# Execution and capability truth MoonFind has no agent runtime. It prepares domain packets, validates research evidence, and projects durable state. MoonClaw remains the only model/agent runtime, MoonFlow remains the only graph activation engine, MoonTown owns civic deliberation, and MoonBook owns reviewed memory and Three-Gap closure. ## Operator call chain The Rabbita operator path is intentionally two-part: 1. The host builds a `moonflow.capability-catalog.v1` from the active installed packs, version-matched `moonflow.adapter.v2` declarations, and unexpired health evidence. 2. MoonFind prepares `moonfind/research.prepare-analysis@0.1.0` with the exact `moonfind/research-run@1.0.0` input, `moonfind/moonclaw-research-command@1.0.0` output, authority, evidence digest, request ID, attempt ID, and idempotency key. The input reference is the workspace-relative `.moonsuite/products/moonfind/runs//run.json`. Its digest uses MoonClaw's exact byte convention: SHA-256 each persisted file, render `|`, join identities with a newline, then SHA-256 the aggregate. JSON is parsed from those same bytes; it is never re-serialized to invent the invocation digest. 3. Before submission, the host appends the `prepared` and `submitting` revisions to `.moonsuite/products/moonfind/runs//capability-execution.jsonl`. 4. The Rabbita client sends the immutable invocation through MoonClaw's generic installed-pack capability port. It never calls a model provider. 5. `submitted`, `pending`, `failed`, `unknown`, and successful receipts are identity-checked and appended. Restart recovery repeats the same invocation only to reconcile the same idempotency key. 6. A successful capability receipt must contain the exact `moonclaw.mooncode.v1` command. Rabbita creates/configures a MoonClaw Cowork session and submits that command through the existing `/messages` path. 7. The returned Cowork session and run IDs are appended before polling. Restart recovery reloads those IDs; it does not create a replacement run. The browser posts only a `moonfind.moonclaw-session-checkpoint-request.v1`; the trusted host appends the session immediately, reads the runtime receipt when a run appears, adds its own observation time and durable result reference, and appends the typed run observation. 8. Only a completed Cowork run with a durable result reference moves the MoonFind execution to `needs-review`. 9. MoonTown challenge, named-human research review, and `moonbook/bookkeeper.outcome.close@0.1.0` remain later, independently receipted stages. Successful execution is never self-acceptance. The browser discovers submit, reconcile, and checkpoint URLs from the MoonFind workspace snapshot. Absence of any required port is an explicit UI blocker, not a fallback to a direct model call. ## Pack-owned deterministic adapter MoonClaw reaches MoonFind's installed `research.prepare-analysis` operation through a pack-owned loopback adapter. This adapter is not another agent runtime: it cannot call a model or provider. It accepts MoonClaw's exact `PackToolExecutionRequest`, verifies the installed binding and immutable attempt against the durable `run.json`, and materializes only the governed `moonclaw.mooncode.v1` command. Start it beside the workspace: ```bash MOONFIND_WORKSPACE="$HOME/Research" \ MOONFIND_PACK_ADAPTER_PORT=4314 \ moon run cmd/pack_adapter_host ``` Its only mutation endpoints are: ```text POST http://127.0.0.1:4314/v1/pack-tool/submit POST http://127.0.0.1:4314/v1/pack-tool/reconcile ``` Configure the corresponding exact route in MoonClaw's host-owned `.moonsuite/products/moonclaw/capability-http-policy.json`: ```json { "route_id": "moonfind-prepare-analysis", "operation_ref": "moonfind/research.prepare-analysis@0.1.0", "adapter_id": "moonfind-pack-local-v1", "transport": "moonclaw.pack-tool-http.v1", "submit_url": "http://127.0.0.1:4314/v1/pack-tool/submit", "reconcile_url": "http://127.0.0.1:4314/v1/pack-tool/reconcile", "base_url": null, "estimated_cost_micros": 0, "provider_route_ref": null } ``` The MoonFind UI host still points to MoonClaw, not directly to this adapter: ```bash MOONCLAW_CAPABILITY_SUBMIT_URL=http://127.0.0.1:18123/v1/capability/submit MOONCLAW_CAPABILITY_RECONCILE_URL=http://127.0.0.1:18123/v1/capability/reconcile ``` Before writing command evidence, the adapter persists the complete immutable request and planned output below `.moonsuite/products/moonfind/pack-tool-attempts//`. Submission replay returns the stored result. Reconciliation never creates a missing attempt; it either resumes the same deterministic local materialization or returns `unknown`. Evidence is restricted to the exact `records/moonclaw-commands/.json` path and symbolic-link escapes are rejected. ## Restart states | Persisted state | Only safe next action | |---|---| | `prepared` | submit once | | `submitting`, `pending` | reconcile the same capability attempt | | `unknown` | investigate and reconcile; never blind retry | | `command-ready` | submit the exact receipt command through Cowork | | `runtime-submitting` | poll the persisted Cowork session for its run ID | | `runtime-pending` | poll the persisted Cowork session/run | | `runtime-unknown` | investigate/reconcile the same Cowork run | | `runtime-failed` | named operator decision | | `needs-review` | named-human review | The append-only journal uses fully synced JSON lines. Recovery ignores a torn last line and returns the latest complete revision. Request, operation, attempt, idempotency, and invocation identities cannot change between rows. ## Cross-product graph [`humanoid-robotics-suite.v1.json`](../workflows/humanoid-robotics-suite.v1.json) is an exact desired graph, not a runnable MoonFlow graph. Version 3 declares 63 stages over 43 canonical operations and contains no null operation reference. It binds MoonBook requirements and Bookkeeper closure, three independent MoonChat decision stages, MoonClaw `wiki` and `mooncode` roles on `gpt-5.6-sol`, MoonFind research, MoonTown challenge, MoonProj planning, MoonMold spatial work, MoonRobo integration/readiness observation, MoonMoon simulation, MoonVis evidence composition, and MoonCast's complete gated production/edit/review/delivery chain. `project_desired_graph` resolves stages exclusively from a supplied `moonflow.capability-catalog.v1`. Its gap suggestions come only from catalog operations for the requested product. `importable_graph` is `None` unless all stages match exact product, version, schema, authority, required-claim ceiling, and health identities; `require_importable_graph` fails closed on any gap. A single missing or expired operation produces diagnostics but no partial MoonFlow work model. MoonBook supplies wiki functionality, so `understand-needs` is not a phantom MoonWiki product. MoonClaw supplies the only agent runtime; both the wiki check and software construction use its exact generic goal operation, with roles `wiki` and `mooncode`, rather than phantom runtimes. MoonFlow is the execution engine. MoonGate supplies policy/catalog support and MoonLib supplies shared contracts; neither is a domain work item. MoonEdit and MoonLeaf are libraries. MoonFish is intentionally excluded from this robotics book. The graph issues no physical command and no publication request. Its terminal boundary is MoonRobo readiness observation plus MoonMoon simulation, immutable MoonCast build/client/final evidence, and a named MoonChat authority decision; only then may MoonBook Bookkeeper close the outcome. The template `source_digest` is the SHA-256 of the canonical desired graph with that self-referential field omitted. `publish-robotics-graph` parses the current v2 contract and rebinds a target digest from the template graph ID, template book ID, template digest, declaration revision, and requested MoonBook ID. It does not search for or replace an old version literal. Every desired stage owns exactly one initial request artifact named `requests/..json`. Predecessor outputs are not listed there: MoonFlow appends accepted dependency artifacts after registration. The publisher now requires a complete `moonfind.robotics-request-materialization.v1` bundle and validates exact stage/schema/path cardinality, payload digests, registered required fields, and constant contract fields before writing anything to the MoonBook. The pack-specific template registry pins all 39 input-schema digests and 21 declared `contract`/`contract_id` constants. Missing or drifted requests reject the whole publication before the target `flow/` directory is created. Publication also requires an explicit compiled capability catalog and canonical evaluation timestamp. The bound v2 intent is preserved as `flow/desired-graph.json`; `flow/work-graph.json` is the projected `moonsuite.work-model.v1` (`items`, not desired `stages`). Missing, expired, authority-drifted, schema-drifted, or claim-exceeding operations therefore fail before any target file is created. The publisher also snapshots the exact catalog used for projection at `flow/capability-catalog.json`; MoonDesk passes that snapshot and a new server-generated evaluation timestamp back to MoonFlow's conformant validator before import. A synthetic or expired snapshot remains non-production evidence and fails according to its own health window. [`humanoid-robotics-requests`](../fixtures/humanoid-robotics-requests) contains 63 JSON-Schema-valid synthetic requests plus their completeness ledger. It is an adapter-preflight scenario fixture, not evidence that provider jobs, customers, publication, or a physical robot are ready. [`humanoid-robotics-capability-catalog.fixture.v1.json`](../fixtures/humanoid-robotics-capability-catalog.fixture.v1.json) contains deterministic, short-lived synthetic health attestations for every operation solely to test projection. Its catalog ID, notice, evidence paths, and every attestation identify it as fixture-only. It is not packaged as a production catalog and cannot prove providers, licensing, publication, or physical readiness. MoonCast's two `external-effect` stages are bounded provider-generation calls (`asset-factory.generate` and `production-coordinator.run-pending`). Each requires an explicit provider grant and records `publication: false`; neither authorizes distribution or a physical effect. ## Installation health [`adapter-declaration.v1.json`](../adapters/moonflow/adapter-declaration.v1.json) publishes the real MoonFind manifest tools under `moonflow.adapter.v2`. [`adapter-health.template.v1.json`](../adapters/moonflow/adapter-health.template.v1.json) is deliberately `unavailable`: only the installed host can observe credentials, provider reachability, and executable adapter wiring and issue a short-lived healthy attestation. A repository fixture cannot satisfy that operational gate.