Audit Impl

Implementation Audit Skill

Audit a completed implementation against its plan(s) before merge. Identifies gaps, missed requirements, scope creep, and unexpected changes. Produces a GO or NO GO verdict.

When to Use

• After implementation completes and tests pass, as the last step before merging
• In single-plan pipelines: audit the worktree against the plan
• In multi-group pipelines: audit the feature branch against all group plans via manifest
• Standalone: /autoskillit:audit-impl {plans_input} {branch_name} {base_branch}

Arguments

{plans_input} {branch_name} {base_branch} [conflict_report_paths]

• plans_input — one of:
• A single plan .md file path
• A comma-separated list of .md plan file paths (no spaces around commas)
• A directory containing *_plan_*.md files
• A manifest_*.json from /autoskillit:make-groups
• branch_name — commit SHA of the pre-implementation base ref (preferred from pipeline,

stable after merge_worktree destroys named refs); a branch name is also accepted for standalone invocations. A live worktree path is accepted for legacy use (Step 0 extracts the branch name automatically).

• base_branch — branch to diff against (default: develop)
• conflict_report_paths (optional) — comma-separated list of absolute paths to conflict

resolution reports produced by resolve-merge-conflicts. When provided and non-empty, cross-reference resolution decisions against plan intent in Step 2.5.

Critical Constraints

NEVER:

• Modify source files, plan files, or any other files — read-only audit only
• Run tests — this skill audits, it does not fix
• Create files outside {{AUTOSKILLIT_TEMP}}/audit-impl/
• Emit a GO verdict when any MISSING or CONFLICT finding exists
• Run subagents in the background (run_in_background: true is prohibited)

ALWAYS:

• Use Explore subagents for all file reads and diff retrieval
• Use model: "sonnet" when spawning all subagents via the Task tool
• Resolve all plan files before starting (abort early if any are missing)
• Write Dry-walkthrough verified = TRUE as the absolute first line of any remediation file
• On a NO GO verdict, after writing the remediation file, emit the absolute path as a

structured output token as your final output. Resolve the relative temp/audit-impl/... save path to absolute by prepending the full CWD: `` verdict = NO GO remediation_path = /absolute/cwd/temp/audit-impl/{filename}.md ` On a GO verdict, emit only verdict = GO` (no remediationpath token). The remediationpath token is MANDATORY on NO GO — the pipeline cannot proceed without it.

Workflow

Step 0 — Parse Arguments

Resolve plans_input:

• Single .md file: no comma, ends in .md → use it directly
• Comma-separated .md paths: value contains , → split on ,, trim whitespace

from each token. Validate that each trimmed token ends in .md; log a warning and skip any token that does not. Use each valid token as a plan file path

• Directory: no comma, does not end in .md or .json → glob for *_plan_*.md

files in the directory

• manifest_*.json: no comma, ends in .json → parse it; extract groups[*].file

paths, resolved relative to the manifest's parent directory

Verify every plan file exists. If any are missing, abort with a clear error listing them.

Determine the diff source from implementation_ref:

1. If implementation_ref is an existing directory path:

• Extract the branch name: git -C {implementation_ref} branch --show-current
• Run: git diff {base_branch}...{branch_name} from the current working directory
• (This handles legacy manual invocations where a live worktree path is passed.)

1. Otherwise, detect whether implementation_ref is a commit SHA or branch name:

• Detect SHA: echo "$implementation_ref" | grep -qE '^[0-9a-f]{40}$'
• If SHA (pre-implementation base ref, preferred from pipeline):
• Run: git diff {implementation_ref}..{base_branch} — two-dot, SHA on the left
• This shows all commits added to base_branch since the pre-implementation snapshot,

covering the full multi-group implementation across all merged worktrees.

• If branch name:
• Run: git diff {base_branch}...{implementation_ref} — three-dot (unchanged)
• If git reports "unknown revision or path not in the working tree", abort with:

> "implementationref '{implementationref}' is neither an existing worktree > directory nor a known git ref. If you are passing a worktree path, ensure > the worktree has not been deleted before calling audit-impl."

The old silent fallthrough (non-existent path treated as branch name without error) is removed. A clear error is emitted instead.

Parse the optional fourth positional argument conflict_report_paths (may be absent or empty string). Split on ,, trim each entry, and filter out any empty strings after splitting; store as conflict_report_path_list. Proceed even if empty — the cross-reference check in Step 2.5 is skipped when the list is empty.

Path-existence guard: Before issuing a Read call on a path that is not guaranteed to exist (e.g., plan file arguments, {{AUTOSKILLIT_TEMP}}/investigate/ reports, external file references), use Glob or ls to confirm the path exists first. This prevents ENOENT errors that cascade into sibling parallel-call cancellations.

Step 1 — Load Plans via Parallel Subagents

Launch one Explore subagent per plan file in parallel. Each returns:

• Plan title and stated scope
• All files the plan said it would create, modify, or delete
• All tests the plan said it would add or modify
• Key requirements and constraints listed in the plan

Aggregate into a unified requirements inventory.

Step 2 — Load Implementation Diff

Stale branch guard (branch name refs only — skip for SHA refs):

Skip this check when implementation_ref is a commit SHA (a SHA-as-ancestor of base_branch is expected after all worktrees are merged — this check is only meaningful for named refs).

# Only run for branch name refs (not SHA):
if ! echo "$implementation_ref" | grep -qE '^[0-9a-f]{40}$'; then
    # Step 1: ref must exist at all
    git rev-parse --verify {implementation_ref} 2>/dev/null
    # Step 2: branch must not already be fully merged into base (fast-forward absorption check)
    git merge-base --is-ancestor {implementation_ref} {base_branch}
fi

• If ref lookup fails (branch ref not found): abort with a clear error —

"branch ref '{implementation_ref}' not found — it may have been absorbed by a fast-forward merge before audit_impl ran. This is a pipeline routing error." Output {"success": false} with this message. Do not proceed to audit.

• If --is-ancestor exits 0 (branch is an ancestor of base — already fully merged): log a

warning, then treat this as GO with note: "Branch '{implementation_ref}' is already an ancestor of '{base_branch}' — absorbed by fast-forward merge prior to audit. No delta to evaluate; returning GO." This is O(1), unambiguous, and distinguishes the stale-fast-forward case from legitimate no-op branches (empty diff is an unreliable guard; --is-ancestor is the correct tool).

Launch one Explore subagent to retrieve:

• git diff {base_branch}...HEAD --stat — file-level summary
• git log {base_branch}..HEAD --oneline — commit history
• git diff {base_branch}...HEAD — full diff

Step 2.5 — Conflict Resolution Context Check

Before running standard audit, check if any plan contains a PR Changes Inventory section (written by merge-pr to document all files changed by the PR).

If a PR Changes Inventory is found:

1. Extract the Category C — Clean Carry-Overs file list from each plan.
2. For each Category C file, verify it appears in the implementation diff.
3. Any Category C file absent from the diff is a MISSING finding — even if no plan

requirement explicitly named it. Missing carry-over files indicate silent data loss in the conflict resolution and always force a NO GO verdict.

Record all Category C MISSING findings alongside the standard audit findings in Step 3.

Conflict Resolution Report Cross-Reference (when conflict_report_path_list is non-empty):

For each path in conflict_report_path_list:

1. Check whether the file exists before reading. If the path does not exist, log a warning

"Warning: conflict report not found at {path} — skipping" and continue to the next path.

1. Read the conflict resolution report.
2. Parse the ## Per-File Resolution Decisions table — extract all rows as

(file, category, confidence, strategy, justification) tuples.

1. For each resolved file, check against the plan:

• Category 3 resolution flagged: Any row with Category = 3 indicates a Category 3

(architectural tension) conflict was resolved rather than escalated. This ALWAYS forces a CONFLICT finding — Category 3 conflicts must never be automatically resolved per the resolve-merge-conflicts escalation contract.

• Strategy contradicts plan intent: If the plan's ## Resolver Contract or

## Implementation Steps sections prescribe a specific outcome for the file that contradicts the recorded strategy (e.g., plan says "preserve the new API signature" but strategy is "ours" which kept the old signature), record a CONFLICT finding.

• Resolved file absent from diff: If a file listed in the report's table is absent

from the implementation diff entirely, record a MISSING finding — the resolved content was not carried into the integration branch.

Record all findings from this cross-reference alongside the standard Step 3 audit findings. Each CONFLICT or MISSING finding here forces a NO GO verdict per the existing verdict logic (Step 4).

Step 3 — Audit via Parallel Subagents

Divide the requirements inventory into up to 3 slices. Launch parallel Explore subagents, each receiving its slice and the full diff. Each subagent checks:

1. Coverage — Is every file and function the plan named present in the diff?
2. Correctness — Does the implementation match the plan's stated intent? Flag inversions,

missing logic, or wrong approaches.

1. Scope creep — What is in the diff that no plan called for? Flag unexpected files or

additions.

1. Test coverage — Were the plan's specified tests added?
2. Cross-plan conflicts (multi-plan only) — Do any two plans' changes interfere or

contradict?

Each subagent returns structured findings:

• COVERED — requirement satisfied in the diff
• MISSING — required change absent from diff
• ODD — change in diff with no plan backing
• CONFLICT — two plans' implementations interfere with each other

Step 4 — Verdict

NO GO if any finding is MISSING or CONFLICT.

GO (with notes) if only ODD findings exist — unexpected additions that do not break correctness.

GO if all findings are COVERED.

Step 5 — Output

If GO or GO with notes

Print:

## Audit Result: GO

### Scope Audited
{list of plan files audited}

### Summary
{2–3 sentences on overall implementation quality}

### Notes
{Minor ODD findings — not blockers. Omit section if none.}

### Verdict
MERGE APPROVED

Exit 0. The pipeline may proceed to merge.

After printing the GO result, emit the following structured output token as the very last line of your text output:

> IMPORTANT: Emit the structured output tokens as literal plain text with no > markdown formatting on the token names. Do not wrap token names in **bold**, > *italic*, or any other markdown. The adjudicator performs a regex match on the > exact token name — decorators cause match failure.

verdict = GO

---

If NO GO

Generate {{AUTOSKILLIT_TEMP}}/audit-impl/remediation_{topic}_{YYYY-MM-DD_HHMMSS}.md:

Dry-walkthrough verified = TRUE

# Remediation Plan: {topic}

## Audit Context

Generated by `/autoskillit:audit-impl` after auditing:
{list of original plan files}

## Findings

{For each MISSING and CONFLICT finding:}

### {Finding type}: {short title}

- **Plan reference:** {plan file + section}
- **Expected:** {what the plan specified}
- **Found:** {what the diff shows, or "not present"}

## Remediation Steps

{For each finding, a concrete fix:}

### Fix: {short title}

- **File:** {path}
- **Change:** {what to add, modify, or remove}
- **Requirement:** {plan requirement this satisfies}

## Verification

After remediation:
- Run:
  ```bash
  CURRENT_BRANCH=$(git rev-parse --abbrev-ref HEAD) && \
  AUTOSKILLIT_TEST_FILTER="${AUTOSKILLIT_TEST_FILTER:-conservative}" \
  AUTOSKILLIT_TEST_BASE_REF=$(cat "{{AUTOSKILLIT_TEMP}}/worktrees/${CURRENT_BRANCH}/base-branch" 2>/dev/null || echo "") \
  task test-all
  ```
- Re-run `/autoskillit:audit-impl` to confirm GO

Then print:

## Audit Result: NO GO

### Scope Audited
{list of plan files audited}

### Findings
{Mirror the findings from the remediation file}

### Remediation File
{absolute path to remediation file}

### Verdict
MERGE BLOCKED — feed remediation file to /autoskillit:retry-worktree or /autoskillit:implement-worktree-no-merge

Exit 1.

After printing the NO GO result, emit the following structured output tokens as the very last lines of your text output:

> IMPORTANT: Emit the structured output tokens as literal plain text with no > markdown formatting on the token names. Do not wrap token names in **bold**, > *italic*, or any other markdown. The adjudicator performs a regex match on the > exact token name — decorators cause match failure.

verdict = NO GO
remediation_path = {absolute_path_to_remediation_file}

The verdict token must be exactly GO or NO GO — this is the value the recipe's on_result: field: verdict routing matches against. The remediation_path token must be the absolute path to the remediation file written in this session (only emitted for NO GO; omit the remediation_path= line entirely on GO).

Output Location

{{AUTOSKILLIT_TEMP}}/audit-impl/
└── remediation_{topic}_{YYYY-MM-DD_HHMMSS}.md    (written on NO GO only)

Related Skills

• /autoskillit:implement-worktree — produces the worktree this skill audits
• /autoskillit:implement-worktree-no-merge — orchestrator-mode implementation
• /autoskillit:retry-worktree — consumes the remediation file on NO GO
• /autoskillit:make-groups — produces the manifest this skill accepts as plans_input

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: TalonT-Org
• Source: TalonT-Org/AutoSkillit
• License: MIT

Install and usage instructions live in the source repository linked above.