Extract adversarial review CI job into a standalone reusable GitHub Action

[keeb]

Summary

Extract the claude-adversarial-review CI job from .github/workflows/ci.yml into a standalone, reusable composite GitHub Action that other repos can call.

Motivation

Community members building swamp extensions want to run the same adversarial code review quality gate on their own PRs. The current implementation is inline in our CI workflow and not reusable.

From Discord — @bixu:

Would y'all be willing to share the adversarial review GH Actions logic you use as a standalone Action? We'd be interested in calling it to ensure our own extensions hit the same quality mark as @swamp

Desired Usage

- uses: systeminit/swamp/.github/actions/adversarial-review@main
  with:
    github_token: ${{ secrets.GITHUB_TOKEN }}

Requirements

• Composite action at .github/actions/adversarial-review/action.yml
• Wraps anthropics/claude-code-action@v1 with the adversarial review prompt
• Parameterizable: extra review dimensions, severity threshold, conventions file path
• swamp's own CI should consume the new action (dogfood)
• No swamp-specific content in the action itself — it should be generic

References

• Current implementation: .github/workflows/ci.yml#L255-L374
• Adversarial review skill docs: .claude/skills/issue-lifecycle/SKILL.md#L114

[keeb]

🔍 Triage started — fetching issue context

[keeb]

📋 Classified as feature (high) — Feature request from community extension author (bixu) to extract the inline adversarial review CI prompt into a reusable composite GitHub Action. The current implementation at ci.yml:255-400 is self-contained and has no swamp-specific dependencies — it's a generic adversarial code review prompt that wraps claude-code-action. No existing .github/actions/ directory exists yet.

[keeb]

📝 Implementation plan generated (v1) — Extract the adversarial review prompt from ci.yml into a reusable composite GitHub Action at .github/actions/adversarial-review/. The action wraps claude-code-action@v1 with the full adversarial review prompt, exposes parameterized inputs (model, conventions file, extra dimensions, allowed tools), and includes a fail step. swamp's own CI dogfoods the new action. External repos like bixu's org can consume it standalone.

[keeb]

🔍 Adversarial review (plan v1)

• ADV-1 [medium/risk]: GitHub composite actions do not support secrets as inputs — anthropic_api_key must be documented as requiring an environment variable (ANTHROPIC_API_KEY) set by the caller, or passed as a regular input with a warning that it appears in logs if the step is verbose. The claude-code-action itself accepts anthropic_api_key as an input, so the composite action can accept it as a regular input and pass it through — GitHub will mask it if the caller passes a secret reference.
• ADV-2 [low/scope]: The plan does not include versioning/tagging strategy for the action. External consumers pinning to @main will get breaking changes. Consider documenting that consumers should pin to a commit SHA or that a tag strategy will come later.
• ADV-3 [low/correctness]: The conventions_file input defaults to CLAUDE.md but some repos won't have this file. The prompt should handle the case where the file doesn't exist gracefully — the current prompt says 'read CLAUDE.md' which Claude will attempt and get an error. Could phrase it as 'if it exists, read it'.

✅ No blocking findings — ready for approval

[keeb]

🔄 Plan revised (v2) — incorporated feedback round 1

[keeb]

✅ Findings resolved (3)

• ADV-1: anthropic_api_key accepted as regular input — GitHub masks secret refs automatically. Documented in README.
• ADV-2: README step 3 now includes guidance to pin to commit SHA until release tagging is established.
• ADV-3: Prompt in step 1 now reads conventions_file conditionally: 'If a file named <conventions_file> exists, read it for project conventions.'

✅ All blocking findings resolved — ready for approval

[keeb]

🔍 Adversarial review (plan v2)

• ADV-4 [low/correctness]: The fail step (touch /tmp/review-failed + check) is currently inline in ci.yml. When dogfooding the composite action, this fail logic should remain in the caller workflow, not inside the action — the action should only run the review, and the caller decides what to do with the outcome. This is already how the plan is structured (step 2 preserves the fail step in ci.yml), so this is just a confirmation, not an issue.

✅ No blocking findings — ready for approval

[keeb]

Approved Implementation Plan (v2)

Summary: Same 3-step plan with ADV-1/2/3 feedback incorporated: step 1 now makes conventions_file conditional in prompt and accepts anthropic_api_key as regular input; step 3 now includes SHA pinning guidance in README.

DDD Analysis

Not applicable — CI/GitHub Actions infrastructure change.

Implementation Steps

1. Create .github/actions/adversarial-review/action.yml composite action wrapping claude-code-action. Inputs: github_token (required), anthropic_api_key (required, passed as regular input — GitHub masks secret refs), model (default claude-opus-4-6), conventions_file (default CLAUDE.md), extra_dimensions (optional), allowed_tools (sensible default). Prompt conditionally reads conventions_file ('if it exists, read it') to handle repos without CLAUDE.md. (risk: Must ensure prompt template correctly interpolates GitHub context variables through the composite action boundary)
2. Replace the inline prompt in ci.yml claude-adversarial-review job with uses: ./.github/actions/adversarial-review, passing github_token and anthropic_api_key. Preserve the fail step, concurrency, and conditional logic. (risk: Must preserve identical behavior)
3. Add README.md documenting standalone usage. Include: required inputs, example workflow, note that consumers should pin to a commit SHA until a tagging/release strategy is established. (risk: None)

Files

• .github/actions/adversarial-review/action.yml
• .github/workflows/ci.yml
• .github/actions/adversarial-review/README.md

Testing Strategy

Dogfood: swamp CI uses the new action. Manual: test PR after merge to verify identical review behavior.

Potential Challenges

• External consumers must supply their own ANTHROPIC_API_KEY secret
• Prompt references github.repository and github.event.pull_request.number which must be threaded through the composite action

Incorporated 1 round(s) of feedback.

Approved via swamp @si/issue-lifecycle