shutil.copytree from default ~/.hermes duplicated ~/.hermes/profiles into
the new profile, causing nested profiles/.../profiles/... and huge disk use.
Match export behavior (_DEFAULT_EXPORT_EXCLUDE_ROOT) by ignoring the sibling
profiles tree at the source root.
Made-with: Cursor
Intended placement per PR #17610 discussion — comfyui belongs in
skills/creative/ alongside other creative built-ins (touchdesigner-mcp,
pretext, sketch), not in optional-skills/.
Pure directory rename, no content changes. History preserved via git mv.
The skip_pre_tool_call_hook flag was added to prevent double-firing of
pre_tool_call when run_agent._invoke_tool pre-checks for a block
directive and then dispatches via handle_function_call. But the
implementation added an else: branch that fired invoke_hook again for
'observers', without noticing that get_pre_tool_call_block_message() in
hermes_cli.plugins already fires invoke_hook('pre_tool_call', ...) as
part of its block-directive poll.
Result: every tool call ran through the run_agent loop fired the hook
twice — reported by community users whose observer / audit plugins
logged each tool invocation twice with identical timestamps.
Fix: delete the else: branch. The single-fire contract is now:
- skip=False (direct handle_function_call): hook fires once inside
get_pre_tool_call_block_message().
- skip=True (run_agent._invoke_tool path): caller fires the hook
once via get_pre_tool_call_block_message(); handle_function_call
must not fire it again.
Tightened the existing skip-flag test (renamed to
test_skip_flag_prevents_double_fire) to assert pre_tool_call fires
zero times when skip=True, and added
test_run_agent_pattern_fires_pre_tool_call_exactly_once to lock in
end-to-end that the full block-check + dispatch sequence fires the
hook exactly once.
Adds Step 0 'Ask Local vs Cloud' as the very first onboarding step, with a
scripted question that spells out the hardware requirements for local
(6 GB VRAM NVIDIA, ROCm AMD on Linux, or M1+ Mac with 16 GB unified)
and routes Cloud users straight to Path A without a hardware check.
Hardware check becomes Step 1, run only when the user picked local.
Layers a programmatic hardware-feasibility check on top of the v4 skill
so the agent doesn't silently push users toward a local install they
can't actually run. The official comfy-cli supports --nvidia / --amd /
--m-series / --cpu, but has no guard against "4 GB laptop GPU on SDXL"
or "Intel Mac falling back to CPU" — both route to comfy-cli paths in
the original table and then fail on first workflow.
- scripts/hardware_check.py: detect OS/arch/GPU (NVIDIA nvidia-smi,
AMD rocm-smi, Apple M1+ via arm64+sysctl, Intel Arc via clinfo),
VRAM, system/unified RAM. Emits JSON
{verdict: ok|marginal|cloud, recommended_install_path, comfy_cli_flag}
with practical thresholds: discrete GPU >=6 GB VRAM minimum,
Apple Silicon >=16 GB unified memory minimum, Intel Mac -> cloud,
no accelerator -> cloud. comfy_cli_flag maps directly to
`comfy install` so the agent can stitch the whole flow together.
- scripts/comfyui_setup.sh: runs hardware_check.py first when no
explicit flag is passed. If verdict=cloud, refuses to install
locally, prints Comfy Cloud URL + an override command, exits 2.
Otherwise auto-selects the right --nvidia/--amd/--m-series flag
for `comfy install`. Surfaces marginal-verdict notes to the user.
- SKILL.md Setup & Onboarding: adds mandatory Step 0 "Check If This
Machine Can Run ComfyUI Locally" ahead of the Path A-E selection.
Documents the verdict thresholds inline, ties verdict + comfy_cli_flag
to the install paths, and updates the path-choice table so
"verdict: cloud" is the first row. Quick-Start "Detect Environment"
block extended to include the hardware check. Verification
checklist gains a hardware-check gate.
- Frontmatter setup.help rewritten to point at hardware_check.py
first. Version bumped 4.0.0 -> 4.1.0.
Capture the reusable layout and animation lessons from the advanced Pretext demo so the skill teaches measured obstacle fields, morphing geometry, and polished browser examples.
Cron is a built-in Hermes feature (CLI `hermes cron`, `cronjob` agent
tool, gateway ticker, scheduler in cron/scheduler.py) but croniter has
been gated behind the [cron] optional extra. Users who do a plain
`pip install hermes-agent` can create jobs via /cron but any recurring
cron schedule silently returns next_run_at=None (HAS_CRONITER=False),
which then gets wrapped into a 'state=error' message only after a tick.
Move croniter into core dependencies so scheduled jobs work out of the
box on any install path. The [cron] extra is kept as an empty
passthrough so existing `pip install hermes-agent[cron]` installs and
the [all]/[termux] extras continue to resolve.
Also update the now-stale user-facing error message in
`compute_next_run()` that still tells users to install `hermes-agent[cron]`.
Salvaged from #17234 (authored by @txbxxx) with a corrected premise:
the original PR claimed [cron] wasn't in [all], but it is (pyproject.toml
line 112). The real UX problem is the plain no-extras install path,
which this fix addresses.
Add a dedicated 'Pinning a skill' section that covers both gating
layers — curator auto-transitions AND the agent's skill_manage tool
— so users know what the flag actually protects against after
PR #17562. Updates the one-line claim in 'How it runs' to cross-link
the new section instead of only mentioning auto-transitions.
Extend curator's pin flag from 'skip auto-transitions' to 'no agent
edits at all'. All five skill_manage mutation actions (edit, patch,
delete, write_file, remove_file) now refuse pinned skills with a
message pointing the user at `hermes curator unpin <name>`.
Motivation: pin used to only stop the curator's own maintenance pass
from touching a skill. Nothing prevented the main agent from editing
or deleting a pinned skill via skill_manage in-session. This gives
users a hard fence against unwanted agent edits — same semantics as
curator pinning, extended to the write tool.
Create is unaffected (you can't pin a name that doesn't exist yet,
and name collisions already error out). Broken sidecars fail open
rather than lock the agent out.
The schema description advertises the new refusal so models know
not to route around it with rename/recreate tricks.
Skill catalog pages (bundled/optional) were drowning out real user-guide
and reference docs in search results. There are ~3100 of them and they
match on almost every generic term.
- Add `ignoreFiles` regexes to docusaurus-search-local for
`user-guide/skills/bundled/` and `user-guide/skills/optional/`.
The two human-written catalog indexes (`reference/skills-catalog`,
`reference/optional-skills-catalog`) remain indexed.
- Add a new feature page `user-guide/features/curator.md` covering the
curator subsystem merged in #16049 and refined in #17307 (per-run
reports): how it runs, config, CLI (`hermes curator status/run/pin/
restore/...`), `.usage.json` telemetry, archival semantics, and
recovery. Slotted into the Core features sidebar next to Skills.
Search index size dropped from 5822 docs to 2704 in the main section;
`user-guide/features/curator` is indexed.
Close integration gaps discovered by auditing qwen-oauth's file coverage.
These are surfaces the original salvage missed — they all existed on
main and were added in the 747 commits since PR #15203 was opened.
Coverage added:
- agent/credential_pool.py: seed pool from auth.json providers.minimax-oauth
so `hermes auth list` reflects logged-in state and
`hermes auth remove minimax-oauth <N>` works through the standard flow.
- agent/credential_sources.py: register RemovalStep for minimax-oauth
with suppression-aware `_clear_auth_store_provider`.
- agent/models_dev.py: PROVIDER_TO_MODELS_DEV mapping (-> 'minimax' family).
- hermes_cli/providers.py: HermesOverlay entry (anthropic_messages transport,
oauth_external auth_type, api.minimax.io/anthropic base).
- hermes_cli/model_normalize.py: add to _MATCHING_PREFIX_STRIP_PROVIDERS so
`minimax-oauth/MiniMax-M2.7` in config.yaml gets correctly repaired.
- hermes_cli/status.py: render MiniMax OAuth block in `hermes doctor`
(logged-in / region / expires_at / error).
- hermes_cli/web_server.py: register in OAUTH_PROVIDER_REGISTRY + dispatch
branch in _resolve_provider_status so the dashboard auth page shows it.
- website/docs/integrations/providers.md: full 'MiniMax (OAuth)' section.
- website/docs/reference/cli-commands.md: --provider enum.
- website/docs/user-guide/features/fallback-providers.md: fallback table row.
- scripts/release.py AUTHOR_MAP: amanning3390 mapping (CI gate).
Add comprehensive documentation for the minimax-oauth provider.
New file: website/docs/guides/minimax-oauth.md
- Overview table (provider ID, auth type, models, endpoints)
- Quick start via 'hermes model'
- Manual login via 'hermes auth add minimax-oauth'
- --region global|cn flag reference
- The PKCE OAuth flow explained step-by-step
- hermes doctor output example
- Configuration reference (config.yaml shape, region table, aliases)
- Environment variables note: MINIMAX_API_KEY is NOT used by
minimax-oauth (OAuth path uses browser login)
- Models table with context length note
- Troubleshooting section: expired token, timeout, state mismatch,
headless/remote sessions, not logged in
- Logout command
Updated: website/docs/getting-started/quickstart.md
- Add MiniMax (OAuth) to provider picker table as the recommended
path for users who want MiniMax models without an API key
Updated: website/docs/user-guide/configuration.md
- Add 'minimax-oauth' to the auxiliary providers list
- Add MiniMax OAuth tip callout in the providers section
- Add minimax-oauth row to the provider table (auxiliary tasks)
- Add MiniMax OAuth config.yaml example in Common Setups
Updated: website/docs/reference/environment-variables.md
- Annotate MINIMAX_API_KEY, MINIMAX_BASE_URL, MINIMAX_CN_API_KEY,
MINIMAX_CN_BASE_URL as NOT used by minimax-oauth
- Add minimax-oauth to HERMES_INFERENCE_PROVIDER allowed values
Wire MiniMax-M2.7 and MiniMax-M2.7-highspeed into the model catalog,
CLI model picker, and agent auxiliary/metadata subsystems.
Changes:
- hermes_cli/models.py:
- Add 'minimax-oauth' to _PROVIDER_MODELS with MiniMax-M2.7 and
MiniMax-M2.7-highspeed
- Add ProviderEntry('minimax-oauth', 'MiniMax (OAuth)', ...) to
CANONICAL_PROVIDERS near existing minimax entries
- Add aliases: minimax-portal, minimax-global, minimax_oauth in
_PROVIDER_ALIASES
- hermes_cli/main.py:
- Add 'minimax-oauth' to provider_labels dict
- Insert 'minimax-oauth' into providers list in
select_provider_and_model() near the other minimax entries
- Add 'minimax-oauth' to --provider argparse choices
- Add _model_flow_minimax_oauth() function: ensures login via
_login_minimax_oauth(), resolves runtime credentials, prompts for
model selection, saves model choice and config
- Add dispatch elif branch for selected_provider == 'minimax-oauth'
- agent/auxiliary_client.py:
- Add 'minimax-oauth': 'MiniMax-M2.7-highspeed' to
_API_KEY_PROVIDER_AUX_MODELS
- Add 'minimax-oauth' to _ANTHROPIC_COMPAT_PROVIDERS set
- agent/model_metadata.py:
- Add 'minimax-oauth' to _PROVIDER_PREFIXES frozenset
- MiniMax-M2.7 context length (200_000) already covered by the
existing 'minimax' substring match in DEFAULT_CONTEXT_LENGTHS
compute_next_run() ignored the last_run_at parameter for cron-type
schedules, always computing from _hermes_now() instead. This was
inconsistent with interval jobs which DO use last_run_at as the anchor.
After a crash or restart, cron jobs would compute next_run_at from
the arbitrary restart time rather than the actual last execution time.
While the stale detection in get_due_jobs() catches most cases, using
last_run_at as the croniter base eliminates edge cases and makes the
behavior consistent across schedule types.
Salvaged from #9014 (authored by @beenherebefore) onto current main.
The original PR branch was 2+ weeks stale and would have reverted
substantial unrelated work (jobs_file_lock, workdir/context_from/
enabled_toolsets, issue #16265 state=error recovery). Kept just the
7-line substantive fix and the regression test.
Bare `float(os.getenv("HERMES_CRON_TIMEOUT", 600))` in `run_job()` raises
a `ValueError` when the env var is set to a non-numeric string (e.g. "abc").
Replace it with the same defensive try/except pattern already used by
`_get_script_timeout()` for `HERMES_CRON_SCRIPT_TIMEOUT`: log a warning
and fall back to the 600 s default instead of crashing.
Also update the existing env-var tests to exercise the new code path and
add two new tests — one for an invalid value, one for an empty string.
Fixes#11319
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Closes#4759, closes#4381.
Mutating actions (patch, edit, write_file, remove_file, delete) used to
refuse skills that lived under `skills.external_dirs` with 'Skill X is in
an external directory and cannot be modified. Copy it to your local skills
directory first.' Faced with that error, the agent would fall back to
action='create', which always writes under ~/.hermes/skills/ — producing
a silent duplicate of the external skill in the local store.
Fix: drop the read-only gate. `skills.external_dirs` is configured by the
user; if they pointed it at a directory, they already said 'these are my
skills, treat them the same.' Filesystem permissions handle the genuine
read-only case (write fails, agent sees the error).
- New _containing_skills_root() resolves whichever dir actually contains
the skill; _delete_skill uses it to bound empty-category cleanup so an
external root is never rmdir'd.
- _create_skill behavior is unchanged: new skills still land in local
SKILLS_DIR only. Fewer moving parts.
- Seven new TestExternalSkillMutations tests covering patch/edit/write_file/
remove_file/delete/create against a mocked two-root layout + a category
rmdir-safety check.
When a user authenticates a built-in provider via env var (e.g. DASHSCOPE_API_KEY
triggers the built-in 'alibaba' row) AND defines a custom_providers entry
pointing at the same endpoint, the picker previously emitted two rows for one
endpoint. The built-in row already carries the canonical slug, curated model
list, and correct auth wiring, so the shadow custom entry is redundant.
Adds a _builtin_endpoints set populated as sections 1/2/2b emit rows. Each
entry is the provider's effective base URL (env override via base_url_env_var
wins over the static inference_base_url, so DASHSCOPE_BASE_URL-overridden
endpoints dedup correctly). Section 4 skips any grouped custom entry whose
base_url matches.
Intentionally does NOT repurpose model_catalog.enabled as a 'hide built-ins'
flag. That config controls the remote curated-manifest fetch (documented on
the model-catalog reference page) and overloading it would silently change
behavior for users who disable it for network/privacy reasons.
Three new tests:
- shadow dedup fires when endpoint matches static inference_base_url
- dedup does NOT hide custom entries on genuinely distinct endpoints
- dedup honors the base_url_env_var override path
Covers the #16748 fix:
- unsigned thinking blocks synthesised from reasoning_content survive replay
- non-latest assistant turns keep their thinking (DeepSeek validates every turn)
- signed Anthropic blocks are stripped (DeepSeek can't validate them)
- cache_control is stripped from thinking blocks
- OpenAI-compat base (api.deepseek.com without /anthropic) is NOT matched
- non-DeepSeek third parties (minimax) keep the generic strip-all behaviour
DeepSeek's /anthropic endpoint requires thinking blocks to be replayed
in multi-turn conversations for reasoning continuity. The existing code
classified api.deepseek.com as a generic third-party endpoint and stripped
ALL thinking blocks, causing HTTP 400 from DeepSeek.
Fix: add _is_deepseek_anthropic_endpoint() detector (following the Kimi
precedent) and a dedicated branch that strips only signed Anthropic blocks
while preserving unsigned ones synthesised from reasoning_content.
This follows the exact same pattern as the Kimi exemption (issue #13848)
and does not change behavior for any other third-party endpoint (Azure,
Bedrock, MiniMax, etc.).
FixesNousResearch/hermes-agent#16748
Follow-up to the cherry-picked PR #17447. The original flush spawned a
bare threading.Thread for the buffer-flush path, overwriting
self._sync_thread — which is aliased to the long-lived writer thread.
Two consequences:
1. No serialization with the writer queue. If old-session retains were
still queued in _retain_queue, the flush ran concurrently with the
writer and both threads could call aretain_batch against the same
document_id.
2. The pre-spawn 'self._sync_thread.join(timeout=5.0)' tried to join the
long-lived writer, which never exits, so the join was a no-op that
just timed out — never actually serialized anything.
Fix: enqueue the flush closure on _retain_queue via _ensure_writer +
put(). Natural FIFO ordering behind any pending retains, no new thread,
no broken join. Shutdown-aware so it doesn't enqueue after teardown.
Tests updated to drain via _retain_queue.join() instead of the stale
_sync_thread.join(). Added regression guard
test_flush_serializes_behind_pending_retains_via_writer_queue that
blocks the writer mid-retain to prove the flush waits in FIFO behind
the old retain.
Also seeds _retain_queue / _shutting_down / stubbed _ensure_writer on
the bare-object test helper in test_memory_session_switch.py so that
path doesn't blow up under the new queue-enqueue.
tests/plugins/memory/test_hindsight_provider.py + tests/agent/test_memory_session_switch.py: 103/103 passing.
Two data-loss / leak gaps in HindsightMemoryProvider.on_session_switch
introduced by #17409.
1. Buffered turns silently lost when retain_every_n_turns > 1.
on_session_switch unconditionally cleared _session_turns without
flushing. Users who batched every N>1 turns and switched mid-batch
(/reset, /new, /resume, /branch, or context compression) had those
buffered turns disappear. Same data-loss class as the shutdown race,
different lifecycle event.
Note commit_memory_session() -> on_session_end() runs *before*
on_session_switch on /reset, but Hindsight doesn't implement
on_session_end so the buffer survives that step and dies at clear
time. /resume, /branch, and compression skip commit_memory_session
entirely so an on_session_end impl wouldn't help them anyway.
Fix: snapshot the old _session_id, _document_id, _parent_session_id,
_turn_index, and _session_turns; spawn one final retain that lands
under the OLD document_id; then rotate state. Metadata is built
synchronously against the old self._* so session_id / lineage tags
on the flushed item all reference the prior session consistently.
2. Stale _prefetch_result leaks across switch.
If queue_prefetch ran in the old session and the result hadn't been
consumed by prefetch() yet, on_session_switch left the cached recall
text in place. The next session's first prefetch() call would return
text mined from the prior session's bank/query.
Fix: join any in-flight _prefetch_thread (3s bounded — matches
shutdown()), then clear _prefetch_result under _prefetch_lock before
rotating session_id.
Tests
-----
- tests/plugins/memory/test_hindsight_provider.py (TestSessionSwitchBufferFlush):
- buffered turns flushed under OLD document_id with OLD lineage tags
- empty buffer => no spurious retain
- _prefetch_result cleared on switch
- in-flight prefetch thread is awaited before clear (no race)
- tests/agent/test_memory_session_switch.py: factory extended to seed the
attrs the new flush path reads (_retain_source, _platform, _bank_id,
prefetch state, etc.) and stub _run_hindsight_operation so existing
switch-state assertions keep passing without network setup.
The ~/.openclaw/ detection banner (#16327) had two problems flagged in #16629:
1. It only pitched 'hermes claw cleanup' (destructive archive) and never
mentioned 'hermes claw migrate' — the actual non-destructive path that
ports config/memory/skills into Hermes.
2. The copy anthropomorphized the bug ('the agent can still get confused',
'dutifully reads') and framed OpenClaw as a competitor to eliminate
('instead of Hermes's').
Rewrite so migrate leads, cleanup is a clearly-labelled follow-up with a
warning that archiving breaks OpenClaw for users still running it.
Closes#16629
Address Copilot review on PR #16666:
1. **Duplicate event on every tool start** — both ``tool_progress_callback``
and ``tool_start_callback`` fire side-by-side in ``run_agent.py``, so
wiring both into chat completions emitted *two* ``hermes.tool.progress``
events per real tool call. Drop the legacy ``_on_tool_progress`` emit
entirely; ``_on_tool_start`` now produces a single unified event that
carries the legacy ``tool``/``emoji``/``label`` fields plus the new
``toolCallId``/``status`` correlation fields. Label is computed inline
via ``build_tool_preview`` so callers do not need to pre-format it.
2. **Weak per-event correlation in the regression test** — the previous
assertion checked that a ``toolCallId`` appeared *somewhere* in the
aggregate, which would have passed even if ``running`` lacked the id.
Collect ``(status, toolCallId)`` per event and assert each event
carries the correct pair, plus exactly two events on the wire (no
silent duplication regression).
The two existing chat-completions tool-progress tests are updated to fire
``tool_start_callback`` instead of ``tool_progress_callback``, matching
production reality where ``run_agent`` always pairs them.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds Vercel Sandbox as a supported Hermes terminal backend alongside
existing providers (Local, Docker, Modal, SSH, Daytona, Singularity).
Uses the Vercel Python SDK to create/manage cloud microVMs, supports
snapshot-based filesystem persistence keyed by task_id, and integrates
with the existing BaseEnvironment shell contract and FileSyncManager
for credential/skill syncing.
Based on #17127 by @scotttrinh, cherry-picked onto current main.
Adds two API server endpoints for external UIs and orchestrators:
- GET /v1/capabilities — machine-readable feature discovery so clients
can detect which Runs API / SSE / auth features this Hermes version
supports before depending on them.
- GET /v1/runs/{run_id} — pollable run status so dashboards can check
queued/running/completed/failed/cancelled/stopping state without
holding an SSE connection open.
Also moves request validation ahead of run allocation so invalid
payloads no longer leave orphaned entries in _run_streams waiting for
the TTL sweep.
task_id is intentionally kept as "default" for the Runs API to
preserve the shared-sandbox model used by CLI, gateway, and the
existing _run_agent_with_callbacks path. session_id is surfaced in
run status for external-UI correlation only.
Salvage of PR #17085 by @Magaav.
The guard that drops Anthropic's `thinking` kwarg for Kimi endpoints was
matched on `https://api.kimi.com/coding` only. Users configuring a
custom Kimi-compatible gateway (or an official Moonshot host) with
`api_mode: anthropic_messages` fall through to the generic third-party
path, which strips thinking blocks AND still sends
`thinking={enabled,...}` → upstream rejects with HTTP 400
"reasoning_content is missing in assistant tool call message at index N"
on the next request after a tool call.
Replace `_is_kimi_coding_endpoint` callers (history replay + thinking
kwarg gate) with `_is_kimi_family_endpoint(base_url, model)` that also
matches the `api.kimi.com` / `moonshot.ai` / `moonshot.cn` hosts and
Kimi/Moonshot family model names (`kimi-`, `moonshot-`, `k1.`, `k2.`,
…) for custom / proxied endpoints. Keeps the UA-header check in
`build_anthropic_client` URL-only — the `claude-code/0.1.0` header is
an official-Kimi contract.
Plumbs optional `model` through `convert_messages_to_anthropic` so
the unsigned reasoning_content→thinking block synthesised for Kimi's
history validation survives the third-party signature-stripping pass
on custom hosts too.
Closes#17057.
The cron schema contracts deliver as a string ("local", "origin",
"telegram", "telegram:chat_id[:thread_id]", or comma-separated combos),
but MCP clients and scripts sometimes pass an array like ['telegram'].
Before this change, the list was written to jobs.json verbatim, and
the scheduler's str(deliver).split(',') then tried to resolve the
literal string "['telegram']" as a platform — returning None and
logging 'no delivery target resolved for deliver=[\'telegram\']'.
Fix on both ends:
- tools/cronjob_tools.py: normalize deliver at the API boundary on
create and update, so storage is always a string.
- cron/scheduler.py: normalize deliver in _resolve_delivery_targets,
so existing jobs.json entries with list-form deliver are handled
gracefully without requiring users to edit the file.
Closes#17139
The normalize_model_name() function unconditionally converted dots to
hyphens in all model names. This caused non-Anthropic models (e.g.
gpt-5.4) to be mangled to gpt-5-4 when routed through the Anthropic
adapter path, resulting in HTTP 404 from the backend.
Now only applies dot-to-hyphen conversion for models starting with
"claude-" or "anthropic/", which are the actual Anthropic model IDs.
Fixes NousResearch/hermes-agent#17171
Related: #7421, #13061, #16417
The plugin used to spawn one daemon thread per sync_turn() to do the
aretain_batch network write. On CLI exit, that pattern raced interpreter
shutdown — the last retain could reach aiohttp after asyncio's
"cannot schedule new futures" guard had fired, producing noisy logs and
silently losing the final unsaved turn:
WARNING ... Hindsight sync failed: cannot schedule new futures after
interpreter shutdown
ERROR asyncio: Unclosed client session
client_session: <aiohttp.client.ClientSession object at 0x...>
Switch to a single-writer model: each provider owns one long-lived
writer thread plus a queue. sync_turn() snapshots state and enqueues a
job; the writer drains sequentially. Once shutdown() is called:
- new sync_turn() / queue_prefetch() calls are dropped, not enqueued
- a sentinel wakes the writer so it finishes in-flight work
- shutdown joins the writer (10s) before nulling the client
Also register an idempotent atexit hook from the first sync_turn(), so
exit paths that don't go through MemoryManager.shutdown_all() (Ctrl-C,
abrupt exit) still get a chance to drain.
Tests: keep _sync_thread as a legacy alias to the writer, swap join()
calls to _retain_queue.join() (canonical wait-for-drain), add a new
TestShutdownRace suite covering single-writer reuse, post-shutdown drop,
queue draining, and shutdown idempotency.
