Skip to content

default-config.md

Latest commit

 

History

History
33 lines (28 loc) · 4.82 KB

default-config.md

File metadata and controls

33 lines (28 loc) · 4.82 KB
Raw

Default config

These defaults are what the project owner personally runs. Change any of them in .local/host.yml or with omp config set.

  • Models: omp runs the agents. Opus 5.5 at high thinking runs the home, with the advisor off. OpenAI requests use the priority tier. See Firstmate dispatch defaults for per-task model choices and thinking levels.
  • Fallbacks: a model that keeps failing falls back to the next one. Opus 5.5 falls back to Fable 5.1 at medium thinking, Fable 5.1 to Opus 5.5, and Sonnet 5.5 to Fable 5.1, then Opus 5.5.
  • Plugins and extensions: the omp plugins ponytail, i-have-adhd, and caveman. Extensions add the turn-end quality check, status-line icons, the Herdr sidebar feed, and a block on kill-by-name commands such as pkill and killall.
  • Gates: the no-mistakes pipeline and ponytail-review check every change. Gates use Sonnet 5.5 at high thinking, and Opus 5.5 for hard work. Pick another model for one run with --model.
  • Skills and rules: blueprint/skills/ and blueprint/config/AGENTS.md serve omp, Claude Code, and Codex. Public files install only with profiles.public_skills (opt-in; a host that already has them keeps them); the upstream skills from GitHub install either way, and private files take precedence (omp). The rules set plain technical writing, always-on skills, and native stacked PRs.
  • Profiles on by default: agents, firstmate
  • Opt-in profiles: development, sgrep, docker, tailscale, desktop, fleet_guards, project_databases, jj, fdu, vphone, fleet_browsers, chat, gateway, forgejo
  • New hosts: ./ship.sh dock writes the core only: agents and firstmate. Pick more with the checklist bunx crewship shows, or --with NAMES (Configuration). A host config made before this keeps its own values.
  • Firstmate backend: see Firstmate backend for the default, alternatives, and existing-checkout behavior.
  • Opt-in settings: data_dir, usage, token_suggestions, firstmate.checklist, mac_ssh, imessage, github_board, board, board.github_sink, ci_pool, runner_up, cloudflare_runner, signal
  • Full lists: blueprint/config/default.yml, blueprint/config/omp.yml, and the wiki.

Optimizations in the default

Each line names the real setting. Details are in the linked pages.

  • Compaction: every session compacts at 200,000 tokens (compaction.thresholdTokens), keeps 32,000 tokens of recent messages (compaction.v2RetainedMessageBudget), and tries snapcompact, remote, handoff, shake, soft in that order (compaction.methodOrder).
  • Tool output spill: output above 20 KB goes to an artifact, and 4 KB head and 4 KB tail stay inline (tools.artifactSpillThreshold, tools.artifactHeadBytes, tools.artifactTailBytes).
  • Cheap helper models: commit, tiny, and memory use Haiku 5.5, and smol uses Sonnet 5.5 with thinking off (modelRoles). omitThinking and hideThinkingBlock are on, and textVerbosity is low.
  • Memory: memory.backend: mnemopi with mnemopi.noEmbeddings: true, so recall needs no embedding calls.
  • Crew advisor: off globally (advisor.enabled: false). Crewmates turn it on at low thinking with syncBacklog: "off" and immuneTurns: 10, so it never blocks a crew (omp).
  • Model fallback: retry.fallbackChains with retry.maxDelayMs: 60000, so a failing model does not stall a run.
  • Provider tier: tier.openai: priority. Models connect to providers directly, with no router.
  • Spawn memory floor: spawn-memory-floor-mb is 8000, so a crewmate does not start on a host low on RAM (Capacity).
  • Idle browsers: browser_prune stops idle AXI browsers after 7200 s (checked every 300 s), and browsers.idle_timeout_min is 30.
  • Docker: docker.init: true, docker.live_restore: true, and fleet.docker_guard.stopped_hours: 24 (Fleet guards).
  • Disk: the storage guard warns at 85% and prunes at 92% (Capacity). Package caches go to <data_dir>/cache/<tool> when data_dir is set (Configuration).
  • CI pool: job_memory_gb: 8 caps each job, and nice can lower CI CPU priority; its default is 0 (CI pool).
  • Terminal UI: tui.tight, tui.vimMode, a borderless composer with composer.tokenRate, and a transparent powerline status line (omp).