--- status: resolved parent: "[[038-implement-skill-map]]" claimed-by: "019fbb66-80d0-7a01-8ac9-01255adfc7b2" claimed-at: "2026-07-31T23:46:29-04:00" blocked-by: [] tags: - ticket/research --- # Implement skill no mistakes research ## Question What should the migrated `implement` skill learn from Kun Chen's `no-mistakes` code review workflow? ## Findings Kun Chen's `no-mistakes` is a local git proxy and validation gate. Its README says pushing to the `no-mistakes` remote spins up a disposable worktree, runs review, tests, docs, lint, push, PR, and CI, and forwards the branch only after every check is green. The migrated `implement` skill should consider separating local implementation from final shipping validation, rather than treating one in-context review pass as the whole quality gate. Source: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/README.md`. `no-mistakes` is explicitly agent-agnostic and supports agents including `pi`. Its skill drives `no-mistakes axi`, a non-interactive TOON interface, and the pipeline can apply safe fixes while escalating intent-touching findings to the human. A local `implement` skill can remain tool-agnostic while optionally detecting and delegating to `no-mistakes` when installed and initialized. Sources: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/README.md`, `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/skills/no-mistakes/SKILL.md`. The `no-mistakes` skill distinguishes validate-only mode from task-first mode. Task-first mode checks scope with `git status`, preserves unrelated pre-existing uncommitted changes, commits only task changes, creates a feature branch when the user is on the default branch, then validates committed work. The migrated `implement` skill should preserve this strict scope boundary and should treat committed branch state as the input to any downstream validation gate. Source: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/skills/no-mistakes/SKILL.md`. `no-mistakes` requires a rich `--intent`, defined as the user's objective in their terms plus decisions, tradeoffs, constraints, and approaches ruled in or out. The migrated `implement` skill should derive review and PR context from the Wayfinder task's `## What to build`, acceptance criteria, parent decisions, and implementation notes, not merely from the diff or file list. Source: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/skills/no-mistakes/SKILL.md`. The `no-mistakes` test-quality rule rejects tests that only grep, parse, snapshot, or inspect implementation source strings as proof of behavior. It requires public or executable interfaces, observable behavior, typed or normalized semantic models for declarative artifacts, and regression tests that fail before the fix when feasible. The migrated `implement` skill should import this as a review and TDD standard because skill and artifact repos are prone to source-string pseudo-tests. Source: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/skills/no-mistakes/SKILL.md`. The `no-mistakes` gate has clear custody rules while a run is active. At a gate, the driver reads structured findings and responds approve, fix, or skip; it does not edit code manually while the pipeline owns findings and fixes. The migrated `implement` skill should either avoid invoking such a pipeline or respect custody strictly once it has delegated validation. Source: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/skills/no-mistakes/SKILL.md`. The anti-slop mechanism is not just "run a code review". The documented claim is that a stable pipeline gives "passed the gate" a fixed meaning: intent, rebase, review, test, document, lint, push, PR, and CI happen in order, and push/PR/CI happen only after local checks pass. This prevents a common agent failure mode where the agent declares done after a narrow local edit or a self-selected test subset. Source: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/docs/src/content/docs/concepts/pipeline.md`. Concretely, it attacks slop by removing the implementing agent's discretion over what "done" means. The implementer cannot choose to skip rebase, skip docs, skip lint, skip CI, or call a cherry-picked test enough. The fixed ordered gate externalizes those checks into a separate executor and stops before public push if any step produces unresolved findings. Source: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/docs/src/content/docs/reference/pipeline-steps.md`. It attacks plausible-but-wrong code by feeding author intent into review, test detection, evidence validation, documentation checks, lint fixes, CI fixes, and PR drafting. The review step treats authoritative intent as enforceable for source-verifiable acceptance criteria, so a diff can fail because it does not do the requested thing even when it compiles and looks tidy. The test step validates the change and requested intent with the smallest relevant evidence-oriented tests or manual checks, and when no targeted check can establish intent it must write or improve a focused test, produce evidence, or report a warning finding. Source: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/docs/src/content/docs/reference/pipeline-steps.md`. It attacks fake tests by making source-content-only tests a first-class violation for agents that write, repair, or review tests. Review flags newly added violations and requires same-pattern tests encountered in the accepted change's scope to be removed or made semantic. The test guidance requires observable behavior, state, output, side effects, failure modes, or a typed or normalized semantic model instead of grepping implementation text. Sources: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/docs/src/content/docs/reference/pipeline-steps.md`, `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/internal/testguidance/guidance.go`. It attacks "review theater" by requiring structured output. The review agent returns severity, file location, description, and action, while approval rules require a pause for error or warning findings and for ask-user findings. That makes defects machine-actionable and prevents them from disappearing inside a flattering narrative summary. Source: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/docs/src/content/docs/reference/pipeline-steps.md`. It attacks unsafe autonomous cleanup by distinguishing mechanical fixes from intent questions. `auto-fix` can be handed back to the pipeline, but `ask-user` is for findings that challenge the author's intent or product behavior, and review auto-fix is disabled by default. That keeps the validation agent from silently rewriting product decisions to make its own review pass. Sources: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/skills/no-mistakes/SKILL.md`, `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/docs/src/content/docs/reference/pipeline-steps.md`. It attacks stale-base and unreviewed-history slop through rebase and push binding. The rebase step rebases onto the latest upstream and reports conflicts, and the push step reloads the durable review-approved commit before remote mutation. Push refuses when the approval binding is missing or when the proposed commit does not equal or descend from the reviewed commit, so later unreviewed changes cannot slip through as if they had passed review. Source: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/docs/src/content/docs/reference/pipeline-steps.md`. It attacks "works locally but fails publicly" by putting PR creation and CI monitoring after local gates. The CI step watches hosted CI, fetches failed job logs where supported, auto-fixes failures within limits, and pauses with findings when failures or merge conflicts persist. This closes the gap between local agent confidence and actual host validation. Source: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/docs/src/content/docs/reference/pipeline-steps.md`. The pipeline requires every step either to complete, return structured findings with severity and action, trigger bounded auto-fix, pause for approval when blocking or ask-user findings remain, or fail fatally. This prevents silent degradation: unresolved review, test, docs, or lint issues become explicit findings instead of prose buried in an agent summary. Source: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/docs/src/content/docs/concepts/pipeline.md`. `no-mistakes` uses intent before review and evidence collection. The review step can compare the diff to the user's actual goal, while the test step can run targeted validation and gather user-facing evidence when intent is known. This targets slop where code is syntactically plausible but does not satisfy the requested behavior. Sources: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/docs/src/content/docs/concepts/pipeline.md`, `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/docs/src/content/docs/reference/repo-config.md`. Repository configuration separates trusted default-branch settings from pushed-branch settings. By default, code-executing fields such as test, lint, format commands and selected agent are read from the trusted default-branch copy, so a feature branch cannot self-authorize weaker validation commands on the daemon host. This prevents a slop or malicious branch from weakening its own gate. Source: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/docs/src/content/docs/reference/repo-config.md`. The review and fix loop classifies findings as auto-fix, ask-user, or no-op, with ask-user reserved for product intent or judgment calls. That design limits autonomous fixing to mechanical work and forces human attention where an AI could otherwise paper over a disagreement with the user's intent. Sources: `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/README.md`, `https://raw.githubusercontent.com/kunchenguid/no-mistakes/main/skills/no-mistakes/SKILL.md`. ## Conclusion The migrated `implement` skill should learn from `no-mistakes` that preventing AI slop requires a fixed ordered gate, rich intent, isolated validation, structured findings, bounded auto-fix, human escalation for judgment calls, trusted validation configuration, and CI-backed proof before public handoff. It should preserve unrelated changes, commit only task scope, provide rich reviewer intent, reject source-grep pseudo-tests, and optionally hand shipping validation to an external gate that runs in a disposable worktree. It should not blindly replace the local workflow with `no-mistakes`, because that would require repository initialization, a configured pipeline agent, and accepting the gate's branch custody model.