Build Inline

Ephemera paths: resolve FEATURE_CODE / <CODE> then follow templates/skills-ephemera-paths.md (feature root .skills/<CODE>/). Resolve pack seeds in this order, first path that exists: (1) templates/ beside this SKILL.md, (2) ${CLAUDE_PLUGIN_ROOT}/templates when that variable is set, (3) ../../../templates relative to this SKILL.md.

Build Inline

Drive an approved plan to completion yourself, task by task, with no implementer (or reviewer) subagent dispatches. Every step is test-first. Progress survives in the ledger. Finish with whole-branch review.

This is not subagent orchestration. Continuous multi-task subagent waves are build-in-waves. Human-gated review units with subagents are build-by-story.

When this skill fires: the user chose the inline route, the environment cannot (or must not) fan out implementers, or plan-tasks offered inline execution and they accepted it. Tool availability does not override the route — if they asked for inline, stay inline even when subagent tools exist.

Narration: at most one short line between tool calls. Ledger + tool results carry the record.

Shared controller recipe: REQUIRED SUB-SKILL: use execute-common. Load ../execute-common/SKILL.md when Setup preflight / ledger / todos or After the last task starts. That file is the one home for those steps and for the close-sequence predicates. This file owns the inline iron law and the controller-as-implementer loop.

The Iron Law

NO IMPLEMENTER SUBAGENT — YOU ARE THE IMPLEMENTER
NO PRODUCTION CODE WITHOUT TDD
NO GUESSING THROUGH A BLOCKER

Dispatching an implementer "because tools exist" or "subagents are faster" violates this skill. Skipping RED because the change is "one line" violates this skill. Inventing an API, path, or acceptance detail when the plan is ambiguous violates this skill — stop and ask.

Scope vs Execution-mode

Invoking this skill selects the inline route. The Execution-mode header is bookkeeping for sibling skills and later handoffs:

Human-gated review units with subagents are build-by-story. Inline already keeps the human in the conversation turn-by-turn.

Setup

1. Route gate. Confirm inline is the intended path (user said so, or no subagent capability). If they want subagent continuous waves → build-in-waves. If they want story-unit barriers → build-by-story. Done when: route is inline.
2. Header bookkeeping. Parse Execution-mode:. If missing/unset/invalid: write Execution-mode: continuous into tasks.md. If already set, leave it. Done when: header is present.
3. Session preflight. Apply ../execute-common/SKILL.md Session preflight. Done when: that section's Done when holds.
4. Runtime binding. Apply ../execute-common/SKILL.md Runtime binding and lease preflight. Inline execution has no worker lease, but it records the active harness/provider/model, pricing policy, context reserve, and effective concurrency in .skills/<CODE>/execution-session.json. Done when: the runtime snapshot exists before the first production edit.
5. Ledger check. Apply ../execute-common/SKILL.md Ledger check. Done when: next task is known.
6. Read the plan. Read tasks.md once. Record the canonical Global Constraints source path and content hash (verify commands live here if docs/agents/project.md is missing — say so and suggest configure-repo). Done when: the source reference and task delta are in hand.
7. Todos — GATE. Apply ../execute-common/SKILL.md Todos — GATE. Done when: the list mirrors the plan and includes the Close branch todo.
8. Pre-flight plan review. One batch question for plan-internal defects before coding. Clean scan → no comment. Done when: conflicts ruled or none.
9. Order. Depends-on topo order; absent Depends-on → every earlier task; none → no prereq. Serial only — never parallel waves / multi-worktree fan-out in this skill. Done when: task sequence is fixed.

Per-task loop

For each Task N in order:

1. Record base. BASE=$(git rev-parse HEAD).
2. Build the task capsule (for yourself). Record Task N, the canonical Global Constraints source path/hash, and the task delta in .skills/<CODE>/task-N-brief.md. That file is the contract you implement against — exact values, paths, signatures, _Requirements: IDs. Include relevant **ARCH-N** when a docs/architecture/ spine exists. WHEN preflight recorded ticket IDs, list them in the brief. Done when: brief exists and you have read it.
3. Clarify first. If anything in the brief is ambiguous — API shape, path, acceptance, dependency — stop and ask now. Do not start RED until clear.
4. Implement with TDD. REQUIRED SUB-SKILL: use test-first for every step in the task. Tests describe domain behavior; map requirement IDs in the report/self-check, not by embedding IDs in source. Work only the files the plan names.
5. Deviations. WHEN territory forces you off the brief: you are the implementer — follow the Deviations recipe in ../execute-common/implementer-prompt.md (nine-field entry under .skills/<CODE>/implementation-notes.md: Unknown class, Map said, Territory showed, Map impact, etc.). Log before finishing the task; append only. IF Map impact is reroute-plan / realign-spec or the fix falsifies the plan / shared contract → stop and REQUIRED SUB-SKILL: use reroute-plan (or ask); do not stretch silently.
6. Commit. Conventional subject explaining the change — no Implements: / Guards: trailer required.
7. Render check. You are the implementer — WHEN the diff touches anything a browser renders, apply the Render check step in ../execute-common/implementer-prompt.md (screenshot to .skills/<CODE>/task-N-render.png, judge against the brief, fix, re-shoot); the Visual check line lands in the step-9 evidence bundle.
8. Self-check (controller, not a subagent). Re-read the brief against the diff: every requirement ID's behavior covered? TDD evidence (RED then GREEN) real? Output pristine? Plan File Structure respected? Fix gaps now — you have no separate task-reviewer pass.
9. Optional evidence bundle. For non-trivial tasks, write .skills/<CODE>/task-N-report.md with what changed, IDs covered, RED/GREEN commands+output, the Visual check line when the render check ran, concerns, deviations path or none. Cheap tasks may skip the file if the commit message and ledger carry enough; when in doubt, write it.
10. Ledger. Append Task N: complete (commits <base7>..<head7>, inline, review self). Mark the todo done.
11. Next. Immediately continue to the next task — no permission pause, no unit barrier. On blocker mid-task: stop the loop and ask (see Stop conditions).

Stop conditions

Stop and ask the user (do not guess, do not dispatch a subagent to invent an answer) when:

• The brief or plan is ambiguous or contradictory
• Verification fails and the fix is not obvious from the brief
• A deviation would falsify plan/design/requirements
• You would need to touch files or contracts outside the task
• main/master consent is missing and the workspace is main/master

After the user answers, resume the same task from the brief — do not skip ledger discipline.

Durable progress

• Start: read .skills/<CODE>/progress.md; trust it and git log over memory.
• Never re-do a task the ledger marks complete.
• After compaction, resume at the first task without a complete line.
• .skills/ is git-ignored; if wiped, reconstruct from git log.

After the last task

Apply ../execute-common/SKILL.md Close sequence in full. You are the fixer under test-first (still no implementer subagent unless the user explicitly lifts the inline route). Point inspect-change at any Minors or notes in the ledger / implementation-notes.md.

Done when: that section's Done when holds.

Red Flags — Never

• Skip the tracker-sync, occupancy, or workspace preflight
• Invent a tracker or ticket set when config is absent or the user declined sync
• Dispatch an implementer or task-reviewer subagent while on this skill
• Skip test-first / write production code before a failing test
• Guess through ambiguity, a red suite, or a plan gap
• Run unit barriers or human unit stops (that is build-by-story)
• Parallel-wave / multi-worktree fan-out
• Re-do ledger-complete tasks after compaction
• Start on main/master without explicit consent
• Create a worktree without asking
• Claim a task complete without a ledger line
• Skip the close sequence, silent-skip polish, or treat EOD/demo as a polish predicate
• Start Task 1 before the todo list exists (tasks and Close branch)
• Hand off mid-plan to build-in-waves subagent loop without an explicit user route change