# AI Consultants - Full Configuration Reference Configuration sources, in order of precedence (highest wins): 1. **CLI flags** — `--preset`, `--strategy`, etc. 2. **Existing env vars** — `export FOO=bar` before invoking 3. **User config dir** (v2.12+) — `~/.config/ai-consultants/{config.sh,.env}` 4. **`config.sh` defaults** — the `${VAR:-default}` fallbacks 5. **Hardcoded defaults** in individual scripts For goal-oriented, copy-paste configurations, start with [`docs/RECIPES.md`](../docs/RECIPES.md). This file is the variable reference; [`scripts/config.sh`](../scripts/config.sh) is the executable source of truth. ## Automatic Configurator The public configurator detects all 11 supported consultants, chooses CLI or API transport from the available CLI binaries and credentials, and writes the persistent XDG configuration: ```bash ai-consultants configure ``` It preserves existing custom values and secrets, refreshes `ENABLE_*` flags from detected availability, creates a timestamped backup, writes the result with mode `600`, and warns when fewer than two consultants are usable. Pin an availability decision with a final `--set ENABLE_=true|false` override. Automatically chosen `*_USE_API` modes are stored with an `# ai-consultants:auto` marker and are recalculated on later runs. Remove the marker or use an environment variable/`--set` to pin a transport explicitly. It never performs a billed authentication probe; use `ai-consultants doctor --live` when you explicitly want a live provider check. ```bash # Guided consultant and transport review ai-consultants configure --interactive # Review every supported persistent parameter ai-consultants configure --advanced # Fully automated, repeatable configuration ai-consultants configure \ --set DEFAULT_PRESET=security \ --set ENABLE_HEALTH_GATE=true \ --set QUORUM_ACTION=stop # Discover the exact accepted keys or preview without exposing secrets ai-consultants configure --show-parameters ai-consultants configure --dry-run ``` `--set KEY=VALUE` is repeatable and fails closed for unknown or removed keys. Use the hidden interactive credential prompts (or exported environment variables) for API keys; command-line arguments may be visible in shell history and process inspection. Model defaults carry an `# ai-consultants:default` marker. On rewrite, `configure` upgrades the exact historical defaults it generated, marked or unmarked, unless they are pinned: `CLAUDE_MODEL=claude-opus-4-8`/`claude-opus-5` to Opus 5.5, plus the previous Gemini (CLI and API), Codex, Mistral API, Qwen, GLM, Grok and DeepSeek defaults, and the fixed timeouts described under [Timeouts, Waiting and Retries](#timeouts-waiting-and-retries). An explicit model override supplied with `--set` is stored with `# ai-consultants:pin`, so `--set CLAUDE_MODEL=claude-opus-5-5` keeps Opus 5.5 (the standard/medium target) as the default. Retired models (5.5.0, listed with their replacements under `retired_models` in `docs/cost_rates.json`) are replaced even when pinned: the marker is kept, the old value is recorded as `migrated-from=`, and `--set` refuses them. A retired or uncatalogued model selected at run time still runs, with a warning, and its cost is reported as unpriced. The exhaustive parameter contract is [`.env.example`](../.env.example); the configurator derives its accepted keys from that template, while [`scripts/config.sh`](../scripts/config.sh) remains the runtime source of truth. A regression test fails when a persistent runtime default is missing from the configurator. ## User Config Dir (v2.12+) Persistent overrides live in `~/.config/ai-consultants/`. The directory and starter files are scaffolded by: ```bash ai-consultants init # creates the dir + .env + config.sh ai-consultants init --force # overwrites existing files ``` `init` only scaffolds files for manual editing. Prefer `configure` when you want automatic detection and a ready-to-use panel. Search order for the directory: | Priority | Source | |----------|--------| | 1 | `$AI_CONSULTANTS_CONFIG_DIR` | | 2 | `$XDG_CONFIG_HOME/ai-consultants` | | 3 | `$HOME/.config/ai-consultants` | Files loaded from that dir (in order, both optional): - **`.env`** — KEY=value lines, parsed and exported. Existing env vars are NOT overridden. `export` prefix and `# comments` supported. Recommended chmod 600 (contains API keys). - **`config.sh`** — full bash, sourced after `.env`. Use `${VAR:-default}` to defer to env. - **`affinity.json`** — picked up by `lib/routing.sh` if present (overrides bundled matrix; superseded by `AFFINITY_FILE`). `./scripts/doctor.sh` reports which user config files are loaded under "Checking User Config (v2.12)". ## Defaults ```bash DEFAULT_PRESET= # Preset when --preset is not given; empty = every enabled consultant DEFAULT_STRATEGY=coverage # Strategy when --strategy is not given ``` `.env.example` leaves `DEFAULT_PRESET=balanced` commented out; set it (for example with `configure --set DEFAULT_PRESET=balanced`) to run three consultants by default instead of every enabled one. ## Context Handoff (v2.14+) The way file context flows from the invoking agent to the consultants. Agents pass file paths as positional arguments to `consult_all.sh` rather than inlining file contents into the query string — this is what lets `build_context.sh` run the AST/chunking optimizer. ### File path syntax — `path@TAG` When passing files to `consult_all.sh` (directly or via slash command), each path may carry a `@TAG` suffix: | Tag | Meaning | |---|---| | `@PRIMARY` | Focus of the question; what consultants should critique. Default if `@TAG` omitted. | | `@CONTEXT` | Ambient reference; read but not the target of the critique. | ```bash ./scripts/consult_all.sh "Why does auth fail under load?" \ src/auth.ts@PRIMARY src/cache.ts@CONTEXT src/logger.ts@CONTEXT ``` Unknown tags fall back to `PRIMARY` with a `log_warn` message. The tag appears in the rendered `### File: ... (TAG)` header so consultants can weigh files differently. ### `--query-file ` flag Use it when the question would exceed the shell's argument limits (128 KiB per argument on Linux, about 1 MiB in total on macOS), contains awkward quoting (JSON, Python dicts with mixed quotes), or should stay out of the process list. Conflicts with a positional question argument. ```bash q=$(mktemp) && chmod 600 "$q" printf '%s\n' "Long multi-paragraph question..." > "$q" ./scripts/consult_all.sh --query-file "$q" src/big.py@PRIMARY rm -f "$q" ``` ### Category-aware project tree `build_context.sh` reads `QUESTION_CATEGORY` (already exported by `consult_all.sh` after `classify_question.sh` runs) to decide whether to include the project-tree listing. | Category | Project tree included? | |---|---| | `ARCHITECTURE`, `CODE_REVIEW`, `API_DESIGN`, `GENERAL` | Yes | | `SECURITY`, `QUICK_SYNTAX`, `ALGORITHM`, `BUG_DEBUG`, `DATABASE`, `TESTING` | No | | Unknown category | Yes (conservative default) | The tree is never listed when the context root is a home directory or `/`: walking it is slow and would send personal file names to every consultant. Override: ```bash FORCE_PROJECT_TREE=true ./scripts/consult_all.sh "..." # Always include ``` ### Token optimization mode (v2.2+, now actually engaged) ```bash TOKEN_OPTIMIZATION_MODE=ast # none, basic, ast (default), full ENABLE_AST_EXTRACTION=true # AST-based code skeleton ENABLE_SYMBOL_COMPRESSION=false # Symbol compression (opt-in) ENABLE_SEMANTIC_CHUNKING=true # Semantic chunking for large files MAX_CONTEXT_FILE_BYTES=8000 # Threshold before optimization kicks in ``` AST extractors are dedicated for **Python, JavaScript, TypeScript, Bash, Go**. Other declared languages (Rust, Java, C, C++, C#, Ruby, PHP, Swift) fall back to a `grep`-based generic extractor. ## Core Features ```bash ENABLE_PERSONA=true # Give each consultant its configured role ENABLE_SYNTHESIS=true # Automatic synthesis SYNTHESIS_CMD=claude # CLI used to synthesize the panel SYNTH_DETAIL_MAX_CHARS=4000 # Per-response detail included in synthesis SYNTHESIS_TIMEOUT=0 # Per synthesis CLI; 0 = no limit (default) SYNTHESIS_TOTAL_TIMEOUT=0 # All provider attempts together; 0 = no limit ENABLE_SMART_ROUTING=false # Category-based consultant selection ENABLE_COST_TRACKING=true # Track API usage costs ``` `SYNTHESIS_CMD` may be `claude`, `codex`, or `gemini`. The invoking agent is excluded automatically; set `INVOKING_AGENT` only for direct integrations that need to declare their host explicitly. ## Classification and Smart Routing ```bash ENABLE_CLASSIFICATION=true # Classify every question before routing CLASSIFICATION_MODE=pattern # pattern (fast) | llm (more accurate, costs a call) ENABLE_SMART_ROUTING=false # Select consultants using the affinity matrix MIN_AFFINITY=7 # Minimum category score, from 1 to 10 ``` The bundled matrix lives in [`references/affinity.json`](affinity.json). Set `AFFINITY_FILE` or place `affinity.json` in the user config directory to override it. ## CLI/API Mode Switching (v2.6+) Seven consultants support switching between CLI and API mode. **The default is the CLI** for every consultant that has one. Grok additionally falls back to the xAI API after a failed CLI launch when a key is available. When API mode is explicitly enabled, the CLI is not used. **Gemini auto-resolution (v2.15.1):** leave `GEMINI_USE_API` unset and the mode is chosen for you — API mode when `GEMINI_API_KEY` is present (no `agy` install or OAuth needed; ideal for `npx`), CLI mode (`agy`) otherwise. Set `GEMINI_USE_API` explicitly only to force a mode (an explicit value disables auto-detection). ```bash # GEMINI_USE_API unset = auto (API if GEMINI_API_KEY set, else agy CLI) GEMINI_USE_API=false # Force agy CLI; set true to force Google AI API CODEX_USE_API=false # Use OpenAI API instead of codex CLI CLAUDE_USE_API=false # Use Anthropic API instead of claude CLI MISTRAL_USE_API=false # Use Mistral API instead of vibe CLI QWEN3_USE_API=false # Use qwen CLI (default) or DashScope API GROK_USE_API=false # Use Grok Build (default); API is fallback MINIMAX_USE_API=false # Use mmx CLI (default) or MiniMax API # META_USE_API unset = auto (API if META_API_KEY set, else muse CLI) ``` ### API Keys | Agent | API Key Variable | Notes | |-------|------------------|-------| | Gemini | `GEMINI_API_KEY` | Google AI API key | | Codex | `OPENAI_API_KEY` | OpenAI API key | | Claude | `ANTHROPIC_API_KEY` | Anthropic API key | | Mistral | `MISTRAL_API_KEY` | Mistral API key | | Qwen3 | `QWEN3_API_KEY` | DashScope API key | | MiniMax | `MINIMAX_API_KEY` | MiniMax API key (API mode only; the mmx CLI uses OAuth) | | Meta | `META_API_KEY` | Meta Model API key (`LLM_...`); Muse Code also reads it headless | | GLM | `GLM_API_KEY` | Required when `ENABLE_GLM=true` | | Grok | `GROK_API_KEY` | xAI API fallback key; Grok Build CLI uses its own login | | DeepSeek | `DEEPSEEK_API_KEY` | Required when `ENABLE_DEEPSEEK=true` | ## Consultant Toggles and Models ```bash # Consultants enabled by default ENABLE_GEMINI=true ENABLE_CODEX=true ENABLE_MISTRAL=true ENABLE_KIMI=true # Kimi CLI - The Eastern Sage ENABLE_QWEN3=true # Qwen CLI/API - The Analyst ENABLE_GROK=true # Grok Build CLI/API fallback - The Provocateur ENABLE_MINIMAX=true # MiniMax CLI/API (mmx) - The Pragmatic Optimizer ENABLE_CLAUDE=true # Claude CLI - The Synthesizer (auto-excluded under Claude Code) # API-only consultants (off by default - require API keys) ENABLE_GLM=false ENABLE_DEEPSEEK=false # Opt-in: Meta (muse CLI or META_API_KEY) - The Systems Thinker ENABLE_META=false ``` ### Model Overrides ```bash GEMINI_MODEL="gemini-4-argon-high" # CLI smoke passed; capability-probed identity; API uses GEMINI_API_MODEL GEMINI_API_MODEL=gemini-4-argon # API only; needs a paid-tier key GEMINI_API_MAX_TOKENS=32768 # API only: Argon thinking + visible output CODEX_MODEL=gpt-6.1-sol # max_quality uses gpt-6-astra CODEX_API_MAX_TOKENS=16384 CLAUDE_MODEL=claude-opus-5-5 # Matching tier defaults: maximum/max, premium/xhigh, standard/medium. # CLAUDE_REASONING_EFFORT overrides the tier effort; Fable remains opt-in. CLAUDE_API_MAX_TOKENS=16384 # API only: adaptive thinking + visible output MISTRAL_MODEL=mistral-large-4 MISTRAL_CLI_MODEL=mistral-medium-3.5 MISTRAL_MAX_TURNS=4 MISTRAL_API_MAX_TOKENS=16384 # API only: Large 4 thinking + visible output KIMI_MODEL=kimi-code/k3.1-pro KIMI_REASONING_EFFORT=max # passed as KIMI_MODEL_THINKING_EFFORT; empty = CLI default QWEN3_MODEL=qwen3.8-max # standard/economy tiers use qwen3.8-flash QWEN3_MAX_QUALITY_TIMEOUT=0 # 0 = no limit QWEN3_API_MAX_TOKENS=16384 GLM_MODEL=glm-5.3-flash GLM_API_MAX_TOKENS=16384 GLM_REASONING_EFFORT= # max_quality sets max GROK_MODEL=grok-4.7 GROK_MAX_TURNS=4 GROK_API_MAX_TOKENS=4096 GROK_REASONING_EFFORT= # max_quality sets xhigh (Grok Build maximum) GROK_OAUTH_MODE=shared # shared (parallel default) | serialized (diagnostic) DEEPSEEK_MODEL=deepseek-flash # DeepSeek-V4.1-Flash API alias DEEPSEEK_API_MAX_TOKENS=16384 DEEPSEEK_MAX_QUALITY_TIMEOUT=0 # 0 = no limit DEEPSEEK_REASONING_EFFORT= # max_quality sets max MINIMAX_MODEL=MiniMax-M3.1-Flash # every tier; always reasons MINIMAX_API_MAX_TOKENS=16384 # reasoning shares the completion budget MINIMAX_MAX_TOKENS=16384 MINIMAX_MAX_QUALITY_TOKENS=32768 MINIMAX_REASONING_EFFORT= # API only: low|medium|high|xhigh|max (default max) MAX_QUALITY_API_MAX_TOKENS=16384 META_CMD=muse META_MODEL=muse-wonder-1.0 # standard/economy: muse-spark-1.3; always reasons META_REASONING_EFFORT=high # minimal..max (none rejected); max_quality sets max META_API_MAX_TOKENS=32768 # reasoning shares the completion budget META_API_URL=https://api.meta.ai/v1/chat/completions ``` Meta always runs in Standard mode, where Meta does not train on the content: the `muse` CLI gets `--contributor-mode off` on every run (subscriptions default to Contributor mode). `muse-spark-1.3-contributor` is a catalogued opt-in for the API and logs a warning whenever it is pinned. Anthropic counts adaptive thinking and visible response tokens against the same `CLAUDE_API_MAX_TOKENS` limit. The default is 16,384; a response that reaches that limit is rejected as truncated instead of entering synthesis as a complete answer. `GROK_MODEL` is passed explicitly to `grok -m`. The default transport is the headless Grok Build CLI using a private `--prompt-file` inside an isolated, tool-free strict sandbox. If the CLI is missing, cannot execute, or has no usable authentication, the same query and model fall back to `GROK_API_URL` when `GROK_API_KEY` is set. Failures after a request launches (including timeouts, model errors, and empty output) are returned without API fallback. Set `GROK_USE_API=true` only to force the API path. The adapter accepts any CLI version that exposes the complete required headless interface and requested model; `metadata.cli_version` is provenance, not a version pin. `GROK_OAUTH_MODE=shared` is the normal CLI path because ai-consultants launches consultants in parallel. Each Grok invocation keeps a private HOME, workspace, prompt, output, permissions, and tool-free agent policy. Concurrent Grok processes share only a runner-owned persistent `GROK_HOME` generation under `$XDG_DATA_HOME/ai-consultants/grok-shared-oauth` (or the standard XDG data fallback). Grok Build's own auth lock coordinates refreshes inside that shared generation; a short ai-consultants lock protects seeding, adoption, atomic runner-policy reconciliation, and atomic publication back to the ambient `~/.grok/auth.json`. An external `grok login` always wins the digest-based CAS: the affected consultation fails temporarily and never overwrites the new login or falls back to a separately billed API request. Malformed credentials and unpersisted refreshes fail closed. Use `GROK_OAUTH_MODE=serialized` only to hold the adapter lock across the complete CLI run for diagnostic comparison. Forced API mode is stateless and never creates or reads the shared OAuth generation. Each new generation also performs exactly one model-inventory bootstrap under the short lock because Grok 1.0.4 lazily initializes home metadata; all later inventory probes and every inference remain concurrent. Grok 1.0.4 retains ownership of its generated `config.toml`; ai-consultants does not replace provider-required sections. Instead it atomically reconciles a separate `.ai-consultants-policy.json`, while the actual enforced surface stays pinned by the per-invocation CLI arguments. `max_quality` sets `GROK_REASONING_EFFORT=xhigh` (the highest value accepted by Grok Build), `GLM_REASONING_EFFORT=max`, and `DEEPSEEK_REASONING_EFFORT=max`. Grok passes the value to the CLI as `--reasoning-effort`; the API transports include the standard `reasoning_effort` field. Unsupported values or CLI capability gaps fail explicitly rather than silently reducing effort. A provider-rejected CLI value fails after launch and never triggers API fallback. Those three effort values are tier-managed: `max_quality` temporarily overrides an ambient lower pin and restores it when another tier is applied. Qwen is the intentional exception because Token Plan access is separately configured: an explicit `QWEN3_REASONING_EFFORT` pin is preserved in the maximum tier. The preset gives Mistral and Grok four bounded advisory turns, applies `QWEN3_MAX_QUALITY_TIMEOUT` and `DEEPSEEK_MAX_QUALITY_TIMEOUT` (default 0, no limit) to Qwen3.8-Max and DeepSeek, sets Meta to `max` effort, and uses the configured per-consultant `*_API_MAX_TOKENS` budget. Maximum temporarily raises Codex, Mistral, Grok, and MiniMax API budgets to `MAX_QUALITY_API_MAX_TOKENS` and restores them when leaving the tier; Qwen, GLM, and DeepSeek already default to their proven 16,384-token budgets. OpenAI-compatible responses ending with `finish_reason=length` fail closed rather than entering synthesis as partial answers. Claude CLI mode is preflighted and dispatched without session persistence, ambient setting sources, tools, or MCP servers. `KIMI_MODEL` is passed directly to `kimi --model`, so `kimi-code/k3.1-pro` overrides any older default stored in the user's Kimi CLI configuration. The Kimi CLI has no effort flag: `KIMI_REASONING_EFFORT` (default `max`) is validated like the other efforts and exported to the CLI as `KIMI_MODEL_THINKING_EFFORT`, its documented override of `[thinking] effort` in `~/.kimi-code/config.toml`. Set it to an empty value to leave the CLI's own effort in place. Kimi CLI compatibility is also capability-probed: prompt mode, `stream-json`, provider inventory, and the requested model must be present. No numeric CLI version is used as a gate. `MISTRAL_MODEL` is used only by the HTTP API. Vibe receives `MISTRAL_CLI_MODEL` through `VIBE_ACTIVE_MODEL` and runs its read-only `plan` agent in a temporary workspace. The current Vibe default is `mistral-medium-3.5`. The API default for every tier is `mistral-large-4` (public preview since 2026-10-06). An authorized consultation smoke returned a structured answer with the provider-reported model `mistral-large-4` on 2026-10-15. The preview is absent from `GET /v1/models` even for keys that can call it, so that listing is not an access check. Mistral Large 3 is retired. The official API IDs `mistral-medium-3-5` and `mistral-small-2603` remain catalogued opt-ins until an authenticated API smoke passes in this project. Gemini 4 Argon (GA 2026-10-13) is the Gemini default and maximum target: `GEMINI_MODEL="gemini-4-argon-high"` on `agy` and `GEMINI_API_MODEL=gemini-4-argon` on the API. An authorized `agy` consultation smoke returned a structured answer on 2026-10-16 (capability-probed identity, about 49 s for a short question). The API transport has offline tests only. On the API, Argon defaults to the `high` thinking level of its CLI target (`GEMINI_REASONING_EFFORT` accepts low|medium|high), sends no sampling parameters, and spends thinking tokens from `GEMINI_API_MAX_TOKENS` (default 32768). Google excludes Argon from the Free tier and the AI Pro plan, so an `agy models` listing is not an access check. Pin `GEMINI_MODEL="gemini-3.8-flash-high" # ai-consultants:pin` or `GEMINI_API_MODEL=gemini-3.8-flash # ai-consultants:pin` to use Gemini 3.8 Flash, the standard and economy target on both transports, by default. The current Vibe CLI exposes only `--prompt TEXT`, not stdin or a prompt-file surface, and the same holds for `agy` and `kimi`. Their adapters keep an isolated workspace, but the prompt remains visible in the CLI process argv, and a prompt above the single-argument limit (`ARGV_PROMPT_MAX_BYTES`, computed by default) fails before dispatch. Avoid secrets in consultation context until those CLIs ship a non-argv input mode. Agentic CLIs (`agy`, `qwen`, `kimi`) run from a private empty directory unless `CLI_WORKSPACE_ISOLATION=false`. `metadata.requested_model` records what ai-consultants asked for, while `metadata.model_identity_source` is `provider-reported`, `capability-probed`, or `requested-only`. The top-level `model` remains the strongest effective identifier the transport can honestly attest. ## Model Quality Tiers (v2.5) ```bash # Set all consultants to a tier programmatically source scripts/config.sh apply_model_tier "premium" # Latest flagship models apply_model_tier "maximum" # max_quality-only / separate-plan models apply_model_tier "standard" # Good quality at reasonable cost apply_model_tier "economy" # Optimized for speed and low cost ``` ## Budget Management (v2.4) ```bash ENABLE_BUDGET_LIMIT=false MAX_SESSION_COST=1.00 # Maximum cost in USD BUDGET_ACTION=warn # warn or stop ``` ## Caching and Optimization (v2.3) ```bash ENABLE_SEMANTIC_CACHE=true # Cache responses by query, context and consultant model/transport/persona CACHE_TTL_HOURS=24 # Cache expiration ENABLE_RESPONSE_LIMITS=false # Deprecated no-op (never applied); use *_API_MAX_TOKENS ENABLE_COST_AWARE_ROUTING=false # Route simple queries to cheaper models ENABLE_COMPACT_REPORT=true # Summaries only in reports ``` ## Health Gate, Quorum, and Retries ```bash ENABLE_HEALTH_GATE=false # Ping each selected consultant before Round 1 HEALTH_GATE_TIMEOUT=30 # Maximum seconds for each parallel ping QUORUM_MIN=2 # Fewer successful responses => failed quorum QUORUM_ACTION=warn # warn | stop MAX_RETRIES=2 RETRY_DELAY_SECONDS=5 ``` The health gate adds one small call per selected consultant. It drops dead or unauthenticated consultants before the full query; `QUORUM_ACTION=stop` aborts when the remaining panel is smaller than `QUORUM_MIN`. ## Timeouts, Waiting and Retries Consultants and synthesis have **no time limit by default** (every `*_TIMEOUT` is 0). Reasoning models can legitimately take minutes, and a call cut off by a client timeout is usually still billed while its answer is lost. The invoking host decides when to stop waiting instead. While `consult_all.sh` waits it logs `Output: ` at the start and `Still waiting after Ns for: ` every `WAIT_HEARTBEAT_SECONDS` (0 = silent), and it accepts: | Request | How | Effect | |---|---|---| | Proceed | `touch /.proceed`, `kill -USR1 `, or a first Ctrl-C | Stop the current phase and continue with the answers received. During Round 1 the consultants still running are stopped and reported as `stopped`; during synthesis the synthesis is skipped. | | Abort | `touch /.abort`, SIGTERM/SIGHUP, or a second Ctrl-C | End the run (exit 130, 143 or 129) and keep the partial output directory. | Stopped work ends with its whole process tree, so no consultant CLI keeps running. Its partial output is discarded and excluded from synthesis and cost; the provider may still bill what it generated. Do not stop a run with SIGKILL: the shell cannot clean up, so consultant CLIs would keep running. Probes keep short limits because they detect a dead service rather than cut paid generation: model inventory, auth status, `HEALTH_GATE_TIMEOUT`, `DOCTOR_LIVE_TIMEOUT` and the LLM classifier's `CLASSIFICATION_TIMEOUT` (which falls back to pattern classification). agy can print its whole answer and then never exit (seen with agy 1.4.2, as a synthesizer). When agy's output is one complete JSON object that has not changed for `AGY_EXIT_GRACE_SECONDS` (default 60), the run ends agy and keeps the answer; prose or a partial answer still waits for agy to exit. 0 disables this and always waits for the exit. ```bash MAX_RETRIES=2 RETRY_DELAY_SECONDS=5 WAIT_HEARTBEAT_SECONDS=30 AGY_EXIT_GRACE_SECONDS=60 # keep a complete agy answer when agy never exits GEMINI_TIMEOUT=0 # 0 = no limit; a positive value restores a hard limit CODEX_TIMEOUT=0 MISTRAL_TIMEOUT=0 MISTRAL_MAX_TURNS=4 KIMI_TIMEOUT=0 CLAUDE_TIMEOUT=0 QWEN3_TIMEOUT=0 QWEN3_MAX_QUALITY_TIMEOUT=0 MINIMAX_TIMEOUT=0 META_TIMEOUT=0 GLM_TIMEOUT=0 GROK_TIMEOUT=0 GROK_MAX_TURNS=4 DEEPSEEK_TIMEOUT=0 DEEPSEEK_MAX_QUALITY_TIMEOUT=0 ``` `configure` migrates the fixed limits it generated through 5.1.2 (180 s per consultant, 240 s for Claude and synthesis, 480 s synthesis total, 600 s for maximum-tier Qwen/DeepSeek) to 0. A value marked `# ai-consultants:pin`, or any other custom value, stays a hard limit.