Repertoire Integration (v3.3+)

Repertoire is preferred. Repertoire (@0xray/repertoire@0.2.9 vendored) is the long-term judgment organ. It matches named laws and enriches orchestrator routing through the MemoryRoutingProvider contract. Wear it. Do not treat Station as a substitute.

Not the successor card. Survive compact / host change is Station (Read .xray/state/STATION.md). 0.1.0–0.1.8 were the same judgment organ; 0.2 cut the factory seed to 8 names. Do not pin 0.1.8. Split: Station vs Repertoire 0.1.

When to use Repertoire

Scenario	Surface
0xRay orchestrator in-process (ExecutionPlanner, thinDispatch)	features.json → memory_routing
External LLM host (Hermes, Grok, OpenCode MCP)	repertoire-mcp stdio server
Ad-hoc signal queries / ingest pipeline	Repertoire CLI (npm run query, pipeline)

Setup

0xray vendors @0xray/repertoire@0.2.9. Fresh npm install 0xray ships memory_routing.enabled: true. Preferred on. Explicit opt-out only: enabled: false with provider: "repertoire". Station then holds the compact ticket — that is fallback, not the preferred brain. This is not an 8th 0xRay MCP.

Shipped xray/features.json:

"memory_routing": {
 "enabled": true,
 "provider": "repertoire",
 "module_path": "node_modules/@0xray/repertoire/dist/provider/memory-routing-provider.js",
 "config": {
 "signalsPath": ".xray/state/repertoire/curated_signals.json",
 "statePath": ".xray/state/repertoire/inference-state.json",
 "feedbackDir": ".xray/state/repertoire/feedback"
 }
}

Optional sibling (framework monorepo only):

git clone https://github.com/0xRayAI/repertoire ../repertoire
cd ../repertoire && npm install && npm run build

Grok session-start writes a one-line repertoireResume into .xray/state/session-boot.json (signal count when the registry is readable). Leftover default-off also resolves at runtime in loadMemoryRoutingProvider / getMemoryRoutingConfig when the module is on disk — ExecutionPlanner and thinDispatch then use Repertoire without a features rewrite. Station heat persists working state (not Bedrock primitives) to .xray/state/repertoire-working.json and a Working: line on STATION.md. Compact may ingestFeedback for non-Bedrock matches. Explicit opt-out remains enabled: false + provider: "repertoire".

features.json fields (memory_routing)

Validated by xray/features.schema.json at load time:

Field	Required	Values	Description
enabled	yes	true / false	Master switch
provider	yes	null, repertoire, custom	Provider kind
module_path	when enabled + repertoire/custom	path string	ESM module exporting createMemoryRoutingProvider()
config	optional	object	Passed to provider factory

Repertoire config keys

Key	Default	Description
dataDir	package data/ (vendored 0.2)	Registry + inference state root
signalsPath	.xray/state/repertoire/curated_signals.json	Project-local registry (factory + stack laws). Dest is named invariants, not hangar repo-* or git slugs.
statePath	data/inference-state.json	Idempotent ingest cursor
logDir	logs/groover-inference	Enriched JSONL log directory

Paths resolve relative to the 0xRay package root unless absolute.

What Repertoire enriches in 0xRay

ExecutionPlanner (v3.3.1 confidence gate)

When getTaskConfidence() is available on the provider:

• calculateTaskComplexity() adds complexityBoost from matched signals
• Trap tasks (ontological-trap, high-confidence primitives) get higher complexity scores
• selectAgent() receives trap hints via TYPE: ontological-trap in operation text
• enrichTasks() attaches memorySignals, memorySignalConfidences, memoryComplexityBoost to task metadata

thinDispatch

• resolveThinDispatch() may override tier-default agent (e.g. → architect when highConfidenceTrapPresent and score ≥ 26)
• provenance-failure tag can route to bug-triage-specialist

Researcher (analyze_proposal)

Module: src/mcps/researcher-confidence.ts

• Triggers on trap language or high-confidence primitive matches
• Calls getTaskConfidence() via MemoryRoutingProvider
• Appends auditable MEMORY_ROUTING: block to governance output

Feedback loop

Per-task ingestFeedback() (v3.3 — not aggregate-only) records orchestrator outcomes back to Repertoire.

MemoryRoutingProvider contract

Custom providers export createMemoryRoutingProvider(config) implementing:

Method	Pipeline stage
buildRoutingContext(operation)	Signal/tag matching
enhanceAgentCapabilities(map)	Agent capability enrichment
enrichTasks(tasks)	Pre-planning metadata
buildInheritedContext(tasks)	→ ExecutionPlan.memoryContext → AsideContext inheritedContext.memoryRouting
selectAgent(...)	ExecutionPlanner assignment
resolveThinDispatch(...)	thinDispatch score + override
getTaskConfidence?(task)	Confidence gate (v3.3.1)
ingestFeedback?(entry)	Per-task learning loop

See src/memory-routing/types.ts and Repertoire's docs/MEMORY-ROUTING-PROVIDER.md.

Repertoire MCP server (external hosts)

Not an 8th 0xRay server. Wear adds a repertoire extra when the module resolves (Hermes / OpenCode / OpenClaw / project .mcp.json / Grok .grok/config.toml).

Server tools are unprefixed so Grok TUI can register them (server__tool → repertoire__get_task_confidence). Hermes/OpenCode see the unprefixed names.

"repertoire": {
 "command": "node",
 "args": ["node_modules/@0xray/repertoire/dist/mcp/server.js"]
}

In-repo launcher (cwd-proof): scripts/mjs/run-repertoire-mcp.mjs.

MCP tools

Server name	Grok TUI name	Purpose
get_high_confidence_signals	repertoire__get_high_confidence_signals	List signals above threshold
get_task_confidence	repertoire__get_task_confidence	Full confidence context for a task
search_primitives	repertoire__search_primitives	Text search against registry
ingest_feedback	repertoire__ingest_feedback	Record orchestrator outcome

In-process orchestrator code uses getMemoryRoutingProviderSync().

Repertoire CLI

cd repertoire
npm run ingest -- --source /path/to/groover/logs
npm run pipeline # ingest + meta-inference
npm run query # ad-hoc confidence query
npm run test:e2e # enriched loop regression
npm run test:mcp # stdio MCP smoke

Data model (summary)

Artifact	Role
data/curated_signals.json	Primitive registry with observation_stats
data/inference-state.json	Ingest idempotency cursor
logs/groover-inference/*.jsonl	Enriched entries (strict enriched-only ingest)

Confidence values come from observation_stats or explicit task metadata — no text-score fallbacks.

Lessons, mint, and caps

A passing grade appends one lesson on the named law and stores that line beside the average. The same task id does not step again.

Cap	Value	What leaves
LESSON_LINE_CAP	20	Oldest lesson text, only after its task id is in the ledger
RETAINED_LESSON_ID_CAP	20	Oldest id, insertion order (not alphabetical)
LEARNED_SIGNAL_CAP	24	Only proposed signals tagged learned (lowest confidence, then oldest last_seen)

Speech that names no stored signal mints one proposed signal tagged learned. The same speech does not mint twice. Factory signals are not eviction candidates.

A plate is the machine schematic (Plates). The npm package ships docs-site/docs/plates/. Station keeps one pointer line for the plate intent names. Pre-compact and MEMORY_ROUTING inject that one schematic.

Testing

0xRay framework repo:

npm test -- src/__tests__/unit/memory-routing-integration.test.ts
npm test -- src/__tests__/unit/memory-routing-provider.test.ts
npm test -- src/__tests__/unit/researcher-confidence.test.ts
npm test -- src/__tests__/unit/researcher-repertoire-wiring.test.ts

Related

• Memory Routing — provider plug-in model
• Features Since 3.1 — full changelog of capabilities
• features.json Reference
• Platform Integrations