Phase 5 — Permanent Loop Install Plan

Mission Control authored. Sandbox restrictions prevent direct writes to ~/.claude/ and launchctl from inside this session. Phase 5 outputs are staged for Raychard to run from Terminal.app on wake.

What “permanent” means

The hackathon system being built IS the deliverable. Phase 5 freezes the parts that should run forever (or until v2 of the hackathon supersedes them).

4 install streams

Stream	What it does	Cadence	Status
A. Identity patches	Apply 7 × `identity-update-patch.md` to live cofounder identity files	One-time (gated on Raychard review)	🟡 staged, awaiting approval
B. Skill promotion	Move 28 codified skills (7 × 4) into `~/.claude/skills/` namespace	One-time per skill	🟡 staged
C. Cadence loops	launchd jobs for daily reflection + weekly retro + monthly synthesis-v2	Continuous	🟡 plist files staged
D. Knowledge pipeline	Continue atomization → linking → sublimation as new cofounder work arrives	Continuous (existing rhcf-* daemons)	🟢 running

Stream A — Identity patches (review-gated)

Each cofounder produced an identity-update-patch at outputs/phase4/{role}/identity-update-patch.md. They are non-breaking, additive. None auto-applies.

Raychard wake-time action (~45 min total, 7 × ~6 min review each):

Read each outputs/phase4/{role}/identity-update-patch.md
Approve / amend / reject per role
Apply approved patches to /Users/raychard/.claude/agents/co-founders/{role}/identity.md

Recommended order (lowest-risk first): CGO → CFO → CPO → CTO → CSO → CEO → CAIO (CAIO last because its patch is gated on first held-out eval at T+45 per CAIO orchestrator-flag).

Stream B — Skill promotion

28 skills produced under outputs/phase4/{role}/skills/. Target paths:

~/.claude/skills/{role}-{skill-name}/SKILL.md (one per skill)

Each skill file already has Claude Code skill frontmatter. Promotion = copy or symlink.

Recommended approach: symlink from outputs/phase4 into ~/.claude/skills so iterations to the corpus auto-reflect. Sandbox blocks ln against ~/.claude/ from inside session, so Raychard runs the staged shell script outputs/phase5/scripts/promote-skills.sh in Terminal.

Stream C — Cadence loops (launchd)

Three launchd jobs staged at outputs/phase5/launchd/:

com.rh.cofounders.daily-reflection.plist — fires 23:30 daily. Runs cofounder-daily.sh which:
- Reads day’s audit/cozempic events
- For each cofounder, asks: “what fired today? what should have but didn’t?”
- Appends to ~/Dev/RH-CoFounders/journal/daily/{YYYY-MM-DD}.md
- No external network. Local Claude session via claude --bg.
com.rh.cofounders.weekly-retro.plist — fires Sunday 22:00. Runs cofounder-weekly.sh which:
- Each cofounder writes 200-word status (what compounded, what didn’t)
- CEO synthesizes into ~/Dev/RH-CoFounders/journal/weekly/{YYYY-WW}.md
- Pairing suggestions for next week’s cross-pollination
com.rh.cofounders.monthly-synthesis-v2.plist — fires 1st of month 02:00. Runs cofounder-monthly.sh which:
- Each cofounder reads journal/daily for past month
- Produces synthesis-v{N+1}.md if material drift
- Otherwise notes “no drift, synthesis-v1 holds”

Install command (Raychard runs in Terminal):

cp ~/Dev/RH-CoFounders/briefs/2026-05-16-36hr-cofounder-hackathon/outputs/phase5/launchd/*.plist ~/Library/LaunchAgents/
launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.rh.cofounders.daily-reflection.plist
launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.rh.cofounders.weekly-retro.plist
launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.rh.cofounders.monthly-synthesis-v2.plist

Stream D — Knowledge pipeline (already running)

The rhcf-* daemons (extractor, curator, evolver, federator) already do atomization → linking → sublimation. Hackathon corpus is now indexed by:

7 × corpus/{role}/atoms/atoms.jsonl (~370 atoms total, ~50 cross-cofounder links)
audit/hackathon.jsonl (event stream)

One-time backfill action: feed atoms.jsonl into rhcf.db. Script staged at outputs/phase5/scripts/backfill-atoms-to-sqlite.sh.

Cross-mesh promotion (CEO synthesis §5 recommendation)

Once Stream A applied, each cofounder can [[cite]] peer atoms. Implementation:

atoms.jsonl links field already populated with cross-cofounder IDs
MCP rh-cofounders server already exposes search_atoms — verify retrieval-bridge health (currently DEGRADED per session start)

Retrieval-bridge fix (per CTO Phase-2 maestros-self-application D1): investigate mcp_not_registered errors in ~/.claude/audit/retrieval-failures.jsonl. This is the prerequisite for cross-mesh promotion to work end-to-end. Mark as blocker in handoff.

What is explicitly NOT being installed

Per debate-5 master verdict (MaestrOS internal-only), the following are deferred:

Public corpus publication (gated on T+45 held-out eval per CAIO)
SaaS productization infrastructure
External growth loop (CGO accepted 12-24mo patience)

Per CPO stop-the-line skill: until external persona + reference customer + opportunity assessment exist, productization features are blocked.