R76.P2 Packet — δ Phase 1.5 Graduation Plan

Execution plan for the 5-step chain. Implementation step (4) writes ONE file (plus two small edits); the 10 prompts inside it are listed here as one-liner summaries so the shape of the deliverable is reviewable before it’s written.

1. Deliverable tree

docs/
├── audits/r76-p2-delta-15-planning-audit.md              (step 1 — written)
├── contracts/r76-p2-delta-15-planning-contract.md        (step 2 — written)
├── packets/r76-p2-delta-15-planning-packet.md            (step 3 — this file)
├── guides/implementation/task-prompts/
│   ├── p1.5-delta-router-graduation.md                   (step 4 — NEW, main deliverable)
│   └── index.md                                          (step 4 — EDIT)
├── architecture/decisions/
│   └── ADR-005-multi-model-defer.md                      (step 4 — EDIT, 1 line)
└── verification/r76-p2-delta-15-planning-verification.md (step 5 — pending)

No src/ tree. No test tree. No concept docs. No roadmap.

2. The 10 Phase 1.5 sub-task one-liners

ID	Title	Effort	Depends on	Files to create	Files to modify
P1.5.1	Real 7-dim intent scoring	M	R91+ κ Rule Engine (weights); P1.5.9 (candidate table schema)	—	`src/domains/router/scoring.ts` (body only; signature frozen)
P1.5.2	Adapter: Kimi K2	M	P1.5.1; `src/domains/integrations/claude.ts` (reference shape)	`src/domains/router/adapters/kimi.ts`, `src/__tests__/domains/router/adapters/kimi.test.ts`	`src/domains/router/index.ts` (re-export)
P1.5.3	Adapter: Codex	M	P1.5.1	`src/domains/router/adapters/codex.ts`, `src/__tests__/domains/router/adapters/codex.test.ts`	`src/domains/router/index.ts`
P1.5.4	Adapter: OpenAI (GPT-4o family)	M	P1.5.1	`src/domains/router/adapters/openai.ts`, `src/__tests__/domains/router/adapters/openai.test.ts`	`src/domains/router/index.ts`
P1.5.5	N-member fallback + circuit breaker	L	P1.5.2, P1.5.3, P1.5.4	`src/domains/router/circuit.ts` (in-memory CB state), `src/__tests__/domains/router/circuit.test.ts`	`src/domains/router/fallback.ts` (body; `ROUTER_PHASE_0_SHAPE` flip); `src/__tests__/domains/router/fallback.test.ts` (drop single-member assertions, add cascade + CB coverage)
P1.5.6	Cost accounting	M	P1.5.5	`src/domains/router/cost.ts` (per-call → USD + aggregates), `src/__tests__/domains/router/cost.test.ts`	`src/domains/router/fallback.ts` (emit cost + tokens in `RouteResult`)
P1.5.7	`router_*` MCP tools	M	P1.5.6	`src/domains/router/tools.ts` (4 tools: `router_score`, `router_call`, `router_fallback`, `router_stats`), `src/__tests__/domains/router/tools.test.ts`	`src/server.ts` (register the 4 tools); `src/domains/router/index.ts` (re-export)
P1.5.8	Cross-model parity test suite	L	P1.5.7	`src/__tests__/domains/router/parity.test.ts` (contract suite across all adapters)	—
P1.5.9	Model candidates table population	S	P1.5.1	`src/db/migrations/NNN-model-candidates-seed.sql` (7 additional rows)	`src/db/schema.sql` (candidate-table columns if not already present)
P1.5.10	ζ decision-trail integration	M	P1.5.7	—	`src/domains/router/fallback.ts` + `src/domains/router/tools.ts` (emit `thought_record` of shape `routing_decision` per concept doc §Decision-trail recording)

Note: P1.5.9 ships BEFORE P1.5.1 in logical order (scoring needs candidates to score against), but the Phase 0 schema already has the single-row candidate table. For Phase 1.5 the sequence is P1.5.9 (seed rows) → P1.5.1 (scoring over seeded rows), and the depends-on DAG in the prompts file reflects that. The order in the table above matches how the scope was given in the R76.P2 brief.

3. Prompt-file skeleton

Each of the 10 blocks below gets fleshed out in step 4 following the p0.9-nu-integrations.md template shape:

Frontmatter: title, tags, type: agent-prompts, status: spec-ready, round: R91-plan, updated: 2026-04-18.
Front-of-doc banner: “Phase 1.5 plan. Do not execute in Phase 0. Activates at R91+ per roadmap §Phase 1.5. Trigger conditions from ADR-005 §Implementation.”
Group summary table (same 10-row table as §2 here, minus the “files to modify” column).
Then for each sub-task:
- Spec source + extraction reference + worktree + branch command + effort + depends-on + unblocks.
- Files to create.
- Acceptance criteria (6–10 checkbox items).
- Pre-flight reading.
- Ready-to-paste agent prompt.
- Verification checklist (for reviewer agent).
- Writeback template.
- Common gotchas.

4. Edits outside the main file

4.1 `docs/guides/implementation/task-prompts/index.md`

Append a new section after “After Phase 0”:

## Phase 1.5 — δ Multi-Model Router (R91+, planned)

See [ADR-005](../../architecture/decisions/ADR-005-multi-model-defer.md)
for the trigger conditions and upgrade path. The prompts below activate
when all four ADR-005 trigger conditions hold.

| Group | File | Sub-tasks |
|-------|------|-----------|
| P1.5 — δ Router Graduation | [p1.5-delta-router-graduation.md](p1.5-delta-router-graduation.md) | P1.5.1 – P1.5.10 |

No edits to the Phase 0 table or the critical-path summary.

4.2 `docs/architecture/decisions/ADR-005-multi-model-defer.md`

Add a single line at the end of the R75 Wave I postscript (the existing trailing sentence stays; this appends a pointer):

**Phase 1.5 graduation prompts:** see
[`docs/guides/implementation/task-prompts/p1.5-delta-router-graduation.md`](../../guides/implementation/task-prompts/p1.5-delta-router-graduation.md)
for the 10-sub-task plan activating at R91+.

5. Step-5 verification plan

Count sub-task blocks in the new file — must equal 10.
Grep for AMS_MODEL / AMS_* — must return zero hits in the new file.
Grep for each of the 4 tool names (router_score, router_call, router_fallback, router_stats) — must each appear at least once in the new file.
Grep for each of the 8 candidate model ids (claude sonnet, claude haiku, gpt-4o, gpt-4o mini, gemini, llama, mixtral, kimi k2) — must appear in the P1.5.9 candidates block.
Confirm index.md has the new Phase 1.5 heading.
Confirm ADR-005 has the one-line pointer at the end.
Run npm run build && npm run lint && npm test and record the pass count (expected: 1085/1085 unchanged; docs-only round).

6. Risk ledger (carry-forward from audit §9)

All risks are mitigated by template fidelity + explicit citation. The only residual risk is that some future round changes the 7-dimension formula or the candidate cohort after this file lands; in that case the Phase 1.5 prompts rebase from the concept doc as part of the R91+ kickoff, not this round.

7. Time budget

Step 1 (audit): ~15 min (done).
Step 2 (contract): ~10 min (done).
Step 3 (packet): ~10 min (this file).
Step 4 (impl): ~60 min — one large file (~800–1100 lines) plus 2 small edits.
Step 5 (verify): ~15 min — counting + gates.

Written 2026-04-18 as step 3 of the R76.P2 5-step chain. Step 4 follows.