) covered the Python surfaces (CLI +
TUI gateway) but the desktop app has its own mic path in voice-barge-in.ts,
which still had both bugs on Windows:
- HALF-DUPLEX GAP: the barge monitor only opened when TTS playback started;
during LLM generation no mic was listening at all.
- PLAYBACK DEAFNESS: the monitor calibrated its noise floor while the
speakers were already playing TTS, baking bleed into the floor. On
Windows, Chromium echoCancellation does not reliably cancel same-app
output (measured live: quiet floor ~35-50 RMS vs playback bleed
600-1700 RMS), making the trigger unreachable.
Changes mirror tools/voice_mode.full_duplex_listen:
- voice-barge-in.ts: phase-aware full-duplex monitor — quiet-only
calibration (floor held through playback, never recalibrated against
bleed), playback min-trigger clamp + ceiling, 500ms grace on playback
onset only, windowed-majority detection.
- use-voice-conversation.ts: monitor arms at turn submit and spans
generation + playback (ensureBargeMonitor, idempotent). Mid-generation
speech fires the new onInterrupt callback; spoken stop-word during a
barge ends the conversation; submit waits for the interrupt to settle.
- use-composer-voice.ts + composer/index.tsx: plumb onInterrupt: haltRun
(same seam as the Stop button).
Tests: use-voice-conversation.test.tsx (6 tests) covering monitor lifecycle
across generation + playback, mid-generation interrupt, stop-word handling.
npm run typecheck + eslint clean; remaining desktop vitest failures
reproduce on a clean tree (pre-existing Windows env failures).
Salvaged from #31671 (@Shizoqua). The config-cache half of that PR was
superseded on main (9b8b054c2d gave _load_config_safe a readonly path),
but this cron-scanner half is still live: _strip_cron_safe_constructs
used re.search + a single str.replace, which only scrubbed occurrences
IDENTICAL to the first match. A cron job loading several GitHub skills
carries heterogeneous auth-header curl forms (-H vs --header, quoting,
token var names) — every non-identical block tripped the
exfil_curl_auth_header detector on every tick, blocking legitimate
GitHub cron jobs.
Now re.sub scrubs every occurrence; the trailing [^\n]* consumes the
URL path so no dangling fragment remains. Sabotage-verified: the old
implementation false-blocks the heterogeneous two-skill prompt the new
regression test pins; exfil to a non-GitHub host is still blocked.
79/79 cron tool tests green.
Follow-up to the #68246 salvage. The backend permission-mode resolution
only checked the DB session_id the tool path passes, but gateway /yolo
keys approval bypass off the gateway session_key (contextvar). Consult
both namespaces so /yolo works on messaging platforms, not just CLI/TUI.
Adds a regression test driving the real approval contextvar + yolo
toggle path E2E.
- Remove tests/-shadowing sys.path.insert(dirname/'..') from 11 test files:
it prepended the tests/ dir itself to sys.path, so 'import agent' /
'import hermes_cli' resolved to the test packages and collection died
with ModuleNotFoundError depending on import order (2 files failed in
every full-suite run; 9 more were latent).
- Patch call_llm in 5 context-compressor tests that called compress()
unmocked: each burned ~50s attempting live LLM traffic through the
relay before falling back (572s file — the slowest in the suite, and
flaky under the 300s per-file timeout). File now runs in ~5s.
- agent/redact.py: fix two catastrophically-backtracking regexes hit by
the compressor's redaction pass on large payloads —
_STRICT_URL_USERINFO_RE anchors on the mandatory '//' (optional-scheme
prefix backtracked O(n^2): ~55s on a 320KB payload, now sub-ms;
output-equivalence fuzz-verified on 20k random strings), and the
_CFG_DOTTED_RE/_CFG_ANCHORED_RE subs gain an exact linear keyword
pre-gate so secret-free text skips the quadratic pattern entirely.
- tests/gateway/test_feishu.py: version-guard the extra_ua_tags SDK
signature check; the repo pins lark-oapi==1.6.8 but stale local
installs (1.5.3) fail the assertion — skip below the pin.
- tests/tools/test_managed_browserbase_and_modal.py: stub
agent.redact + agent.credential_persistence in the fake agent package
(empty __path__ blocks all real agent.* imports added since the fake
was written).
- tests/gateway/test_startup_restart_race.py: raise wait_for timeouts
2s -> 30s; 2s wall-clock on a loaded 40-worker box flaked in the
baseline run (passes instantly when the box is quiet).
GatewayRunner carried ~19 separate Dict[str, ...] attributes keyed by
session_key, each with an ad-hoc lifecycle. They now live in one
`self._sessions: dict[str, SessionState]` (gateway/session_state.py) with
three lifecycle scopes and a `_session_state(key)` get-or-create accessor.
Mechanical refactor: same state, same semantics, new container.
Migration table (dict -> old decl line -> current clear path -> new home):
| legacy dict | decl | cleared by (before) | SessionState field |
|------------------------------------------|-------|--------------------------------------------------|----------------------------------------|
| _running_agents | 3505 | _release_running_agent_state; stop() .clear() | turn.agent |
| _running_agents_ts | 3506 | same | turn.started_ts |
| _active_session_leases | 3507 | same (+ lease.release()) | turn.lease |
| _busy_ack_ts | 3539 | same | turn.busy_ack_ts |
| _turn_lease_tokens ((key, gen)-keyed) | 3518 | _release_turn_lease (generation-guarded) | turn.lease_token + turn.lease_generation |
| _session_model_overrides | 3585 | _CONVERSATION_SCOPED_STATE funnel | conversation.model_override |
| _pending_one_turn_model_restores | 3586 | funnel; one-shot pop in turn finally | conversation.one_turn_restore |
| _session_reasoning_overrides | 3589 | funnel; lazy-init dict swap ~5692 (RACE) | conversation.reasoning_override |
| _session_service_tier_overrides | 3592 | funnel; lazy-init dict swap ~5738 (RACE) | conversation.service_tier_override |
| _last_resolved_model ("*" = process-wide)| 3527 | funnel | conversation.last_resolved_model |
| _queued_events | 3537 | funnel; lazy-init dict swap ~5204 (RACE) | conversation.queued_events |
| _pending_turn_sidecar_notes | 3597 | funnel; lazy-init dict swap ~19832 (RACE) | conversation.sidecar_notes |
| _session_ephemeral_pin | 3601 | agent-cache evict pop; lazy swap ~19885 (RACE) | conversation.ephemeral_pin |
| _session_vc_last | 3604 | agent-cache evict pop; lazy swap ~19864 (RACE) | conversation.vc_last |
| _pending_approvals | 3611 | boundary security funnel; stop() .clear() | persistent.approvals |
| _update_prompt_pending | 3623 | security funnel; update watcher pops | persistent.update_prompt_pending |
| _pending_native_image_paths_by_session | 3538 | one-shot consume; lazy swap ~12944 (RACE) | persistent.native_image_paths |
| _pending_messages (runner-level, str) | 3519 | _interrupt_and_clear_session pop; stop() flush | persistent.pending_command_text |
| _session_run_generation | 3540 | NEVER (monotonic, #28686) | persistent.run_generation (never reset)|
Races eliminated: every `self._X = {}` lazy-init/reset replaced the WHOLE
dict, so a writer on session A racing a lazy init triggered by session B
could lose its entry. All six such sites (_session_reasoning_overrides
~5692, _session_service_tier_overrides ~5738, _queued_events ~5204,
_pending_turn_sidecar_notes ~19832, _session_ephemeral_pin ~19885,
_session_vc_last ~19864, _pending_native_image_paths_by_session ~12944,
_turn_lease_tokens ~13644) are now per-session field writes on an existing
SessionState; a reset can no longer cross sessions structurally.
Registry successors:
- _release_running_agent_state -> state.turn.clear() (one structured reset
instead of the drifting pop-list; still pops the slot lease and calls
lease.release() first; still generation-guarded).
- _CONVERSATION_SCOPED_STATE funnel -> state.conversation.clear(); the
tuple is retained for legacy plain-dict stores not yet folded in
(_pending_model_notes) and for the public test contract.
- _turn_lease_tokens' (key, generation) tuple key -> lease_token +
lease_generation fields; release/rebind only match when the generation is
current, preserving the #28686/#64934 ownership check.
- _session_run_generation stays monotonic on persistent.run_generation and
is never cleared (conversation boundaries and turn releases don't touch it).
Compatibility adapters: tests (and a few mixin call sites) access the old
dict names directly (137 direct assignments to _running_agents alone), so
each legacy name is kept as a thin @property returning a live MutableMapping
view over the corresponding SessionState field (legacy_dict_property /
legacy_lease_token_property in session_state.py). Setter accepts a plain
dict (the `runner._X = {...}` test pattern); views support ==, in, len,
.get/.pop/.clear. The shutdown path in _stop_impl deliberately keeps
duck-typed legacy-attribute access because test fakes borrow it with plain
dicts.
Name collision noted (NOT touched, out of scope): gateway/platforms/base.py
has its own _pending_messages Dict[str, MessageEvent] (adapter-level slot);
the runner-level Dict[str, str] of the same name is what moved to
persistent.pending_command_text.
Entry leaks preserved (follow-up, no new eviction in this PR): SessionState
entries in self._sessions are never evicted, matching the old dicts — e.g.
_last_resolved_model, _session_run_generation, _session_vc_last entries for
dead sessions leaked before and their fields still occupy a SessionState now.
Verification: 123 tests/gateway files referencing the old names +
_release_running_agent_state + _CONVERSATION_SCOPED_STATE all pass (sole
failure test_feishu.py::test_websocket_sdk_accepts_channel_ua_tag is
pre-existing, stash-verified); 8 non-gateway test files touching the names
pass (266 tests); `import gateway.run` subprocess smoke OK; ruff clean;
post-migration grep shows zero non-comment `self._<oldname>` references in
run.py outside the property adapters and the duck-typed shutdown block.
The earlier scrub covered adapter.py; nine internal QA-campaign tracker IDs
remained in the two new test files, including a module docstring and an
assertion message. They mean nothing to a future reader — describe the
behavior instead. Comments only, no assertion changes.
Both relay Slack knobs read their value through bool(), while the native
adapter they mirror uses str(raw).strip().lower() in {"1","true","yes","on"}.
A YAML-quoted string diverges:
dm_top_level_threads_as_sessions: "false" → relay True, native False
Non-empty strings are truthy, so the escape hatch is silently ignored in
exactly the shape an operator writes to switch it OFF. reply_in_thread has the
same defect and gates reply placement, session keying and run.py's progress
resolver, so one quoted "false" misfires three ways.
Route both through a shared _coerce_flag mirroring native's predicate. Real
booleans pass through untouched; None falls back to the default. Contract §8
documents the accepted spellings.
Tests: both knobs parametrized over the true/false spellings native accepts,
plus the absent-key default.
The DM thread-anchor contract was resolved only in send(). _send_media() —
backing send_image, send_image_file, send_voice, send_video and send_document
— passed reply_to straight to the frame and never touched metadata, so
attachments egressing through the same connector-side Slack sender got both
failure shapes this branch set out to remove:
flat mode → reply_to survives, the image threads UNDER the user's DM
message (the original reported symptom)
thread mode → no metadata.thread_id, and threadTs() never reads reply_to,
so the image lands in the home channel instead of the
per-message thread
Both are reachable: gateway/run.py delivers agent artifacts through
send_voice/send_document.
Extract the three steps that must always happen together (mode gate, mirrored
reply_to_message_id strip, metadata promotion) into
_apply_slack_thread_anchor and route BOTH lanes through it, so text and media
cannot drift again. The media lane copies caller metadata rather than mutating
it — these helpers are called in loops with a shared mapping.
Also fold send_typing/stop_typing's duplicated status-anchor blocks into
_with_status_thread_anchor. They had already drifted (stop_typing omitted the
platform check) and the clear must target the thread the heartbeat set or the
status line sticks until Slack's own timeout.
Tests: media lane pinned in both modes plus the channel and
no-caller-mutation cases; verified as real by reverting the fix and watching
them fail.
Three provably-safe optimizations for O(n)-per-iteration history walks:
1. sanitize_tool_call_arguments: optional identity-keyed cursor (strong
refs to the exact validated message objects) skips re-json.loads-ing
already-validated history each loop iteration. Any list rewrite
(compression, repair, undo, steer) breaks the identity prefix match
and forces re-scan from the divergence point. Wired via a per-agent
cursor dict in conversation_loop.
2. estimate_messages_tokens_rough: per-message memo keyed on a deep
identity fingerprint (strings pinned by strong reference so id()
aliasing is impossible; scalars by value; dicts/lists structurally
with key order). Equal fingerprints imply identical str(shadow)
bytes, hence identical estimates. Unfingerprintable shapes fall
through to direct compute. Bounded FIFO cache (4096 entries).
3. _flush_messages_to_session_db_unlocked: bounded scan that skips the
identity-matched prefix of the previous successful flush's snapshot.
Snapshot only taken on full success; cleared on exception. Compression
rewrites use fresh copies, breaking identity and forcing full re-scan.
Parity proven in tests/agent/test_cursor_optimizations_parity.py:
500-message synthetic histories with tool calls, malformed args, unicode,
element-wise old==new across 3 iterations incl. simulated compression.
Measured (median of 5): sanitize 0.097ms->0.011ms, tokens 1.145ms->0.853ms,
persist-scan 179.5us->10.0us at 500 messages.
