Skip to content

Required skill changes

The current skills contain useful behavior, but several rules conflict with the approved architecture.

No skill should be edited until its exact diff is planned and approved.

agent-routing

Required changes:

  • preserve Claude as the main reasoning loop;
  • preserve read-only Claude subagents for research and review;
  • route every write task through Laneward;
  • remove the blanket requirement that every subagent creates its own write worktree;
  • make Laneward the sole owner of write worktrees;
  • require plan approval before write dispatch;
  • verify all delegated output.

codex-fleet

Required changes:

  • remove every exception that permits the agent to commit;
  • remove independent worktree creation for normal write execution;
  • turn fleet execution into a Laneward worker adapter;
  • retain bounded prompts, owned paths, model selection, and local validation;
  • keep direct use only as an explicitly approved emergency or diagnostic path;
  • default agent network access to denied.

Codex Fleet must no longer act as a competing orchestrator.

research-first

Required changes:

  • retain the no-edit research phase;
  • retain explicit approval before implementation;
  • replace Obsidian outputs with project-local plan and decision records;
  • record research assumptions and unresolved questions in the plan revision.

session-handoff

Required changes:

  • remove Obsidian as the state destination;
  • load and update project-local handoff data;
  • include active Laneward plan, pending approvals, ready-for-review lanes, and runtime state;
  • treat Laneward as live operational truth;
  • avoid copying transient status into committed decision files.

explain-plainly

Keep and extend:

  • use it for plan explanations;
  • use it for approval requests;
  • use it for verification-layer findings;
  • separate user impact from technical details;
  • never hide risk behind simplified wording.

ship-repo

Required changes:

  • preserve explicit approval for merge and push;
  • integrate with Laneward plan and integration-branch state;
  • reject publication when integration gates are incomplete;
  • record the user approval before remote mutation.

ship-service

Required changes:

  • preserve separate approval for runtime effects;
  • use .laneward/project.json for installation and verification contracts;
  • require rollback for consequential service changes;
  • record real service evidence before declaring completion.

gptpro-handoff

Status: disabled.

Reason: there is no active GPT Pro subscription. The skill may remain archived or installed, but routing must never select it automatically and no workflow may depend on it.

Status: excluded.

OMP has been removed from the architecture and roadmap. No fallback, adapter, or placeholder should be added.

Prompt style

The existing prompt-style guidance may remain a reference for difficult prompts. It is not an execution component and must not become a mandatory layer for routine work.

New laneward-orchestrator skill

A focused skill should eventually define how Claude:

  • reads project and Laneward state;
  • performs research;
  • constructs a plan;
  • requests approval;
  • creates lanes;
  • waits for or inspects results;
  • reviews and commits;
  • integrates;
  • requests runtime approval;
  • verifies real operation;
  • closes the plan.

This skill describes the standard workflow. Hooks and Laneward enforce the hard boundaries.