Operations
Capability inventory
A maintained map of LMP capabilities, evidence, and interaction boundaries.
Capability inventory
This page is the compact status record for the implementation. It separates what runs locally from what needs an external service or an independent human. It is intentionally more precise than a feature list: a capability is only marked verified when a reproducible test or artifact exists in this checkout.
Status vocabulary
| Status | Meaning |
|---|---|
| Verified locally | A current test, build, or artifact proves the stated bounded behavior. |
| Implemented; evidence pending | The code path exists, but the acceptance evidence is not complete. |
| External gate | Completion depends on deployment, credentials, a network service, or a real maintainer identity. |
| Human gate | An independent person must perform and record the action. |
| Not implemented | The capability is not part of the current product path. |
Local protocol surface
| Capability | Entry point | Current evidence | Interaction? |
|---|---|---|---|
| Mind package validation | lmp skill validate and @lending-mind/skill | Schema tests cover canonical fields, provenance, community authorship, artifacts, and proposals. | No; deterministic local command. |
| Profile and rule contract | evidence.json, rules/manifest.json, orchestrator/profile_fixture_gate.ts | Canonical validation requires positive, negative, and exception fixtures plus a machine-checkable contract for every declared policy file. Each rule records classification, rationale, assertion, scope, remediation, limitations, and evidence; the gate executes all six bundled profiles through the evaluator and compares observed findings with the contract. | No; deterministic local contract. |
| Mind compilation and guidance | lmp skill compile, agent instructions | Compiles a package into visible instructions; does not claim private reasoning access. | No; output is deterministic for the same input. |
| Public-evidence harvesting | orchestrator/mind_compiler.py, orchestrator/mind_crawler.py, orchestrator/mind_harvest_artifact_gate.py | The stable compiler command delegates to one canonical engine that separates textual, implementation, and critique evidence; supports bounded RSS/Atom, public YouTube timed text, and public GitHub discovery; records language choices, bounded dependency-manifest names, digests, and rights metadata; clusters repeated signals; extracts only explicit numeric bounds; records contradictions; and emits a draft-only proposal. The artifact gate rejects raw source fields and promotion/enforcement flags. | No; explicit sources are compiled deterministically, with human review required for contradictions. |
| Signature creation and verification | lmp skill sign, lmp skill verify, lmp skill rotate-key, lmp share | Ed25519 digest/signature tests reject altered bytes; signer rotation requires the current package signature, records a proof signed by the previous key over the exact new public key, re-signs with the new key, and returns the old key for explicit revocation; Rust verification can reject an explicit revoked-key list. | No; requires an explicit command. |
| Workspace onboarding | lmp init, create-lmp, orchestrator/onboarding_benchmark.py | Additive .gitignore handling, brownfield/greenfield mismatch protections, checksum-verified platform runtime downloads, and measured greenfield/brownfield bootstrap timing under a 3,000 ms fail-closed target are tested. | Yes; interactive selection is user-controlled and runtime failure is reported explicitly. |
| Changed-file evaluation | Rust evaluator and lmp evaluate --changed-only | Scope, artifact, redaction, and failure-state tests exist. | No; bounded by the selected workspace and profile. |
| Parser boundary reporting | Rust evaluator artifacts | Each result records analyzed languages, parser adapters, versions, and checked-file count; recognized but unsupported Go, Python, JVM, C-family, SQL, and HCL/Terraform files are explicit skipped checks rather than silent omissions. | No; deterministic and bounded by the selected files. |
| Remediation lifecycle | packages/evaluator/ | Deterministic state transitions enforce retry budgets, repeated-finding stops, security/regression escalation, and explicit human attestation. | Yes; an agent or person supplies remediation and evaluation events. |
| Proposal acceptance | lmp proposal approve | Acceptance validates candidate changes, expected benefit, false-positive risk, required tests, benchmark plans, declared version-bump class, artifact schema, profile identity/version, and rejects blocked or evaluation-error evidence. Promotion enforces the declared patch/minor/major semver floor. | No; explicit maintainer action is required. |
| Autonomy authorization | packages/evaluator/ | A0-A6 policy resolution independently gates commands, package installation, network, and deployment; denied and enabled paths are tested. | No; each action requires an explicit policy decision. |
| External tool execution boundary | packages/evaluator/src/tool-adapters.ts | Adapter plans require explicit authorization, validate adapter identity and executable names, require an existing absolute working directory, optionally confine execution to a workspace root, use non-shell bounded processes, cap timeouts at 120 seconds, pass only a sanitized environment, redact output, and reject boundary violations before spawning. | Yes; the caller must authorize execution and provide a boundary when isolation is required. |
| Bounded fleet evidence | Rust lmp fleet-validate and lmp fleet-merge | Versioned role DAGs reject cycles and unknown handoffs; aggregate/per-role retry, handoff, command, and time budgets are enforced; reports merge in deterministic order and preserve disagreements as first-class evidence. | Yes; external agent hosts submit role reports, while Rust owns validation and merge semantics. |
| Daemon watching | lmpd | Notify watcher, initial evaluation, --once, PID-file --status/--stop/--reload lifecycle controls, clean signal shutdown, and a readiness-aware resource gate are implemented. Created, modified, and removed paths trigger re-evaluation. A real Linux process smoke test verified status, SIGHUP reload, stop, and PID-file cleanup; the retained resource artifact measured 3.926 MiB RSS and 0.0% CPU for 5 seconds in an isolated workspace. In enforced mode, watcher backend errors and event-channel disconnects fail closed. | Yes; reacts to filesystem events and explicit lifecycle signals. |
| MCP compatibility | lmp-mcp, TypeScript adapter, and create-lmp | Stdio JSON-RPC lifecycle, accepted and rejected initialization, state locking after invalid protocol negotiation, malformed input, concurrency, duplicate in-flight ID rejection, strict schemas, typed context, rule explanation, changed-file scope, loop-status evidence, notification forwarding, and the no-command/no-network/no-deployment surface are tested. The TypeScript adapter preserves one persistent Rust session instead of spawning one process per request; onboarding adds tested project-scoped Claude Code, Cursor, Cline, and Roo Code entries without overwriting existing servers or global settings, and writes a copy-ready Claude Desktop entry without editing its global settings. | Yes; host sends JSON-RPC requests and onboarding selects the host. |
| Evidence artifacts | .lending-mind/artifacts | Strict artifact schema records identity, scope, findings, skipped checks, loop transitions, limitations, and privacy; Rust-generated artifacts emit canonical RFC3339 timestamps, hashed workspace identity, Git state, and changed-scope metadata. | No; artifacts are produced by runs. |
| Public npm packages | orchestrator/public_package_gate.ts | Dry-run packing verifies the scoped runtime packages, the unscoped lmp npx launcher, and the unscoped create-lmp bootstrap package contain only their intended runtime files, without tests, build state, or internal workspace dependencies. | No; release gate is deterministic. |
| Optional sandbox | Python orchestrator and Docker checks | Docker capability detection and selected-path checks exist; unavailable Docker is reported as blocked. | No; requires explicit opt-in. |
The documentation portal's Fumadocs search is also static-first: the build
exports an immutable search index at /api/search, and the browser performs
queries locally. It is a static asset route, not a custom application backend
or registry service.
External and human boundaries
| Capability | Required evidence | Current state |
|---|---|---|
| Public registry release | Real maintainer signatures, valid digests, immutable CIDs, and a passing registry gate | External gate: checked-in packages now have locally verified signatures and digests, and canonical/bundled profile drift is checked; the gate still rejects entries without independently retrievable immutable CIDs. |
| Static portal deployment | A successful request to the public portal and registry index | Verified: GitHub Pages serves the portal and /registry.json with HTTP 200. Immutable package manifest retrieval remains a separate distribution gate. |
| OCI/IPFS distribution | Static registry sync is implemented for complete signed packages: the generated public index can describe every package file with an HTTPS URL and SHA-256 digest, and the Rust sync client fetches, verifies, and installs the complete package. Legacy manifest-only entries remain compatible but are explicitly incomplete for full enforcement. Public IPFS pinning, immutable CID verification, and failure recovery remain external release gates. | Local implementation verified: generated registry/public-index parity, complete package-file digests, local static deployment fetch, signature validation, and install. External gate: local/static sync is not a substitute for a public pin. |
| 64-scenario real-world benchmark | Complete technical artifact with Docker, ordinary compiler controls, source verification, versions, digests, and all scenarios passing | Hosted and locally verified: run 34721679791 succeeded for commit 2e9a70459040eef38fd742acade397fbea889ec7; lmp-test-results/real-world-clean-current/real-world-benchmark.json self-binds to its producing commit. Evidence records 64/64 transitions, 128/128 Docker gates, 128/128 ordinary compiler controls, and 15/15 provenance sources. Independent review remains a separate publication gate. |
| Adoption pilot | An independent developer completes install, use, evaluation, remediation, and explains the result | Human gate: automation cannot provide independent adoption evidence. |
| Profile authorship and endorsement | Maintainer authorization or explicit community-archetype labeling and provenance | Governance gate: no profile is treated as an official person's or organization's endorsement without proof. |
How to use this inventory
Before a release, run the local verification suite, inspect the generated artifacts, and then close each external or human row with its own evidence. Do not promote an “implemented; evidence pending” row to verified merely because a test fixture passes. The roadmap and versioning and release notes page defines the release decision boundary.
