agenticros 0.1.18 → 0.2.0

package/runtime/packages/core/src/mission.ts

@@ -0,0 +1,574 @@
+/**
+ * Mission runner — Phase 1.c of the AgenticROS strategy.
+ *
+ * A mission is a declarative sequence of capability invocations. Each step
+ * names a capability (from the registry exposed by `ros2_list_capabilities`)
+ * with literal or templated inputs, where templates can reference outputs
+ * from earlier steps via `{{stepId.outputs.fieldName}}`.
+ *
+ * The runner here is transport-agnostic — it takes a `dispatcher` callback
+ * that knows how to invoke a single tool for the host adapter. Adapters
+ * pass their existing tool dispatch path (handleToolCall in claude-code,
+ * tool.execute() in OpenClaw, executeTool() in Gemini), and the runner
+ * stays the same across all three.
+ *
+ * Phase 1.d adds multi-robot routing: `mission.robot_id` (optional) and
+ * per-step `inputs.robot_id` are auto-injected into every dispatched
+ * tool call when the binding's tool args don't already specify one. The
+ * adapter-side tool handlers then resolve the robot via
+ * `resolveRobotFromArgs(config, args)` and route the underlying
+ * publish/subscribe/etc. through that robot's namespace.
+ *
+ * Precedence (highest wins):
+ *   1. Tool args produced by the binding (`binding.buildArgs(inputs)`)
+ *      — if it sets `robot_id`, that's final.
+ *   2. Per-step `inputs.robot_id` (lets a mission mix robots step-by-step).
+ *   3. Mission-level `mission.robot_id` (the default for every step).
+ *   4. None — the adapter falls back to the active robot.
+ *
+ * Phase 1.f extends the runner with two additive capabilities:
+ *   - `cancellation`: a token the runner checks BEFORE each step. The
+ *     adapter holds the token in its `MissionRegistry` so a sibling
+ *     tool call (`mission_cancel(mission_id)`) can flip
+ *     `cancellation.cancelled` mid-run and stop the mission gracefully.
+ *     Steps that have already completed keep their results; the current
+ *     step finishes naturally (the runner doesn't preempt
+ *     `await dispatcher(...)` because that requires per-tool cancel
+ *     support — out of scope for Phase 1.f); subsequent steps are
+ *     marked `"cancelled"` and skipped.
+ *   - `transcript`: a callback invoked after every step with the
+ *     step's `MissionStepResult`. Adapters wire this to the shared
+ *     memory subsystem so a second agent can `recall(namespace=
+ *     "mission:<id>")` and inspect what's run so far. The runner
+ *     itself owns no I/O — the callback can be a no-op when memory
+ *     is disabled.
+ *
+ * What's still deferred:
+ *   - Parallel step execution (today: sequential only).
+ *   - Per-tool cancellation (cancel TIES to step boundaries today).
+ *   - Natural-language plan compilation (today: declarative only).
+ *   - Retry / backoff policies.
+ *
+ * See: docs/strategy-ai-agents-plus-ros.md §4 (Phase 1.c / 1.f).
+ */
+
+import type { Capability } from "./capabilities.js";
+
+/**
+ * Cancellation token consumed by `runMission`.
+ *
+ * Plain object (not AbortController) so it's easy to share across
+ * processes via a registry without pulling Web platform shims in. The
+ * runner only reads `cancelled`; `reason` is surfaced in the result
+ * for traceability.
+ */
+export interface MissionCancellationToken {
+  cancelled: boolean;
+  /** Optional free-text reason — bubbled up into the cancelled step results. */
+  reason?: string;
+}
+
+/**
+ * Per-step transcript sink. Called immediately after a step finishes
+ * (including cancelled / skipped steps) so an external store sees the
+ * whole timeline, not just the post-mortem `MissionResult`.
+ *
+ * The callback is best-effort: thrown errors are logged but never
+ * propagate up — losing a transcript entry must not abort an
+ * otherwise-healthy mission.
+ */
+export type MissionTranscriptSink = (
+  entry: MissionTranscriptEntry,
+) => Promise<void> | void;
+
+/** One transcript entry per executed step. */
+export interface MissionTranscriptEntry {
+  /** Unique id of the running mission (assigned by the adapter, not the runner). */
+  mission_id: string;
+  /** Free-text mission label. */
+  mission_name?: string;
+  /** Adapter that ran the mission (e.g. "claude-code", "openclaw", "gemini"). */
+  adapter?: string;
+  /** When the step started (ms since epoch). */
+  started_at: number;
+  /** Mission-level robot id (empty when unset). */
+  robot_id?: string;
+  /** Index of this step in mission.steps (0-based). */
+  step_index: number;
+  /** Total steps in the mission. */
+  step_total: number;
+  /** Snapshot of the per-step result. */
+  result: MissionStepResult;
+}
+
+/** One step in a mission. */
+export interface MissionStep {
+  /** Unique id within the mission, used by later steps for `{{id.outputs.x}}`. */
+  id: string;
+  /**
+   * Capability id (e.g. `drive_base`, `find_object`). The runner looks this
+   * up in the capability registry and uses a dispatch map to find the
+   * underlying tool to invoke.
+   */
+  capability: string;
+  /**
+   * Inputs to the capability. Values can be literals or `{{stepId.outputs.x}}`
+   * templates that resolve from prior steps' outputs.
+   */
+  inputs?: Record<string, unknown>;
+  /**
+   * Optional behaviour when this step fails. Defaults to "stop".
+   *   - "stop":     halt the mission, mark it failed
+   *   - "continue": record the failure and run the next step anyway
+   */
+  on_fail?: "stop" | "continue";
+}
+
+/** A complete mission plan. */
+export interface Mission {
+  /** Free-text label for logs / chat replies (e.g. "find chair and approach"). */
+  name?: string;
+  /**
+   * Optional planning notes from the agent — surfaced in the result for
+   * downstream summarisation. The runner ignores it.
+   */
+  goal?: string;
+  /**
+   * Optional default robot id for every step. Each step's tool args get
+   * `robot_id` auto-injected from this when the step doesn't supply its
+   * own and the binding's tool args don't already include one. Unknown
+   * robot ids surface as a tool error from the adapter, not from the
+   * runner. Empty/whitespace is ignored (uses the active robot).
+   */
+  robot_id?: string;
+  steps: MissionStep[];
+}
+
+/** Result of a single step. */
+export interface MissionStepResult {
+  id: string;
+  capability: string;
+  /**
+   * Final step outcome:
+   *   - "ok":       tool returned cleanly, outputs (if any) parsed.
+   *   - "error":    tool returned an error or the binding/build threw.
+   *   - "skipped":  earlier step failed with on_fail=stop.
+   *   - "cancelled": mission was cancelled (Phase 1.f) before this step ran.
+   */
+  status: "ok" | "error" | "skipped" | "cancelled";
+  /** Resolved inputs (with `{{...}}` templates substituted). */
+  inputs: Record<string, unknown>;
+  /** Outputs parsed from the tool response (may be `undefined` for fire-and-forget). */
+  outputs?: Record<string, unknown>;
+  /** Free-form text the tool returned (e.g. for human display). */
+  message?: string;
+  /** Error message when status is "error". */
+  error?: string;
+  /** Wall time in ms. */
+  duration_ms: number;
+}
+
+/** Result of a mission run. */
+export interface MissionResult {
+  /**
+   * Overall mission outcome:
+   *   - "ok":        every step succeeded (or skipped on_fail=continue).
+   *   - "error":     one or more steps errored with on_fail=stop.
+   *   - "cancelled": mission was cancelled mid-run (Phase 1.f). Some
+   *                  steps may have completed successfully — see steps[].
+   */
+  status: "ok" | "error" | "cancelled";
+  /** Number of steps that ran (skipped/cancelled steps are NOT counted as run). */
+  steps_run: number;
+  steps_total: number;
+  /** Per-step results, in declaration order. */
+  steps: MissionStepResult[];
+  /** Total wall time across all steps. */
+  duration_ms: number;
+  /** Brief one-line summary suitable for chat reply. */
+  summary: string;
+  /**
+   * Mission-level default robot id used during this run (surfaced for
+   * traceability — useful in chat replies and logs). Empty string when
+   * the mission didn't pin a robot.
+   */
+  robot_id?: string;
+  /**
+   * Phase 1.f — the mission id the adapter registered before running.
+   * Empty when the adapter didn't pass an id (e.g. ad-hoc dispatcher
+   * tests). Use this with `mission_cancel(mission_id)` to abort a
+   * mission mid-run, or `memory_recall(namespace="mission:<id>")` to
+   * read the transcript later.
+   */
+  mission_id?: string;
+  /**
+   * Phase 1.f — set when the mission was cancelled. Carries the free-text
+   * reason the canceller supplied (or "cancelled" when none was given).
+   */
+  cancellation_reason?: string;
+}
+
+/**
+ * A dispatcher invokes a single MCP tool by name and returns:
+ *   - `text`: the tool's text response (free-form, for human display)
+ *   - `outputs`: optional structured outputs (parsed JSON when the text
+ *                payload is a JSON blob). The runner uses these for
+ *                `{{...}}` template resolution by later steps.
+ *
+ * Adapters wrap their existing tool entry points to satisfy this contract.
+ */
+export type MissionToolDispatcher = (
+  toolName: string,
+  args: Record<string, unknown>,
+) => Promise<{ text: string; outputs?: Record<string, unknown>; isError?: boolean }>;
+
+/**
+ * Capability → MCP tool mapping. Each entry says "to satisfy capability X,
+ * call tool Y with these arguments". The transform receives the step's
+ * (already template-resolved) inputs and returns the args object the tool
+ * expects.
+ *
+ * Why a separate map (not a field on the capability itself)?
+ *   - Some capabilities map onto different tools per adapter
+ *     (claude-code's `ros2_find_object` is a single tool call; the OpenClaw
+ *     adapter equivalent runs in a different process).
+ *   - Skill-declared capabilities will eventually carry their own
+ *     implementation hint (`implementation.kind = "in_process"|"external_ros_node"`),
+ *     but until that wire-up is real, mapping happens here.
+ */
+export interface CapabilityToolBinding {
+  /** The MCP tool name to invoke. */
+  tool: string;
+  /** Transform the step's resolved inputs into the tool's expected arguments. */
+  buildArgs: (inputs: Record<string, unknown>) => Record<string, unknown>;
+  /**
+   * Optional: extract structured outputs from the tool's text response
+   * for `{{...}}` template references by later steps. When omitted, the
+   * runner attempts to JSON-parse the text and use the result.
+   */
+  parseOutputs?: (text: string) => Record<string, unknown> | undefined;
+}
+
+export type CapabilityToolBindings = Record<string, CapabilityToolBinding>;
+
+const TEMPLATE_RE = /\{\{\s*([a-zA-Z0-9_]+)\.outputs\.([a-zA-Z0-9_]+)\s*\}\}/g;
+
+/**
+ * Substitute `{{stepId.outputs.field}}` references in `value` using the
+ * given step output map. Recurses into nested objects/arrays.
+ *
+ * String values containing exactly one template reference are replaced with
+ * the raw value (preserving type — number stays number, etc.). Strings with
+ * interpolated text get string-substituted.
+ */
+function resolveTemplates(
+  value: unknown,
+  outputs: Record<string, Record<string, unknown> | undefined>,
+): unknown {
+  if (typeof value === "string") {
+    const matches = [...value.matchAll(TEMPLATE_RE)];
+    if (matches.length === 0) return value;
+    // Whole-string single template: preserve type.
+    if (matches.length === 1) {
+      const m = matches[0];
+      if (m[0] === value.trim()) {
+        const stepOuts = outputs[m[1]];
+        return stepOuts ? stepOuts[m[2]] : undefined;
+      }
+    }
+    return value.replace(TEMPLATE_RE, (_, stepId: string, field: string) => {
+      const stepOuts = outputs[stepId];
+      const v = stepOuts ? stepOuts[field] : undefined;
+      return v === undefined ? "" : String(v);
+    });
+  }
+  if (Array.isArray(value)) return value.map((v) => resolveTemplates(v, outputs));
+  if (value && typeof value === "object") {
+    const out: Record<string, unknown> = {};
+    for (const [k, v] of Object.entries(value as Record<string, unknown>)) {
+      out[k] = resolveTemplates(v, outputs);
+    }
+    return out;
+  }
+  return value;
+}
+
+function tryParseJsonOutputs(text: string): Record<string, unknown> | undefined {
+  // Tools often return "Summary line.\n{json}" — try parsing each line plus
+  // the whole payload, returning the first JSON object we find.
+  const candidates = [text, text.trim(), text.split("\n").pop() ?? ""];
+  for (const c of candidates) {
+    if (!c || !c.trim().startsWith("{")) continue;
+    try {
+      const parsed = JSON.parse(c);
+      if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
+        return parsed as Record<string, unknown>;
+      }
+    } catch {
+    }
+  }
+  return undefined;
+}
+
+/** Phase 1.f optional behaviour bundle for `runMission`. */
+export interface RunMissionOptions {
+  /**
+   * Stable identifier for this mission run. Adapters generate it and
+   * register a cancellation token under this key, so a sibling
+   * `mission_cancel(mission_id)` tool call can flip the token mid-run.
+   * Surfaced in the result + every transcript entry.
+   */
+  mission_id?: string;
+  /**
+   * Cancellation token the runner checks BEFORE each step. When
+   * `cancelled` is true at a step boundary, the remaining steps are
+   * marked "cancelled" and the mission returns with
+   * `status: "cancelled"`.
+   */
+  cancellation?: MissionCancellationToken;
+  /**
+   * Best-effort transcript sink — called after every step finishes
+   * (including skipped/cancelled). Adapters wire this to the memory
+   * subsystem so a downstream agent can read the timeline. Thrown
+   * errors are swallowed; transcripts are not gated by the mission's
+   * success.
+   */
+  transcript?: MissionTranscriptSink;
+  /** Adapter label (e.g. "claude-code", "openclaw", "gemini") — copied into transcripts. */
+  adapter?: string;
+}
+
+/**
+ * Execute a mission sequentially.
+ *
+ * @param mission     The declarative plan.
+ * @param capabilities Capability registry — used to validate that each
+ *                    requested capability id is real.
+ * @param bindings    Capability → tool mapping; entries missing here yield
+ *                    an "unsupported capability" error for that step.
+ * @param dispatcher  Adapter-supplied function that invokes a tool by name.
+ * @param options     Phase 1.f options (cancellation token, transcript sink, mission id).
+ *                    Omit for legacy callers.
+ */
+export async function runMission(
+  mission: Mission,
+  capabilities: Capability[],
+  bindings: CapabilityToolBindings,
+  dispatcher: MissionToolDispatcher,
+  options?: RunMissionOptions,
+): Promise<MissionResult> {
+  const capIds = new Set(capabilities.map((c) => c.id));
+  const stepOutputs: Record<string, Record<string, unknown> | undefined> = {};
+  const results: MissionStepResult[] = [];
+  const t0Mission = Date.now();
+  let aborted = false;
+  let cancelled = false;
+
+  // Phase 1.f — best-effort transcript emit. Synchronous helper so the
+  // mission loop stays linear; the caller's sink may itself be async
+  // but we don't block the next step on it.
+  const emitTranscript = (idx: number, started_at: number, result: MissionStepResult): void => {
+    if (!options?.transcript) return;
+    try {
+      const entry: MissionTranscriptEntry = {
+        mission_id: options.mission_id ?? "",
+        mission_name: mission.name,
+        adapter: options.adapter,
+        started_at,
+        robot_id: typeof mission.robot_id === "string" ? mission.robot_id : "",
+        step_index: idx,
+        step_total: mission.steps.length,
+        result,
+      };
+      // Fire-and-forget — never throw out of the runner. We catch
+      // synchronous throws AND attach a no-op .catch on the promise so
+      // a rejected async sink doesn't surface as an unhandled rejection.
+      const ret = options.transcript(entry);
+      if (ret && typeof (ret as Promise<unknown>).catch === "function") {
+        (ret as Promise<unknown>).catch(() => {});
+      }
+    } catch {
+      // Transcript loss never breaks the mission.
+    }
+  };
+
+  for (let idx = 0; idx < mission.steps.length; idx++) {
+    const step = mission.steps[idx];
+    const t0 = Date.now();
+
+    // Phase 1.f — honour cancellation token at the step boundary. We
+    // check BEFORE the step runs so that a cancel arriving while
+    // step N is in-flight only affects step N+1 onward (per-tool
+    // preemption is out of scope; see the module header).
+    if (!cancelled && options?.cancellation?.cancelled) {
+      cancelled = true;
+    }
+
+    if (cancelled) {
+      const result: MissionStepResult = {
+        id: step.id,
+        capability: step.capability,
+        status: "cancelled",
+        inputs: {},
+        message: `Cancelled: ${options?.cancellation?.reason ?? "cancelled"}`,
+        duration_ms: 0,
+      };
+      results.push(result);
+      emitTranscript(idx, t0, result);
+      continue;
+    }
+
+    if (aborted) {
+      const result: MissionStepResult = {
+        id: step.id,
+        capability: step.capability,
+        status: "skipped",
+        inputs: {},
+        message: "Skipped: earlier step failed and on_fail=stop.",
+        duration_ms: 0,
+      };
+      results.push(result);
+      emitTranscript(idx, t0, result);
+      continue;
+    }
+
+    if (!capIds.has(step.capability)) {
+      const result: MissionStepResult = {
+        id: step.id,
+        capability: step.capability,
+        status: "error",
+        inputs: step.inputs ?? {},
+        error: `Capability "${step.capability}" not found in registry. Use ros2_list_capabilities to see what's available.`,
+        duration_ms: Date.now() - t0,
+      };
+      results.push(result);
+      emitTranscript(idx, t0, result);
+      if ((step.on_fail ?? "stop") === "stop") aborted = true;
+      continue;
+    }
+
+    const binding = bindings[step.capability];
+    if (!binding) {
+      const result: MissionStepResult = {
+        id: step.id,
+        capability: step.capability,
+        status: "error",
+        inputs: step.inputs ?? {},
+        error: `Capability "${step.capability}" is registered but has no mission-runner tool binding yet. See docs/strategy-ai-agents-plus-ros.md §4 Phase 1.c.`,
+        duration_ms: Date.now() - t0,
+      };
+      results.push(result);
+      emitTranscript(idx, t0, result);
+      if ((step.on_fail ?? "stop") === "stop") aborted = true;
+      continue;
+    }
+
+    const resolvedInputs = (resolveTemplates(step.inputs ?? {}, stepOutputs) as Record<string, unknown>) ?? {};
+
+    let toolArgs: Record<string, unknown>;
+    try {
+      toolArgs = binding.buildArgs(resolvedInputs);
+    } catch (err) {
+      const result: MissionStepResult = {
+        id: step.id,
+        capability: step.capability,
+        status: "error",
+        inputs: resolvedInputs,
+        error: `Failed to build args for tool ${binding.tool}: ${err instanceof Error ? err.message : String(err)}`,
+        duration_ms: Date.now() - t0,
+      };
+      results.push(result);
+      emitTranscript(idx, t0, result);
+      if ((step.on_fail ?? "stop") === "stop") aborted = true;
+      continue;
+    }
+
+    // Inject robot_id when the binding didn't set one. Per-step inputs
+    // take precedence over the mission-level default; empty strings
+    // mean "use the active robot" so we ignore them. This is what
+    // makes per-tool fleet routing work end-to-end from `run_mission`.
+    if (!("robot_id" in toolArgs) || typeof toolArgs.robot_id !== "string" || toolArgs.robot_id.trim().length === 0) {
+      const stepRid = typeof resolvedInputs.robot_id === "string" ? resolvedInputs.robot_id.trim() : "";
+      const missionRid = typeof mission.robot_id === "string" ? mission.robot_id.trim() : "";
+      const effective = stepRid || missionRid;
+      if (effective) toolArgs.robot_id = effective;
+    }
+
+    try {
+      const dispatched = await dispatcher(binding.tool, toolArgs);
+      const outputs =
+        dispatched.outputs ??
+        binding.parseOutputs?.(dispatched.text) ??
+        tryParseJsonOutputs(dispatched.text);
+      stepOutputs[step.id] = outputs;
+      const status: MissionStepResult["status"] = dispatched.isError ? "error" : "ok";
+      const result: MissionStepResult = {
+        id: step.id,
+        capability: step.capability,
+        status,
+        inputs: resolvedInputs,
+        outputs,
+        message: dispatched.text,
+        ...(dispatched.isError ? { error: dispatched.text } : {}),
+        duration_ms: Date.now() - t0,
+      };
+      results.push(result);
+      emitTranscript(idx, t0, result);
+      if (status === "error" && (step.on_fail ?? "stop") === "stop") aborted = true;
+    } catch (err) {
+      const msg = err instanceof Error ? err.message : String(err);
+      const result: MissionStepResult = {
+        id: step.id,
+        capability: step.capability,
+        status: "error",
+        inputs: resolvedInputs,
+        error: msg,
+        duration_ms: Date.now() - t0,
+      };
+      results.push(result);
+      emitTranscript(idx, t0, result);
+      if ((step.on_fail ?? "stop") === "stop") aborted = true;
+    }
+  }
+
+  // "Ran" counts steps that actually executed — skipped (on_fail=stop)
+  // and cancelled (Phase 1.f) are explicitly excluded.
+  const ran = results.filter((r) => r.status !== "skipped" && r.status !== "cancelled").length;
+  const failed = results.filter((r) => r.status === "error").length;
+  let overallStatus: MissionResult["status"] = failed === 0 ? "ok" : "error";
+  if (cancelled) overallStatus = "cancelled";
+  const summary = summariseMission(mission, results, overallStatus);
+
+  return {
+    status: overallStatus,
+    steps_run: ran,
+    steps_total: mission.steps.length,
+    steps: results,
+    duration_ms: Date.now() - t0Mission,
+    summary,
+    robot_id: typeof mission.robot_id === "string" ? mission.robot_id : "",
+    ...(options?.mission_id ? { mission_id: options.mission_id } : {}),
+    ...(cancelled
+      ? { cancellation_reason: options?.cancellation?.reason ?? "cancelled" }
+      : {}),
+  };
+}
+
+function summariseMission(
+  mission: Mission,
+  results: MissionStepResult[],
+  status: MissionResult["status"],
+): string {
+  const label = mission.name ?? "Mission";
+  const ran = results.filter((r) => r.status !== "skipped" && r.status !== "cancelled").length;
+  const cancelled = results.filter((r) => r.status === "cancelled").length;
+  const failed = results.filter((r) => r.status === "error");
+  if (status === "ok") {
+    return `${label}: completed ${ran}/${results.length} step(s) successfully.`;
+  }
+  if (status === "cancelled") {
+    return `${label}: cancelled after ${ran} step(s) — ${cancelled} remaining step(s) skipped.`;
+  }
+  const first = failed[0];
+  return `${label}: failed at step "${first?.id ?? "?"}" (${first?.capability ?? "?"}): ${first?.error ?? "unknown error"}.`;
+}