Long Horizon Jobs

Long-Horizon Jobs

Progress must never live only in the conversation. Context limits kill sessions; SIGKILL kills processes. A job is durable only if a kill at ANY instant loses zero items and duplicates zero work on resume.

The contract — three files per job

Every long-horizon job owns a job directory (convention: data/jobs/<job-name>/) containing exactly:

Recovery invariant: on load, the completed set is rebuilt as the UNION of state.json.completed and the _hash fields found in findings.jsonl. A kill between the WAL append and the state write therefore cannot duplicate (the hash is recovered from the WAL) and cannot lose (the finding is already on disk). A torn final JSONL line from a mid-append kill is skipped, and its item re-runs.

Implementation

Python: scripts/lib/checkpoint.py (CheckpointedJob). Contract tests: scripts/tests/test_checkpoint.py (includes a 3x SIGKILL harness).

from checkpoint import CheckpointedJob

job = CheckpointedJob("data/jobs/sydecar-docs", "sydecar-docs")
for item in list_items(start=job.cursor):          # cursor bounds re-listing
    if job.is_done(item.id):                       # hash-set skip, O(1)
        continue
    try:
        result = process(item)                     # the expensive part
        job.record_finding(item.id, result)        # WAL append -> mark done -> checkpoint
    except Exception as e:
        job.record_error(item.id, str(e))          # error queue, attempts counted
    job.set_cursor({"last": item.id})
job.finish()

Rules:

1. Item keys must be stable across runs (URL, document id, post id) — never list indices or timestamps. The key is hashed (item_hash) for the completed set.
2. Do the expensive work between is_done and record_finding, nothing after. Side effects (file downloads) go to a deterministic path derived from the item key, written atomically (tmp + rename), so a re-run overwrites rather than duplicates.
3. Errors don't block progress. record_error queues the item and the loop continues; a later successful record_finding clears it. Retry the error queue at the start of the next run, bounded by attempts.
4. Never buffer batches in memory. One item = one durable write. If an API forces batching, keep the batch small and checkpoint per batch.
5. JS jobs follow the same file contract: fs.appendFileSync + fsync for the WAL, write-tmp-then-fs.renameSync for state/summary.

Agent resume protocol

When an agent (re)starts a long-horizon job:

1. Read SUMMARY.md ONLY. Do not re-read findings.jsonl, raw transcripts, task outputs, or prior conversation. The summary carries counts, cursor, and the error tail — everything needed to decide what to do next.
2. Launch/continue the worker process. The library skips completed items itself.
3. On context pressure or phase end, ensure the last checkpoint happened, then summarize position in one sentence. The next session starts at step 1.

The PostToolUse hook (~/.claude/hooks/lh-observer-digest.sh, registered in ~/.claude/settings.json) injects a checkpoint reminder every 10 tool calls — when it fires mid-job, flush state before continuing.

Memory observer

The memory-observer agent (~/.claude/agents/memory-observer.md) distills durable observations at phase boundaries. It reads ONLY the pre-truncated event digest at ~/.claude/observer/events/<session_id>.jsonl (each event capped at 400 chars by the hook; observer reads tail -c 24000, 6k tokens hard budget), emits at most 10 observations per batch, and appends them to `/.claude/observer/observations.jsonl` BEFORE replying. Never feed it raw transcripts; never raise its budget — bounded-and-lossy beats complete-and-dead.

Anti-patterns (each one has already cost us)

• Progress tracked as a number in the conversation ("done 340 of 900") — the KB Phase-1 backfill's escalating --limit values were the only record of progress.
• Re-listing and re-diffing the whole corpus every run instead of a cursor (KB ingest re-reads ~4,900 docs hourly; crash = full re-fetch).
• Buffering expensive results in RAM until a batch write (KB ingest's in-flight 200-doc batch discards up to 200 paid LLM calls on crash).
• Feeding an observer/summarizer untruncated tool output (claude-mem: unbounded JSON.stringify + never-compacted history = "Prompt is too long" on every batch, zero records captured).
• Rewriting a results file wholesale instead of appending (torn-write risk; newsfeed posts.json uses plain writeFile).
• Kill-testing skipped. A harvester is not done until it has been SIGKILLed mid-run at least 3 times and resumed with zero duplicates and zero losses (see test_sigkill_mid_run_resumes_exactly_once).