ACTIVE-TASKS.md projection (facts ledger)

When activeTask.ledger is facts, the canonical task state lives in hybrid-memory category:project facts (same store as memory_store). ACTIVE-TASKS.md is an optional read-only projection for humans and agents.

Regenerate

openclaw hybrid-mem active-tasks render

Requires activeTask.ledger: facts. With the default markdown ledger, the file is the ledger—no render step.

Fast snapshot/render via gateway

The public gateway route /plugins/memory-public/active-tasks returns a DB-backed snapshot directly from category:project facts (no CLI cold start).

GET /plugins/memory-public/active-tasks returns active rows + projection status (stale, reasons, timestamps).
GET /plugins/memory-public/active-tasks?render=1 does a best-effort projection refresh before returning.
memory_store writes to category:project also trigger a best-effort ACTIVE-TASKS.md refresh when activeTask.ledger: facts; failures mark the projection stale.

Timestamp semantics

The projection must not invent “when work started” or “last touch” from the render clock.

Started uses fact fields in order: started, task_started, created_at (parseable), then the earliest SQLite createdAt among fact rows for that task entity.
Updated uses: task_updated, updated, updated_at, then the latest createdAt in the row.
If nothing can be resolved, the markdown shows Unknown. Those rows are treated as stale for activeTask.staleThreshold (no trustworthy “last update” time).

Prompt injection

Per-turn prependContext uses the same readable projection filters (excludeGenericTitle, titleMinChars, dedupeBy) and optional row caps (injectionMaxTasks or projection.maxRowsPerSection). Tasks are sorted non-stale first, then by relevance to the user message / session, then by Updated. Stale rows are detailed in the stale-warning block, not the <active-tasks> summary. All blocks share activeTask.injectionBudget (default 500 tokens).

Facts-backed default output also strips obviously placeholder next text such as Task [label] next: and suppresses non-actionable subagent placeholder rows (for example a subagent session key row with no real next step). This keeps stale orphan bookkeeping rows out of the default projection and prompt injection until an explicit reconcile/render pass closes them.

When plugin debug logging is enabled, facts-ledger selection and projection emit bounded diagnostics with fetched row counts, grouped entities / duplicate collapse, stale handling, rendered or injected task counts, and elapsed milliseconds. The detailed metrics stay on debug to keep normal production logs low-noise.

Sections and filters (`activeTask.projection`)

Key	Meaning
`mode`	`readable` (default): drop generic titles, optional dedupe. `full`: skip those filters; list everything (caps still apply if set).
`excludeGenericTitle`	Default `true`: omit rows whose title is only the placeholder “Project task”.
`titleMinChars`	If > 0, drop shorter titles (readable mode).
`dedupeBy`	`none` (default), `label`, or `normalizedTitle`.
`maxRowsPerSection`	Cap Active, Stale — revisit, and Completed separately; a short footer line counts omitted rows.
`sectioned`	Default `true`: Active (not stale), Stale — revisit, Completed. Set `false` for the legacy flat `## Active Tasks` layout.

Empty sections are omitted (no blank headings), except the all-empty case which shows a single Active block with No active tasks.

Operator playbook

Close or update work via memory_store / facts (set status to done, failed, etc., or update task_updated and narrative keys).
Run hybrid-mem active-tasks reconcile when subagent Session: references point at missing transcripts—rows can be completed automatically.
Run hybrid-mem active-tasks hygiene --dry-run to detect stale failed rows and duplicate normalized entities; use --apply to mark stale/superseded rows without deleting history.
Run hybrid-mem active-tasks render to refresh the markdown.

See also TASK-HYGIENE.md for heartbeat nudges and active_task_propose_goal.

Optional goal links

Project facts may include related_goal or goal_id on a task entity; the projection adds a Related goal line when present.