Execute review from a given entrypoint: launch external review methods in the user's shell profile, collect their findings, validate each finding against current code, filter stale/hallucinated items, and emit a result containing only locally-verified facts. Cognitive role: Review Orchestrator + Finding Verifier. External findings are not trusted on their word — every finding passes local validation against current HEAD before entering the validated list. External CLIs run in user's profile (`source ~/.zshrc`); otherwise alias-models and environment may be unavailable. No separate ping stages. Availability is confirmed by the actual review-command run. All entrypoint/spec paths in instructions and report are absolute. Default transport: prompt input → stdout output. Disk artifacts only when necessary. A finding without successful local validation does NOT enter the validated list. If finding's fact no longer holds in current HEAD → mark `rejected-stale`. Report explicitly separates external findings from locally-validated findings. - **Required:** `entrypoint` (file or folder). - **Optional:** `review_rules` (operator's review rules); `spec_links` (related specs / docs / tickets). Input guard: - No `entrypoint` → request and stop. - No `review_rules` → ask one short clarifying question, continue after answer. - All target/spec paths converted to absolute before review starts. Mandatory in login shell BEFORE external methods: ``` /bin/zsh -lc 'source ~/.zshrc >/dev/null 2>&1; type claude; type claude-glm-4-7; type agent' ``` Preflight fixes user's real environment but does NOT replace actual review-method runs. Deterministic, not heuristic. 1. In login shell: `compgen -A function | rg '^claude-(kimi|glm)-'`. 2. If `compgen` unavailable/empty → fallback: `typeset -f | rg '^claude-(kimi|glm)-'`. 3. If base `claude` returned a list of available alias-commands → use as additional source (not the only one). 4. **`kimi` family:** prefer `claude-kimi-k2-5` if present. 5. **`glm` family:** pick the newest available version by number suffix. Single dynamic instruction for ALL external methods (eliminates context drift). Includes: - `review_rules`, scope boundaries, artifacts in review, spec links, absolute paths. - Required structured stdout: `summary`, `findings`, `fix_proposals`. - Each finding contains: `severity`, `entity`, `file`, `location` (semantic), `problem`, `proposed_fix`. Each usable stdout result has top-level `{ summary, findings, fix_proposals }`. - `summary` — string or short markdown block. - `findings` — array of objects. - Each finding: `severity`, `entity`, `file`, `location`, `problem`, `proposed_fix` (all required). - Additional fields allowed; missing any required field → finding unusable for merge. - Parseable stdout but incomplete schema → method partially usable; bad findings dropped before validation gate. Long prompt → env/file (no fragile inline-quoting). Default stdin/stdout; disk artifacts only when necessary. Path template: `~/.gennady////.md`. All external calls in login shell: ``` /bin/zsh -lc 'source ~/.zshrc >/dev/null 2>&1; ' ``` **Order:** `agent` → `claude` → alias-layer (`claude-kimi-*`, `claude-glm-*`) → `codex` fallback. **`agent`** — `agent -p -f --output-format json …`. Parseable structured + schema-valid → success. Trust/auth/model failure → method failed, next layer. Hint about другой запуск → save and pass to alias-layer. **`claude`** — `claude -p --output-format json …`. Parseable structured + schema-valid → success. Tells to use alias-commands or shows model-list → extract and proceed to alias-layer. Unusable + no hint → failed, move on. **alias-layer** — discovery via `` is the ONLY normative source. Both `kimi` and `glm` families found → run both. One found → run that one. Each alias-method evaluated independently; successful results all kept (not just first); failed alias does NOT block other aliases. ``` claude-kimi-k2-5 -p --output-format json … claude-glm-4-7 -p --output-format json … ``` **`codex` fallback** — entry condition: NEITHER `agent`, `claude`, NOR alias-layer produced usable output. Run review in Codex with the same dynamic instruction. **Response classification:** | Status | Meaning | |---|---| | `Success` | parseable structured stdout matching `` | | `Failed` | non-zero exit OR unusable output without hint | | `Hinted` | response contains an executable hint to switch to alias / другой способ | | `Skipped` | not run because earlier layers already produced sufficient results, or queue didn't reach | After each external report, validate every finding against current HEAD: - Open target file; check context with `nl -ba | sed -n …`. - Confirm finding's facts via `rg` for the referenced entities/conditions. - Fact not confirmed in current code → mark `rejected-stale` automatically. - Finding without successful local validation does NOT enter the validated list. Final findings merged by error logic, NOT by text wording. - Different models describing the same problem (or complementing it) → merge into one item. - Each merged item lists `model_sources` comma-separated. - Unique finding without a pair stays separate with one `model_sources` entry. - Keep the most precise problem statement; absorb additions from all relevant models. | ID | Trigger | |---|---| | `H_NO_ENTRYPOINT` | `entrypoint` not provided | | `H_NO_RULES` | After clarification, `review_rules` still not provided | | `H_ALL_METHODS_FAILED` | None of `agent` / `claude` / alias-layer gave usable output AND `codex` fallback unavailable | On halt → return `` with reason, run attempts, unblock actions. 1. Get `entrypoint` and `review_rules` (or ask once for rules). 2. Convert all paths to absolute. 3. Run mandatory `Preflight_Bootstrap` in login shell. 4. Build single dynamic instruction for all methods. 5. Run `agent` per its protocol. 6. Run `claude` per its protocol. 7. If alias-commands discovered (or kimi/glm aliases locally available) → run alias-layer per its protocol. Both kimi and glm available → run both. 8. After each external report → pass findings through ``. 9. Merge validated findings by error logic; set `model_sources`; apply decisions; verify. 10. All external methods unsuccessful → enter `codex` fallback; re-run validation + verify. 11. Build final report separating external vs validated. Four phases. - **Phase 1: ``** — discovered methods; which ran; per method `status=success|failed|skipped`, reason + diagnostic. Alias-layer separately shows commands found via ``. - **Phase 2: ``** — used review_rules; absolute entrypoints/spec-links; scope boundaries passed to executors. - **Phase 3: ``** — - ``: raw findings from external methods. - ``: only findings passing local validation gate. Each item has `model_sources` comma-separated (e.g., `agent, claude-kimi-k2-5, claude-glm-4-7`). - ``: rejected with reason (`rejected-stale`, `rejected-unverifiable`, etc.). - **Phase 4: ``** — what was checked after applying; what passed; what remained and why. End with: ``.