@archmax-ai/harness 0.0.0

package/dist/authoring-skill/archmax-harness/references/backend-integration.md

@@ -0,0 +1,626 @@
+# Wiring @archmax-ai/harness into a backend — integration reference
+
+How to embed the `@archmax-ai/harness` package in a service so a worker can
+**execute** workflows and an API/frontend can **visualize** run state. All names
+below are the real public API (`src/index.ts`); confirm against the installed
+package's `dist/index.d.ts`.
+
+## Contents
+
+- [Install (registry & local tarball)](#install)
+- [Public exports](#public-exports)
+- [createAgent options](#createworkflowagent-options)
+- [Agent](#workflowruntime)
+- [Running a workflow](#running-a-workflow)
+- [Human-in-the-loop](#human-in-the-loop)
+- [Lifecycle events (live visualization)](#lifecycle-events)
+- [Session artifacts (historical visualization)](#run-artifacts)
+- [Model configuration](#model-configuration)
+- [Custom backends](#custom-backends)
+- [Cases against your agent](#cases-against-your-agent)
+- [Suggested worker / API / frontend split](#suggested-split)
+
+---
+
+## Install
+
+**Registry** (once published): `npm i @archmax-ai/harness`.
+
+**Local / pre-publish** — install exactly what the registry would ship (a real copy
+in `node_modules`, not a symlink), so behavior matches production:
+
+```bash
+# in the harness repo
+npm run build
+npm pack --pack-destination /tmp        # → /tmp/archmax-ai-harness-<version>.tgz
+
+# in the consuming service
+npm  add /tmp/archmax-ai-harness-<version>.tgz
+# or pnpm --filter <pkg> add /tmp/archmax-ai-harness-<version>.tgz
+```
+
+The published tarball contains only `dist/` + `README.md` (the package `files`
+allowlist), so `dist/index.d.ts` types travel with it — a shared types package can
+re-export the run-artifact/event shapes for an API/frontend without importing the
+runtime. Because the version rarely changes during development, `remove` then `add`
+(or prune the store) so a repack of the same version actually reinstalls.
+
+**This skill ships inside the package** (staged into `dist/authoring-skill/archmax-harness/`):
+resolve it from the install via the `BUNDLED_AUTHORING_SKILL_DIR` export — the
+`dist/authoring-skill` directory holding `archmax-harness/`, so this file is
+`<BUNDLED_AUTHORING_SKILL_DIR>/archmax-harness/SKILL.md` — e.g.
+`npx skills add ./node_modules/@archmax-ai/harness/dist/authoring-skill/archmax-harness`.
+(`BUNDLED_SKILLS_DIR` is a deprecated alias with the same value, kept one
+release.) Prefer this over
+syncing a copy from a source checkout, so the skill can never drift from the
+installed SDK's validator. Tooling that needs the install root without importing
+the runtime can use `require.resolve("@archmax-ai/harness/package.json")` (exposed in the
+`exports` map).
+
+## Public exports
+
+From `@archmax-ai/harness` (`src/index.ts`):
+
+| Export | Kind | Purpose |
+| --- | --- | --- |
+| `createAgent` | fn | Assemble a runnable workflow-governed agent |
+| `Agent` | type | Returned runtime (`graph`, `decide`, `deliver`, `resolveSession`, `runs`, `emitRunArtifacts`, `dispose`, `resolveTrigger`) |
+| `CreateAgentParams` | type | Assembly options |
+| `DecideOutcome` | type | Return of `decide()` |
+| `DeliverOutcome`, `TriggerDelivery` | type | Argument / return of `deliver()` |
+| `WorkflowLoadError` | class | Thrown when the workflow spec (`workflow.yaml`) is missing/invalid (fails closed) |
+| `DEFAULT_WORKFLOW` | const | `"order-lookup"` |
+| `BUNDLED_AUTHORING_SKILL_DIR` | const | The packaged `dist/authoring-skill` directory holding `archmax-harness/` (`<dir>/archmax-harness/SKILL.md`). `BUNDLED_SKILLS_DIR` is a deprecated alias with the same value. The example workspace is not shipped in `dist/`; clone the repo to use it. |
+| `WORKFLOW_STATUSES` | const | `running` \| `completed` \| `rejected` \| `awaiting_decision` (`awaitingDecision`) \| `awaiting_input` (`awaitingInput`) |
+| `isFinished`, `classifyStatus` | fn | The open/finished partition — a parked session is **open** |
+| `WorkflowLifecycleEvent`, `WorkflowEventHandler` | type | Event stream types |
+| `createWorkflowEventEmitter`, `renderEventLine` | fn | Event emitter + console formatter |
+| `createWorkspaceContext` | fn | Compose the workspace backends (session zone + mounts + authoring backend) |
+| `createFilesystemSessionStore`, `createBackendSessionStore`, `createMemorySessionStore` | fn | Run-store factories (physical storage for the session zone) |
+| `SessionStoreRequiredError`, `SessionStoreCapabilityError` | class | Assembly / capability errors for run storage |
+| `workflowStateSchema`, `workflowPaths`, `sessionPaths`, `runArtifactPaths` | schema/fn | State schema + path helpers |
+| `runTests`, `RunTestsOptions`, `CaseResult` | fn/type | Run a workflow's cases — [against your own agent](#cases-against-your-agent) |
+
+## createAgent options
+
+All optional (`CreateAgentParams`, `src/assembly/index.ts`):
+
+| Option | Type | Notes |
+| --- | --- | --- |
+| `rootDir` | `string` | Workspace root for the default filesystem backend. Defaults to cwd. |
+| `workflow` | `string` | Named workflow under `workflows/<name>/`. |
+| `trigger` | `TriggerInput` | `{ id, args? }`; defaults to `{ id: "manual" }`. Selects the start state. |
+| `model` | `BaseChatModel` | Defaults to the env-configured OpenAI-compatible model. Inject your own to bypass env. |
+| `backend` | `BackendProtocolV2` | Serves the authored workspace. Defaults to a `FilesystemBackend` over `rootDir`. |
+| `authoring` | `BackendProtocolV2` | Serves the **authoring plane** — `workflows/**` (machine specs including the rubrics their states declare, `WORKFLOW.md`, `hooks/`, `tests/`) — to the runtime only. Never a route in the agent's workspace, so no tool call or PTC call can reach it. Defaults to `backend` when you supply one, else a `FilesystemBackend` over the resolved root. Handing it to a **writable** mount throws `AuthoringBackendExposedError`; a read-only mount may share it. |
+| `mounts` | `Record<string, MountSpec>` | Authored mounts as a `CompositeBackend` route table: key → backend, or `{ backend, readOnly, governed, searchable }`. Trailing slash = directory mount; no slash = exact-path file mount (`"/AGENTS.md"`). Read-only unless declared otherwise. `governed: true` hands the spec's `mounts` block the decision of **which states** may reach it (closed by default, like a skill bundle) and whether each may write there; a mount not so declared is visible in every state exactly as today, so a table that marks nothing behaves unchanged. `readOnly` is a **ceiling** the spec cannot lift: a grant may narrow a writable mount to reads in a state, never open a read-only one. A key may be several segments deep (`"/catalogs/eu/"`); its **first** segment is what is reserved, so `/workflows/x/` fails as `/workflows/` does. Omitted on the default backend: `defaultMounts(rootDir)` (the conventional table — `/skills/`, `/.platform/`, `/AGENTS.md`, none governed — extend with `{ ...defaultMounts(root), "/templates/": … }`). Omitted with a custom `backend`: nothing authored is served; expose subtrees with `mountSubtree(backend, "skills")`. A key colliding with a run area — or naming the authoring-plane prefix (`/workflows/`) — throws `MountCollisionError`. `searchable: false` declares a mount the runtime never searches on its own initiative: a root-wide `grep`/`glob` fans out over the other mounts only, while a search the agent addresses at the mount (or a path inside it) still reaches its backend and returns the backend's own answer — matches or `{ error }` — verbatim. Use it for a backend that must refuse searches (a remote folder served live). Listing and reading are unchanged, the prompt marks it "browse only", names arrive as `MountPrefixes.unsearchable`, and it is ignored on a file mount. |
+| `sessionStore` | `SessionStore` | Physical storage for run state (checkpoints, artifacts, the per-session `scratchpad/`). Zero-config default: `createFilesystemSessionStore({ dir: "<rootDir>/sessions" })`. Required with a custom `backend` (else `SessionStoreRequiredError`). Build with `createBackendSessionStore({ backend, prefix? })` for S3-style storage or `createMemorySessionStore()` for ephemeral runs. |
+| `systemPrompt` | `string` | Appended after workspace `AGENTS.md`. |
+| `checkpointer` | `BaseCheckpointSaver` | Custom-adapter escape hatch; takes precedence over the session store for checkpoint persistence. Defaults to a durable `BackendCheckpointSaver` writing through the session store. Use `MemorySaver` for ephemeral runs. |
+| `middleware` | `AgentMiddleware[]` | Extra middleware appended after workflow instrumentation. |
+| `onEvent` | `WorkflowEventHandler` | Subscribe to all lifecycle diagnostics; suppresses console output when set. |
+| `promptCache` | `PromptCacheOptions` | `{ enabled?, ttl?: "5m" \| "1h" }`. Provider prompt caching for the stable prefix; on by default. Falls back to `settings.prompt_cache`, then `ARCHMAX_PROMPT_CACHE`/`_TTL`. |
+| `pricing` | `PricingTable` | USD per 1M tokens, keyed by model id (`default` matches any): `{ input?, output?, cacheRead?, cacheWrite? }`. Looked up by the id the response reported, else the id the runtime was configured to run — so a keyed table prices a workload behind a proxy that echoes no model back. Input rate applies to the input tokens the cache did not serve. Makes `costUsd` appear on usage events, run metadata, and the CLI footer. Falls back to `ARCHMAX_PRICE_*`. Unpriced → tokens only. |
+
+Fails closed: throws `WorkflowLoadError` if the named workflow's `workflow.yaml`
+is missing or has no valid machine spec (never silently runs ungoverned; a
+`WORKFLOW.md` carrying frontmatter with no `workflow.yaml` — the removed
+single-file layout — is a load error with migration guidance).
+
+## Agent
+
+```ts
+interface Agent {
+  graph: /* LangGraph runnable */;
+  workflow: string;
+  runtimeContract: ResolvedRuntimeContract;
+  resolveTrigger(trigger?: TriggerInput): ResolvedTrigger;
+  emitRunArtifacts(runId: string, trajectory: Trajectory): Promise<string | null>;
+  runs: {                                // run handles over the configured session store
+    list(): Promise<SessionSummary[]>;    // [] when the store/checkpointer can't list
+    get(sessionId: string): Promise<SessionSummary | null>;
+    delete(sessionId: string): Promise<boolean>;  // throws SessionStoreCapabilityError
+  };
+  decide(sessionId: string, resolution: DecisionResolution): Promise<DecideOutcome>;
+  dispose(sessionId: string): void;
+}
+```
+
+## Running a workflow
+
+```ts
+import { createAgent, WORKFLOW_STATUSES } from "@archmax-ai/harness";
+
+const agent = await createAgent({ workflow, onEvent, workspace: { rootDir } });
+const sessionId = `run-${runId}`;
+
+try {
+  const result = await agent.invoke(
+    { messages: [{ role: "user", content: prompt }] },
+    { configurable: { thread_id: sessionId }, recursionLimit: 50 },
+  );
+  // result: { messages, status?, workflowState?, pendingDecision? }
+  if (result.status === WORKFLOW_STATUSES.awaitingDecision) {
+    // parked — see Human-in-the-loop
+  } else {
+    return lastAssistantText(result.messages);
+  }
+} finally {
+  agent.dispose(sessionId);   // cleans up sandbox/PTC state for the session
+}
+```
+
+Use `agent.stream(...)` instead of `invoke` when you want to record a
+trajectory or push incremental updates.
+
+`lastAssistantText` walks `result.messages` from the end and returns the last
+assistant/AI message's text.
+
+## Multi-turn conversations (chat history)
+
+`agent.invoke` accepts an **array of role-tagged messages**, not just one user
+string — the `messages` channel is a LangChain `add_messages` channel, so
+whatever you pass is appended to the session's transcript. There is no need to
+concatenate a prior chat into one giant user prompt; give the model real turns.
+
+**Preferred — session persistence.** Sessions are durable by default (the
+`BackendCheckpointSaver` writes each turn to the session store). Drive a live
+conversation by re-invoking the **same `thread_id`** once per user message,
+sending only the new message each time — prior turns are already in the
+checkpoint:
+
+```ts
+// turn 1
+await agent.invoke(
+  { messages: [{ role: "user", content: "Look up order 1234" }] },
+  { configurable: { thread_id: "chat-42" }, recursionLimit: 50 },
+);
+// turn 2 — SAME thread_id; only the new message
+await agent.invoke(
+  { messages: [{ role: "user", content: "And when did it ship?" }] },
+  { configurable: { thread_id: "chat-42" }, recursionLimit: 50 },
+);
+```
+
+The `init` state re-runs on each turn, resetting `workflowState` to the trigger's
+start state while **preserving the accumulated messages** — exactly right for
+chat: each user message re-enters the workflow with full history retained.
+
+**Alternative — seed prior history on the first invoke.** When the transcript
+lives in your own store (not a harness session), pass it as the initial `messages`
+array with the newest user turn last:
+
+```ts
+await agent.invoke(
+  {
+    messages: [
+      { role: "user", content: "Test" },
+      { role: "assistant", content: "Test complete…" },
+      { role: "user", content: "who are you?" },
+      { role: "assistant", content: "I'm an agent that…" },
+      { role: "user", content: "what was my first message?" }, // the new turn
+    ],
+  },
+  { configurable: { thread_id: "chat-42" }, recursionLimit: 50 },
+);
+```
+
+Gotchas:
+
+- **A hook reads the transcript from `messages`, newest last.** A hook that keys
+  off "the current request" should take the *last* `user` entry, not the first —
+  the first is the message that opened the session. (A grading rubric still
+  receives the compact `userRequest`/`history` payload its rubric is written
+  against; there `userRequest` is the *first* user message.)
+- **A definition edit is adopted, not refused.** A session takes its next turn
+  under the `workflow.yaml` in force now; the recorded `specHash` says which
+  definition each turn ran under but gates nothing. A turn positioned in a state
+  the edit removed reopens at its trigger's entry state rather than failing.
+- **Seed assistant turns as text only.** Don't fabricate dangling `tool_calls`
+  in seeded history, or the provider may reject the turn.
+- **Not every message in a session's transcript is a person's.** The runtime
+  writes **runtime notes** of its own — `[decision]`, `[event]`, `[error]`,
+  `[after]`, `[sub-workflow: …]` — carried as a synthetic `archmax_note` tool call
+  and its result, so a model cannot read them as a person speaking. Ask
+  `isRuntimeNote(message)` / `runtimeNoteKind(message)` rather than inferring
+  authorship from role or a `[bracket]` prefix: a turn's lower bound is a human
+  message **or** a note, and a transcript renderer should draw a note as a
+  runtime line, not as the user's own bubble. The one message a person really
+  sends into a running session — a reply to a run parked at a human state — is an
+  ordinary human message with no marker.
+
+```ts
+import { isRuntimeNote, runtimeNoteKind } from "@archmax-ai/harness";
+
+const opensTurn = (m: unknown) => isRuntimeNote(m) || m.role === "user";
+const label = (m: unknown) => runtimeNoteKind(m); // "decision" | "event" | … | null
+```
+
+## Human-in-the-loop
+
+```ts
+if (result.status === WORKFLOW_STATUSES.awaitingDecision && result.pendingDecision) {
+  const { state, transitions, evidence } = result.pendingDecision;
+  // Present `transitions` (each { to, description?, type? }) + evidence to a human.
+  // `evidence` is the state's declared paths followed by whatever the advancing
+  // agent attached via `archmax_advance({ evidence })` — one merged list, no
+  // host-side assembly, and no need to re-read the spec.
+  // When they pick:
+  const outcome = await agent.workflow.decide(sessionId, { target: chosenTo, comment });
+  // DecideOutcome: { status?, workflowState?, reparked, state?, messages }
+  if (outcome.reparked) { /* parked again at another human state */ }
+}
+```
+
+`decide` throws `SessionNotParkedError` if the session isn't awaiting a decision, and
+`InvalidDecisionTargetError` if `target` isn't a declared transition of the parked
+state. `sessions.list()` enumerates durable sessions (empty with a `MemorySaver` or a non-listable session store).
+
+## Event-in-the-loop: a run that parked itself
+
+The agent can park the run with `archmax_wait({ reason })` — in any state, terminal
+ones included — when the work cannot continue without something from outside. The
+session's status becomes `awaiting_input` (classified **open**, never finished) and
+`result.pendingInput` carries `{ state, title?, reason, parkedAt }`: the state it
+stopped in and, in its own words, what it is waiting for.
+
+```ts
+if (result.status === WORKFLOW_STATUSES.awaitingInput && result.pendingInput) {
+  const { state, reason } = result.pendingInput;   // show "waiting in <state>: <reason>"
+}
+
+// When the event arrives, deliver it. The run continues in the state it parked in;
+// the trigger id becomes its current trigger and the variables arrive locked.
+const outcome = await agent.workflow.deliver(sessionId, {
+  trigger: { id: "email_reply" },
+  variables: { reply_body: "It's ORD-1001." },
+});
+// DeliverOutcome: DecideOutcome + { variables, parkedChannel? }
+if (outcome.reparked) { /* parked again — outcome.parkedChannel says which channel */ }
+```
+
+### Waking a scheduled park
+
+A park may declare when it wants to be resumed (`archmax_wait({ reason, until: "1d" })`),
+which lands on the record, the summary, and the park event as an absolute
+`resumeAt`. **The SDK schedules nothing** — no timer, no cron — so this is your
+job, and it is a query:
+
+```ts
+const now = new Date().toISOString();
+for (const t of await agent.sessions.list()) {
+  if (t.status === WORKFLOW_STATUSES.awaitingInput && t.resumeAt && t.resumeAt <= now) {
+    await agent.workflow.deliver(t.sessionId, { trigger: { id: "timer" }, variables: { now } });
+  }
+}
+```
+
+A duplicate firing is safe: `deliver` refuses a session that is no longer parked. A
+`resumeAt` gates nothing — an event arriving earlier resumes the park normally — and
+a state that polls can cap its own retry loop with `budget.maxParks`, whose overrun
+routes through the state's `on_error`.
+
+`deliver` throws `SessionNotAwaitingInputError` if the session isn't parked awaiting
+an event and `MissingDeliveryTriggerError` if no id is supplied. The id is
+**recorded, never matched**: any trigger resumes a park, so there is no awaited-id
+list to check against — an id no state declares is deliverable exactly like a
+declared one. `resolveSession({ trigger, variables })` tells you which of
+the two a firing is — a `turn` to invoke, or a `resume` to deliver into — always on
+the conversation's own id.
+
+An event that must **never start** a run is therefore declared nowhere in the spec:
+every trigger a state declares starts a session at that state. Name the session
+yourself — `send(sessionId, { delivery })` / `deliver(sessionId, …)` — or pass
+`sessionPath` with the firing (`resolveSession({ trigger, variables, sessionPath })`)
+and let the runtime read the id out of the firing's own variables. A firing that
+names neither, on a trigger no state declares, throws
+`UnknownSessionTriggerError`: there is nowhere for it to go.
+
+One turn per invoke, on **one** session. A finished session re-invoked with a new
+message takes its next turn on the same id, continuing in the state the previous
+turn ended in — the position, variables, transcript and files are already there, so
+**never seed a previous transcript into the `messages` channel**. Pass only the new
+message.
+
+The runtime does not reposition a later turn, so nothing is injected into the
+transcript to explain a jump. When a turn needs a path the current state cannot
+reach, the agent calls `archmax_reset` to return to the state the conversation began
+in (available in every state). To continue *mid-state* instead — inside the state
+that stopped, with the arrival in its transcript — the turn must park rather than
+finish.
+
+## Lifecycle events
+
+Pass `onEvent`; every `WorkflowLifecycleEvent` carries a `level`
+(`info`/`warn`/`error`) plus a typed payload. Map these to a live run UI:
+
+| Event `type` | Meaning |
+| --- | --- |
+| `state-enter` / `state-leave` | state entered / left (`{ state, next? }`) |
+| `advance` | transition committed (`{ from, to, reason?, callId? }` — `callId` names the `archmax_advance` call that drove it; absent on the turn's opening arrival) |
+| `hook-start` / `hook-output` / `hook-passed` / `hook-verdict` / `hook-rejected` | lifecycle hook activity + verdict (`ok`/`correct`/`veto`) |
+| `parked` / `decided` | human state parked / resolved (`{ state, sessionId, awaiting, to? }` — a human park carries no `callId`, since no call asked for it) |
+| `parked` (`awaiting: "input"`) / `delivered` | the agent parked the run (`{ state, sessionId, reason, resumeAt?, callId }` — the `archmax_wait` call) / an event resumed it (`{ state, trigger, to }`, where `to` is the same state) |
+| `agent-text` | complete assistant message in a state (`{ state, text, messageId }` — always the id the message is stored under; absent only on a `partial: true` event from a failed turn) |
+| `agent-text-delta` | streaming text chunk (`{ state, text, messageId? }`) — flows for every run, even when the graph is driven with `invoke` |
+| `tool-called` / `tool-result` / `tool-blocked` | tool passed governance (`callId`, `args`) / settled (`callId`, `status`, `durationMs`, `output` preview) / denied. **Every** governed call, `archmax_advance` included — so a UI drawing a row per call draws one per transition; filter it if you already draw transitions from `advance`/`state-leave` |
+| `rubric-start` / `rubric-result` | grading-rubric dispatch bracketed by `dispatchId`, with the rubric's `name` |
+| `model-usage` | per model call provider token usage: `inputTokens` (the total, of which the cache counts are a breakdown), `outputTokens`, `cacheReadTokens`, `cacheCreationTokens`, `model?` (the id the call was priced against), `costUsd?` |
+| `prompt-shaping` | assembly-time payload summary: resolved `profile`, prompt-cache mechanism, `withheld` built-ins |
+| `skills-loaded`, `hooks-summary`, `graph-topology`, `workflow-reset`, `warning` | assembly-time diagnostics |
+
+For a browser client, forward these over SSE/WebSocket keyed by `sessionId`. To
+render a live chat transcript: accumulate `agent-text-delta` into a bubble
+keyed by `messageId`, replace it with the final `agent-text` bearing the same
+`messageId`, and interleave tool cards from `tool-called`/`tool-result` paired
+by `callId` (order everything by `seq`).
+
+Building a **durable record** rather than a live view? The ids join the stream to
+the stored messages exactly: `callId` equals the issuing AI message's
+`tool_calls[].id` and the settling `ToolMessage.tool_call_id`, and
+`agent-text.messageId` equals the assistant message's `id`. Every event a call
+*causes* carries that call's id too (`advance`, `parked`, `variables-set`,
+`title-set`, and the sub-workflow pair's `toolCallId`), so the events one call
+produced group from ids alone — no scanning back to a message boundary. An event
+of those types with no `callId` is one the runtime caused, not a call: a turn
+opening, a human park, run-start seeding.
+
+`title-set` is the one event carrying a variable's **value** — the reserved
+`title`, whenever it changes. `variables-set` still carries names only; the value
+rides its own event so that rule needs no exception, and it is safe there because
+a title is bounded to one short line by its write check. Indexing a run's title on
+your own record is a handler for this event and nothing else — no checkpoint read,
+which matters when you install your own checkpointer.
+
+### Reading tokens and cost
+
+Usage figures are the **provider's own** (LangChain `usage_metadata`), not
+estimates — including how many input tokens were served from the prompt cache.
+Use the exported tracker instead of summing events yourself:
+
+```ts
+import { createAgent, createUsageTracker } from "@archmax-ai/harness";
+
+// Wrap your own subscriber, or omit `onEvent` if you have none.
+const usage = createUsageTracker({ onEvent: forwardToClient });
+const agent = await createAgent({ onEvent: usage.handler, /* … */ });
+
+await agent.invoke(input, { configurable: { thread_id: sessionId } });
+
+const totals = usage.totals(sessionId);
+// { inputTokens, outputTokens, cacheReadTokens, cacheCreationTokens, costUsd? }
+usage.bySession();            // Map<sessionId, UsageSummary> for concurrent runs
+usage.reset(sessionId);       // drop a finished run's totals
+```
+
+Rising `cacheReadTokens` is how you confirm prompt caching is actually working;
+`costUsd` is present only when pricing is configured — report tokens rather than a
+guessed cost when it is absent. `formatUsage`, `addSummaries`, `emptyUsage` and
+`hasUsage` are exported for rendering and combining summaries; the extraction and
+pricing internals are not part of the public surface.
+
+### Reading a run's cost
+
+A run's usage is **checkpointed**, so a session handle answers it for a run that
+parked and resumed — including in another process:
+
+```ts
+const summary = await agent.sessions.get(sessionId);
+summary?.usage;
+// { inputTokens, outputTokens, cacheReadTokens, cacheCreationTokens, costUsd? }
+```
+
+`costUsd` appears only when pricing is configured; without it, tokens are
+reported and cost is omitted, never guessed. A parked run reports what it has
+spent so far and keeps accumulating on resume. The same totals land in
+`sessions/<sessionId>/artifacts/metadata.json` as `usage`, and `createUsageTracker()`
+gives the same figure live in process.
+
+### Measuring anything else — fold the stream
+
+The SDK keeps no metrics ledger. Tool calls, governance refusals, the path a run
+took, corrections and vetoes, rubric dispatches, and how long a session
+sat waiting for a person are all already on the typed event stream, and a backend
+counts what it needs:
+
+```ts
+const tally = { toolCalls: 0, toolBlocks: 0, path: [] as string[] };
+const agent = await createAgent({
+  onEvent: (e) => {
+    forwardToClient(e);
+    if (e.type === "tool-called") tally.toolCalls++;
+    if (e.type === "tool-blocked") tally.toolBlocks++;
+    if (e.type === "state-enter") tally.path.push(e.state);
+  },
+  /* … */
+});
+```
+
+Two properties make this the right home for it: the events are the ones the
+runtime acts on, so a fold cannot drift from behavior the way a parallel ledger
+can; and a backend counts what it cares about — a per-tenant tool budget, a
+per-state SLA — rather than what the SDK guessed.
+
+A fold is a **live view**: it sees only events this process observed. Cost is the
+deliberate exception, because it is checkpointed.
+
+## Session artifacts
+
+Each run persists under its own folder in the session store — which is also the
+agent's workspace root for that run (`sessions/<sessionId>/` on the filesystem default):
+
+```
+sessions/<sessionId>/
+  checkpoints/               # LangGraph checkpoint records (durable resume; also
+                             # the sole record of a parked human decision)
+  scratchpad/                # the run's working area — intermediate files AND the
+                             # artifacts a state produces (agent-visible as
+                             # `scratchpad/…`; the only such area)
+  large_tool_results/        # runtime-offloaded oversized tool results (agent reads only)
+  conversation_history/      # runtime-offloaded history (agent reads only)
+  artifacts/
+    graph.json               # SerializedGraph (topology; includes mermaid)
+    graph.mmd                # Mermaid diagram
+    trajectory.json          # Trajectory (per-state messages + tool calls)
+    trail.json               # Audit trail (committed transitions + rationale)
+    variables.json           # The run's variables: name -> { value, locked }.
+                             # `locked` is the attribution: locked = host-established
+                             # (variables, a delivery's seeds, the built-in
+                             # `trigger`), unlocked = set by the agent. Written for a
+                             # governed run; an empty store still writes `{}`.
+    metadata.json            # RunMetadata (runtime contract, package version,
+                             # specHash, and the run's `usage` summary — tokens
+                             # by kind plus costUsd when priced)
+```
+
+Schemas:
+
+```ts
+interface SerializedGraph {
+  nodes: string[];
+  edges: { source: string; target: string; conditional: boolean; label?: string }[];
+  entry?: string; finals: string[]; mermaid: string;
+}
+interface Trajectory {
+  runId: string; workflow: string; prompt?: string; finalAnswer: string;
+  segments: { node: string; enteredState?: string; exitedState?: string;
+              items: (MessageItem | ToolCallItem)[] }[];
+}
+// MessageItem: { kind:"message", role, content }
+// ToolCallItem: { kind:"tool_call", name, args, result: string|null }
+interface RunMetadata { runId: string; workflow: string;
+  runtimeContract: ResolvedRuntimeContract; packageVersion: string; }
+// trail.json: { runId, workflow, steps: TrailStep[] } — the audit trail,
+// re-derived from checkpointed state on every artifact write.
+// TrailStep: { to, kind: "trigger"|"agent"|"human"|"reset"|"on_error",
+//              reason?, ts }
+```
+
+An **API** reads these files (through the same session store, or via `runArtifactPaths(sessionId)`
+path helpers) to render historical runs: `graph.json`/`graph.mmd` for the diagram,
+`trajectory.json` for the step-by-step timeline, `trail.json` for the decision
+path (which edges committed, by whom, and why), `variables.json` for the run's
+inputs and established facts (also what a test author reads to reproduce a real
+firing's payload — see
+[hook-and-test-scripts.md](hook-and-test-scripts.md#recovering-a-cases-inputs-from-previous-runs)).
+Note it holds whatever the run's variables held, verbatim: a host that must not
+persist an input should not seed it. To persist artifacts yourself,
+stream the run and call `agent.emitRunArtifacts(runId, trajectory)`.
+
+## Model configuration
+
+Env (loaded from the workspace root `.env`, real process env wins):
+`ARCHMAX_API_BASE_URL`, `ARCHMAX_API_KEY`, `ARCHMAX_MODEL`, optional
+`ARCHMAX_TEMPERATURE`, `ARCHMAX_MAX_TOKENS` — any OpenAI-compatible endpoint.
+Or inject a model directly:
+
+```ts
+import { ChatOpenAI } from "@langchain/openai";
+const model = new ChatOpenAI({
+  model: "claude-sonnet-5",
+  apiKey: process.env.MY_KEY,
+  configuration: { baseURL: "https://api.example.com/v1" },
+});
+const agent = await createAgent({ model });
+```
+
+## Custom backends
+
+`BackendProtocolV2` (from `deepagents`) decouples storage — implement
+`readRaw`/`write`/`edit`/`ls` over filesystem, a store, a sandbox, S3, or a remote.
+Pass it as `backend` (authored workspace); run state goes through a `sessionStore`
+(e.g. `createBackendSessionStore({ backend: s3Backend, prefix: "tenants/a/runs" })`).
+A custom `backend` **requires** an explicit `sessionStore` — assembly throws
+`SessionStoreRequiredError` instead of inferring storage. The SDK owns the logical
+`sessions/<sessionId>/…` namespace; the store owns where it physically lives.
+Persisting the session zone off-box (an S3-backed session store) is how you share run
+state across a worker fleet and an API.
+
+## Cases against your agent
+
+`runTests` runs a workflow's cases — declarative YAML documents
+under `workflows/<name>/tests/*.test.yaml`, interpreted host-side — with the
+same discovery, case interpretation, judge, and verdicts `archmax test` uses.
+By default it builds its own agent, which is *not* your agent: no connection
+tools, no mounts, no session store, no checkpointer — and it reads the suite off
+the local filesystem under `rootDir`, which is *not* your authored tree if that
+tree lives in a store. Supply both:
+
+```ts
+import { runTests, createAgent, createToolMockMiddleware } from "@archmax-ai/harness";
+
+const { results, exitCode } = await runTests({
+  workflow,
+  rootDir,
+  // Where the spec, the case documents and their `from:` fixtures are read
+  // from. Omit it only when they really are files under `rootDir`.
+  authoring,
+  createTarget: async ({ workflow, rootDir, onEvent }) =>
+    createAgent({
+  workflow,
+  onEvent,
+  backend,
+  tools: await resolveConnectionTools(tenant),
+  checkpointer,
+  middleware: [createToolMockMiddleware()],
+  // REQUIRED — see below,
+  workspace: { rootDir, sessionStore },
+}),
+  sessionIdForCase: (file) => `${hostRunId}/${file}`,
+});
+```
+
+| Option | Why |
+| --- | --- |
+| `authoring` | The backend the **suite itself** is read from: the machine spec, `workflows/<name>/tests/*.test.yaml`, and the fixtures their `from:` entries name. The same option `createAgent` takes, and normally the same backend — the agent under test runs the authored tree the cases were read from. Defaults to a filesystem backend over the resolved root, so a host whose authored tree lives in a store and omits it discovers **no cases at all** and gets an empty run rather than a failure. |
+| `createTarget` | The agent under test — an already-built runtime, or a factory that builds one. Must be **workflow-governed** — a plain agent is refused before any case runs. In the factory form, `rootDir` in the context is the resolved workspace root. |
+| `sessionIdForCase` | Names the *agent* session the case runs on — a case is one conversation on one session (default `test-default-<runId>-<file>`), invoked once per case file — so its checkpoints, `scratchpad/…` files, and token usage join your own run record. |
+
+**Your target owns its mocks — and the runtime checks.** A case's `mocks:`
+list intercepts agent-initiated tool calls through
+`createToolMockMiddleware()`, which the default target wires for you.
+`createAgent` stamps `capabilities.toolMocks` from the middleware
+actually wired; a case that declares mocks against a target without the
+capability **fails loudly before the agent runs**, naming the fix — never run
+with partial interception (scripts' PTC calls mocked, the agent's own calls
+real). A bespoke target that intercepts by other means sets
+`capabilities.toolMocks` itself. Mockless cases run against any
+workflow-governed target.
+
+**Sub-workflows inherit the assembly.** When a workflow delegates (by advancing
+an `archmax_workflow_<slug>` tool), the child runtime is
+composed from the *same* resolved inputs as the parent — model and
+`modelFactory`, `backend` and `mounts`, `sessionStore` and `checkpointer`, host
+`tools` and `essentialTools`, `sandboxRuntime`, `hookExecutors`,
+`policyRules`, `promptCache` and `pricing` — differing only in
+the machine that governs it and the system prompt that machine renders. So a
+custom backend serves the child's spec too, and a `modelFactory` supplies its
+model. Composition is **lazy and memoized per workflow slug**: a workflow that is
+never delegated to is never loaded, and one delegated from many places is
+composed once. A child whose spec is broken surfaces as a *dispatch* failure, not
+an assembly failure — a parent still assembles.
+
+The child runs on the parent's **session id** (which is what gives it the same
+run zone) under a nested checkpoint namespace, so `sessions.list()` reports one
+session and the parent session's metrics include the child's tokens.
+`capabilities.subWorkflows` is stamped `true` for a governed assembly and `false`
+for the plain one.
+
+Omit both options and behavior is identical to `archmax test`. Cases run
+sequentially (`tests.maxConcurrency` above 1 is rejected, not silently
+serialized) and there is no cancellation signal: call `runTests` once per file
+(`filter`) if you need to interrupt between cases; a single case is bounded by
+`tests.caseTimeoutMs`.
+
+## Suggested split
+
+- **Worker**: `createAgent` + `agent.invoke`/`stream`; `agent.workflow.decide` for
+  approvals. Owns execution and the durable session store/checkpointer.
+- **API**: reads `sessions/<sessionId>/artifacts/*` (shared session store) and/or relays
+  the `onEvent` stream; exposes `sessions.list()`, run detail, and a decide endpoint.
+- **Frontend**: renders `graph.json`/`graph.mmd`, the `trajectory.json` timeline,
+  the `variables.json` inputs, and the live event stream; posts human decisions
+  back to the API.
+
+Keep the heavy runtime import (LangChain/LangGraph) in the worker; the API/frontend
+only need the artifact/event *types*, which ship in `dist/index.d.ts`.