Skip to main content
Capabilities

Capabilities, with their current state

See what is shipped today, what remains beta or experimental, and the evidence and limitations behind each claim.

Parallel Subagent Waves

Shipped

Scoped dispatches can fan work out to subagents and collect distilled results while the harness controls concurrency.

Agentic Orchestration Evidence

Signed Action Receipts

Shipped

Receipt-enabled routed mutations attempt to persist Ed25519-signed evidence and report receipt failures loudly.

Verifiable Delivery Evidence

Rule-Based Expert Search Routing

Shipped

A deterministic classifier selects configured experts by question class and fuses provenance-ranked results.

Code Intelligence Evidence

Rubric-Routed Edits

Shipped

A deterministic rubric selects an edit strategy, supports preview, applies scope checks, and records verification outcomes.

Safe Writes Evidence

Structured Option Pickers

Shipped

Typed options can present tradeoffs and a recommendation for explicit operator selection.

Agentic UI Evidence

Progressive-Disclosure Skills Library

Shipped

Operational procedures ship as skills that compatible harnesses load when their task matches.

Packaging & Distribution Evidence

Living Agentic Components in Action

Svelte 5 Runes Live
Open-Question Deliberative Synth

Illustrative Decision Stepper

Recommendation confidence: high
Run a deterministic preflight first

Prefer a proven structural edit when one applies, with an explicit fallback when it does not.

Trust surfaces stay state-labelled

The receipt JSON widget is currently a structure inspector, not a cryptographic verifier. It remains listed below as a scaffold until hashing, key resolution, and Ed25519 verification are implemented.

Capabilities by current state

This catalog includes shipped, beta, experimental, and scaffold work. Only entries labelled Shipped are presented as current product capabilities. Evidence links open the live console surface for each capability — on hosted deploys they require sign-in.

Agentic Orchestration

Coordinate concurrent work with explicit ownership and visible state.

  • Scoped dispatches can fan work out to subagents and collect distilled results while the harness controls concurrency.

    Available: Configured coding-agent harnesses

    Boundaries
    • Concurrency does not replace occupancy, file-conflict checks, or review before integration.
  • One active unsequenced occupancy coordinates a repository and base; contributors join its branch and pull request.

    Boundaries
    • The invariant is per repository and base, not a promise that every branch has an occupancy or pull request.
  • A local release board coordinates pickup of unblocked work through renewable task leases.

    Available: Local cubicle runtime

    Boundaries
    • Expired work can be re-leased; occupancy remains the write lock and exactly-once execution is not implied.
  • Session Health Gate Beta

    Session checks report readiness of the console, MCP builds, sidecars, and declared workspace state.

    Available: Local cubicle runtime

    Boundaries
    • A green sidecar probe does not by itself prove that an already-running harness attachment recovered.
  • Long-Running Role Loops Beta

    Product-grooming and health roles can use bounded in-turn sleep and wake loops for continuing work.

    Available: Role-specific skills and an active agent session

    Boundaries
    • Roles still require an active harness session and explicit operator authority.

Verifiable Delivery

Attach evidence to supported actions and state the coverage boundary.

  • Receipt-enabled routed mutations attempt to persist Ed25519-signed evidence and report receipt failures loudly.

    Available: Configured native MCP receipt paths

    Boundaries
    • Host-native actions and dry-run refusals can fall outside receipt coverage.
    • Retention is not indefinite and remains an operator policy.
  • Approval-aware paths verify signed authority and may narrow an existing scope before execution.

    Boundaries
    • Only tools wired to the authority contract enforce this approval model.
  • Tool-Definition Pinning CLI Experimental

    An experimental CLI captures MCP tool-definition hashes and can verify a signed pin against a later live surface, a trusted Ed25519 signer, and source, artifact, configuration, and protocol bindings.

    Available: Manual local CLI

    Boundaries
    • Capture can still produce a loudly marked unsigned, non-authoritative file; verification rejects it.
    • Automatic startup enforcement is not wired.
    • Trust-store file provenance and permission checks, plus complete dirty-worktree and runtime-dependency binding, remain open independent-audit gaps.
  • MCP Interface Snapshots Experimental

    Root CI and full-validation gates compare committed tools/list snapshots with freshly spawned native MCP tool catalogs and fail when a declared snapshot or build is missing.

    Available: Root CI and full-validation gates with built native MCPs

    Boundaries
    • Task-support declarations and native-server directory completeness are not yet enforced by the inventory until the follow-up gate lands.
  • Article 12 Evidence Export Experimental

    A paginated local JSON endpoint exports selected signed-receipt metadata. Integrity, envelope binding, and signer trust are reported as separate flags. Not a legal-compliance determination.

    Available: Local runtime only

    Boundaries
    • Not a compliance certification or legal conclusion. Qualified legal confirmation remains separate.
    • Only receipt-plane operations appear. Host-native actions outside MCP are not covered.
    • Retention (Article 19) is not demonstrated by this exporter.

Code Intelligence

Select the right retrieval paths and preserve their provenance.

  • A deterministic classifier selects configured experts by question class and fuses provenance-ranked results.

    Available: Local native search MCP

    Boundaries
    • The current classifier is rules-based; trained or LLM routing is not shipped.
  • Selected semantic, lexical/history, symbol, graph, and documentation adapters can run in parallel and return per-item provenance.

    Boundaries
    • Routing is question-specific, so not every expert runs on every query.
    • Coverage depends on configured, healthy, and current indexes.
  • Explicit wrapped reads can return outlines and neighboring context under a token budget selected for the reading intent.

    Boundaries
    • Bare host reads and unindexed files do not automatically receive wrapped context.

Safe Writes

Route edits, record verification, and make rollback failures loud.

  • A deterministic rubric selects an edit strategy, supports preview, applies scope checks, and records verification outcomes.

    Boundaries
    • Available strategies and verification commands depend on repository configuration.
  • Affected files are snapshotted and individually replaced atomically by default; failed batches attempt scoped rollback.

    Boundaries
    • Cross-file visibility is not transactional.
    • Rollback can fail and is then reported as a classified error.
  • Writes run the configured in-scope verification rungs and retain per-rung pass, failure, or skip reasons.

    Boundaries
    • Rungs without a configured command or outside the selected ladder are explicitly skipped.

Memory Plane

Persist selected evidence and proposals beyond one session.

  • Verified, receipt-linked episodes can be distilled into human-reviewable draft improvement proposals.

    Boundaries
    • Distillation does not create tickets automatically; approval and ticket mutation are separate.
  • Instrumented sessions can persist utilization, quality, compaction, and role metadata for later analysis.

    Boundaries
    • Reporting is not automatic for every harness or every session.
  • Convention Promotion Workflow Beta

    An operator-invoked skill helps promote repeated instructions into the appropriate law, skill, script, or research owner.

    Boundaries
    • There is no autonomous compiler that detects and publishes every repeated instruction.

Delivery Analytics

Analyze the events the platform actually records.

  • Instrumented tool events support error-rate, throughput, token, and latency summaries.

    Boundaries
    • Uninstrumented calls are absent and cost is observable only when emitters report it.
  • Persisted session reports expose utilization, compaction, role, and quality distributions.

    Boundaries
    • Results inherit emitter coverage and endpoint scan limits.
  • Occupancy lifecycle events provide cycle, dwell, closeout, and fix-forward proxies for delivery flow.

    Boundaries
    • Merge, deployment, and incident events are not yet fully recorded or joined.
  • Context Utilization Monitoring Beta

    Reported session checkpoints can surface utilization, compaction, truncation, and quality notes.

    Boundaries
    • Only sessions that emit checkpoints are observable.

Fleet Operations

Keep registered repositories, workspaces, and shared services visible.

  • A registry view lists enrolled repositories, branches, index status, and known heads.

    Boundaries
    • It is not a live process, gate-state, or occupancy view of every checkout on the machine.
  • Release Manifest Generator Experimental

    An on-demand CLI walks an explicitly selected immutable commit tree and emits schema-versioned repository version and content-digest metadata for release tooling.

    Available: Manual release CLI

    Boundaries
    • The selected commit is not yet bound to an expected ref or release channel.
    • External output-path safety, artifact publication, and deployment enforcement are not yet wired.
  • Repository Drift Checks Beta

    Cubicle-open checks can fast-forward eligible clean clones and report dirty or diverged repositories for operator action.

    Boundaries
    • Dirty, stashed, or diverged work is reported rather than automatically rewritten.
  • Supervised Console Lifecycle Beta

    The shared local console is launchd-managed and records one owning cubicle for lifecycle decisions.

    Available: Local macOS runtime

    Boundaries
    • Recovery is macOS/launchd-specific and requires bootstrap, datastore, and supervisor participants to use the same installed ownership-lock protocol.

Agentic UI

Use structured interaction patterns and label prototypes honestly.

  • Typed options can present tradeoffs and a recommendation for explicit operator selection.

    Boundaries
    • The caller still owns the quality and completeness of the options presented.
  • Receipt Structure Inspector Scaffold

    A prototype UI accepts receipt-shaped JSON and demonstrates fields intended for a future browser verifier.

    Boundaries
    • It does not currently perform payload hashing, public-key resolution, or Ed25519 verification.
  • A scaffold documents the consent, origin validation, and bundle-pinning requirements for a future sandboxed app host.

    Boundaries
    • No interactive MCP Apps widget is mounted today.
  • Configured agents can emit optional spoken progress updates through the local voice MCP.

    Available: Local runtime with a configured voice provider

    Boundaries
    • Provider credentials, audio output, and user preferences must be configured.

Research Engine

Keep product research, experiments, and evidence close to implementation.

  • Curated engineering research is browsable beside the product surfaces it informs.

    Boundaries
    • Each paper carries its own evidence quality, date, and applicability boundaries.
  • Configured strategy comparisons can persist runs, receipts, rewards, and Thompson-sampling arm selections.

    Boundaries
    • Experiments require configured engines, fixtures, and sufficient observations.
  • Case-study pages connect bounded product claims to dated implementation artifacts.

    Boundaries
    • A case study supports only the scope and evidence stated on its page.

Integrations & Platform

Expose protocol-level capabilities while naming current providers as data.

  • Git Hosting Integration Beta

    The current GitHub backend supports repository, pull-request, issue, comment, and check-run reads plus approval-gated comment mutations.

    Available: Current provider: GitHub

    Boundaries
    • Receipt settlement applies to supported mutations, not every read or the entire pull-request lifecycle.
  • The current Linear backend supports backlog inventory, specification linting, structured issue creation, comments, and grooming operations.

    Available: Current provider: Linear

    Boundaries
    • Live mutations require configured credentials and the product MCP approval path.
  • Cloud Runtime Hosting Beta

    The current Cloudflare backend provides Workers deployments and D1 persistence alongside local development.

    Available: Current provider: Cloudflare Workers and D1

    Boundaries
    • Local sidecars remain the primary development path; cloud capability varies by package and environment.
  • The current KYM backend can publish and resolve configured agent-card identifiers for supported sessions.

    Available: Current provider: Know Your Model

    Boundaries
    • Identity-card binding is not yet universal across every agent and runtime.

Packaging & Distribution

Install deliberately and keep harness-specific adapters visible.

  • Operational procedures ship as skills that compatible harnesses load when their task matches.

    Boundaries
    • Skill discovery and context cost depend on the active harness integration.
  • Generated configuration templates attach the same native MCP packages to supported coding-agent harnesses.

    Boundaries
    • Harness capabilities and reconnect behavior differ; configuration parity does not imply runtime parity.
  • A canonical registry records available server packages, transports, installation steps, and purpose.

    Boundaries
    • Generated harness configuration and runtime health are separate from catalog membership.

Ready to experience AI-native product development?

Open an authorized workspace or launch the local runtime to explore shipped capabilities.

Sign in to Ozzy Dev