* fix(curator): defer first run and add --dry-run preview (#18373)
Curator was meant to run 7 days after install, not on the very first
gateway tick. On a fresh install (no .curator_state), should_run_now()
returned True immediately because last_run_at was None — so the gateway
cron ticker fired Curator against a fresh skill library moments after
'hermes update'. Combined with the binary 'agent-created' provenance
model (anything not bundled and not hub-installed), this consolidated
hand-authored user workflow skills without consent.
Changes:
- should_run_now(): first observation seeds last_run_at='now' and returns
False. The next real pass fires one full interval_hours later (7 days
by default), matching the original design intent.
- hermes curator run --dry-run: produces the same review report without
applying automatic transitions OR permitting the LLM to call
skill_manage / terminal mv. A DRY-RUN banner is prepended to the
prompt and the caller skips apply_automatic_transitions. State is
NOT advanced so a preview doesn't defer the next scheduled real pass.
- hermes update: prints a one-liner on fresh installs pointing at
--dry-run, pause, and the docs. Silent on steady state.
- Docs: curator.md and cli-commands.md explain the deferred first-run
behavior and warn that hand-written SKILL.md files share the
'agent-created' bucket, with guidance to pin or preview before the
first pass.
Tests:
- test_first_run_defers replaces the old 'first run always eligible'
assertion — same fixture, inverted expectation.
- test_maybe_run_curator_defers_on_fresh_install covers the gateway tick
path end-to-end.
- Three new dry-run tests cover state-advance suppression, prompt
banner injection, and apply_automatic_transitions skipping.
Fixes#18373.
* feat(curator): pre-run backup + rollback (#18373)
Every real curator pass now snapshots ~/.hermes/skills/ into
~/.hermes/skills/.curator_backups/<utc-iso>/skills.tar.gz before calling
apply_automatic_transitions or the LLM review. If a run consolidates or
archives something the user didn't want touched, 'hermes curator
rollback' restores the tree in one command. Dry-run is skipped — no
mutation means no snapshot needed.
Changes:
- agent/curator_backup.py (new): tar.gz snapshot + safe rollback. The
snapshot excludes .curator_backups/ (would recurse) and .hub/ (managed
by the skills hub). Extract refuses absolute paths and .. components,
and uses tarfile's filter='data' on Python 3.12+. Rollback takes a
pre-rollback safety snapshot FIRST, stages the current tree into
.rollback-staging-<ts>/ so the extract lands in an empty dir, and
cleans the staging dir on success. A failed extract restores the
staged contents.
- agent/curator.py: run_curator_review() calls curator_backup.
snapshot_skills(reason='pre-curator-run') before apply_automatic_
transitions. Best-effort — a failed snapshot logs at debug and the
run continues (a transient disk issue shouldn't silently disable
curator forever).
- hermes_cli/curator.py: new 'hermes curator backup' and 'hermes curator
rollback' subcommands. rollback supports --list, --id <ts>, -y.
- hermes_cli/config.py: curator.backup.{enabled, keep} config block
with sane defaults (enabled=true, keep=5).
- Docs: curator.md gets a 'Backups and rollback' section; cli-commands
.md table gets the new rows.
Tests (new file tests/agent/test_curator_backup.py, 16 cases):
- snapshot creates tarball + manifest with correct counts
- snapshot excludes .curator_backups/ (recursion guard) and .hub/
- snapshot disabled via config returns None without creating anything
- snapshot uniquifies ids within the same second (-01 suffix)
- prune honors keep count, newest-first
- list_backups + _resolve_backup cover newest-default and unknown-id
- rollback restores a deleted skill with content intact
- rollback is itself undoable — safety snapshot shows up in list_backups
- rollback with no snapshots returns an error
- rollback refuses tarballs with absolute paths or .. components
- real curator runs take a 'pre-curator-run' snapshot; dry-runs do not
All curator tests: 210 passing locally.
Prevents ghost sessions from accumulating in state.db when the TUI/web
dashboard is opened and closed without sending a message.
Changes:
- run_agent.py: Add _ensure_db_session() gate method, called at
run_conversation() entry. Remove eager create_session() from __init__.
Handle compression rotation flag correctly.
- tui_gateway/server.py: Remove eager db.create_session() in
_start_agent_build(). Add post-first-message pending_title re-apply.
- hermes_state.py: Extract _insert_session_row() shared helper (DRY).
Add prune_empty_ghost_sessions() for one-time migration.
- cli.py: One-time ghost session prune on startup. Fix _pending_title
to call _ensure_db_session() before set_session_title().
- hermes_cli/main.py: Guard TUI exit summary on message_count > 0.
- tests: Update test_860_dedup to call _ensure_db_session() before
direct _flush_messages_to_session_db() calls.
Closes: ghost session clutter in hermes sessions list and web dashboard.
Telegram's client does not display empty forum topics in the chat's
topic list. After createForumTopic succeeds, send a short pin message
into the new topic so it becomes immediately visible to the user.
Only fires for newly created topics (no thread_id in config yet).
Failure to send the seed is non-fatal (debug-logged, topic still works).
The bot-owner identity check inside OwnerCommandMiddleware was commented
out and replaced with a hardcoded `is_owner = True`, so any group member
could trigger allowlisted privileged commands (/approve, /deny, /stop,
/reset, /retry, /undo, /new, /background, /bg, /btw, /queue, /q) by
sending the slash command without @-mentioning the bot. The most severe
case is /approve: a non-owner could approve a dangerous tool call the
bot was waiting on the owner to confirm.
Re-enable the documented identity check (push.from_account ==
push.bot_owner_id) so only the configured owner can issue these
commands.
Adds a new top-of-sidebar docs page at /docs/user-stories that is a
masonry-style collage of 99 real user stories sourced from X/Twitter,
GitHub issues/PRs, Reddit, Hacker News, YouTube, blogs (Medium, Substack,
dev.to), podcasts, LinkedIn, GitHub Gists, and Product Hunt.
Every tile links to the original post/issue/video/gist where someone
described a specific use case: personal assistants, dev workflows,
trading bots, research briefs, family WhatsApp agents, Kubernetes
deployments, legal-domain self-hosted setups, and more.
- docs/user-stories.mdx: MDX entry mounting the collage component
- src/components/UserStoriesCollage: React component with category +
source filters, CSS-columns masonry layout, per-category accent colors
- src/data/userStories.json: source-of-truth dataset (force-added; the
root .gitignore's unanchored 'data/' rule would otherwise swallow it,
same reason skills.json is explicitly listed in website/.gitignore)
- sidebars.ts: link added at the top of the docs sidebar
Four callsites hardcoded Path.home() / '.hermes' with no HERMES_HOME
check, breaking Docker deployments and profile isolation (hermes -p):
- plugins/hermes-achievements/dashboard/plugin_api.py:
state_path(), snapshot_path(), checkpoint_path() bare-literal paths
- scripts/profile-tui.py:
DEFAULT_STATE_DB and DEFAULT_LOG defaults ignored HERMES_HOME
- hermes_cli/slack_cli.py:
except-Exception fallback for slack-manifest.json dump
- optional-skills/migration/openclaw-migration/scripts/openclaw_to_hermes.py:
--target argparse default
Use get_hermes_home() (with an ImportError shim for the standalone
scripts) or 'os.environ.get("HERMES_HOME") or str(Path.home()/".hermes")'
where importing hermes_constants is impractical.
E2E-verified: with HERMES_HOME=/tmp/x all three achievements paths and
both profile-tui defaults route under /tmp/x.
Salvaged from #18068 (original scope was broader mechanical cleanup
claiming 23 callsites were buggy; most were already respecting
HERMES_HOME via os.environ.get(key, default) — only these 4 had no env
check at all). Credit: @web-dev0521.
Two machine-readable entry points to the Hermes Agent docs:
/llms.txt curated index of every doc page, one link per page
with short descriptions. ~17 KB, safe to load into
an LLM context window.
/llms-full.txt every page under website/docs/ concatenated as markdown.
~1.8 MB. For one-shot ingestion by coding agents and
RAG pipelines.
Both files are also served from /docs/llms.txt and /docs/llms-full.txt
(Docusaurus serves website/static/ under baseUrl=/docs/). Some agents and
IDE plugins probe the classic site-root path; the deploy workflow now copies
both files to _site root so either URL works.
Conforms to the emerging llmstxt.org spec: H1 project name, blockquote
summary, short install command, GitHub link, then curated sections
mirroring the docs-site navigation (Getting Started, Using Hermes,
Features, Messaging, Integrations, Guides, Developer Guide, Reference).
Generated by website/scripts/generate-llms-txt.py. Wired into prebuild.mjs
so every 'npm run build' and 'npm run start' refreshes the files alongside
the existing skills.json extraction. Both outputs are gitignored (same
precedent as src/data/skills.json).
Descriptions in llms.txt are pulled from each page's frontmatter, so they
stay current automatically. All ~80 section slugs are validated against
the filesystem at generation time; an invalid slug would fail the prebuild.
Adds a proper feature page at user-guide/features/goals.md covering
the /goal slash command — Hermes' take on the Ralph loop shipped in
PR #18262. The slash-commands reference table had two table rows but
no narrative doc walking through the judge model, fail-open semantics,
turn budget, persistence, user-message preemption, or the aux-model
config override.
Adds a walkthrough example showing a multi-turn goal running to
completion, covers the two judge failure modes with how to recover,
and credits Codex CLI 0.128.0 / Eric Traut as prior art.
Also cross-links both slash-commands.md rows to the new page so
readers discovering /goal from the command reference can dive in.
The anyOf collapse in _repair_schema returned early, skipping the
nullable-strip and enum-cleanup steps. When a schema had anyOf
[{enum: [..., null, '']}, {type: null}] alongside a parent-level
'nullable: true', collapsing to the single non-null branch produced a
merged node that still had both 'nullable' and the bad enum values —
Moonshot would still 400 on it.
Fix: fall through to Rules 1/3 when the collapse produces a single
merged node; only return early for the multi-branch case (pure
anyOf preservation) or when there was no null branch to remove.
Adds a test that locks in the combined-case expectation.
When a schema node inside anyOf has enum values but no explicit 'type',
Rule 3 (enum cleanup) ran before _fill_missing_type, so node_type was
None and the enum was never cleaned. Moonshot then rejected the schema
with 'enum value (<nil>) does not match any type in [string]'.
Fix: reorder operations — fill missing type first, strip nullable,
then clean enum. This ensures enum cleanup always has a type to check.
Also fixes test expectation: empty string in enum is now correctly
stripped (Moonshot rejects it too).
Closes#16875
Add a standing-goal slash command that keeps Hermes working toward a
user-stated objective across turns until it is achieved, paused, or
the turn budget runs out. Our take on the Ralph loop — cf. Codex CLI
0.128.0's /goal.
After each turn, a lightweight auxiliary-model judge call asks 'is
this goal satisfied by the assistant's last response?'. If not, and
we're under the turn budget (default 20), Hermes feeds a continuation
prompt back into the same session as a normal user message. Any real
user message preempts the continuation loop automatically.
Judge failures fail OPEN (continue) so a flaky judge never wedges
progress — the turn budget is the real backstop.
### Commands
- `/goal <text>` — set a standing goal (kicks off the first turn)
- `/goal` or `/goal status` — show current state
- `/goal pause` — pause the continuation loop
- `/goal resume` — resume (resets turn counter)
- `/goal clear` — drop the goal
Works on both CLI and gateway platforms via the central CommandDef
registry.
### Design invariants preserved
- **Prompt cache**: continuation prompts are regular user-role
messages appended to history. No system-prompt mutation, no toolset
swap.
- **Role alternation**: continuation is a user turn, never injected
mid-tool-loop.
- **Session persistence**: goal state lives in SessionDB.state_meta
keyed by `goal:<session_id>`, so `/resume` picks it up.
- **Mid-run safety**: on the gateway, `/goal status|pause|clear` are
allowed mid-run (control-plane only); setting a new goal requires
`/stop` first so we don't race a second continuation prompt against
the current turn.
### Files
- `hermes_cli/goals.py` (new, 380 lines) — GoalManager + judge + state
- `hermes_cli/commands.py` — CommandDef entry
- `hermes_cli/config.py` — `goals.max_turns` default
- `hermes_cli/web_server.py` — dashboard category merge
- `cli.py` — /goal handler + post-turn continuation hook in
process_loop
- `gateway/run.py` — /goal handler + post-turn continuation hook
wrapping _handle_message_with_agent
- `tests/hermes_cli/test_goals.py` (new, 26 tests) — judge parsing,
fail-open semantics, lifecycle, persistence, budget exhaustion
- `website/docs/reference/slash-commands.md` — docs entry
* docs(sidebar): collapse exploding skills tree to a single Skills node
The Skills sub-tree in the left sidebar expanded to 200+ entries
(22 bundled categories + 15 optional categories, every skill a page).
That's most of the nav on a first visit — docs for the actual product
get drowned in it.
Collapse the sidebar to:
Skills
godmode (hand-written spotlight)
google-workspace (hand-written spotlight)
Bundled catalog (reference/skills-catalog — table of all bundled)
Optional catalog (reference/optional-skills-catalog — table of all optional)
Per-skill pages still generate and are still reachable at their URLs;
they're linked from the two catalog tables and from the Skills overview
page. They just don't appear in the left nav anymore.
sidebars.ts goes from 649 lines to 247. generate-skill-docs.py loses
the bundled/optional sidebar render helpers.
Also picks up incidental generator output drift on current main
(comfyui skill content refresh; 4 new skill pages for
devops-kanban-orchestrator, devops-kanban-worker,
productivity-here-now, productivity-shopify; two catalog refreshes).
These are what the generator produces on main today — keeping them
committed avoids the next docs build showing 'working tree dirty'.
* docs(sidebar): drop godmode and google-workspace spotlight pages
Keep the Skills sidebar node strictly principled: two catalog links,
nothing else. There was no rule for which skills got spotlight pages
and which got auto-generated pages — just that these two happened to
be hand-written first.
Both pages still build and are still reachable at
/docs/user-guide/skills/godmode and
/docs/user-guide/skills/google-workspace. They're linked from the
catalog tables and the Skills overview page.
Sidebar Skills node now:
Skills
├── Bundled catalog
└── Optional catalog
hermes update had two interactive [Y/n] prompts with no bypass:
1. Config migration (after new env/config options are added)
2. Autostash restore (when uncommitted work was stashed before pull)
hermes uninstall already has --yes/-y; mirrors that.
Under --yes:
- Config-migrate prompt → auto-yes, migrate_config(interactive=False)
so new config fields are applied but API-key prompts are skipped
(user runs 'hermes config migrate' later for those). Matches
gateway-mode semantics.
- Stash-restore prompt → auto-yes, git stash apply runs automatically.
Closes the 'can I hermes update -y, No ! Fix' gap reported by @murelux.
Adds opt-in auto-deletion for slash-command reply messages like
"New session started!", "Restarting gateway…", "Stopped.", and
YOLO toggles. After the TTL elapses the gateway calls the adapter's
delete_message; on platforms without a delete API (everything except
Telegram today) the TTL is silently ignored and the message stays.
Requested on Twitter by @charlesmcdowell — tool-call bubbles are useful
real-time, but system notices clutter the thread once the agent finishes.
Implementation:
- EphemeralReply(str) sentinel in gateway/platforms/base.py. Subclasses
str so existing 'X' in response / response.startswith(...) checks in
tests and call sites keep working unchanged; isinstance() still
distinguishes it for the send path.
- _process_message_background and both busy-session bypass paths
(in base.py) call _unwrap_ephemeral() on the handler return, send
the unwrapped text, and schedule a detached delete task when the
TTL > 0 AND the adapter class overrides delete_message.
- display.ephemeral_system_ttl (default 0 = disabled) in DEFAULT_CONFIG.
Handler can pass ttl_seconds explicitly to override.
- Wrapped the highest-noise return sites: /new, /reset, /stop,
/yolo on/off, /restart success + "already in progress". Draining
notices and /help output left as plain strings — those are
informational and users want to read them.
Backward-compat: default TTL 0 → no scheduling, no behavior change
for existing users. Platforms without delete_message silently no-op.
When the curator consolidates skill X into umbrella Y, any cron job
that listed X in its skills field would fail to load X at run time —
the scheduler logs a warning and skips it, so the scheduled job runs
without the instructions it was scheduled to follow.
cron.jobs.rewrite_skill_refs(consolidated, pruned) now updates jobs
in-place: consolidated names route to the umbrella target (dedup
when umbrella is already present), pruned names are dropped.
agent.curator._write_run_report calls it after classification,
best-effort so a cron-side failure never breaks the curator itself.
Results are recorded in run.json (counts.cron_jobs_rewritten + full
cron_rewrites payload), a separate cron_rewrites.json for convenience
when jobs were touched, and a section in REPORT.md.
Reported by @tombielecki.
DeepSeek V4 Pro tightened thinking-mode validation and rejects empty-string
reasoning_content with HTTP 400:
The reasoning content in the thinking mode must be passed back to the API.
run_agent.py injected "" at three fallback sites — the tool-call pad in
_build_assistant_message and both injection branches of
_copy_reasoning_content_for_api (cross-provider poison guard + unconditional
thinking pad). All three now emit " " (single space), which satisfies the
non-empty check on V4 Pro without leaking fabricated reasoning.
Also upgrades stale empty-string placeholders on replay: sessions persisted
before this change have reasoning_content="" pinned at creation time; when
the active provider enforces thinking-mode echo, the replay path now rewrites
"" -> " " so existing users don't 400 on their first V4 Pro turn after
updating. Non-thinking providers still round-trip "" verbatim.
Updates 9 existing assertions + adds 2 regression tests (stale-placeholder
upgrade, non-thinking verbatim preservation).
Refs #15250, #17400.
Closes#17341.
The user-visible /compress banner and the post-compression last_prompt_tokens
writeback both counted only the raw message transcript (chars/4). With a 15KB
system prompt and 30 tool schemas (~26KB), a 4-message transcript that looks
like ~45 tokens to the transcript-only estimator is really ~10.5K tokens of
request pressure — a 234x gap.
Two user-facing consequences:
- Banner shows 'Compressing … (~45 tokens)…' while compression is actually
firing on 10K+ tokens of real pressure, confusing users about why
compression triggered (reported by @codecovenant on X; #6217).
- Post-compression last_prompt_tokens writeback omits tool schemas, so the
next should_compress() check compares real usage against a stale
underestimate — compression triggers late, potentially past the model's
context limit on small-context models (#14695).
Swap estimate_messages_tokens_rough() for estimate_request_tokens_rough()
at every user-visible banner and at the post-compression writeback.
estimate_request_tokens_rough() already existed for exactly this purpose
and includes system prompt + tool schemas.
Touched call sites:
- run_agent.py: post-compression last_prompt_tokens writeback, post-tool
call should_compress() fallback when provider usage is missing
- cli.py: /compress banner + summary
- gateway/run.py: gateway /compress banner + summary
- tui_gateway/server.py: TUI /compress status + summary
- acp_adapter/server.py: ACP /compact before/after
Left intentionally alone:
- Session-hygiene fallback and the 'no agent' /status path in gateway/run.py
— no agent instance is in scope to query for system prompt/tools, and the
existing 30-50% overestimate wobble on hygiene is safety-accepted.
- Verbose-mode 'Request size' logging — informational only, already counts
system prompt via api_messages[0].
Also relabels the feedback line from 'Rough transcript estimate' to
'Approx request size' so the metric label matches what it actually measures.
Credits: diagnoses from @devilardis (#14695) and @Jackten (#6217);
user report @codecovenant on X (2026-04-30).
Closes#14695Closes#6217
When a user types /steer <text> on an ACP session that isn't actively
running a turn (and there's no interrupted-prompt salvage available),
_cmd_steer silently appended to state.queued_prompts and replied
"No active turn — queued for the next turn". That looks identical to
/queue output even though the user never typed /queue — @EddyLeeKhane
reported this as "/steer never works, gets queued instead".
Rewrite the payload to a plain user prompt before the slash-intercept
fires, matching the gateway's idle-/steer fallthrough in
gateway/run.py ~L4898.
`hermes update` ran the config migration (11 → 17) successfully then
crashed at `agent/skill_utils.py:340` during the post-migration
skill-config prompt. User @FlockonUS reported this on Twitter.
Root cause: `get_missing_skill_config_vars` in hermes_cli/config.py
only guarded the import of `discover_all_skill_config_vars`, not the
call. Any runtime exception inside the skill scan (malformed SKILL.md,
unreadable external skill dir, etc.) propagated up through
`migrate_config` and aborted `hermes update` after the version bump.
Wrap the call in try/except so skill-config prompting — which is a
post-migration nicety — can never block the migration itself.
The initial guardrail PR consolidated failure classification by pointing
display._detect_tool_failure at the new classify_tool_failure helper,
which was strictly broader: it flagged any JSON result with
"success": false / "failed": true / non-empty "error", plus plain-text
"traceback" and "error:" prefixes. That would uptick the user-visible
[error] tag on tools that return {"success": false} as a benign signal
(memory fullness, todo state, etc.) and feed the failure-streak counter
at the same time.
Restore display._detect_tool_failure to its pre-PR semantics verbatim.
Tighten classify_tool_failure (the guardrail's internal safety-fallback
used only when callers don't pass failed=) to match _detect_tool_failure
exactly, so the two never disagree. Production callers in run_agent.py
already pass an explicit failed= derived from _detect_tool_failure, so
the guardrail counter is driven by the same signal the CLI shows.
- Emit providers in CANONICAL_PROVIDERS order (matching hermes model)
with user-defined/custom providers appended after
- Remove digit quick-select (1-9,0) handler — inconsistent with
absolute row numbering and already removed from hint text
- Remove unused windowOffset import
_process_message_background snapshotted callback_generation from the
interrupt event at the TOP of the task — before the handler ran.
_hermes_run_generation is only set on the event by
GatewayRunner._bind_adapter_run_generation during
_handle_message_with_agent, which runs DURING the handler await. The
early snapshot always captured None, which then flowed into
pop_post_delivery_callback(..., generation=None) in the finally block.
In pop_post_delivery_callback, generation=None with a tuple-registered
entry (generation, callback) bypasses the ownership check — it pops and
fires the callback regardless of which run owns it. Result: a stale run
could fire a fresher run's post-delivery callback (e.g. a
background-review notification attributed to the wrong turn).
Fix: move the snapshot into the finally block, after the handler has
run and _hermes_run_generation has been bound to the current run.
Regression test added: simulates a stale handler at generation=1 and a
fresher callback registered at generation=2. Pre-fix: snapshot=None →
pop fires the generation=2 callback under generation=1's ownership
("newer" fires). Post-fix: snapshot=1 → pop skips the mismatched
entry, callback stays in the dict for the correct run to claim.
Verified: test FAILS on current main (captures "newer" in fired list),
PASSES with this fix.
Salvaged from PR #12565 (the callback-ownership portion only; the
/status totals portion was already fixed on main in 7abc9ce4d via #17158).
Co-authored-by: Oxidane-bot <1317078257maroon@gmail.com>
Widens #16528 to two sibling sites that had the same quoted-boolean
bug: a YAML string "false" (or "0", "no", "off") silently evaluated
truthy under bool() / if-check.
- gateway/run.py _load_show_reasoning: is_truthy_value wrap
- tools/skill_manager_tool.py _guard_agent_created_enabled: is_truthy_value wrap
- regression tests for both
SELECT in get_messages_as_conversation() was missing finish_reason, so
assistant messages round-tripped through replay (including /branch copies)
silently dropped the provider's stop signal. Adds it to the SELECT, restores
it on assistant rows, and locks it in with a round-trip test.
When running on a host with sudoers NOPASSWD configured for the current
user, interactive Hermes sessions were unnecessarily entering the
password prompt path before executing sudo commands. Outside Hermes,
`sudo -n true` exits 0 for that user.
Add `_sudo_nopasswd_works()` that probes `sudo -n true` and, when it
succeeds, lets `_transform_sudo_command()` return the command unchanged
with no stdin password. The probe:
- Is scoped to the `local` terminal backend only, so Docker/SSH/Modal
and other remote backends do not inherit host sudo state.
- Re-probes every call (no process-lifetime cache) so an expired sudo
timestamp cannot silently make a later command block waiting for a
password that Hermes never prompts for.
- Is bypassed entirely when `SUDO_PASSWORD` is configured or a cached
password already exists, preserving existing explicit-password flows.
Co-authored-by: Junting Wu <juntingpublic@gmail.com>
The fix for this bug (isinstance guard) was merged via commit 3ff9e010,
but test coverage was not included. Adding 4 tests:
- dict metadata with hermes keys (normal case)
- string metadata (bug case — previously caused AttributeError)
- None metadata
- missing metadata key
Proves token A's detected capabilities do not leak to token B after the
fix in the preceding commit. Before the fix this test would have seen
both tokens return token A's cached value.
_capability_cache was a single module-level dict shared across all
tokens. If the bot token rotates or multiple tokens are used in one
process, capabilities detected for token A would be returned for
token B, causing wrong schema gating and incorrect runtime behavior.
Replace the single Optional cache with a Dict keyed by token so each
token gets its own isolated capability entry.
