Harness Router

Route useful session learnings into durable, reviewable agent knowledge surfaces without silently rewriting them.

Scope

Use this skill when the task is to:

• Decide where new agent/harness knowledge should live after a session.
• Turn user corrections or repeated agent mistakes into an enhancement proposal.
• Propose updates across agent instruction files, skills, docs, work records, memory, hooks, CI, evals, or structured local artifacts.
• Review whether a learning is durable enough to promote from chat/session context into a versioned artifact.

Do not use this skill for:

• Directly editing a requested target when the destination is already clear.
• Generic research summarization without a persistence decision.
• Creating or refactoring a skill's content; use skill-builder.
• Operating .work/; use work-cli.
• Changing repo convention contracts; use convention-engineering.

MCP is out of scope as a durable destination for this skill.

First Actions

1. Identify the session learning, correction, or proposed enhancement.
2. Gather only the nearby evidence needed to classify it:
   • user correction or session summary
   • changed files and tests
   • active work item or research space
   • relevant agent instruction files
   • related or used skills
3. Read references/externalization-model.md, then references/routing-taxonomy.md.
4. If the request mines agent session history for reusable value, read references/session-workflow-packaging.md.
5. Produce a proposal before making durable edits.

Operating Loop

1. Extract candidate learnings in compact form.
2. Classify the externalized burden and artifact class using references/externalization-model.md.
3. Classify authority, scope, durability, load policy, consumer, enforcement need, sensitivity, provenance, and cost.
4. Pick candidate target surfaces using references/target-surfaces.md.
5. Format recommendations with references/proposal-format.md.
6. Ask for approval before changing instructions, skills, hooks, CI, evals, or memory.

Routing Rule

• Route by externalized burden first, then by durable surface.
• Required always-on repo behavior goes to the repo agent instructions; use harness adapter files only when runtime-specific guidance is needed.
• Repeatable procedures go to skills or skill reference files.
• Interaction contracts go to protocol owners such as CLI contracts, schemas, adapters, Taskfiles, generated-surface rules, or convention docs.
• Bulky durable knowledge goes to docs or research pages.
• Active follow-up goes to .work records and spaces.
• Personal low-authority context goes to memory.
• Hard guarantees go to hooks, settings, CI, lint, or evals.
• Temporary state stays in session context, compaction, sandbox/workspace files, or run logs unless it proves durable value.

Output

Default output is a reviewable proposal in a hybrid layout — a narrow Recommendations index table followed by one section per recommendation:

• short summary of the durable lesson
• recommended destination(s)
• externalized burden and artifact class
• compact proposed change
• compact evidence and reasoning tied to routing dimensions
• confidence and risks

Use the full layout (index table + per-recommendation sections + Not Promoted table) from references/proposal-format.md. Add the structured YAML block only when another tool or agent will parse the result.

Gaps

This is a proposal skill. It does not yet own a script, CLI, load verifier, or automatic mutation workflow. Add code only after the proposal pattern has worked across multiple real sessions.