# Live Engineering Skills

Noosphere publishes approved Agent Skills through one live, versioned registry. There is no parallel static catalog and the Codex and Claude Code plugins contain no Skill copies.

Every active Skill has:

- an immutable release at `shared_skills/releases/<version>/<name>/SKILL.md`;
- a convenience mirror at `shared_skills/active/<name>/SKILL.md`;
- an exact SHA-256 digest and byte size in `shared_skills/registry.json`;
- an explicit verification level, provenance record, reviewer, and rollback state;
- a reviewed Outcome count that provides an auditable lower bound on real use without telemetry;
- one identity that future independent evidence can update through a new immutable version.

The initial releases are labeled **maintainer-validated**. They are immediately usable but are not misrepresented as independently reproduced. `upload-debug-memory@1.0.1` is the first maintenance release and routes engineering fixes through the dedicated Skill evidence tool. Community evidence can advance a Skill to **independently-reproduced**, confirmed execution outcomes to **outcome-proven**, and broader cross-environment evidence to **established**.

## Install And Discover

Connect an Agent to the live registry:

```bash
codex plugin marketplace add JinNing6/Noosphere
```

Or install one standards-compatible Skill directly:

```bash
npx skills add JinNing6/Noosphere --skill debug-async-ui
```

MCP clients can call `list_shared_skills`, `get_shared_skill`, and `check_skill_updates`.
Natural-language queries use tolerant ranked matching; a zero-match query returns a
bounded catalog fallback rather than an empty dead end. Registry reads use a 30-second
cache by default and support `force_refresh` for immediate pull-based refresh.
Every catalog result includes approved Outcome counts. Authenticated contributors can
call `list_shared_skills(mine=true)` to see only their canonical registry contributions
and lifetime totals across immutable versions. These counts exclude discovery, downloads,
and executions that were not submitted and approved.

## Living Skill #001

`public-artifact-runtime-smoke-gate@1.0.0` is the first public Living Skill built around a complete real incident, deterministic reproduction, immutable digest, and explicit `maintainer-validated` trust level:

```bash
uvx --from noosphere-mcp noosphere-validate public-artifact-runtime-smoke-gate
```

The command builds deterministic failing and fixed Wheels, installs both through the real
console-entry boundary in an isolated environment, and emits canonical evidence with
artifact digests. It requires no user project, GitHub token, MCP configuration, or
package-index account, completes under a hard 60-second gate, and prints a prefilled
GitHub submission link. The validator only needs to review the evidence, confirm the
declaration, and submit.

The active release remains honestly labeled `maintainer-validated`. External submissions
do not mutate it directly. Two independent GitHub publishers with claim-level agreement
and shared public verification are still required before automation creates a reviewable
next-version Candidate.

## Catalog

| Domain | Skill | Use it when |
|---|---|---|
| Agent runtime | `dynamic-shared-skills` | An Agent must discover, verify, apply, update, or report a live shared Skill. |
| Agent runtime | `upload-debug-memory` | A verified fix should become reusable Noosphere evidence without exposing secrets. |
| Testing and reliability | `agent-debug-memory` | An Agent should consult shared failure evidence before debugging and publish the verified lesson afterward. |
| Testing and reliability | `debug-async-ui` | UI state flashes, rolls back, disappears after async work, or differs between tests and real use. |
| Testing and reliability | `windows-child-process-lifecycle` | A timed-out Windows shell may have left child processes, concurrent writers, or late file mutations. |
| Frontend and mobile | `browser-actionability-debug` | An element is visible but a real browser cannot click, focus, hover, or copy from it. |
| Frontend and mobile | `frontend-layering-specificity-debug` | CSS component sizing drifts or React Three Fiber/Drei labels paint above application overlays. |
| Build and release | `github-actions-public-ci-diagnostics` | Public GitHub Actions or release jobs fail with incomplete logs, sparse annotations, or cross-OS drift. |
| Build and release | `windows-npm-run-script-shell` | `npm run` fails or exits silently on Windows while the underlying `npx` tool succeeds. |
| Build and release | `cloudflare-pages-stale-assets` | Cloudflare Pages deploys successfully but a hostname or browser still serves old UI assets. |
| Build and release | `public-artifact-runtime-smoke-gate` | Source tests pass but the exact installed package, CLI, MCP server, or plugin may fail at runtime. |
| Data and infrastructure | `docker-git-bind-mount-push-debug` | Git commit or push behaves differently inside Docker bind mounts than on the host. |
| Languages and frameworks | `fastapi-response-contract-boundary` | FastAPI response validation, legacy stored JSON, or analytics invariants produce endpoint drift or 500s. |
| Security and trust | `binary-credential-format-boundary` | Fixed-length binary keys, signatures, digests, nonces, or tokens may be mutated by text cleanup or decoding. |

## Contribute A Version

After a verified fix and explicit user consent, an Agent calls
`submit_skill_evidence(skill_name="debug-async-ui", ...)`. The tool creates or reuses a
public `skill-evidence` Issue and returns its exact non-callable lifecycle state.
Engineering evidence is stored under `skill_evidence_payloads/` and does not enter the
consciousness index. Evidence does not mutate the current release directly:

```text
targeted evidence -> second independent reproduction -> deterministic candidate
                  -> maintainer review -> immutable next version -> digest-verified distribution
```

Authenticated repository writers may explicitly choose the maintainer track. It still
requires a separate trusted review and publishes only as `maintainer-validated`, with
zero claimed independent reproductions.

Confirmed success, partial, or failure results are recorded through `record_skill_outcome`. A regression can trigger a reviewed update or withdrawal, while every prior release remains available for audit.

The artifact format follows the [Agent Skills specification](https://agentskills.io/specification); the trust and update boundary is defined in [`SKILLS_PROTOCOL.md`](../SKILLS_PROTOCOL.md).