Orchestrate Saga

---
description: Generate strict <3900-char goal-orchestrator prompt for a saga and lock via /goal
argument-hint: <sagaId> [--docs slug1,slug2] [--decisions D001,D002] [free-text intent]
allowed-tools: Bash, Read, Write, SlashCommand, Task, AskUserQuestion
model: opus
---

# Orchestrate Saga

Generate a strict, structured Prime-Orchestrator goal prompt for the saga in `$ARGUMENTS` and lock it via `/goal`. Generated prompt MUST be ≤3900 characters. Target 3700 for safety.

## Argument parsing

First token = saga ID (e.g. `T10176`). Optional flags:

- `--docs <comma-list>` — research doc slugs to include
- `--decisions <comma-list>` — memory decision IDs to include
- Anything else = free-text intent appended to the generated prompt's final line

If `--docs` omitted, auto-discover via `cleo docs find "<sagaTitle>"`. If `--decisions` omitted, auto-discover via `cleo memory find "<sagaTitle>" --type decision`.

## Execute in order

### Step 1 — Gather saga context (read-only, batched in ONE bash block)

```bash
SAGA="<sagaId from $ARGUMENTS>"
cleo show $SAGA 2>&1 | python3 -c "import json,sys; d=json.load(sys.stdin)['data']; t=d['task']; print(f\"TITLE={t['title']}\"); print(f\"AC_COUNT={len(t.get('acceptance',[]))}\"); print(f\"STATUS={t['status']}\")"
cleo saga rollup $SAGA 2>&1 | python3 -c "import json,sys; d=json.load(sys.stdin)['data']; print(f\"TOTAL={d['total']} DONE={d['done']} PCT={d['completionPct']}\")"
# Try saga groups relation first; fall back to parent-edge enumeration if empty.
MEMBERS=$(cleo saga members $SAGA 2>&1 | python3 -c "import json,sys; d=json.load(sys.stdin)['data']
ms = d.get('members', d.get('items', []))
out = []
for m in ms:
    if isinstance(m, dict) and m.get('id'):
        out.append(f\"EPIC {m['id']} {(m.get('title') or '')[:60]}\")
print('\n'.join(out))")
if [ -z "$MEMBERS" ]; then
  # Fallback: enumerate child Epics via parent edge.
  # T9658 wired `cleo list --parent <saga>` through saga.groups in packages/core/src/tasks/list.ts,
  # so this branch + the primary `cleo saga members` cover every shipped saga shape.
  MEMBERS=$(cleo list --parent $SAGA --limit 50 2>&1 | python3 -c "import json,sys; d=json.load(sys.stdin)['data']
for t in d['tasks']:
    if t.get('type') == 'epic':
        rollup = t.get('childRollup', {})
        nt = rollup.get('total', 0) if isinstance(rollup, dict) else 0
        print(f\"EPIC {t['id']} {t['title'][:55]} ({nt}t)\")")
fi
if [ -z "$MEMBERS" ]; then
  echo "WARN: could not auto-enumerate Epic members for $SAGA. Ask user to supply --epics T9978,T9979,..."
  exit 2
fi
echo "$MEMBERS"

For each --docs <slug>: cleo docs fetch <slug> — capture description ONLY (not full body — too large for the generated prompt; we reference by slug).

For each --decisions <id>: cleo memory fetch <id> — capture decision text 1-line summary.

If auto-discover: cleo docs find "$SAGA" + cleo memory find "$SAGA_TITLE" --type decision. Pick top-1 of each.

Step 2 — Compose the generated prompt

Emit the prompt below substituting <placeholders> from Step 1. Keep epic list to ID + 60-char title. Cap at 10 epics displayed; if more, say "+ N more — see cleo saga members".

SAGA <sagaId> <short-title> (<epicCount>E/<taskCount>T): ship 100% + install + dogfood.

ROLE: Prime Orchestrator. NEVER solo. TeamCreate per active Epic (lead + 3-5 workers). AskUserQuestion ONLY when blocked, 3 ranked researched options — never make user dig.

SESSION START:
1. cleo briefing
2. cleo show <sagaId> ; cleo saga rollup <sagaId>
3. cleo docs fetch <primary-research-slug>
4. cleo memory fetch <primary-decision-id>

EPICS (dep order):
<auto: one line per Epic "Eid — title (Nt)">

EXEC: TeamCreate continuous loop, never idle.
- Pull: `cleo orchestrate ready --epic <Eid>`. Spawn: `cleo orchestrate spawn <Tid>`.
- After every merge, re-pull next ready. Loop until rollup = N/N.
- Cross-team handoff via SendMessage on wave-block clear.

WORKTREE (NON-NEGOTIABLE):
- Orchestrator + teams stay on local main.
- Worker: `git worktree add -b feat/<Tid>-<slug> ~/.local/share/cleo/worktrees/<projectHash>/<Tid> origin/main` (projectHash via @cleocode/paths — NEVER literal name, NEVER /mnt/projects/* siblings, NEVER .claude/worktrees/).
- Worker's FIRST action: `cd <worktree>` BEFORE any Edit/Write/Bash.

PER-TASK FLOW (one PR/Task):
worktree → implement → `pnpm biome check --write . && pnpm run build && pnpm exec vitest run <scope>` → `.changeset/<Tid>.md` → conventional commit (Closes <Tid>; Saga: <sagaId>; Decision: <Did>) → push → `gh pr create` → `gh pr checks <num> --watch` → `gh pr merge --merge --delete-branch` → `cleo verify <Tid> --gate {implemented,testsPassed,qaPassed} --evidence "pr:<num>" --shared-evidence` → `cleo complete <Tid>` → `git worktree remove` → re-pull.

ZERO-TOLERANCE:
- SOLID + DRY + boundary per ADR-078 (slug adr-078-boundary-registry); TOOLS-in-CORE = crates/<X>-core → <X>-napi → packages/<X> thin → @cleocode/core → @cleocode/cleo (dispatch-only).
- DELETE Rust↔TS dupes; no compat shims unless ADR amendment.
- File bugs vs EXISTING epics first (`cleo find "<symptom>" --status pending`); new epic only if no fit.
- NO admin-merge unless flake is documented saga-precedent (T9407/T9775). Diagnose, fix, re-push.
- Hierarchy (ADR-073): Saga→Epic→Task→Subtask. Task MUST fit one context window else split to Subtasks.

FINAL EPIC POST-MERGE:
1. `cleo release plan v<calver> --epic <sagaId>`
2. `cleo release open v<calver>` → release PR → merge → tag → npm publish
3. `npm i -g @cleocode/cleo@<calver>` → `cleo --version` confirms
4. DOGFOOD: `cleo orchestrate spawn <real-task>` → p50<5s, canonical projectHash path, zero new orphans
5. `cleo docs add closure-report` + `cleo memory observe`

NEVER STOP UNTIL: `cleo saga rollup <sagaId>` = N/N + release installed + dogfood green.
<free-text intent if any>

Step 3 — Self-check character budget

GENERATED_PROMPT_FILE=/tmp/orchestrate-saga-prompt-${SAGA}.txt
echo "<composed prompt>" >