Long-Running Harness (Anthropic's durable-state loop)

Anthropic's empirical finding: "compaction isn't sufficient … even Opus 4.5 running in a loop across multiple context windows will fall short if it's only given a high-level prompt." Durable on-disk artifacts — not summarization — are what survive. This skill operationalizes their canonical harness. (Source: anthropic.com/engineering/effective-harnesses-for-long-running-agents; the full prescription is in plans/anthropic-loop-harness-playbook.md.)

The two failure modes this kills

1. Over-ambition — the agent tries to one-shot the app, runs out of context mid-feature, leaves it half-done. → fixed by one feature at a time.
2. Premature completion — a later session "sees that progress had been made, and declares the job done." → fixed by the features.json passes ledger (verify-before-flip).

The two-prompt split (same system prompt + tools; only the first user prompt differs)

Initializer (runs ONCE). Expand the high-level goal into the scaffold, then STOP — do no feature work:

• features.json — every feature seeded passes:false (the full outline of "done").
• init.sh — boots the dev server + runs a basic end-to-end smoke test (saves tokens every session).
• claude-progress.txt — the running work log (seed it).
• the initial git commit (the baseline to revert to).

Templates: templates/init.sh, templates/claude-progress.txt, and the verify-before-flip schema at ../loop-designer/references/features-json.template.json.

Coding agent (every session after). Runs the session-start ritual below, works ONE feature, leaves clean state. Identical prompt each session — the files, not the context, carry continuity.

The session-start ritual (put this in the coding-agent prompt / CLAUDE.md / loop.md)

1. pwd — confirm the only-editable directory.
2. Read git log + claude-progress.txt — reconstruct what happened, don't re-explore.
3. Liveness health-check: run ./init.sh, drive the app end-to-end (Playwright/Puppeteer MCP: real user action → observe result). If broken, fix that FIRST before any new feature.
4. Read features.json; pick the single highest-priority passes:false feature.
5. Implement that one feature only.
6. Verify against its steps end-to-end; flip passes:false→true only on success.
7. Commit with a descriptive message; append a claude-progress.txt entry; leave merge-clean state.

The features.json contract (the anti-premature-completion ledger)

• One object per feature: {category, description, steps[], verify{command,expected}, passes}.
• The coding agent may edit ONLY the passes boolean — never add, remove, or reword entries. Header rule (verbatim Anthropic): "It is unacceptable to remove or edit tests because this could lead to missing or buggy functionality."
• Flip to true ONLY after end-to-end verification, never when code is merely written.
• JSON, not Markdown — "the model is less likely to inappropriately change or overwrite JSON files."
• A "done" claim while any passes:false remains is a premature-completion violation (the Ralph-Wiggum guard). Now hook-enforced: fabric_stop.py emits a FEATURES GUARD advisory at every Stop when an in-scope features.json has passes:false items (file-gated → dormant when absent). Also check on demand with ~/.claude/bin/check-features-complete (exit 1 = features still failing), or wire it as a hard Stop block via --stop-hook mode (exit 2) per the script's header.

Durable state (disk > context)

• git = the revert/recover state machine (descriptive commits; revert bad changes to a working state).
• claude-progress.txt = human-readable session log (see template for the per-session schema).
• features.json = the machine-checkable completion ledger.
• (SDK/API path) mirror these under the memory tool's /memories and treat ASSUME-INTERRUPTION as law: anything not written to disk/memory is assumed lost on the next context reset.

Context economy (keep the loop inside the window)

Smallest high-signal token set; CLAUDE.md < 200 lines (it re-injects after /compact; nested ones don't); push detail into path-scoped .claude/rules/ loaded on demand. Strip harness components on each model upgrade — "every component encodes an assumption about what the model can't do; stress-test them."

Wire it to this setup

• Drive the per-session loop with native /loop (bare /loop → ~/.claude/loop.md) or the bespoke bin/autonomous-loop --run (preemptive budget caps).
• Use evaluator-optimizer for the per-feature verify (separate skeptical checker).
• Feed accept/reject into the CAPC ledger (probe_cost_per_accepted_change).
• Decision layer: loop-designer. Bespoke 2-agent generator: autonomous-skill / nonstop-agent.