name: argus description: >- Multi-model code review. Dispatches a diff / PR / files to a configurable roster of reviewers (GLM-5.2, MiniMax M3, Kimi K2.6, MiMo-V2-Pro, Qwen3.6-Plus, Grok 4.20, DeepSeek V4 Pro, Gemini CLI, Codex CLI, Claude CLI, OpenCode CLI) in parallel, applies a confidence filter with cross-reviewer corroboration, and produces one merged review. Includes benchmark mode that runs the full fixture suite to establish preferred reviewers. argument-hint: "[--profile NAME | --custom LIST | --models LIST] [--pr URL | --files GLOB | -] [--route-pref {openrouter,direct} | --prefer-direct | --prefer-openrouter] [--benchmark] [--stats] [--dry-run] [--yes-cost] [--allow-free] [--save-as NAME] [--output {md,json,gsd}]" allowed-tools:

Bash
Read
Write
Edit
Glob
Grep

<objective> Run a multi-model code review using the Argus skill at `d:/projects/.skills/argus/`. </objective> <context> - Skill root: `d:/projects/.skills/argus/` (env: `ARGUS_HOME`) - Config: `$ARGUS_HOME/config.yaml` - Scripts: `$ARGUS_HOME/scripts/*.py` (Python 3.12) - Runs: `$ARGUS_HOME/runs/<ISO-ts>-<id>/` - Benchmarks: `$ARGUS_HOME/benchmarks/<ts>.{md,json}` - History: `$ARGUS_HOME/history.db` (SQLite, created on first run) </context> <execution>

1. Parse arguments

Extract from $ARGUMENTS:

| Flag | Effect | |---|---| | --profile NAME | Use named profile (quick / standard / panel / security / deep / favorites / direct / leaderboard-top5 / saved-custom-name) | | --route-pref {openrouter,direct} | Route preference for dual-route reviewers (glm-5.2 / minimax-m3 / deepseek-v4-pro). openrouter (default) tries OpenRouter first; direct tries each provider's own API first. Shorthands: --prefer-direct, --prefer-openrouter. Env: ARGUS_ROUTE_PREF. | | --custom "a,b,c" | One-off roster; combine with --save-as NAME to persist | | --models "a,b,c" | Alias for --custom, never saved | | --pr URL | Diff via gh pr diff URL | | --files GLOB | Scope via git diff -- GLOB | | - or no scope | Stdin, else git diff HEAD | | --benchmark [--runs N] | Run fixture suite (default N=3), produce leaderboard | | --stats | Per-reviewer accuracy, cost, agreement matrix from history.db | | --dry-run | Print roster + cost estimate, do not dispatch | | --yes-cost | Bypass $2 hard block for this run | | --allow-free | Allow free-tier reviewers | | --allow-logging | Include reviewers that log prompts/completions | | --save-as NAME | Persist --custom list as a named profile | | --output {md,json,gsd} | Output format (default: md) | | --refresh | Check for model ID updates on OpenRouter | | --compare RUN_A RUN_B | Diff two previous runs |

Env overrides: ARGUS_PROFILE, ARGUS_YES_COST=1, ARGUS_OUTPUT=gsd, ARGUS_ROUTE_PREF={openrouter,direct}.

2. Detect host CLI and adapt roster

HOST=$(python "$ARGUS_HOME/scripts/detect_host.py")

Apply host_rules[$HOST] from config.yaml: remove skip list, append add list.

3. Resolve the roster

Load config.yaml
Base list from --profile / --custom / --models / default profile
Apply host rules (step 2 above). For explicitly-named rosters (--custom / --models) apply only the skip list — never inject add reviewers the user didn't ask for.
Apply filters:
- drop reviewers with disabled: true unless explicitly named
- drop reviewers with tier: free unless --allow-free
- drop reviewers with privacy: LOGS unless --allow-logging
- drop reviewers with custom_only: true unless explicitly named
Dedupe

If roster is empty, abort with message and list disabled reasons.

Defense in depth: dispatch.py and benchmark.py enforce this same policy internally via _common.resolve_roster (drops recorded in dispatch_summary.json under "dropped"), and estimate_cost.py rejects unknown reviewer names — so a skipped filter here fails safe downstream.

4. Gather scope → diff

Produce a single diff.patch in the new run directory:

--pr URL → gh pr diff URL > diff.patch
--files GLOB → git diff -- GLOB > diff.patch
stdin present → read stdin into diff.patch
else → git diff HEAD > diff.patch

Apply .argusignore (project root) to strip matching paths from diff.

If diff is empty after filtering, abort.

5. Pre-flight cost estimate

python "$ARGUS_HOME/scripts/estimate_cost.py" --roster "$ROSTER" --diff "$RUN_DIR/diff.patch"

If the user passed --route-pref / --prefer-direct / --prefer-openrouter (or ARGUS_ROUTE_PREF is set), forward the same flag to estimate_cost.py, dispatch.py, and benchmark.py so the cost/balance gate and dispatch agree on which provider each dual-route reviewer hits first.

Exit codes: 0 OK, 1 WARN (≥ warn threshold or low OpenRouter balance), 2 BLOCK (≥ block threshold) or invalid roster — check stderr to tell them apart; --yes-cost downgrades a cost BLOCK to WARN but cannot fix an invalid roster. On BLOCK without --yes-cost: stop. Report estimate to user.

If --dry-run: stop here and print roster + estimate. Do not dispatch.

6. Dispatch

Argus supports two dispatch modes. Subagent-per-reviewer is the default recommended pattern — single-process is the legacy / quick path.

6a. Default: subagent-per-reviewer (recommended)

When more than one reviewer is in the roster, the calling agent should dispatch ONE subagent per reviewer via the parent platform's Agent / subagent API (Claude Code: Agent tool with run_in_background: true on the parent's Agent call — so the parent fans out concurrent subagents without blocking). Each subagent runs:

python "$ARGUS_HOME/scripts/dispatch.py" \
  --run-dir "$RUN_DIR" \
  --roster "<single-reviewer-name>" \
  --diff   "$RUN_DIR/diff.patch" \
  [--overlay security|deep]

CRITICAL: subagent execution semantics

The subagent MUST run dispatch.py synchronously in its OWN session (foreground) and wait for the process to exit before ending. If the subagent uses its own run_in_background: true to spawn dispatch.py, the subagent's session ends BEFORE dispatch.py completes, and $RUN_DIR/reviews/<name>.json is never written. The parent then sees a "completed" subagent with no result.

This is observed when a subagent reports back with a plan ("I'll wait for the dispatch...") instead of actual results — that's the fingerprint of the bailout bug.

Always include explicit synchronous-execution wording in the subagent prompt. A minimal safe template:

You are dispatching a SINGLE argus reviewer. Do not do any other work.

Run this command IN THE FOREGROUND (synchronously) and WAIT for it to
exit. Do NOT use run_in_background. This may take 30s–5min depending on
the reviewer; that is expected — block your session until it returns.

    python "$ARGUS_HOME/scripts/dispatch.py" \
      --run-dir "<RUN_DIR>" \
      --roster "<reviewer-name>" \
      --diff   "<RUN_DIR>/diff.patch" \
      [--overlay security|deep]

After it exits, read `<RUN_DIR>/reviews/<reviewer-name>.json` and
report back in under 100 words:

- exit code from the dispatch command
- latency_sec field from the JSON
- count of findings (length of .findings array, 0 if missing)
- skipped status + skip_reason if set
- error field if present
- fallback_used flag

Do NOT run merge.py — the parent session handles that.
Do NOT modify any files. Do NOT do any other work.

If a subagent's result-reporting indicates it never read the review JSON (or the file is missing on disk), re-dispatch that single reviewer either as a fresh subagent with stricter prompt wording, or directly in the parent session with run_in_background: true on the parent's Bash call (parent stays responsive while waiting for completion).

Concurrency cap

At most 4 subagents running in parallel. If the roster has more than 4 reviewers, dispatch the first 4 in a wave, then queue the rest and dispatch as each completes. This matches defaults.max_parallel: 4 in config.yaml and keeps per-reviewer visibility (independent timestamps, streaming progress, isolated failure modes).

6b. Legacy / quick path: single-process dispatch

For a roster of ≤4 reviewers OR when per-reviewer subagent visibility isn't needed, the calling agent MAY run argus's in-process dispatch:

python "$ARGUS_HOME/scripts/dispatch.py" \
  --run-dir "$RUN_DIR" \
  --roster "$ROSTER" \
  --diff   "$RUN_DIR/diff.patch" \
  [--overlay security|deep]

This still uses argus's internal max_parallel: 4 cap from config.yaml.

6c. Failure semantics (both modes)

Each reviewer writes $RUN_DIR/reviews/<name>.json. Timeouts, non-zero exits, and parse failures are logged but do not abort the run; other reviewers continue. After all reviewers return (subagent or in-process), run merge.py exactly once in the main session (step 7).

7. Merge + filter

python "$ARGUS_HOME/scripts/merge.py" --run-dir "$RUN_DIR"

Applies the confidence threshold, corroboration boost, and anchor-based ±3-line clustering (defaults.merge_line_tolerance / --line-tolerance), writes merged.md and metrics.json. Also appends a row to history.db.

8. Present

Print merged.md. Include summary line: N findings, corroborated/solo breakdown, wall time, total estimated cost, run directory path.

9. Optional: benchmark mode

python "$ARGUS_HOME/scripts/benchmark.py" --runs 3 [--profile panel | --roster LIST]

Runs every roster member against every fixture N times. Produces benchmarks/<ts>.md (leaderboard) and benchmarks/<ts>.json. Uses the separate benchmark cost gate ($10 warn / $30 block).

10. Optional: stats mode

python "$ARGUS_HOME/scripts/stats.py"

Per-reviewer: runs, avg cost, avg latency, accuracy (if benchmarks present), pairwise agreement matrix. Flags redundant pairs (>85% agreement) as candidates for demotion to fallbacks/custom.

</execution> <rules> - Respect `privacy_mode: true` by default. Never include logger reviewers without `--allow-logging`. - Never pass free-tier reviewers without `--allow-free`. - Never exceed review cost block threshold without `--yes-cost` (env: `ARGUS_YES_COST=1`). - Never rewrite config.yaml without explicit user approval (only `--save-as` adds profiles, and only after showing the diff). - Reviewer timeouts or bad output never abort the run; they're logged and the run continues with surviving reviewers. - After every run, append to `history.db`. </rules>

Argus: Revue de code multi-modèle

Recommandé pour