tune-my-harness

name	tune-my-harness
description	Install or upgrade a model-routing and delegation layer in the current project. Use when the user says "tune my harness", "add model routing", "set up delegation", or after a harness cleanup, to add pinned-model agent rosters, a routing table, and QA policy to an existing Claude Code or Codex project.

Tune My Harness

Where the cleanup kits prune, this skill installs. It audits how the current project routes work between model tiers and proposes the missing pieces: a delegation roster with pinned models, a routing table in the instructions file, Codex tier profiles, and a QA policy. Nothing is written without the user approving the numbered proposal.

This skill is self-contained: everything it installs ships in its own references/ folder, so it works on a machine with no prior roster anywhere.

• references/claude-agents/: the five Claude Code roles (fetcher, summarizer, implementer, qa-reviewer, architect) with pinned models
• references/codex-agents/: the same five roles as Codex TOML definitions (luna/terra/sol tiers)
• references/routing-table.md: paste-ready Model Routing, QA Policy, and Codex profile blocks

The method and rationale live in guides/model-routing-guide.md (in the AI Workbench; read it if available). If the user already has a roster at ~/.claude/agents/ or ~/.codex/agents/, prefer consistency with theirs over the bundled references.

Step 1: Inventory (read-only)

Establish what exists before proposing anything:

1. Instruction files: AGENTS.md, CLAUDE.md, .claude/CLAUDE.md. Note which is canonical, whether a routing table or QA policy already exists.
2. Rosters: .claude/agents/ and ~/.claude/agents/ (frontmatter model: per agent); .codex/agents/ and ~/.codex/agents/; [profiles.*] in ~/.codex/config.toml.
3. Commands and skills: any model: frontmatter, any skill descriptions long enough to strain Codex's discovery budget (keep descriptions to roughly two sentences about when to select).
4. Which runtimes this project actually uses (presence of .claude/, .codex/, .agents/, or ask once).

Step 2: Propose (numbered, nothing applied)

Present a numbered proposal covering only the gaps found:

1. Roster: which of the five roles to install, at which level (user level covers all projects and is the default; project level only when the project needs different tools or models). Never create a same-named agent at both levels.
2. Routing table: the task-class table plus escalation rule for the canonical instructions file, with tiers matched to the runtimes in use.
3. QA policy: change classes with review levels, cross-vendor off by default.
4. Codex profiles: [profiles.luna|terra|sol] if Codex is used and they are absent.
5. Skill description tightening: any descriptions to shorten for the discovery budget.

For each item: what changes, why, and how to reverse it. Recommend, do not pad; a project that already routes well should get a short "nothing to install" report.

Step 3: Apply approved items only

• Back up every file before touching it (state where the backups are).
• Copy roster definitions from this skill's references/claude-agents/ and references/codex-agents/ folders (or from the user's existing user-level roster if one exists), adjusting tools and stack references to this project. Keep each description selection-oriented: what the role is for and what it must not be used for.
• Insert the routing table and QA policy using the blocks in references/routing-table.md as the starting point, with the change classes edited to what this project actually ships.
• Insert the routing table and QA policy into the canonical instructions file (never into a duplicate).
• Do not change permission or approval settings while installing; note them for the user if they conflict with the roster (for example, a sandbox that blocks subagent spawning).

Step 4: Verify and receipt

1. Confirm new agent files parse (frontmatter present, model values valid: haiku, sonnet, opus, inherit for Claude; exact runtime-exposed IDs for Codex).
2. Confirm the instructions file still passes the project's checks if it has any (in the AI Workbench: python3 scripts/check-ai-harness-rules.py).
3. Leave a short WHAT-CHANGED note beside the instructions file: what was installed, what was skipped, where backups are.
4. Remind the user of the two standing rules that make routing safe: escalation for auth, money, and irreversible changes; and no self-certification by cheap tiers.