Configuration

OpenWolf is configured through .wolf/config.json. Every setting has a sensible default; nothing needs changing for normal use. openwolf update merges newly introduced keys into existing configs without touching your values.

json

{
  "version": 1,
  "openwolf": {
    "enabled": true,
    "reads": { ... },
    "bash": { ... },
    "context": { ... },
    "anatomy": { ... },
    "token_audit": { ... },
    "cron": { ... },
    "memory": { ... },
    "cerebrum": { ... },
    "daemon": { ... },
    "dashboard": { ... }
  }
}

`reads`

Duplicate-read handling and read hints.

Key	Default	Description
`duplicate_mode`	`"warn"`	`"warn"` advises on duplicate full reads; `"deny"` blocks each duplicate once (with a pass-through retry so the model is never stranded); `"off"` disables
`skeleton_hints`	`true`	For large indexed files, serve a signature outline in the pre-read hint instead of the largest-sections line

`bash`

The Bash channel: the pre-run suggestion filter and the output governor.

Key	Default	Description
`filter_mode`	`"suggest"`	Pre-run note for flood-prone commands: `"suggest"` or `"off"`
`governor.mode`	`"replace"`	Master switch: `"replace"`, `"suggest"`, or `"off"`
`governor.threshold_tokens`	`2000`	Only results above this size are governed
`governor.families`	(below)	Per-family action override

Default family actions:

json

"families": {
  "grep_flood": "replace",
  "file_print": "replace",
  "git_show": "replace",
  "test": "suggest",
  "build": "suggest",
  "unknown": "suggest"
}

Only the three losslessly recoverable families replace by default. Test and build output is never replaced unless you opt in, and stderr is never modified by any family. Full outputs are always preserved at .wolf/cache/bash/.

`context`

Session digest, rule re-injection, and state budgets.

Key	Default	Description
`session_digest_budget_tokens`	`1500`	Upper bound for the session-start injection (the index targets ~400)
`budgets`	per agent	Per-agent digest budgets (`claude` 1500, `codex` 1200, `cursor` 800, ...)
`reinjection_interval`	`25`	Re-surface the top Do-Not-Repeat rules every N tool batches; `0` disables
`state_budgets`	`{".wolf/cerebrum.md": 2000, ".wolf/STATUS.md": 1000}`	Token budgets enforced with one warning per session on over-budget writes

`anatomy`

The project scanner.

Key	Default	Description
`auto_scan_on_init`	`true`	Run a full scan during `openwolf init`
`rescan_interval_hours`	`6`	Daemon rescan cadence. Rescans are skipped when the index is provably fresh (stat sweep + git HEAD)
`max_description_length`	`100`	Max characters per file description
`max_files`	`500`	Stop scanning after this many files
`exclude_patterns`	node_modules, .git, dist, ...	Directories and globs to skip

Lockfiles, OS junk, caches, coverage, minified files, and agent-config directories (.claude, .codex, and friends) are excluded built-in, regardless of this list.

`token_audit`

Key	Default	Description
`enabled`	`true`	Token tracking
`chars_per_token_code`	`3.5`	Estimation ratio for code
`chars_per_token_prose`	`4.0`	Estimation ratio for prose
`waste_threshold_percent`	`15`	Waste-alert threshold

Estimates are the softest tier; measured and verified numbers come from transcripts and are unaffected by these ratios.

`cron`

Key	Default	Description
`enabled`	`true`	Scheduled tasks
`max_retry_attempts`	`3`	Retries before dead-lettering

`memory` and `cerebrum`

Key	Default	Description
`memory.consolidation_after_days`	`7`	Compress session blocks older than this
`cerebrum.max_tokens`	`2000`	Target size for cerebrum.md

`daemon` and `dashboard`

Key	Default	Description
`daemon.port`	per project	Daemon HTTP API port
`dashboard.port`	per project	Dashboard HTTP and WebSocket port
`dashboard.enabled`	`true`	Serve the dashboard

Ports are assigned per project at init so multiple dashboards never collide, and they survive updates.

Configuration ​

reads ​

bash ​

context ​

anatomy ​

token_audit ​

cron ​

memory and cerebrum ​

daemon and dashboard ​