Flags and incremental state (all jobs, desktop and online)

Rule (Atul, 18 Oct): every job — desktop (collector, harness, local models such as Gemini or Ollama jobs) or online (Claude check-in, MMDEP Assistant, hosted CI fallback) — follows the heartbeat and does incremental work only: it reads its state, processes only what changed since its last run, and writes its state back, so no other job repeats it. A new job is not merged until it has a row here.

Shared flags

Flag	Where	Written by	Meaning
Heartbeat	control branch: heartbeat.json, HEARTBEAT.md	Collector (hourly, wake, unlock, reconnect)	ACTIVE while now < active_until_utc: the laptop is up and desktop jobs own laptop work
Laptop runner status	inside heartbeat.json	Collector, from GitLab	online / offline / unknown
Stop switch	control branch: harness/STOP	Atul or MMDEP-PM	Harness does nothing while present
Check-in late	control branch: heartbeat.json (checkin_late, checkin_last_utc), HEARTBEAT.md, and the MMDEP row of latest.md (check-in late: last <time or none>)	Collector, from the last commit to check-ins/ on reports (or docs/status/check-ins/ on main)	The Claude check-in (08:50 and 19:50 IST) has left no report for over 14 hours
Data source failing	control branch: data_sources.json (per ID: status, ms, content type, bytes, fails in a row), latest.md section "Data sources (daily)" and the MMDEP row's flags (data source <ID> failing <n> days in a row (<status>))	Collector, once a day, a read-only GET of each URL row in docs/registers/data-sources.md (#66)	A registry source failed (non-2xx or timeout) two daily checks in a row; one failure is shown but not flagged
Memory: sessions stale	control branch: memory.json, the repository's Flags in latest.md (memory: sessions stale)	Collector, every 6 h (backlog B14, #92), read-only	docs/sessions/ has no commit within 24 h of the repository's latest commit, or is missing; a repository it can't read shows "not visible" and no flag
Memory: CURRENT.md stale	as above (memory: CURRENT.md stale); counts of open decision-log and question-atul items in the section "Memory (rules §21)"	Collector, every 6 h (#92)	docs/knowledge/CURRENT.md is missing or older than the last commit outside docs/

Job state (cursor per job)

Job	Runs on	State file	Incremental rule
Collector	Laptop	control: latest.json, history/	Snapshot each run; daily and weekly checks reuse cached results (site.json, certs.json) until stale
Harness	Laptop	control: harness/state.json, ledger.csv, LATEST.md	Skips issues already pushed (waiting for review) and issues that failed until their body changes
Claude check-in	Cloud (scheduled)	main: docs/status/check-ins/state.json	Only issues, PRs, branches and panel answers changed since last_run_utc; run-jobs only when heartbeat ACTIVE
MMDEP Assistant	ChatGPT (attended)	reports: state/mmdep-assistant.json	Digest covers only items changed since last_digest_utc; ONLINE MODE only when heartbeat INACTIVE
Hosted CI fallback	GitLab / GitHub cloud	the pipeline itself + minutes in the digest	Only for security, release or critical items, only while heartbeat INACTIVE, within the group budget (DR-011)
Actions sync (#57)	Laptop, after each collector snapshot	reports: actions/state.json (last Atul answers comment handled, every decision and whether it was commented and reached main)	Only answer comments after the cursor; each decision commented and relayed once; a relay row lost before main is written again
Knowledge extract (B12)	Laptop (by hand or scheduled)	branch knowledge/<date>: docs/knowledge/state.json (hash per chat ID)	Only chats new or changed since their last extract; cost to harness/spend.csv, stops before the monthly cap
Future model jobs (triage, preflight, quota probe)	Laptop or cloud	control: jobs/<job>/state.json	Same pattern: read cursor, process new items, write cursor; check heartbeat first

State file shape

{"job": "mmdep-assistant", "last_run_utc": "2026-10-18T06:30+00:00", "cursor": {"issues_updated_after": "...", "answers_rows": 142}, "handled": ["MMDEP#21", "PE#337"]}