Crystallize agent behavior into replayable skills.
A skill workshop and proof harness for the [&] Stack. Teach an agent once, capture every observable step, and turn the trace into a signed, replayable SkillBundle that PRISM can score and that any [&] body (browser / host / simulator) can re-execute.
The [&] Stack has eleven protocols (OS-001..OS-011), nine MCP servers, and roughly five hundred Elixir / TypeScript tests showing the parts work in isolation. What it didn't have — until now — was a user-facing artifact that exercises the whole pipeline and produces something a person actually wants. Workbench is that artifact.
A SkillBundle is valuable to the user (a reusable, replayable thing they can share). The same JSON is valuable to PRISM (it satisfies the EvidenceBundle contract from STACK_ARCHITECTURE_GAP_REVIEW.md Finding 8). Two audiences, one file.
See `docs/spec/README.md` for the authoritative spec.
workbench/
├── docs/spec/ # authoritative v0.1 spec + 6 JSON schemas
├── fixtures/scenarios/ # 5 canonical fixture scenarios
├── app/ # SvelteKit static app (Cloudflare Pages)
│ ├── src/lib/
│ │ ├── types.ts # TypeScript mirrors of every schema
│ │ ├── hash.ts # canonical JSON + sha256 + bundleContentHash
│ │ ├── storage.ts # IndexedDB + sessionStorage
│ │ ├── openrouter.ts # BYO-key chat client (via /api/llm/openrouter)
│ │ ├── trace.ts # InteractionTrace builder + stateHash
│ │ ├── ingest/ # Claude Code transcript → trace / manifest / bundle (pure)
│ │ ├── mcp/ # streamable-HTTP MCP client + per-server wrappers
│ │ └── bundle/ # builder / verify / replay
│ ├── src/routes/ # +layout, +page (landing), record, teach, skills, compare, prism, import
│ └── tools/ # ingest-transcript.ts — the only reader of a real transcript
└── functions/
├── llm/openrouter.ts # CF Pages Function: OpenRouter pass-through
├── mcp/ # CF Pages Functions: delegatic / body-browser / prism MCP proxies
└── prism/leaderboard.ts # CF Pages Function: PRISM read-only leaderboard proxyv0.4.0-alpha — Record. Workbench no longer needs you to drive an agent through a browser textarea to produce a bundle. Drop a Claude Code session transcript (~/.claude/projects/<project>/<session-id>.jsonl) on /record and it becomes a scored SkillBundle: streamed and indexed in the tab, segmented at each user turn, verified against the same six gates, and saved to the same library. No model call, no API key, nothing uploaded.
The parser (app/src/lib/ingest/) is pure and browser-safe; the CLI at app/tools/ingest-transcript.ts is the only thing that reads a transcript off disk. Findings, defects fixed, and measured numbers are in `docs/INGEST_SPIKE.md` — including three bugs the real transcripts surfaced that the fixtures never did (a canonicalize crash on undefined-valued keys, and two live secret-leak paths in the redaction profiles).
Also in v0.4: checkPhaseOrder now closes a cycle when feedback arrives. It previously read every tool call in a loop as a learn → act violation — 124 of them on a 124-tool-call segment — which measured its own segmentation rather than the trace.
v0.3.0-alpha — Invariant Arithmetic v0.3 is wired into the verifier and replay surface. Every proof gate's verdict is now computed by an InvariantArithmetic.consume(value, requirements) call (app/src/lib/ia/), and every gate carries law (e.g. 'L12/L14') and invariant_family (topological | spatial | temporal | governance) annotations. Replay surprise signals carry the same projection. The three relevant JSON schemas (PROOF_RESULT, REPLAY_REPORT, EVIDENCE_BUNDLE) declare the optional law / invariant_family / ia_substrate / invariants fields.
v0.2.3 baseline — authority gate supports delegatic-mcp HMAC/TTL verification when enabled, body-browser-mcp is wired for browser-body teach flows, and the PRISM page stages SkillBundle hand-off by content_hash.
Persisted bundles are now self-describing: verifyBundle returns an ia_substrate block (version + laws_exercised + families_exercised projected from the gates' annotations) which the teach flow stamps into bundle.proof.ia_substrate before re-sealing the content_hash. The compare flow leaves it ephemeral (no persistence).
152 / 152 vitest tests pass as of 2026-08-11 (132 through v0.3 + 18 ingest + 2 rewritten phase-order tests). npm run check reports 27 TypeScript/Svelte diagnostics; 26 predate v0.4 and the one added is the same "no declaration file for uuid" error that trace.ts, builder.ts, replay.ts and charter.ts already carry — @types/uuid would clear five at once. Do not present the app as typecheck-clean until those are resolved.
| Surface | Status |
|---|---|
| JSON schemas (6) | ✅ complete + IA-annotated |
| Fixture scenarios | ✅ 5 / 5 |
| SvelteKit app shell | ✅ scaffolded |
| Teach (simulator) | ✅ end-to-end |
| Fixture chooser | ✅ 5 / 5 (AC #3) |
| Crystallize bundle | ✅ working |
| Bundle verify | ✅ 6 / 6 gates, IA-derived (v0.3) |
| Bundle replay | ✅ deterministic, IA-annotated surprises (v0.3) |
| Redaction profiles | ✅ none / transcript_pii / full |
| Compare tab (2 models) | ✅ AC #13 (live LLM) |
| Live manifest validation | ✅ AC #6 |
| Cross-session re-import | ✅ AC #8 |
| Body-browser teach | ✅ v0.2.2 (perceive + encode_state edges) |
| Delegatic authority | ✅ v0.2.1 (put_policy → authorize → verify) |
| PRISM connectivity | ✅ v0.2.3 (leaderboard + list_systems + bundle hand-off) |
| Invariant Arithmetic | ✅ v0.3.0 (15 laws, 1000-trial property tests, consume-derived gates) |
| Record (transcript ingest) | ✅ v0.4.0 (streamed, segmented, 6/6 gates, redacted-on-save) |
| Phase-order check | ✅ v0.4.0 (cycles close at feedback; authorize-after-act still caught) |
cd app
npm install
npm run dev # http://localhost:5180The dev server expects an OpenRouter key in sessionStorage (set via the landing page). For local development without a real proxy, set up a Vite middleware that forwards /api/llm/openrouter/* to https://openrouter.ai/api/v1/*, or configure Cloudflare's wrangler pages dev to mount functions/.
Cloudflare Pages, zero-budget tier. The static SvelteKit build lives under app/build/; the Pages Function under functions/ is auto-discovered.
docs/spec/README.md — Workbench v0.1 spec (authoritative)
../AmpersandBoxDesign/SPEC.md — [&] Protocol
../PULSE/schemas/pulse-loop-manifest.v0.1.json — PULSE manifests
../PRISM/docs/ — PRISM benchmark contract
../STACK_ARCHITECTURE_GAP_REVIEW.md Finding 8 — the Evidence Bundle gap this closes