Three tests for the #61099 salvage: OpenRouter attribution headers
present after switching to openrouter.ai, Kimi User-Agent sentinel
present after switching to api.kimi.com, and stale headers cleared
when switching to a provider with no URL-specific headers.
2/3 fail on unpatched main (DID NOT ATTACH), confirming the bug.
Follow-ups for salvaged PR #43819: the registry key was
str(Path(db_path).expanduser()) — a symlinked or relative path to the
same DB file got its own connection, silently reintroducing the exact
multi-writer contention the registry prevents. Key on Path.resolve()
(OSError-tolerant fallback). Adds a symlink regression test and the
AUTHOR_MAP entry for adambiggs.
Every MemoryStore instance opened its own SQLite connection guarded by
its own RLock. Several providers coexist in one process (the main agent
plus every delegate_task subagent), so instances pointing at the same
memory_store.db raced as independent WAL writers. Combined with writes
that were not rolled back on error, one connection could leave an open
write transaction that pinned the write lock and made every other
connection's writes fail with "database is locked" for the full busy
timeout.
Instances for the same database now share ONE process-wide connection
and ONE re-entrant lock, so access is fully serialized and
cross-connection contention is impossible. The shared connection is
refcounted: closing one instance never tears it out from under a live
sibling, and the last close releases it. The connection runs in
autocommit (isolation_level=None) so a write that raises mid-method can
never leave a dangling transaction holding the write lock; the existing
explicit commit() calls become harmless no-ops.
The provider's shutdown() now calls the refcount-guarded close() instead
of just dropping the reference: leaving finalization to GC kept the
connection (and its write lock) alive indefinitely on long-running
gateways, prolonging the exact contention this fix removes. The last
provider now releases the connection deterministically while siblings
stay live; regression tests fail without the wiring.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Flip the two tests that pinned the old buggy behavior (rewrite_transcript
called after in-place compaction) to assert the corrected invariant from
#61145: archive_and_compact() already persisted, so the handler must NOT
call rewrite_transcript — its replace_messages(active_only=False) would
DELETE the just-archived rows.
E2E-verified against a real SessionDB: 6 soft-archived rows are wiped by
replace_messages' default path, confirming the data-loss premise.
test_accepted_at_every_position spawned 11 separate
'python -m hermes_cli.main' subprocesses, each cold-importing the full
CLI module tree under a 15s TimeoutExpired deadline. On a loaded CI
worker the import alone can exceed that (slice 2/8 flaked exactly here
on PR #61726's run, TimeoutExpired at subprocess.py:1253), failing PRs
that never touched the CLI.
Replace with ONE driver subprocess that imports hermes_cli.main once
and parses all 11 argvs in-process (catching SystemExit per argv),
reporting JSON results. Same assertions per argv, identical semantics
(verified the --help-before-unknown-flag exit behavior matches the old
method), ~11x less import work, and the 180s timeout only trips on a
genuine hang.
Structural completion of the malformed-job freeze fixes (#61382 id-less,
#61525 non-dict schedule, #61581 bad next_run_at): wrap the per-job body
of _get_due_jobs_locked in try/except so any FUTURE malformed-field
variant degrades to skipping that one job for the tick instead of
aborting the scan before save_jobs() and freezing the whole profile's
scheduler.
Also: restore test_repeated_concurrent_runs_accumulate_completed_count
to TestMarkJobRunConcurrency (accidentally re-parented by the #61581
diff), add a containment regression test, and AUTHOR_MAP for hydracoco7.
E2E: one jobs.json carrying all five malformed shapes (drifted job_id,
missing id, null schedule, garbage next_run_at, non-string last_run_at)
plus a healthy sibling — single tick contains all five, sibling fires,
repairs persist, second tick stable. 670 cron tests green.
One bad next_run_at value in jobs.json aborts the due-jobs scan with
ValueError from fromisoformat, before any save_jobs, so siblings lose
progress (fast-forwards etc).
Early normalization in _get_due_jobs_locked + defensive parses in
compute_next_run / _recoverable_oneshot_run_at.
Added test_bad_next_run_at_does_not_crash_or_block_sibling_jobs.
A job record in jobs.json can have a non-dict 'schedule' value (null, string,
etc.) from direct edit or old writers.
In _get_due_jobs_locked:
schedule = job.get('schedule', {})
kind = schedule.get('kind')
This (and direct schedule['kind'] in compute_next_run etc.) raises and
aborts the entire due-jobs scan before save_jobs() or advancing next_run_at
for healthy jobs. Exactly the same failure mode as the id-less job P1.
Fix: normalize non-dict schedules to {} early (before any use), matching the
defense added for id-less records. Also added defensive guards in compute
functions.
Added regression test that a bad schedule does not crash and healthy sibling
is still returned.
Refs similar pattern in #61382.
A cron record authored by a direct jobs.json edit that bypassed
add_job() can lack an "id" key (older writers used "job_id"). Every
site in _get_due_jobs_locked indexes job["id"] eagerly — both the
logging helpers (job.get("name", job["id"]) evaluates the default
argument unconditionally) and the 'for rj in raw_jobs: if rj["id"] ==
job["id"]' persistence loops. A single malformed record therefore
raised KeyError mid-tick, aborting the entire scan before save_jobs()
ran. Result: healthy jobs' fast-forwarded next_run_at was computed in
memory then discarded on the exception unwind, freezing the whole
profile's scheduler in a per-minute loop (observed dormant for weeks).
Fix: normalize id-less records at the top of _get_due_jobs_locked before
anything keys off job["id"] — recover the id from a drifted "job_id"
key when present, else synthesize one via uuid4, and persist. This
repairs the whole bug class at the source rather than guarding each of
the ~12 downstream index sites.
Adds a regression test that fails with KeyError on the current code and
passes with the fix, asserting a healthy sibling job is still returned
when an id-less record shares the store.
The HTML session export interpolated the tool-call name into the page
without escaping, while every sibling field went through _escape_html. A
tool-call name is attacker-influenced, so a prompt-injected model can emit
a name containing HTML that executes when the export is opened in a browser.
Escape the tool-call name like the other fields.
switch_model() unconditionally set agent.provider but only set
agent.base_url when the resolved value was truthy. When a real
provider change resolved an empty base_url (e.g. minimax after
copilot), the agent ended up with provider="minimax" but
base_url still pointing at api.githubcopilot.com. That incoherent
pair then got snapshotted into agent._primary_runtime, so it kept
re-applying on every subsequent turn via restore_primary_runtime()
until the process restarted.
try_activate_fallback() and _swap_credential() were audited and
confirmed unaffected: both always derive base_url from an actually
constructed client, never from a possibly-empty resolver hint.
Fix: when base_url is empty AND the provider is genuinely changing,
raise ValueError instead of silently keeping the old provider's URL.
This routes through switch_model()'s existing snapshot/rollback
path, and callers (tui_gateway/server.py's _apply_model_switch)
already catch and surface a clean "switch failed, staying on X"
message. Re-selecting the SAME provider with an empty base_url
(credential-only refresh) still keeps the current URL, unchanged.
Fixes#47828
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Port from openai/codex#31188: a parse failure in a policy-bearing config
file must not silently replace the effective policy with an empty/default
one. Codex's load_exec_policy_with_warning replaced the whole exec policy
with Policy::empty() when a .rules file failed to parse, silently dropping
managed prompt/forbidden rules; the fix preserves the managed policy while
still warning.
Hermes had the same bug shape in load_config(): a YAML parse error made
_load_config_impl() fall through to DEFAULT_CONFIG, dropping every user
override — including approvals.deny rules, which are documented to block
commands even under --yolo. In a long-running gateway, a user mid-editing
config.yaml into broken YAML silently disarmed their own deny rules on the
next load.
Now, when the process has a last successfully loaded config for that path
(_LAST_EXPANDED_CONFIG_BY_PATH), a parse failure keeps serving it (cached
under the corrupt file's signature so the broken file isn't re-parsed) and
the warning says edits are being ignored until the YAML is fixed. Fresh
processes with no last-known-good keep the existing DEFAULT_CONFIG
fallback and warning.
E2E-verified: deny rule 'curl*evil.com*' still blocks after mid-process
corruption; fixed file reloads normally; fresh-process fallback unchanged.
The earlier fix gated _resolve_use_tui, but the EARLY launcher
(_wants_tui_early) decides TUI from display.interface before cmd_chat
runs — so a `display.interface: tui` default still booted the Ink UI for
headless spawns (kanban workers), whose no-TTY bail-out exits 0 →
"protocol violation". Gate the early resolver on a real TTY: headless
stdio never boots the TUI regardless of config; explicit --tui still does.
Dragging a task done→ready did nothing: the respawn guard saw a run that
completed within the success window and deferred forever, unable to tell a
deliberate operator re-run from a status flap. Now a re-queue event
(status change, promote, unblock, reclaim) AFTER the completion bypasses
the recent_success guard, so an explicit done→ready runs again.
A retried task (→ running) kept showing "crashed Nx": the in-flight run
has no outcome yet, so the trailing crash scan skipped it and kept
counting the prior streak, and the consecutive_failures counter lingers.
Exempt `running` from both repeated_failures and repeated_crashes so a
fresh attempt clears the banner until it itself resolves (re-fires if the
new run also fails).
An inherited HERMES_TUI=1 or a `display.interface: tui` config default sent
kanban workers into the Ink TUI, whose no-TTY bail-out exits 0 without doing
the task — every attempt ended in "protocol violation". Two layers:
- _default_spawn pins `--cli` (highest-precedence interface flag) and strips
HERMES_TUI from the child env (covers older builds on PATH).
- _resolve_use_tui gates ambient TUI prefs (env/config) behind a real TTY;
an explicit --tui still wins so the informative bail-out stays reachable.
A manual done (dashboard/desktop drag) runs complete_task but ends no
run, so a trailing crashed/crashed run history never gains the
'completed' outcome that breaks the repeated_crashes streak — the card
kept flagging "needs attention" forever after being finished.
repeated_failures had the same hole via a stale counter. Terminal
statuses are now exempt from both: done means done; the history stays
on the event log for audit. Regression test included.
Adding the 6 gpt-5.6 slugs to DEFAULT_CODEX_MODELS grew the curated codex
catalog to 11, above the test's max_models=10 cap. That truncated the
picker list to 10 while total_models reported 11, breaking the
total_models == len(models) assertion. The cap was an implicit
change-detector on catalog size; raise it to 100 so the list is never
truncated and the count-consistency invariant stays meaningful as new
gpt-5.x slugs land.
Close the remaining end-to-end gaps so the full gpt-5.6 family (sol/
terra/luna + their -pro high-effort modes, 6 slugs) works on every
surface a user can reach them through:
- agent/auxiliary_client.py: the Codex OAuth backend hard-caps context
at 272K for gpt-5.6 exactly as it does for 5.4/5.5, but the default
50% compaction trigger would summarize at ~136K and waste half the
usable window. Extend the existing _is_codex_gpt54_or_gpt55 chokepoint
(single enforced predicate feeding _compression_threshold_for_model)
to match gpt-5.6* on the openai-codex route so those sessions get the
same 0.85 auto-raise. Direct-API/OpenRouter routes (full 1.05M window)
are unaffected; the historical codex_gpt55_autoraise opt-out still
applies. The one-time notice banner is model-dynamic and already
renders the correct slug/cap.
- hermes_cli/config.py, agent/agent_init.py: refresh the autoraise
comments/notice to mention the 5.6 family.
- hermes_cli/codex_models.py: add the -pro variants to DEFAULT_CODEX_MODELS
+ forward-compat so ChatGPT-OAuth (openai-codex) Pro users see the full
family in /model, not just the base tiers.
Supersedes the earlier commit's note that 5.6 was intentionally kept out
of the codex catalog: the slugs are confirmed routable (OpenRouter live
+ codex backend), so they belong there like every other codex-capable
gpt-5.x slug.
E2E verified across all 6 slugs: direct-API ctx 1.05M, codex ctx 272K,
pricing reachable from openai + openai-api routes, codex compaction
override 0.85 (and None on direct-API + when opted out), present in
openai-api picker + codex catalog, /model gpt resolves to sol on both
native routes. Guard tests added for the compaction route matrix.
PR #61587 adds sol-pro/terra-pro/luna-pro to the aggregator lists.
Complete those on the native surfaces the same way this PR completes
the base tiers:
- hermes_cli/models.py: -pro variants in _PROVIDER_MODELS[openai-api].
- agent/usage_pricing.py: alias ("openai", "gpt-5.6-*-pro") onto the
base-tier PricingEntry rows — the -pro high-effort modes bill at the
SAME per-token rates (verified against OpenRouter live pricing
2026-07-09: identical prompt/completion prices for base and -pro);
they cost more per task by consuming more tokens, not a higher rate.
- Context lengths need no new entries: "gpt-5.6-sol" et al. are
substrings of their -pro variants and both lookup tables match
longest-key-first (verified: sol-pro -> 1.05M direct / 272K codex).
- model_switch sort: -pro variants parse as suffix "sol-pro" (rank 1),
so /model gpt still defaults to base sol — pinned by test.
- Not added to DEFAULT_CODEX_MODELS: only confirmed routable via API/
OpenRouter so far; codex live discovery will surface them if ChatGPT
exposes them, same policy as other unconfirmed codex slugs.
Tests: invariant tests extended (pro aliases share base entries, base
sol outranks sol-pro); 191 targeted tests pass.
Phase-2 review findings addressed:
- resolve_billing_route: normalize the "openai-api" picker slug to the
"openai" billing provider — without this the ("openai", <model>)
_OFFICIAL_DOCS_PRICING keys (incl. every pre-existing gpt-4o/gpt-4.1
entry, not just 5.6) were unreachable when the provider is openai-api.
- pricing_version: drop the "preview" tag (GA 2026-07-09 at same rates).
- model_metadata comment: dict order is cosmetic — lookups length-sort
keys at match time; the old comment implied a positional invariant.
- model_switch comment: note "sol" is a series codename, not a generic
quality word.
- tests/hermes_cli/test_gpt56_registration.py: behavior contracts (no
list snapshots) — sol > terra/luna > 5.5 sort invariant, pricing
reachability from both openai and openai-api routes, cache-write
1.25x / cache-read 0.10x input relation.
hermes send "MEDIA:/x.png This Caption" now arrives as one native captioned
bubble instead of a separate text message followed by an uncaptioned bubble.
Root cause: the standalone senders (hermes send / cron / send_message tool)
stripped the MEDIA: tag, sent the remaining text as its own message, and
called the media send with no caption -- even though hermes send's help
advertises the captioned form and the bridges/adapters already support a
caption. Signal already captioned correctly.
- tools/send_message_tool.py: new _media_caption_split() chokepoint decides
caption-vs-separate-body (single captionable non-voice file within the
platform's message-length cap). Wired into the Telegram, WhatsApp and
Discord dispatch paths.
- Telegram/WhatsApp/Discord: when the single captioned file is missing, the
caption text is delivered as a plain message so it is never silently lost.
- Telegram caption send gets a MarkdownV2->plain parse fallback.
- Tests: _media_caption_split unit tests + per-platform caption tests
(ride, multi-file fallback, voice exclusion, over-limit fallback,
missing-file text fallback); updated the 3 tests that asserted the old
text-then-media split.
Closes the gap reported against #58911 (the MEDIA_CAPTION directive PR);
credit to @ferreiraesilva for surfacing the caption behavior.
Review finding: callers mutate the returned dicts in place —
hermes_cli/web_server.py annotates s['enabled']/s['usage'] on the skills
list — so handing out the cached objects poisons the cache for every
subsequent caller (and is a cross-thread shared-mutable hazard in the
gateway). Return [dict(s) for s in cached] on both hit and miss paths;
warm-path cost is negligible (241x speedup retained on a 300-skill
fixture). Regression test mutates a returned list/dict and asserts the
next cached call is clean.
Review findings on the cherry-picked cache (follow-up to #58985):
- The cache key was the max mtime of only the TOP-LEVEL scan dirs.
Adding/removing a skill inside a category subdir bumps the category
dir's mtime, NOT the root's, so the cache served a stale list
indefinitely. Replace with a per-dir signature covering roots +
immediate children (one scandir per dir; mirrors
hermes_cli/profiles.py::_count_skills from d5eee133e).
- The disabled-set is config-driven and changes with no filesystem
mtime bump; fold it into the signature so /skills disable takes
effect without a restart.
- Platform is part of the signature (gateway processes serve multiple
platform scopes; scan results are platform-filtered).
- Add a 30s TTL to bound staleness from in-place SKILL.md edits (file
mtime is invisible to any directory signature).
- The original also keyed dirs off the module-level SKILLS_DIR constant;
the scan itself uses _skills_dir() (live profile HERMES_HOME) — use
the same resolution for the signature.
Mutation-verified: nested-add, disabled-set, and TTL tests fail against
the pre-fix cache and pass with it.
The salvaged PR added a standalone test_reasoning_timeouts.py that duplicated
the structure of the existing parametrized test_reasoning_stale_timeout_floor.py.
Fold the v4-flash/v4-pro/-free positive cases and deepseek-chat negative cases
into the canonical parametrized tables and remove the redundant file.
DeepSeek V4 models (deepseek-v4-flash, deepseek-v4-pro) emit
reasoning_content in a separate delta field before final content,
requiring the same 600s stale timeout floor as R1. Without this,
streams hang for 30–50s with APITimeoutError on providers like
opencode-go while direct calls succeed in ~3s.
Fixes#60338.
Phase-2 review finding: the validation branch expanduser()s the path but
web_server.py reads os.environ['HERMES_WEB_DIST'] raw at import — a
'~/dist' value would validate here and still 404 there. Write the
expanded path back before the web_server import. Adds a regression test
asserting the env var holds the expanded path after cmd_dashboard.
Review finding: PR #40632's branch had silently dropped gemma-3-27b
from the keep-extra_content test loop (part of its Gemma narrowing,
which this salvage reverts). Restore main's original coverage so a
future narrowing back to Gemini-only fails loudly.
Trim the salvaged commit to its two still-valid conversions:
- ChatCompletionsTransport.convert_messages (copy-on-write sanitize)
- QwenProfile.prepare_messages (copy-on-write normalize + cache_control)
Dropped from the original PR:
- agent/prompt_caching.py selective-copy: superseded by #57229 which
already rewrites apply_anthropic_cache_control on current main.
- Gemma extra_content narrowing (_model_consumes_thought_signature
'gemini or gemma' -> 'gemini' only) + its two tests: unrelated
behavior change reverting deliberate e8c3ac2f5; belongs in its own
PR with its own justification if pursued.
Conflict resolution: preserved main's newer timestamp-stripping
(#47868) inside the copy-on-write path.
Mock _preflight_user_systemd and _select_systemd_scope in
test_systemd_start_refreshes_outdated_unit and
test_systemd_restart_refreshes_outdated_unit. These tests target
unit-file refresh logic, not D-Bus reachability, so the preflight
check was causing spurious UserSystemdUnavailableError on macOS,
WSL, and Docker where systemd is unavailable.
(cherry picked from commit 34113300a1)
A custom HERMES_WEB_DIST without --skip-build skipped BOTH the web UI
build and any validation: cmd_dashboard fell through the build gate and
started the server against a dist that may not exist, serving 404s with
no obvious cause. This is the same failure mode issue #23817 fixed for
the --skip-build branch — the env-var branch was left unvalidated.
Add the missing else-branch: fail fast with actionable guidance when
HERMES_WEB_DIST has no index.html, proceed (still without building) when
it does.
Credit: @Caelier (#17845) originally proposed dist validation for the
dashboard startup path; the --skip-build half of that PR's scope has
since landed via the #23817 fix, this covers the remaining env-var path
on the rewritten cmd_dashboard surface.
Review finding: the substring check ('kimi' or 'moonshot' in model)
under-matches bare release slugs like k2-thinking that the repo's
canonical _model_name_is_kimi_family matcher (anthropic_adapter.py)
already covers. Reuse it instead of a second ad-hoc matcher; add a
regression test for the bare-slug case.
Adapted from PR #26014's test file to the canonical seam
(tests/run_agent/test_anthropic_prompt_cache_policy.py _make_agent
helper) instead of a new top-level file with sys.path manipulation.
Covers: kimi-k2.6 + moonshot-v1 on OpenRouter (envelope layout),
kimi via Nous Portal, and the non-OpenRouter negative case.
Structured review (2a/2b/2c) findings, all fixed:
- MAJOR: detect_local_server_type memo was process-lifetime with no
invalidation, permanently pinning a URL's server type. Now a bounded
1h TTL ((type, monotonic) tuples) so a backend swap on the same port
is re-detected. Test covers ollama->lm-studio swap after expiry.
- MAJOR: legacy disk-row compat was one-way. get_cached_context_length
and _invalidate_cached_context_length now consult the same key-shape
set {canonical, literal, canonical+slash} in both directions, so an
old slashed row is found (and cleared) when the runtime passes the
normalized URL. Tests pin both migration directions.
- MINOR: _localhost_to_ipv4 did whole-string replacement, which could
corrupt a proxy URL embedding http://localhost in its query. Now a
scheme-anchored host-only regex; localhost.example.com and embedded
substrings pass through. Tests added.
- MINOR: _invalidate_cached_context_length now also drops the
in-memory TTL probe rows for the pair, so a resolution inside the
TTL window can't re-persist the value just declared stale.
- Test gaps closed: detect-type cache hit + TTL-expiry re-detection,
ollama-show TTL expiry re-probe, reverse legacy-row lookups.
- Attribution gate: added zhchl@hermes-agent.local -> 8294 (PR #50572
author) to AUTHOR_MAP; the strict CI grep needs bare non-plus emails
literal in release.py.
Gates: ruff clean; targeted suites 202 passed / 0 failed; full
tests/agent 5426 passed with 17 failures identical on clean
upstream/main (pre-existing env-dependent anthropic/bedrock/credpool
tests); mypy delta vs base: 0 new errors; live smoke 6/6 PASS.
#37595 fixed the Windows dual-stack IPv6 timeout only inside
detect_local_server_type. The same 2s-per-probe penalty existed at every
other helper that builds a probe URL from base_url. Extract the rewrite
into _localhost_to_ipv4() and apply it at:
- query_ollama_num_ctx
- query_ollama_supports_vision
- _query_ollama_api_show (server_url derivation)
- _query_local_context_length (server root + LM Studio native URL)
Tests cover the helper's URL forms, non-localhost passthrough, and that
the ollama probes actually POST to 127.0.0.1.
Follow-up hunks completing the probe-cache cluster:
1. _query_ollama_api_show now goes through the existing
_LOCAL_CTX_PROBE_CACHE (30s TTL, positive-only, namespaced key) —
it was the one remaining per-resolution POST not covered by the
#56431-era wrapper. Failures are never memoized so a server that
comes up mid-startup is re-probed. Idea credit: #42081 (@Morad37),
reworked to comply with the positive-only rule.
2. Persistent context-cache keys are normalized through
_context_cache_key (trailing-slash strip) so http://host/v1 and
http://host/v1/ share one entry; reads and invalidation honor
legacy un-normalized rows. Idea credit: #37905 (@stevenau21).
Tests: TTL hit collapses to one POST, failure-not-memoized
(mutation-verified: unconditional caching makes it fail), namespace
no-collision vs the sibling probe, slash-variant dedup, legacy-row
read, dual-shape invalidation.
Remaining hunk of the PR-branch fixup commit: the LM Studio first-probe
assertion now expects the IPv4-resolved URL (the code half — applying
the rewrite to `normalized` before deriving lmstudio_url — was folded
into the previous cherry-pick's conflict resolution).
(cherry picked from commit 7d324b0e47)
On Windows, `localhost` resolves to both ::1 (IPv6) and 127.0.0.1 (IPv4).
httpx tries IPv6 first, hanging 2 sec per probe when the server binds IPv4
only. detect_local_server_type() is called 3+ times during init, each with
a new httpx.Client, compounding to ~14s of dead time.
Replace localhost with 127.0.0.1 inside the function before connecting.
The function is only called for local endpoints (callers guard with
is_local_endpoint()), so IPv6 loopback adds no diagnostic value.
Measured: 19.9s → 4.0s on Windows with a local proxy on 127.0.0.1:8317.
(cherry picked from commit a075d3194b)
Follow-up to HexLab98's fix. The sync-before-regenerate invariant was
enforced by convention across 6 callsites (~3 idempotent unit-file reads
per command). Consolidate it into the one function every compare/regenerate
path funnels through — systemd_unit_is_current — and drop the now-redundant
callsite pre-syncs in refresh_systemd_unit_if_needed / systemd_start /
systemd_restart / systemd_status.
Kept the systemd_install pre-sync: the --force path bypasses the
is_current gate and calls generate_systemd_unit() directly, so it needs
its own sync to avoid baking /root/.hermes under sudo.
Reworked the two callsite-ordering tests into a chokepoint-invariant guard
(test_is_current_syncs_before_reading_unit) + a delegation test proving
start/restart no longer pre-sync. Both fail if the chokepoint sync is
removed; the pre-existing behavior test still passes.
Follow-up to the salvaged str(tools) fix. The id()-keyed
_TOOLS_TOKENS_CACHE had no eviction, so a long-lived gateway/desktop
backend could accumulate an unbounded number of stale entries as it
builds transient tool lists. Cap it at 256 with oldest-first eviction
(insertion-ordered dict) and add a regression test asserting the cache
never exceeds the cap.
