epic(harness): multi-model Agent/Skill operating system and experience flywheel

## Goal

Build a model- and machine-independent development Agent/Skill operating system for WorkspaceX. GitHub issues/PRs and coord-service remain the authority; Claude Code, Codex and future clients are interchangeable execution surfaces.

## Current audit baseline — 2026-08-04

- `origin/main` has 7 persistent Directory-backed roles but only 8 generic generated subagent specs.
- `.harness/agents/registry.yaml` and the live Directory duplicate identity facts; runtime roles have no generated Claude/Codex specifications.
- every live Directory agent reports `lifecycle=active`, `parent=null`, and no heartbeat timestamp, so registration is not proof of live presence or hierarchy.
- 18 repository Skills exist, all under 500 lines and using `name`/`description`, but none records license, compatibility, version/provenance, owner or validation status.
- the only module knowledge Skill is `mod-_template`; actual module experience is not being accumulated in `mod-*` Skills.
- `PROJECT.md` remains a template, so project-specific deployment/module facts are scattered.
- generated Claude/Codex subagent drift is checked in CI, but there is no Agent Skills spec validator, provenance catalog, permission/risk tier, trigger/eval suite or upstream upgrade gate.

## Architectural boundaries

1. Separate persistent Directory roles from ephemeral subagents.
2. Separate developer harness Skills from product/org Skills; promotion between them is explicit and provenance-preserving.
3. Repository YAML/Skill folders are declarative source; Directory ULIDs and tokens are deployment state. Never commit tokens.
4. GitHub is the durable work/evidence bus; coord-service is the task/lease/presence authority; platform-native messages are notifications only.
5. External Skills enter quarantine first and never execute bundled scripts during review.

## Upstream reference set (pin on implementation)

- Agent Skills spec: `agentskills/agentskills` @ `27a9f0c075e876ad632fc2e88b8866c5dc8ca15c` (Apache-2.0/CC-BY-4.0).
- Superpowers @ `44c9b2d6e889982ac18c27d05a19fefe335194e1` (MIT): selectively adapt systematic-debugging, test-driven-development, verification-before-completion, receiving-code-review, and writing-skills pressure evaluation.
- GitHub Spec Kit @ `ab468c4db7d760cb15238a3718d85a2803921837` (MIT): reference artifact consistency and cross-agent generation; do not replace the existing harness or install its CLI by default.
- AGENTS.md @ its reviewed pinned commit (MIT): keep scoped directory instructions portable.
- Anthropic Skills @ `b29e7cf65e5cb78a5ac33d582270551bc74a14eb` and OpenAI Plugins @ `11c74d6ba24d3a6d48f54a194cd00ef3beea18f9`: directory-by-directory reference only; each selected directory needs its own license review.
- LangChain Deep Agents @ `f4cc5160c75eb44e8ddee8b049048690ea0f8616` (MIT): reference isolated subagent context, explicit permissions and memory namespaces; no runtime adoption in this epic.

## Waves

### A — identity and portable roles

- reuse #436 for single-source Agent specifications, Claude/Codex generation, Directory ULID resolution and uniqueness.
- define persistent versus ephemeral role criteria and a minimal role/reviewer topology.
- make live presence, parent/reports-to and capability drift mechanically observable.

### B — Skill supply chain

- add a catalog with upstream repo, commit, subpath, archive/content hash, license, local owner, risk tier, declared tools/network/filesystem, evaluation status and last-reviewed date.
- quarantine remote sources; adapt into `.agents/skills` only after review.
- validate Agent Skills naming/frontmatter/references plus WorkspaceX-specific permissions, provenance and no-secret rules.

### C — verified engineering core pack

- adapt only the five selected Superpowers methods to WorkspaceX issue/lease/signoff semantics.
- add positive trigger, negative trigger, pressure/rationalization and cross-model behavior tests.
- reject upstream parallel-dispatch or auto-merge semantics that bypass coord-service, one-issue/PR or independent review.

### D — experience flywheel

- every issue/PR/handoff may emit structured lesson candidates.
- curator routes a lesson to module Skill, general Skill, instruction, ADR or test; raw chat is never the durable memory.
- promotion requires a reproduced incident or repeated pattern, owner review, pressure test and provenance.
- add deprecation, supersession and periodic load-bearing audits so Skills can become smaller as models improve.

### E — product boundary

- reviewed developer Skills are not automatically visible to WorkspaceX organizations.
- any product Skill/Agent pack uses the explicit admin import path, immutable versions and separate design signoff.

## Completion

- child issues follow one issue/branch/PR and are dispatched to workers, not coord-main.
- at least Claude Code and Codex run the same canonical role/Skill fixtures and produce equivalent policy outcomes.
- a malicious/unlicensed/over-privileged Skill and a stale Directory mapping are mechanically rejected.
- the first module Skills contain evidence-backed lessons from current Auth, Chat, AI runtime, platform and harness work.

Start with child issue #436, then inspect .harness/agents/registry.yaml, .agents/skills, and PROJECT.md alongside the stated upstream commits and existing CI drift checks. Break the epic into separately scoped child issues; completion requires equivalent Claude Code/Codex policy outcomes, mechanical rejection of unsafe Skills and stale Directory mappings, and evidence-backed module Skills.

Written by the indexing model from the issue text.