3.0 KiB

Raw Blame History

name	description	nextStepFile
step-05-adversarial-review	Construct diff and invoke adversarial review skill	./step-06-resolve-findings.md

Step 5: Adversarial Code Review

Goal: Construct diff of all changes, invoke adversarial review skill, present findings.

AVAILABLE STATE

From previous steps:

{baseline_commit} - Git HEAD at workflow start (CRITICAL for diff)
{execution_mode} - "tech-spec" or "direct"
{tech_spec_path} - Tech-spec file (if Mode A)

1. Construct Diff

Build complete diff of all changes since workflow started.

If `{baseline_commit}` is a Git commit hash:

Tracked File Changes:

git diff {baseline_commit}

New Untracked Files: Only include untracked files that YOU created during this workflow (steps 2-4). Do not include pre-existing untracked files. For each new file created, include its full content as a "new file" addition.

If `{baseline_commit}` is "NO_GIT":

Use best-effort diff construction:

List all files you modified during steps 2-4
For each file, show the changes you made (before/after if you recall, or just current state)
Include any new files you created with their full content
Note: This is less precise than Git diff but still enables meaningful review

Capture as {diff_output}

Merge all changes into {diff_output}.

Note: Do NOT git add anything - this is read-only inspection.

2. Invoke Adversarial Review

With {diff_output} constructed, invoke the bmad-review-adversarial-general skill. If possible, use information asymmetry: invoke the skill in a separate subagent or process with read access to the project, but no context except the {diff_output}.

Pass {diff_output} as the content to review. The skill should return a list of findings.

3. Process Findings

Capture the findings from the skill output. If zero findings: HALT - this is suspicious. Re-analyze or request user guidance. Evaluate severity (Critical, High, Medium, Low) and validity (real, noise, undecided). DO NOT exclude findings based on severity or validity unless explicitly asked to do so. Order findings by severity. Number the ordered findings (F1, F2, F3, etc.). If TodoWrite or similar tool is available, turn each finding into a TODO, include ID, severity, validity, and description in the TODO; otherwise present findings as a table with columns: ID, Severity, Validity, Description

NEXT STEP

With findings in hand, read fully and follow: {nextStepFile} for user to choose resolution approach.

SUCCESS METRICS

Diff constructed from baseline_commit
New files included in diff
Skill invoked with diff as input
Findings received
Findings processed into TODOs or table and presented to user

FAILURE MODES

Missing baseline_commit (can't construct accurate diff)
Not including new untracked files in diff
Invoking skill without providing diff input
Accepting zero findings without questioning
Presenting fewer findings than the review skill returned without explicit instruction to do so

3.0 KiB Raw Blame History