subagenttasks

.com the task primitive
3 routines, live from D1

Routines

Standing work this ecosystem already runs on a schedule -- real recurring jobs, not invented ones. See the grounding for what each column means.

subagentjobs-cron board crawl + eviction

cron schedule schedule: 0 */6 * * * (every 6 hours, Cloudflare Cron Trigger)

Crawls 73 Greenhouse + 7 Lever + 1 Ashby job boards, upserts fact_job_posting + dim_board via ON CONFLICT, then runs a 3-way platform-dispatch eviction pass to remove stale postings.

scope

workers/cron (subagentjobs-cron Worker) — runs independently of any interactive session

connections used

D1 write access to subagentjobs-dwh; outbound HTTP to each ATS board's public API

retry policy

Not documented in this repo today — unlike proactivity.md's claude-tag routines (skip-and-retry, then disable after three consecutive failures spanning an hour), workers/cron has no recorded retry-then-disable policy of its own. An honest gap, not silently assumed to exist.

grounded in

CLAUDE.md, "workers/cron"; workers/cron/wrangler.toml, "[triggers] crons"

mcp__scheduled-tasks session routines

session tool call schedule: Cron expression or one-time fireAt, set per-task via create_scheduled_task

Runs a described piece of work automatically later — repeating ("every morning", "daily at 6am") or once at a specific future time — the same shape proactivity.md's "describe the schedule you want and the work Claude should do in one message" pattern.

scope

A Cowork session (this exact session has this tool available)

connections used

Whatever connectors/tools the session itself has access to when the scheduled run fires

retry policy

Not independently verified in this repo — the tool exists and this session can call it, but no failure/retry semantics were inspected before writing this row.

grounded in

This session's own mcp__scheduled-tasks tool description ("create_scheduled_task", "list_scheduled_tasks", "update_scheduled_task")

citation-service-loop heartbeat pipeline

heartbeat dispatch schedule: Dispatched on a heartbeat tick, not a fixed cron expression

Orchestrates crawl:vendors → ingest:citations → cache:warm → dw:load → dreams:consolidate as one chained 24/7 pipeline, emitting DurableTasks to engineering.jsonl on any step failure.

scope

The knowledge-engineering sister repo's cowork chassis (see CLAUDE.md, "Sister repo: knowledge-engineering")

connections used

durable-lru-dreams (dreams:consolidate step), semantic-cache (cache:warm step)

retry policy

Failures are captured as DurableTask envelopes in engineering.jsonl rather than silently retried — closer to proactivity.md's "the routine's page shows why" transparency requirement than workers/cron's undocumented behavior.

grounded in

This repo's available skill catalog, "citation-service-loop" description