Roadmap: thin-PAI implementation phases (P0 → P8)

[eccentricnode]

Plan

Full strategy in STRATEGY.md. Divergence catalog in DIVERGENCE.md. Upstream sync procedure in .github/UPSTREAM_SYNC.md.

Phases

• P0 — Fork + meta — strategy, README, divergence catalog, upstream-sync procedure (commit e7bb47e)
• P1 — Audit catalog — every upstream file classified KEEP / REBUILD / DELETE in a machine-readable manifest (MANIFEST.yaml at repo root). First Ralph mission.
• P2 — Delete pass — remove all DELETE-class files; verify a normal CC session boots clean. Ralph mission.
• P3 — Native rebuild — convert REBUILD-class to native CC primitives (hooks → native handler types, memory → auto-memory, security → native permissions). Target: hook count ≤ 7, startup context ≤ 14K tok. Ralph mission.
• P4 — CLAUDE.md surgery — trim to <100 lines; move PROJECTS/TELOS to .claude/rules/ with paths: frontmatter.
• P5 — MEMORY.md restructure — split into index + topic files following CC's auto-memory pattern; ensure nothing past 200-line cap is silently dropped.
• P6 — Mode classifier reliability — resolve 25s-timeout failure mode (see STRATEGY.md § Open Questions #1).
• P7 — Dogfood — daily-driver for 1 week, log issues, zero P0/P1 bugs after 7 days.
• P8 — v0.1 release — tag, release notes, publish.

Success criteria (overall)

Metric	Current PAI 5.0	pai-thin target
Hook count	25+	≤ 7
Startup context tokens	~22.7K	≤ 14K
@-imports at CLAUDE.md root	5	2
Parallel storage systems	2 (MEMORY/* + auto-memory)	1 (auto-memory only)
MEMORY.md lines (index only)	262 (over CC cap)	≤ 150

Open questions deferred to specific phases

1. Mode classifier rewrite — deferred to P6 (needs telemetry first)
2. Pulse retention shape — deferred to P3
3. Path-scoped rule portability — deferred to P4
4. pai-lite skill consumption via pi install — deferred to P2/P3
5. Upstream-pull mechanism formalization — P0 deliverable, done (see .github/UPSTREAM_SYNC.md)

Anti-scope

See STRATEGY.md § Anti-scope. This issue does NOT track:

• New PAI ideas (those go upstream or to a separate experiment repo)
• General CC improvements (file with Anthropic)
• pi.dev substrate work (pai-lite)

Next action

Spin up the P1 Ralph mission (Audit catalog). High parallelism, read-only, low blast radius — best first run.