Evidence-driven SPEC -> BUILD -> REVIEW -> REPAIR loop for coding agents.
SpecLoop helps an agent deliver software changes against an explicit specification, with recorded evidence, verification, review scoring, and bounded repair iterations. It is intended for implementation, fix, refactor, migration, and audit tasks where "done" must mean requirements and project checks actually passed.
- Discovers the repository state, relevant instructions, conventions, tools, and checks before editing.
- Writes or resumes a task specification with stable requirement IDs.
- Builds only against the approved scope.
- Records evidence that maps requirements to files, behavior, checks, and remaining uncertainty.
- Reviews the real project state against hard gates.
- Repairs failed findings until
SPECLOOP PASSor an honest blocked state.
Use the specloop folder as the skill path.
python3 ~/.codex/skills/.system/skill-installer/scripts/install-skill-from-github.py \
--repo ivankuprin/SpecLoopSkill \
--path specloopRestart Codex after installation so the skill is discovered.
Use $specloop to implement this feature from spec to passing review.
Use $specloop to fix this bug end to end, including tests and repair iterations if review fails.
Use $specloop to audit and repair this migration until required checks pass.
SpecLoop finishes with exactly one of:
SPECLOOP PASSSPECLOOP FAILSPECLOOP BLOCKED
PASS requires all hard gates to pass. A score alone cannot override a failed requirement, failed acceptance criterion, failed required check, incomplete evidence, unresolved blocker, critical or high finding, functional regression, or unapproved scope change.