**English** | [Русский](/ru/docs/configuration)

# TAUSIK Configuration Reference

<!-- doc-map: reader=user; zone=configuration -->

All knobs live in `.tausik/config.json` at project root. Anything not set falls back to the documented default. To override, add the key under the top-level object (NOT under `bootstrap` — that section is bootstrap-managed).

See also: [environment.md](https://github.com/Kibertum/tausik-core/blob/v1.11.0/docs/en/environment.md) — env vars, [permissions.md](https://github.com/Kibertum/tausik-core/blob/v1.11.0/docs/en/permissions.md) — permission modes.

## Session signals (SENAR Rule 9.2 — advice, not gates since 1.10)

| Key | Default | Purpose |
|---|---|---|
| `session_max_minutes` | `180` | Advisory threshold on session ACTIVE minutes: above it `task start`, `status` and the Stop hook print advice; nothing refuses. `0` switches it off. |
| `session_idle_threshold_minutes` | `10` | Gap (in minutes) above which a pause between `events` rows is treated as AFK (excluded from active-time sum). |
| `session_warn_threshold_minutes` | `150` | Stop-hook reminder threshold in `session_cleanup_check.py`. Should be < `session_max_minutes`. |
| `session_capacity_calls` | `200` | Per-session tool-call budget. A task `call_budget` above what remains is an advisory line in `task start`, not a refusal. |
| `answer_budget_words` | `200` | Word budget for the agent's final answer. On the next prompt, an answer over it (or without a verdict in its first line) gets one advisory line with the numbers; nothing is blocked. Measure with `tausik metrics answers`. |
| `checkpoint_calls` | `40` | Calls since the last handoff before checkpoint advice (derived from the ledger). `0` switches it off. |
| `journal_freshness_calls` | `40` | Calls since an active task's last log entry before journal advice. `0` switches it off. |
| `audit_every_closures` | `17` | Task closures since the last audit mark before the SENAR 9.5 audit is overdue. |

## Host context budget

`host_context_budget` guards the real host thread. Each threshold has three positive integer fields: `responses`, cumulative `input_tokens`, and the latest response's `context_tokens`. Crossing any advisory field writes one checkpoint and returns a copy-ready new-window prompt. Crossing any hard field sets `allow_continue=false`; an old thread without a bound TAUSIK session is not opened. Defaults are advisory `{24, 2500000, 140000}` and hard `{32, 4000000, 180000}` in that field order. Every advisory value must be below its hard value.

Codex supplies native thread identity and usage from its local journal. A host without those signals returns `level=unavailable` and a reason, then fails open to the existing time/call advice. Ending and reopening a TAUSIK session in the same host thread never makes `fresh_context` true.

## Code search (RAG) languages

| Key | Default | Meaning |
|-----|---------|---------|
| `rag.extra_extensions` | `{}` | Extra file types to index, `{".unity": "unity-scene"}`. A built-in extension cannot be overridden. Godot (`.gd`, `.gdshader`, `.tscn`, `.tres`, `.godot`) is built in since 1.10. |
| `rag.boundaries` | `{}` | Where to cut a language into chunks, `{"unity-scene": "^--- !u!"}` (a multiline regex). A bad entry is skipped and named in `rag_status` under `language_config.problems`. |

## Update check

| Key | Default | Meaning |
|-----|---------|---------|
| `updates.check` | `true` | At most once a day, SessionStart runs `tausik update-check` detached: one anonymous GET to GitHub's `releases/latest` of `Kibertum/tausik-core`, nothing about the project in it. `status` then names a newer release. `false` switches it off; `tausik doctor` shows the state. |

## Verification cache (SENAR Rule 5)

| Key | Default | Purpose |
|---|---|---|
| `verify_cache_ttl_seconds` | `600` | How long a green verify run is reused before re-running gates. Lower for security-critical projects. |

## Settings met in the quick-start

All keys sit at the root of `.tausik/config.json` unless a path says otherwise.

| Key | Default | Purpose |
|---|---|---|
| `context_tier` | `"minimal"` | Size of the generated rules file: `"minimal"` (default, bounded), `"standard"`, `"full"` (extended pointers). Bootstrap refuses an unknown value; `tausik doctor` checks drift against the saved tier. |
| `output_mode` | `"off"` | `"caveman"` adds terse prose. Every mode ships the same controlled-prose rules: name actor/action, prefer active voice where natural, keep one action per sentence and one term per concept. This is multilingual guidance, not ASD-STE100 compliance. Code, tool output and SENAR evidence stay whole. A bad value falls back to `"off"`. |
| `model_profile` | unset | Host profile slug (`a-z`, digits, hyphens), e.g. `claude`, `codex`. Bootstrap writes it when `TAUSIK_MODEL_PROFILE` is set; an invalid value aborts bootstrap. `python bootstrap/bootstrap.py --refresh` refreshes config only. |
| `mcp.compact_tool_list` | `true` after bootstrap | Advertise complete definitions for the core workflow and short definitions for other tools; retrieve a full deferred schema with `tausik_tool_schema`. Explicit `false` is preserved. Missing or malformed config fails open to the full list. |
| `task_done.auto_verify` | `false` | `true` makes `task done` run the gates inline and close **without a signed receipt**. It is not a bypass of the gates — they still run — but it gives up the receipt, so the config trust check treats it as a weakening and a trusted tier sets it back (`scripts/config_trust.py`). Prefer `task done <slug> --ac-verified --verify`: one call, with the receipt. |

For explanation-shaped prompts, the prompt hook adds a conditional format matrix. Concise prose remains the fallback; tables serve repeated exact comparisons, Mermaid serves connected structure, and HTML requires an explicit or named interactive need. The router never selects video. Ordinary prompts do not pay for this extra instruction.

**Rules files are preserved once written.** `CLAUDE.md`, `AGENTS.md`, `.cursorrules`, `QWEN.md` and the OpenCode rules file are not rewritten by bootstrap when they exist, so changing `context_tier` or `output_mode` later does not reach them. Bootstrap says so; delete the generated file and run bootstrap again, or edit it by hand.

**The MCP close returns structure.** `tausik_task_done` answers with `stage`, `gate_results` and `blocking_failures`, so an agent can fix what failed without parsing prose.

## Stacks

| Key | Default | Purpose |
|---|---|---|
| `custom_stacks` | `[]` | List of custom stack slugs accepted by `task add --stack X`. |

## Gates

| Key | Default | Purpose |
|---|---|---|
| `gates` | `{}` | Per-gate overrides: `{ "pytest": { "enabled": true }, "filesize": { "max_lines": 600 } }`. Merges over `default_gates.py`. |

## Agent-instruction files (the DYNAMIC block)

`tausik update-claudemd` rewrites the block between `<!-- DYNAMIC:START -->`
and `<!-- DYNAMIC:END -->` in CLAUDE.md and, when one sits beside it, in
AGENTS.md. AGENTS.md is tracked in most projects, so what goes into it is a
policy, not an accident (GitLab #14):

| Key | Default | Purpose |
|---|---|---|
| `claudemd.sibling_dynamic` | `true` | `true`: AGENTS.md is refreshed with the block **minus** "Shared knowledge — from other projects" — other projects' knowledge never enters this repository's history, and a memory tail left with nothing under it is dropped. `false`: AGENTS.md is not written at all; CLAUDE.md is refreshed as before. Only the JSON boolean `false` turns it off — the string `"false"` or `0` is read as on. The key is read from the `.tausik/` beside the file being written. |

The `claudemd_state` commit gate judges each file against the same plan the
writer follows: a sibling the policy does not write is not judged, and a
sibling whose only knowledge would be foreign owes no tail.

## Publication (what a redacted export hides)

The shared store (`~/.tausik-knowledge`) needs no configuration: `--global` on
`decide` / `memory add` writes to it, `$TAUSIK_HOME` moves it. These keys feed
`knowledge export --redacted` only (see [knowledge-store.md](https://github.com/Kibertum/tausik-core/blob/v1.11.0/docs/en/knowledge-store.md)).

| Key | Default | Purpose |
|---|---|---|
| `publication.project_names` | `[]` | Project names to replace with `[REDACTED:project]`, in addition to this project's directory name. |
| `publication.private_url_patterns` | `[]` | Regex strings; a URL matching one becomes `[REDACTED:url]`. |
| `brain.local_mirror_path` | `~/.tausik-brain/brain.db` | Only read by the one-off `knowledge import-brain`: where the retired Notion transport left its local mirror. |

## Codex subscription-credit rates

`codex_subscription_credit_rates` is optional and has no framework default.
Set a dated card only when `metrics tokens --host codex` should add a
credit-weighted accepted-task view to its raw counters:

```json
{
  "codex_subscription_credit_rates": {
    "source": "https://developers.openai.com/docs/pricing",
    "as_of": "2026-10-02",
    "unit": "credits_per_million_tokens",
    "models": {
      "gpt-5.6-sol": {"input": 100, "cached_input": 10, "output": 500}
    }
  }
}
```

Use the values applicable to your account and date. These are subscription
credits, not API USD rates or a conversion of remaining included allowance.

## Example

```json
{
  "session_max_minutes": 240,
  "session_idle_threshold_minutes": 15,
  "host_context_budget": {
    "advisory": {"responses": 20, "input_tokens": 2000000, "context_tokens": 120000},
    "hard": {"responses": 28, "input_tokens": 3500000, "context_tokens": 170000}
  },
  "verify_cache_ttl_seconds": 1200,
  "custom_stacks": ["ruby", "elixir"],
  "gates": {
    "filesize": { "max_lines": 500 },
    "ruff": { "enabled": false }
  },
  "publication": {
    "project_names": ["acme"],
    "private_url_patterns": ["acme\\.internal"]
  },
  "claudemd": { "sibling_dynamic": false }
}
```

## Health check

`tausik doctor` (v1.3+) verifies that resolved config + venv + DB + skills are coherent and surfaces actionable next steps.