Skip to content

aiskillstore/marketplace / skills/aakashh242/gauntlet

gauntlet

Verified

Evidence-driven adversarial review and root-cause repair of code, designs, APIs, infrastructure, tests, and agent workflows. Use for deep reviews, red-team reviews, edge-case audits, production-readiness checks, or requests to find and fix defects iteratively. Supports sub-agents or a tracked solo workflow. Not for ordinary implementation or cosmetic edits unless explicitly invoked.

references/closure.md

View on GitHub
# Closure, recursion, and stop conditions

## What recursion means here

After a confirmed defect, revisit the causal mechanism, sibling sites, the repair, and any affected assumptions or contracts. This is bounded iteration over an impact graph, not infinite self-reflection or recursive agent spawning. The purpose of another round must be a new test, newly affected surface, contradictory evidence, or a distinct attack angle.

Track the number of unique confirmed mechanisms, unresolved findings, failing/blocked required checks, changed surfaces, and the clean streak. Do not invent numerical confidence from the number of agents or the number of rounds.

## Valid clean re-attack

All selected tasks have evidence for the final snapshot, required checks have passed using the agreed method, all candidates are resolved, all repairs are verified, and no new substantiated defects were found during the re-attack. It must not be the round that modified the target. For two required clean rounds, use different falsification strategies; rerunning the same checklist with synonymous instructions does not count.

A dismissed hypothesis may occur in a clean round when concrete evidence rules it out. A verified new defect still makes its discovery round non-clean, even if repaired quickly. Any subsequent target change resets the streak. An accepted low/medium exception can permit `PASS_WITH_EXCEPTIONS`, never an unqualified pass. Critical/high findings block completion in the helper.

## Outcomes

**REVIEW_COMPLETE:** A read-only review has completed its selected tasks and checks; confirmed findings may remain by design. State findings and hypotheses separately. Do not imply the artifact is ready to release. In the helper, unresolved candidates and unavailable required checks still prevent complete status.

**PASS_WITHIN_SCOPE:** Repair mode, no unresolved candidates or defects, every fix verified on the final snapshot, required clean streak achieved. Means the declared gates were met, not all bugs absent.

**PASS_WITH_EXCEPTIONS:** Same scoped verification, with specifically approved lower-risk residual findings. Include owner, approval reference, rationale, and consequences. An agent cannot self-approve a risk acceptance.

**BLOCKED:** A needed decision, permission, environment, unresolved defect, incomplete verification, or missing evidence prevents completion. List exact next actions and continue independent useful work while budget remains.

**BUDGET_EXHAUSTED:** Maximum rounds or another agreed budget reached without meeting gates. Stop, persist state, and show what remains. Never start another batch merely because the skill says “deep.” An extension needs explicit authorization, a new maximum, and a concrete next strategy.

## Oscillation and diminishing returns

After two failed repair attempts on the same mechanism, revisit the oracle, requirements, and root-cause hypothesis. A fix that alternates between two regressions is not progress; preserve both counterexamples. Concurrency fixes may need a model or a storage-boundary invariant rather than another timing change. If reviewers disagree without evidence, design an adjudicating experiment or retain the uncertainty.

If two rounds produce no new evidence and remaining blockers are external, stop early rather than spending the full budget. If clean criteria are met early, stop early. Larger models or more reviewers are optional choices subject to budget, privacy, and host support, not automatic escalation rights.

## Final artifact and handoff

Write a short verdict plus scope/revision, review lenses, verified fixes with root causes, confirmed remaining findings, tests/checks actually run, unavailable checks, accepted risk references, excluded surfaces, and next action. Link evidence rather than embedding huge logs. Include the tracker location and current round for resumption. Keep the difference between a locally verified patch and a deployed fix explicit.
rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first.Stars: 194kUpdated Claude CodeCodexVerifiedInstall