Teamwork init

Teamwork Init

Make project-local agent context accurate, small, and maintainable. This skill owns only the named project; it never refreshes global Teamwork installations. Ordinary natural discussion or brainstorming does not activate Init unless the existing project-local outcome trigger applies.

Authority

An audit request is read-only. An explicit initialize, repair, migrate, or slim request authorizes only the corresponding files inside the named project. It does not authorize edits under user-global config directories, host settings, credentials, plugin catalogs, global skills or agents, dependency installation, remote services, Git publication, or release work. Root alone asks users through the current host's native surface; leaf roles return proposed questions or blockers to Root. Ask only for missing mutation authority or unavoidable manual host action. Answers do not expand authority.

Resolve the exact project root before writing. Preserve unrelated content and managed-block boundaries. If ownership is unclear and safe merging is impossible, stop with the concrete conflict instead of overwriting.

Full Bootstrap And Candidate Inputs

Emit the complete Capability Matrix only for an explicit full bootstrap. Audit, repair, migration, or slimming must not manufacture that broad matrix merely from some platform inspection.

Treat external-memory or docs-graph output as candidate-only context, never as project or Teamwork truth. Candidate-promotion gates (all must pass): currentness | scope | direct evidence | privacy/protected-data review | Root authority. Do not promote it into instructions, memory, routing, or durable artifacts until those five gates pass. A logged, partial, or permissive gate result is not promotion; a failed or missing gate leaves the candidate unpromoted and is reported as a concrete limitation. No logged, partial, permissive, fallback, or exception path promotes candidate material.

Project Workflow

1. Dispatch Explorer for the read-only audit of instruction hierarchy, platform surfaces, human docs, source/configuration, test commands, runbooks, and docs/teamwork/ context. Ground every durable rule in project evidence; do not invent commands, paths, architecture, model mappings, or capabilities.
2. Keep only stable behavior-changing facts: project purpose, canonical owners, required commands, boundaries, source-of-truth paths, and local tool policy. Leave volatile progress, experiments, and temporary failures in their tracker.
3. Give each fact one canonical owner. Merge duplicates and use short pointers or platform deltas elsewhere. Do not copy external docs, schemas, or manuals into project instructions.
4. Keep local inspection and clear authorized implementation native. Describe special tools such as CodeGraph only when configured or requested, with use and unavailable behavior.
5. For Teamwork memory, preserve ordinary retrieval metadata separately from the single optional Grill record. Never rebuild a discussion transaction, hidden lifecycle, or skill-reference graph.
6. Initialize a repository-local CodeGraph index only when the requested setup includes it and the CLI is available. Otherwise report the exact gap; do not install external tooling or change host configuration without separate authority.
7. Only after the audit identifies an authorized change, dispatch Worker with exact project-local ownership. Worker applies the smallest complete mutation, preserves recovery state for the whole operation, and fails closed on ownership or migration conflicts. Root retains any privileged host action.
8. Re-read every changed instruction surface and exercise the nearest real project-local validation or command. If no real activation check is available, state that limit instead of treating syntax or file presence as live proof.

An audit/check-only Init remains read-only and conversational. After a mutating Init in an initialized writable project, a receipt is a completion companion and defaults through Writer unless the user says no files, off-record, read-only, no writes, or equivalent. Freeze a bounded receipt packet: purpose/audience, facts/sources, frozen decision/status, style/structure, artifact kind/consumer, preserve/forbid, changed surfaces, evidence, validation, and human action. Dispatch one low-cost Writer; Root may do only answer-invariant handoff work while Writer runs and must join and read back before claiming the receipt is saved or durable. Writer uses artifact-inspect -> artifact-schema <create|update|supersede> -> artifact-apply; the transaction derives the destination and registers the ordinary index. Writer is disposable compute and the transaction owns destination, compare-and-swap, journal recovery, atomic apply, and readback. If interrupted before generic artifact apply begins, there is no durable claim; recover only from surviving workflow evidence or report unsaved. Missing project memory, Writer, brief, authority, consumer, or transaction blocks only persistence: deliver the receipt and report it unsaved/blocked. No Root or Worker fallback writes it.

If the Explorer audit finds no decision-relevant change, dispatch no Worker and write nothing. Report the selected project surface, changed or proposed files, canonical ownership decisions, verification strength, conflicts, and any remaining human action. Never invoke a global update as part of project initialization.