Checkpoint and resume an expensive research workflow

Multi-agent research runs (deep-research and similar Workflow fan-outs) are the costliest thing a session does, and any of the account's usage limits can run out mid-run — the 5-hour session window, the weekly all-models allowance, or the weekly per-model (e.g. Fable) allowance. This applies to all three meters equally. A weekly meter can force a multi-day pause, so the checkpoint must be durable enough to survive it.

.claude/rules/workflow.md rule 12 owns the standing policy; this skill holds the mechanics.

Standing practice

• Checkpoint at launch. When launching a long research workflow, immediately save a checkpoint file (scratchpad or session dir) with the run ID, script path, and the exact args string, so a later turn — possibly after context summarization — can resume without reconstructing anything.
• Stop before the limit, not after. If the maintainer pauses for cost or any of the three meters is near its limit, stop the run (TaskStop) rather than letting it burn into overage. This is safe: every completed agent is journaled (journal.jsonl in the workflow transcript dir) and replays from cache at zero cost; a stop loses only in-flight agents.
• Resume from cache, byte-identical. Resume with Workflow({scriptPath, resumeFromRunId, args}) where args is byte-identical to the original — omitting args aborts instantly without touching the cache (Jul 2026: a resume without args died with "No research question provided"), and changed args miss the cache and re-pay for everything. Same-session only.
• Run the tail cheap. The resumed remainder (final verifies, synthesis) inherits the session model — switch to the cheapest model adequate for those stages before resuming.
• Never re-run whole when a journal exists. Before diagnosing an empty or failed result as lost work, read the journal — the cached results are usually all there.
• Never promise a threshold stop — pick a mechanism before ending the turn. Claude cannot see any of the usage meters and does not run between turns, so "I'll stop it at N%" is unenforceable (Jul 18 2026: promised a 97% stop with no mechanism; the maintainer had to stop the run manually at 98%). When the maintainer names a threshold on any meter or reports one running hot, do one of two things before ending the turn — and say which one was done: TaskStop the run immediately (the journal cache makes stopping cheap), or arm a timed one-shot stop sized to the remaining tail (send_later, e.g. 15–20 minutes out: on fire, stop the run if it's still going and re-checkpoint). A stated intention with no armed mechanism is the failure mode, not a plan. Model choice is a lever here too: the weekly per-model meter drains only for that model's usage, so running research tails on a cheaper model spares the premium meter for judgment work.

Related

Before committing to any expensive/slow/unattended run at all, ~/.claude/CLAUDE.md's cheaper-and-faster-frontier gate applies first: present the frontier options (speed + cost + quality + Net) up front, and treat a large asymmetry against an existing cheaper path as a hard stop.