source-command-pipeline-dev-start

name	source-command-pipeline-dev-start
description	Start a focused session on one objective — locks scope, blocks opportunistic refactors.

source-command-pipeline-dev-start

Use this skill when the user asks to run the migrated source command dev-start.

Command Template

🎯 Mode: Focused Dev

You just entered dev-start mode with the objective:

$ARGUMENTS

Posture for this session (until told otherwise)

1. Sync preflight — look at GitHub before coding [ADR-0026]. Run:

   node contextkit/tools/scripts/sync-check.mjs preflight
   

   It reports ahead/behind, recent in-flight branches, and open PRs with their CI/review status (flagging any awaiting status). If an open PR or a recent branch overlaps this objective, surface it and confirm with the user before duplicating work — coordinate or /claim first. gh missing/unauthed degrades to the git-only view; it never blocks. Behind upstream is a fact to surface; do not pull automatically.

2. Resolve the economy plan before broad context. Pass the objective as one explicit argv value — it is classified as data and is never executed:

   node contextkit/tools/scripts/economy/dev-start-bootstrap.mjs --objective -- "$ARGUMENTS"
   

   The fail-open, read-mostly bootstrap reports ordered stages and structured lever states for resume/checkpoint, Project Map freshness + focused path/symbol lookup, task intake, RequestOrchestrator, the dev-start context profile and run-compact. It never regenerates Project Map, runs free-form text, changes git state or moves a card. Use --json for cdk-dev-start-bootstrap/1.

   Economy canaries to apply from that plan:

   • If the bootstrap or Project Map reports an exact symbol/path match, run Task Compiler in read/compile-only mode: node contextkit/tools/scripts/economy/task-compiler.mjs --symbol <exact-symbol> --objective "$ARGUMENTS". If the match is partial/ambiguous/missing, report task-compiler skipped: no exact Project Map match.
   • If the bootstrap reports a previous checkpoint/run id, render the resume pack before re-reading old context: node contextkit/tools/scripts/economy/resume-pack.mjs <runId>. If no checkpoint exists, report resume-pack skipped: no checkpoint.
   • If RequestOrchestrator suggests subagent work, resolve the bounded profile: node contextkit/tools/scripts/economy/subagent-profile.mjs, then attach that profile to each subagent packet. This is advisory and must not increase dispatch count.

3. Read the current state first — run:

   node contextkit/tools/scripts/context-pack.mjs --profile dev-start
   

   [ADR-0027]: one bounded bundle (latest-session digest + [Unreleased] + immutable rules + open backlog + recent ADRs) in a single call instead of opening each file. Open a full source only if the pack flags something to inspect.

4. Right-size the work [ADR-0030]. Classify the objective before committing to a process — don't over-engineer a typo or under-plan a migration:

   node contextkit/tools/scripts/complexity-rubric.mjs classify "$ARGUMENTS"
   

   It returns a tier (trivial → no ADR/no story · feature → story · architectural → /new-adr FIRST) and detects a regulated domain. If it flags a domain (LGPD / fintech / healthcare …), auto-route to the named agents (e.g. @privacy-lgpd + @security) and treat the work as architectural. The tier is advisory, not a cage — state it and adjust with the user if it misreads.

   Auto-start a referenced task [ADR-0034]: if the objective names a backlog task id (e.g. "fix 042" / "ticket 058"), use the bootstrap's correlated id, move it into working/, and attach it to this session:

   node contextkit/tools/scripts/pipeline.mjs start <id>
   

   While you work, the task heartbeat is renewed on every edit. At completion, check off its acceptance criteria so the Stop hook can auto-conclude it.

   Model routing is canary/active when configured [ADR-0094] — the boot banner shows the actual mode (canary by default in templates; dogfood may use active). Apply the posture Haiku operates · Sonnet executes · Opus decides without being re-prompted: run ≤3 simple deterministic commands directly (runner-first — never spawn an agent for one trivial command); batch mechanical investigation (grep/glob/tests/lint/log triage) to a Haiku agent that returns a compact pack; delegate bounded low/medium-risk implementation to Sonnet under a short contract; keep Opus for decisions and for implementing high/critical-risk code directly (auth/RLS, migrations, concurrency, public contracts). Never auto-select Fable. Minimize total task cost, not just per-token price — don't delegate when coordination costs more than direct execution. Disable per session via routing.enabled=false.

5. Define IN-SCOPE / OUT-OF-SCOPE explicitly from the objective. Show the user:

   ✅ IN-SCOPE: <what we will touch>
   ❌ OUT-OF-SCOPE: <what we will NOT touch, even if tempting>
   

   Ask for confirmation before proceeding if there is ambiguity.

6. Scope lock during the session:

   • Do NOT suggest refactors in files outside IN-SCOPE.
   • Do NOT "while we're here" rename/reorganize adjacent code.
   • Do NOT add new dependencies without asking.
   • If you spot a problem out of scope, note it and mention it at the END ("for next session: X, Y, Z") — do not act on it now.
   • Correct-course checkpoint [ADR-0030]: if mid-session the work clearly outgrows the agreed scope (a feature turns out to need a migration, a new dependency, or an auth change), STOP and re-classify with the user before continuing — don't silently let scope creep past the tier you agreed on.

7. Break the objective into 3–7 concrete tasks and track them with task plan/checklist.

8. Workflow spec pack context [ADR-0057]: if the objective names a workflow slug or the task card has workflow: / spec: metadata, read contextkit/memory/workflows/<slug>/prd.md, spec.md, tasks.md, and memory.md before editing. Do not duplicate those artifacts in the task; link them and keep implementation evidence in the card/report.

9. Per-task scratch (optional): if you accumulate ephemeral notes while a ticket is in contextkit/pipeline/testing/, drop them in a sibling file named NNN-*.scratch.md next to the ticket. The pipeline's .gitignore excludes *.scratch.md — scratches are local-only. At conclude time, summarise the useful parts into the ticket body and let the scratch be discarded.

10. Before opening a PR — re-check sync [ADR-0026]. Run node contextkit/tools/scripts/sync-check.mjs prepr --fetch (or just use /git pr, which runs it): it re-confirms you are not behind main and that no open PR already exists for this branch before you create one. (--fetch refreshes remote refs — read-only checks skip the fetch by default, ticket 065.) Don't duplicate a PR; push to update the existing one.

11. At the end: offer /log-session (or /new-adr if an architectural decision was made).

Why this mode exists

Sessions without a defined focus tend to become giant refactors that mix intentional change with incidental cleanup — impossible to review, and the changelog becomes a patchwork. Scope-locking fixes that at the root.