Starter kits

codex-completion-contract — the requirement-to-evidence audit

For long-running work where partial completion could be mistaken for done: defines every required outcome and its evidence before acting, then audits…

For long-running work where partial completion could be mistaken for done: defines every required outcome and its evidence before acting, then audits each requirement as proved, weak, missing, or contradicted. Complete only when everything is proved. Budget exhaustion never counts as success.

Category: I · Self-improvement / meta · Heartbeat: conditional (run-until-done, on-demand — only when explicitly asked) · Level: L1 (report-only) · Cost: Low

Source: Loop #28, "The Codex completion-contract loop," from Forward Future's Loop Library, original prompt by 3goblack (@Dis_Trackted).

Quickstart

1. Read LOOP.md — the six parts, the prompt, the limits, the three stops,
   and the source's verbatim prompt.
2. Only invoke this loop when explicitly asked for a completion audit —
   the source's own gate. Ask before filling in codex-completion-contract-
   state.md's "The Goal" section, then commit it.
3. Claude Code: run it in-session on the named Goal, pointed at
   .claude/skills/loop-task/SKILL.md. OpenCode: rename
   opencode.json.example → opencode.json, add the AGENTS.md line, run
   skills/loop-task.md.
4. Rehearse once, by hand, on a real (small) Goal before trusting it on a
   big one: "Run the codex-completion-contract skill now on [Goal]."
5. Grade the rehearsal with the loop-verifier agent — it fails any "proved"
   requirement it can't independently re-verify, and any run that calls
   exhaustion "complete." PASS → trust it on bigger Goals. FAIL → fix the
   skill, never the audit.

What's in this kit

File	Role
`LOOP.md`	the definition — six parts, prompt, limits, ownership, three stops
`codex-completion-contract-state.md`	the spine — the Goal, the requirement-to-evidence table
`loop-budget.md`	caps: 20 runs, 200k tokens — added by this library; the source specified none
`loop-constraints.md`	the constitution — never starts itself, never reports exhaustion as success
`loop-run-log.md`	append-only, one line per beat
`.claude/skills/loop-task/SKILL.md`	the procedure (Claude Code)
`.claude/agents/loop-verifier.md`	the read-only reviewer — re-checks "proved" claims against real evidence
`opencode.json.example` · `skills/loop-task.md`	the OpenCode twin — same skill text

Loop Ready checklist

Provable success condition — the source's own spec: every requirement proved, audit has no weak/missing/contradicted item
Run limit — 20 runs (added by this library)
Spine written first, committed — only with explicit approval, per the source
Report-only (L1) start — the catalog's own permanent target for this loop
Human gate placed — the source's own explicit "only when asked" gate, before the loop even begins
One log line per beat, no silent runs
Kill switch — loop-pause-all in loop-budget.md

How this relates to the course

This kit is close to a formalization of CLAUDE.md rule 12, "green ≠ done," and of Spec-Driven Development's stopping-condition-as-spec discipline (S3). The source's own vocabulary — proved / weak / missing / contradicted, and complete / blocked / stalled / exhausted — is more precise than this course's own three stops, and its single sharpest line — "budget exhaustion never counts as success" — is exactly the distinction this library draws everywhere between the Limit stop and the Success stop: they are never allowed to be confused for each other.

Full attribution: resources/sources.md.