jayden-dang 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 By Story
Drive an approved story-unit plan to completion: derive review units from requirement stories, run the shared continuous scheduler inside each unit, stop for a human after each unit, resume from the unit ledger, and finish with whole-branch review.
Continuous multi-task orchestration without unit barriers is build-in-waves.
Controller-implements-without-subagents is build-inline. Invoking this skill
selects story-unit execution; continuous-header handoff is in Mode
ownership below. If the user wants no subagents, name build-inline.
Shared controller recipe: REQUIRED SUB-SKILL: use execute-common — the
one home for preflight/ledger/todos/close, applied where each Setup step below
says. This file owns story-unit barriers and unit derivation; load
../execute-common/task-lifecycle.md for task dispatch/review and use the
continuous scheduler rules from build-in-waves inside each unit.
Narration: at most one short line between tool calls. Ledger + tool results carry the record.
The Iron Law
NO NEXT UNIT WITHOUT HUMAN UNLOCK (OR A WRITTEN MODE CHANGE)
Agent-clean is not human unlock. EOD, demo pressure, and "whole PR later" do
not drop the barrier. "Stop stopping" is a mode change: write
Execution-mode: continuous into tasks.md before dropping barriers.
Iron laws (story-unit)
- Review units are derived, never authored in
tasks.md(ignore PM "Human review order" comments for partition). - After each unit's tasks are clean: unit agent review → STOP for human → unlock.
- "continue" after looking = next unit only. "stop stopping" /
"just run it all" = write
Execution-mode: continuousintotasks.mdfirst, then hand remaining work tobuild-in-waves. Chat-only is not a mode change. - Ledger
Unit <k>: complete (tasks …, range <base>..<head>)— resume reads it. - Whole-branch agent review still runs after the last unit. Human unit reviews are not a substitute.
| Thought | Reality |
|---|---|
| "User said keep going — no need to edit tasks.md" | "Stop stopping" is a mode change; write the header or the plan lies |
| "Unit agent review was clean — skip human barrier" | Agent unit review is not the human stop |
| "EOD / whole PR later — drop unit stops" | Time and later review do not rewrite Execution-mode |
| "continue means run everything left" | "continue" unlocks one next unit; run-it-all needs mode-change language |
| "PM review waves are clearer than stories" | Partition is derived from requirement story IDs, not authored lists |
Mode ownership
Align tasks.md to the story-unit route and continue:
Execution-mode: on tasks.md |
Action |
|---|---|
story-unit |
Proceed |
continuous |
REQUIRED SUB-SKILL: use build-in-waves |
missing / unset / invalid |
Write Execution-mode: story-unit into tasks.md (commit if tracked). Proceed |
Setup
- Mode ownership. Apply the table above. Done when: header is
story-unitand you stay on this skill, or you have handed off tobuild-in-waves. - Session preflight. Apply
../execute-common/SKILL.mdSession preflight. Done when: that section's Done when holds. - Ledger check. Apply
../execute-common/SKILL.mdLedger check (its story-unit clause covers skipping units already ledgered complete). Done when: next task/unit is known. - Read the plan. Read
tasks.mdonce. Record the canonical Global Constraints path and content hash; dispatches reference and take verify commands from it instead of pasting it into every reviewer prompt. Ifdocs/agents/project.mdis missing, say so and suggestconfigure-repo. When## Teamhas roster/band, load band packaging only — never skip dual-verdict review for Solo. Done when: constraints captured word-for-word. - Derive units — GATE. Load
story-unit-mode.mdbeside this file. Run Derive partition + File count + print the Unit table. Hard-fail blocks dispatch. Done when: table printed; hard-fails resolved or reported. - Todos — GATE. Apply
../execute-common/SKILL.mdTodos — GATE. Done when: the list mirrors the plan and includes the Close branch todo. - Pre-flight plan review. One batch question for plan-internal defects before dispatch. Clean scan → no comment. Done when: conflicts ruled or none.
- Unit order. Fixed already by step 5's Derive partition. Inside each unit, use the shared ready-set scheduler: serial when only one task is ready or surfaces overlap; parallel only for disjoint surfaces with safe worktree isolation. Done when: concurrency is fixed per unit.
Per-unit loop
For each unit U in order:
A. Tasks in the unit
Run ../execute-common/task-lifecycle.md for every ready task in U. The
scheduler supplies the lane, worker/reviewer lease IDs, base revision, brief,
report, and diff-package paths. A clean task review is ledgered before the
unit barrier; that file's single-task close-out and no-duplicate-reviewer rule
governs whether B below still needs to dispatch.
B. Unit barrier (after every task in U is ledgered)
Load story-unit-mode.md Per-unit barrier and run it in full. The human-facing
STOP message MUST fill every REQUIRED slot in that recipe — no freeform
abbreviation under time pressure. A multi-task unit's one unit-scope dispatch
(from A) is its synthesis; a single-task unit already closed in A and is not
re-reviewed here.
Unlock:
| User says | Action |
|---|---|
| continue / LGTM / looks good (after looking) | Ledger Unit <k>: complete (tasks …, range <base>..<head>). Next unit only. Mode stays story-unit. |
| stop stopping / just run it all | Write Execution-mode: continuous into tasks.md (commit if tracked). Ledger this unit complete. REQUIRED SUB-SKILL: use build-in-waves for remaining tasks (no further unit barriers). |
| stop / hold / changes needed | Stay stopped; do not start the next unit. |
C. Next unit
Resume the per-unit loop. After the last unit's unlock and ledger line → After the last unit.
Implementer status handling
| Status | Your move |
|---|---|
| DONE | Package diff → task review. |
| DONE_WITH_CONCERNS | Read concerns and .skills/<CODE>/implementation-notes.md. Entries must match the nine-field Deviations recipe in ../execute-common/implementer-prompt.md. Missing notes path, incomplete fields, or five-field-only while claiming deviation → incomplete — re-dispatch to log first. Map impact reroute-plan / realign-spec or plan-falsifying → REQUIRED SUB-SKILL: use reroute-plan (do not accept as clean DONE). |
| NEEDS_CONTEXT | Supply what was named; re-dispatch same model. |
| BLOCKED | Context → supply; ceiling → stronger model; too large → split; plan wrong → REQUIRED SUB-SKILL: use reroute-plan. |
Cap redispatches at 2 per task; still not DONE → escalate to the user.
Model tiering
State the model explicitly on every dispatch.
- Cheap: transcription / single-file mechanical fixes.
- Mid floor: every reviewer; every implementer working from prose.
- Top: design judgment, broad codebase understanding, whole-branch review.
Reviewer-prompt hygiene
- Never pre-judge findings in the dispatch ("do not flag X", "Minor at most").
- A plan-mandated defect is still a finding — ask the user which governs.
- No open-ended "check everything" directives without a concrete reason.
- Do not re-run tests the implementer already evidenced in the report.
Durable progress
- Never re-dispatch a task or re-open a unit the ledger marks complete.
- After compaction, resume at the first task without a complete line; unit complete lines mean the human barrier already closed for that unit.
.skills/is git-ignored scratch; if wiped, reconstruct fromgit log.
After the last unit
Apply ../execute-common/SKILL.md Close sequence in full. Inspect base is
git merge-base main HEAD — never a last-unit-only range.
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
- Start the next unit (or ledger unit complete as if unlocked) without human unlock or a written mode change
- Treat "stop stopping" as chat-only without writing
Execution-mode: continuous - Over-read "continue" as run-all-remaining-units
- Author or prefer PM review-order lists over derived story units
- Skip unit agent review and send the raw unit to the human first
- Skip whole-branch review because every unit was human-approved
- Use
HEAD~1as a review base - Hand a subagent the whole plan file
- Move on with open Critical/Important findings
- Fix reviewer findings in the controller context
- Re-dispatch work the ledger marks complete
- Dispatch before the todo list exists (tasks and Close branch)
- Skip the close sequence, silent-skip polish, or treat EOD/demo as a polish predicate
- Implement on main/master without explicit consent
- Create a worktree without asking
- Silently switch to continuous rules while the header still says story-unit