Everything lives in one file, config.yaml, seeded on first run from the tracked template config.default.yaml. Most of it is adjustable live with /config <dotted.key> <value>; edits persist by default, --session keeps one for this session only. Trust keys are the exception — see below.

::tiers and roles

The agent references model roles, never concrete model ids. A hardware tier maps each role to a model, so swapping hardware is a one-line active_tier change:

config.yaml

active_tier: laptop

tiers:
  laptop:
    provider: ollama
    roles:
      planner: gemma4:e4b
      tool_caller: gemma4:e4b
      synthesizer: gemma4:e4b
      utility: gemma4:e4b
      judge: gemma4:e4b
    embedder: qwen3-embedding:8b
  workstation:
    provider: ollama
    roles:
      planner: qwen3.6:27b
      tool_caller: qwen3.6:27b
      synthesizer: qwen3.6:27b
      utility: qwen3.6:27b
      judge: qwen3.6:27b
    embedder: qwen3-embedding:8b

planner

drafts and redrafts the plan (structured output; keeps its hidden rationale)

tool_caller

the execute node: per-step reasoning and single-tool constrained calls

synthesizer

the final streamed answer

judge

the engine's judgment calls: rectify verdicts, the semantic write gate

utility

cheap background tasks

embedder

the RAG embedding model (separate from the roles)

The engine requires native tool-calling and structured output; capabilities: declares each model's context window and features, and /config setup warns when a bound model lacks one. /models lists installed Ollama models and rebinds a role, all roles, the embedder, or a whole tier (--session for this session only). Cloud providers are shelved: a non-Ollama binding refuses to run.

::runtime

key	default	meaning
`max_iterations`	16	execute-pass cap per turn
`auto_approve`	read_only	tools at or below this tier run unprompted (trust key)
`num_ctx`	null (auto)	Ollama context window; null uses each model's declared window
`llm_timeout`	120	read timeout per model call, guards a wedged daemon
`auto_compact` / `compact_threshold`	true / 0.85	fold older turns into a summary once the window fills
`citations`	true	inline `[n]` citations and the Sources footer
`airgap`	false	block and record every network exit (trust key)
`redaction`	off	`off` / `warn` / `redact` at the network boundary (trust key)
`quarantine`	gate	`off` / `warn` / `gate` for untrusted tool output (trust key)
`receipt`	true	the trust segment on the per-answer receipt
`confidence`	true	capture logprobs and mark low-confidence runs
`confidence_threshold`	auto	per-model calibrated threshold, or a pinned number
`confidence_exit_threshold`	derived	hysteresis exit for an open low-confidence run
`grant_scope`	task	how long a gate `a` grant lives: `task` / `session` / `persist` (trust key)

::other sections

web.max_results

results per web_search (default 5); the backend is fixed and keyless

rag.chunk_size / chunk_overlap / k

1000 / 150 / 6 — chunking changes force a re-embed

mcp.servers / connect_timeout / call_timeout

see mcp servers

shell.timeout / shell.env_scrub

60 s per command; secret-shaped env fragments removed from children (trust key)

paths.*

database, documents, workspace, cache, memory, sessions, snapshots, permissions, exports

::what persists

–Ordinary keys set with /config persist to config.yaml by default; --session keeps the change for this session.
–Trust keys — auto_approve, airgap, quarantine, redaction, grant_scope, shell.env_scrub — and the /policy and /privacy toggles are session-only unless you pass an explicit --save. A loosened posture is never written to disk silently.
–/policy risk overrides and the shell prefix allowlist persist in database/permissions.json. /policy allow <prefix> (the explicit command) always persists; a gate a grant follows grant_scope.
–The live config.yaml is user data and not tracked by git, so persisted settings never dirty the repo or break /update.

note :: careful

Migration for older clone installs: pulling the change that untracked config.yaml removes an unmodified copy (recreated from the template on next launch). If you had edited it, git refuses the pull once — back the file up, git checkout -- config.yaml, pull again, and re-apply your settings.

configuration & models

::tiers and roles

::runtime

::other sections

::what persists