Night Shift — rotating-brains overnight improvement loop

1. Verdict

Build both. A new companion skill (working name night-shift) that runs the rotating-brains overnight improvement loop, plus five surgical upgrades folded into recursive-10-10-product-hardening. They occupy different seams:

night-shift piggybacks rather than duplicates: it consumes coverage-matrix scores, invokes recursive-10-10 semantics for verification, and reuses /bhag's two-gate auto-merge safety model verbatim.

2. What was surveyed

Also checked and deliberately not copied: claude-flow / ruflo (enterprise swarm theater), Spec-Kit / OpenSpec / GSD (phase-gate spec tooling; the vibe-to-spec interview is /blindspot's job already).

3. The cautionary evidence

4. The steal list

5. Proposal A: the night-shift skill

flowchart TB
  A[Wake: read HOOK.md + queue] --> B{Queue has ready items?}
  B -- no --> C[Rotate to next brain in roster]
  C --> D[Brain: best-in-class research via firecrawl
+ product walk
+ findings with citations]
  D --> E[Findings → queue items
NNN-ready-pN-lens-slug.md
dedup vs ALL prior findings]
  E --> B
  B -- yes --> F[Pick highest-priority ready item]
  F --> G[Negotiate done-contract
generator ↔ evaluator]
  G --> H[Generator implements
on branch, tests green]
  H --> I[Skeptical evaluator:
browser-drive + judge vs contract]
  I -- fail --> H
  I -- pass --> J[Land: PR → checks appear AND pass
→ merge serialized vs fresh main]
  J --> K[Compound: write back learnings
note → rule → new reviewer]
  K --> L[Update HOOK.md + morning report]
  L --> B

wake → (queue empty? rotate brain: research best-in-class → findings → queue) → pick item → done-contract → implement → skeptical evaluator (loop until pass) → evidence-gated land → compound learnings → update hook → repeat

5.1 The two alternating phases

Patrol phase (divergent). When the queue has no ready items, the next brain in the roster takes the wheel for one full iteration. It (a) researches best-in-class for its lens in this product category via firecrawl, citing sources and scoping every statistic to its dataset; (b) walks the actual product (browser for UI brains, code for engineering brains, both for most); (c) emits findings as queue items with a decision-brief each: what, why, evidence, effort, the brain's recommendation, and its lens signature. Dedup runs against ALL prior findings (accepted, rejected, and open), so judge-rejected ideas do not resurrect every night.

Build phase (convergent). While ready items exist, iterations consume the queue: pick highest priority, negotiate a done-contract between generator and evaluator (Anthropic's sprint contract), implement on a branch with tests, then the skeptical evaluator browser-drives and judges against the contract. Pass lands through the evidence gate; fail loops with feedback, subject to the stuck-only backstop (3 consecutive no-progress attempts → mark blocked, move on).

5.2 The brain roster (default; per-repo overridable)

5.3 State substrate (all on disk, in-repo, git-ignored scratch where noted)

.night-shift/
  HOOK.md                 # current-step pointer: brain index, active item, phase
  roster.md               # brain roster for this repo (generated on first run, editable)
  queue/                  # NNN-{ready|building|blocked|landed|rejected}-{p1|p2|p3}-{lens}-{slug}.md
  findings-ledger.md      # every finding ever, with disposition; the dedup reference
  learnings/              # compound-phase output, YAML frontmatter, escalation status
  morning-report.html     # user-challenge decisions queued + the night's ledger (regenerated each run)
  scratch/                # ephemeral bookkeeping, git-ignored, squashed to a digest at run end

5.4 Non-negotiables carried over from house doctrine

• Backstop is stuck-detection only; success never halts the run.
• The four-check isolation gate before any browser write, verbatim from recursive-10-10 Step 0b.
• No auto-merge on bare invocation; explicit auto-merge argument + production guard, verbatim from Step 0a.
• Evidence, never assertion: checks appear AND pass; browser evidence for UI claims; "no issues found" states what was examined.
• Chain-of-command lanes: security and design judgment on Fable, volume build/research on Opus, pure sweeps on Haiku/Sonnet.
• User-Challenge decisions (anything overruling Jack's stated direction, destructive beyond git, or spending real money) are never auto-decided; they go to the morning report.

6. Proposal B: upgrades to recursive-10-10-product-hardening

7. Open design choices (Jack's call)

8. Build plan

• Author jacked/data/skills/night-shift/SKILL.md + persona files (guidance-style, superpowers:writing-skills discipline)
• Fold the five upgrades into jacked/data/skills/recursive-10-10-product-hardening/SKILL.md
• Wire install/generation (Claude + Codex parity), regen installed copies in ~/.claude/skills and ~/.agents/skills
• Tests: skill packaging/manifest tests per repo convention, full local CI mirror
• /dcr multi-lens review + /cso pass (the skill grants broad autonomy; audit the guardrails)
• Branch, PR, version bump per repo release convention

Generated with the jacked HTML artifact template.