P1.5.7 — `router_*` MCP Tools Verification

1. Gates

Gate	Command	Result
Build	`npm run build`	PASS — no TypeScript errors
Lint	`npm run lint`	PASS — no ESLint errors
Test	`npm test`	PASS — 3353/3353 tests across 77 suites

Baseline on cf6221c9 (post-P1.5.6): 3325 tests across 76 suites. Delta after P1.5.7: +28 tests (the new src/__tests__/domains/router/tools.test.ts suite). +1 suite.

2. Files changed

docs/audits/p1-5-7-mcp-tools-audit.md             (new, 241 lines)
docs/contracts/p1-5-7-mcp-tools-contract.md       (new, 315 lines)
docs/packets/p1-5-7-mcp-tools-packet.md           (new, 403 lines)
docs/verification/p1-5-7-mcp-tools-verification.md (new, this file)
src/domains/router/tools.ts                        (new, 502 lines)
src/__tests__/domains/router/tools.test.ts        (new, 339 lines)
src/domains/router/index.ts                        (+1 line — barrel re-export)
src/server.ts                                      (+2 lines import, +14 lines register call)

3. MCP tool surface count proof

3.1. Pre-P1.5.7 surface (23 tools on main `cf6221c9`)

Wave	Domain	Tools
Phase 0 P0.2.1 / P0.7 / P0.6 / P0.3 / P0.8	α / ε / ζ / β / η	`server_ping`, `server_health`, `thought_record`, `thought_record_list`, `audit_verify_chain`, `skill_list`, `task_create`, `task_get`, `task_update`, `task_list`, `task_next_actions`, `audit_session_start`, `merkle_finalize`, `merkle_root`
Total (Phase 0)		14
R89 Phase A (P2.5.1)	λ	`reputation_get`, `reputation_history`, `reputation_leaderboard`, `reputation_check_gates`
Total (R89A)		+4 = 18
R89 Phase B (P3.7.1)	θ	`consensus_propose`, `consensus_vote`, `consensus_finality`, `consensus_gossip`, `vrf_eval`
Total (R89B)		+5 = 23

3.2. Post-P1.5.7 surface (27 tools)

Wave	Domain	Tools
P1.5.7 (this slice)	δ	`router_score`, `router_call`, `router_fallback`, `router_stats`
Total (Phase 1.5 W6)		+4 = 27

3.3. Registration call site

src/server.ts:594 (after registerConsensusTools(ctx);):

// P1.5.7: register δ Router MCP tools — router_score, router_call,
// router_fallback, router_stats. Phase 1.5 W6 graduation: first δ
// MCP surface (Phase 0 P0.5.1/P0.5.2 shipped library-only stubs per
// ADR-005). Closes ADR-004 R75 Wave H tool-surface amendment for δ.
// Tool count moves from 23 → 27 (Phase 0: 14, λ R89A: 4, θ R89B: 5,
// δ Phase 1.5: 4). Underlying handlers wrap REAL functions
// (scoreIntent / routeRequest / getCircuitBreakerState /
// resetCircuitBreaker / getRouterStats); no stubs. apiKey +
// injection seams (completionFn, fetchFn, scoringFn, delayFn,
// logger, nowFn) are rejected at the strict-Zod input boundary;
// secrets come from COLIBRI_*_API_KEY env vars only.
registerRouterTools(ctx);

4. Tool-by-tool acceptance verification

4.1. `router_score`

AC	Status	Evidence
Wraps `scoreIntent(prompt, context ?? {})`	PASS	`tools.ts:300-321`
Returns `{scores, winner, rule_version_hash}`	PASS	Output type at `tools.ts:266`, schema at `tools.ts:209-216`
Zod input strict + min(1) on prompt	PASS	`tools.ts:138-150`
Output schema `rule_version_hash` validates `^sha256:[0-9a-f]{64}$`	PASS	`tools.ts:213`
Test 1: empty-candidate path returns `winner: 'claude'`	PASS	Test `√ 1`
Test 2: rule_version_hash matches sha256 pattern	PASS	Test `√ 2`
Test 3: determinism	PASS	Test `√ 3`
Test 4: empty prompt rejected	PASS	Test `√ 4`
Test 5: extra key rejected (strict)	PASS	Test `√ 5`
Test 6: negative tokens rejected	PASS	Test `√ 6`
Test 7: output passes output schema	PASS	Test `√ 7`

4.2. `router_call`

AC	Status	Evidence
Wraps `routeRequest(prompt, options)`	PASS	`tools.ts:340-371`
Returns full `RouteResult` (including costUsd, modelsAttempted)	PASS	Return type `RouterCallOutput = RouteResult` at `tools.ts:269`
Zod schema rejects `apiKey`	PASS	Test `√ 10`
Zod schema rejects `completionFn`	PASS	Test `√ 11`
Zod schema rejects empty prompt	PASS	Test `√ 12`
Zod schema accepts valid options subset	PASS	Tests `√ 8`, `√ 9`
Output schema OMITTED (variable shape)	NOTE	Documented in contract §3.2 and source §5
`FallbackChainExhaustedError` propagates to MCP `HANDLER_ERROR`	PASS by inheritance	Middleware Stage-4 catch wraps any throw (`src/server.ts:361-374`); error.message preserves attempted-model list (`fallback.ts:312-323`)

4.3. `router_fallback`

AC	Status	Evidence
Wraps `getCircuitBreakerState()`	PASS	`tools.ts:376-393`
Wraps `resetCircuitBreaker(modelId?)`	PASS	`tools.ts:378-384`
`{reset: true, model_id: X}` clears one model	PASS	Test `√ 15`
`{reset: true}` clears all models	PASS	Test `√ 16`
Read-only call does NOT mutate state	PASS	Test `√ 17`
Zod schema rejects unknown model_id	PASS	Test `√ 18`
Output is plain object (not Map)	PASS	Test `√ 13`, `√ 14` (uses `toEqual({...})`)

4.4. `router_stats`

AC	Status	Evidence
Wraps `getRouterStats()`	PASS	`tools.ts:401-407`
Returns `{models: Record<ModelId, RouterStats>}`	PASS	Test `√ 23`
Empty state → `{models: {}}` (sparse)	PASS	Test `√ 19`
After success record → `calls_total = 1`	PASS	Test `√ 20`
After success+failure → `success_rate = 0.5`	PASS	Test `√ 21`
Zod input strict (rejects extras)	PASS	Test `√ 22`
Output passes output schema	PASS	Test `√ 23`

4.5. `registerRouterTools`

AC	Status	Evidence
Adds exactly 4 names	PASS	Test `√ 24`
Idempotence guard throws on second call	PASS	Test `√ 25`

4.6. Cross-cutting

AC	Status	Evidence
Outputs frozen	PASS	Test `√ 26`
No forbidden keys in input schemas (source grep)	PASS	Test `√ 27`
MODEL_IDS tuple length 9 matches ModelId union	PASS	Test `√ 28`

5. Forbiddens check

All forbiddens listed in the round prompt and the slice file’s FORBIDDENS section verified:

Forbidden	Status
Edit main checkout (`E:\AMS`)	NOT VIOLATED — all edits in `.worktrees/claude/p1-5-7-mcp-tools`
Push to main	NOT VIOLATED — push step uses `feature/p1-5-7-mcp-tools`
Force push	NOT VIOLATED — no force push
`--no-verify` or `--amend`	NOT VIOLATED — 5 sequential commits
Edit underlying router function	NOT VIOLATED — `scoring.ts`, `fallback.ts`, `circuit.ts`, `cost.ts`, `adapters/*` untouched
Edit any adapter (W3 territory)	NOT VIOLATED
Use Zod v4	NOT VIOLATED — uses Zod v3.23 surface (`.strict()`, `.safeParse()`)
AMS_* env vars	NOT VIOLATED — only `COLIBRI_*` referenced
Implement parity tests (P1.5.8)	NOT VIOLATED — scope limited to schema + handler tests
ζ integration (P1.5.10)	NOT VIOLATED — no `thought_record` call in this module
Promote δ frontmatter `partial` → `complete`	NOT VIOLATED — no frontmatter edits
Invent tools beyond 4	NOT VIOLATED — exactly 4 tools registered
Accept `apiKey` in MCP input	NOT VIOLATED — strict schemas reject it (Test `√ 10`)
Accept injection seams (completionFn / scoringFn / fetchFn / logger / delayFn / nowFn)	NOT VIOLATED — strict schemas reject them (Test `√ 11`, `√ 27` static grep)

6. Known pre-existing flakes (out of scope)

These flakes were documented in the round prompt and observed during the test runs:

consensus/parity-harness G7.1 perf budget — CI-load sensitive, retry-clean.
- Hit once during the initial full-suite run (6363 ms > 5000 ms budget).
- Retry was clean (1637 ms).
- Out of P1.5.7 scope.
reputation/tools.test.ts parallel-migration prefix race — retry-clean.
- Not observed in this round’s runs.
- Out of P1.5.7 scope.
kimi.test.ts ● injection seams › 7. latency measurement: 50ms delay → latencyMs >= 50 — timer-imprecision under CI load (introduced by P1.5.2 W3).
- Not observed in this round’s runs.
- Out of P1.5.7 scope; documented as candidate for P1.5.8 parity suite to address.

7. Test tail

Final clean full-suite run:

Test Suites: 77 passed, 77 total
Tests:       3353 passed, 3353 total
Snapshots:   0 total
Time:        38.999 s, estimated 53 s
Ran all test suites.

8. Tool summary table

Tool	Input	Output	Wraps
`router_score`	`{prompt: string≥1, context?: {task?, operatorPreference?}}`	`{scores: Record<string,number>, winner: ModelId, rule_version_hash: sha256:hex64}`	`scoreIntent` + `computeScoringRuleVersionHash`
`router_call`	`{prompt: string≥1, options?: {maxTokens?, systemPrompt?, model?, task?, operatorPreference?}}`	`RouteResult` (full shape from `fallback.ts:232-242`, includes `costUsd`, `modelsAttempted`)	`routeRequest`
`router_fallback`	`{model_id?: ModelId, reset?: boolean}`	`{circuitState: Record<ModelId, {failures: number, openedAt: number\\|null}>}`	`getCircuitBreakerState` + optional `resetCircuitBreaker`
`router_stats`	`{}`	`{models: Record<ModelId, {calls_total, successes, failures, avg_cost_usd, p50_latency_ms, success_rate}>}`	`getRouterStats`

9. Compliance with slice file (P1.5.7 §”Ready-to-paste agent prompt”)

Slice requirement	Status
Export 4 MCP tool factories	PASS
`router_score` Zod input `{prompt, context?}`	PASS
`router_call` Zod input mirrors RouteOptions (apiKey NOT accepted)	PASS — apiKey + injection seams rejected
`router_fallback` Zod input `{model_id?, reset?}`	PASS
`router_stats` Zod input `{}`	PASS
Tool count 14 → 18 (slice file wording, predated R89 A+B)	EFFECTIVELY: 23 → 27 — slice was written before R89 closed and is stale. Net effect on δ axis is +4 tools, matching slice intent.
Tests: Zod validation + bad-input rejection + ζ emission	PASS for Zod + bad-input. ζ emission deferred to P1.5.10 per round prompt’s explicit forbidden.
Single global completion callable rejection from MCP	PASS (Test `√ 11`)
`npm run build && npm run lint && npm test` green	PASS

10. Deviations from slice / contract

router_call output schema OMITTED: contract §3.2 documents this — RouteResult.content is variable across upstream adapters. The handler stays type-safe at the TS level. This deviates from the consensus exemplar (which includes output schemas on all 5 tools).
rule_version_hash format: contract §2.2 initially specified ^[0-9a-f]{64}$. The live computeVersionHash implementation prefixes with sha256:, so the schema (and Test 2) was adjusted to ^sha256:[0-9a-f]{64}$. Documented in §4.1 above.
ζ emission deferred: The slice file’s acceptance criterion “All 4 tools emit a thought_record (type ‘decision’)” is explicitly deferred to P1.5.10 per the round prompt’s “ζ integration (P1.5.10 scope)” forbidden. The middleware’s audit-enter/audit-exit stages already wrap every tool call; P1.5.10’s SQLite-backed audit sink will produce the thought_record entries without further router-tools edits.
tools param omitted from router_call: Slice file mentions tools in router_call’s RouteOptions mirror. The Anthropic tool-shape pass-through is too complex to validate at the MCP boundary without re-vendoring the schema; deferred to a future round (forward-compatible additive change).
Slice’s “14 → 18” tool-count wording is stale: The slice file was authored before R89 Phase A+B added 9 tools (4 λ + 5 θ). Pre-P1.5.7 surface is actually 23 (14 + 4 λ + 5 θ); post-P1.5.7 is 27. The net δ contribution (+4 tools) matches slice intent.

P1.5.7 — router_* MCP Tools Verification