GSD-style cross-AI peer review: dispatch review to independent AI CLI

[azalio]

Summary

map-framework map-review runs within the same model session. GSD can dispatch reviews to independent AI CLIs (Codex, Gemini, Copilot, OpenCode) for truly independent second opinions. This is cross-AI peer review — the reviewer is a different model, possibly from a different vendor, with no shared context.

What GSD does

/gsd-review --cross-ai

• Dispatches the review task to an independent AI CLI (different from the current runtime)
• Supports: Claude Code, Gemini CLI, Codex, Copilot, OpenCode, Cursor, Windsurf, Kilo, Augment Code, Trae, Cline, Antigravity
• Delivers prompt via stdin-based delivery to the external CLI
• Review result is parsed and presented in the orchestrator session

Why cross-AI matters

• Same-model review is "inbred" — the model reviews its own patterns and biases
• Cross-AI review catches model-specific blind spots
• Independent CLI = true fresh context, no shared session state
• Different vendors = different strengths (Claude for reasoning, Gemini for breadth, Codex for code generation patterns)

Current map-framework state

• map-review operates within a single model session
• Review agents (Monitor, Predictor, Evaluator) share the same model and context
• No mechanism to dispatch to external AI runtimes
• No cross-model or cross-vendor comparison

Proposed design

map-review --cross-ai <runtime>

map-review --cross-ai codex       # Dispatch to Codex CLI
map-review --cross-ai gemini      # Dispatch to Gemini CLI
map-review --cross-ai all         # Dispatch to all available, aggregate

External CLI dispatch

1. Detect available AI CLIs on the system
2. Format prompt for target CLI format (Claude: markdown, Gemini: extension JSON, Codex: slash command)
3. Execute: claude -p "$(cat review-prompt.md)" --output-format json
4. Parse result, normalize to common review schema
5. Present alongside in-session review for comparison

Multi-vendor aggregation

• When --cross-ai all: dispatch to all detected CLIs
• Vote/consensus: which issues do all reviewers agree on?
• Disagreements: highlight contradictions for user judgment
• Format: unified report with per-reviewer sections + consensus section

Integration

• .map/config.yaml:

  review:
    cross_ai:
      enabled: true
      runtimes: [codex, gemini]   # Or "all" / "detect"
      fallback: claude             # If target unavailable

References

• GSD: commands/gsd/gsd-review.md, cross-AI flag implementation
• GSD Architecture: docs/ARCHITECTURE.md (cross-AI peer review section)

Acceptance criteria

• map-review --cross-ai <runtime> flag
• Auto-detect available AI CLIs on system
• Prompt formatting for target CLI (Claude, Codex, Gemini, OpenCode)
• Result parsing + normalization to common schema
• --cross-ai all: dispatch to all, aggregate with consensus
• .map/config.yaml cross_ai section
• Fallback: if target unavailable, use in-session review

[azalio]

Беру в работу.

Планируемый срез 1 (config-driven, opt-in, off by default):

• review.cross_ai.* секция в .map/config.yaml (dotted-key alias → snake_case dataclass, по образцу sofa.enabled)
• Python step-runner subcommand для dispatch во внешний AI CLI: shell=False argv + timeout, парсинг JSON-envelope, нормализация в общий finding-schema (producer-owns-parse — как skills_eval/dispatcher.py для claude -p)
• runtime-adapter registry: {detect_on_PATH, build_argv, parse_envelope, normalize_findings}; слайс 1 — codex / gemini / claude / generic stdin-JSON
• inbound trust boundary: результат внешнего CLI оборачивается в EXTERNAL UNTRUSTED REFERENCE fence (механизм из map-so-search/SOFA)
• egress-guard: dispatch только по явному флагу, never auto-run; в скилле явное уведомление, что код/дифф уходит во внешний vendor
• graceful degradation: CLI не установлен / не залогинен / timeout / не-JSON → non-blocking статус + fallback на in-session review
• map-review --cross-ai <runtime> флаг в SKILL.md (по образцу свежего --adversarial, PR feat: adversarial multi-perspective code review (map-review --adversarial) #294)
• тесты + render generated trees

Срез 2 (отдельный заход): --cross-ai all + consensus/disagreement aggregation.

Сверяю дизайн с llm-council, дальше имплементирую. Part of #288 на PR среза 1 (issue остаётся открытым до consensus-среза).

[azalio]

✅ Slice 1 landed via #295 (squash-merged to main).

map-review --cross-ai <runtime> dispatches the review to an INDEPENDENT external AI CLI (codex/gemini/claude/opencode) for a true second opinion. Shipped:

• Python step runner owns dispatch+parse+normalize+untrusted-wrap (run_cross_ai_review/dispatch_cross_ai_review, producer-owns-parse); skill handles only consent+presentation
• runtime adapter registry (hardcoded allowlist) {binary, argv, envelope, independent_vendor}
• double-consent egress (off by default): --cross-ai flag AND review.cross_ai.enabled: true
• outbound secret scan blocks dispatch before the subprocess (pattern name only, never value)
• shell=False literal-argv + configurable timeout
• inbound EXTERNAL UNTRUSTED REFERENCE fence (link/injection scan, applied in Python so the model can't skip it); findings advisory-only (source: cross_ai)
• honest independent_vendor: false for same-vendor (claude)
• non-blocking degradation → fall back to in-session review on disabled/unavailable/timeout/error/unparsed/secret_blocked
• config keys review.cross_ai.{enabled,runtime,timeout_seconds}; tests + docs (USAGE/ARCHITECTURE/CHANGELOG)

make check green locally (ruff/mypy/pyright 0, 2880 tests, render check) and GitHub CI green (build + test matrix macos/ubuntu × 3.11/3.12). Design was llm-council-reviewed (conv 92a7f159).

Deferred to slice 2 (issue stays open): --cross-ai all — multi-runtime fan-out with consensus/disagreement aggregation. The Finding/source shape is already designed to support it without a schema migration.