Configuration

Project settings are stored in .wolf/config.json under openwolf. A package update adds new defaults while preserving existing values. Use the installed template as the complete reference.

Read guidance

Setting	Default	Effect
reads.duplicate_mode	warn	Advise on eligible repeated full reads. deny denies once with a retry path. off disables the check.
reads.skeleton_hints	true	Offer symbol outlines for eligible large indexed files.

Ranged reads and changed files are handled separately from unchanged full reads. Results depend on the agent delivering the relevant tool event.

Command output

Setting	Default	Effect
bash.filter_mode	suggest	Suggest output limits before selected commands.
bash.governor.mode	replace	Request supported output replacement. Alternatives are suggest and off.
bash.governor.threshold_tokens	2000	Estimated output size above which controls may apply.

Default command-family settings:

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

Place these values under openwolf.bash.governor.families. Replacement depends on supported Claude hook output. A cache pointer accompanies a preserved original. Standard error is not rewritten.

Context and project notes

Setting	Default	Effect
context.session_digest_budget_tokens	1500	Estimated size limit for the startup digest.
context.budgets	Per agent	Override the digest budget for a named agent.
context.reinjection_interval	25	Interval for selected approved rules on supported tool-batch events. 0 disables it.
context.state_budgets	Cerebrum: 2000; STATUS: 1000	Estimated budgets used for size warnings.
memory.consolidation_after_days	7	Age used by configured memory maintenance tasks.
cerebrum.max_tokens	2000	Target size for candidate convention notes.

Archival retains the latest block, pinned notes and tracked active sessions. It keeps restore pointers rather than deleting their history. The explicit memory archive --days command sets the age for that run. Protected approval is required for automatic durable instruction injection.

Project scanning

Setting	Default	Effect
anatomy.auto_scan_on_init	true	Scan during initial setup.
anatomy.rescan_interval_hours	6	Scheduled rescan interval.
anatomy.max_description_length	100	Maximum description length in characters.
anatomy.max_files	500	Scan limit. A limited scan must not be reported as complete.
anatomy.exclude_patterns	Template list	Exclude specified paths and patterns.
anatomy.respect_gitignore	true	Apply repository ignore rules.

Built-in exclusions also cover common generated files, caches and agent configuration paths. Review the resulting index before sharing it.

Usage estimates

token_audit.chars_per_token_code defaults to 3.5; chars_per_token_prose defaults to 4.0. These ratios estimate local content size. They do not change recorded provider counters. The dashboard keeps estimates and recorded usage separate.

Activity and updates

{
  "openwolf": {
    "visibility": { "mode": "quiet" },
    "updates": { "mode": "compatible" }
  }
}

This is a partial example. Merge it into the existing configuration.

Visibility modes are quiet, verbose and off. OPENWOLF_VISIBILITY=off disables new receipts and notices for that process. Functional protection and update policy are separate settings.

Update modes are compatible, all, notify and off. The default prepares compatible stable runtimes for new sessions and reports major upgrades without installing them. OPENWOLF_NO_UPDATE=1 disables checks and new runtime selection for that process. Existing session pins remain in place. See automatic updates.

Daemon and dashboard

The dashboard is enabled by default and binds to 127.0.0.1. Port settings are dashboard.port and daemon.port; setup and launch can select available project ports. Keep the project token private.

cron.enabled controls scheduled tasks. Retry and heartbeat settings are in cron. Restart the daemon after changes that affect its runtime configuration. Data polling does not apply every configuration change to a running process.

Disabled projects and selective memory

Set openwolf.enabled to false for a manual-memory project. Cached hook calls exit silently before reading stdin, selecting runtimes, or writing bookkeeping. OpenCode dispatch also honors the setting. openwolf init and openwolf update skip disabled projects, preserving their configuration and instructions; explicitly re-enable the project before reinstalling. Stop any already-running daemon when changing this setting; a newly started daemon exits without starting background work. Static module imports still run before the hook guard, and already-injected conversation context cannot be withdrawn.

Anatomy lookup is optional. Use direct search when the index is stale or unhelpful. Keep current handoffs compact and retrieve detailed evidence on demand. Hook observations are historical records, not proof that tests still pass after later edits.

Maintenance after suspend

The daemon persists the next due time for each schedule and checks overdue maintenance every 30 seconds and on startup. scan_project, consolidate_memory and generate_token_report catch up once; missed occurrences are not replayed individually. Set catch_up: false on a task in cron-manifest.json to disable catch-up. Disabled tasks remain disabled. A changed schedule starts a new cursor.

Catch-up coverage begins when the new scheduler first records its cursor; it does not infer missed runs from old, incomplete execution logs. Claims are persisted before execution: a process crash between claim and execution can skip that attempt, rather than executing a potentially completed action twice. Existing retry policy handles reported action failures. This is bounded maintenance recovery, not an exactly-once distributed scheduler.

Init/update move retired ai_task definitions to retired_tasks, with enabled: false, preserving their parameters. They are never scheduled or restored as model calls. Restart the daemon after changing its manifest or watcher exclusions.

Optional task review suggestions

Set openwolf.task_review.enabled to true to enable advisory context for Claude and Codex hooks (default: false). Two consecutive reported command failures or five minutes between active task events can prompt a brief review of the current evidence and approach. The agent can continue unchanged: these suggestions never deny a tool, terminate a task, require approval, or launch an extra model turn.

failure_threshold defaults to 2, interval_minutes to 5, and cooldown_minutes to 5. Only commands with a stable tool ID and an explicit numeric exit status count; duplicate deliveries and unknown outcomes do not. A successful command clears the failure count. Elapsed time is observed at hook events, excluding waiting between a Stop event and the next user prompt; it is not a background deadline. Pending suggestions are delivered through the next supported Bash or user-prompt context response. An explicit handoff objective change resets the task; otherwise the scope is the current session. OpenCode does not yet emit these suggestions.