@smthrs/gateway 0.35.0 → 1.0.0-rc.3

package/src/GatewayProjection.ts

@@ -0,0 +1,948 @@
+/**
+ * The stable wire rows the gateway serves for each projection, and the pure
+ * folds that compute them from control-plane facts.
+ *
+ * These rows are the whole read contract. Nothing here exposes a store row, a
+ * database column, or an engine type: every field is either a
+ * `@smthrs/control` projection value or something folded out of the ordered
+ * `ControlEvent` deltas `Control.watch` publishes. Wire names are the flows
+ * names (`flowId`, `createdAt`, and the `ApprovalTarget.Node` envelope), so a
+ * client written against the control plane reads these rows without a
+ * translation table.
+ *
+ * @since 1.0.0
+ */
+
+import { ControlFacts, ControlSchema, Health, Monitor } from "@smthrs/control"
+import * as Fault from "@smthrs/flow/Fault"
+import { ExecutionFact } from "@smthrs/journal"
+import { Schema } from "effect"
+import * as Diagnosis from "./Diagnosis.ts"
+import { callScope, openCallIndex, uniqueCallEvents } from "./internal/callEvents.ts"
+import * as NativeResolution from "./internal/nativeResolution.ts"
+import * as NodeEvents from "./internal/nodeEvents.ts"
+
+/**
+ * One run, everything a run card displays, and the diagnosis of what happened
+ * to it.
+ *
+ * `flowId` carries what the old wire split across `run.workflowKey` and
+ * `run.workflow`, and `createdAt` carries `run.createdAtMs`: one flows name
+ * each, rather than two spellings of one fact.
+ *
+ * @since 1.0.0
+ * @category models
+ */
+export const RunSummaryRow = Schema.Struct({
+  runId: Schema.String,
+  flowId: Schema.String,
+  statusRollup: Schema.optional(Health.StatusRollup),
+  status: ControlSchema.RunStatus,
+  lifecycleProvenance: Schema.optional(ControlFacts.LifecycleProvenance),
+  executionProvenance: Schema.optional(ExecutionFact.Provenance),
+  createdAt: Schema.Number,
+  updatedAt: Schema.Number,
+  planId: Schema.optional(Schema.String),
+  planDigest: Schema.optional(Schema.String),
+  parentRunId: Schema.optional(Schema.String),
+  lineageId: Schema.optional(Schema.String),
+  roundOrdinal: Schema.optional(Schema.Number),
+  waitingReason: Schema.optional(Schema.String),
+  steeringPending: Schema.optional(Schema.Number),
+  cancellation: Schema.optional(ControlSchema.Cancellation),
+  /** When the run's approved deadline passes; absent for a run with none. */
+  deadlineAt: Schema.optional(Schema.Number),
+  /** The model seat the run's last opened turn ran on. */
+  seat: Schema.optional(Schema.String),
+  turns: Schema.Number,
+  calls: Schema.Number,
+  callsFailed: Schema.Number,
+  editsAttempted: Schema.Number,
+  editsSucceeded: Schema.Number,
+  inputTokens: Schema.Number,
+  outputTokens: Schema.Number,
+  /**
+   * One line: the run row's status plus the reason that most explains it.
+   *
+   * Lifecycle provenance distinguishes a covered control fact from a legacy
+   * snapshot or a separate engine observation. Missing facts never silently
+   * replace a newer observed state with the last recorded status.
+   */
+  verdict: Schema.String,
+  /** The whole diagnosis card, which the old wire called `whatHappened`. */
+  diagnosis: Schema.String,
+  finalOutput: Schema.optional(Schema.String),
+  /** The failed run's fault class and `<_tag>/<code>`, stamped when it failed. */
+  failureFault: Schema.optional(Fault.Class),
+  failureTag: Schema.optional(Schema.String)
+})
+
+/**
+ * One run, everything a run card displays, and its diagnosis.
+ *
+ * @since 1.0.0
+ * @category models
+ */
+export type RunSummaryRow = typeof RunSummaryRow.Type
+
+/**
+ * One node of a run, as a tree view renders it.
+ *
+ * `label` carries what the old wire called `node.cardLabel`, and `seat`
+ * carries `node.agent`: the model seat is the flows name for who ran a node.
+ *
+ * `nodeId` is the ordinal the call opened on, because the emitter names no
+ * node. {@link runTree} says why.
+ *
+ * @since 1.0.0
+ * @category models
+ */
+export const RunTreeRow = Schema.Struct({
+  runId: Schema.String,
+  nodeId: Schema.String,
+  label: Schema.String,
+  status: Schema.Literals(["running", "completed", "failed"]),
+  seat: Schema.optional(Schema.String),
+  startedAt: Schema.Number,
+  endedAt: Schema.optional(Schema.Number),
+  parentRunId: Schema.optional(Schema.String)
+})
+
+/**
+ * One node of a run, as a tree view renders it.
+ *
+ * @since 1.0.0
+ * @category models
+ */
+export type RunTreeRow = typeof RunTreeRow.Type
+
+/**
+ * One approval a run is parked on, carrying the exact payload that decides it.
+ *
+ * `title` carries the old `approval.requestTitle` and `request` the old
+ * `approval.request`; `payload` is the `ApprovalTarget.Node` envelope a client
+ * submits back unchanged, so no client reconstructs authority.
+ *
+ * @since 1.0.0
+ * @category models
+ */
+export const ApprovalRow = Schema.Struct({
+  runId: Schema.String,
+  /**
+   * The execution holding the wait, when this gate is a question a person owes
+   * an answer to and it is held below `runId`.
+   *
+   * `runId` is the run a person opened and the run a decision is addressed to;
+   * a `HumanTask` in a nested flow parks its own execution, and the two are
+   * different runs. Present only on such a gate, so a reader can tell a
+   * question rolled up from a tree from one the named run raised itself, and
+   * an inbox can list one question once however many ancestors carry it.
+   */
+  waitRunId: Schema.optional(Schema.String),
+  questionProvenance: Schema.optional(Schema.Literals(["events", "legacy-observation", "unverified-observation"])),
+  requestId: Schema.String,
+  title: Schema.String,
+  request: Schema.Json,
+  payload: ControlSchema.ApprovalPayload,
+  requestedAt: Schema.Number,
+  status: Schema.Literals(["pending", "approved", "denied"])
+})
+
+/**
+ * One approval a run is parked on.
+ *
+ * @since 1.0.0
+ * @category models
+ */
+export type ApprovalRow = typeof ApprovalRow.Type
+
+/**
+ * The output one node produced.
+ *
+ * @since 1.0.0
+ * @category models
+ */
+export const NodeOutputRow = Schema.Struct({
+  runId: Schema.String,
+  nodeId: Schema.String,
+  outcome: Schema.Literals(["success", "failure"]),
+  output: Schema.String,
+  settledAt: Schema.Number
+})
+
+/**
+ * The output one node produced.
+ *
+ * @since 1.0.0
+ * @category models
+ */
+export type NodeOutputRow = typeof NodeOutputRow.Type
+
+/**
+ * How long one flow's nodes take, per action tag.
+ *
+ * `samples` is the number of measured executions behind the two percentiles,
+ * and it is on the wire because a percentile over one sample and a percentile
+ * over twenty are different claims. A tag with no measured execution has no
+ * row at all: there is no shape here for an unmeasured prediction.
+ *
+ * @since 1.0.0
+ * @category models
+ */
+export const FlowDurationRow = Schema.Struct({
+  flowId: Schema.String,
+  actionTag: Schema.String,
+  samples: Schema.Number,
+  p50Ms: Schema.Number,
+  p90Ms: Schema.Number
+})
+
+/**
+ * How long one flow's nodes take, per action tag.
+ *
+ * @since 1.0.0
+ * @category models
+ */
+export type FlowDurationRow = typeof FlowDurationRow.Type
+
+/**
+ * One measured execution of one node, as a duration fold's sample.
+ *
+ * The action tag is the only part of a node's key material that survives a
+ * re-key, which is why history is grouped by it rather than by node id or by
+ * dispatch key.
+ *
+ * @since 1.0.0
+ * @category models
+ */
+export interface NodeDuration {
+  readonly actionTag: string
+  readonly durationMs: number
+}
+
+/**
+ * One line of a run's turn-by-turn transcript.
+ *
+ * @since 1.0.0
+ * @category models
+ */
+export const TranscriptRow = Schema.Struct({
+  runId: Schema.String,
+  sequence: Schema.Number,
+  turn: Schema.Number,
+  at: Schema.Number,
+  kind: Schema.String,
+  /** Dispatch identity on current call records; absent on legacy history. */
+  callId: Schema.optional(Schema.String),
+  text: Schema.String
+})
+
+/**
+ * One line of a run's turn-by-turn transcript.
+ *
+ * @since 1.0.0
+ * @category models
+ */
+export type TranscriptRow = typeof TranscriptRow.Type
+
+/** Copies only the optional fields a run summary actually carries. */
+const optional = <A>(key: string, value: A | undefined): Record<string, A> =>
+  value === undefined ? {} : { [key]: value }
+
+/**
+ * Folds one run's summary and events into the served `run-summary` row.
+ *
+ * @param run the control-plane run summary
+ * @param events that run's ordered control events
+ * @param now the instant the row is rendered at
+ * @param carry the digest of the events a bounded reader dropped before them
+ * @since 1.0.0
+ * @category projections
+ */
+export const runSummary = (
+  run: ControlSchema.RunSummary,
+  events: ReadonlyArray<ControlSchema.ControlEvent>,
+  now: number = Math.max(run.updatedAt, events.at(-1)?.occurredAt ?? 0),
+  carry?: Diagnosis.Digest | undefined
+): RunSummaryRow => {
+  const projected = ControlFacts.fold(events, run)
+  // Supplying a snapshot guarantees a run: uncovered facts retain that
+  // observation; only the event-only fold can return an absent run.
+  run = projected.run!
+  const native = ExecutionFact.foldControl(events, run.runId, run.executionView)
+  // The control fold remains control authority. Native lifecycle is independently
+  // replayed/verified against the coherent executor observation when available.
+  if (run.executionObservation === "observed" && native?.view !== undefined) {
+    const { root, current } = native.view
+    run = {
+      ...run,
+      status: current.status === "pending" ? "accepted" : current.status === "suspended"
+        ? current.waiting?.reason === "approval" ||
+            (native.view.humanWaits?.length ?? run.pendingWaits?.length ?? 0) > 0
+          ? "waiting-approval" :
+          "parked"
+        : current.status === "running" && (native.view.humanWaits?.length ?? run.pendingWaits?.length ?? 0) > 0
+        ? "waiting-approval" :
+        current.status,
+      waitingReason: current.status === "suspended" ? current.waiting?.reason : undefined,
+      parentRunId: root.parentRunId ?? undefined,
+      lineageId: root.lineageId,
+      roundOrdinal: root.roundOrdinal
+    }
+  }
+  // `carry` is the digest of the events a bounded reader dropped ahead of
+  // `events`. Combining it keeps turns, calls, edits and tokens exact for a run
+  // whose journal no longer fits one window, so the counters a run card shows
+  // describe the whole run and not just its tail.
+  const window = Diagnosis.digest(events, run.runId)
+  const facts = { ...(carry === undefined ? window : Diagnosis.combine(carry, window)), status: run.status }
+  return {
+    runId: run.runId,
+    flowId: run.flowId,
+    status: run.status,
+    lifecycleProvenance: projected.provenance,
+    ...optional("executionProvenance", native?.provenance),
+    statusRollup: statusRollup(run, events, now),
+    createdAt: run.createdAt,
+    updatedAt: run.updatedAt,
+    ...optional("planId", run.planId),
+    ...optional("planDigest", run.planDigest),
+    ...optional("parentRunId", run.parentRunId),
+    ...optional("lineageId", run.lineageId),
+    ...optional("roundOrdinal", run.roundOrdinal),
+    ...optional("waitingReason", run.waitingReason),
+    ...optional("steeringPending", run.steering?.pending),
+    ...optional("cancellation", run.cancellation),
+    ...optional("deadlineAt", run.deadlineAt),
+    ...optional("seat", facts.seat),
+    turns: facts.turns,
+    calls: facts.calls,
+    callsFailed: facts.callsFailed,
+    editsAttempted: facts.editsAttempted,
+    editsSucceeded: facts.editsSucceeded,
+    inputTokens: facts.inputTokens,
+    outputTokens: facts.outputTokens,
+    verdict: Diagnosis.verdict(facts),
+    diagnosis: Diagnosis.render({ runId: run.runId, ...optional("flowId", run.flowId) }, facts),
+    ...optional("finalOutput", Diagnosis.resolvedOutput(facts) ?? (
+      run.status === "failed" && facts.fault?.tag.endsWith("/source_refused") &&
+        facts.cause?.match(/source_refused: ([a-z_]+)/)
+        ? JSON.stringify({ message: facts.cause.match(/source_refused: ([a-z_]+)/)![0] })
+        : undefined
+    )) ,
+    ...(run.status === "failed" ? optional("failureFault", facts.fault?.class) : {}),
+    ...(run.status === "failed" ? optional("failureTag", facts.fault?.tag) : {})
+  }
+}
+
+/**
+ * Claims a call by its durable dispatch identity. Legacy history has no id:
+ * match only legacy starts by flow name (or FIFO when unnamed). A new
+ * settlement can close a legacy start after an old parked run resumes, but
+ * neither fallback may steal an identified call.
+ */
+const takeOpenCall = <
+  A extends { readonly flowName: string; readonly callId: string | undefined; readonly scope: string | undefined }
+>(
+  open: Array<A>,
+  callId: string | undefined,
+  flowName: string | undefined,
+  scope: string | undefined
+): A | undefined => {
+  const found = openCallIndex(open, callId, flowName, scope)
+  return found < 0 ? undefined : open.splice(found, 1)[0]
+}
+
+/**
+ * One agent call, as every node projection sees it: the ordinal it opened on,
+ * the seat that ran it, and the settlement that closed it, if one did.
+ */
+interface CallRecord {
+  readonly nodeId: string
+  readonly flowName: string
+  readonly seat: string | undefined
+  readonly startedAt: number
+  readonly settlement: CallSettlement | undefined
+}
+
+/** What closed a call, and where among the run's settlements it closed. */
+interface CallSettlement {
+  /** The rank of this settlement among the run's, counting from one. */
+  readonly order: number
+  /** The run the settlement named, which a malformed journal can omit. */
+  readonly runId: string | undefined
+  readonly at: number
+  readonly outcome: "success" | "failure"
+  readonly output: string
+}
+
+/**
+ * Folds one run's events into the agent calls it made, in the order they
+ * opened.
+ *
+ * This is the only fold that assigns a call its `call-N` identity, records the
+ * seat and time it opened on, and pairs a settlement with the call it closed.
+ * A client asks for a node's output by the id a tree row carries, so the two
+ * projections have to agree on that id; when they counted separately they
+ * could disagree, and once did, because skipping an event before the ordinal
+ * advanced shifted one fold's keys against the other's. The ordinal therefore
+ * advances on every distinct start, whatever else that event is missing.
+ * Identified duplicate starts and settlements are ignored. Old idless rows
+ * retain their original ordinal/FIFO interpretation; their missing dispatch
+ * identity cannot be reconstructed when same-name calls overlapped.
+ *
+ * `uniqueCallEvents` also normalizes native step facts, so this fold reads the
+ * run's own records and one stream per step a module run dispatched. The three
+ * classes `Diagnosis` names apply here too, field by field:
+ *
+ * | Field                   | Class           | Why                                                    |
+ * | ----------------------- | --------------- | ------------------------------------------------------ |
+ * | `nodeId` (`call-N`)     | aggregate       | one tree lists every call the run made, steps included |
+ * | `flowName`, `startedAt` | aggregate       | read off the call's own record                         |
+ * | `settlement`            | aggregate       | claimed by call identity within the recorded scope     |
+ * | `seat`                  | scope dependent | who ran THIS call, so the turn must be this call's own |
+ *
+ * There is no root-only field here: a tree row describes a call, never the
+ * run's answer or state. A new field has to name its class.
+ */
+const callHistory = (
+  events: ReadonlyArray<ControlSchema.ControlEvent>
+): ReadonlyArray<CallRecord> => {
+  const calls = new Map<string, CallRecord>()
+  const open: Array<
+    {
+      readonly callId: string | undefined
+      readonly flowName: string
+      readonly call: CallRecord
+      readonly scope: string | undefined
+    }
+  > = []
+  // The seat of the last turn opened in each scope. A module run interleaves
+  // steps in one journal, so the run-wide reading attributed one step's seat
+  // to another step's calls. A prompt run records one unscoped stream and
+  // reads exactly as it did.
+  const seats = new Map<string | undefined, string | undefined>()
+  let ordinal = 0
+  let settlements = 0
+
+  for (const event of uniqueCallEvents(events)) {
+    const payload = Diagnosis.asRecord(event.payload)
+    if (event.kind === "control.agent.turn-opened") {
+      const scope = callScope(event)
+      seats.set(scope, Diagnosis.asString(payload.seat) ?? seats.get(scope))
+      continue
+    }
+    if (event.kind === "control.agent.cell-call-started") {
+      const callId = Diagnosis.asString(payload.callId)
+      ordinal += 1
+      const nodeId = `call-${ordinal}`
+      const flowName = Diagnosis.asString(payload.flowName) ?? nodeId
+      const scope = callScope(event)
+      // This scope's own turn, else the run's. A step that opened no turn of
+      // its own borrows nothing from a sibling step.
+      const seat = seats.get(scope) ?? seats.get(undefined)
+      const call: CallRecord = { nodeId, flowName, seat, startedAt: Diagnosis.timeOf(event), settlement: undefined }
+      open.push({ callId, flowName, call, scope })
+      calls.set(nodeId, call)
+      continue
+    }
+    if (event.kind !== "control.agent.cell-call-settled") continue
+    const callId = Diagnosis.asString(payload.callId)
+    const claimed = takeOpenCall(open, callId, Diagnosis.asString(payload.flowName), callScope(event))
+    if (claimed === undefined) continue
+    const failed = Diagnosis.asString(payload.outcome) === "failure"
+    settlements += 1
+    // Rewriting the entry keeps the call in the position it opened on, which
+    // is the order a tree renders.
+    calls.set(claimed.call.nodeId, {
+      ...claimed.call,
+      settlement: {
+        order: settlements,
+        runId: Diagnosis.asString(payload.runId) ?? event.runId,
+        at: Diagnosis.timeOf(event),
+        outcome: failed ? "failure" : "success",
+        output: failed
+          ? Diagnosis.asString(payload.message) ?? ""
+          : Diagnosis.asString(payload.value) ?? JSON.stringify(payload.value ?? null)
+      }
+    })
+  }
+  return [...calls.values()]
+}
+
+/** How a tree row reports a call: settled the way it settled, or still running. */
+const treeStatus = (settlement: CallSettlement | undefined): RunTreeRow["status"] => {
+  if (settlement === undefined) return "running"
+  return settlement.outcome === "failure" ? "failed" : "completed"
+}
+
+/**
+ * Folds one run's events into its node rows.
+ *
+ * A node opens on `control.agent.cell-call-started` and settles on the
+ * matching `control.agent.cell-call-settled`.
+ *
+ * AgentSession journals the same `callId` from the harness's dispatch identity
+ * on both records. Settlements join on it even when same-name calls overlap
+ * and finish out of order. The distinct-start ordinal remains the published
+ * `call-N` node key, preserving existing node links. Legacy records without
+ * callId retain same-name FIFO pairing, restricted to other legacy records
+ * (including a newly identified settlement of an old parked call).
+ *
+ * The durable engine's own `flows.engine.*` records are not folded here, but
+ * that is a choice rather than a limit. A host keeps the control plane and the
+ * engine in two databases with two journals (`@smthrs/cli`
+ * `NodeControl.databasePath` and `executionDatabasePath`), and
+ * `EngineJournalProjection` copies every engine entry of the run into the
+ * control journal as a `control.engine.event` envelope, which is how
+ * {@link nodeDurations} reads the node records. A tree row is an agent cell
+ * call, so what an engine step did reaches this projection as the call that
+ * made it.
+ *
+ * A node that never settled stays `running`, which is how a live tree renders
+ * work in flight.
+ *
+ * @param run the control-plane run summary
+ * @param events that run's ordered control events
+ * @since 1.0.0
+ * @category projections
+ */
+export const runTree = (
+  run: ControlSchema.RunSummary,
+  events: ReadonlyArray<ControlSchema.ControlEvent>
+): ReadonlyArray<RunTreeRow> =>
+  callHistory(events).map((call): RunTreeRow => ({
+    runId: run.runId,
+    nodeId: call.nodeId,
+    label: call.flowName,
+    status: treeStatus(call.settlement),
+    ...optional("seat", call.seat),
+    startedAt: call.startedAt,
+    ...optional("endedAt", call.settlement?.at),
+    ...optional("parentRunId", run.parentRunId)
+  }))
+
+/**
+ * The declared question of a human wait, as far as it can be read.
+ *
+ * `request` is whatever the wait wrote about itself, which for `HumanTask` is
+ * `{task, name, kind, prompt, attempt, maxAttempts}`. A wait that declared
+ * nothing — an older park, or a plugin's own — still produces a row, because a
+ * gate a person cannot see is a run that waits forever; it just has less to
+ * render.
+ */
+const questionOf = (wait: ControlSchema.PendingWait): Record<string, unknown> => Diagnosis.asRecord(wait.request)
+
+/**
+ * One pending row for a human wait held anywhere in a run's tree.
+ *
+ * `runId` is the ROOT: this is the run a person opened, and the run a decision
+ * is addressed to. `waitRunId` is the execution actually holding it, because a
+ * client routes by run and the control plane routes within the tree.
+ *
+ * `requestId` is the wait point's own name (`coding-clarification#1`), which is
+ * unique within a tree, stable across a re-read, and the name a `Signal`
+ * addresses. The token travels in the request too, for a client that would
+ * rather submit the exact durable address back.
+ */
+const humanWaitRow = (
+  run: ControlSchema.RunSummary,
+  wait: ControlSchema.PendingWait,
+  projected?: ExecutionFact.Observation,
+  questionProvenance: "events" | "legacy-observation" | "unverified-observation" = "legacy-observation"
+): ApprovalRow => {
+  const question = projected === undefined ? questionOf(wait) : Diagnosis.asRecord(projected.waiting?.request)
+  const name = wait.name ?? Diagnosis.asString(question["name"]) ?? wait.token
+  const requestId = wait.attempt === undefined ? name : `${name}#${wait.attempt}`
+  const prompt = Diagnosis.asString(question["prompt"])
+  const waitFlowId = projected?.flowName ?? wait.flowId
+  return {
+    runId: run.runId,
+    waitRunId: wait.runId,
+    questionProvenance,
+    requestId,
+    title: prompt ?? `Answer needed — ${name}`,
+    request: {
+      ...question,
+      kind: Diagnosis.asString(question["kind"]) ?? "ask",
+      name,
+      ...(wait.attempt === undefined ? {} : { attempt: wait.attempt }),
+      ...(waitFlowId === undefined ? {} : { waitFlowId }),
+      token: wait.token
+    },
+    payload: {
+      target: {
+        _tag: "Node",
+        runId: run.runId,
+        requestId,
+        digest: wait.token,
+        // A human answer grants no capabilities: the envelope a decision binds
+        // to is the empty one, the same shape a flow with none carries.
+        envelope: { capabilities: [], flows: [], budget: {} }
+      },
+      scope: "once",
+      // The identity is the PARK, not its name: a loop that re-asks one
+      // question parks again under the same run, name and attempt, and a
+      // name-keyed answer would deduplicate the second decision into the
+      // first — accepted, never delivered, and parked forever.
+      idempotencyKey: `answer:${run.runId}:${requestId}:${wait.tokenDigest ?? wait.token}`
+    },
+    requestedAt: projected?.createdAtMs ?? wait.createdAt,
+    status: "pending"
+  }
+}
+
+/**
+ * Folds one run's events into its approval rows.
+ *
+ * A request opens a pending row carrying the submit-ready payload; the
+ * matching `control.approval.approved` or `control.approval.denied` closes it
+ * without discarding the request, so a decided gate stays readable.
+ *
+ * A `HumanTask` declares its question on the parked row. Native execution
+ * facts carry the redacted question and token digest; verified facts supply
+ * its display metadata here. `@smthrs/control` rolls current open waits onto
+ * the root as `pendingWaits`, which supply each row's protected wake address.
+ * History alone cannot create an answerable human wait. Older observations
+ * remain visible with explicit question provenance when native fact coverage
+ * is unavailable. These rows come first because they still need an answer.
+ *
+ * A decision names the gate it closed by `tokenId`. `@smthrs/control`
+ * `SqlControlRuntime.lookupApproval` mints that token id from the target, and
+ * for the `Node` target a run parks on it is the request id itself, so the two
+ * records join on one field. A decision whose `tokenId` or `requestId` names
+ * no row is retained for an exact later request. Only a legacy decision that
+ * names neither field closes the oldest pending legacy row in that run.
+ * Current facts bind the target digest; duplicates cannot reopen a decision.
+ *
+ * @param events the run's ordered control events
+ * @since 1.0.0
+ * @category projections
+ */
+export const approvals = (
+  events: ReadonlyArray<ControlSchema.ControlEvent>,
+  run?: ControlSchema.RunSummary | undefined
+): ReadonlyArray<ApprovalRow> => {
+  const native = run === undefined ? undefined : ExecutionFact.foldControl(events, run.runId, run.executionView)
+  const observed = run?.pendingWaits?.map((wait) => {
+    const projected = native?.provenance.humanWaits === "events" ?
+      native.view?.humanWaits?.find((candidate) =>
+        candidate.executionId === wait.runId && candidate.waiting?.tokenDigest === wait.tokenDigest
+      ) :
+      undefined
+    return humanWaitRow(
+      run,
+      wait,
+      projected,
+      projected === undefined
+        ? native?.provenance.humanWaits === "events" ?
+          "unverified-observation" :
+          native?.provenance.humanWaits ?? "legacy-observation"
+        : "events"
+    )
+  })
+  // The current observation supplies wake authority. History supplies only the
+  // display metadata that verifies against that exact observed token digest.
+  return ControlFacts.fold(events, undefined, observed).approvals
+}
+
+/**
+ * Folds one run's events into its node outputs, keyed the way {@link runTree}
+ * keys its rows: by the ordinal the call opened on.
+ *
+ * A settled call carries the value it produced, so the row carries that value.
+ * A run's own result reaches a client as `RunSummaryRow.finalOutput` instead.
+ *
+ * @param events the run's ordered control events
+ * @since 1.0.0
+ * @category projections
+ */
+export const nodeOutput = (
+  events: ReadonlyArray<ControlSchema.ControlEvent>
+): ReadonlyArray<NodeOutputRow> => {
+  const settled: Array<{ readonly order: number; readonly row: NodeOutputRow }> = []
+  for (const call of callHistory(events)) {
+    const settlement = call.settlement
+    if (settlement === undefined) continue
+    // A row names the run it belongs to. A settlement that names none closes
+    // its call without producing one.
+    if (settlement.runId === undefined) continue
+    settled.push({
+      order: settlement.order,
+      row: {
+        runId: settlement.runId,
+        nodeId: call.nodeId,
+        outcome: settlement.outcome,
+        output: settlement.output,
+        settledAt: settlement.at
+      }
+    })
+  }
+  // A tree reads in the order calls opened; outputs read in the order they
+  // settled, which is the order a client watched them arrive.
+  return settled.sort((left, right) => left.order - right.order).map((entry) => entry.row)
+}
+
+/** The node id every execution gives the flow it was started to run. */
+const rootNodeId = "root"
+
+/**
+ * The native execution the host bound this control run to, when one is known.
+ *
+ * Only the host's own binding record names the root, and two records that
+ * disagree name nothing: an ambiguous binding cannot tell a callee's
+ * execution apart from the root's, so it authorizes no skip.
+ *
+ * The binding is written once, when the host takes the root up, so a window
+ * bounded to a run's newest events is the one window it falls out of. The
+ * carried digest keeps it across that eviction, and reading from the carry is
+ * what keeps a long run's calls counted the same as a short run's.
+ */
+const boundExecution = (
+  events: ReadonlyArray<ControlSchema.ControlEvent>,
+  carry: Diagnosis.Digest | undefined
+): string | undefined => {
+  let resolution = carry?.nativeResolution
+  for (const event of events) resolution = NativeResolution.combine(resolution, NativeResolution.fromEvent(event))
+  return resolution === undefined || resolution.conflict === true ? undefined : resolution.binding?.executionId
+}
+
+/**
+ * A callee's own view of a call its caller already measured.
+ *
+ * Every execution below the bound root was started by a `FlowCall` node in
+ * the execution above it, and that node carries the same action tag over the
+ * same span. The caller's node is the one kept, because it is the node a plan
+ * carries and the one a graph can join a prediction onto.
+ *
+ * With no binding nothing is echoed: a fold that cannot name the root cannot
+ * tell which `root` record belongs to it, and dropping samples on a guess
+ * would answer a smaller history than the flow ran.
+ */
+const echoedRoot = (bound: string | undefined, executionId: string, nodeId: string): boolean =>
+  bound !== undefined && nodeId === rootNodeId && executionId !== bound
+
+/**
+ * Folds one run's events into the executions it measured.
+ *
+ * A node's work starts when the engine recorded it scheduled and ends when it
+ * recorded it settled, so a duration is the distance between those two native
+ * stamps. Four of the five outcomes contribute nothing: `clean` was served
+ * from records rather than run, `failed` measures a collapse, `skipped` was
+ * never reached, and `deferred` is scheduling debt. Only `built` ran.
+ *
+ * A settlement pairs with the LAST schedule of the same execution and node,
+ * which is what makes a retried node measure its final attempt rather than
+ * the span across every attempt. Two executions of one node id stay apart
+ * because the pairing key carries the execution, and a settlement no schedule
+ * opened measures nothing rather than measuring from zero.
+ *
+ * A record that names no action tag contributes nothing either: history is
+ * grouped by tag, and a sample with no tag has nowhere honest to go.
+ *
+ * A called flow is measured once. The engine records one call twice: the
+ * caller admits and settles the node that made it, and the callee's own
+ * execution admits and settles its `root` over the same span with the same
+ * action tag. {@link echoedRoot} drops the callee's copy, so one invocation
+ * is one sample rather than two, and `samples` stays the count of executions
+ * the flow performed.
+ *
+ * @param events the run's ordered control events
+ * @param carry the digest of the events a bounded window dropped, when it dropped any
+ * @since 1.0.0
+ * @category projections
+ */
+export const nodeDurations = (
+  events: ReadonlyArray<ControlSchema.ControlEvent>,
+  carry?: Diagnosis.Digest | undefined
+): ReadonlyArray<NodeDuration> => {
+  const bound = boundExecution(events, carry)
+  const started = new Map<string, number>()
+  const measured: Array<NodeDuration> = []
+  for (const event of events) {
+    const schedule = NodeEvents.nodeScheduled(event)
+    if (schedule !== undefined) {
+      if (echoedRoot(bound, schedule.executionId, schedule.payload.nodeId)) continue
+      started.set(`${schedule.executionId}\u0000${schedule.payload.nodeId}`, schedule.emittedAtMs)
+      continue
+    }
+    const settlement = NodeEvents.nodeSettled(event)
+    if (settlement === undefined) continue
+    if (echoedRoot(bound, settlement.executionId, settlement.payload.nodeId)) continue
+    if (settlement.payload.outcome !== "built") continue
+    const actionTag = settlement.payload.action
+    if (actionTag === undefined) continue
+    const key = `${settlement.executionId}\u0000${settlement.payload.nodeId}`
+    const startedAt = started.get(key)
+    if (startedAt === undefined) continue
+    started.delete(key)
+    const durationMs = settlement.emittedAtMs - startedAt
+    // Wall clocks can move backwards, and a completion that predates its own
+    // start is a clock fact rather than a duration.
+    if (durationMs < 0) continue
+    measured.push({ actionTag, durationMs })
+  }
+  return measured
+}
+
+/** The nearest-rank percentile of an ascending sample, which is one observed value. */
+const nearestRank = (ascending: ReadonlyArray<number>, percentile: number): number =>
+  ascending[Math.ceil((percentile / 100) * ascending.length) - 1]!
+
+/**
+ * Folds every measured execution of one flow into one row per action tag.
+ *
+ * Both percentiles are nearest-rank, so each answer is a duration the flow
+ * really took rather than an interpolation between two it did not. Rows are
+ * ordered by tag, so two reads of the same history answer in the same order.
+ *
+ * A tag reaches this function only by having been measured, so every row
+ * carries at least one sample, and a flow nothing has measured answers with
+ * no rows at all.
+ *
+ * @param flowId the flow whose history was read
+ * @param samples every measured execution of that flow's runs
+ * @since 1.0.0
+ * @category projections
+ */
+export const flowDurations = (
+  flowId: string,
+  samples: ReadonlyArray<NodeDuration>
+): ReadonlyArray<FlowDurationRow> => {
+  const byTag = new Map<string, Array<number>>()
+  for (const sample of samples) {
+    const held = byTag.get(sample.actionTag)
+    if (held === undefined) byTag.set(sample.actionTag, [sample.durationMs])
+    else held.push(sample.durationMs)
+  }
+  return [...byTag.entries()]
+    .sort(([left], [right]) => left < right ? -1 : 1)
+    .map(([actionTag, observed]) => {
+      const ascending = [...observed].sort((left, right) => left - right)
+      return {
+        flowId,
+        actionTag,
+        samples: ascending.length,
+        p50Ms: nearestRank(ascending, 50),
+        p90Ms: nearestRank(ascending, 90)
+      }
+    })
+}
+
+/** Events the transcript reports verbatim rather than as agent activity. */
+const transcriptKinds: ReadonlySet<string> = new Set(["control.approval.requested"])
+
+/**
+ * Folds one run's events into a turn-numbered transcript.
+ *
+ * A row is reported for every `control.run.*` and `control.agent.*` event and
+ * for an approval request, and for nothing else: the transcript is what a
+ * reader follows, not the whole journal, which `run-events` already serves.
+ * The turn counter advances on `control.agent.turn-opened`, so every row
+ * carries the turn it belongs to, and each row's text is one display line, so
+ * a multi-line seat, message, or question cannot split one row into several.
+ *
+ * `uniqueCallEvents` normalizes native step facts, so a module run's steps
+ * report here too. The classes `Diagnosis` names apply, field by field:
+ *
+ * | Field  | Class     | Why                                                       |
+ * | ------ | --------- | --------------------------------------------------------- |
+ * | rows   | aggregate | every row is one event this run recorded, wherever it ran |
+ * | `turn` | aggregate | a turn a step opened is a turn this run opened            |
+ *
+ * No row is root only, because no row claims to be the run's answer or state:
+ * a `control.agent.resolved` row says one agent resolved, and the run's own
+ * answer is served by `runSummary` from `Diagnosis`. A new field has to name
+ * its class.
+ *
+ * @param events the run's ordered control events
+ * @since 1.0.0
+ * @category projections
+ */
+export const transcript = (
+  events: ReadonlyArray<ControlSchema.ControlEvent>
+): ReadonlyArray<TranscriptRow> => {
+  const rows: Array<TranscriptRow> = []
+  let turn = 0
+  for (const event of uniqueCallEvents(events)) {
+    const payload = Diagnosis.asRecord(event.payload)
+    const runId = Diagnosis.asString(payload.runId) ?? event.runId
+    if (runId === undefined) continue
+    const reported = event.kind.startsWith("control.run.") ||
+      event.kind.startsWith("control.agent.") ||
+      transcriptKinds.has(event.kind)
+    if (!reported) continue
+    if (event.kind === "control.agent.turn-opened") turn += 1
+    rows.push({
+      runId,
+      sequence: event.sequence,
+      turn,
+      at: Diagnosis.timeOf(event),
+      kind: event.kind,
+      ...optional("callId", Diagnosis.asString(payload.callId)),
+      text: line(event.kind, payload)
+    })
+  }
+  return rows
+}
+
+/** One event as a single transcript line. */
+const line = (kind: string, payload: Record<string, unknown>): string => {
+  switch (kind) {
+    case "control.agent.turn-opened":
+      return `turn opened · ${Diagnosis.firstLine(Diagnosis.asString(payload.seat) ?? "")}`
+    case "control.agent.model-settled": {
+      const usage = Diagnosis.asRecord(payload.usage)
+      return `model ${Diagnosis.asNumber(usage.inputTokens) ?? 0} in / ${
+        Diagnosis.asNumber(usage.outputTokens) ?? 0
+      } out`
+    }
+    case "control.agent.cell-call-started":
+      return `call ${Diagnosis.firstLine(Diagnosis.asString(payload.flowName) ?? "?")}`
+    case "control.agent.cell-call-settled":
+      return Diagnosis.asString(payload.outcome) === "failure"
+        ? `  -> FAIL ${Diagnosis.clip(Diagnosis.firstLine(Diagnosis.asString(payload.message) ?? ""), 100)}`
+        : "  -> ok"
+    case "control.agent.resolved":
+      return `resolved ${Diagnosis.clip(Diagnosis.firstLine(Diagnosis.asString(payload.text) ?? ""), 100)}`
+    case "control.approval.requested":
+      return `approval requested: ${Diagnosis.firstLine(Diagnosis.asString(payload.question) ?? "")}`
+    default:
+      return kind.slice("control.".length)
+  }
+}
+
+/** Fold health from the same authoritative summary and committed event buffer as the run card.
+ * Invalid observations and foreign incarnations cannot color the current owner.
+ * @category projections
+ * @since 1.0.0
+ */
+export const statusRollup = (
+  run: ControlSchema.RunSummary,
+  events: ReadonlyArray<ControlSchema.ControlEvent>,
+  now: number
+): Health.StatusRollup => {
+  const incarnation = Health.runIncarnation(run)
+  const candidates: Array<Health.RecordedObservation> = []
+  const decode = Schema.decodeUnknownOption(Health.HealthObservation)
+  let evidenceSeq = 0
+  for (const event of events) {
+    if (event.kind === Health.statusObservedEventType) {
+      const reading = decode(event.payload)
+      if (reading._tag === "Some") candidates.push({ observation: reading.value, sequence: event.sequence })
+    } else if (!Monitor.isBookkeepingEvent(event.kind)) evidenceSeq = Math.max(evidenceSeq, event.sequence)
+  }
+  const latest = Health.latestObservation(candidates, `run:${run.runId}`, incarnation)
+  return Health.rollup({
+    subjectId: `run:${run.runId}`,
+    state: run.status,
+    incarnation,
+    waitingReason: run.waitingReason,
+    baseHealth: latest?.observation.baseHealth ??
+      Monitor.classify({ summary: run, events, beatsWithoutProgress: 0, stallBeats: 3 }),
+    latest,
+    now,
+    evidenceSeq,
+    updatedAt: run.updatedAt
+  })
+}