Use sdlc-implementation to execute
an approved plan in docs/execplans/<slug>.md, validate coherent milestones,
and hand PR delivery to pr-babysitting. The existing execplans skill owns
plan authoring and the living-document contract; sdlc-implementation owns
implementation, not another copy of the PR lifecycle.
install-skills installs the new skill and its references alongside the other
skills in ${HOME}/.codex/skills and ${HOME}/.claude/skills; the
rust-entrypoint home phase runs that installer too. Its managed helper
checkout retains both skills/ and agents/, so the workflow can read the
provider-neutral agents/subagents.yml manifest. The installer copies skills,
not agent definitions, into the user skill directories. Installation supplies
instructions, not CodeRabbit or Firecrawl credentials, Lody, agent processes,
or permission to publish or merge.
For example, after approving the identified plan:
Use sdlc-implementation to implement docs/execplans/<slug>.md in the assigned
PR. Follow its approved scope and tolerances, and use pr-babysitting to take
the PR through review and merge when the required gates hold.
State a narrower endpoint explicitly for implementation-only, review-only, or
no-merge work. The agent reads the plan's linked documentation, skills, and
decisions first. It normalizes branch/upstream and PR metadata safely, removes
only the leading Plan: prefix and its following space. It links the active
Lody session in the PR's final ## References section when applicable. It
commits frequently, keeps the plan's findings and progress current, and checks
only an explicitly associated roadmap item whose implementation activity is
complete.
Firecrawl provides bounded, version-matched external documentation; wyverns answer tightly scoped read-only repository questions; artisans implement one fully specified, independently testable task. The supervisor or journeyman retains architecture, integration, and acceptance. Read the effective agent manifest and actual tool access; do not widen a worker's permissions to make a packet succeed. See the delegation reference.
Scrutineers run sequential deterministic gates before each explicitly requested
milestone coderabbit review --agent. The
CLI backoff reference
uses the service's requested wait plus ten minutes, randomized up to ninety
minutes, or ten minutes above that minimum when it exceeds ninety. Hosted
review dispatch retains the managed queue's separate cooldown policy.
The implementation handoff contains the accepted plan, candidate, gate and
milestone evidence, proof inventory, roadmap state, PR metadata, and mutation
authority. It can arrive with CI red and the PR draft. pr-babysitting owns CI
monitoring and repair, the hosted completeness/correctness and applicable proof
assessments, readiness, rebasing, reviews, and authorized merge. It uses
monitoring-only scrutineers for CI rather than starting duplicate gate or
review runs. The implementation skill supplies bounded repairs and keeps its
plan current without running competing watchers or conversations.
The two initial assessment questions and all proof-evidence requirements live
only in the babysitting skill's
ExecPlan assessment reference.
For an implementation assignment, completeness/correctness must clear before
readiness. Introduced or materially changed CrossHair, Kani, Verus,
LemmaScript, or comparable proofs also require the separate substantive,
rigorous, well-founded assessment, with actual paths and named references.
Generic approval cannot replace it. Initial questions and necessary
candidate-bound follow-ups use the assigned manual-token procedure as
assessment conversations, not formal review requests. The agent actively awaits
and reads the substantive issue-comment replies before readiness. Assessment
conversations never enter comenq; formal review dispatch and its rate-limit
recovery belong after the verified ready transition. See
awaiting CodeRabbit.