config.yaml

# Argus — multi-model code review skill
# Registry of reviewers, named profiles, and host-CLI adaptation rules.
#
# Routes:
#   aichat        — invoke via `aichat -m CLIENT:MODEL -S` with prompt piped to stdin.
#                   CLIENT must match an entry under `aichat_clients:` below and a
#                   corresponding client in ~/.config/aichat/config.yaml (generated by
#                   scripts/install_aichat.py).
#   gemini-cli    — invoke the Gemini CLI (uses your paid sub)
#   codex-cli     — invoke the Codex CLI (uses your paid sub)
#   claude-cli    — invoke the Claude CLI (auto-added when host != claude)
#   opencode-cli  — invoke the OpenCode CLI (uses your Go sub)
#
# Every reviewer: primary route + optional fallback.
# Fallbacks are attempted only when primary returns non-zero or times out.

version: 1

defaults:
  profile: standard
  # ───────────────────────────────────────────────────────────────────
  # Route preference — which provider a reviewer tries FIRST when it has
  # BOTH a direct-API route and an OpenRouter route (currently glm-5.2,
  # minimax-m3, deepseek-v4-pro).
  #   openrouter  → OpenRouter first, direct API as fallback  (PUBLIC DEFAULT:
  #                 one OPENROUTER_API_KEY covers most reviewers)
  #   direct      → each provider's own API first, OpenRouter as fallback
  #                 (cheaper / your own subs; what you'd use if your OR
  #                 balance is depleted)
  # CLI reviewers (codex/claude/opencode/gemini) are NEVER reordered — their
  # CLI sub stays primary and OpenRouter stays a true fallback.
  # Override per-run with --prefer-direct / --prefer-openrouter / --route-pref,
  # or the ARGUS_ROUTE_PREF environment variable. Precedence:
  #   CLI flag  >  ARGUS_ROUTE_PREF  >  this config value.
  route_preference: openrouter
  privacy_mode: true              # exclude reviewers with privacy: LOGS
  allow_free: false               # require --allow-free to include free-tier routes
  allow_logging: false            # require --allow-logging to include privacy: LOGS
  confidence_threshold: 80        # drop findings with effective confidence < this
  corroboration_boost: 15         # confidence bonus when 2+ reviewers agree on same file:line
  merge_line_tolerance: 3         # cluster findings within +/- this many lines (merge.py)
  ctx_safety_ratio: 0.70          # skip reviewer if prompt_tokens > ctx * this
  reviewer_timeout_sec: 360       # bumped from 180 on 2026-05-06 after a 4-doc
                                  # ~80K-token bundle had 5/7 reviewers timeout.
                                  # 360s handles bundles up to ~150K tokens; for
                                  # larger bundles, callers should split or
                                  # raise per-call via --timeout in dispatch.py.
  max_parallel: 4                 # cap on concurrent reviewers within a single
                                  # dispatch.py invocation; also the recommended
                                  # cap on parallel subagents when the calling
                                  # agent fans out subagent-per-reviewer.
  review_cost_warn_usd: 0.50
  review_cost_block_usd: 2.00
  benchmark_cost_warn_usd: 10.00
  benchmark_cost_block_usd: 30.00
  daily_soft_cap_usd: 10.00
  default_output_tokens_est: 800
  prompt_overhead_tokens: 500
  # OR balance pre-flight: require available ≥ estimate × this factor.
  or_balance_safety_factor: 2.0

# ─────────────────────────────────────────────────────────────────────
# Reviewers
# ─────────────────────────────────────────────────────────────────────
reviewers:

  glm-5.2:
    display: "GLM-5.2 (Zhipu / z.ai — direct + OpenRouter dual route)"
    role: [security, logic, code]
    ctx: 131072
    tier: paid
    privacy: ok
    cost_per_m: { input: 0.60, output: 2.20 }
    # Dual route. Declared order is direct-first; `defaults.route_preference`
    # (or --prefer-*) decides which is actually tried first at dispatch time.
    # The OpenRouter reasoning-only trap (content=null on large prompts via
    # Io Net) is mitigated by aichat_clients.openrouter.patch below.
    primary:  { route: aichat, client: zai,        model: "glm-5.2" }
    fallback: { route: aichat, client: openrouter, model: "z-ai/glm-5.2" }

  minimax-m3:
    display: "MiniMax M3 (MiniMax — direct + OpenRouter dual route)"
    role: [style, correctness, agent]
    ctx: 204800
    tier: paid
    privacy: ok
    cost_per_m: { input: 0.30, output: 1.20 }
    primary:  { route: aichat, client: minimax,    model: "MiniMax-M3" }
    fallback: { route: aichat, client: openrouter, model: "minimax/minimax-m3" }

  opencode-minimax-m3:
    display: "MiniMax M3 via OpenCode (minimax coding plan)"
    role: [correctness, code, agent]
    ctx: 200000
    tier: paid
    privacy: ok
    custom_only: true
    cost_per_m: null   # opencode sub
    primary: { route: opencode-cli, model: "minimax-coding-plan/MiniMax-M3" }

  opencode-glm-5.2:
    display: "GLM-5.2 via OpenCode (Ollama Cloud)"
    role: [security, logic, code]
    ctx: 131072
    tier: paid
    privacy: ok
    custom_only: true
    cost_per_m: null   # opencode sub
    primary: { route: opencode-cli, model: "ollama-cloud/glm-5.2" }

  kimi-k2.6:
    display: "Kimi K2 (OR; k2.5 GA — k2.6 preview not yet on OR)"
    role: [long-context, agent, code]
    ctx: 262144
    tier: paid
    privacy: ok
    cost_per_m: { input: 0.60, output: 2.50 }
    # KIMI_API_KEY is for chat.kimi.com (consumer), not Moonshot Platform. Use OR.
    primary:  { route: aichat, client: openrouter, model: "moonshotai/kimi-k2.5" }
    fallback: { route: aichat, client: openrouter, model: "moonshotai/kimi-k2-thinking" }

  mimo-v2-pro:
    display: "Xiaomi MiMo-V2-Pro"
    role: [long-context, code]
    ctx: 1048576
    tier: paid
    privacy: ok
    cost_per_m: { input: 1.00, output: 3.00 }
    primary:  { route: aichat, client: openrouter, model: "xiaomi/mimo-v2-pro" }

  qwen-3.6-plus:
    display: "Qwen3.6 Plus (Alibaba)"
    role: [style, agent, code]
    ctx: 1048576
    tier: paid
    privacy: ok
    cost_per_m: { input: 0.50, output: 2.00 }
    primary:  { route: aichat, client: openrouter, model: "qwen/qwen3.6-plus" }

  grok-4.20:
    display: "Grok 4.20 (xAI)"
    role: [long-context, reasoning]
    ctx: 2000000
    tier: paid
    privacy: ok
    cost_per_m: { input: 3.00, output: 15.00 }
    primary:  { route: aichat, client: openrouter, model: "x-ai/grok-4.20" }

  # Superseded by deepseek-v4-pro (kept for benchmark-history continuity).
  # No longer in any default profile; name it via --custom/--models to use it.
  deepseek-v3.2:
    display: "DeepSeek V3.2 (superseded by deepseek-v4-pro)"
    role: [reasoning, security, code]
    ctx: 131072
    tier: paid
    privacy: ok
    custom_only: true
    cost_per_m: { input: 0.27, output: 1.10 }
    primary:  { route: aichat, client: openrouter, model: "deepseek/deepseek-v3.2" }

  deepseek-v4-pro:
    display: "DeepSeek V4 Pro (1.6T MoE, 49B active, 1M ctx — direct + OpenRouter dual route)"
    role: [reasoning, security, code, long-context]
    ctx: 1048576
    tier: paid
    privacy: ok
    cost_per_m: { input: 0.435, output: 0.87 }
    # Dual route: DeepSeek's own platform (api.deepseek.com) + OpenRouter.
    # `route_preference` / --prefer-* selects which is tried first.
    primary:  { route: aichat, client: deepseek,   model: "deepseek-v4-pro" }
    fallback: { route: aichat, client: openrouter, model: "deepseek/deepseek-v4-pro" }

  # Custom-only — not included in any default profile. Available via --custom or --models.
  hermes-4.3:
    display: "Nous Hermes 4.3 36B"
    role: [style, reasoning]
    ctx: 131072
    tier: paid
    privacy: ok
    custom_only: true
    cost_per_m: { input: 1.50, output: 4.50 }
    # Nous Portal model ID — only used when NOUSRESEARCH_API_KEY is set
    # (it isn't in this env, so fallback via OR is what actually runs).
    # Verified OR slug; direct slug unconfirmed — update on first successful call.
    primary:  { route: aichat, client: nous,       model: "Hermes-4.3-36B" }
    fallback: { route: aichat, client: openrouter, model: "nousresearch/hermes-4-405b" }

  # Paid-CLI reviewers — no cost attached (use host subscription)
  gemini:
    display: "Gemini (Google CLI — ON HOLD — Windows .cmd shim doesn't tree-kill, hangs)"
    role: [frontier, long-context]
    ctx: 2000000
    tier: paid
    privacy: ok
    disabled: true      # on hold — run_subprocess now tree-kills on timeout; re-test on Windows before re-enabling (use gemini-or meanwhile)
    cost_per_m: null
    primary:  { route: gemini-cli }
    fallback: { route: aichat, client: openrouter, model: "google/gemini-3.1-pro-preview" }

  gemini-or:
    display: "Gemini 2.5 Flash via OpenRouter (pro variant needed more OR credits than available)"
    role: [frontier, long-context]
    ctx: 1048576
    tier: paid
    privacy: ok
    cost_per_m: { input: 0.30, output: 2.50 }
    primary:  { route: aichat, client: openrouter, model: "google/gemini-2.5-flash" }

  codex:
    display: "GPT-5.4 via Codex CLI"
    role: [frontier, reasoning, code]
    ctx: 400000
    tier: paid
    privacy: ok
    cost_per_m: null
    primary:  { route: codex-cli }
    fallback: { route: aichat, client: openrouter, model: "openai/gpt-5.4" }

  claude:
    display: "Claude via claude CLI"
    role: [frontier, reasoning]
    ctx: 200000
    tier: paid
    privacy: ok
    cost_per_m: null
    primary: { route: claude-cli }

  opencode:
    display: "OpenCode (your Go sub)"
    role: [agent, code]
    ctx: 200000
    tier: paid
    privacy: ok
    cost_per_m: null
    primary: { route: opencode-cli }

  # ELIMINATED: Copilot CLI's agent harness returns prose instead of our JSON
  # schema, yielding F1=0 across all non-clean fixtures. Not a retry issue —
  # fundamental prompt/harness mismatch. Left here as a reference entry only.
  copilot-gpt5:
    display: "GitHub Copilot CLI (GPT-5.2) — ELIMINATED (harness returns prose not JSON)"
    role: [frontier, code]
    ctx: 128000
    tier: paid
    privacy: ok
    custom_only: true
    disabled: true
    cost_per_m: null
    primary: { route: copilot-cli, model: "gpt-5.2" }

# ─────────────────────────────────────────────────────────────────────
# Named profiles
# ─────────────────────────────────────────────────────────────────────
profiles:
  quick:
    description: "2-reviewer smoke test"
    members: [glm-5.2, gemini-or]
  standard:
    description: "Balanced everyday review — your favorites + paid CLIs + fast Gemini"
    members: [glm-5.2, minimax-m3, gemini-or, codex]
  panel:
    description: "Maximum coverage (host CLI auto-skipped; copilot, gemini-direct, grok disabled from default)"
    members: [glm-5.2, minimax-m3, kimi-k2.6, mimo-v2-pro, qwen-3.6-plus, deepseek-v4-pro, gemini-or, codex, claude, opencode]
  security:
    description: "Auth / crypto / input-handling changes"
    members: [glm-5.2, deepseek-v4-pro, codex, claude]
    prompt_overlay: security
  deep:
    description: "Long-context, large diffs, cross-file correctness (grok excluded — use --custom if 2M ctx needed)"
    members: [mimo-v2-pro, gemini-or, kimi-k2.6, deepseek-v4-pro, codex]
    prompt_overlay: deep
  favorites:
    description: "Your picks (GLM-5.2 + MiniMax M3 direct subs)"
    members: [glm-5.2, minimax-m3]
  direct:
    description: "Direct-API subs only, no Gemini — pair with route_preference=direct (or --prefer-direct) when your OpenRouter balance is depleted"
    members: [glm-5.2, minimax-m3, deepseek-v4-pro, codex, claude, opencode]
  leaderboard-top5:
    description: "Post-bench top-5 by F1 (opencode, qwen3.6+, glm-5.2, gemini-or, minimax)"
    members: [opencode, qwen-3.6-plus, glm-5.2, gemini-or, minimax-m3]

# ─────────────────────────────────────────────────────────────────────
# Host-CLI adaptation
# When Argus detects it's running inside one of these CLIs, the matching
# reviewer is removed from the roster and (optionally) another is added.
# ─────────────────────────────────────────────────────────────────────
host_rules:
  claude:   { skip: [claude],   add: [] }
  codex:    { skip: [codex],    add: [claude] }
  gemini:   { skip: [gemini],   add: [claude] }
  opencode: { skip: [opencode], add: [claude] }
  unknown:  { skip: [],         add: [] }

# ─────────────────────────────────────────────────────────────────────
# aichat client definitions — materialized into ~/.config/aichat/config.yaml
# by scripts/install_aichat.py. API keys NEVER written to disk; passed via
# subprocess env as AICHAT_<CLIENT>_API_KEY from the source env vars listed.
# ─────────────────────────────────────────────────────────────────────
aichat_clients:
  openrouter:
    type: openai-compatible
    api_base: https://openrouter.ai/api/v1
    api_key_env: OPENROUTER_API_KEY
    # Suppress reasoning field (we just want content) and route away from
    # providers that return reasoning-only with content: null on large prompts.
    # Confirmed: aichat 0.30 honors `patch.chat_completions` body merges.
    patch:
      chat_completions:
        ".*":
          body:
            # Exclude the reasoning field from responses entirely.
            reasoning:
              exclude: true
              # Cap the provider's internal reasoning budget. Without this,
              # reasoning models (z-ai/glm-5.2, minimax/minimax-m3) can
              # spend 3+ minutes thinking on a 30K-token diff even when we
              # don't want to see the thoughts. 2000 tokens ≈ 30-60s budget.
              max_tokens: 2000
            # Route away from known reasoning-only providers that return
            # content=null when reasoning is trimmed or budget-capped.
            provider:
              ignore: ["io.net", "together.ai"]
  zai:
    type: openai-compatible
    api_base: https://api.z.ai/api/coding/paas/v4   # Coding Plan endpoint (your sub); standard /paas/v4 requires separate billing
    api_key_env: ZAI_API_KEY
  minimax:
    type: openai-compatible
    api_base: https://api.minimaxi.chat/v1
    api_key_env: MINIMAX_API_KEY
  deepseek:
    type: openai-compatible
    api_base: https://api.deepseek.com/v1
    api_key_env: DEEPSEEK_API_KEY
  moonshot:
    type: openai-compatible
    api_base: https://api.moonshot.ai/v1
    api_key_env: KIMI_API_KEY                 # also accepts MOONSHOT_API_KEY
  nous:
    type: openai-compatible
    api_base: https://inference-api.nousresearch.com/v1
    api_key_env: NOUSRESEARCH_API_KEY

# ─────────────────────────────────────────────────────────────────────
# CLI invocation templates
# {client} and {model} are substituted for aichat; others take prompt on stdin.
# ─────────────────────────────────────────────────────────────────────
cli_commands:
  aichat:       ["aichat", "-m", "{client}:{model}", "-S"]   # prompt via stdin
  gemini-cli:   ["gemini", "--yolo", "-p", ""]               # prompt via STDIN (empty -p forces non-interactive; Windows ARG_MAX broke {prompt} for >32KB diffs, fixed 2026-05-06)
  codex-cli:    ["codex", "exec", "--skip-git-repo-check", "-"]   # stdin input; skip git-repo trust check
  claude-cli:   ["claude", "-p", "--output-format", "text", "--bare"]   # stdin
  opencode-cli: ["opencode", "run", "-"]                     # stdin input
  copilot-cli:  ["copilot", "-p", "Follow the review instructions provided on stdin.",
                 "--model", "{model}", "--allow-all-tools", "--no-color",
                 "--output-format", "text"]                  # prompt via STDIN (Windows ARG_MAX; same fix as gemini-cli)