MoonSuite AutoResearch domain pack

moon add vectie/moonfind@0.1.0

Author

vectie

Version

0.1.0

License

Apache-2.0

Repository

github.com/vectie/moonfind

Last updated

14 days ago

Downloads

Dependencies

README

#MoonFind

Domain pack · evidence-backed research alpha. Read the product contract for provider licensing, hostile-document isolation, the closed loop and release gates.

MoonFind is a MoonSuite domain pack for evidence-backed AutoResearch. It tracks research questions, acquires and normalizes papers, asks MoonClaw to create cited summaries and falsifiable ideas, uses MoonTown to challenge bridges across paper-books, and routes reviewed outcomes to MoonBook Bookkeeper for Three-Gap learning.

The first specialization is humanoid robotics model research over arXiv. Domain semantics stay in humanoid/; the reusable research graph stays in research/. The implementation is MoonBit-native and the application UI is Rabbita. There is no Python runtime or secondary agent runtime.

#Product loop

Research watch
  → arXiv point-in-time paper records
  → deterministic domain relevance ranking
  → MoonClaw gpt-5.6-sol analysis packet
  → evidence-backed summaries, claims, questions, ideas
  → Moontown research salon across paper-books
  → named-human review
  → MoonBook outcome binding
  → Bookkeeper Three-Gap analysis
  → reviewed capability proposal
  → MoonFlow canary/shadow activation
  → next MoonClaw run resolves the assigned capability
  → observed experiment outcome returns to Bookkeeper

Moonfind never installs a capability or approves its own output. Model ideas are proposals until evidence and a named human review say otherwise.

MoonFind also exposes an optional public listing search adapter for MoonTown's exact sanitized moonsuite.public-listing.v1 cards. It performs deterministic, explainable organic retrieval only; Town continues to own publication, recommendations, sponsorship, notifications and messaging.

See Responsibility and testability for the executable delegation boundary and layered test model. See Execution and capability truth for the restart-safe installed-pack → MoonClaw Cowork call chain.

#Run the complete fixture

moon run cmd/main -- run-fixture /tmp/moonfind-workspace fixtures
moon run cmd/main -- status /tmp/moonfind-workspace humanoid-cross-paper-demo
moon run cmd/main -- export-ui /tmp/moonfind-workspace humanoid-cross-paper-demo ui/rabbita-moonfind/public/moonfind-workspace.json

The run is written below .moonsuite/products/moonfind/runs/<run-id>/ with:

the normalized and ranked paper set;
a MoonClaw command bound to gpt-5.6-sol and moonfind.humanoid.synthesis;
the validated analysis and human-review state;
a Moontown research-salon scenario;
a versioned Moontown handoff bound to the Moonfind run;
a MoonBook Bookkeeper deliverable candidate, ready for its own durable review before outcome binding;
a MoonFlow capability observation;
a Rabbita workspace projection.

For UI-to-UI qualification, stop before the legacy fixture applies its direct terminal review:

MF_QUAL="$(mktemp -d /tmp/moonfind-ui-qualification.XXXXXX)"
moon run cmd/main -- prepare-fixture "$MF_QUAL/suite" fixtures

This leaves humanoid-cross-paper-demo at human-review, with a typed MoonTown handoff and no human-review.json. Follow the UI-to-UI guide to exercise the governed MoonTown, MoonChat, and MoonBook path.

#Use live arXiv intake

moon run cmd/main -- prepare-live ~/Research moonfind-2026-07-20

The pending run is durable. The production operator path first invokes moonfind/research.prepare-analysis@0.1.0 through MoonClaw's generic installed-pack seam, appends every attempt/receipt revision, and only then submits the exact resulting command through the existing MoonClaw Cowork gateway. The Cowork session/run identity is appended before polling. Save its durable JSON result, then:

moon run cmd/main -- complete ~/Research moonfind-2026-07-20 analysis-result.json
moon run cmd/main -- review ~/Research moonfind-2026-07-20 human-review.json

Before full text can enter the MoonClaw packet, record a moonfind.full-text-evidence.v1 artifact. The command fails closed unless the provider-use decision allows the exact use, the hostile-PDF isolation receipt is bound to the PDF digest, active content was removed with networking disabled, the expansion/page limits pass, and page/section evidence locations exist:

moon run cmd/main -- record-fulltext ~/Research moonfind-2026-07-20 full-text-evidence.json
moon run cmd/main -- resume ~/Research moonfind-2026-07-20

resume rebuilds a missing MoonClaw packet from durable run/full-text records and writes recovery.json with the next idempotent action. It never invents a completed stage.

The arXiv adapter remains a public development/research adapter. Provider policy records for Crossref, OpenAlex, Semantic Scholar, and a licensed full-text port are present, but are deny-by-default until a named review records usable terms. Validate an intended use with:

moon run cmd/main -- validate-provider provider-policy.json use-request.json

#Run the current Moontown handoff

Moonfind and Moontown remain separate products. Moonfind writes a versioned handoff; Moontown consumes it through its existing civic runtime and writes a machine-readable receipt. MoonDesk now exposes the generic MoonFlow composition canvas and can invoke authority-bound actions published by MoonFlow. Moonfind still does not call a model provider directly: the installed MoonClaw host remains the only agent runtime. The equivalent operator path is:

RUN_DIR="$HOME/Research/.moonsuite/products/moonfind/runs/moonfind-2026-07-20"
RECEIPT="/tmp/moonfind-2026-07-20-moontown-receipt.json"

cd ../moontown
moon run src/cmd/main -- civic protocols pattern-handoff \
  "$RUN_DIR/moontown-handoff.json" "$RECEIPT" "$HOME/Research/.moontown"

cd ../moonfind
moon run cmd/main -- ingest-moontown \
  "$HOME/Research" moonfind-2026-07-20 "$RECEIPT" \
  "$HOME/Research/.moontown/books/wenyu-social-square/raw/bootstrap/civic-communication-results.json"

The production route uses Moontown's existing MoonClaw reducer. pattern-handoff-fixture exists only for deterministic integration testing. The normal moonfind review command refuses to proceed until both the run-bound receipt and typed reducer output have been ingested. A completed receipt must still say review_status: pending; Bookkeeper/human review is a separate authority gate. The self-contained run-fixture command bypasses the cross-repository prerequisite only as a pack-local fixture.

#Compose a cross-suite robotics job

Publish Moonfind's typed robotics graph into a MoonBook, open that book in MoonDesk, and switch to MoonFlow:

moon run cmd/main -- publish-robotics-graph \
  /path/to/moonbook \
  fixtures/humanoid-robotics-requests \
  fixtures/humanoid-robotics-capability-catalog.fixture.v1.json \
  2026-07-31T12:00:00Z

The JSON is a 63-stage exact desired graph, never a runnable graph. Every stage uses a canonical operation, schema, authority, and claim class. MoonBook owns wiki requirements and Bookkeeper closure; MoonClaw is the only runtime and runs the wiki and mooncode roles with gpt-5.6-sol. MoonChat owns the owner, stakeholder, and final authority decisions. The MoonCast tail is a complete staged production and cut-editor flow through immutable delivery, client decision, and final evidence—not a monolithic video or publication call.

project_desired_graph shows catalog-derived suggestions for every gap and emits no moonsuite.work-model.v1 object unless the entire selected graph is exactly conformant, healthy, and within its claim ceiling. One missing or expired operation blocks the entire import. MoonFlow executes; MoonGate and MoonLib are support-only; MoonEdit and MoonLeaf remain libraries. MoonFish is outside the robotics book. There is no physical command and no publication stage.

The checked-in capability catalog beside the workflow is deterministic fixture-only data. It is deliberately marked synthetic and is not packaged as production health evidence.

Publication is fail-closed on request materialization. Every one of the 63 stages has exactly one first, pack-owned typed request; MoonFlow appends dependency outputs later. The publisher requires a complete moonfind.robotics-request-materialization.v1 ledger, checks exact stage/schema/path coverage, payload digests, required top-level fields, and registered constants, then copies the requests and graph into the target MoonBook. The bundled request set is schema-valid synthetic scenario data for projection/adapter preflight only. Replace it with an equivalently complete validated bundle for real work.

The catalog and evaluation timestamp are explicit. The publisher preserves bound intent as flow/desired-graph.json, projects it against that exact healthy catalog, writes the resulting moonsuite.work-model.v1 to flow/work-graph.json, and snapshots the exact evaluated catalog as flow/capability-catalog.json for MoonDesk/MoonFlow validation and replay. The timestamp above is valid only inside the short-lived synthetic fixture health window.

The pack contains enforcement contracts and local paths for full-text evidence, reducer-output ingestion, restart recovery, malicious-document isolation, provider licensing, and MoonDesk-to-MoonFlow/MoonClaw execution. Production use still requires real provider contracts, credentials, a deployed isolated extractor, conformant product adapters and explicit authority grants. The desired composition JSON is not itself proof of those external facts.

#Rabbita UI

cd ui/rabbita-moonfind
npm install
npm run build
npm run dev

The workspace answers one primary question: which cross-paper humanoid model idea is worth falsifying next, and why? Papers, bridges, counter-hypotheses, falsifiers, open evidence gaps, and the Bookkeeper state remain visible together.

#
product_home

source

fn product_home(workspace : String, book_id : String) -> String

#
product_id

source

fn product_id() -> String

#
version

source

fn version() -> String

Source Files

moonfind.mbt

Site source Report issue Packages Build queue Skills Statistics

moonfind

#MoonFind

#Product loop

#Run the complete fixture

#Use live arXiv intake

#Run the current Moontown handoff

#Compose a cross-suite robotics job

#Rabbita UI

#product_home

#product_id

#version

Source Files

#
product_home

#
product_id

#
version