| name | implement |
| description | TRIGGER when: user asks to implement, fix, build, or work on something — whether from a
docs/wip plan OR a standalone task (bug fix, GitHub issue, one-off change).
Examples: "work on task 1", "fix this bug", "implement feature X from the issue".
Provides structured execution with profile detection, dependency handling, review checkpoints.
|
Implementing Work
Conventions
- Read capy knowledge base conventions at shared-capy-knowledge-protocol.md.
- Read profile detection at shared-profile-detection.md. When the sub-task's target files activate a profile that contributes an
implement/ subdirectory (e.g., ../../profiles/k8s/implement/), its index.md lists per-task gotchas the skill must consult BEFORE writing. See Step 2.
Modes
Two modes, determined automatically: plan mode when the user references a docs/wip feature or task number; standalone mode otherwise (bug fix, GitHub issue, one-off change). When ambiguous, ask.
Both modes share the same execution core (Step 2 onward) — profile detection, dependency handling, verification, review.
Required Outputs
After each execution + review cycle, verify all outputs:
Indexing ownership: Review skills ($kk:review-code, $kk:review-spec) index their own findings. This skill only indexes kk:project-conventions for non-obvious patterns discovered during implementation. Do NOT duplicate review indexing here.
Review Mode
By default, review checkpoints use isolated mode (kk:review-code:isolated, kk:review-spec:isolated). This is mandatory because the implementing session has authorship bias — the same model that wrote the code produces weaker reviews of it. Isolated mode spawns an independent sub-agent with no prior exposure to the implementation.
The user can override at any checkpoint ("use standard review for this one") to fall back to in-session $kk:review-code.
Workflow
Mandatory order — understand before executing. The flow below is strictly sequential. Do not read source files to modify, write code, edit files, run tests, or otherwise act on any task until you have loaded full context (design, implementation plan, task list in plan mode, or full problem understanding in standalone) and completed profile detection and loaded all resolved profile content. The only early contact with the codebase is the task's target filenames — enough to drive profile detection, not enough to pattern-match implementation.
The Process
Step 1: Load Context
Determine mode (see §Modes), then read the appropriate mode file and follow its entry procedure:
- Plan mode: Read plan-mode.md — loads tasks.md, design.md, implementation.md, identifies next task.
- Standalone mode: Read standalone-mode.md — parses the problem, explores relevant code, forms an approach.
After completing the mode's entry procedure, continue with Step 2.
Step 2: Execute
Mandatory order — instructions before action. Steps 1–3 load instructions; step 4 is the first step that touches subject matter. Do not write code, edit files, or otherwise act until steps 1–3 have been performed in order. If a later step reveals that an instruction was missed, return to step 1.
- (Plan mode only) Update
tasks.md: set the task's status to in-progress.
- Profile-aware per-task gotchas (pre-write). Run the
shared-profile-detection.md against the target files (and any diff-so-far). For each active profile that contributes an implement/ subdirectory, load ../../profiles/<name>/implement/index.md and read the always-load + any matching conditional content. Apply those gotchas to the upcoming edits — they exist to prevent mistakes the post-write reviewer would otherwise catch. If no active profile contributes an implement/ subdirectory, skip this step.
- Dependency-handling (pre-write). Whenever the task introduces or changes a dependency — new import, version bump, unfamiliar call, and per the widened trigger also: a Kubernetes API version, a CRD, a Helm chart or chart dependency, or a container image tag/digest — apply the
$kk:dependency-handling skill BEFORE writing the call. Do not guess signatures, API versions, or configuration; look them up via capy/context7 per that skill's rules. Per-profile lookup cascades live in each profile's overview.md (e.g., ../../profiles/k8s/overview.md §Looking up Kubernetes dependencies).
- Make the changes. (Plan mode: follow the plan exactly.)
- (Plan mode only) Check off subtasks (
- [x]) in tasks.md as you complete them.
- Run verifications; run
$kk:test skill.
Step 3: Report and Review
- Show what was implemented
- Show verification output
- Load
kk:review-code:isolated skill — this handles both sub-agent and pal codereview internally with independent reviewers. Do NOT run a separate pal codereview call, as it is already included in the isolated workflow.
- Based on user and code-review feedback: apply changes if needed and finalize
- (Plan mode only) Update
tasks.md: set the task's status to done
After finalizing, verify all items in the Required Outputs section above:
If any item is unchecked, go back and complete it. Do NOT proceed to the next task with incomplete outputs.
Step 4: Continue (plan mode only)
Follow the iteration procedure in plan-mode.md — move to next task, repeat Steps 1–3.
Step 5: Complete (plan mode only)
Follow the completion procedure in plan-mode.md — final validation, documentation, reflection.
When to Stop and Ask for Help
STOP executing immediately when:
- Hit a blocker (missing dependency, test fails, instruction unclear)
- (Plan mode) Plan has critical gaps preventing starting
- You don't understand a requirement or instruction is ambiguous
- Verification fails repeatedly
IMPORTANT! Always ask for clarification rather than guessing.
When to Revisit Earlier Steps
Return to Step 1 when:
- Partner updates the plan or clarifies the problem
- Fundamental approach needs rethinking
IMPORTANT! Don't force through blockers — stop and ask.
Remember
- Review plan critically first
- Follow plan steps exactly
- Don't skip verifications
- Use skills when applicable (
$kk:dependency-handling, $kk:test, $kk:review-code:isolated) (Plan mode: also when the plan says to do so)
- Between batches: just report and wait
- Stop when blocked, don't guess
