AIMEAT 1.17.0 — Tasks triage, Quality tab & per-agent concurrency

Everything since v1.15.0 (1.15.1 → 1.17.0). Companion: aimeat-crewai 0.3.8 on PyPI (separate tag).

[1.17.0] - 2026-06-01

Agent Tasks tab triage — long task lists are split into three buckets so the
working set stays small without pagination, plus on-demand search and a per-task
view of the memory entries a task produced. Full design:
docs/plans/agent-tasks-triage-plan.md.

Added

• Triage buckets — the Tasks tab now has Recent (non-terminal + recently
  finished), Keep (manually pinned), and Archive (manually archived +
  auto-archived old terminal tasks), each with a live count. Per-task Keep /
  Archive / Restore actions. One field triage ('kept'|'archived'|null) on
  AgentTaskRecord (SQLite + MongoDB); PATCH /v1/agents/{name}/tasks/{id}/triage.
• On-demand search — a 🔍 toggle reveals a search box (title + description)
  plus time chips (Today / 7d / 30d / All). GET /v1/agents/{name}/tasks gained
  bucket, q, updated_after, updated_before params and returns per-bucket
  counts for the tab badges. No pagination — buckets + search handle length.
• Auto-archive node config (admin Config tab) — tasks.auto_archive
  (default on) and tasks.archive_after_hours (default 24): un-triaged terminal
  tasks fall to Archive once older than the window. AIMEAT derives buckets from
  these; off = tasks stay in Recent until archived manually.
• Per-task memory entries — the expanded task lists the memory entries that
  belong to it: entries tagged task:<id>, the task's deliverableKey, and the
  live-status key prefix (agents.<name>.tasks.<id>.), deduped by key. Each entry
  is collapsible, and JSON values render as a structured key/value tree
  (indented nested blocks, type-coloured values) rather than raw text. (A
  tagging-aware runner populates the task:<id> entries; until then the
  deliverableKey + live key show.)

Changed

• Blur-title eye is hover-only in the Tasks list — the per-task "hide title"
  toggle now appears on row hover/focus instead of always being visible,
  de-cluttering long lists; rows whose title is already hidden keep their eye.

Fixed

• No more empty flash on live updates — tabs that refetch on the live-update
  signal no longer blank-then-repaint; they swap fresh data in place (first mount
  still shows the loading state, via a showSpinner option on the loaders).
  Applied across the agent Quality, Activity, Agent Config, Services tabs, the
  profile Knowledge, Packages, Notifications tabs, and the admin Memory,
  Stats, Packages, Cortex, Sharing Groups, Agent Tasks tabs.

Companion

• aimeat-crewai 0.3.8 (PyPI, versioned independently via the
  aimeat-crewai-v* tag) implements the worker pool that consumes the
  max_concurrent_tasks config from 1.16.2: run_crew_daemon runs EXECUTE tasks
  on a bounded thread pool, each with its own liaison/MCP, reading the value from
  the integration kit. Default 1 = serial (unchanged).

[1.16.2] - 2026-06-01

Follow-on to the Quality tab (1.16.0): an owner-facing way to rate from the UI,
richer rating metadata, and a new per-agent runner-concurrency config.

Added

• max_concurrent_tasks per-agent runner config — PATCH /v1/agents/{name}/max-concurrent-tasks (1–20, default 1) and an editable
  number field at the top of the agent Tasks tab. Default 1 = serial (safe
  for any engine, backward-compatible); higher values require a runner that can
  process tasks in parallel (e.g. a CrewAI daemon with a per-task liaison/worker
  pool). AIMEAT only stores and exposes the number — it does not enforce
  concurrency server-side. The runner reads it from the integration kit's
  watchdog_spec.max_concurrent_tasks; the field is also on the agents list
  (max_concurrent_tasks). Persisted on AgentRecord (SQLite + MongoDB).
• Owner "Rate deliverable" control — completed (done) tasks now show a
  Rate / Re-rate button in the Tasks tab (next to the deliverable) and in a
  new "Rate deliverables" list on the Quality tab (unrated first). Opens a
  shared modal (1–5 stars, context, source-grounded flag, comment) that calls
  POST /v1/agents/{name}/tasks/{id}/rate; the per-context rollups refresh on
  submit.
• Optional metadata on a rating — free-form evaluation context
  (temperature, top_p, max_tokens, tokens, cost, …) stored on the rating for
  later slicing (size-capped at 4 KB serialized).

Fixed

• Quality tab custom metrics are now read from the agent's GAII namespace
  (where an agent's own memory writes land) instead of the owner GHII, so
  agent-published agents.<name>.statistics.custom.* entries actually surface.

[1.16.0] - 2026-05-31

Agent Quality tab — per-context peer/owner reviews of task deliverables,
turned into a quality picture alongside recomputed performance statistics.
Reviews attach to tasks (not the work/actions system) and carry a quality
context dimension, so an agent's strengths and weaknesses surface per area
(e.g. creative: 4.5★ / factual: 2.1★). Full design and rationale:
docs/plans/agent-quality-tab-plan.md.

Added

• POST /v1/agents/{name}/tasks/{id}/rate — review a completed (done)
  task's deliverable with a 1–5 star rating and a context
  (factual/creative/code/planning/summarization/research/communication/other).
  Authorization: the task's owner, or a same-owner agent (e.g. a parent
  orchestrator that delegated the work) — an agent cannot rate its own
  deliverable. Logged as a rating task event.
• Source-grounding hard gate — for the factual family
  (factual/research/code/summarization) an agent rater must set
  source_grounded: true (the rating was checked against the inputs/sources),
  otherwise the request is rejected with 422 GROUNDING_REQUIRED. Human owners
  are exempt; creative accepts an output-alone craft rating. Stops stars from
  rewarding confabulation over faithfulness.
• GET /v1/agents/{name}/statistics — recomputes a performance rollup (task
  counts, success rate, completion durations by context) and a per-context
  review rollup (stars, distribution, variance, low-confidence flag below n=10,
  rater mix, per-model slice) from the agent's tasks — anyone can recompute,
  so the result is not forgeable. Written to the owner's public cache keys
  agents.<name>.statistics.performance / .reviews, and surfaces any custom
  metrics published under agents.<name>.statistics.custom.*.
• Quality tab in the expanded agent view (profile → Agents → agent →
  Quality) — performance cards + durations by context, per-context star bars
  with sample size and a low-confidence flag, and a custom-metrics list. Quality
  reads statistics + reviews; the existing Activity tab (event log / pulse)
  is unchanged.
• Optional metadata on a rating — free-form evaluation context
  (temperature, top_p, max_tokens, tokens, cost, …) stored on the rating for
  later slicing (size-capped at 4 KB serialized). evaluated_model is a
  first-class field so stars are not compared blindly across model changes.
• AgentTaskRating on the task record (SQLite + MongoDB), surfaced on the
  AgentTask API schema.

Notes

• New E2E suite test/e2e-agent-quality.ts (15 checks: happy paths, grounding
  hard gate, self-rating + non-done + validation rejections, rollup, public
  cache) — green on SQLite and MongoDB.

[1.15.1] - 2026-05-31

Added

• Cooperative task cancellation — owners can Cancel an active/stalled
  task from the Tasks tab. It writes an owner-scoped agents.cancel.task.<id>
  marker (a same-owner runner self-skips before its next kickoff) and natively
  pauses an active task for immediate effect — a circuit breaker for abandoned or
  over-delegated subtasks. (The aimeat-crewai daemon honours the marker as of
  its 0.3.7 release.)
• aimeat-agents.js browser library — commission and observe AIMEAT agents
  from a web page (new served lib-agents.js helper + demo HTML).

Fixed

• Integration tab no longer flashes on live updates — a showSpinner option
  on its loader stops the tab from re-rendering its loading state every time
  server-side data changes (first mount still shows it). (Extended to the rest of
  the tabs in 1.17.0.)