Implement the CRs in $ARGUMENTS. This command does not author CRs; authoring goes through the governance skill.
Usage: /implement-cr <CR-ID | docs/cr/CR-*.md> [more CRs ...]
Preconditions
The procedure per CR is the change_flow of the operating_model block, with the subagent_creation, task_delegation, continuous_commit, testing, and user_scenario_testing blocks of the consumer's agent instructions. This skill depends on those blocks; when they are loaded, apply them, do not restate them.
Fallback. When those blocks are absent from the agent instructions, follow skills/governance/reference/cr-implementation-workflow.md instead, so a consumer without them still gets a working implementation step.
Tools
- Create subagents only through the
Workflowtool. Load the workflow-authoring skill before the first workflow. - Every
agent()call passes modelclaude-opus-5-5and effortlow. The oracle of themandateblock is the exception: it uses modelclaude-fable-5-1and efforthigh. A tier that the task states stays permitted. - Commit through the
/checkpoint-commitworkflow and recover context through/checkpoint-read.
1. Parse
- Each argument is a CR ID (
CR-NNNN) or a path underdocs/cr/. - If
$ARGUMENTSis empty, stop and print:usage: /implement-cr <CR-ID | docs/cr/CR-*.md> [more CRs ...].
2. Resolve all arguments first
- Each argument must match exactly one
docs/cr/CR-NNNN-*.mdthat is not a-validation-report,-rollback-log, or-iteratefile. - Files under
docs/cr/iterations/are iteration ledgers and MUST NOT be resolved as a CR. - If any argument fails, stop before any work, name the failed argument, and make no commit.
3. Recover context
Run /checkpoint-read before the first CR, and again whenever the run resumes after context loss. When a CR in the argument list already has checkpoint commits, continue that CR from its last checkpoint rather than from its start, and do not repeat a CR that already completed.
4. Run each CR, strictly sequentially
In argument order. A CR starts only after the previous CR completes, so it reads the tree the previous CR left.
- Layer 0: write the intent statement from the CR.
- Review: a separate agent in a workflow, which did not author the CR, reviews it. On approval, set frontmatter
status: approved. - Layer 2: update the docs first.
- Layer 3: implement the code from the spec diff. Delegate each unit of work per
task_delegation, with the fixed 8-section prompt. - Commit each coherent step through the
/checkpoint-commitworkflow with the scopeCR-NNNN, so the subject ischeckpoint(CR-NNNN): .... - Gate destructive changes with a rollback log per
core_principles. - Verify against the CR acceptance criteria, then against the intent. Use a user scenario per
testingwhere the CR has a user surface.
If a CR fails review or verification, stop. Do not start later CRs.
5. Report
- One line per CR:
CR-NNNN status: completed|failed|not-started. - For a failed CR, add the failed step.
- Set frontmatter
status: implementedon each completed CR. - End the report with the
/iterate-crcommand that covers every completed CR, for example/iterate-cr CR-XXXX CR-YYYY, as the next step for post-implementation adjustment.