@marcoscale98/pi-extensible-workflows 5.17.0-fork.1 → 6.2.1-fork.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (128) hide show
  1. package/CHANGELOG.md +145 -0
  2. package/README.md +2 -2
  3. package/dist/src/agent-execution.d.ts +43 -10
  4. package/dist/src/agent-execution.js +442 -162
  5. package/dist/src/background-widget.d.ts +5 -1
  6. package/dist/src/background-widget.js +5 -10
  7. package/dist/src/decoders.d.ts +2 -3
  8. package/dist/src/decoders.js +59 -37
  9. package/dist/src/execution.d.ts +6 -2
  10. package/dist/src/execution.js +48 -7
  11. package/dist/src/host-delivery.js +5 -8
  12. package/dist/src/host-navigator.js +5 -2
  13. package/dist/src/host-phases.d.ts +1 -0
  14. package/dist/src/host-recovery.d.ts +11 -8
  15. package/dist/src/host-recovery.js +14 -22
  16. package/dist/src/host-runtime.js +56 -5
  17. package/dist/src/host-view.d.ts +6 -1
  18. package/dist/src/host-view.js +46 -20
  19. package/dist/src/host.d.ts +2 -2
  20. package/dist/src/host.js +246 -118
  21. package/dist/src/index.d.ts +1 -2
  22. package/dist/src/index.js +1985 -1690
  23. package/dist/src/index.js.map +3 -3
  24. package/dist/src/paths.d.ts +8 -5
  25. package/dist/src/paths.js +24 -13
  26. package/dist/src/persistence.d.ts +1 -0
  27. package/dist/src/persistence.js +1 -0
  28. package/dist/src/pi-runtime-runner.js +4 -1
  29. package/dist/src/registry.d.ts +9 -7
  30. package/dist/src/registry.js +29 -71
  31. package/dist/src/runtime/agent-runner.d.ts +2 -0
  32. package/dist/src/script-tools.d.ts +27 -0
  33. package/dist/src/script-tools.js +87 -0
  34. package/dist/src/session-lease.d.ts +1 -1
  35. package/dist/src/session-lease.js +45 -11
  36. package/dist/src/settings.d.ts +19 -3
  37. package/dist/src/settings.js +63 -50
  38. package/dist/src/store.d.ts +3 -1
  39. package/dist/src/store.js +17 -1
  40. package/dist/src/trajectory.d.ts +12 -3
  41. package/dist/src/trajectory.js +32 -11
  42. package/dist/src/types.d.ts +129 -45
  43. package/dist/src/types.js +5 -8
  44. package/dist/src/utils.d.ts +21 -1
  45. package/dist/src/utils.js +46 -6
  46. package/dist/src/validation.d.ts +14 -3
  47. package/dist/src/validation.js +143 -87
  48. package/dist/starter/index.js +8 -11
  49. package/dist/starter/index.js.map +1 -1
  50. package/dist/starter/review-loop.d.ts +0 -13
  51. package/dist/starter/review-loop.js +8 -9
  52. package/dist/subagents/index.js +904 -956
  53. package/dist/subagents/index.js.map +3 -3
  54. package/dist/subagents/src/contracts.d.ts +48 -31
  55. package/dist/subagents/src/contracts.js +20 -9
  56. package/dist/subagents/src/index.d.ts +1 -1
  57. package/dist/subagents/src/index.js +14 -8
  58. package/dist/subagents/src/manager.js +52 -73
  59. package/dist/subagents/src/navigator.js +3 -14
  60. package/dist/subagents/src/view.d.ts +1 -1
  61. package/dist/subagents/src/view.js +15 -68
  62. package/dist/subagents/src/worktree.js +1 -1
  63. package/dist/trajectory/assets/index.html +35 -13
  64. package/dist/trajectory/index.js +576 -932
  65. package/dist/trajectory/index.js.map +3 -3
  66. package/dist/trajectory/src/assets/index.html +35 -13
  67. package/dist/trajectory/src/index.js +134 -104
  68. package/dist/trajectory/src/server.js +268 -27
  69. package/dist/trajectory/src/server.js.map +3 -3
  70. package/dist/trajectory/src/startup-mutex.d.ts +17 -0
  71. package/dist/trajectory/src/startup-mutex.js +172 -0
  72. package/examples/workflow-extension-template/README.md +4 -7
  73. package/examples/workflow-extension-template/extension.test.mjs +4 -8
  74. package/examples/workflow-extension-template/index.js +0 -3
  75. package/package.json +10 -10
  76. package/skills/pi-extensible-workflows/SKILL.md +12 -8
  77. package/src/agent-execution.ts +412 -164
  78. package/src/background-widget.ts +5 -11
  79. package/src/decoders.ts +51 -37
  80. package/src/execution.ts +46 -9
  81. package/src/host-delivery.ts +5 -8
  82. package/src/host-navigator.ts +4 -2
  83. package/src/host-phases.ts +1 -1
  84. package/src/host-recovery.ts +23 -28
  85. package/src/host-runtime.ts +44 -6
  86. package/src/host-view.ts +41 -16
  87. package/src/host.ts +229 -142
  88. package/src/index.ts +1 -2
  89. package/src/paths.ts +18 -9
  90. package/src/persistence.ts +1 -0
  91. package/src/pi-runtime-runner.ts +3 -1
  92. package/src/registry.ts +28 -64
  93. package/src/runtime/agent-runner.ts +2 -0
  94. package/src/script-tools.ts +76 -0
  95. package/src/session-lease.ts +34 -6
  96. package/src/settings.ts +56 -50
  97. package/src/store.ts +18 -2
  98. package/src/trajectory.ts +32 -15
  99. package/src/types.ts +82 -34
  100. package/src/utils.ts +42 -5
  101. package/src/validation.ts +121 -80
  102. package/starter/index.ts +0 -1
  103. package/starter/review-loop.ts +8 -9
  104. package/starter/roles/developer.md +0 -1
  105. package/starter/roles/oracle.md +0 -1
  106. package/starter/roles/researcher.md +0 -1
  107. package/starter/roles/reviewer.md +0 -1
  108. package/starter/roles/scout.md +0 -1
  109. package/subagents/README.md +10 -16
  110. package/subagents/src/contracts.ts +21 -11
  111. package/subagents/src/index.ts +17 -8
  112. package/subagents/src/manager.ts +48 -128
  113. package/subagents/src/navigator.ts +4 -13
  114. package/subagents/src/view.ts +16 -63
  115. package/subagents/src/worktree.ts +1 -1
  116. package/trajectory/src/assets/index.html +35 -13
  117. package/trajectory/src/index.ts +112 -91
  118. package/trajectory/src/server.ts +46 -10
  119. package/trajectory/src/startup-mutex.ts +127 -0
  120. package/dist/src/roles.d.ts +0 -70
  121. package/dist/src/roles.js +0 -325
  122. package/dist/starter/roles/developer.md +0 -6
  123. package/dist/starter/roles/oracle.md +0 -18
  124. package/dist/starter/roles/researcher.md +0 -20
  125. package/dist/starter/roles/reviewer.md +0 -19
  126. package/dist/starter/roles/scout.md +0 -18
  127. package/examples/workflow-extension-template/roles/reviewer.md +0 -4
  128. package/src/roles.ts +0 -326
package/src/host.ts CHANGED
@@ -4,8 +4,8 @@ import { copyFile, mkdir } from "node:fs/promises";
4
4
  import { dirname, join } from "node:path";
5
5
  import { fileURLToPath } from "node:url";
6
6
  import { Type, type Api, type Model, type Static, type TSchema } from "@earendil-works/pi-ai";
7
- import { copyToClipboard, getAgentDir, ModelSelectorComponent, type ExtensionAPI, type ExtensionContext, type ModelRuntime, type ToolDefinition } from "@earendil-works/pi-coding-agent";
8
- import { FairAgentScheduler, getAgentAttempts, WorkflowAgentExecutor, localAgentTransport, type AgentActivity, type AgentAttempt, type AgentDefinition, type AgentExecutionRoot, type AgentProgress, type AgentProviderFailure, type AgentProviderRecovery } from "./agent-execution.js";
7
+ import { copyToClipboard, getAgentDir, ModelSelectorComponent, type ExtensionAPI, type ExtensionContext, type ExtensionToolContext, type ModelRuntime, type ToolDefinition } from "@earendil-works/pi-coding-agent";
8
+ import { FairAgentScheduler, getAgentAttempts, WorkflowAgentExecutor, localAgentTransport, type AgentActivity, type AgentAttempt, type AgentExecutionRoot, type AgentProgress, type AgentProviderFailure, type AgentProviderRecovery } from "./agent-execution.js";
9
9
  import { RunLifecycle, WorkflowEventPublisher, hostSessionContext, nextNamedOccurrence, withWorkflowFunctions, withoutActiveShells, workflowRunContext, type WorkflowEventSink, type WorkflowRunRecord, type WorkflowToolUpdate } from "./host-runtime.js";
10
10
  import { createWorkflowRecovery, persistedFailure, type ModelRegistryCapability } from "./host-recovery.js";
11
11
  import { registerWorkflowNavigator, uiHostCapabilities } from "./host-navigator.js";
@@ -13,53 +13,22 @@ import { acquireSessionLease, isPersistedRun, listPersistedSessionIds, listRunId
13
13
  import { retainTerminalRuns } from "./retention.js";
14
14
  import type { PersistedRun, WorktreeReference } from "./persistence.js";
15
15
  import { validateBudget, WorkflowBudgetRuntime } from "./budget.js";
16
- import { SerialLane, asWorkflowError, createLaunchSnapshot, errorCode, errorText, fail, isNodeError, jsonValue, mergeWorkflowExtensionSettings, modelAliasErrorName, modelCapability, object, parseModelReference, positiveInteger, sanitizeDisplayText, validateModelAliases } from "./utils.js";
17
- import { loadSettings, preflight, resolveAgentResourcePolicy, resolveWorkflowSettings, validateCheckpoint, validateModelAliasAvailability, validateWorkflowLaunchWithRegistry, workflowProjectSettingsPath, workflowSettingsPath } from "./validation.js";
18
- import { loadAgentDefinitions, loadProjectAgentDefinitions } from "./roles.js";
16
+ import { SerialLane, asWorkflowError, createLaunchSnapshot, errorCode, errorText, fail, isNodeError, jsonValue, modelAliasErrorName, object, parseModelReference, positiveInteger, reachableTools, resolveModelReference, VIRTUAL_MODEL_PROVIDER, sanitizeDisplayText, validateModelAliases } from "./utils.js";
17
+ import { loadCodemodeToolsSetting, preflight, resolveAgentResourcePolicy, staticAgentPreparationOptions, resolveWorkflowSettings, validateCheckpoint, validateModelAliasAvailability, validateWorkflowLaunchWithRegistry, workflowProjectSettingsPath, workflowSettingsPath, workflowToolExposure } from "./validation.js";
19
18
  import { beginWorkflowExtensionLoading, loadingRegistry, resetWorkflowRegistryIfIdle, retainWorkflowRegistry, type WorkflowRegistryApi } from "./registry.js";
20
- import { agentHandleTurnPath, agentIdentityPath, agentWorktree, encoded, executeShellCommand, persistActiveAgentAttempt, persistAgentAttempts, readShellResult, runWorkflow, shellIdentityPath } from "./execution.js";
19
+ import { agentHandleTurnPath, agentIdentityPath, agentWorktree, encoded, executeShellCommand, persistActiveAgentAttempt, persistAgentAttempts, readShellResult, runWorkflow, shellIdentityPath, toolIdentityPath } from "./execution.js";
20
+ import { prepareScriptToolLoadout, scriptTool, scriptToolContext, scriptToolValue, validateScriptToolReferences } from "./script-tools.js";
21
21
  import backgroundWidget, { type BackgroundWidgetAPI } from "./background-widget.js";
22
22
  import { showChangelogNotice } from "./changelog.js";
23
23
  import { createTrajectoryRunLoader, createTrajectoryRunMetadataLoader, createTrajectorySubagentLoader, createTrajectorySubagentMetadataLoader, createTrajectoryTranscriptLoader, type TrajectoryActionRequest, type TrajectoryActionResult, type TrajectorySubagent } from "./trajectory.js";
24
24
  import { getTrajectoryHost, type TrajectoryPublisherProvider } from "./trajectory-host-handle.js";
25
25
  import { getSubagentManager } from "./subagent-manager-handle.js";
26
- import { HARD_TERMINAL_RUN_STATES, LAUNCH_SNAPSHOT_IDENTITY_VERSION, WORKFLOW_BLOCKED_EVENT, WorkflowError, isContextFileScope, isExternallyEndedRunState, isHardTerminalRunState, roleNameOf, type AgentAccounting, type AgentIdentity, type AgentRecord, type AgentResourcePolicy, type AgentTransport, type JsonValue, type LaunchSnapshot, type LiveSessionHandoff, type HardTerminalRunState, type ModelSpec, type PreparedAgentSession, type RunState, type ShellIdentity, type ShellOptions, type ShellResult, type WorkflowAgentSession, type WorkflowErrorCode, type WorkflowMetadata, type WorkflowModelAliasResolverContext, type WorkflowSettings, type WorkflowSettingsResolution, type WorkflowWorktreeReference } from "./types.js";
26
+ import { HARD_TERMINAL_RUN_STATES, LAUNCH_SNAPSHOT_IDENTITY_VERSION, THINKING_LEVELS, WORKFLOW_BLOCKED_EVENT, WorkflowError, type WorkflowFunction, isContextFileScope, isExternallyEndedRunState, isHardTerminalRunState, type AgentAccounting, type AgentIdentity, type AgentRecord, type AgentResourcePolicy, type AgentTransport, type JsonValue, type LaunchSnapshot, type LiveSessionHandoff, type HardTerminalRunState, type ModelSpec, type PreparedAgentSession, type RunState, type ShellIdentity, type ShellOptions, type ShellResult, type ToolIdentity, type WorkflowAgentSession, type WorkflowErrorCode, type WorkflowMetadata, type WorkflowModelAliasResolverContext, type WorkflowSettings, type WorkflowScriptCall, type WorkflowSettingsResolution, type WorkflowWorktreeReference } from "./types.js";
27
27
  import type { SubagentManagerContext, SubagentRunRequest, SubagentStatus } from "../subagents/src/contracts.js";
28
- import {
29
- SETTLED_AGENT_STATES,
30
- catalogResultValue,
31
- formatWorkflowCatalog,
32
- styledTextBlock,
33
- textBlock,
34
- workflowCatalogBlock,
35
- workflowControlCall,
36
- workflowControlResult,
37
- workflowProgressBlock,
38
- type WorkflowProgressRenderState,
39
- } from "./host-view.js";
40
- import {
41
- DELIVERY_LIMIT_BYTES,
42
- ForegroundDeliveryController,
43
- markWorkflowFailureDiagnostics,
44
- WORKFLOW_LOG_ENTRY,
45
- completionDescriptor,
46
- completionDeliveryFromStore,
47
- createWorkflowFailureDiagnostics,
48
- failureDiagnosticsFrom,
49
- formatWorkflowFailure,
50
- formatWorkflowFailureDelivery,
51
- formatWorkflowFailureDeliveryFallback,
52
- formatWorkflowFailureDiagnostics,
53
- isWorkflowFailureDiagnostics,
54
- serializeWorkflowFailureDiagnostics,
55
- utf8Prefix,
56
- type CompletionDeliveryContext,
57
- type ForegroundDelivery,
58
- type ForegroundDetachResult,
59
- type WorkflowLogEntry,
60
- } from "./host-delivery.js";
28
+ import { SETTLED_AGENT_STATES, catalogResultValue, formatWorkflowCatalog, styledTextBlock, textBlock, workflowCatalogBlock, workflowControlCall, workflowControlResult, workflowProgressBlock, type WorkflowProgressRenderState } from "./host-view.js";
29
+ import { DELIVERY_LIMIT_BYTES, ForegroundDeliveryController, markWorkflowFailureDiagnostics, WORKFLOW_LOG_ENTRY, completionDescriptor, completionDeliveryFromStore, createWorkflowFailureDiagnostics, failureDiagnosticsFrom, formatWorkflowFailure, formatWorkflowFailureDelivery, formatWorkflowFailureDeliveryFallback, formatWorkflowFailureDiagnostics, isWorkflowFailureDiagnostics, serializeWorkflowFailureDiagnostics, utf8Prefix, type CompletionDeliveryContext, type ForegroundDelivery, type ForegroundDetachResult, type WorkflowLogEntry } from "./host-delivery.js";
61
30
 
62
- export type WorkflowExtensionAPI = Pick<ExtensionAPI, "appendEntry" | "getActiveTools" | "getThinkingLevel" | "on" | "registerCommand" | "registerTool" | "sendMessage"> & Pick<BackgroundWidgetAPI, "events" | "registerEntryRenderer" | "registerShortcut">;
31
+ export type WorkflowExtensionAPI = Pick<ExtensionAPI, "appendEntry" | "getActiveTools" | "getThinkingLevel" | "on" | "registerCommand" | "registerTool" | "sendMessage"> & Partial<Pick<ExtensionAPI, "getAllTools">> & Pick<BackgroundWidgetAPI, "events" | "registerEntryRenderer" | "registerShortcut">;
63
32
 
64
33
  export {
65
34
  agentBreadcrumb,
@@ -117,6 +86,13 @@ function mainAgentError(error: unknown): WorkflowError {
117
86
  Object.assign(presented, typed);
118
87
  return presented;
119
88
  }
89
+ /** The structured counterpart of {@link completionControlContent}: the result without the persisted run, with the full value. */
90
+ function completionControlValue(result: Readonly<Record<string, unknown>>, controlRunId?: string): Record<string, unknown> {
91
+ const record: Record<string, unknown> = { ...(controlRunId === undefined ? {} : { runId: controlRunId }), ...result };
92
+ delete record.run;
93
+ delete record.completion;
94
+ return record;
95
+ }
120
96
  function completionControlContent(result: unknown, controlRunId?: string): string {
121
97
  const record = object(result) ? { ...result } : undefined;
122
98
  if (record && controlRunId !== undefined && record.runId === undefined) record.runId = controlRunId;
@@ -133,33 +109,13 @@ function completionControlContent(result: unknown, controlRunId?: string): strin
133
109
  delete record.run;
134
110
  return JSON.stringify({ ...record, value });
135
111
  }
136
- type RoleCapture = { readonly snapshot: Readonly<LaunchSnapshot>; capture(role: string, model: ModelSpec): Promise<void> };
137
- /** Persists a role definition (and its model) into the launch snapshot the first time an agent uses it, so later replays and retries see the definition the run actually ran with. */
138
- function roleCapture(store: RunStore, initial: Readonly<LaunchSnapshot>, definitions: Readonly<Record<string, AgentDefinition>>, projectDefinitions: Readonly<Record<string, AgentDefinition>>): RoleCapture {
139
- let persisted = initial;
140
- return {
141
- get snapshot() { return persisted; },
142
- async capture(role, model) {
143
- const definition = definitions[role];
144
- if (!definition) return;
145
- const modelName = `${model.provider}/${model.model}`;
146
- const hasProjectRole = projectDefinitions[role] !== undefined;
147
- if (persisted.roles?.[role] !== undefined && (!hasProjectRole || persisted.projectRoles?.includes(role)) && persisted.models.includes(modelName)) return;
148
- const roles = { ...(persisted.roles ?? {}), [role]: definition };
149
- const projectRoles = hasProjectRole ? [...new Set([...(persisted.projectRoles ?? []), role])] : persisted.projectRoles ?? [];
150
- const models = [...new Set([...persisted.models, modelName])];
151
- persisted = createLaunchSnapshot({ ...persisted, models, roles, projectRoles });
152
- await store.saveSnapshot(persisted);
153
- },
154
- };
155
- }
156
112
  export function formatWorkflowPreview(args: { script?: unknown; scriptPath?: unknown; name?: unknown; description?: unknown }): string {
157
113
  const name = typeof args.name === "string" && args.name.trim() ? args.name.trim() : "workflow";
158
114
  if (typeof args.script !== "string" || !args.script.trim()) return `workflow ${name}`;
159
115
  return [`workflow ${name}`, typeof args.description === "string" && args.description.trim() ? args.description.trim() : ""].filter(Boolean).join("\n");
160
116
  }
161
117
  export const WORKFLOW_TOOL_LABEL = "Workflow";
162
- export const WORKFLOW_TOOL_DESCRIPTION = "Run a deterministic JavaScript workflow with a named inline or file-backed parallel-to-summary path by default"
118
+ export const WORKFLOW_TOOL_DESCRIPTION = "Run a deterministic JavaScript workflow with a named inline or file-backed parallel-to-summary path by default. Scripts call this session's tools as `await tools.<name>(args)`, outside withWorktree: a tool resolves to its text unless its description gives a `tools.<name>(args)` result shape, and failures reject"
163
119
  export const WORKFLOW_TOOL_PROMPT_SNIPPET = "Run a deterministic, resumable JavaScript workflow. Prefer a named inline script that fans out independent work with parallel(...), awaits the keyed results before interpolating them into one summarizing agent(...), and returns. Provide exactly one of script or scriptPath and a non-empty name. Registered catalog functions are available as globals inside the script; call them there, for example return await someFunction(args). Advanced controls include registered functions, outputSchema, budgets, checkpoints, worktrees, retry/resume, CLI export, and pipelines. Runs are in the background by default; completion arrives as a follow-up message. Set foreground: true when the caller must wait for the final value. Manage runs from the interactive /workflow picker; use workflow_status, workflow_resume, workflow_retry, workflow_stop, and workflow_respond for explicit tool controls. If a foreground call detaches before its result is accepted, its terminal success or failure is promoted to one follow-up message. Foreground results include the completed run ID. Recovery inherits the source launch mode; legacy snapshots without launchMode recover in the background. Set foreground: true or false on workflow_resume/workflow_retry to override it; foreground recovery waits for terminal value and run details, while background recovery returns immediately and delivers completion or failure as a follow-up. After failure follow-ups, especially CANCELLED or interrupted runs, call workflow_status({ runId }) before recovery or replacement work, then pass its state as expectedState to workflow_retry/workflow_resume so recovery cannot act on a state that changed. Recovery map: agent(..., { retries }) reruns one agent call in the same run for transient failures; workflow_retry({ runId, expectedState?, foreground? }) replays a failed run into a child; workflow_resume({ runId, expectedState?, budget?, foreground? }) continues a budget_exhausted run; parentRunId on a new launch only borrows named worktrees and never replays or resumes."
164
120
  export const WORKFLOW_TOOL_PARAMETERS = Type.Object({
165
121
  name: Type.String({ description: "Required non-empty workflow name" }),
@@ -174,6 +130,21 @@ export const WORKFLOW_TOOL_PARAMETERS = Type.Object({
174
130
  }, { additionalProperties: false });
175
131
  export const WORKFLOW_STATUS_PARAMETERS = Type.Object({ runId: Type.String({ description: "Workflow run ID visible in the current project" }) }, { additionalProperties: false });
176
132
  export const WORKFLOW_RETRY_PARAMETERS = Type.Object({ runId: Type.String({ description: "Explicit failed workflow run ID" }), expectedState: Type.Optional(Type.String({ description: "Persisted source state observed before recovery" })), foreground: Type.Optional(Type.Boolean({ description: "Override the source launch mode for this recovery" })) });
133
+ // Structured results codemode scripts receive instead of the JSON text; nested records stay loose so they cannot drift from persistence.
134
+ const WORKFLOW_ERROR_OUTPUT = Type.Object({ code: Type.String(), message: Type.String() });
135
+ const WORKFLOW_LAUNCH_OUTPUT = Type.Object({ runId: Type.String(), state: Type.String(), detached: Type.Optional(Type.Boolean()), value: Type.Optional(Type.Unknown()) });
136
+ const WORKFLOW_STATUS_OUTPUT = Type.Object({ runId: Type.String(), workflowName: Type.String(), state: Type.String(), error: Type.Optional(WORKFLOW_ERROR_OUTPUT), failedAt: Type.Optional(Type.String()), budget: Type.Optional(Type.Unknown()), usage: Type.Optional(Type.Unknown()), phase: Type.Optional(Type.String()), delivery: Type.Optional(Type.Object({ mode: Type.String(), state: Type.String() })), agents: Type.Array(Type.Object({ id: Type.String(), label: Type.Optional(Type.String()), path: Type.String(), state: Type.String(), lastEventAt: Type.Optional(Type.Number()), accounting: Type.Optional(Type.Unknown()) })) });
137
+ const WORKFLOW_STOP_OUTPUT = Type.Object({ runId: Type.String(), state: Type.String(), stopped: Type.Boolean(), reason: Type.Optional(Type.String()) });
138
+ const WORKFLOW_RECOVERY_OUTPUT = Type.Object({ state: Type.String(), runId: Type.Optional(Type.String()), parentRunId: Type.Optional(Type.String()), proposalId: Type.Optional(Type.String()), value: Type.Optional(Type.Unknown()) });
139
+ const WORKFLOW_RESPOND_OUTPUT = Type.Object({ state: Type.String(), approved: Type.Boolean(), reason: Type.String(), accepted: Type.Optional(Type.Boolean()), runId: Type.Optional(Type.String()), value: Type.Optional(Type.Unknown()) });
140
+ const WORKFLOW_CATALOG_ALIAS_OUTPUT = Type.Object({ name: Type.String(), kind: Type.String(), provenance: Type.String(), version: Type.Optional(Type.String()), headline: Type.Optional(Type.String()) });
141
+ // The index without a name; with one, the function, the model alias, or { error }.
142
+ const WORKFLOW_CATALOG_OUTPUT = Type.Union([
143
+ Type.Object({ functions: Type.Array(Type.Object({ name: Type.String(), description: Type.String(), input: Type.Unknown() })), modelAliasEntries: Type.Optional(Type.Array(WORKFLOW_CATALOG_ALIAS_OUTPUT)), settings: Type.Optional(Type.Unknown()) }),
144
+ Type.Object({ name: Type.String(), version: Type.String(), headline: Type.String(), description: Type.String(), input: Type.Unknown(), output: Type.Unknown() }),
145
+ WORKFLOW_CATALOG_ALIAS_OUTPUT,
146
+ Type.Object({ error: Type.Object({ code: Type.String(), name: Type.String(), message: Type.String() }) }),
147
+ ]);
177
148
 
178
149
  function workflowToolUpdate(run: PersistedRun): WorkflowToolUpdate {
179
150
  //NOTE: renderers read details.run; the partial text stays O(1) so high-frequency updates never format the whole tree.
@@ -257,9 +228,46 @@ function modelInventory(root: ModelSpec | undefined, registry: ModelRegistryCapa
257
228
  if (rootName) { knownModels.add(rootName); availableModels.add(rootName); }
258
229
  return { knownModels, availableModels };
259
230
  }
260
- function resumeHostContext(ctx: unknown): { model: { provider: string; id: string } | undefined; modelRegistry: ModelRegistryCapability | undefined; deliveryContext: CompletionDeliveryContext } {
231
+ function resumeHostContext(ctx: unknown): { model: { provider: string; id: string } | undefined; modelRegistry: ModelRegistryCapability | undefined; deliveryContext: CompletionDeliveryContext; toolContext: ExtensionToolContext | undefined } {
261
232
  const model = object(ctx) && object(ctx.model) && typeof ctx.model.provider === "string" && typeof ctx.model.id === "string" ? { provider: ctx.model.provider, id: ctx.model.id } : undefined;
262
- return { model, modelRegistry: contextHostCapabilities(ctx).modelRegistry, deliveryContext: completionContext(ctx) };
233
+ return { model, modelRegistry: contextHostCapabilities(ctx).modelRegistry, deliveryContext: completionContext(ctx), toolContext: scriptToolContext(ctx) };
234
+ }
235
+ // State is rewritten on every update, so large shell inputs and errors are kept as prefixes.
236
+ const SCRIPT_CALL_TEXT_LIMIT_BYTES = 16 * 1024;
237
+ function scriptCallInput(input: JsonValue): Pick<WorkflowScriptCall, "input" | "inputBytes"> {
238
+ const text = JSON.stringify(input);
239
+ const bytes = Buffer.byteLength(text);
240
+ return bytes <= SCRIPT_CALL_TEXT_LIMIT_BYTES ? { input } : { input: utf8Prefix(text, SCRIPT_CALL_TEXT_LIMIT_BYTES), inputBytes: bytes };
241
+ }
242
+ /** Records a script call as started, replacing the record of an attempt that an interruption left unfinished. */
243
+ function startScriptCall(run: PersistedRun, call: WorkflowScriptCall): PersistedRun {
244
+ const recorded = { ...call, ...(run.phase === undefined ? {} : { phase: run.phase }) };
245
+ const calls = run.scriptCalls ?? [];
246
+ return { ...run, scriptCalls: calls.some(({ path }) => path === call.path) ? calls.map((current) => current.path === call.path ? recorded : current) : [...calls, recorded] };
247
+ }
248
+ function finishScriptCall(run: PersistedRun, path: string, error?: WorkflowError): PersistedRun {
249
+ const finishedAt = Date.now();
250
+ return { ...run, scriptCalls: (run.scriptCalls ?? []).map((call) => call.path === path ? { ...call, finishedAt, ...(error ? { error: { code: error.code, ...(call.kind === "shell" ? { message: utf8Prefix(error.message, SCRIPT_CALL_TEXT_LIMIT_BYTES) } : {}) } } : {}) } : call) };
251
+ }
252
+ type AliasVirtualModel = { provider: string; id: string; name: string; thinkingLevels: readonly string[]; route: (request: { thinkingLevel: string }, ctx: ExtensionContext) => { model: Model<Api>; thinkingLevel: string } };
253
+ /** Lists workflow model aliases in `/model` as virtual models; needs a Pi host with `registerVirtualModel`. */
254
+ function registerAliasVirtualModels(pi: unknown, names: readonly string[], settingsPath: string): void {
255
+ const register = object(pi) ? asFn(pi.registerVirtualModel) : undefined;
256
+ if (!register) return;
257
+ for (const name of names) {
258
+ const definition: AliasVirtualModel = {
259
+ provider: VIRTUAL_MODEL_PROVIDER, id: name, name: `${name} (workflow alias)`, thinkingLevels: THINKING_LEVELS,
260
+ route: (request, ctx) => {
261
+ // Settings are read per request so alias edits apply without a reload.
262
+ const aliases = resolveWorkflowSettings(ctx.cwd, projectTrusted(ctx), settingsPath).effective.modelAliases ?? {};
263
+ const target = resolveModelReference(name, aliases, new Set(ctx.modelRegistry.getAll().map((model) => `${model.provider}/${model.id}`)), settingsPath);
264
+ const model = ctx.modelRegistry.find(target.provider, target.model);
265
+ if (!model) throw new WorkflowError("UNKNOWN_MODEL", `Unknown model alias ${name} resolved to ${target.provider}/${target.model}`);
266
+ return { model, thinkingLevel: request.thinkingLevel };
267
+ },
268
+ };
269
+ Reflect.apply(register, pi, [definition]);
270
+ }
263
271
  }
264
272
  async function resolveLaunchAliases(registry: WorkflowRegistryApi, staticAliases: Readonly<Record<string, string>>, context: Readonly<WorkflowModelAliasResolverContext>, availableModels: ReadonlySet<string>, knownModels: ReadonlySet<string>, settingsPath: string): Promise<{ aliases: Readonly<Record<string, string>>; dynamicNames: readonly string[] }> {
265
273
  const dynamic = typeof registry.resolveModelAliases === "function" ? await registry.resolveModelAliases(context, new Set(Object.keys(staticAliases))) : {};
@@ -276,6 +284,20 @@ async function resolveLaunchAliases(registry: WorkflowRegistryApi, staticAliases
276
284
  }
277
285
  }
278
286
 
287
+ /**
288
+ * Prepares each completely static agent configuration in inspection mode, so preparation plugins resolve their own aliases
289
+ * and options without persisting a configuration. Any preparation failure rejects launch before effects; cancellation stops
290
+ * the remaining inspections.
291
+ */
292
+ async function inspectStaticAgentConfigurations(executor: WorkflowAgentExecutor, script: string, workflowName: string, signal: AbortSignal, functions: Readonly<Record<string, WorkflowFunction>>): Promise<void> {
293
+ // NOTE: inspection uses the launch cwd; agents inside withWorktree(...) prepare against their checkout at runtime.
294
+ for (const agentOptions of staticAgentPreparationOptions(script, functions)) {
295
+ if (signal.aborted) throw new WorkflowError("CANCELLED", "Workflow launch cancelled");
296
+ await executor.prepare({ label: "agent", workflowName, agentOptions }, undefined, signal, "inspection");
297
+ }
298
+ if (signal.aborted) throw new WorkflowError("CANCELLED", "Workflow launch cancelled");
299
+ }
300
+
279
301
  export default function workflowExtension(pi: WorkflowExtensionAPI, home?: string, clipboard = copyToClipboard, transport: AgentTransport = localAgentTransport, agentDir?: string, additionalSkillPaths: readonly string[] = []) {
280
302
  beginWorkflowExtensionLoading();
281
303
  const registry = loadingRegistry();
@@ -290,8 +312,14 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
290
312
  return textBlock(data ? `Warning: ${data.message}` : "");
291
313
  });
292
314
  let backgroundWidgetEnabled = true;
293
- try { backgroundWidgetEnabled = loadSettings(workflowSettingsPath(extensionAgentDir)).backgroundWidget ?? true; } catch { /* Keep the optional UI enabled; the launch path reports settings errors. */ }
315
+ try {
316
+ const globalSettings = resolveWorkflowSettings(process.cwd(), false, workflowSettingsPath(extensionAgentDir)).effective;
317
+ backgroundWidgetEnabled = globalSettings.backgroundWidget ?? true;
318
+ // Registered at load, before Pi resolves --model and restores the session model. Project aliases need project trust, which is unknown here.
319
+ registerAliasVirtualModels(pi, Object.keys(globalSettings.modelAliases ?? {}), workflowSettingsPath(extensionAgentDir));
320
+ } catch { /* Keep the optional UI enabled; the launch path reports settings errors. */ }
294
321
  const backgroundWidgetController = backgroundWidget(pi, backgroundWidgetEnabled);
322
+ const codemodeTools = loadCodemodeToolsSetting(extensionAgentDir);
295
323
  const logBridge = (store: RunStore, lifecycle: RunLifecycle, workflowName: string) => async (message: string) => {
296
324
  const timestamp = Date.now();
297
325
  const bounded = utf8Prefix(message, DELIVERY_LIMIT_BYTES);
@@ -314,6 +342,9 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
314
342
  return skillPath ? { skillPaths: [skillPath] } : undefined;
315
343
  });
316
344
  const runs = new Map<string, WorkflowRunRecord>();
345
+ // Extension load failures a headless launcher reported for this host. They hold for the whole process, so every executor,
346
+ // including those that paused resume, cold resume and retry create, keeps failing ownerless options closed.
347
+ let extensionLoad: AgentExecutionRoot["extensionLoad"];
317
348
  const deliveryController = new ForegroundDeliveryController({ runs, ...(typeof pi.sendMessage === "function" ? { deliver: (content: string) => { deliver(pi, content); } } : {}) });
318
349
  let releaseWorkflowRegistry: (() => void) | undefined;
319
350
  const providerRecoveryLane = new SerialLane();
@@ -534,7 +565,11 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
534
565
  if (replayed) return readShellResult(replayed.value);
535
566
  const shellStartedAt = Date.now();
536
567
  let shellPhaseIndex = -1;
537
- const started = await persistRunState(store, metadata, (current) => {
568
+ // Env values stay out of the record: they may carry secrets, and a shared trajectory publishes run state.
569
+ const call: WorkflowScriptCall = { kind: "shell", name: "shell", path, ...scriptCallInput({ command, ...(options.timeoutMs === undefined ? {} : { timeoutMs: options.timeoutMs }), ...(options.env === undefined ? {} : { env: Object.keys(options.env) }), ...(identity.worktreeOwner ? { worktree: identity.worktreeOwner } : {}) }), startedAt: shellStartedAt };
570
+ let failure: WorkflowError | undefined;
571
+ const started = await persistRunState(store, metadata, (run) => {
572
+ const current = startScriptCall(run, call);
538
573
  const history = current.phaseHistory ?? [];
539
574
  if (current.phase !== undefined) shellPhaseIndex = current.phaseHistoryIndex ?? (history.length ? history.length - 1 : 0);
540
575
  const phaseActivities = [...(current.activeShellsByPhase ?? [])];
@@ -552,8 +587,12 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
552
587
  if (!jsonValue(result)) fail("SHELL_FAILED", "Shell result is not JSON-compatible");
553
588
  await store.complete(path, result);
554
589
  return result;
590
+ } catch (error) {
591
+ failure = asWorkflowError(error);
592
+ throw error;
555
593
  } finally {
556
- const stopped = await persistRunState(store, metadata, (current) => {
594
+ const stopped = await persistRunState(store, metadata, (run) => {
595
+ const current = finishScriptCall(run, path, failure);
557
596
  const phaseActivities = [...(current.activeShellsByPhase ?? [])];
558
597
  const phaseActivityIndex = phaseActivities.findIndex(({ phaseIndex }) => phaseIndex === shellPhaseIndex);
559
598
  const phaseActivity = phaseActivities[phaseActivityIndex];
@@ -572,6 +611,44 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
572
611
  }
573
612
  } finally { await lifecycle.leave(); }
574
613
  };
614
+ /** A script `tools.<id>(args)` call: replayed from the journal, otherwise run through the session's tool pipeline and journaled. */
615
+ const toolForRun = async (store: RunStore, lifecycle: RunLifecycle, context: ExtensionToolContext | undefined, identifier: string, args: Readonly<Record<string, JsonValue>>, signal: AbortSignal, identity: ToolIdentity): Promise<JsonValue> => {
616
+ await lifecycle.enter();
617
+ try {
618
+ const path = toolIdentityPath(identifier, args, identity);
619
+ const replayed = await store.replay(path);
620
+ if (replayed) return replayed.value;
621
+ if (!context) fail("UNKNOWN_TOOL", `tools.${identifier} needs a run launched or recovered by a Pi tool call; runs resumed from /workflow or at session start, and headless launches, only replay journaled tool calls`);
622
+ // Rendering is best-effort: it must never lose or mask a completed call.
623
+ const render = async () => { try { const update = runs.get(store.runId)?.update; if (update) update(workflowToolUpdate(withLiveActivities((await store.load()).run))); } catch { /* best-effort progress */ } };
624
+ // Inspection writes must not prevent execution, mask errors, or reject a journaled success.
625
+ const record = async (update: (run: PersistedRun) => PersistedRun) => { try { await store.updateState(update); } catch { /* best-effort call metadata */ } };
626
+ let value: JsonValue;
627
+ try {
628
+ const tool = scriptTool(context, identifier);
629
+ if (!tool) fail("UNKNOWN_TOOL", `tools.${identifier} is not a tool this session can call`);
630
+ const cancelled = (): boolean => signal.aborted;
631
+ if (cancelled()) fail("CANCELLED", "Workflow cancelled");
632
+ await record((run) => startScriptCall(run, { kind: "tool", name: tool.name, path, startedAt: Date.now() }));
633
+ const finished = liveAgents.toolStarted(store.runId, tool.name);
634
+ try {
635
+ await render();
636
+ const outcome = await context.executeTool(tool.name, args, { signal });
637
+ if (cancelled()) fail("CANCELLED", "Workflow cancelled");
638
+ value = scriptToolValue(tool, outcome);
639
+ } finally { finished(); }
640
+ } catch (error) {
641
+ // A stale context throws plain errors from ctx.tools; report them as tool failures.
642
+ const failure = error instanceof WorkflowError ? error : new WorkflowError("TOOL_FAILED", `tools.${identifier}: ${errorText(error)}`);
643
+ await record((run) => finishScriptCall(run, path, failure));
644
+ throw failure;
645
+ }
646
+ await store.complete(path, value);
647
+ await record((run) => finishScriptCall(run, path));
648
+ await render();
649
+ return value;
650
+ } finally { await lifecycle.leave(); }
651
+ };
575
652
  const lifecycleFor = (store: RunStore, state: RunState, metadata: WorkflowMetadata) => new RunLifecycle(state, async (next, previous, reason) => {
576
653
  const run = runs.get(store.runId);
577
654
  if (!run) throw new WorkflowError("INTERNAL_ERROR", `Unknown production run: ${store.runId}`);
@@ -646,7 +723,7 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
646
723
  const persisted = await persistRunState(run.store, run.metadata, (current) => ({ ...current, ...run.budget.snapshot(), agents: current.agents.map((agent) => agent.id === id ? { ...agent, lastEventAt } : agent) }));
647
724
  run.update?.(workflowToolUpdate(withLiveActivities(persisted)));
648
725
  };
649
- const result = await run.executor.execute(prompt, { label: options.label, workflowName: run.metadata.name, tuiIndex, tuiLabel: options.requestedLabel ?? options.label, agentNodeId: id, ...(options.extensionSettings === undefined ? {} : { inheritedExtensionSettings: options.extensionSettings }), onProgress, onAttempt, budget, ...(run.providerErrorRecovery ? { providerErrorRecovery: run.providerErrorRecovery } : {}), ...(parentId ? { parent: parentId, cwd: options.cwd, ...(options.worktreeOwner ? { worktreeOwner: options.worktreeOwner } : {}) } : options.worktreeOwner ? { worktreeOwner: options.worktreeOwner } : {}), ...(options.model ? { model: options.model } : {}), ...(options.role ? { role: options.role } : {}), ...(options.contextFiles ? { contextFiles: options.contextFiles } : {}), tools: options.tools, ...(options.skills ? { skills: options.skills } : {}), ...(options.extensions ? { extensions: options.extensions } : {}), effectiveTools: options.tools, ...(options.schema ? { schema: options.schema } : {}), ...(options.retries === undefined ? {} : { retries: options.retries }), ...(options.timeoutMs === undefined ? {} : { timeoutMs: options.timeoutMs }), ...(options.sessionPath ? { sessionPath: options.sessionPath } : {}), ...(options.agentOptions ? { agentOptions: options.agentOptions } : {}), ...(options.agentIdentity ? { agentIdentity: options.agentIdentity } : {}) }, signal, scheduler.toolsFor(id, (role, tools, model, inheritedTools, skills, extensions) => run.executor.resolve({ label: "child", workflowName: run.metadata.name, ...(model ? { model } : {}), ...(role ? { role } : {}), ...(tools !== undefined ? { tools } : {}), ...(skills !== undefined ? { skills } : {}), ...(extensions !== undefined ? { extensions } : {}) }, inheritedTools).tools), setSteer, () => { scheduler.cancelChildren(id); scheduler.retry(id); });
726
+ const result = await run.executor.execute(prompt, { label: options.label, workflowName: run.metadata.name, tuiIndex, tuiLabel: options.requestedLabel ?? options.label, agentNodeId: id, ...(options.extensionSettings === undefined ? {} : { inheritedExtensionSettings: options.extensionSettings }), onProgress, onAttempt, budget, ...(run.providerErrorRecovery ? { providerErrorRecovery: run.providerErrorRecovery } : {}), ...(parentId ? { parent: parentId, cwd: options.cwd, ...(options.worktreeOwner ? { worktreeOwner: options.worktreeOwner } : {}) } : options.worktreeOwner ? { worktreeOwner: options.worktreeOwner } : {}), ...(options.model ? { model: options.model } : {}), ...(options.contextFiles ? { contextFiles: options.contextFiles } : {}), tools: options.tools, ...(options.skills ? { skills: options.skills } : {}), ...(options.extensions ? { extensions: options.extensions } : {}), effectiveTools: options.tools, ...(options.schema ? { schema: options.schema } : {}), ...(options.retries === undefined ? {} : { retries: options.retries }), ...(options.timeoutMs === undefined ? {} : { timeoutMs: options.timeoutMs }), ...(options.sessionPath ? { sessionPath: options.sessionPath } : {}), ...(options.agentOptions ? { agentOptions: options.agentOptions } : {}), ...(options.agentIdentity ? { agentIdentity: options.agentIdentity } : {}), ...(options.configuration ? { configuration: options.configuration } : {}), ...(options.capabilities ? { capabilities: options.capabilities } : {}), ...(options.projectTrusted === undefined ? {} : { projectTrusted: options.projectTrusted }), ...(options.inheritedExcludeTools ? { inheritedExcludeTools: options.inheritedExcludeTools } : {}), onConfiguration: (configuration) => scheduler.setConfiguration(id, configuration) }, signal, scheduler.toolsFor(id), setSteer, () => { scheduler.cancelChildren(id); scheduler.retry(id); });
650
727
  cancelDeferredProgress();
651
728
  const before = (await run.store.load()).run;
652
729
  await persistAgentAttempts(run.store, id, result.attempts);
@@ -682,17 +759,16 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
682
759
  const existing = new Map(current.agents.map((agent) => [agent.id, agent]));
683
760
  const agents = ownership.map((node) => {
684
761
  const previous = existing.get(node.id);
685
- const requested = { label: node.options.label, workflowName: run.metadata.name, ...(node.options.skills ? { skills: node.options.skills } : {}), ...(node.options.extensions ? { extensions: node.options.extensions } : {}), ...(node.options.model ? { model: node.options.model } : {}), ...(node.options.role ? { role: node.options.role } : {}), ...(node.options.contextFiles ? { contextFiles: node.options.contextFiles } : {}) };
762
+ const requested = { label: node.options.label, workflowName: run.metadata.name, ...(node.options.skills ? { skills: node.options.skills } : {}), ...(node.options.extensions ? { extensions: node.options.extensions } : {}), ...(node.options.model ? { model: node.options.model } : {}), ...(node.options.contextFiles ? { contextFiles: node.options.contextFiles } : {}), ...(node.options.configuration ? { configuration: node.options.configuration } : {}) };
686
763
  let effective: { model: ModelSpec; requestedModel?: string; tools: readonly string[] };
687
764
  try { effective = { ...run.executor.resolve({ ...requested, effectiveTools: node.options.tools }), tools: node.options.tools }; }
688
765
  catch { effective = previous ? { model: previous.model, ...(previous.requestedModel ? { requestedModel: previous.requestedModel } : {}), tools: previous.tools } : { model: node.options.model ? modelSpec(node.options.model, run.model) : run.model, ...(node.options.model ? { requestedModel: node.options.model } : {}), tools: node.options.tools }; }
689
766
  const resultPath = !node.parentId && node.options.agentIdentity ? agentIdentityPath(node.options.agentIdentity) : undefined;
690
- const nodeRole = roleNameOf(node.options.role);
691
767
  const now = Date.now();
692
768
  const lastEventAt = node.state === "running" ? previous?.state === "running" && previous.lastEventAt !== undefined ? previous.lastEventAt : now : previous?.lastEventAt;
693
769
  const startedAt = previous?.startedAt ?? (node.state === "running" ? now : undefined);
694
770
  const durationMs = previous?.durationMs ?? (SETTLED_AGENT_STATES.has(node.state) && startedAt !== undefined ? Math.max(0, now - startedAt) : undefined);
695
- return { ...(previous?.systemPrompt === undefined ? {} : { systemPrompt: previous.systemPrompt }), ...(node.prompt !== undefined ? { prompt: node.prompt } : previous?.prompt !== undefined ? { prompt: previous.prompt } : {}), id: node.id, name: node.label, ...(node.options.requestedLabel ? { label: node.options.requestedLabel } : {}), path: node.id, state: node.state, ...(node.parentId ? { parentId: node.parentId } : {}), structuralPath: [...(node.options.agentIdentity?.structuralPath ?? [])], ...(resultPath ? { resultPath } : {}), ...(node.options.parentBreadcrumb ? { parentBreadcrumb: node.options.parentBreadcrumb } : {}), ...(node.options.worktreeOwner ? { worktreeOwner: node.options.worktreeOwner } : {}), ...(node.options.agentIdentity?.handle === undefined ? {} : { handle: node.options.agentIdentity.handle, ...(node.options.agentIdentity.turn === undefined ? {} : { turn: node.options.agentIdentity.turn }), ...(node.options.continuity ? { continuity: node.options.continuity } : {}) }), ...(nodeRole ? { role: nodeRole } : {}), ...(effective.requestedModel ? { requestedModel: effective.requestedModel } : {}), model: effective.model, tools: effective.tools, attempts: previous?.attempts ?? 0, ...(startedAt === undefined ? {} : { startedAt }), ...(durationMs === undefined ? {} : { durationMs }), ...(previous?.attemptDetails ? { attemptDetails: previous.attemptDetails } : {}), ...(previous?.accounting ? { accounting: previous.accounting } : {}), ...(previous?.toolCalls ? { toolCalls: previous.toolCalls } : {}), ...(previous?.activity ? { activity: previous.activity } : {}), ...(lastEventAt === undefined ? {} : { lastEventAt }) };
771
+ return { ...(previous?.systemPrompt === undefined ? {} : { systemPrompt: previous.systemPrompt }), ...(node.prompt !== undefined ? { prompt: node.prompt } : previous?.prompt !== undefined ? { prompt: previous.prompt } : {}), id: node.id, name: node.label, ...(node.options.requestedLabel ? { label: node.options.requestedLabel } : {}), path: node.id, state: node.state, ...(node.parentId ? { parentId: node.parentId } : {}), structuralPath: [...(node.options.agentIdentity?.structuralPath ?? [])], ...(resultPath ? { resultPath } : {}), ...(node.options.parentBreadcrumb ? { parentBreadcrumb: node.options.parentBreadcrumb } : {}), ...(node.options.worktreeOwner ? { worktreeOwner: node.options.worktreeOwner } : {}), ...(node.options.agentIdentity?.handle === undefined ? {} : { handle: node.options.agentIdentity.handle, ...(node.options.agentIdentity.turn === undefined ? {} : { turn: node.options.agentIdentity.turn }), ...(node.options.continuity ? { continuity: node.options.continuity } : {}) }), ...(effective.requestedModel ? { requestedModel: effective.requestedModel } : {}), model: effective.model, tools: effective.tools, attempts: previous?.attempts ?? 0, ...(startedAt === undefined ? {} : { startedAt }), ...(durationMs === undefined ? {} : { durationMs }), ...(previous?.attemptDetails ? { attemptDetails: previous.attemptDetails } : {}), ...(previous?.accounting ? { accounting: previous.accounting } : {}), ...(previous?.toolCalls ? { toolCalls: previous.toolCalls } : {}), ...(previous?.activity ? { activity: previous.activity } : {}), ...(lastEventAt === undefined ? {} : { lastEventAt }) };
696
772
  });
697
773
  return { ...current, agents };
698
774
  });
@@ -822,17 +898,20 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
822
898
  label: string,
823
899
  description: string,
824
900
  parameters: P,
825
- run: (params: Static<P>, signal: AbortSignal, ctx: unknown) => Promise<{ text: string; details: unknown }>,
901
+ outputSchema: TSchema,
902
+ run: (params: Static<P>, signal: AbortSignal, ctx: unknown) => Promise<{ text: string; details: unknown; structured: unknown }>,
826
903
  ) => {
827
904
  pi.registerTool({
828
905
  name,
829
906
  label,
830
907
  description,
831
908
  parameters,
909
+ outputSchema,
910
+ ...workflowToolExposure(name, codemodeTools),
832
911
  async execute(_id, params, signal: AbortSignal, _onUpdate, ctx) {
833
912
  try {
834
913
  const result = await run(params, signal, ctx);
835
- return { content: [{ type: "text" as const, text: result.text }], details: result.details };
914
+ return { content: [{ type: "text" as const, text: result.text }], details: result.details, structuredContent: result.structured as JsonValue };
836
915
  } catch (error) {
837
916
  throw mainAgentError(error);
838
917
  }
@@ -847,15 +926,18 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
847
926
  "Workflow Respond",
848
927
  "Approve or reject one pending workflow checkpoint or budget decision",
849
928
  Type.Object({ runId: Type.String(), name: Type.Optional(Type.String()), proposalId: Type.Optional(Type.String()), approved: Type.Boolean() }, { additionalProperties: false }),
929
+ WORKFLOW_RESPOND_OUTPUT,
850
930
  async (params, signal, ctx) => {
851
931
  if (params.proposalId) {
852
932
  const result = await recovery.answerBudgetDecision(params.runId, params.proposalId, params.approved, false, ctx, signal);
853
- if (!result) { const denied = { state: "budget_exhausted" as const, approved: false, reason: "proposal_not_pending" }; return { text: JSON.stringify(denied), details: denied }; }
854
- return { text: completionControlContent(result, params.runId), details: { ...result, reason: params.approved ? "approved" : "rejected" } };
933
+ if (!result) { const denied = { state: "budget_exhausted" as const, approved: false, reason: "proposal_not_pending" }; return { text: JSON.stringify(denied), details: denied, structured: denied }; }
934
+ const reason = params.approved ? "approved" : "rejected";
935
+ return { text: completionControlContent(result, params.runId), details: { ...result, reason }, structured: { ...completionControlValue(result, params.runId), reason } };
855
936
  }
856
937
  if (!params.name) throw new WorkflowError("INVALID_METADATA", "workflow_respond requires name or proposalId");
857
938
  const accepted = await answerCheckpoint(params.runId, params.name, params.approved);
858
- return { text: accepted ? "Checkpoint response accepted." : "Checkpoint is not awaiting a response.", details: { accepted, state: accepted ? "checkpoint_answered" : "not_pending", approved: params.approved, reason: "checkpoint" } };
939
+ const details = { accepted, state: accepted ? "checkpoint_answered" : "not_pending", approved: params.approved, reason: "checkpoint" };
940
+ return { text: accepted ? "Checkpoint response accepted." : "Checkpoint is not awaiting a response.", details, structured: details };
859
941
  },
860
942
  );
861
943
  registerControlTool(
@@ -863,9 +945,10 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
863
945
  "Workflow Stop",
864
946
  "Stop an active workflow run by ID",
865
947
  Type.Object({ runId: Type.String() }, { additionalProperties: false }),
948
+ WORKFLOW_STOP_OUTPUT,
866
949
  async (params) => {
867
950
  const result = await stopWorkflowRun(params.runId);
868
- return { text: JSON.stringify(result), details: result };
951
+ return { text: JSON.stringify(result), details: result, structured: result };
869
952
  },
870
953
  );
871
954
  registerControlTool(
@@ -873,9 +956,10 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
873
956
  "Workflow Status",
874
957
  "Read a compact summary of a workflow run in the current project",
875
958
  WORKFLOW_STATUS_PARAMETERS,
959
+ WORKFLOW_STATUS_OUTPUT,
876
960
  async (params, _signal, ctx) => {
877
961
  const result = await workflowStatusRun(params.runId, ctx);
878
- return { text: JSON.stringify(result), details: result };
962
+ return { text: JSON.stringify(result), details: result, structured: result };
879
963
  },
880
964
  );
881
965
  let catalogRegistered = false;
@@ -888,13 +972,15 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
888
972
  if (!catalog.functions.length && !hasAliases && !hasSettings) return;
889
973
  pi.registerTool({
890
974
  name: "workflow_catalog",
975
+ ...workflowToolExposure("workflow_catalog", codemodeTools),
976
+ outputSchema: WORKFLOW_CATALOG_OUTPUT,
891
977
  label: "Workflow Catalog",
892
978
  description: "List reusable workflow functions and model aliases; pass `name` to load one entry in full",
893
979
  parameters: Type.Object({ name: Type.Optional(Type.String({ description: "Registered function or model alias name for full detail" })) }, { additionalProperties: false }),
894
980
  async execute(_id, params = {}) {
895
981
  const context = { cwd, projectTrusted: trustedProject, globalSettingsPath: workflowSettingsPath(extensionAgentDir) };
896
982
  const result = params.name === undefined ? registry.catalogIndex(context) : registry.catalogDetail(params.name, context);
897
- return { content: [{ type: "text" as const, text: JSON.stringify(result) }], details: result };
983
+ return { content: [{ type: "text" as const, text: JSON.stringify(result) }], details: result, structuredContent: result as unknown as JsonValue };
898
984
  },
899
985
  renderCall(args, theme) {
900
986
  const title = theme.fg("toolTitle", theme.bold("workflow_catalog"));
@@ -906,22 +992,9 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
906
992
  });
907
993
  catalogRegistered = true;
908
994
  };
909
- const createAgentExecutor = (root: Omit<AgentExecutionRoot, "agentDir" | "agentSetupHooks">) => new WorkflowAgentExecutor({ ...root, agentDir: extensionAgentDir, ...(additionalSkillPaths.length ? { additionalSkillPaths } : {}), agentSetupHooks: registry.agentSetupHooks(), validateExtensionSettings: registry.validateExtensionSettings, onResourceWarning: (message) => { deliverWarning(pi, message); }, onAgentSettings: (agentId, settings) => { scheduler.setExtensionSettings(agentId, settings); } }, transport);
910
- // Resumed runs keep the role definitions frozen in their snapshot and fall back to the current definitions for roles a
911
- // registered function reaches for the first time after the restart; those are captured into the snapshot like at launch.
912
- const resumeRoles = (store: RunStore, snapshot: Readonly<LaunchSnapshot>, cwd: string, trustedProject: boolean) => {
913
- let current: Readonly<Record<string, AgentDefinition>> = {};
914
- let project: Readonly<Record<string, AgentDefinition>> = {};
915
- try {
916
- current = loadAgentDefinitions(cwd, extensionAgentDir, trustedProject, registry.roleDirectoryRegistrations());
917
- project = trustedProject ? loadProjectAgentDefinitions(cwd) : {};
918
- } catch { /* Unreadable current roles leave the resume with the snapshot roles, as before. */ }
919
- const definitions = { ...current, ...(snapshot.roles ?? {}) };
920
- const captured = roleCapture(store, snapshot, definitions, project);
921
- return { definitions, capture: (role: string, model: ModelSpec) => captured.capture(role, model) };
922
- };
995
+ const createAgentExecutor = (root: Omit<AgentExecutionRoot, "agentDir" | "agentSetupHooks">) => new WorkflowAgentExecutor({ ...root, ...(extensionLoad ? { extensionLoad } : {}), agentDir: extensionAgentDir, ...(additionalSkillPaths.length ? { additionalSkillPaths } : {}), agentPreparationHooks: registry.agentPreparationHooks(), agentSetupHooks: registry.agentSetupHooks(), validateExtensionSettings: registry.validateExtensionSettings, onResourceWarning: root.onResourceWarning ?? ((message) => { deliverWarning(pi, message); }), onAgentSettings: (agentId, settings) => { scheduler.setExtensionSettings(agentId, settings); }, onAgentCapabilities: (agentId, capabilities, projectTrusted) => { scheduler.setAttemptCapabilities(agentId, capabilities, projectTrusted); } }, transport);
923
996
  const activeSnapshotTools = (tools: readonly string[], active: ReadonlySet<string> | "session") => active === "session"
924
- ? new Set(tools.filter((tool) => pi.getActiveTools().includes(tool) && tool !== "workflow_catalog"))
997
+ ? new Set(tools.filter((tool) => reachableTools(pi).includes(tool) && tool !== "workflow_catalog"))
925
998
  : new Set(tools.filter((tool) => active.has(tool) || tool === "workflow_catalog"));
926
999
  const resumeLaunchPrologue = async (input: {
927
1000
  snapshot: Readonly<LaunchSnapshot>;
@@ -931,11 +1004,12 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
931
1004
  modelRegistry?: ModelRegistryCapability | undefined;
932
1005
  signal: AbortSignal;
933
1006
  resolvedAliases?: Readonly<Record<string, string>>;
1007
+ resolvedDynamicAliasNames?: readonly string[];
934
1008
  blockedAliases?: ReadonlySet<string>;
935
1009
  blockedAliasTargets?: Readonly<Record<string, string>>;
936
1010
  withPreflight: boolean;
937
1011
  }) => {
938
- const active = new Set(pi.getActiveTools().filter((tool) => !INTERNAL_WORKFLOW_TOOLS.includes(tool)));
1012
+ const active = new Set(reachableTools(pi).filter((tool) => !INTERNAL_WORKFLOW_TOOLS.includes(tool)));
939
1013
  const missing = input.snapshot.tools.filter((tool) => tool !== "workflow_catalog").find((tool) => !active.has(tool));
940
1014
  if (missing) throw new WorkflowError("RESUME_INCOMPATIBLE", `Required tool is unavailable: ${missing}`);
941
1015
  const settingsPath = workflowSettingsPath(extensionAgentDir);
@@ -943,26 +1017,27 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
943
1017
  const extensionSettings = input.snapshot.settings.extensionSettings;
944
1018
  const extensionSettingsPath = input.snapshot.settingsSources?.extensionSettings ?? input.snapshot.settingsPath ?? settingsPath;
945
1019
  registry.validateExtensionSettings(extensionSettings, { source: "effective", cwd: input.cwd, projectTrusted: input.trustedProject, settingsPath: extensionSettingsPath });
946
- for (const [role, definition] of Object.entries(input.snapshot.roles ?? {})) registry.validateExtensionSettings(mergeWorkflowExtensionSettings(extensionSettings, definition.extensionSettings), { source: "role", cwd: input.cwd, projectTrusted: input.trustedProject, settingsPath: extensionSettingsPath, role });
947
1020
  const currentPolicy = resolveAgentResourcePolicy(input.cwd, input.trustedProject, settingsPath);
948
1021
  const staticAliases = resolution.effective.modelAliases ?? {};
949
1022
  const previousAliases = input.snapshot.modelAliases ?? input.snapshot.settings.modelAliases ?? {};
950
1023
  const inventory = modelInventory(input.rootModel, input.modelRegistry);
951
1024
  const knownModels = input.modelRegistry ? inventory.knownModels : new Set([...input.snapshot.models, ...inventory.knownModels]);
952
1025
  const availableModels = input.modelRegistry ? inventory.availableModels : new Set([...input.snapshot.models, ...inventory.availableModels]);
953
- const currentAliases = input.resolvedAliases ?? (await resolveLaunchAliases(registry, staticAliases, { cwd: input.cwd, projectTrusted: input.trustedProject, rootModel: input.rootModel, knownModels, availableModels, signal: input.signal }, availableModels, knownModels, settingsPath)).aliases;
1026
+ const resolved = input.resolvedAliases ? { aliases: input.resolvedAliases, dynamicNames: input.resolvedDynamicAliasNames ?? [] } : await resolveLaunchAliases(registry, staticAliases, { cwd: input.cwd, projectTrusted: input.trustedProject, rootModel: input.rootModel, knownModels, availableModels, signal: input.signal }, availableModels, knownModels, settingsPath);
1027
+ const currentAliases = resolved.aliases;
1028
+ const dynamicAliasNames = resolved.dynamicNames;
954
1029
  const blockedAliases = input.blockedAliases ?? new Set(Object.keys(previousAliases).filter((name) => !Object.prototype.hasOwnProperty.call(currentAliases, name)));
955
1030
  const blockedAliasTargets = input.blockedAliasTargets ?? Object.fromEntries(Object.entries(previousAliases).filter(([name]) => !Object.prototype.hasOwnProperty.call(currentAliases, name)));
956
1031
  const script = input.withPreflight ? input.snapshot.script : undefined;
957
1032
  if (script !== undefined) {
958
1033
  const resumeAliases = { ...previousAliases, ...currentAliases };
959
- preflight(script, { models: availableModels, tools: active, agentTypes: new Set(input.snapshot.agentTypes), modelAliases: resumeAliases, knownModels, settingsPath, skipModelAvailability: true }, input.snapshot.schemas, input.snapshot.metadata, true);
1034
+ preflight(script, { models: availableModels, tools: active, modelAliases: resumeAliases, knownModels, settingsPath, skipModelAvailability: true }, input.snapshot.schemas, input.snapshot.metadata, true);
960
1035
  }
961
1036
  const refreshed = resumedSnapshotSettings(input.snapshot, resolution, currentAliases);
962
1037
  const snapshot = createLaunchSnapshot({ ...input.snapshot, settingsPath, ...refreshed, modelAliases: currentAliases });
963
- return { active, settingsPath, resolution, currentPolicy, previousAliases, knownModels, availableModels, currentAliases, blockedAliases, blockedAliasTargets, snapshot, script };
1038
+ return { active, settingsPath, resolution, currentPolicy, previousAliases, knownModels, availableModels, currentAliases, dynamicAliasNames, blockedAliases, blockedAliasTargets, snapshot, script };
964
1039
  };
965
- const workflowAgentHandler = (store: RunStore, metadata: WorkflowMetadata, lifecycle: RunLifecycle, executor: WorkflowAgentExecutor, cwd: string, runId: string, captureRole?: (role: string, model: ModelSpec) => Promise<void>) => {
1040
+ const workflowAgentHandler = (store: RunStore, metadata: WorkflowMetadata, lifecycle: RunLifecycle, executor: WorkflowAgentExecutor, cwd: string, runId: string) => {
966
1041
  // Replay lookup and spawn are serialised per run so concurrent agent(...) calls reach the scheduler in call order.
967
1042
  const admission = new SerialLane();
968
1043
  return async (prompt: string, options: Readonly<Record<string, JsonValue>>, agentSignal: AbortSignal, identity: AgentIdentity) => {
@@ -974,20 +1049,21 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
974
1049
  if (replayed) return { replayed };
975
1050
  const worktree = agentWorktree(identity);
976
1051
  const agentCwd = worktree.worktreeOwner ? (await persistWorktree(store, metadata, worktree.worktreeOwner)).cwd : cwd;
977
- const role = typeof options.role === "string" ? options.role : undefined;
978
1052
  const model = typeof options.model === "string" ? options.model : undefined;
979
1053
  const requestedLabel = typeof options.label === "string" ? options.label : undefined;
980
1054
  const skills = Array.isArray(options.skills) ? options.skills as string[] : undefined;
981
1055
  const extensions = Array.isArray(options.extensions) ? options.extensions as string[] : undefined;
982
1056
  const contextFiles = Array.isArray(options.contextFiles) && options.contextFiles.every(isContextFileScope) ? options.contextFiles : undefined;
983
- const resolved = executor.resolve({ label: requestedLabel ?? role ?? "agent", workflowName: metadata.name, ...(model ? { model } : {}), ...(role ? { role } : {}), ...(contextFiles ? { contextFiles } : {}), ...(Array.isArray(options.tools) ? { tools: options.tools as string[] } : {}), ...(skills ? { skills } : {}), ...(extensions ? { extensions } : {}) });
984
- if (role) await captureRole?.(role, resolved.model);
985
- const label = displayAgentName(requestedLabel, role, resolved.model);
1057
+ // A paused resume replaces the run executor with refreshed aliases; new identities must prepare against it.
1058
+ const current = runs.get(runId)?.executor ?? executor;
1059
+ const configuration = await current.prepare({ label: requestedLabel ?? "agent", workflowName: metadata.name, agentOptions: options, agentIdentity: identity, ...(worktree.worktreeOwner ? { worktreeOwner: worktree.worktreeOwner } : {}) }, agentCwd, agentSignal);
1060
+ const resolved = configuration;
1061
+ const label = requestedLabel ?? configuration.label ?? resolved.model.model;
986
1062
  const tools = resolved.tools;
987
1063
  const schema = object(options.outputSchema) ? options.outputSchema : undefined;
988
1064
  const sessionPath = identity.handle !== undefined && identity.turn !== undefined ? await handleTurnInput(store, identity.handle, identity.turn) : undefined;
989
1065
  const continuity = identity.handle === undefined ? undefined : sessionPath ? "continued" as const : "fresh" as const;
990
- return { spawned: scheduler.spawn(runId, prompt, { label, ...(requestedLabel ? { requestedLabel } : {}), ...(identity.parentBreadcrumb ? { parentBreadcrumb: identity.parentBreadcrumb } : {}), cwd: agentCwd, tools, ...(skills ? { skills } : {}), ...(extensions ? { extensions } : {}), ...worktree, ...(model ? { model } : {}), ...(role ? { role } : {}), ...(contextFiles ? { contextFiles } : {}), ...(schema ? { schema } : {}), ...(typeof options.retries === "number" ? { retries: options.retries } : {}), ...(positiveInteger(options.timeoutMs) || options.timeoutMs === null ? { timeoutMs: options.timeoutMs } : {}), ...(sessionPath ? { sessionPath } : {}), ...(continuity ? { continuity } : {}), agentOptions: options, agentIdentity: identity }) };
1066
+ return { spawned: scheduler.spawn(runId, prompt, { label, ...(requestedLabel ? { requestedLabel } : {}), ...(identity.parentBreadcrumb ? { parentBreadcrumb: identity.parentBreadcrumb } : {}), cwd: agentCwd, tools, ...(skills ? { skills } : {}), ...(extensions ? { extensions } : {}), ...worktree, ...(model ? { model } : {}), ...(contextFiles ? { contextFiles } : {}), ...(schema ? { schema } : {}), ...(typeof options.retries === "number" ? { retries: options.retries } : {}), ...(positiveInteger(options.timeoutMs) || options.timeoutMs === null ? { timeoutMs: options.timeoutMs } : {}), ...(sessionPath ? { sessionPath } : {}), ...(continuity ? { continuity } : {}), agentOptions: options, agentIdentity: identity, configuration }) };
991
1067
  });
992
1068
  if ("replayed" in admitted) return admitted.replayed.value;
993
1069
  const { spawned } = admitted;
@@ -1002,7 +1078,7 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
1002
1078
  };
1003
1079
  };
1004
1080
  const recovery = createWorkflowRecovery({
1005
- pi, home, runs, scheduler, eventPublisher, persistRunState, projectTrusted, resumeHostContext, ensureSessionLease, coordinateRunMutation, createAgentExecutor, activeSnapshotTools, frozenResourcePolicy, resolveLaunchPrologue: resumeLaunchPrologue, resumeRoles, workflowAgentHandler, shellForRun, resolveWorktree, checkpointBridge, phaseBridge, logBridge, lifecycleFor, createProviderErrorRecovery, cleanupTerminalRun, deliver: (content) => { deliver(pi, content); }, deliverTerminal: deliveryController.deliverTerminal, workflowToolUpdate, registry, modelSpec,
1081
+ pi, home, runs, scheduler, eventPublisher, persistRunState, projectTrusted, resumeHostContext, ensureSessionLease, coordinateRunMutation, createAgentExecutor, activeSnapshotTools, frozenResourcePolicy, resolveLaunchPrologue: resumeLaunchPrologue, workflowAgentHandler, shellForRun, toolForRun, resolveWorktree, checkpointBridge, phaseBridge, logBridge, lifecycleFor, createProviderErrorRecovery, cleanupTerminalRun, deliver: (content) => { deliver(pi, content); }, deliverTerminal: deliveryController.deliverTerminal, workflowToolUpdate, registry, modelSpec,
1006
1082
  });
1007
1083
  const resumeSelectedWorkflow = async (runId: string, foreground: boolean, context: unknown, budgetPatch?: unknown): Promise<{ workflowName: string; state: "running" | "completed" | "awaiting_approval"; attached: boolean; value?: JsonValue }> => {
1008
1084
  const run = runs.get(runId);
@@ -1021,9 +1097,8 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
1021
1097
  if (!foreground && wasAttached) await deliveryController.moveForegroundToBackground(runId);
1022
1098
  if (foreground && !wasAttached) {
1023
1099
  run.foreground = true;
1024
- const loaded = await run.store.load();
1025
1100
  await persistRunState(run.store, run.metadata, (current) => ({ ...current, delivery: { ...(current.delivery ?? {}), mode: "foreground", state: "attached" } }));
1026
- await run.store.saveSnapshot(createLaunchSnapshot({ ...loaded.snapshot, launchMode: "foreground" }));
1101
+ await run.store.setLaunchMode("foreground");
1027
1102
  } else if (!foreground) run.foreground = false;
1028
1103
  else run.foreground = true;
1029
1104
  await recovery.refreshPausedRunAliases(run, { ...recoveryContext, projectTrusted: projectTrusted(context) });
@@ -1058,9 +1133,10 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
1058
1133
  "Workflow Retry",
1059
1134
  "Retry a failed workflow run by replaying its completed structural operations",
1060
1135
  WORKFLOW_RETRY_PARAMETERS,
1136
+ WORKFLOW_RECOVERY_OUTPUT,
1061
1137
  async (params, signal, ctx) => {
1062
1138
  const result = await recovery.retryWorkflowRun(params.runId, ctx, signal, params.foreground, params.expectedState);
1063
- return { text: completionControlContent(result), details: result };
1139
+ return { text: completionControlContent(result), details: result, structured: completionControlValue(result) };
1064
1140
  },
1065
1141
  );
1066
1142
  registerControlTool(
@@ -1068,9 +1144,10 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
1068
1144
  "Workflow Resume",
1069
1145
  "Resume an exhausted workflow with unchanged or patched aggregate budgets",
1070
1146
  Type.Object({ runId: Type.String(), expectedState: Type.Optional(Type.String({ description: "Persisted source state observed before recovery" })), budget: Type.Optional(Type.Unknown()), foreground: Type.Optional(Type.Boolean({ description: "Override the source launch mode for this recovery" })) }, { additionalProperties: false }),
1147
+ WORKFLOW_RECOVERY_OUTPUT,
1071
1148
  async (params, signal, ctx) => {
1072
1149
  const result = await recovery.resumeWorkflowRun(params.runId, params.budget, ctx, signal, params.foreground, true, params.expectedState);
1073
- return { text: completionControlContent(result), details: result };
1150
+ return { text: completionControlContent(result), details: result, structured: completionControlValue(result) };
1074
1151
  },
1075
1152
  );
1076
1153
  const deliverStaleTerminal = async (store: RunStore, run: PersistedRun): Promise<void> => {
@@ -1127,10 +1204,9 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
1127
1204
  const budgetRuntime = new WorkflowBudgetRuntime(budget, loaded.run.budgetVersion ?? 1, loaded.run.usage, loaded.run.budgetEvents, { active: loaded.run.state === "running" });
1128
1205
  const lifecycle = lifecycleFor(store, loaded.run.state, loaded.snapshot.metadata);
1129
1206
  const providerPause = async () => { deliver(pi, `Workflow ${loaded.snapshot.metadata.name} paused: provider limit.`); await lifecycle.providerPause(); };
1130
- const roleDefinitions = loaded.snapshot.roles ?? {};
1131
1207
  const abortController = new AbortController();
1132
1208
  const providerErrorRecovery = createProviderErrorRecovery(ctx, new Set(loaded.snapshot.models), () => { abortController.abort(); });
1133
- runs.set(runId, { executor: createAgentExecutor({ cwd: ctx.cwd, model, tools: activeSnapshotTools(loaded.snapshot.tools, "session"), resourceSelectors: snapshotResourcePolicy(loaded.snapshot, store.cwd, projectTrusted(ctx), workflowSettingsPath(extensionAgentDir)).effective, extensionSettings: loaded.snapshot.settings.extensionSettings, availableModels: new Set(loaded.snapshot.models), knownModels: new Set(loaded.snapshot.models), ...(loaded.snapshot.modelAliases ?? loaded.snapshot.settings.modelAliases ? { modelAliases: loaded.snapshot.modelAliases ?? loaded.snapshot.settings.modelAliases } : {}), ...(loaded.snapshot.settingsSources?.modelAliases ? { settingsPath: loaded.snapshot.settingsSources.modelAliases } : loaded.snapshot.settingsPath ? { settingsPath: loaded.snapshot.settingsPath } : {}), agentDefinitions: roleDefinitions, runStore: store, providerPause, agentResourcePolicy: frozenResourcePolicy(snapshotResourcePolicy(loaded.snapshot, store.cwd, projectTrusted(ctx), workflowSettingsPath(extensionAgentDir))) }), store, metadata: loaded.snapshot.metadata, model, lifecycle, budget: budgetRuntime, abortController, projectTrusted: () => projectTrusted(ctx), checkpointResolvers: new Map(), ...(providerErrorRecovery ? { providerErrorRecovery } : {}) });
1209
+ runs.set(runId, { executor: createAgentExecutor({ cwd: ctx.cwd, projectTrusted: projectTrusted(ctx), model, tools: activeSnapshotTools(loaded.snapshot.tools, "session"), resourceSelectors: snapshotResourcePolicy(loaded.snapshot, store.cwd, projectTrusted(ctx), workflowSettingsPath(extensionAgentDir)).effective, extensionSettings: loaded.snapshot.settings.extensionSettings, availableModels: new Set(loaded.snapshot.models), knownModels: new Set(loaded.snapshot.models), ...(loaded.snapshot.modelAliases ?? loaded.snapshot.settings.modelAliases ? { modelAliases: loaded.snapshot.modelAliases ?? loaded.snapshot.settings.modelAliases } : {}), ...(loaded.snapshot.settingsSources?.modelAliases ? { settingsPath: loaded.snapshot.settingsSources.modelAliases } : loaded.snapshot.settingsPath ? { settingsPath: loaded.snapshot.settingsPath } : {}), runStore: store, providerPause, agentResourcePolicy: frozenResourcePolicy(snapshotResourcePolicy(loaded.snapshot, store.cwd, projectTrusted(ctx), workflowSettingsPath(extensionAgentDir))) }), store, metadata: loaded.snapshot.metadata, model, lifecycle, budget: budgetRuntime, abortController, projectTrusted: () => projectTrusted(ctx), checkpointResolvers: new Map(), ...(providerErrorRecovery ? { providerErrorRecovery } : {}) });
1134
1210
  for (const checkpoint of await store.awaitingCheckpoints()) deliver(pi, `Workflow ${loaded.snapshot.metadata.name} checkpoint ${checkpoint.name}: ${checkpoint.prompt}\nContext: ${JSON.stringify(checkpoint.context)}\nRespond with workflow_respond.`);
1135
1211
  for (const decision of await store.pendingWorkflowDecisions()) deliver(pi, recovery.budgetDecisionDelivery(loaded.snapshot.metadata, decision));
1136
1212
  scheduler.restoreRun(runId, loaded.snapshot.settings.concurrency, loaded.snapshot.identityVersion === LAUNCH_SNAPSHOT_IDENTITY_VERSION ? await store.loadOwnership() : [], () => runs.get(runId)?.budget.checkAgentLaunch(), loaded.snapshot.settings.extensionSettings);
@@ -1159,19 +1235,14 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
1159
1235
  throw error;
1160
1236
  }
1161
1237
  });
1162
- pi.on("before_agent_start", (event, ctx) => {
1163
- if (!pi.getActiveTools().includes("workflow")) return;
1164
- const roles = Object.entries(loadAgentDefinitions(ctx.cwd, extensionAgentDir, projectTrusted(ctx), typeof registry.roleDirectoryRegistrations === "function" ? registry.roleDirectoryRegistrations() : undefined)).filter(([, definition]) => definition.description);
1165
- if (!roles.length) return;
1166
- const content = `Workflow role descriptions:\n${roles.map(([name, definition]) => `- \`${name}\`: ${String(definition.description)}`).join("\n")}`;
1167
- return { systemPrompt: `${event.systemPrompt}\n\n${content}` };
1168
- });
1169
1238
  const workflowTool: ToolDefinition<typeof WORKFLOW_TOOL_PARAMETERS, WorkflowToolResult, WorkflowProgressRenderState> = {
1170
1239
  name: "workflow",
1171
1240
  label: WORKFLOW_TOOL_LABEL,
1172
1241
  description: WORKFLOW_TOOL_DESCRIPTION,
1173
1242
  promptSnippet: WORKFLOW_TOOL_PROMPT_SNIPPET,
1174
1243
  parameters: WORKFLOW_TOOL_PARAMETERS,
1244
+ outputSchema: WORKFLOW_LAUNCH_OUTPUT,
1245
+ prepareLoadout: prepareScriptToolLoadout,
1175
1246
  async execute(toolCallId, params, signal, onUpdate, ctx) {
1176
1247
  let resolveDetached: ((result: ForegroundDetachResult) => void) | undefined;
1177
1248
  let foregroundStore: RunStore | undefined;
@@ -1179,7 +1250,7 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
1179
1250
  const detachedResult = params.foreground ? new Promise<ForegroundDetachResult>((resolve) => { resolveDetached = resolve; }) : undefined;
1180
1251
  const detachedToolResult = (run: PersistedRun) => {
1181
1252
  const detached = { runId: run.id, state: "running" as const, detached: true as const };
1182
- return { content: [{ type: "text" as const, text: JSON.stringify(detached) }], details: { ...detached, run, preview: `Moved workflow ${run.id} to background.` } };
1253
+ return { content: [{ type: "text" as const, text: JSON.stringify(detached) }], details: { ...detached, run, preview: `Moved workflow ${run.id} to background.` }, structuredContent: detached };
1183
1254
  };
1184
1255
  try {
1185
1256
  const headless = object(ctx) && ctx.headless === true;
@@ -1192,7 +1263,7 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
1192
1263
  const inventory = modelInventory(rootModel, modelRegistry);
1193
1264
  const knownModels = inventory.knownModels;
1194
1265
  const availableModels = inventory.availableModels;
1195
- const rootTools = pi.getActiveTools().filter((name) => !INTERNAL_WORKFLOW_TOOLS.includes(name));
1266
+ const rootTools = reachableTools(pi).filter((name) => !INTERNAL_WORKFLOW_TOOLS.includes(name));
1196
1267
  const trustedProject = projectTrusted(ctx);
1197
1268
  const launchCwd = typeof ctx.cwd === "string" ? ctx.cwd : process.cwd();
1198
1269
  const launch = workflowLaunchSettings(launchCwd, trustedProject, settingsPath, params.concurrency);
@@ -1206,23 +1277,31 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
1206
1277
  if (launch.resolution.global.extensionSettings !== undefined) registry.validateExtensionSettings(launch.resolution.global.extensionSettings, { source: "global", cwd: ctx.cwd, projectTrusted: trustedProject, settingsPath: launch.resolution.globalSettingsPath });
1207
1278
  if (trustedProject && launch.resolution.project.extensionSettings !== undefined) registry.validateExtensionSettings(launch.resolution.project.extensionSettings, { source: "project", cwd: ctx.cwd, projectTrusted: true, settingsPath: launch.resolution.projectSettingsPath });
1208
1279
  const validated = validateWorkflowLaunchWithRegistry(params, { cwd: ctx.cwd, agentDir: extensionAgentDir, projectTrusted: trustedProject, availableModels, rootTools: new Set(rootTools), modelAliases, knownModels, settingsPath, ...(settings.extensionSettings === undefined ? {} : { extensionSettings: settings.extensionSettings }) }, registry);
1209
- const { script, checked, agentDefinitions, projectAgentDefinitions, roleNames } = validated;
1280
+ const { script, checked } = validated;
1281
+ const toolContext = scriptToolContext(ctx);
1282
+ validateScriptToolReferences(script, toolContext);
1283
+ // Headless hosts that tolerated extension load failures report them with the extensions they loaded.
1284
+ const strings = (value: unknown) => Array.isArray(value) ? value.filter((entry): entry is string => typeof entry === "string") : [];
1285
+ const loadErrors = object(ctx) ? strings(ctx.extensionLoadErrors) : [];
1286
+ if (loadErrors.length) extensionLoad = Object.freeze({ errors: loadErrors, loaded: object(ctx) ? strings(ctx.loadedExtensionPaths) : [] });
1287
+ const executorRoot = { cwd: ctx.cwd, projectTrusted: launch.resourcePolicy.projectTrusted, model: rootModel, tools: new Set(rootTools), resourceSelectors: launch.resourcePolicy.effective, extensionSettings: settings.extensionSettings, availableModels, knownModels, modelAliases, dynamicModelAliasNames: resolvedAliases.dynamicNames, settingsPath, agentResourcePolicy: frozenResourcePolicy(launch.resourcePolicy) };
1288
+ // Runtime emits the same resource warnings, so inspection stays silent.
1289
+ await inspectStaticAgentConfigurations(createAgentExecutor({ ...executorRoot, onResourceWarning: () => undefined }), script, checked.metadata.name, runController.signal, registry.functions());
1210
1290
  await ensureSessionLease(ctx.cwd, ctx.sessionManager.getSessionId());
1291
+ // Cancellation can arrive while the lease or a queued parent validation is pending; no run may be created after it.
1292
+ const assertLaunchActive = () => { if (runController.signal.aborted) throw new WorkflowError("CANCELLED", "Workflow launch cancelled"); };
1293
+ assertLaunchActive();
1211
1294
  const runId = randomUUID();
1212
1295
  const args = params.args ?? null;
1213
1296
  encoded(args);
1214
1297
  const runContext = workflowRunContext(ctx.cwd, ctx.sessionManager.getSessionId(), runId, checked.metadata, args, runController.signal);
1215
1298
  const store = new RunStore(ctx.cwd, ctx.sessionManager.getSessionId(), runId, home);
1216
1299
  const parentRunId = params.parentRunId;
1217
- const roles = Object.fromEntries(roleNames.map((role) => [role, agentDefinitions[role]])) as Record<string, AgentDefinition>;
1218
- const projectRoles = roleNames.filter((role) => projectAgentDefinitions[role] !== undefined);
1219
- const roleModels = roleNames.flatMap((role) => { const model = agentDefinitions[role]?.model; return model ? [modelCapability(model, modelAliases, knownModels, settingsPath)] : []; });
1220
- const snapshotModels = [...new Set([rootModelName, ...checked.referenced.models, ...roleModels])];
1221
- const snapshot = createLaunchSnapshot({ script, args, metadata: checked.metadata, launchMode: params.foreground ? "foreground" : "background", settings, settingsPath, settingsSources: { ...launch.resolution.sources, concurrency: params.concurrency === undefined ? launch.resolution.sources.concurrency : "per-run options" }, ...(Object.keys(modelAliases).length ? { modelAliases } : {}), ...(budget ? { budget } : {}), ...(checked.referenced.phases.length ? { phases: checked.referenced.phases } : {}), models: snapshotModels, tools: rootTools, agentTypes: checked.referenced.agentTypes, roles, projectRoles, schemas: checked.schemas });
1222
- const capturedRoles = roleCapture(store, snapshot, agentDefinitions, projectAgentDefinitions);
1300
+ const snapshotModels = [...new Set([rootModelName, ...checked.referenced.models])];
1301
+ const snapshot = createLaunchSnapshot({ script, args, metadata: checked.metadata, launchMode: params.foreground ? "foreground" : "background", settings, settingsPath, settingsSources: { ...launch.resolution.sources, concurrency: params.concurrency === undefined ? launch.resolution.sources.concurrency : "per-run options" }, ...(Object.keys(modelAliases).length ? { modelAliases } : {}), ...(budget ? { budget } : {}), ...(checked.referenced.phases.length ? { phases: checked.referenced.phases } : {}), models: snapshotModels, tools: rootTools, agentConfigurations: {}, schemas: checked.schemas });
1223
1302
  const budgetRuntime = new WorkflowBudgetRuntime(budget);
1224
1303
  const initialBudget = budgetRuntime.snapshot();
1225
- const createRun = () => store.create({ id: runId, workflowName: checked.metadata.name, cwd: ctx.cwd, sessionId: ctx.sessionManager.getSessionId(), state: "running", ...(parentRunId !== undefined ? { parentRunId } : {}), agents: [], agentSessions: [], delivery: params.foreground ? { mode: "foreground", state: "attached", toolCallId } : { mode: "background", state: "pending" }, ...(budget ? { budget } : {}), budgetVersion: 1, ...initialBudget }, snapshot);
1304
+ const createRun = () => { assertLaunchActive(); return store.create({ id: runId, workflowName: checked.metadata.name, cwd: ctx.cwd, sessionId: ctx.sessionManager.getSessionId(), state: "running", ...(parentRunId !== undefined ? { parentRunId } : {}), agents: [], agentSessions: [], delivery: params.foreground ? { mode: "foreground", state: "attached", toolCallId } : { mode: "background", state: "pending" }, ...(budget ? { budget } : {}), budgetVersion: 1, ...initialBudget }, snapshot); };
1226
1305
  // Parent validation and persistence share the mutation lane so a manual deletion cannot remove the parent between the two.
1227
1306
  if (parentRunId !== undefined) await coordinateRunMutation(async () => { await store.validateParentRun(parentRunId); await createRun(); });
1228
1307
  else await createRun();
@@ -1243,7 +1322,7 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
1243
1322
  delivery.detached = true;
1244
1323
  const activeRun = runs.get(runId);
1245
1324
  if (activeRun) { activeRun.foreground = false; delete activeRun.update; }
1246
- await store.saveSnapshot(createLaunchSnapshot({ ...capturedRoles.snapshot, launchMode: "background" }));
1325
+ await store.setLaunchMode("background");
1247
1326
  for (const checkpoint of await store.awaitingCheckpoints()) deliverBackgroundCheckpoint(checked.metadata.name, runId, checkpoint);
1248
1327
  signal?.removeEventListener("abort", onForegroundAbort);
1249
1328
  const run = (await store.load()).run;
@@ -1258,12 +1337,12 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
1258
1337
  const backgroundLaunch = !params.foreground;
1259
1338
  const providerPause = async () => { if (!foregroundAttached) deliver(pi, `Workflow ${checked.metadata.name} paused: provider limit.`); await lifecycle.providerPause(); };
1260
1339
  const providerErrorRecovery = createProviderErrorRecovery(ctx, availableModels, () => { runController.abort(); });
1261
- const executor = createAgentExecutor({ cwd: ctx.cwd, model: rootModel, tools: new Set(rootTools), resourceSelectors: launch.resourcePolicy.effective, extensionSettings: settings.extensionSettings, availableModels, knownModels, modelAliases, settingsPath, agentDefinitions, runStore: store, providerPause, agentResourcePolicy: frozenResourcePolicy(launch.resourcePolicy), runContext });
1340
+ const executor = createAgentExecutor({ ...executorRoot, runStore: store, providerPause, runContext });
1262
1341
  const runRecord: WorkflowRunRecord = { executor, store, metadata: checked.metadata, model: rootModel, lifecycle, budget: budgetRuntime, abortController: runController, foreground: foregroundAttached, projectTrusted: () => projectTrusted(ctx), checkpointResolvers: new Map(), ...(providerErrorRecovery ? { providerErrorRecovery } : {}), ...(params.foreground && onUpdate ? { update: onUpdate } : {}) };
1263
1342
  runs.set(runId, runRecord);
1264
1343
  if (params.foreground && onUpdate) onUpdate(workflowToolUpdate((await store.load()).run));
1265
1344
  scheduler.addRun(runId, settings.concurrency, () => runs.get(runId)?.budget.checkAgentLaunch(), settings.extensionSettings);
1266
- const execution = runWorkflow(script, args, withWorkflowFunctions({ shell: (command, options, signal, identity) => shellForRun(store, checked.metadata, lifecycle, command, options, signal, identity), agent: workflowAgentHandler(store, checked.metadata, lifecycle, executor, ctx.cwd, runId, (role, model) => capturedRoles.capture(role, model)), worktree: async (owner) => resolveWorktree(store, checked.metadata, owner), checkpoint: checkpointBridge(runId, store, checked.metadata, () => runs.get(runId)?.foreground ?? foregroundAttached, ctx.hasUI ? ctx.ui : undefined, headless), phase: phaseBridge(store, checked.metadata, lifecycle), log: logBridge(store, lifecycle, checked.metadata.name) }, store, runContext, registry, settings.extensionSettings), runController.signal);
1345
+ const execution = runWorkflow(script, args, withWorkflowFunctions({ tool: (identifier, toolArgs, signal, identity) => toolForRun(store, lifecycle, toolContext, identifier, toolArgs, signal, identity), shell: (command, options, signal, identity) => shellForRun(store, checked.metadata, lifecycle, command, options, signal, identity), agent: workflowAgentHandler(store, checked.metadata, lifecycle, executor, ctx.cwd, runId), worktree: async (owner) => resolveWorktree(store, checked.metadata, owner), checkpoint: checkpointBridge(runId, store, checked.metadata, () => runs.get(runId)?.foreground ?? foregroundAttached, ctx.hasUI ? ctx.ui : undefined, headless), phase: phaseBridge(store, checked.metadata, lifecycle), log: logBridge(store, lifecycle, checked.metadata.name) }, store, runContext, registry, settings.extensionSettings), runController.signal);
1267
1346
  runRecord.execution = execution;
1268
1347
  await eventPublisher.runStarted(store, checked.metadata);
1269
1348
  const finish = execution.result.then(async (value) => {
@@ -1302,7 +1381,7 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
1302
1381
  }, async (error: unknown) => {
1303
1382
  await deliveryController.deliverTerminal(store, deliverFailureContent(error), true);
1304
1383
  });
1305
- return { content: [{ type: "text" as const, text: JSON.stringify({ runId, state: "running" }) }], details: { runId, preview: `Started workflow ${runId}.` } };
1384
+ return { content: [{ type: "text" as const, text: JSON.stringify({ runId, state: "running" }) }], details: { runId, preview: `Started workflow ${runId}.` }, structuredContent: { runId, state: "running" } };
1306
1385
  }
1307
1386
  void completion.then(async (result) => {
1308
1387
  await deliveryController.deliverDetachedTerminal(toolCallId, completionContent("background", result));
@@ -1327,7 +1406,7 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
1327
1406
  }
1328
1407
  deliveryController.foregroundDeliveries.delete(toolCallId);
1329
1408
  const run = (await store.load()).run;
1330
- return { content: [{ type: "text" as const, text: foregroundDelivery.content }, ...(foregroundDelivery.inlined ? [{ type: "text" as const, text: `Workflow run ID: ${runId}` }] : [])], details: { runId, value, run } };
1409
+ return { content: [{ type: "text" as const, text: foregroundDelivery.content }, ...(foregroundDelivery.inlined ? [{ type: "text" as const, text: `Workflow run ID: ${runId}` }] : [])], details: { runId, value, run }, structuredContent: { runId, state: "completed", value } };
1331
1410
  } catch (error) {
1332
1411
  if (params.foreground && foregroundStore && completionInstalled) {
1333
1412
  const claim = await deliveryController.claimForegroundDelivery(foregroundStore, toolCallId);
@@ -1406,7 +1485,7 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
1406
1485
  return textBlock(isPartial ? "Workflow starting..." : runDetails?.preview ?? (content?.type === "text" ? content.text : "Workflow finished"));
1407
1486
  },
1408
1487
  };
1409
- pi.registerTool(workflowTool);
1488
+ pi.registerTool({ ...workflowTool, ...workflowToolExposure(workflowTool.name, codemodeTools) });
1410
1489
  registerWorkflowNavigator({ pi, home, clipboard, extensionAgentDir, runs, terminalRunStates, hardTerminalRunStates: HARD_TERMINAL_RUN_STATES, ensureSessionLease, coordinateRunMutation, answerCheckpoint, recovery, stopWorkflowRun, moveForegroundToBackground: deliveryController.moveForegroundToBackground, isForegroundAttached: deliveryController.isForegroundAttached, liveAgents, registry, projectTrusted, resumeHostContext, resumeSelectedWorkflow, reportBlocked: reportWorkflowBlocked, trajectoryProvider, setNavigatorOpen: (open) => { if (open) backgroundWidgetController.suspend(); else backgroundWidgetController.resume(); } });
1411
1490
  pi.on("session_shutdown", async () => {
1412
1491
  try {
@@ -1436,6 +1515,7 @@ export default function workflowExtension(pi: WorkflowExtensionAPI, home?: strin
1436
1515
  type LiveAgentState = { session?: WorkflowAgentSession; prepared?: Readonly<PreparedAgentSession>; handoff?: LiveSessionHandoff; activity?: AgentActivity; lastEventAt?: number };
1437
1516
  class LiveAgentRegistry {
1438
1517
  readonly #byRun = new Map<string, Map<string, LiveAgentState>>();
1518
+ readonly #tools = new Map<string, Set<{ name: string; startedAt: number }>>();
1439
1519
  get(runId: string, agentId: string): Readonly<LiveAgentState> | undefined { return this.#byRun.get(runId)?.get(agentId); }
1440
1520
  setSession(runId: string, agentId: string, session?: WorkflowAgentSession): void {
1441
1521
  if (session) this.#state(runId, agentId).session = session;
@@ -1452,17 +1532,26 @@ class LiveAgentRegistry {
1452
1532
  setEventTime(runId: string, agentId: string, timestamp?: number): void {
1453
1533
  if (timestamp !== undefined) this.#state(runId, agentId).lastEventAt = timestamp;
1454
1534
  }
1455
- deleteRun(runId: string): void { this.#byRun.delete(runId); }
1456
- /** Overlays live activity and freshness onto a persisted run; returns the input when nothing is live. */
1535
+ /** Marks a script tool call in flight until the returned callback runs. */
1536
+ toolStarted(runId: string, name: string): () => void {
1537
+ const call = { name, startedAt: Date.now() };
1538
+ const calls = this.#tools.get(runId) ?? new Set<{ name: string; startedAt: number }>();
1539
+ this.#tools.set(runId, calls.add(call));
1540
+ return () => { calls.delete(call); if (!calls.size && this.#tools.get(runId) === calls) this.#tools.delete(runId); };
1541
+ }
1542
+ deleteRun(runId: string): void { this.#byRun.delete(runId); this.#tools.delete(runId); }
1543
+ /** Overlays live activity, freshness, and script tool calls onto a persisted run; returns the input when nothing is live. */
1457
1544
  overlay(run: PersistedRun): PersistedRun {
1545
+ const tools = this.#tools.get(run.id);
1546
+ const withTools = tools?.size ? { ...run, activeTools: [...tools].map((call) => ({ ...call })) } : run;
1458
1547
  const agents = this.#byRun.get(run.id);
1459
- if (!agents?.size) return run;
1548
+ if (!agents?.size) return withTools;
1460
1549
  const next = run.agents.map((agent) => {
1461
1550
  const live = agents.get(agent.id);
1462
1551
  if (!live || (live.activity === undefined && live.lastEventAt === undefined)) return agent;
1463
1552
  return { ...agent, ...(live.activity === undefined ? {} : { activity: live.activity }), ...(live.lastEventAt === undefined ? {} : { lastEventAt: live.lastEventAt }) };
1464
1553
  });
1465
- return next.some((agent, index) => agent !== run.agents[index]) ? { ...run, agents: next } : run;
1554
+ return next.some((agent, index) => agent !== run.agents[index]) ? { ...withTools, agents: next } : withTools;
1466
1555
  }
1467
1556
  #state(runId: string, agentId: string): LiveAgentState {
1468
1557
  const agents = this.#byRun.get(runId) ?? new Map<string, LiveAgentState>();
@@ -1515,9 +1604,7 @@ async function handleTurnInput(store: RunStore, handle: string, turn: number): P
1515
1604
  return target;
1516
1605
  }
1517
1606
 
1518
- function displayAgentName(label: string | undefined, role: string | undefined, model: ModelSpec): string {
1519
- return label ?? role ?? model.model;
1520
- }
1607
+
1521
1608
 
1522
1609
  function modelSpec(value: string, fallback: ModelSpec): ModelSpec {
1523
1610
  try {