start-dev

Compatibility entry point for "implement this". Routes the request to the right piece of the workflow and stays out of the way.

Direction of travel. Implementing and delivering are now separate skills: worker-dev (implement → validate → commit → handoff, inside an objective) and deliver-dev (review the integrated diff → gates → push → one PR for the objective). /start-dev remains a fully supported front door: for a simple standalone task it still runs the same implement→PR flow it always has, and for work an objective already owns it hands off to worker-dev and opens no PR. Nothing you invoke today stops working. Over time, reach for /orchestrate-dev for multi-part work, and worker-dev/deliver-dev directly when you want one half of an objective's flow.

Why this skill exists

start-dev used to run one monolithic implementation→PR flow, which quietly assumed every task is its own pull request. That assumption breaks the moment a single user objective is split across several tasks: you get five PRs for one feature, five CI runs, five review cycles, and no place where the combined diff is ever reviewed. The routing below fixes that while preserving the one-task-one-PR path for the case where it is genuinely correct.

Input

Parse $ARGUMENTS as one of:

• GitHub issue number(s): #258 or 258 259 260
• A free-text task description: "Add rate limiting to the /login endpoint"
• A file path to a spec or plan: path/to/plan.md

If $ARGUMENTS is empty, ask "What do you want to implement?" and stop.

Step 0 — Route (do this before anything else)

REPO_ROOT="$(git rev-parse --show-toplevel)"
SKILL_DIR="${CLAUDE_SKILL_DIR:-$REPO_ROOT/skills/dev-workflow/start-dev}"

ROUTE="$(bash "$SKILL_DIR/scripts/detect-objective.sh" "$REPO_ROOT")"
MODE="$(echo "$ROUTE" | jq -r .mode)"
OBJECTIVE_ID="$(echo "$ROUTE" | jq -r '.objective_id // empty')"
echo "$ROUTE" | jq -r '"route: \(.mode) — \(.reason)"'

Then pick exactly one of three routes:

The multi-part test (route B)

Skip this test entirely — route C is the answer — when orchestrate-dev sent the work here. Its fall-through is one-directional: it already applied the orchestration test and decided against it, so re-applying the test here would bounce the request between the two skills forever. Recognise the hand-off by any of:

• TAMIRS_ROUTED_FROM=orchestrate-dev in the environment;
• the invocation says the orchestration decision is already made, or "do not route back";
• this run was reached from orchestrate-dev in the same session.

If the work genuinely looks bigger than orchestrate-dev judged, say so in one line and let the user decide — never re-route on your own.

Otherwise, route to orchestrate-dev when two or more of these hold:

• the work splits into parts that could progress independently;
• it needs more than one role (e.g. implementer and test-engineer and reviewer);
• it plausibly touches more than ~10 files or more than one subsystem;
• the user's phrasing is objective-shaped ("build the auth system", "migrate X to Y") rather than task-shaped;
• a spec/plan file was passed containing multiple distinct phases.

Say so in one line and offer the route rather than silently switching:

This looks like a multi-part objective (auth API + migration + tests, 3 roles).
Running /orchestrate-dev — one objective, one PR at the end.
Say "just do it directly" and I'll run it as a single task instead.

If the user declines, continue on route C — never refuse the work.

Route A — worker only (objective already active)

The objective owns delivery. This run ends at commit + handoff, exactly like worker-dev invoked directly.

Skill: worker-dev
Arguments: <objective-id> <task-id>

worker-dev needs both ids and reads the task file for its scope, role and validation_tier. Resolve them before dispatching:

SHARED_DIR="$REPO_ROOT/skills/dev-workflow/_shared/scripts"
OBJECTIVE_ID="${OBJECTIVE_ID:-$(bash "$SHARED_DIR/objective-state.sh" active)}"
TASK_ID="$(bash "$SHARED_DIR/objective-state.sh" next "$OBJECTIVE_ID")"

If no task matches the user's request — they asked for something the objective does not cover — say so and ask whether to add a task to the objective or run it standalone (route C). Do not invent a task id.

Hard stop: do not gh pr create, do not gh pr merge --auto, do not merge the base branch, and do not run the full repo suite unless the task's validation_tier says to. Delivery for the whole objective happens once, in deliver-dev.

Report at the end:

Task complete — committed on <branch>, handoff written to
.dev-files/objectives/<id>/handoffs/<task-id>.json
No PR opened: objective <id> owns delivery (run /orchestrate-dev or /deliver-dev to ship it).

Route B — multi-part objective

Skill: orchestrate-dev
Arguments: <the original $ARGUMENTS>

orchestrate-dev builds the task DAG, runs the workers (in parallel where the platform supports subagents, sequentially where it does not), integrates, and calls deliver-dev once. Nothing further is required here.

Route C — simple standalone task (the classic /start-dev behaviour)

This is the original flow, unchanged. worker-dev and deliver-dev are objective-scoped — they need a task file and an objective branch respectively — so a genuine one-off task runs here rather than through them. This is exactly what deliver-dev means when it points non-objective work back at "start-dev's delivery path".

Sequence:

1. Understand — gh issue view <n>, or read the spec file; confirm your understanding in one sentence. When working from a GitHub issue, read the Resume block first and stop if it reports a blocker:

   SHARED_DIR="$REPO_ROOT/skills/dev-workflow/_shared/scripts"
   bash "$SHARED_DIR/parse-issue-resume.sh" <issue_number>

2. Workspace — derive a slug branch (feat/add-rate-limiting) and resolve a worktree:

   WT_JSON="$(bash "$SHARED_DIR/resolve-worktree.sh" "$REPO_ROOT" "$BRANCH")"
   WORKTREE="$(echo "$WT_JSON" | jq -r .worktree_path)"

   All edits happen inside $WORKTREE, never in the main checkout.
3. Implement — Read before Edit; one conventional commit per logical unit; git add -p, never git add ..
4. Validate — Tier 1 first, then Tier 2 before pushing:

   bash "$SKILL_DIR/scripts/detect-stack.sh" "$WORKTREE"

   Run each command it prints, one at a time, from $WORKTREE, and stop at the first failure:

   if bash "$SHARED_DIR/detect-multi-platform-repo.sh" "$WORKTREE"; then
     bash "$SHARED_DIR/run-pre-pr-gates.sh" "$WORKTREE"
   fi

   Abort on the first failure. Never --no-verify, never skip because "it's urgent".
5. Push and open one PR — read references/pr-templates.md for the right body template and fill every […] placeholder:

   git -C "$WORKTREE" push -u origin HEAD
   gh pr create --title "feat(auth): add rate limiting to login endpoint" --body "<filled template, incl. Closes #N>"

6. Hand to pr-dev — merge policy (auto-merge or not) is resolved by pr-dev against repository policy, not forced here. Run /pr-dev <N>.

Want the task tracked like an objective anyway? Create a one-task objective first and the work flows through the same machinery as everything else:

bash "$SHARED_DIR/objective-state.sh" init "<slug>" --title "<task>"
bash "$SHARED_DIR/objective-state.sh" task-add "<slug>" --role implementer --scope '<glob>' --tier worker

Then route A applies on the next invocation, and deliver-dev ships it. This is optional — a one-off task does not need objective state.

Validation tiers

Tier definitions live in core/policies/validation.md.

Hard rules

• Never route back to orchestrate-dev when it routed here. The hand-off is one-directional by design; a mutual bounce would loop forever and is invisible until it happens in real use.
• Never open a PR on route A. An active objective means delivery is somebody else's job; a PR per task is the exact defect this split exists to remove.
• Never push directly to the default branch. Resolve it — bash skills/dev-workflow/_shared/scripts/default-branch.sh — rather than comparing against a list of conventional names; a repo whose default is trunk or dev goes unguarded by any such list. Always a feature branch.
• Never skip validation, even if the user says "just push it". Tier 1 always; Tier 2 before any push on route C.
• Never commit with git add . blindly — stage selectively so secrets and build artifacts stay out.
• Never make architectural decisions silently — a choice that changes the public API, schema, or module structure gets surfaced (use /decision) before coding.
• Never force auto-merge from here. pr-dev resolves merge policy from the repository and the user's configuration.
• Commit messages follow conventional commits (feat:, fix:, chore:, docs:, refactor:).

What NOT to do

Quick reference

Supporting files

Scope boundary

This skill routes. It does NOT merge PRs, monitor CI, or resolve post-review conflicts — that is /pr-dev.