Self-augmentation loop

Goal

Autonomously identify, select, implement, and verify one improvement that makes the repository materially better. A report-only analysis or test-only run is not enough.

Required distinction

• Self-verification loop: verifies that the service/harness behaves as intended, including tests and QA. Deterministic gate: ./bin/issueops self-verify --seed=100 --target-score=95 --llm-eval=false --json; collect-all evidence pass: add --collect-all-steps --progress=jsonl.
• Self-augmentation loop: directly implements one needed feature, performance improvement, quality improvement, or documentation improvement, then verifies it with the self-verification loop.

Exit criteria

Completion requires all goals below to exceed target_score. The default target is 95. If any score is 95 or below, keep improving or report the blocker.

1. Improvement selection: use repo evidence and GENIUS_THINK.md to produce at least 10 candidates, then score value, feasibility, and risk.
2. Implementation: the selected candidate appears as an actual code, docs, or skill diff. Cosmetic-only changes do not count.
3. Verification and QA: targeted tests, QA checks, and self-verify pass.
4. Learning capture: record decisions, failure lessons, and next candidates in state or docs where appropriate.

Workflow

1. Baseline

   • Read the nearest AGENTS.md/CLAUDE.md, GENIUS_THINK.md, and skills/self-augment/SELF_AUGMENTATION.md when present.
   • Run or inspect ./bin/issueops self-augment --json for the current candidate curriculum.
   • Use ./bin/issueops self-augment --save-state --state-key self-augment-latest --json when the selected plan should become durable memory for the next cycle.
   • Use ./bin/issueops self-augment --cycles 3 --json to plan multiple improvement cycles at once (default 1).
   • Use ./bin/issueops self-augment lesson --lesson "..." --next-action "..." --json to store reusable Reflexion lessons.
   • Run a baseline self-verification loop when feasible; otherwise capture why it cannot run.

2. Candidate curriculum

   • Generate at least 10 concrete improvement candidates.
   • Use at least two GENIUS_THINK.md formulas, preferring problem reframing, innovative solution generation, thought evolution, and complexity resolution.
   • Score each candidate by impact, feasibility, novelty, risk, verification cost, and user value.
   • Treat candidates marked already_satisfied by the planner as audit history only; select the highest-value open candidate so repeated cycles keep moving to the next necessary improvement.
   • Prefer high-value, low-risk, reversible improvements over broad rewrites.

3. Select and implement

   • Choose one candidate whose expected score can exceed 95 after implementation.
   • Make small, reviewable diffs. Do not add dependencies unless the user explicitly asked or evidence shows they are necessary.
   • Preserve host-neutral core boundaries: shared behavior in Go contract/domain/application/ports, host-specific details in Codex/Claude/Omo/omp/opencode adapters.

4. Feedback and retry

   • Convert every failing test, QA issue, or design concern into a short Reflexion-style lesson.
   • Apply the lesson and retry until the goal scores exceed 95 or a hard blocker remains.

5. Verify

   • Run targeted tests for the changed behavior.
   • Run go test ./... -count=1, relevant golden tests, risk-tier QA checks (go vet ./... / go test -race ./... -count=1 when Go risk is present), skill validation, and build checks as applicable.
   • Finish with ./bin/issueops self-verify --seed=100 --target-score=95 --llm-eval=false --json when practical; add --collect-all-steps --progress=jsonl when every deterministic step must run after an earlier failure.

6. Capture

   • Store durable lessons only when reusable: issueops state, .issueops/; prefer self-augment --save-state for the selected candidate curriculum and self-augment lesson for reusable failure/QA/design lessons.
   • Final report includes selected candidate, implemented diff, goal scores, verification evidence, and remaining risk.

Reference

For the detailed loop contract and research rationale, read SELF_AUGMENTATION.md in this skill directory. The self-verification candidate catalog in skills/self-verify/CANDIDATES.md is the source of truth.