api and worker processes each keep adapter state in memory. After the
upstream reaped an idle tab and the api process reopened one, the agent
worker kept pointing at the lost tab and every screenshot / snapshot in
solve_challenge failed with not_found (REAL-2 task 4). loaded() now
restores from the database when its current tab differs. Needs image
rebuild.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
With only a browser.solve_challenge grant, the vision child run was
denied browser.click_at, so the loop stopped after the first screenshot
(REAL-2 task 4). Steps inside a run spawned by solve_challenge are now
allowed; ancestors' allowlist and step budget still apply. Needs image
rebuild.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Built-in agent checks kept the config stored at first registration, so
check.pending never received decision.list / subscription.list added in
AG-3 and the patrol could not read pending items (REAL-2 scenario g).
Track config_customized (migration 0045); sync replaces config with the
code default unless the user PATCHed it. Needs image rebuild.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Uses the desktop iframe when the backend reports it available and the page is
a secure context; otherwise falls back to screenshot mode and says why.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
/api/v1/browser/desktop/{sessionId}/* proxies Selkies HTTP and WebSocket,
console sessions only. The holder of the active human lease gets the control
password, everyone else the view-only one; the WS bridge rechecks every 3s
and drops the connection on role change so the client reconnects downgraded.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Selkies baseimage (debiantrixie, digest-pinned) with camofox 1.17.0 copied
from camofox-browser:1.17.0, run headed under labwc so the desktop shows the
same Firefox the agent drives. s6 service waits for wayland-0 and depends on
svc-de and svc-dbus so shutdown persists login state. Selkies control/view
passwords derive from CAMOFOX_ACCESS_KEY. Compose defaults to the new image;
CAMOFOX_IMAGE=camofox-browser:1.17.0 switches back. Optional caddy (profile tls).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- tool results carry a budget reminder near limits; text before the last tool
call is dropped as report when cut by budget/abort
- usage persisted when a run suspends, so waiting sessions show spent tokens
- task page shows budget_exceeded via business.outcome instead of completed
- budget-exceeded decisions merge by root cause (decision / schedule / goal
template); new summaries added to investigate allowlist
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- three read-only summary capabilities (38 registered); task.inspect points
download task ids to download.inspect
- probe queries reject empty / single-letter / wildcard-only terms; externalId
documented as string; every description says when to use it
- download.plan requiresApproval uses coveringGrant (shared with authorize)
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- browser.screenshot returns the page as an image content block (≤1280px wide, JPEG quality
configurable) for vision-capable upstreams; non-vision upstreams get a redacted text snapshot.
Login pages / password fields are marked sensitive and withheld unless a
browser.sensitive_screenshot grant exists and allowSensitive is passed.
- browser.click_at / type_at / scroll / wait act in screenshot pixels; the server converts via
viewport/DPR, requires automation control and an allowlisted host, and writes a thumbnail
evidence file (DATA_DIR/tmp/evidence) referenced from agent_run_event.
- browser.solve_challenge runs a bounded child agent_run (vision caps only, maxSteps ≤ 12,
budget), judges success from page facts, stops immediately on 2FA / SMS / device confirmation,
and otherwise moves the session to waiting with evidence and a takeover link.
- PiAgentRuntime passes modelImage outputs as pi-ai image content when model.input has image;
the ledger never stores image data; coordinate actions use per-invocation idempotency keys.
- Tests: image codec, runtime image content, E14 e2e (login solved, 2FA stop, step limit,
sensitive withheld). Real camofox 1.17.0 check on example.com recorded in the runbook.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- POST /pi/upstreams/{id}/models/refresh calls the provider's model list
endpoint (OpenAI-shaped /models, Anthropic /v1/models, Google
/v1beta/models, OpenRouter /api/v1/models; Radius via pi-ai refresh),
normalizes {id, displayName, contextWindow, supportsVision,
supportsTools, pricing} and stores it in pi_upstream.models (0044)
- GET .../models returns {source, fetchedAt, models[]}: remote merged
with the static catalog (remote first); failures (401/network/
unsupported) fall back to static with a one-line error
- POST /pi/models/probe for trying credentials before creating
- wizard and edit drawer: fetch button, searchable dropdown with
vision/tool tags, manual id, source line; default upstream needs a model
- tests: mock /models, 401, network failure; smoke step; runbook
Co-Authored-By: Claude Opus 5.5 <[email protected]>
When download.recheck / task.retry was executed but errored (e.g. the
downloader was unreachable), applyRecoveryOutcome only looked at
succeeded actions and recorded outcome=not_attempted with the summary
"没有执行恢复动作", which is false. Record outcome=failed and name the
failed action instead. Needs image rebuild to take effect.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
torrentSource threw plain Error for a missing site or a site without an
adapter definition, so submitDownload retried to max attempts, left the
plan and slot reserved, and sweepReservedPlans re-enqueued it forever
(agent runs stayed waiting on goal-settled). Throw SiteError
(not_found / unsupported) so the plan fails and the slot is released.
Needs image rebuild to take effect.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The search page 订阅 button and the new-subscription wizard threw away
the TMDB id and created a title-only subscription, which stays pending
(metadata_pending) with no seasons, episode totals or coverage matrix
until some release happens to be identified. Call
/files/entity-from-candidate first and subscribe the entity; fall back
to title tracking only if that fails. Picking a TMDB row in the wizard
now creates the library entity even if the dialog is then cancelled
(idempotent per TMDB id).
Found in REAL-1b on the v0.2.2 test stack; needs an image rebuild to
take effect.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
catalog.lookup only returned the zh-CN localized title, so an English
release name (Dune.Part.Two.2024...) was rejected against "沙丘2" and
files/identify came back with zero candidates. Pass originalTitle
through catalog.lookup and use it as an extra title in ranking.
Found in REAL-1b on the v0.2.2 test stack; needs an image rebuild to
take effect.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- compose: camofox service on by default (internal only, volume umt-camofox,
healthcheck, CAMOFOX_REPLICAS=0 to skip); app depends on it with
service_started/required=false and auto-registers camofox-env from
BROWSER_PROVIDER_URL / BROWSER_ACCESS_KEY / BROWSER_ENABLED
- docker/camofox/Dockerfile: builds upstream tag v1.17.0 (389c996); fixes the
upstream Dockerfile (missing postinstall.js before npm ci, sh -c not
forwarding SIGTERM so logins were never persisted on restart)
- adapter fixes from the real container: coordinate click via evaluate +
selector, focus typing with mode=keyboard, password values redacted from
ARIA snapshots/extract, downloads field names, 410 after browser restart
treated as lost tab; navigate reopens a tab in the same session
- verified-capability table keyed by upstream version (1.17.0 all ten);
other versions stay unverified; probe reports auth failures
- settings: browser providers block (create/edit/delete/probe, capability
table); API POST/PATCH/DELETE /browser/providers, POST .../probe
- scripts/browser-verify: umtbr stack + mock NexusPHP run of REQ 4.7.4 items
- docs: runbook browser §1/§7 (real-instance record), deploy §2.1
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- GET /system/health splits processes into current instances (hosts with
a beat in the last 10 min, newest per host+kind, not stopped) and
retiredProcesses[]; 45s stale and level only consider current ones
- heartbeat writer marks stopped_at on graceful exit (migration 0043),
upsert clears it; pure api role now stops its writer on SIGTERM
- scheduled job system.heartbeat_cleanup (maintain, 1h) deletes rows
without a beat for 24h and samples older than 7 days
- status page: per-role alarms ("io worker 45 秒没有心跳"), collapsible
retired instances; glossary/runbook updated
- bump all members to 0.2.2, CHANGELOG [0.2.2]
Co-Authored-By: Claude Opus 5.5 <[email protected]>
shadcn (web devDependency) pulled a second peer copy of
@modelcontextprotocol/[email protected].1 (zod@3), so the server's zod@4 copy
lived in the npm cache as sdk/1.30.1_1. `deno compile --exclude-unused-npm`
embeds that _1 folder but the pruned snapshot resolves it as sdk/1.30.1,
so every SDK import failed at startup with ERR_MODULE_NOT_FOUND.
- drop the shadcn CLI dependency; vendor its tailwind.css (the only
runtime use) as apps/web/app/shadcn-tailwind.css, removing the duplicate
- scripts/check-compiled-binary.sh: run the compiled binary with an empty
env and `--role preflight`; it must load the full module graph and exit 2
on config validation. Wired into `deno task ci` (check:binary) and the
Dockerfile build stage.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
GET /img/tmdb/{size}/{file} ported from MP ImageHelper / fetch_image
(sanitized cache path under DATA_DIR/cache/images, GLOBAL_IMAGE_CACHE
read switch, md5 ETag / 304, day-based cleanup job). Cache module adds
cache.sweep / cache.images_clean jobs and /cache/tmdb, /cache/images
list and clear endpoints. Library list/detail and the search palette
show posters with the neutral placeholder as fallback; metadata settings
get TMDB cache and image cache cards. Runbook docs/runbook/cache.md and
porting notes docs/porting/moviepilot-tmdb-cache.md.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Migration 0042_cache adds UNLOGGED kv_cache. PgCache mirrors MP
runtime/cache.py (region, per-key TTL, TLRU maxsize eviction, @cached as
cachedCall); only the storage moves from memory to PG. TMDB request
responses (MP TMDb.request @cached) and recognize results (MP TmdbCache:
same key, {id:0} negative cache with EMPTY_RESULT_CACHE_TTL, type-conflict
checks, backfill after manual assignment) live in kv_cache, shared by all
processes. Settings cache.tmdbTtlHours / tmdbMaxSize / emptyResultTtlSec
map to META_CACHE_EXPIRE / CONF.tmdb / EMPTY_RESULT_CACHE_TTL. TMDB
details now also store posterPath / backdropPath.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Migration 0041_rate_limit adds rate_limit_bucket. PgTokenBucket (@umt/db)
refills and takes tokens under SELECT ... FOR UPDATE on the DB clock and
polls with jitter within a wait budget; penalize() carries Retry-After.
Site adapters get createSharedSiteRateLimiter (key site:<id>), shared by
api / worker / worker:agent, with an in-process serial queue in front and
a local fallback when the bucket is unavailable.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A suspended agent.propose_options / approval call is replayed on resume with the
same toolCallId; foldEvents pushed a second item with the same key, which also
inflated the tool call meter. Also use 「」 in the report footnote.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Auto-fix of 602 string literals across 74 files (half-width punctuation next to
Chinese, parenthesised Chinese, quotes); biome reflow; e2/e6 assertions follow
the new full-width summaries. check:copy now runs with --strict in ci.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
