omp-config

Omp harness configuration for Phaedrus / Misty Step. Source of truth for how agents run on this machine: model roles, global policy, skills, themes. ./install deploys everything.

Layout

Path	Purpose
install	Deployment script. Validates sources, copies the allowlist into $(omp config path), deploys agents and skills atomically, merges live MCP auth, and self-installs the pre-push hook
bin/omp-grievances.ts	Manual grievance inbox CLI; reads OMP intake without mutation and stores private acknowledgements under XDG state
config.yml	Model roles and fallback chains, theme/statusline/TUI display, providers (web search routed through exa), task/LSP settings
models.yml	Local Ollama provider discovery; cloud models come from omp's bundled catalog
mcp.json	Declared MCP servers; install merges per-server auth/oauth from the live copy
global/AGENTS.md	Session-opening context for architecture, technology, review, operations, and communication
global/RULES.md	Short core rules reattached near every turn: deletion, reversibility, cutovers, proof, failures, and expert judgment
agents/	Read-only specialist agents; installed globally and routed by each agent's model role
global/WATCHDOG.md, WATCHDOG.yml	Continuous Steward advisor: a high-reasoning model reviews intent, design, correctness, proof, and operations; roster configured in YAML
themes/	TUI themes (tokyonight, everforest, everforest-light)
skills/	Skill packages, copied wholesale on install — see below
.githooks/pre-push	Hook source installed into this repo's git dir by ./install
.agents/skills/writing-for-agents/	Agent-writing guidance served via skill://writing-for-agents; not deployed by install
extensions/loc/	Session-resident LOC status and commands; deployed by install
CANON.md	Concise, implementation-independent operating philosophy. Reference only — never deployed or auto-loaded

Install

./install   # requires jq

Checks every allowlisted source exists and parses, deploys config files with mode 600, folds live MCP credentials into declared servers, swaps skills/ atomically, and installs the git hook. Run after every change; sessions pick up deployed state on their next start.

LOC extension

The globally deployed LOC extension provides /loc, /loc-trend, and a committed-HEAD status row. Session and turn hooks read .git/loc_cache only; a missing or stale cache hides the row rather than calculating during interactive work. Explicit LOC commands may calculate on demand.

Disable only LOC for a large repository with project-local configuration:

# .omp/config.yml
disabledExtensions:
  - extension-module:loc

Grievance inbox

omp-grievances treats OMP's grievance database as a read-only inbox. Its acknowledgement ledger defaults to $XDG_STATE_HOME/omp-config/grievances.sqlite3 or ~/.local/state/omp-config/grievances.sqlite3.

omp-grievances status
omp-grievances inbox --limit 20
omp-grievances show 294
omp-grievances ack 294 --outcome ticketed --ref POW-123
omp-grievances ack --through 250 --outcome historic --note "pre-ledger backlog"
omp-grievances unack 294

Outcomes are ticketed, no-action, and historic. ticketed requires an opaque external reference such as a Powder or Habitat item. The ledger stores grievance IDs, outcomes, references, and notes; raw reports remain owned by ~/.omp/autoqa.db. A salted source fingerprint prevents acknowledgements from silently attaching to a replaced or rewritten grievance history.

Skills

Two invocation classes:

Operator-invoked (disable-model-invocation: true) — human-directed flows that run only on explicit request: agent-ergonomics, audit-choices, audit-observability, audit-simplifications, backlog, brief, capture, code-review, core-docs, deliver, diagnose, executive, explore-design, extract-module, field-station, forest-executive, foundation, groom, install-anti-slop, now-next, polish, product-description, pulse, refactor, release, resilience, security-review, shape, tidy, torvalds-design-review, understand.

Model-invocable — unambiguous triggers, bounded cost, cheap wrong-fire: ast-grep, custom-linters, dispatch, evidence-packet, exocortex, frontend-design, herdr, research, show-me.

The exocortex skill is generated from misty-step/exocortex skills/exocortex/SKILL.md. Do not hand-edit skills/exocortex/SKILL.md. Regenerate with that repo's scripts/install-skill.sh and prove with skills/exocortex/check-source.sh.

Canon relationship

CANON.md states the implementation-independent operating philosophy. Configuration, standing policy, and skills are concrete mechanisms that should express it without turning the canon into an inventory of the current harness. Implementation details and routing live in this README and the configuration. Never paste the canon wholesale into agent context; distill only the judgment a surface needs.

Ecosystem

• Powder — work ledger for all Misty Step work; skill managed in misty-step/powder and scoped to ~/Development/misty-step/ via workspace discovery.
• Iron Forest — headless Builder/Verifier/Fixer factory. Mechanical enforcement belongs there and in CI, not in prose.
• Landmark — release pipeline: conventional commits become semantic versions, technical changelogs, synthesized user-facing notes, and machine-readable evidence.
• Estate — private infrastructure map; consult before any infrastructure claim.