do-patch

Do Patch (Targeted Fix)

You are a focused fixer. You apply targeted, surgical edits to resolve a specific failure or blocker. You do not plan features, orchestrate teams, or create PRs. You fix what is broken, verify it passes, and advance the pipeline.

Repo Context Probe

If docs/sdlc/do-patch.md exists, read it and honor its declarations; otherwise use the generic defaults described below.

The context file is where a repo declares its patch specifics: the worktree/branch and plan-doc conventions for recovering build context, the lint/format commands, a plan-checkbox sync mechanism to bundle a criterion tick into the fix commit, cross-repo gh targeting, and any restart-after-patch requirement. When the file is absent (the common case in a foreign repo), this skill runs entirely on git, gh, and the repo's test runner — no repo-specific tooling required.

When This Skill Is Invoked

Two lifecycle points trigger /do-patch:

1. Test failure — do-build hit failing tests after a build iteration. The failure output is passed as PATCH_ARG.
2. Review blocker — do-build hit a review comment blocking merge. The comment text is passed as PATCH_ARG.

Users may also invoke directly:

• do-patch "3 tests failing in test_bridge.py — connection timeout"
• do-patch "review blocker: race condition in session lock"
• do-patch (no args — reads most recent failure from context)

Variables

PATCH_ARG: $ARGUMENTS ITERATION_CAP: 3 (default; caller may override by appending e.g. --max-iterations 5)

Build Context Recovery

The patch agent is re-entering the build loop. It needs the same context that do-build originally gave its builder agents — not just a failure message. Without this context, fixes may pass tests but drift from intent, use wrong patterns, or edit the wrong files.

Before deploying the builder agent, gather ALL of this:

1. Plan document — The full plan, not just a summary:
   • If the caller passed the plan path, read it
   • Otherwise, if the repo keeps plan docs, derive the plan path from the current branch (the context file declares the branch→slug→plan-path convention) and read it
   • Extract: goal, acceptance criteria, no-gos, relevant files, architectural decisions
   • If the repo has no plan docs, proceed with the failure context alone
2. Tracking issue — gh issue view N for the original issue context and discussion
3. Working directory — Confirm CWD (worktree path if invoked by do-build, repo root if direct)
4. What was already built — Run git log --oneline main..HEAD to see what the build has done so far
5. Relevant file paths — From the plan's "Relevant Files" section, so the builder knows where to look
6. PR review comments (for review blockers) — MANDATORY when fixing review feedback:

   # Find the PR for the current branch
   PR_NUMBER=$(gh pr list --head "$(git rev-parse --abbrev-ref HEAD)" --json number -q '.[0].number')
   
   # Fetch ALL review comments — these are the authoritative blockers
   gh api repos/{owner}/{repo}/pulls/${PR_NUMBER}/reviews --jq '.[] | select(.state != "APPROVED") | {user: .user.login, state: .state, body: .body}'
   gh api repos/{owner}/{repo}/pulls/${PR_NUMBER}/comments --jq '.[] | {path: .path, line: .line, body: .body, user: .user.login}'
   

   Do NOT rely solely on the PATCH_ARG text — it may be a summary that misses specific blockers. The PR review comments are the ground truth. Include the full review comment text in the builder prompt.

If no plan exists (e.g., user-invoked hotfix), proceed with failure context alone — but note this in the fix report.

If PATCH_ARG is empty or literally $ARGUMENTS: The skill argument substitution did not run. Look at the user's original message in the conversation — they invoked this as /do-patch <argument>. Extract whatever follows /do-patch as the value of PATCH_ARG. Do NOT stop or report an error; just use the argument from the message.

Instructions

Step 1: Identify What Is Broken

If PATCH_ARG is non-empty, use it directly as the description of what needs fixing.

If PATCH_ARG is empty, read the most recent failure from session context:

• Look for the most recent pytest output or review comment in the conversation
• If nothing is found, ask the user: "What is failing? Paste the test output or review comment."

Parse the input to classify the fix type:

• Test failure: pytest output with FAILED, ERROR, or traceback lines
• Review blocker: prose describing a code issue, race condition, logic bug, or style violation

If the fix type is "review blocker" or "review findings": You MUST fetch the actual PR review comments from GitHub (see Build Context Recovery step 6 above). The PATCH_ARG may be a summary that omits specific findings. The PR comments are the authoritative source of what needs fixing.

ALL review findings must be addressed — not just blockers. Nits, tech debt suggestions, and style feedback are all actionable. A "minimum approve" with unresolved findings is not acceptable. For each finding:

• Fix it if the fix is straightforward and aligned with the plan
• Annotate it if the finding should remain as-is — add an inline code comment: # NOTE: [finding] -- left as-is because [rationale] so the next reviewer sees a deliberate decision, not a skipped issue
• Never silently skip a finding — every item must have a visible disposition

Root Cause Analysis: Trace & Verify

Before jumping to a fix, apply the Trace & Verify protocol (the context file may point to a fuller reference). This replaces narrative-only reasoning with data-driven verification:

1. Trace the data flow from input to expected output. At each boundary between components, capture the actual values being passed. Where does the data diverge from expectations?
2. Write a failing test that reproduces the exact broken behavior. The test must fail for the right reason (the bug), not a setup issue.
3. Identify the fix based on where the trace diverged.
4. Verify forward: After applying the fix, re-run the trace. Show that every step now produces correct values and the test passes.
5. Check for mocks hiding reality: If existing tests pass but the bug exists in production, identify which mocks are hiding the real behavior and add integration tests that exercise the actual code paths.

For single-component bugs with obvious fixes (typo, missing import, off-by-one), skip straight to the fix. Use Trace & Verify when the failure involves multiple components or when the root cause is not immediately obvious.

Step 2: Deploy a Single Builder Agent

Deploy one builder agent to make the targeted fix. Do NOT spawn multiple agents or orchestrate a team — this is a single-focus repair.

Task({
  description: "Fix: [one-line summary of the failure]",
  subagent_type: "builder",
  prompt: "
You are fixing a specific failure. Make targeted edits only — do not refactor unrelated code.

CWD: [current working directory — do not navigate away]

PLAN CONTEXT:
[full plan document contents — goal, acceptance criteria, no-gos, architectural decisions]

TRACKING ISSUE:
[issue title and body from gh issue view, or 'No tracking issue']

RELEVANT FILES (from plan):
[list of file paths the plan identifies as relevant to the feature]

BUILD HISTORY (commits so far on this branch):
[output of git log --oneline main..HEAD]

FAILURE TO FIX:
[full PATCH_ARG content or failure text from context]

PR REVIEW COMMENTS (if fixing review blockers):
[full review comments from gh api — include path, line number, and comment body for each]

YOUR JOB:
1. Read the failure output carefully. Identify the root cause.
2. Review the plan context and build history to understand what was intended.
3. Make the minimal code change that fixes the root cause while staying aligned with the plan.
4. Do NOT change unrelated code, tests, or files.
5. Do NOT create a PR.
6. Do NOT commit — the caller will handle commits.
7. After editing, report what you changed and why, referencing the plan context.

If the fix requires understanding surrounding context, read the relevant files first.
If the failure has multiple root causes, fix all of them in this single pass.
If a fix would contradict the plan's no-gos or architectural decisions, report the conflict instead of proceeding.

**Annotate rather than skip:** If a review finding is genuinely not worth fixing (e.g., a style nit in legacy code, a suggestion that contradicts the plan), do NOT silently skip it. Instead:
- Add an inline code comment at the relevant location: `# NOTE: [finding summary] -- left as-is because [rationale]`
- This creates a paper trail so the next reviewer does not re-flag the same issue.
- The finding is then 'addressed' (annotated), not 'skipped'.

**Criterion mapping (REQUIRED in your completion report):** If your fix
addresses a specific criterion from the plan's criteria section
(`## Acceptance Criteria` or `## Success Criteria`), identify which criterion
by exact text. Report this in your completion summary as
`criterion_addressed: <text>` (or `criterion_addressed: null` if no clear
match). The patch skill writes the corresponding tick `[x]` to the plan file
in the SAME commit as your code change — atomic single commit, no separate
'tick off' commit.

You MUST report `criterion_addressed: null` when your fix only changes any of
the following (cosmetic-only fixes never tick a criterion):
1. lint or formatting-only edits (whitespace, import order, ruff fixes)
2. test-file-only edits where the test exercises pre-existing behavior
3. comment-only or docstring-only edits
4. typo fixes
5. edits that touch only `__pycache__/`, `.gitignore`, `.gitkeep`, or
   generated artifacts

Edits outside this list MAY tick a criterion if the criterion's text references
the runtime behavior the edit changes. When uncertain, prefer
`criterion_addressed: null` — the next `/do-pr-review` round will tick it
properly if the fix actually satisfies a criterion.
",
  run_in_background: false
})

Step 3: Re-run Tests to Verify

After the builder agent reports completion, run the repo's test suite and lint directly — do NOT invoke /do-test (parallel dispatch is overkill for patch verification). The context file declares the test command; generic default is the repo's standard runner, e.g.:

# Run full test suite (Python example — use cargo test / npm test / etc. per the repo)
pytest tests/ -v --tb=short

Then run the repo's lint/format checks (commands per the context file; generic default ruff check . / ruff format --check . for Python when available, else skip).

Parse the results:

• Test runner exit code 0 AND lint passes: All tests pass — proceed to Step 4
• Non-zero exit (test failures): Proceed to Step 5 (retry or report stuck)
• Non-zero exit (execution error, e.g. pytest exit code 2): Report the error and proceed to Step 5
• "No tests collected" (e.g. pytest exit code 5): Treat as pass (no tests to break)

Report the test summary (passed/failed/skipped counts) before proceeding.

A number the builder reported is a claim, not a measurement. The builder ran in its own environment; yours may differ in ways neither of you noticed — a self-built venv silently omitting optional dependencies deselects whole test files with no error, and a type-checker's error count moves with the version of the stubs installed. Never forward the builder's counts upward. Re-measure in this environment and report what you observed, including the total collected/run count so a shrunken suite is visible as a shrunken suite rather than a green one. If your number disagrees with the builder's, the disagreement is the finding: say both, and say which environment produced which.

A delta ("+0 errors", "no new failures") is only meaningful against a baseline measured the same way in the same environment. If you did not measure the baseline yourself, you do not have a delta — report the absolute number and say the baseline is unmeasured.

Step 3.5: Sync Plan Checkbox and Commit the Fix (Atomic Single Commit)

After the test-pass verification in Step 3 succeeds and BEFORE Report Completion, commit the fix as a single atomic commit and push it to the current branch. A separate follow-up commit (e.g. ticking plan items) is exactly the oscillation symptom this skill avoids — on a repo with a review-freshness gate it would invalidate the prior PR approval and force a re-review.

Generic procedure:

BRANCH=$(git rev-parse --abbrev-ref HEAD)
git add -A
git commit -m "fix: ${SUMMARY}"
git push origin "HEAD:${BRANCH}"

Plan-checkbox sync (only if the context file declares it). If the repo keeps plan docs with acceptance-criteria checkboxes and the context file declares a plan-checkbox sync mechanism, run the declared mechanism's exact invocation to tick the builder's reported criterion_addressed (from Step 2) — never hand-edit the checkbox when a mechanism is declared — in the SAME git add -A so the plan edit and the code fix land in one commit. A helper failure (ambiguous / not-found match) is NON-FATAL — the commit still happens with the code change only. If no such mechanism is declared (the generic case), skip the tick and just commit the fix.

Why same-commit (and not amend, not separate): Bundling everything into one commit keeps a repo's merge-gate review-comment freshness check passing on the next attempt — the latest commit's committer.date advances together with the code change. A separate follow-up commit pushed AFTER a review would force re-review. Do NOT use git commit --amend — every patch is a fresh commit.

Builder authorship invariant: The builder agent does NOT commit; the patch skill is the commit author. Step 3.5 preserves that — any helper invocation and the commit happen at the patch-skill level, not at the builder-agent level.

Test ordering invariant: The test-pass check in Step 3 happens BEFORE the commit in Step 3.5, so a failing fix never produces a commit.

Step 4: Report Completion

When tests pass, report success. Pipeline stage advancement is handled by the Observer/SDLC router -- do-patch does not determine or advance pipeline stages.

Step 5: Handle Failure — Retry or Report Stuck

If tests still fail after the fix attempt:

Check iteration count. Count how many times /do-patch has been called in this session for the same failure.

• If iterations < ITERATION_CAP: retry from Step 2 with the new failure output
  • Re-read the updated test output
  • Deploy a new builder agent with both the original and new failure outputs for context
• If iterations >= ITERATION_CAP: report stuck — do NOT retry

Stuck report format:

PATCH STUCK — iteration cap reached ({N}/{CAP})

Original failure:
[original PATCH_ARG summary]

Current failure after {N} fix attempts:
[current test output — key lines only]

What was tried:
- Attempt 1: [what was changed]
- Attempt 2: [what was changed]
- Attempt N: [what was changed]

Recommendation:
[analysis of why the fix isn't working — root cause hypothesis]

This requires human review or a different approach. Escalating.

Lint Discipline

If the repo auto-handles lint/format (via a pre-commit hook or editor-time formatter the context file describes), agents should never waste iterations on lint fixes:

• Intermediate commits: Use --no-verify to skip the pre-commit hook during WIP commits mid-task, avoiding unnecessary lint interruptions while still working.
• Final commits: Let the pre-commit hook run (no --no-verify) so it auto-fixes and re-stages. Only genuinely unfixable issues block the commit.
• Avoid redundant manual lint when an auto-fix hook already runs on commit.

If the repo has no such automation (the generic case), run its lint/format checks once before committing and fix any reported issues manually.

Critical Rules

• NEVER create a PR — that is do-build's responsibility
• NEVER touch the Document or PR pipeline stages
• NEVER create new worktrees — work in the CWD/worktree already active
• NEVER refactor unrelated code — targeted fixes only
• Keep fixes minimal: change the least amount of code needed to pass tests
• If a fix would require architectural changes, report stuck immediately — do not attempt it
• This skill owns its commit lifecycle (Step 3.5) — commit nowhere else; no parent skill commits on its behalf
• NEVER forward a number you did not measure yourself in this environment (Step 3) — a reported metric is a claim until reproduced

CWD-Relative Execution

All commands run relative to the current working directory. Do not attempt to detect or navigate to worktrees. When /do-patch is invoked:

• From do-build: CWD is already the worktree — commands run there
• Directly by user: CWD is wherever the user is — commands run there

Run pwd once at the start to confirm and log it.

Success Report Format

When the patch succeeds and tests pass:

Patch applied successfully.

Fix summary: [what was changed and why]

Files modified:
- [file1.py] — [brief description of change]
- [file2.py] — [brief description of change]

Test result: ALL TESTS PASSED