Dream

dream — Overnight Batch Execution

Runtime syntax: invoke the plugin copy through the active runtime's owner-qualified skill selector; use a bare selector only when intentionally selecting a user-tier copy (a separately installed personal copy; Genie no longer seeds this tier). Cross-skill prose below uses bare names as portable semantic routes; the orchestrator resolves the selector for the active runtime.

Pick SHIP-ready wishes, build a dependency-ordered plan, dispatch one worker subagent per wish, review PRs, merge to dev, run the QA loop, and write a wake-up report. The dream orchestrator dispatches — it never executes wish work directly.

This is a high-impact, explicit-only workflow. The user must approve the selected wishes, the generated plan, PR creation, and merge-to-dev authority. Never merge to main or master, deploy, send external messages, or expand scope without separate authority.

When to Use

• Human wants to queue multiple wishes for autonomous overnight execution
• Multiple WISH.md files have persisted status APPROVED

Flow

1. Pick wishes (Picker below); human confirms the selection.
2. Generate .genie/DREAM.md — dependency-ordered plan; human may edit before the run.
3. Phase 1 — Execute: dispatch workers layer by layer, collect outcomes.
4. Phase 2 — Review + PR: review every PR, fix valid gaps, CI green.
5. Phase 3 — Merge + QA: merge to dev in order, QA loop until criteria proven.
6. Phase 4 — Report: write .genie/DREAM-REPORT.md, the wake-up artifact.

Picker

1. Read .genie/wishes/*/WISH.md and select only wishes whose Status field is exactly APPROVED. The brainstorm jar is historical/discovery context, never readiness authority. A Poured entry without an existing approved WISH.md is skipped and reported as drift. No matches → print No APPROVED wishes found under .genie/wishes/ and stop.
2. List matches numbered by slug: 1. <slug> — <one-line description>.
3. Human picks by number (1 3 5) or all.

DREAM.md

1. Read the wish-level **depends-on:** value from each selected WISH.md's ## Dependencies section (none means no edge).
2. Topologically sort into merge_order layers 1..N — layer 1 has no selected dependencies; same-layer wishes are parallel.
3. Per-wish entry: slug, branch: feat/<slug>, wish-path: .genie/wishes/<slug>/WISH.md, depends-on, merge-order. Keep the canonical hyphenated keys so the plan can be checked directly against each wish.
4. Write .genie/DREAM.md in the shared worktree; present for human confirmation before executing.

Phase 1: Execute

For each merge_order layer, in order:

• Spawn one worker subagent per wish via the native delegation surface — all of the layer's spawns in ONE message so they run in parallel (background; each notifies you with its final message).
• Every brief carries the Worker Contract below plus curated wish context (goal, groups, acceptance criteria, validation commands — see work § Context Curation).
• Follow-ups to a running worker go through native follow-up messaging; completion is push (the final-message notification), never a sleep-poll.
• Inspect state on demand: genie board --wish <slug> / genie task list --wish <slug>. If a wish has no task rows, drive it off WISH.md directly — task tracking is an enhancement, never a blocker.
• The layer is done when every worker has reported; then dispatch the next layer.

Worker Contract

Each worker, independently:

1. Work in a dedicated branch and worktree for feat/<slug>, using runtime-managed or ordinary Git worktrees according to the active environment. The contract governing parallel writers and repo-level git state is stated once in AGENTS.md and the work skill's Dispatch section — follow it there.
2. Execute the wish per work (its dispatch, review-gate, and task-state rules govern): dispatched engineers claim via genie task checkout; the worker leaves task state in_progress and reports evidence. Only the dream PM/orchestrator runs genie task done after clean review and passing validation.
3. Run review per group against acceptance criteria.
4. Run CI; on failure fix and retry (max 3 attempts; poll CI status, never sleep-loop). After 3 failures → blocked.
5. Only after CI green and authorized PR creation: create a PR targeting dev, preferring the GitHub connector.
6. Final message is the completion signal, every claim audited against tool output:
   • done — PR <url>, CI green, groups N/N
   • blocked — <reason>, groups N/N

Phase 2: Review + PR

Trigger: all workers in the layer reported done or blocked.

1. Dispatch one reviewer subagent per PR via the native delegation surface (reviewer ≠ worker) to run review against the wish's acceptance criteria.
2. Read bot comments critically — never blindly accept automated findings.
3. On FIX-FIRST: diagnose first; return an overdesigned plan to wish/design review, otherwise dispatch fix for valid gaps (max 2 loops per PR). On another architectural issue: escalate in the report, no fix attempt.
4. CI must be green before proceeding — poll status, do not sleep.
5. On SHIP: mark the PR review-complete.

Phase 3: Merge + QA

Trigger: all PRs marked SHIP.

1. After explicit merge authorization, merge PRs to dev in merge_order; never merge to main or master.
2. Dispatch a qa subagent on dev to test against each wish's QA criteria.
3. Each failure: report → trace → fix → retest. Every fix is a new PR through review and merge.
4. Continue until all criteria are proven or blocked.

Phase 4: Report

Write .genie/DREAM-REPORT.md — always, even if every wish blocked:

# Dream Report — <date>

## Per-Wish Status
| merge_order | slug | PR | CI | Review | Merged | QA |
|-------------|------|----|----|--------|--------|----|

## Blocked Wishes
- `<slug>`: <blocking reason>

## QA Findings
- `<slug>`: <criterion failed — root cause, fix PR>

## Follow-ups
- <items requiring human intervention>

Grounded Progress

The report is an audit, not a recollection. Every cell traces to tool output from the run: PR URLs, CI results, review verdicts, genie task list --wish <slug> state, worker final messages. State per wish exactly what is verified, what failed, and what was skipped. Never report a wish shipped until its merge and QA evidence is in hand — dispatched is not done.

Rules

• Never early-stop: a blocked wish is recorded and the remaining wishes continue.
• Never skip Phase 2 or Phase 3 — every PR is reviewed, every merge is QA-tested against wish criteria.
• The orchestrator never executes wish work — always dispatch worker subagents.
• No scope beyond what each WISH.md defines.
• Poll CI status — never sleep in retry loops.