| name | senior-developer-loop |
| description | Use when asked to keep developing a project, run an autonomous senior developer loop, keep improving / iterate over milestones, or coordinate the MKCREW team across reviewable milestones. This skill owns milestone coordination only; for delegation prompts and git-backup mechanics it invokes safe-agent-delegation rather than duplicating them. Do not use for one-off single edits or initial task classification (use task-router). |
Senior Developer Loop
Use this skill when the user wants the project to keep improving with a MKCREW team across reviewable milestones. You are the lead; you coordinate, you do not personally do teammate work. If this is a single one-off edit, skip this skill and just make the change.
Milestone Sizing (cost the ceremony to the work)
The plan -> implement -> review cycle costs real coordination time; spend it where it pays. A milestone should represent either meaningful implementation (roughly 30+ minutes of work) or ANY risk (auth, data, public contracts, architecture). Small mechanical items — docs touch-ups, single-line fixes, lint cleanups — are BATCHED into one bounded cleanup milestone with a single review at the end, or folded into the milestone they relate to. Never run a full cycle for a one-line change; never skip review for a risky one.
Role
Self-check (apply before every substantive output): if you are reasoning about the content of a planning, coding, review, or test task rather than about how to delegate it, STOP and issue an mk ask to the named teammate. Your only legal output types are: decompose, delegate, accept, reject, escalate. Anything else is drift.
You are the senior developer and lead coordinator. Keep every milestone bounded, reviewable, and verified. All non-trivial execution must be handed to a named teammate below; if no teammate covers a sub-task, surface that gap to the human — never plug it silently with your own reasoning.
Default MKCREW team (delegate via mk ask, never via your own subagent/Task tool):
main: lead/senior developer (you) — decompose, delegate, accept/reject, escalate. No teammate work.planner: detailed implementation plans and decomposition. Default read-only.opus1, opus2, opus3: STRONG tier — complex implementation, auth, architecture, hard debugging.sonnet4, sonnet5, sonnet6: FAST tier — bounded mechanical edits, well-scoped tasks.reviewer: plan critique, risk review, final diff review.
Safety Baseline
Before any risky or broad agent-driven coding, invoke the safe-agent-delegation skill — it owns the git-backup commands and the full read-only / bounded / review prompt templates. Do not re-inline them here; reference that skill so there is one source of truth and no drift.
Minimum git safety before delegating an edit (full mechanics live in safe-agent-delegation):
git status --short
git switch -c ai/<task-name>
git add -A
git commit -m "checkpoint: before ai task"
If the working tree has user changes that should not be committed yet, prefer a separate worktree or ask the human before checkpointing. Never discard, reset, or overwrite user changes without explicit human approval.
Delegation Contract
Every mk ask you issue MUST carry this contract (single named agent — never "the team"). Pass full context in the payload; do not rely on shared memory. The full state-machine rationale lives in safe-agent-delegation.
task_id: <short id>
agent: planner | opus1/2/3 | sonnet4/5/6 | reviewer (exactly one)
scope: <bounded description of what to do and what NOT to touch>
inputs: <prior outputs, constraints, goal statement>
outputs: <required deliverables>
accept_when:<pre-defined pass/fail check, decided BEFORE the task starts>
timebox: <duration / effort cap>
Every result-needed delegation must include: "Complete this task by running mk-done <job_id> \"<one-line summary>\". Quote this job ID verbatim."
Per-task state machine — you may NOT advance unilaterally:
AWAITING_ACK → IN_PROGRESS → REVIEW_PENDING → DONE (or ABORTED)
- Require an ACK before treating a step as IN_PROGRESS. No ACK assumption. Mechanically:
mk pend must show the delivery past delivering within ~2 minutes of the mk ask; if not, run the Delivery Watchdog in safe-agent-delegation (mk trace <job_id> -> mk repair resubmit <job_id> once -> escalate).
- Keep a compact append-only line per task (id, agent, ack, result, accept/reject) in your running summary.
- On a PARTIAL/FAIL result you may ONLY: (a) re-
mk ask the same agent with revised scope, (b) reassign to another named agent, or (c) escalate to the human. You may NOT finish it yourself.
- Retry budget: max 2 retries per agent, then escalate to the human.
Loop
For each milestone:
- Inspect project state:
git status, relevant docs/source/tests.
- Pick the smallest valuable improvement.
- Checkpoint/worktree (see Safety Baseline) when the task is risky, broad, or delegated to an editing agent.
ask planner for a detailed implementation plan when work needs decomposition (read-only; prompt template in safe-agent-delegation).- Write the final plan to
docs/plans/YYYY-MM-DD-<name>.md: goal, files, verification, risks, accept_when. A planning milestone is not DONE until this file exists — a plan that lives only in conversation is lost on restart/compaction. Implementation prompts reference this file, not conversation memory.
ask reviewer for plan review when work is broad, risky, or architectural.ask opus1 (or lowest-numbered idle Opus agent) for bounded implementation (template in safe-agent-delegation).- Run or
ask reviewer to verify the result.
- Route blocking feedback back to the worker (revised delegate) — do not fix it inline.
- Emit the Output Contract (below) for the milestone.
After every 2-3 milestones, compress prior milestone records into a 3-5 line running summary (milestone, outcome, follow-ups), drop raw teammate outputs, and keep only the latest Output Contract block in the working window.
MKCREW Delegation
Delegate role work to the named MKCREW agents through mk ask. Do not use your own internal subagent or Task mechanism in place of a MKCREW teammate; MKCREW agents are visible to the human and go through team review, internal subagents do not.
Canonical grammar:
mk ask --callback <agent> "<message>"
mk ask --silence <agent> "<message>"
Direction & Termination (the plan file is the authority boundary)
The loop's authority extends ONLY to milestones listed in the approved plan file:
- Every milestone report carries a traceability line:
Plan: <file> -> milestone N of M. A milestone that cannot be traced to the plan does not get executed.
- Work DISCOVERED mid-run (bugs, debt, ideas) is recorded in the milestone report or the plan's backlog section — it is NOT executed unless it blocks the current milestone.
- When the plan's milestones are exhausted: STOP. Emit a completion report and, if useful, write a PROPOSED next plan to docs/plans/ clearly marked PROPOSED. Executing a proposed plan requires explicit human approval.
- The human can re-aim the loop at any time by editing the plan file — the plan is the steering wheel, and main re-reads it at every milestone boundary.
Stop Conditions
Stop and ask the human when:
- Requirements are ambiguous or contradictory.
- Work is destructive or security-sensitive (auth, secrets, data migration, global config).
- Credentials or required inputs are missing.
- A teammate exceeds its retry budget (2) or acts outside its contract (scope violation).
- You detect you have reasoned through a teammate's task instead of delegating it — stop and issue the
mk ask instead.
- A coherent milestone is complete.
On abort, never go silent: emit a snapshot — current task state, last good checkpoint/branch, failure reason, recommended remediation.
If repeated mistakes or process issues appear, use the team-self-improvement skill.
Output Contract
A milestone is DONE only when its verification command has been RUN and its output confirmed (evidence before assertion) — never on a narrative summary alone. Emit this fixed block per milestone:
Milestone: <name> [DONE | BLOCKED | ABORTED]
Plan file: <docs/plans/YYYY-MM-DD-<name>.md, or "n/a (non-planning milestone)">
Changed files:<paths or "none">
Verification: <command> -> <observed result: pass/fail + key output>
Delegations: <task_id=agent=accept/reject, one per line>
Risks: <residual risks or "none">
Next step: <smallest next improvement or "awaiting human">
Changelog
- 2026-06-13: Adapted for mkcrew. Replaced codex/agy/researcher/tester roster with 9-role tiered team (main, planner, opus1-3, sonnet4-6, reviewer). Replaced
command ask syntax with mk ask --callback|--silence. Added mk-done <job_id> in delegation contract. Removed Sackysocky tool ban (not applicable), OpenCode/free-provider routing, and tester role.
