Add replay workflow for gh-aw threat detection runs

[Copilot]

• Copy root-level prompt.txt files into aw-prompts/prompt.txt during replay normalization
• Remove GH_AW_REPLAY_TOKEN handling and rely on the dispatching repository token
• Update README to document internal/current-token replay behavior
• Re-run workflow syntax checks, lint/tests, secret scan, and validation

[Copilot]

Pull request overview

Adds a manual GitHub Actions workflow to replay threat-detect runs against artifacts from prior github/gh-aw workflow runs, enabling maintainers to reproduce and compare historical threat-detection results using different detector sources (current checkout, released binary, or container image).

Changes:

• Adds a new Replay Threat Detection workflow_dispatch workflow that downloads prior run artifacts, normalizes them into the detector input contract, runs the detector, and uploads sanitized replay outputs.
• Expands README documentation for the artifact directory contract and documents replay workflow usage and required tokens.

Show a summary per file

File	Description
README.md	Documents replay workflow usage and updates the artifacts directory contract to include replay-relevant files.
.github/workflows/replay-detection.yml	Implements a workflow_dispatch “replay” pipeline (download, normalize, run detector via current/release/image, sanitize, compare, upload).

Copilot's findings

Tip

Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

• Files reviewed: 2/2 changed files
• Comments generated: 3

[Copilot]

Copilot's findings

Comments suppressed due to low confidence (3)

.github/workflows/replay-detection.yml:440

• In image mode, the container is run as a non-root user (uid 1000 in the Dockerfile), but the bind-mounted ${REPLAY_OUTPUT} directory is created on the host and may not be writable by uid 1000. Consider running the container with the host UID/GID (or adjusting permissions) so --output /workspace/output/result.json can be written reliably.

              -e "CUSTOM_PROMPT=${CUSTOM_PROMPT}" \
              -e "WORKFLOW_NAME=${WORKFLOW_NAME}" \
              -e "WORKFLOW_DESCRIPTION=${WORKFLOW_DESCRIPTION}" \
              -v "${REPLAY_INPUT}:/workspace/artifacts:ro" \
              -v "${REPLAY_OUTPUT}:/workspace/output" \

.github/workflows/replay-detection.yml:438

• The docker replay path only passes CUSTOM_PROMPT/WORKFLOW_* into the container. If image mode is intended to work with authenticated engines, it also needs to pass through the relevant engine auth env vars (e.g., GH_AW_COPILOT_TOKEN/ANTHROPIC_API_KEY/OPENAI_API_KEY, or a token expected by the engine CLI). Otherwise the container run will fail or run unauthenticated.

            docker run --rm \
              -e "CUSTOM_PROMPT=${CUSTOM_PROMPT}" \
              -e "WORKFLOW_NAME=${WORKFLOW_NAME}" \
              -e "WORKFLOW_DESCRIPTION=${WORKFLOW_DESCRIPTION}" \

.github/workflows/replay-detection.yml:497

• Original result extraction assumes a raw THREAT_DETECTION_RESULT:{...} line; it won’t successfully parse Claude’s stream-json output (the JSON will be escaped inside a JSON envelope). Since this workflow supports engine=claude, consider extending this step to handle stream-json logs (e.g., parse each JSON line, read the .result field, then extract the prefix) so comparisons work for Claude runs too.

          candidates = list(root.rglob('detection.log')) + list(root.rglob('*.log'))
          marker = re.compile(r'THREAT_DETECTION_RESULT:(\{.*\})')
          for path in candidates:
              for line in path.read_text(errors='replace').splitlines():
                  match = marker.search(line)

• Files reviewed: 2/2 changed files
• Comments generated: 1

[Copilot]

Copilot's findings

• Files reviewed: 2/2 changed files
• Comments generated: 2

[Copilot]

Copilot's findings

• Files reviewed: 2/2 changed files
• Comments generated: 3