| name | graft-boot |
| description | Repository-specific startup workflow for Graft. Use when the task starts from a short prompt such as "boot", "continue", "read AGENTS", or "next step", and Codex should first ground itself in AGENTS.md, the ai-plan/ documents, the current repo state, assess whether multi-agent work is justified, and then enter implementation with the repository's closeout and commit workflow in place. |
Graft Boot
Use this skill to start or resume work in Graft with minimal prompting.
Treat AGENTS.md as the source of truth. This skill performs startup plus the mandatory workflow hooks that follow
startup; it does not replace repository rules.
Startup Workflow
- Run the startup preflight defined in
AGENTS.md 4.1 Startup Governance.
- Emit the minimum startup receipt from
AGENTS.md before substantive work:
governance source
task class
recovery source
- If the current turn needs recovery context, read
ai-plan/public/README.md only after preflight, then follow the
mapped parent topic and relevant subtopic recovery files for the current task shape.
- Read the relevant repository-wide design and roadmap truth needed by the task.
- Inspect the current repository state before assuming toolchains or entrypoints exist.
- If the request introduces new long-running work and no active topic already owns it, route through
graft-work-intake before creating a new topic, design, roadmap, or ADR.
- Identify the first concrete boundary decision before editing.
- If the task touches live schema or migration files, re-read the
server/AGENTS.md migration/comment rules and
treat table/column comment completeness as a mandatory implementation item, including core-owned handwritten
migration directories such as server/internal/httpx/migrations/**.
- Assess whether
graft-multi-agent-batch is justified:
- use it only when the task is large enough, write scopes stay disjoint, and the current slice owner can keep the
immediate blocking step local
- when the task explicitly uses
graft-multi-agent-loop, the outer loop owner may instead delegate the whole
bounded round to one worker subagent and keep only orchestration, budget tracking, acceptance, and stop
conditions local
- do not enable it for small, overlapping, or review-hostile slices
- Before edits, tell the user what you read, how you classified the task, whether multi-agent work is justified, and
the first implementation step.
- When the current slice reaches a stop, completion, or handoff point, route the ending through
graft-task-closeout
instead of relying on an implicit wrap-up path.
graft-task-closeout must evaluate commit eligibility through graft-commit rules:
- if validation and ownership allow a safe scoped commit, use
graft-commit
- if they do not, report the exact blocker and keep the handoff status honest
- If the current turn ends by proposing a next task, include one explicit next-task startup prompt that restates the
startup receipt fields needed by the next turn instead of assuming boot state carries across turns.
Recovery Rules
- recovery follows startup preflight; it does not replace it
- prefer repository truth over assumptions
- if the repo still lacks a stable build or runtime contract, say so explicitly and keep validation expectations honest
- if docs and code diverge, update the docs first or in the same change