The cloud agent checks out a branch like copilot/review-pr-NNN which
may not have upstream/main or origin/main refs available for merge-base.
Changes:
- Use gh pr diff as primary method (always works in PR context)
- Fall back to git merge-base for local non-PR usage
- Remove path filters (*.rs *.toml examples/) — review full diff
- Remove head -2000 truncation — let agents see everything
- Explicitly fetch origin/main in copilot-setup-steps.yml as backup
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Add Copilot review skills, project instructions, and coding agent setup
for automated code review on regorus PRs.
Files added:
- .github/copilot-instructions.md — project context (no_std, 9 bindings,
dual execution paths, deny lints, security-critical evaluation)
- .github/skills/code-review/SKILL.md — fast single-agent review (~2 min)
- .github/skills/deep-review/SKILL.md — multi-agent deep review (~12 min)
- .github/copilot-setup-steps.yml — minimal coding agent environment
Development and testing methodology:
The skills were developed iteratively (v3 through v11.4) against a
460-line SARIF output module on the feature/sarif-output branch, which
served as a controlled test bed with 25 known issues of varying severity
(correctness, safety, API design, platform, security, performance).
Each version was tested by running the skill via the Copilot CLI, then
mapping discovered findings against the ground truth set to measure
recall and precision. Key iterations:
- v3: baseline single-agent (8/25 recall, 32%)
- v7: 3 parallel agents + verification (14/25, 56%)
- v10c: model diversity + adversarial pass (10/25, 40%)
- v11.3: merged adversarial-verifier architecture (12/25 + 2 novel, 0 noise)
- v11.4: domain expertise prompting (12/25 + 2 novel, 0 noise, full report)
The final architecture uses 3 parallel discovery agents (with cross-model
diversity and context asymmetry), risk-triggered micro-passes, and a
single adversarial verifier that both validates candidates via disproval
and hunts blind spots. Agents are prompted to reason from policy-author
perspective across Rego/OPA, Azure Policy, and RVM workloads.
Combined CR+DR catches 16-17/25 ground truth with zero false positives
and produces verified findings with confidence levels, test gap analysis,
and agent performance metrics.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
