@odla-ai/harness 0.10.1 → 0.10.3

package/dist/{chunk-U324RQ4N.js → chunk-CPV7ZVHH.js}

@@ -6,4 +6,4 @@ export {
   HARNESS_PROTOCOL_VERSION,
   DEFAULT_AI_ROUTE
 };
-//# sourceMappingURL=chunk-U324RQ4N.js.map
+//# sourceMappingURL=chunk-CPV7ZVHH.js.map

package/dist/{chunk-U324RQ4N.js.map → chunk-CPV7ZVHH.js.map}

@@ -1 +1 @@
-{"version":3,"sources":["../src/types.ts"],"sourcesContent":["import type { ChatInput, OracleResponse } from \"@odla-ai/ai\";\nimport type { CodeToolPresentation } from \"./code-session-event-types\";\nexport type { CodeToolLocationPreview, CodeToolPresentation } from \"./code-session-event-types\";\n\n/** Current JSONL protocol version exchanged between a runner and an agent container. */\nexport const HARNESS_PROTOCOL_VERSION = 1 as const;\n\n/** Default control-plane route used for model inference requested by coding agents. */\nexport const DEFAULT_AI_ROUTE = \"coding\" as const;\n\n/** Lifecycle state reported for a harness task and its active attempt. */\nexport type HarnessTaskStatus =\n  | \"queued\"\n  | \"running\"\n  | \"cancel_requested\"\n  | \"completed\"\n  | \"failed\"\n  | \"cancelled\";\n\n/** Lifecycle state of one execution attempt for a task. */\nexport type HarnessAttemptStatus = HarnessTaskStatus;\n\n/** Trusted or untrusted participant that emitted a harness event. */\nexport type HarnessActor = \"operator\" | \"runner\" | \"agent\" | \"model\" | \"system\";\n\n/** Resource and isolation limits enforced while an untrusted task executes. */\nexport interface HarnessPolicy {\n  network: \"none\";\n  timeoutMs: number;\n  maxOutputBytes: number;\n  maxPatchBytes: number;\n}\n\n/** Immutable task instructions and execution policy delivered with a lease. */\nexport interface HarnessTaskSpec {\n  taskId: string;\n  attemptId: string;\n  title: string;\n  prompt: string;\n  workspace: string;\n  aiRoute: string;\n  policy: HarnessPolicy;\n  parentAttemptId?: string | null;\n  checkpointSeq?: number | null;\n}\n\n/** Time-bound assignment authorizing a runner to execute one task attempt. */\nexport interface HarnessLease {\n  protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n  leaseId: string;\n  generation: number;\n  expiresAt: number;\n  task: HarnessTaskSpec;\n}\n\n/** Runner-supplied event before the control plane assigns sequence and task metadata. */\nexport interface HarnessEventInput {\n  eventId: string;\n  kind: string;\n  actor: HarnessActor;\n  payload: unknown;\n  createdAt: number;\n}\n\n/** Persisted, ordered event associated with a specific task attempt. */\nexport interface HarnessEvent extends HarnessEventInput {\n  seq: number;\n  taskId: string;\n  attemptId: string;\n}\n\n/** List-view metadata for a task and its current attempt. */\nexport interface HarnessTaskSummary {\n  taskId: string;\n  attemptId: string;\n  appId: string;\n  env: string;\n  title: string;\n  prompt: string;\n  workspace: string;\n  aiRoute: string;\n  status: HarnessTaskStatus;\n  parentAttemptId: string | null;\n  checkpointSeq: number | null;\n  runnerId: string | null;\n  generation: number;\n  createdAt: number;\n  updatedAt: number;\n}\n\n/** List-view metadata for one attempt, including retry ancestry and runner ownership. */\nexport interface HarnessAttemptSummary {\n  attemptId: string;\n  taskId: string;\n  parentAttemptId: string | null;\n  checkpointSeq: number | null;\n  status: HarnessAttemptStatus;\n  runnerId: string | null;\n  generation: number;\n  createdAt: number;\n  updatedAt: number;\n}\n\n/** Complete task view including attempts, events, result, and generated patch. */\nexport interface HarnessTaskDetail extends HarnessTaskSummary {\n  attempts: HarnessAttemptSummary[];\n  events: HarnessEvent[];\n  patch: string | null;\n  result: unknown;\n}\n\n/** Public control-plane view of a registered harness runner. */\nexport interface HarnessRunnerView {\n  runnerId: string;\n  appId: string;\n  env: string;\n  name: string;\n  createdAt: number;\n  lastSeenAt: number | null;\n  revokedAt: number | null;\n}\n\n/** Credential-free normalized model request sent from an agent through the runner. */\nexport interface HarnessInferenceRequest {\n  requestId: string;\n  /** Exact initial, follow-up, or resume command whose budget this call consumes. */\n  interactionId?: string;\n  call: ChatInput;\n}\n\n/** Normalized model response plus auditable provider, policy, and token metadata. */\nexport interface HarnessInferenceResponse {\n  requestId: string;\n  response: OracleResponse;\n  receipt: {\n    provider: string;\n    model: string;\n    policyVersion: number;\n    inputTokens: number;\n    outputTokens: number;\n    /**\n     * USD charged for this call, priced against the model the control plane\n     * actually resolved.\n     *\n     * ABSENT when the live catalog has no price for that model — never zero.\n     * An unpriced call is unknown spend, and reporting it as free is what let\n     * a goal's `maxUsd` look enforced while nothing enforced it. The runtime\n     * cannot compute this itself: it asks for `brokered` and only the control\n     * plane knows which model answered.\n     */\n    costUsd?: number;\n  };\n}\n\n/** Bounded, content-minimized session activity emitted by a Code runtime.\n * Message bodies are projected separately into the app's owner-private\n * odla-db chat. `interactionId` is optional so stored v1 events remain valid. */\nexport type CodeSessionEventData = (\n  | { type: \"message\"; actor: \"agent\" | \"system\"; body: string }\n  | { type: \"diagnostic\"; level: \"error\"; message: string }\n  | { type: \"thinking\"; available: true; durationMs: number }\n  | {\n      type: \"tool\";\n      phase: \"started\";\n      tool: HarnessToolName;\n      operationId?: string;\n      presentation?: CodeToolPresentation;\n    }\n  | {\n      type: \"tool\";\n      phase: \"completed\";\n      tool: HarnessToolName;\n      ok: boolean;\n      durationMs: number;\n      operationId?: string;\n      /** Why the call failed, bounded. Absent when `ok`. Without this a watcher\n       * saw that a tool failed and never why, which is what made an 84%\n       * apply_patch failure rate impossible to diagnose (PM bug 515655ec). */\n      failureReason?: string;\n      presentation?: CodeToolPresentation;\n    }\n  | {\n      type: \"collaboration\";\n      phase: \"started\";\n      skill: string;\n      tool: string;\n      operationId: string;\n    }\n  | {\n      type: \"collaboration\";\n      phase: \"completed\";\n      skill: string;\n      tool: string;\n      ok: boolean;\n      durationMs: number;\n      operationId: string;\n    }\n  | {\n      type: \"usage\"; provider: string; model: string;\n      inputTokens: number; outputTokens: number; durationMs: number;\n      interactionTokens?: number; interactionMaxTokens?: number;\n      /** USD for this call; absent when the model is unpriced, never zero. */\n      costUsd?: number;\n      /** Cumulative USD for this owner interaction, when every call in it was\n       *  priced. Absent the moment one was not, so a partial total can never be\n       *  mistaken for the whole. */\n      interactionCostUsd?: number;\n    }\n  | {\n      type: \"status\"; status: \"running\" | \"idle\" | \"failed\" | \"checkpointed\";\n      durationMs?: number;\n    }\n) & { interactionId?: string };\n\n/** Registry-assigned cursor and timestamp for an owner-visible Code event. */\nexport type CodeSessionEvent = CodeSessionEventData & {\n  eventId: string; sequence: number; createdAt: number;\n};\n\n/** Closed set of effects an agent container may request from its trusted broker. */\nexport type HarnessToolName =\n  | \"sandbox.read\"\n  | \"sandbox.list\"\n  | \"sandbox.search\"\n  | \"sandbox.overview\"\n  | \"sandbox.where_is\"\n  | \"sandbox.who_imports\"\n  | \"sandbox.who_touches\"\n  | \"sandbox.apply_patch\"\n  | \"sandbox.run_recipe\";\n/** Correlated, structured tool request emitted by an untrusted agent container. */\nexport interface HarnessToolRequest {\n  requestId: string;\n  tool: HarnessToolName;\n  input: Record<string, unknown>;\n}\n/** Bounded tool result returned to an agent after trusted policy evaluation. */\nexport interface HarnessToolResponse {\n  requestId: string;\n  ok: boolean;\n  content: string;\n  details?: Record<string, unknown>;\n}\n\n/** Validated JSONL message emitted by an untrusted agent container. */\nexport type HarnessAgentOutput =\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"event\";\n      kind: string;\n      payload?: unknown;\n    }\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"inference.request\";\n      requestId: string;\n      call: ChatInput;\n    }\n  | ({ protocolVersion: typeof HARNESS_PROTOCOL_VERSION; type: \"tool.request\" } & HarnessToolRequest)\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"attempt.complete\";\n      status: \"completed\" | \"failed\" | \"cancelled\";\n      result?: unknown;\n    };\n\n/** JSONL command or inference result written by the trusted runner to an agent. */\nexport type HarnessAgentInput =\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"task.start\";\n      task: HarnessTaskSpec;\n    }\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"inference.response\";\n      requestId: string;\n      response: OracleResponse;\n    }\n  | ({ protocolVersion: typeof HARNESS_PROTOCOL_VERSION; type: \"tool.response\" } & HarnessToolResponse)\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"attempt.cancel\";\n      reason: string;\n    };\n\n/** Terminal attempt report submitted by a runner to the control plane. */\nexport interface HarnessCompletion {\n  status: \"completed\" | \"failed\" | \"cancelled\";\n  result?: unknown;\n  patch?: string;\n  error?: string;\n}\n\n/** Operations a credentialed runner may perform against the harness control plane. */\nexport interface HarnessControlPlane {\n  lease(workspaces: string[]): Promise<HarnessLease | null>;\n  heartbeat(attemptId: string, leaseId: string): Promise<{ cancelRequested: boolean; expiresAt: number }>;\n  appendEvents(attemptId: string, leaseId: string, events: HarnessEventInput[]): Promise<void>;\n  infer(attemptId: string, leaseId: string, request: HarnessInferenceRequest): Promise<HarnessInferenceResponse>;\n  complete(attemptId: string, leaseId: string, completion: HarnessCompletion): Promise<void>;\n}\n\n/** Trusted inference bridge used to keep model credentials outside agent containers. */\nexport interface HarnessAiConnection {\n  infer(request: HarnessInferenceRequest): Promise<HarnessInferenceResponse>;\n}\n\n/** Trusted tool boundary. Implementations must evaluate CaMeL policy before effects. */\nexport interface HarnessToolBroker {\n  execute(\n    context: { lease: HarnessLease; workspaceDir: string; signal?: AbortSignal },\n    request: HarnessToolRequest,\n  ): Promise<HarnessToolResponse>;\n}\n"],"mappings":";AAKO,IAAM,2BAA2B;AAGjC,IAAM,mBAAmB;","names":[]}
+{"version":3,"sources":["../src/types.ts"],"sourcesContent":["import type { ChatInput, OracleResponse } from \"@odla-ai/ai\";\nimport type { CodeToolPresentation } from \"./code-session-event-types\";\nexport type { CodeToolLocationPreview, CodeToolPresentation } from \"./code-session-event-types\";\n\n/** Current JSONL protocol version exchanged between a runner and an agent container. */\nexport const HARNESS_PROTOCOL_VERSION = 1 as const;\n\n/** Default control-plane route used for model inference requested by coding agents. */\nexport const DEFAULT_AI_ROUTE = \"coding\" as const;\n\n/** Lifecycle state reported for a harness task and its active attempt. */\nexport type HarnessTaskStatus =\n  | \"queued\"\n  | \"running\"\n  | \"cancel_requested\"\n  | \"completed\"\n  | \"failed\"\n  | \"cancelled\";\n\n/** Lifecycle state of one execution attempt for a task. */\nexport type HarnessAttemptStatus = HarnessTaskStatus;\n\n/** Trusted or untrusted participant that emitted a harness event. */\nexport type HarnessActor = \"operator\" | \"runner\" | \"agent\" | \"model\" | \"system\";\n\n/** Resource and isolation limits enforced while an untrusted task executes. */\nexport interface HarnessPolicy {\n  network: \"none\";\n  timeoutMs: number;\n  maxOutputBytes: number;\n  maxPatchBytes: number;\n}\n\n/** Immutable task instructions and execution policy delivered with a lease. */\nexport interface HarnessTaskSpec {\n  taskId: string;\n  attemptId: string;\n  title: string;\n  prompt: string;\n  workspace: string;\n  aiRoute: string;\n  policy: HarnessPolicy;\n  parentAttemptId?: string | null;\n  checkpointSeq?: number | null;\n}\n\n/** Time-bound assignment authorizing a runner to execute one task attempt. */\nexport interface HarnessLease {\n  protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n  leaseId: string;\n  generation: number;\n  expiresAt: number;\n  task: HarnessTaskSpec;\n}\n\n/** Runner-supplied event before the control plane assigns sequence and task metadata. */\nexport interface HarnessEventInput {\n  eventId: string;\n  kind: string;\n  actor: HarnessActor;\n  payload: unknown;\n  createdAt: number;\n}\n\n/** Persisted, ordered event associated with a specific task attempt. */\nexport interface HarnessEvent extends HarnessEventInput {\n  seq: number;\n  taskId: string;\n  attemptId: string;\n}\n\n/** List-view metadata for a task and its current attempt. */\nexport interface HarnessTaskSummary {\n  taskId: string;\n  attemptId: string;\n  appId: string;\n  env: string;\n  title: string;\n  prompt: string;\n  workspace: string;\n  aiRoute: string;\n  status: HarnessTaskStatus;\n  parentAttemptId: string | null;\n  checkpointSeq: number | null;\n  runnerId: string | null;\n  generation: number;\n  createdAt: number;\n  updatedAt: number;\n}\n\n/** List-view metadata for one attempt, including retry ancestry and runner ownership. */\nexport interface HarnessAttemptSummary {\n  attemptId: string;\n  taskId: string;\n  parentAttemptId: string | null;\n  checkpointSeq: number | null;\n  status: HarnessAttemptStatus;\n  runnerId: string | null;\n  generation: number;\n  createdAt: number;\n  updatedAt: number;\n}\n\n/** Complete task view including attempts, events, result, and generated patch. */\nexport interface HarnessTaskDetail extends HarnessTaskSummary {\n  attempts: HarnessAttemptSummary[];\n  events: HarnessEvent[];\n  patch: string | null;\n  result: unknown;\n}\n\n/** Public control-plane view of a registered harness runner. */\nexport interface HarnessRunnerView {\n  runnerId: string;\n  appId: string;\n  env: string;\n  name: string;\n  createdAt: number;\n  lastSeenAt: number | null;\n  revokedAt: number | null;\n}\n\n/** Credential-free normalized model request sent from an agent through the runner. */\nexport interface HarnessInferenceRequest {\n  requestId: string;\n  /** Exact initial, follow-up, or resume command whose budget this call consumes. */\n  interactionId?: string;\n  call: ChatInput;\n}\n\n/** Normalized model response plus auditable provider, policy, and token metadata. */\nexport interface HarnessInferenceResponse {\n  requestId: string;\n  response: OracleResponse;\n  receipt: {\n    provider: string;\n    model: string;\n    policyVersion: number;\n    inputTokens: number;\n    outputTokens: number;\n    /**\n     * USD charged for this call, priced against the model the control plane\n     * actually resolved.\n     *\n     * ABSENT when the live catalog has no price for that model — never zero.\n     * An unpriced call is unknown spend, and reporting it as free is what let\n     * a goal's `maxUsd` look enforced while nothing enforced it. The runtime\n     * cannot compute this itself: it asks for `brokered` and only the control\n     * plane knows which model answered.\n     */\n    costUsd?: number;\n  };\n}\n\n/** Bounded, content-minimized session activity emitted by a Code runtime.\n * Message bodies are projected separately into the app's owner-private\n * odla-db chat. `interactionId` is optional so stored v1 events remain valid. */\nexport type CodeSessionEventData = (\n  | { type: \"message\"; actor: \"agent\" | \"system\"; body: string }\n  | { type: \"diagnostic\"; level: \"error\"; message: string }\n  | { type: \"thinking\"; available: true; durationMs: number }\n  | {\n      type: \"tool\";\n      phase: \"started\";\n      tool: HarnessToolName;\n      operationId?: string;\n      presentation?: CodeToolPresentation;\n    }\n  | {\n      type: \"tool\";\n      phase: \"completed\";\n      tool: HarnessToolName;\n      ok: boolean;\n      durationMs: number;\n      operationId?: string;\n      /** Why the call failed, bounded. Absent when `ok`. Without this a watcher\n       * saw that a tool failed and never why, which is what made an 84%\n       * apply_patch failure rate impossible to diagnose (PM bug 515655ec). */\n      failureReason?: string;\n      presentation?: CodeToolPresentation;\n    }\n  | {\n      type: \"collaboration\";\n      phase: \"started\";\n      skill: string;\n      tool: string;\n      operationId: string;\n    }\n  | {\n      type: \"collaboration\";\n      phase: \"completed\";\n      skill: string;\n      tool: string;\n      ok: boolean;\n      durationMs: number;\n      operationId: string;\n    }\n  | {\n      type: \"usage\"; provider: string; model: string;\n      inputTokens: number; outputTokens: number; durationMs: number;\n      interactionTokens?: number; interactionMaxTokens?: number;\n      /** USD for this call; absent when the model is unpriced, never zero. */\n      costUsd?: number;\n      /** Cumulative USD for this owner interaction, when every call in it was\n       *  priced. Absent the moment one was not, so a partial total can never be\n       *  mistaken for the whole. */\n      interactionCostUsd?: number;\n    }\n  | {\n      type: \"status\"; status: \"running\" | \"idle\" | \"failed\" | \"checkpointed\";\n      durationMs?: number;\n    }\n) & { interactionId?: string };\n\n/** Registry-assigned cursor and timestamp for an owner-visible Code event. */\nexport type CodeSessionEvent = CodeSessionEventData & {\n  eventId: string; sequence: number; createdAt: number;\n};\n\n/** Closed set of effects an agent container may request from its trusted broker. */\nexport type HarnessToolName =\n  | \"sandbox.read\"\n  | \"sandbox.list\"\n  | \"sandbox.search\"\n  | \"sandbox.overview\"\n  | \"sandbox.where_is\"\n  | \"sandbox.who_imports\"\n  | \"sandbox.who_touches\"\n  | \"sandbox.apply_patch\"\n  | \"sandbox.edit\"\n  | \"sandbox.run_recipe\";\n/** Correlated, structured tool request emitted by an untrusted agent container. */\nexport interface HarnessToolRequest {\n  requestId: string;\n  tool: HarnessToolName;\n  input: Record<string, unknown>;\n}\n/** Bounded tool result returned to an agent after trusted policy evaluation. */\nexport interface HarnessToolResponse {\n  requestId: string;\n  ok: boolean;\n  content: string;\n  details?: Record<string, unknown>;\n}\n\n/** Validated JSONL message emitted by an untrusted agent container. */\nexport type HarnessAgentOutput =\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"event\";\n      kind: string;\n      payload?: unknown;\n    }\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"inference.request\";\n      requestId: string;\n      call: ChatInput;\n    }\n  | ({ protocolVersion: typeof HARNESS_PROTOCOL_VERSION; type: \"tool.request\" } & HarnessToolRequest)\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"attempt.complete\";\n      status: \"completed\" | \"failed\" | \"cancelled\";\n      result?: unknown;\n    };\n\n/** JSONL command or inference result written by the trusted runner to an agent. */\nexport type HarnessAgentInput =\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"task.start\";\n      task: HarnessTaskSpec;\n    }\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"inference.response\";\n      requestId: string;\n      response: OracleResponse;\n    }\n  | ({ protocolVersion: typeof HARNESS_PROTOCOL_VERSION; type: \"tool.response\" } & HarnessToolResponse)\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"attempt.cancel\";\n      reason: string;\n    };\n\n/** Terminal attempt report submitted by a runner to the control plane. */\nexport interface HarnessCompletion {\n  status: \"completed\" | \"failed\" | \"cancelled\";\n  result?: unknown;\n  patch?: string;\n  error?: string;\n}\n\n/** Operations a credentialed runner may perform against the harness control plane. */\nexport interface HarnessControlPlane {\n  lease(workspaces: string[]): Promise<HarnessLease | null>;\n  heartbeat(attemptId: string, leaseId: string): Promise<{ cancelRequested: boolean; expiresAt: number }>;\n  appendEvents(attemptId: string, leaseId: string, events: HarnessEventInput[]): Promise<void>;\n  infer(attemptId: string, leaseId: string, request: HarnessInferenceRequest): Promise<HarnessInferenceResponse>;\n  complete(attemptId: string, leaseId: string, completion: HarnessCompletion): Promise<void>;\n}\n\n/** Trusted inference bridge used to keep model credentials outside agent containers. */\nexport interface HarnessAiConnection {\n  infer(request: HarnessInferenceRequest): Promise<HarnessInferenceResponse>;\n}\n\n/** Trusted tool boundary. Implementations must evaluate CaMeL policy before effects. */\nexport interface HarnessToolBroker {\n  execute(\n    context: { lease: HarnessLease; workspaceDir: string; signal?: AbortSignal },\n    request: HarnessToolRequest,\n  ): Promise<HarnessToolResponse>;\n}\n"],"mappings":";AAKO,IAAM,2BAA2B;AAGjC,IAAM,mBAAmB;","names":[]}

package/dist/{chunk-OZBJNTML.js → chunk-ESLFS3FY.js}

@@ -1,13 +1,13 @@
 import {
   runContainerAttempt,
   stageWorkspace
-} from "./chunk-5LRYJKUI.js";
+} from "./chunk-FUDTSAOZ.js";
 import {
   makeHarnessEvent
-} from "./chunk-VDY5V7ZG.js";
+} from "./chunk-JV45JAAW.js";
 import {
   HARNESS_PROTOCOL_VERSION
-} from "./chunk-U324RQ4N.js";
+} from "./chunk-CPV7ZVHH.js";
 
 // src/runner.ts
 import { setTimeout as delay } from "timers/promises";
@@ -165,4 +165,4 @@ export {
   runLeasedAttempt,
   runHarnessRunner
 };
-//# sourceMappingURL=chunk-OZBJNTML.js.map
+//# sourceMappingURL=chunk-ESLFS3FY.js.map

package/dist/{chunk-5LRYJKUI.js → chunk-FUDTSAOZ.js}

@@ -1,10 +1,10 @@
 import {
   encodeAgentInput,
   parseAgentOutput
-} from "./chunk-VDY5V7ZG.js";
+} from "./chunk-JV45JAAW.js";
 import {
   HARNESS_PROTOCOL_VERSION
-} from "./chunk-U324RQ4N.js";
+} from "./chunk-CPV7ZVHH.js";
 
 // src/container.ts
 import { execFile, spawn } from "child_process";
@@ -548,4 +548,4 @@ export {
   stageWorkspacePair,
   safeWorkspaceLabel
 };
-//# sourceMappingURL=chunk-5LRYJKUI.js.map
+//# sourceMappingURL=chunk-FUDTSAOZ.js.map

package/dist/{chunk-VDY5V7ZG.js → chunk-JV45JAAW.js}

@@ -1,6 +1,6 @@
 import {
   HARNESS_PROTOCOL_VERSION
-} from "./chunk-U324RQ4N.js";
+} from "./chunk-CPV7ZVHH.js";
 
 // src/protocol.ts
 var CONTROL = /[\u0000-\u0008\u000b\u000c\u000e-\u001f\u007f]/;
@@ -51,7 +51,7 @@ function parseAgentOutput(line) {
   if (message.type === "tool.request") {
     const input = record(message.input);
     const tool = String(message.tool);
-    if (!input || !["sandbox.read", "sandbox.apply_patch", "sandbox.run_recipe"].includes(tool)) {
+    if (!input || !["sandbox.read", "sandbox.apply_patch", "sandbox.edit", "sandbox.run_recipe"].includes(tool)) {
       throw new HarnessProtocolError("tool.request requires a registered tool and object input");
     }
     return {
@@ -90,4 +90,4 @@ export {
   encodeAgentInput,
   makeHarnessEvent
 };
-//# sourceMappingURL=chunk-VDY5V7ZG.js.map
+//# sourceMappingURL=chunk-JV45JAAW.js.map

package/dist/chunk-JV45JAAW.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/protocol.ts"],"sourcesContent":["import type { ChatInput } from \"@odla-ai/ai\";\nimport {\n  HARNESS_PROTOCOL_VERSION,\n  type HarnessAgentInput,\n  type HarnessAgentOutput,\n  type HarnessEventInput,\n} from \"./types\";\n\nconst CONTROL = /[\\u0000-\\u0008\\u000b\\u000c\\u000e-\\u001f\\u007f]/;\n\n/** Error raised when an agent emits malformed, oversized, or unsupported protocol data. */\nexport class HarnessProtocolError extends Error {\n  override readonly name = \"HarnessProtocolError\";\n}\n\nfunction record(value: unknown): Record<string, unknown> | null {\n  return value !== null && typeof value === \"object\" && !Array.isArray(value)\n    ? value as Record<string, unknown>\n    : null;\n}\n\nfunction boundedText(value: unknown, label: string, max: number): string {\n  if (typeof value !== \"string\" || !value || value.length > max || CONTROL.test(value)) {\n    throw new HarnessProtocolError(`${label} must be a non-empty string of at most ${max} characters`);\n  }\n  return value;\n}\n\n/** Parse and validate one newline-delimited message emitted by an agent container. */\nexport function parseAgentOutput(line: string): HarnessAgentOutput {\n  if (Buffer.byteLength(line, \"utf8\") > 1_000_000) throw new HarnessProtocolError(\"agent message exceeds 1 MB\");\n  let value: unknown;\n  try { value = JSON.parse(line); } catch { throw new HarnessProtocolError(\"agent emitted invalid JSON\"); }\n  const message = record(value);\n  if (!message || message.protocolVersion !== HARNESS_PROTOCOL_VERSION) {\n    throw new HarnessProtocolError(`agent protocolVersion must be ${HARNESS_PROTOCOL_VERSION}`);\n  }\n  if (message.type === \"event\") {\n    return {\n      protocolVersion: HARNESS_PROTOCOL_VERSION,\n      type: \"event\",\n      kind: boundedText(message.kind, \"event.kind\", 120),\n      ...(message.payload === undefined ? {} : { payload: message.payload }),\n    };\n  }\n  if (message.type === \"inference.request\") {\n    const call = record(message.call);\n    if (!call || !Array.isArray(call.messages) || !Number.isSafeInteger(call.maxTokens)) {\n      throw new HarnessProtocolError(\"inference.request.call requires messages and maxTokens\");\n    }\n    return {\n      protocolVersion: HARNESS_PROTOCOL_VERSION,\n      type: \"inference.request\",\n      requestId: boundedText(message.requestId, \"requestId\", 180),\n      call: call as unknown as ChatInput,\n    };\n  }\n  if (message.type === \"tool.request\") {\n    const input = record(message.input);\n    const tool = String(message.tool);\n    if (!input || ![\"sandbox.read\", \"sandbox.apply_patch\", \"sandbox.edit\", \"sandbox.run_recipe\"].includes(tool)) {\n      throw new HarnessProtocolError(\"tool.request requires a registered tool and object input\");\n    }\n    return {\n      protocolVersion: HARNESS_PROTOCOL_VERSION,\n      type: \"tool.request\",\n      requestId: boundedText(message.requestId, \"requestId\", 180),\n      tool: tool as \"sandbox.read\" | \"sandbox.apply_patch\" | \"sandbox.edit\" | \"sandbox.run_recipe\",\n      input,\n    };\n  }\n  if (message.type === \"attempt.complete\") {\n    if (!new Set([\"completed\", \"failed\", \"cancelled\"]).has(String(message.status))) {\n      throw new HarnessProtocolError(\"attempt.complete.status is invalid\");\n    }\n    return {\n      protocolVersion: HARNESS_PROTOCOL_VERSION,\n      type: \"attempt.complete\",\n      status: message.status as \"completed\" | \"failed\" | \"cancelled\",\n      ...(message.result === undefined ? {} : { result: message.result }),\n    };\n  }\n  throw new HarnessProtocolError(\"agent message type is unsupported\");\n}\n\n/** Serialize one trusted runner message as a newline-terminated JSONL record. */\nexport function encodeAgentInput(message: HarnessAgentInput): string {\n  return `${JSON.stringify(message)}\\n`;\n}\n\n/** Create a timestamped, uniquely identified event for control-plane submission. */\nexport function makeHarnessEvent(\n  kind: string,\n  actor: HarnessEventInput[\"actor\"],\n  payload: unknown,\n  now = Date.now(),\n  id = crypto.randomUUID(),\n): HarnessEventInput {\n  boundedText(kind, \"event.kind\", 120);\n  return { eventId: id, kind, actor, payload, createdAt: now };\n}\n"],"mappings":";;;;;AAQA,IAAM,UAAU;AAGT,IAAM,uBAAN,cAAmC,MAAM;AAAA,EAC5B,OAAO;AAC3B;AAEA,SAAS,OAAO,OAAgD;AAC9D,SAAO,UAAU,QAAQ,OAAO,UAAU,YAAY,CAAC,MAAM,QAAQ,KAAK,IACtE,QACA;AACN;AAEA,SAAS,YAAY,OAAgB,OAAe,KAAqB;AACvE,MAAI,OAAO,UAAU,YAAY,CAAC,SAAS,MAAM,SAAS,OAAO,QAAQ,KAAK,KAAK,GAAG;AACpF,UAAM,IAAI,qBAAqB,GAAG,KAAK,0CAA0C,GAAG,aAAa;AAAA,EACnG;AACA,SAAO;AACT;AAGO,SAAS,iBAAiB,MAAkC;AACjE,MAAI,OAAO,WAAW,MAAM,MAAM,IAAI,IAAW,OAAM,IAAI,qBAAqB,4BAA4B;AAC5G,MAAI;AACJ,MAAI;AAAE,YAAQ,KAAK,MAAM,IAAI;AAAA,EAAG,QAAQ;AAAE,UAAM,IAAI,qBAAqB,4BAA4B;AAAA,EAAG;AACxG,QAAM,UAAU,OAAO,KAAK;AAC5B,MAAI,CAAC,WAAW,QAAQ,oBAAoB,0BAA0B;AACpE,UAAM,IAAI,qBAAqB,iCAAiC,wBAAwB,EAAE;AAAA,EAC5F;AACA,MAAI,QAAQ,SAAS,SAAS;AAC5B,WAAO;AAAA,MACL,iBAAiB;AAAA,MACjB,MAAM;AAAA,MACN,MAAM,YAAY,QAAQ,MAAM,cAAc,GAAG;AAAA,MACjD,GAAI,QAAQ,YAAY,SAAY,CAAC,IAAI,EAAE,SAAS,QAAQ,QAAQ;AAAA,IACtE;AAAA,EACF;AACA,MAAI,QAAQ,SAAS,qBAAqB;AACxC,UAAM,OAAO,OAAO,QAAQ,IAAI;AAChC,QAAI,CAAC,QAAQ,CAAC,MAAM,QAAQ,KAAK,QAAQ,KAAK,CAAC,OAAO,cAAc,KAAK,SAAS,GAAG;AACnF,YAAM,IAAI,qBAAqB,wDAAwD;AAAA,IACzF;AACA,WAAO;AAAA,MACL,iBAAiB;AAAA,MACjB,MAAM;AAAA,MACN,WAAW,YAAY,QAAQ,WAAW,aAAa,GAAG;AAAA,MAC1D;AAAA,IACF;AAAA,EACF;AACA,MAAI,QAAQ,SAAS,gBAAgB;AACnC,UAAM,QAAQ,OAAO,QAAQ,KAAK;AAClC,UAAM,OAAO,OAAO,QAAQ,IAAI;AAChC,QAAI,CAAC,SAAS,CAAC,CAAC,gBAAgB,uBAAuB,gBAAgB,oBAAoB,EAAE,SAAS,IAAI,GAAG;AAC3G,YAAM,IAAI,qBAAqB,0DAA0D;AAAA,IAC3F;AACA,WAAO;AAAA,MACL,iBAAiB;AAAA,MACjB,MAAM;AAAA,MACN,WAAW,YAAY,QAAQ,WAAW,aAAa,GAAG;AAAA,MAC1D;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,MAAI,QAAQ,SAAS,oBAAoB;AACvC,QAAI,EAAC,oBAAI,IAAI,CAAC,aAAa,UAAU,WAAW,CAAC,GAAE,IAAI,OAAO,QAAQ,MAAM,CAAC,GAAG;AAC9E,YAAM,IAAI,qBAAqB,oCAAoC;AAAA,IACrE;AACA,WAAO;AAAA,MACL,iBAAiB;AAAA,MACjB,MAAM;AAAA,MACN,QAAQ,QAAQ;AAAA,MAChB,GAAI,QAAQ,WAAW,SAAY,CAAC,IAAI,EAAE,QAAQ,QAAQ,OAAO;AAAA,IACnE;AAAA,EACF;AACA,QAAM,IAAI,qBAAqB,mCAAmC;AACpE;AAGO,SAAS,iBAAiB,SAAoC;AACnE,SAAO,GAAG,KAAK,UAAU,OAAO,CAAC;AAAA;AACnC;AAGO,SAAS,iBACd,MACA,OACA,SACA,MAAM,KAAK,IAAI,GACf,KAAK,OAAO,WAAW,GACJ;AACnB,cAAY,MAAM,cAAc,GAAG;AACnC,SAAO,EAAE,SAAS,IAAI,MAAM,OAAO,SAAS,WAAW,IAAI;AAC7D;","names":[]}

package/dist/{chunk-FAN2R3GW.js → chunk-KXFI3WM2.js}

@@ -68,6 +68,10 @@ function codeToolRequestPresentation(request) {
   if (request.tool === "sandbox.apply_patch") {
     return { kind: "patch", ...patchStats(input.patch) };
   }
+  if (request.tool === "sandbox.edit") {
+    const lines = (value) => typeof value === "string" ? value.slice(0, 262144).split("\n").length : 0;
+    return { kind: "patch", additions: lines(input.newText), deletions: lines(input.oldText) };
+  }
   const recipeId = text(input.recipeId, 120);
   return recipeId ? { kind: "recipe", recipeId } : void 0;
 }
@@ -163,4 +167,4 @@ export {
   toolFailureReason,
   observedBroker
 };
-//# sourceMappingURL=chunk-FAN2R3GW.js.map
+//# sourceMappingURL=chunk-KXFI3WM2.js.map

package/dist/chunk-KXFI3WM2.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/code-tool-presentation.ts","../src/code-runtime-observer.ts"],"sourcesContent":["import type {\n  CodeToolLocationPreview, CodeToolPresentation, HarnessToolRequest, HarnessToolResponse,\n} from \"./types\";\n\nconst text = (value: unknown, maximum: number): string | undefined => {\n  if (typeof value !== \"string\") return undefined;\n  const bounded = value.replace(/[\\u0000-\\u0008\\u000b\\u000c\\u000e-\\u001f\\u007f]/g, \" \").trim();\n  return bounded ? bounded.slice(0, maximum) : undefined;\n};\nconst integer = (value: unknown): number | undefined =>\n  Number.isSafeInteger(value) && Number(value) >= 0 ? Number(value) : undefined;\nconst record = (value: unknown): Record<string, unknown> | undefined =>\n  value && typeof value === \"object\" && !Array.isArray(value)\n    ? value as Record<string, unknown> : undefined;\nconst excerpt = (value: unknown, tail = false): string | undefined => {\n  if (typeof value !== \"string\") return undefined;\n  const safe = value.replace(/[\\u0000-\\u0008\\u000b\\u000c\\u000e-\\u001f\\u007f]/g, \" \");\n  const source = (tail ? safe.slice(-10_000) : safe.slice(0, 10_000)).trim();\n  if (!source) return undefined;\n  const lines = source.split(\"\\n\").filter((line) => line.trim()).map((line) => line.slice(0, 240));\n  const selected = tail ? lines.slice(-10) : lines.slice(0, 10);\n  return text(selected.join(\"\\n\"), 2_400);\n};\nconst paths = (value: unknown): string[] | undefined => {\n  if (!Array.isArray(value)) return undefined;\n  const items = value.flatMap((item) => {\n    const path = text(item, 1_024);\n    return path ? [path] : [];\n  }).slice(0, 12);\n  return items.length ? items : undefined;\n};\n\nfunction patchStats(value: unknown): { additions?: number; deletions?: number } {\n  if (typeof value !== \"string\") return {};\n  let additions = 0;\n  let deletions = 0;\n  for (const line of value.slice(0, 262_144).split(\"\\n\")) {\n    if (line.startsWith(\"+++\") || line.startsWith(\"---\")) continue;\n    if (line.startsWith(\"+\")) additions += 1;\n    else if (line.startsWith(\"-\")) deletions += 1;\n  }\n  return { ...(additions ? { additions } : {}), ...(deletions ? { deletions } : {}) };\n}\n\nfunction searchResults(value: unknown): CodeToolLocationPreview[] | undefined {\n  if (typeof value !== \"string\") return undefined;\n  const results = value.split(\"\\n\").flatMap((line) => {\n    const match = /^([^:\\n]{1,1024}):(\\d+):\\s?(.*)$/.exec(line);\n    if (!match) return [];\n    const lineNumber = Number(match[2]);\n    const itemText = text(match[3], 240);\n    if (!Number.isSafeInteger(lineNumber) || lineNumber < 1) return [];\n    return [{ path: match[1]!, line: lineNumber, ...(itemText ? { text: itemText } : {}) }];\n  }).slice(0, 5);\n  return results.length ? results : undefined;\n}\n\n/** Produce the bounded presentation available as soon as a broker call starts. */\nexport function codeToolRequestPresentation(request: HarnessToolRequest): CodeToolPresentation | undefined {\n  const input = request.input;\n  if (request.tool === \"sandbox.read\") {\n    const path = text(input.path, 1_024);\n    if (!path) return undefined;\n    const startLine = integer(input.startLine);\n    const endLine = integer(input.endLine);\n    return { kind: \"read\", path, ...(startLine ? { startLine } : {}), ...(endLine ? { endLine } : {}) };\n  }\n  if (request.tool === \"sandbox.list\") {\n    const scope = text(input.prefix, 1_024);\n    return { kind: \"list\", ...(scope ? { scope } : {}) };\n  }\n  if (request.tool === \"sandbox.search\" || request.tool === \"sandbox.overview\"\n    || request.tool === \"sandbox.where_is\" || request.tool === \"sandbox.who_imports\"\n    || request.tool === \"sandbox.who_touches\") {\n    const query = text(input.query, 512);\n    const scope = request.tool === \"sandbox.search\" ? text(input.prefix, 1_024) : undefined;\n    if (request.tool === \"sandbox.search\" && !query) return undefined;\n    return { kind: \"query\", ...(query ? { query } : {}), ...(scope ? { scope } : {}) };\n  }\n  if (request.tool === \"sandbox.apply_patch\") {\n    return { kind: \"patch\", ...patchStats(input.patch) };\n  }\n  if (request.tool === \"sandbox.edit\") {\n    const lines = (value: unknown) => (typeof value === \"string\" ? value.slice(0, 262_144).split(\"\\n\").length : 0);\n    return { kind: \"patch\", additions: lines(input.newText), deletions: lines(input.oldText) };\n  }\n  const recipeId = text(input.recipeId, 120);\n  return recipeId ? { kind: \"recipe\", recipeId } : undefined;\n}\n\n/** Enrich a started presentation with only the broker's bounded result shape. */\nexport function codeToolResultPresentation(\n  request: HarnessToolRequest,\n  response: HarnessToolResponse,\n): CodeToolPresentation | undefined {\n  const started = codeToolRequestPresentation(request);\n  if (!started || !response.ok) return started;\n  const details = record(response.details);\n  if (started.kind === \"read\") {\n    return {\n      ...started,\n      ...(integer(details?.startLine) ? { startLine: integer(details?.startLine) } : {}),\n      ...(integer(details?.endLine) ? { endLine: integer(details?.endLine) } : {}),\n      ...(excerpt(response.content) ? { excerpt: excerpt(response.content) } : {}),\n    };\n  }\n  if (started.kind === \"list\") {\n    const listed = response.content.split(\"\\n\")\n      .filter((line) => line && !line.startsWith(\"…\") && !line.startsWith(\"Workspace \"))\n      .map((line) => text(line, 1_024)).filter((line): line is string => Boolean(line)).slice(0, 8);\n    return {\n      ...started,\n      ...(integer(details?.count) !== undefined ? { count: integer(details?.count) } : {}),\n      ...(listed.length ? { paths: listed } : {}),\n    };\n  }\n  if (started.kind === \"query\") {\n    const results = request.tool === \"sandbox.search\" ? searchResults(response.content) : undefined;\n    const resultExcerpt = request.tool === \"sandbox.search\" ? undefined : excerpt(response.content);\n    return {\n      ...started,\n      ...(integer(details?.count) !== undefined ? { count: integer(details?.count) } : {}),\n      ...(results ? { results } : {}),\n      ...(resultExcerpt ? { excerpt: resultExcerpt } : {}),\n    };\n  }\n  if (started.kind === \"patch\") {\n    return { ...started, ...(paths(details?.paths) ? { paths: paths(details?.paths) } : {}) };\n  }\n  const output = response.content.replace(/^Recipe [^\\n]*\\.?\\s*/u, \"\");\n  return {\n    ...started,\n    ...(integer(details?.exitCode) !== undefined ? { exitCode: integer(details?.exitCode) } : {}),\n    ...(typeof details?.timedOut === \"boolean\" ? { timedOut: details.timedOut } : {}),\n    ...(typeof details?.outputLimitExceeded === \"boolean\"\n      ? { outputLimitExceeded: details.outputLimitExceeded } : {}),\n    ...(excerpt(output, true) ? { excerpt: excerpt(output, true) } : {}),\n  };\n}\n","// Reporting a brokered tool effect onto the session event stream.\n//\n// Split out of code-runtime-engine.ts, which was one line under the 250 LOC cap\n// and so could not absorb the failure reason below (PM bug c775485c: the same\n// two lines left PR #684 red and unlandable). The seam is real rather than\n// convenient — this is the only code in the engine that translates a broker\n// call into what a watcher sees, and it is the only code that needs to know how\n// a failure is described.\n\nimport { codeToolRequestPresentation, codeToolResultPresentation } from \"./code-tool-presentation\";\nimport type { CodeSessionEventData, HarnessToolBroker, HarnessToolResponse } from \"./types\";\n\n/** Longest failure reason forwarded onto a session event. A reason is a short\n * diagnostic for a human or an agent reading the stream, never a transcript. */\nexport const MAX_FAILURE_REASON = 240;\n\n/** The bounded reason a failed tool response carries, if it carries one.\n *\n * Only `response.ok` used to reach the stream, so a watcher saw that a call\n * failed and could never see why — which is what made an 84% apply_patch\n * failure rate undiagnosable (PM bugs 515655ec and 38d92f1d). */\nexport function toolFailureReason(response: HarnessToolResponse): string | undefined {\n  if (response.ok) return undefined;\n  // Guaranteed here rather than only where the response is built, because this\n  // is the boundary that emits the event: a failure reaching the stream without\n  // a reason is the defect, whoever constructed the response.\n  const supplied = response.details?.failureReason;\n  const reason = (typeof supplied === \"string\" && supplied)\n    || DEFAULT_FAILURE_REASON[response.content]\n    || response.content\n    || \"tool request failed; inspect the tool input and workspace state\";\n  return reason.slice(0, MAX_FAILURE_REASON);\n}\n\n/** Reasons for the two refusals the broker states as bare content. */\nconst DEFAULT_FAILURE_REASON: Record<string, string> = {\n  \"tool denied by CaMeL policy\": \"tool denied by CaMeL policy\",\n  \"review sessions are read-only\": \"review session is read-only; workspace changes are not permitted\",\n};\n\n/** Wrap `broker` so every effect reports its start and its finish.\n *\n * `emit` is the engine's own event sink; it already swallows its own errors, so\n * reporting can never fail the effect it is reporting on.\n */\nexport function observedBroker(input: {\n  broker: HarnessToolBroker;\n  operationIdFor: (requestId: string) => string;\n  emit: (event: CodeSessionEventData) => Promise<void>;\n  now?: () => number;\n}): HarnessToolBroker {\n  const now = input.now ?? Date.now;\n  return {\n    execute: async (context, request) => {\n      const startedAt = now();\n      // Stable and opaque: an older container may choose any bounded request\n      // id, while the owner stream admits only a closed printable id shape.\n      const operationId = input.operationIdFor(request.requestId);\n      const startedPresentation = codeToolRequestPresentation(request);\n      await input.emit({\n        type: \"tool\", phase: \"started\", tool: request.tool, operationId,\n        ...(startedPresentation ? { presentation: startedPresentation } : {}),\n      });\n      const response = await input.broker.execute(context, request);\n      const completedPresentation = codeToolResultPresentation(request, response);\n      const failureReason = toolFailureReason(response);\n      await input.emit({\n        type: \"tool\", phase: \"completed\", tool: request.tool,\n        ok: response.ok, durationMs: now() - startedAt, operationId,\n        ...(failureReason ? { failureReason } : {}),\n        ...(completedPresentation ? { presentation: completedPresentation } : {}),\n      });\n      return response;\n    },\n  };\n}\n"],"mappings":";AAIA,IAAM,OAAO,CAAC,OAAgB,YAAwC;AACpE,MAAI,OAAO,UAAU,SAAU,QAAO;AACtC,QAAM,UAAU,MAAM,QAAQ,mDAAmD,GAAG,EAAE,KAAK;AAC3F,SAAO,UAAU,QAAQ,MAAM,GAAG,OAAO,IAAI;AAC/C;AACA,IAAM,UAAU,CAAC,UACf,OAAO,cAAc,KAAK,KAAK,OAAO,KAAK,KAAK,IAAI,OAAO,KAAK,IAAI;AACtE,IAAM,SAAS,CAAC,UACd,SAAS,OAAO,UAAU,YAAY,CAAC,MAAM,QAAQ,KAAK,IACtD,QAAmC;AACzC,IAAM,UAAU,CAAC,OAAgB,OAAO,UAA8B;AACpE,MAAI,OAAO,UAAU,SAAU,QAAO;AACtC,QAAM,OAAO,MAAM,QAAQ,mDAAmD,GAAG;AACjF,QAAM,UAAU,OAAO,KAAK,MAAM,IAAO,IAAI,KAAK,MAAM,GAAG,GAAM,GAAG,KAAK;AACzE,MAAI,CAAC,OAAQ,QAAO;AACpB,QAAM,QAAQ,OAAO,MAAM,IAAI,EAAE,OAAO,CAAC,SAAS,KAAK,KAAK,CAAC,EAAE,IAAI,CAAC,SAAS,KAAK,MAAM,GAAG,GAAG,CAAC;AAC/F,QAAM,WAAW,OAAO,MAAM,MAAM,GAAG,IAAI,MAAM,MAAM,GAAG,EAAE;AAC5D,SAAO,KAAK,SAAS,KAAK,IAAI,GAAG,IAAK;AACxC;AACA,IAAM,QAAQ,CAAC,UAAyC;AACtD,MAAI,CAAC,MAAM,QAAQ,KAAK,EAAG,QAAO;AAClC,QAAM,QAAQ,MAAM,QAAQ,CAAC,SAAS;AACpC,UAAM,OAAO,KAAK,MAAM,IAAK;AAC7B,WAAO,OAAO,CAAC,IAAI,IAAI,CAAC;AAAA,EAC1B,CAAC,EAAE,MAAM,GAAG,EAAE;AACd,SAAO,MAAM,SAAS,QAAQ;AAChC;AAEA,SAAS,WAAW,OAA4D;AAC9E,MAAI,OAAO,UAAU,SAAU,QAAO,CAAC;AACvC,MAAI,YAAY;AAChB,MAAI,YAAY;AAChB,aAAW,QAAQ,MAAM,MAAM,GAAG,MAAO,EAAE,MAAM,IAAI,GAAG;AACtD,QAAI,KAAK,WAAW,KAAK,KAAK,KAAK,WAAW,KAAK,EAAG;AACtD,QAAI,KAAK,WAAW,GAAG,EAAG,cAAa;AAAA,aAC9B,KAAK,WAAW,GAAG,EAAG,cAAa;AAAA,EAC9C;AACA,SAAO,EAAE,GAAI,YAAY,EAAE,UAAU,IAAI,CAAC,GAAI,GAAI,YAAY,EAAE,UAAU,IAAI,CAAC,EAAG;AACpF;AAEA,SAAS,cAAc,OAAuD;AAC5E,MAAI,OAAO,UAAU,SAAU,QAAO;AACtC,QAAM,UAAU,MAAM,MAAM,IAAI,EAAE,QAAQ,CAAC,SAAS;AAClD,UAAM,QAAQ,mCAAmC,KAAK,IAAI;AAC1D,QAAI,CAAC,MAAO,QAAO,CAAC;AACpB,UAAM,aAAa,OAAO,MAAM,CAAC,CAAC;AAClC,UAAM,WAAW,KAAK,MAAM,CAAC,GAAG,GAAG;AACnC,QAAI,CAAC,OAAO,cAAc,UAAU,KAAK,aAAa,EAAG,QAAO,CAAC;AACjE,WAAO,CAAC,EAAE,MAAM,MAAM,CAAC,GAAI,MAAM,YAAY,GAAI,WAAW,EAAE,MAAM,SAAS,IAAI,CAAC,EAAG,CAAC;AAAA,EACxF,CAAC,EAAE,MAAM,GAAG,CAAC;AACb,SAAO,QAAQ,SAAS,UAAU;AACpC;AAGO,SAAS,4BAA4B,SAA+D;AACzG,QAAM,QAAQ,QAAQ;AACtB,MAAI,QAAQ,SAAS,gBAAgB;AACnC,UAAM,OAAO,KAAK,MAAM,MAAM,IAAK;AACnC,QAAI,CAAC,KAAM,QAAO;AAClB,UAAM,YAAY,QAAQ,MAAM,SAAS;AACzC,UAAM,UAAU,QAAQ,MAAM,OAAO;AACrC,WAAO,EAAE,MAAM,QAAQ,MAAM,GAAI,YAAY,EAAE,UAAU,IAAI,CAAC,GAAI,GAAI,UAAU,EAAE,QAAQ,IAAI,CAAC,EAAG;AAAA,EACpG;AACA,MAAI,QAAQ,SAAS,gBAAgB;AACnC,UAAM,QAAQ,KAAK,MAAM,QAAQ,IAAK;AACtC,WAAO,EAAE,MAAM,QAAQ,GAAI,QAAQ,EAAE,MAAM,IAAI,CAAC,EAAG;AAAA,EACrD;AACA,MAAI,QAAQ,SAAS,oBAAoB,QAAQ,SAAS,sBACrD,QAAQ,SAAS,sBAAsB,QAAQ,SAAS,yBACxD,QAAQ,SAAS,uBAAuB;AAC3C,UAAM,QAAQ,KAAK,MAAM,OAAO,GAAG;AACnC,UAAM,QAAQ,QAAQ,SAAS,mBAAmB,KAAK,MAAM,QAAQ,IAAK,IAAI;AAC9E,QAAI,QAAQ,SAAS,oBAAoB,CAAC,MAAO,QAAO;AACxD,WAAO,EAAE,MAAM,SAAS,GAAI,QAAQ,EAAE,MAAM,IAAI,CAAC,GAAI,GAAI,QAAQ,EAAE,MAAM,IAAI,CAAC,EAAG;AAAA,EACnF;AACA,MAAI,QAAQ,SAAS,uBAAuB;AAC1C,WAAO,EAAE,MAAM,SAAS,GAAG,WAAW,MAAM,KAAK,EAAE;AAAA,EACrD;AACA,MAAI,QAAQ,SAAS,gBAAgB;AACnC,UAAM,QAAQ,CAAC,UAAoB,OAAO,UAAU,WAAW,MAAM,MAAM,GAAG,MAAO,EAAE,MAAM,IAAI,EAAE,SAAS;AAC5G,WAAO,EAAE,MAAM,SAAS,WAAW,MAAM,MAAM,OAAO,GAAG,WAAW,MAAM,MAAM,OAAO,EAAE;AAAA,EAC3F;AACA,QAAM,WAAW,KAAK,MAAM,UAAU,GAAG;AACzC,SAAO,WAAW,EAAE,MAAM,UAAU,SAAS,IAAI;AACnD;AAGO,SAAS,2BACd,SACA,UACkC;AAClC,QAAM,UAAU,4BAA4B,OAAO;AACnD,MAAI,CAAC,WAAW,CAAC,SAAS,GAAI,QAAO;AACrC,QAAM,UAAU,OAAO,SAAS,OAAO;AACvC,MAAI,QAAQ,SAAS,QAAQ;AAC3B,WAAO;AAAA,MACL,GAAG;AAAA,MACH,GAAI,QAAQ,SAAS,SAAS,IAAI,EAAE,WAAW,QAAQ,SAAS,SAAS,EAAE,IAAI,CAAC;AAAA,MAChF,GAAI,QAAQ,SAAS,OAAO,IAAI,EAAE,SAAS,QAAQ,SAAS,OAAO,EAAE,IAAI,CAAC;AAAA,MAC1E,GAAI,QAAQ,SAAS,OAAO,IAAI,EAAE,SAAS,QAAQ,SAAS,OAAO,EAAE,IAAI,CAAC;AAAA,IAC5E;AAAA,EACF;AACA,MAAI,QAAQ,SAAS,QAAQ;AAC3B,UAAM,SAAS,SAAS,QAAQ,MAAM,IAAI,EACvC,OAAO,CAAC,SAAS,QAAQ,CAAC,KAAK,WAAW,QAAG,KAAK,CAAC,KAAK,WAAW,YAAY,CAAC,EAChF,IAAI,CAAC,SAAS,KAAK,MAAM,IAAK,CAAC,EAAE,OAAO,CAAC,SAAyB,QAAQ,IAAI,CAAC,EAAE,MAAM,GAAG,CAAC;AAC9F,WAAO;AAAA,MACL,GAAG;AAAA,MACH,GAAI,QAAQ,SAAS,KAAK,MAAM,SAAY,EAAE,OAAO,QAAQ,SAAS,KAAK,EAAE,IAAI,CAAC;AAAA,MAClF,GAAI,OAAO,SAAS,EAAE,OAAO,OAAO,IAAI,CAAC;AAAA,IAC3C;AAAA,EACF;AACA,MAAI,QAAQ,SAAS,SAAS;AAC5B,UAAM,UAAU,QAAQ,SAAS,mBAAmB,cAAc,SAAS,OAAO,IAAI;AACtF,UAAM,gBAAgB,QAAQ,SAAS,mBAAmB,SAAY,QAAQ,SAAS,OAAO;AAC9F,WAAO;AAAA,MACL,GAAG;AAAA,MACH,GAAI,QAAQ,SAAS,KAAK,MAAM,SAAY,EAAE,OAAO,QAAQ,SAAS,KAAK,EAAE,IAAI,CAAC;AAAA,MAClF,GAAI,UAAU,EAAE,QAAQ,IAAI,CAAC;AAAA,MAC7B,GAAI,gBAAgB,EAAE,SAAS,cAAc,IAAI,CAAC;AAAA,IACpD;AAAA,EACF;AACA,MAAI,QAAQ,SAAS,SAAS;AAC5B,WAAO,EAAE,GAAG,SAAS,GAAI,MAAM,SAAS,KAAK,IAAI,EAAE,OAAO,MAAM,SAAS,KAAK,EAAE,IAAI,CAAC,EAAG;AAAA,EAC1F;AACA,QAAM,SAAS,SAAS,QAAQ,QAAQ,yBAAyB,EAAE;AACnE,SAAO;AAAA,IACL,GAAG;AAAA,IACH,GAAI,QAAQ,SAAS,QAAQ,MAAM,SAAY,EAAE,UAAU,QAAQ,SAAS,QAAQ,EAAE,IAAI,CAAC;AAAA,IAC3F,GAAI,OAAO,SAAS,aAAa,YAAY,EAAE,UAAU,QAAQ,SAAS,IAAI,CAAC;AAAA,IAC/E,GAAI,OAAO,SAAS,wBAAwB,YACxC,EAAE,qBAAqB,QAAQ,oBAAoB,IAAI,CAAC;AAAA,IAC5D,GAAI,QAAQ,QAAQ,IAAI,IAAI,EAAE,SAAS,QAAQ,QAAQ,IAAI,EAAE,IAAI,CAAC;AAAA,EACpE;AACF;;;AC5HO,IAAM,qBAAqB;AAO3B,SAAS,kBAAkB,UAAmD;AACnF,MAAI,SAAS,GAAI,QAAO;AAIxB,QAAM,WAAW,SAAS,SAAS;AACnC,QAAM,SAAU,OAAO,aAAa,YAAY,YAC3C,uBAAuB,SAAS,OAAO,KACvC,SAAS,WACT;AACL,SAAO,OAAO,MAAM,GAAG,kBAAkB;AAC3C;AAGA,IAAM,yBAAiD;AAAA,EACrD,+BAA+B;AAAA,EAC/B,iCAAiC;AACnC;AAOO,SAAS,eAAe,OAKT;AACpB,QAAM,MAAM,MAAM,OAAO,KAAK;AAC9B,SAAO;AAAA,IACL,SAAS,OAAO,SAAS,YAAY;AACnC,YAAM,YAAY,IAAI;AAGtB,YAAM,cAAc,MAAM,eAAe,QAAQ,SAAS;AAC1D,YAAM,sBAAsB,4BAA4B,OAAO;AAC/D,YAAM,MAAM,KAAK;AAAA,QACf,MAAM;AAAA,QAAQ,OAAO;AAAA,QAAW,MAAM,QAAQ;AAAA,QAAM;AAAA,QACpD,GAAI,sBAAsB,EAAE,cAAc,oBAAoB,IAAI,CAAC;AAAA,MACrE,CAAC;AACD,YAAM,WAAW,MAAM,MAAM,OAAO,QAAQ,SAAS,OAAO;AAC5D,YAAM,wBAAwB,2BAA2B,SAAS,QAAQ;AAC1E,YAAM,gBAAgB,kBAAkB,QAAQ;AAChD,YAAM,MAAM,KAAK;AAAA,QACf,MAAM;AAAA,QAAQ,OAAO;AAAA,QAAa,MAAM,QAAQ;AAAA,QAChD,IAAI,SAAS;AAAA,QAAI,YAAY,IAAI,IAAI;AAAA,QAAW;AAAA,QAChD,GAAI,gBAAgB,EAAE,cAAc,IAAI,CAAC;AAAA,QACzC,GAAI,wBAAwB,EAAE,cAAc,sBAAsB,IAAI,CAAC;AAAA,MACzE,CAAC;AACD,aAAO;AAAA,IACT;AAAA,EACF;AACF;","names":[]}

package/dist/{chunk-WM34GGTK.js → chunk-TMBYA5JX.js}

@@ -1,16 +1,16 @@
 import {
   observedBroker
-} from "./chunk-FAN2R3GW.js";
+} from "./chunk-KXFI3WM2.js";
 import {
   SKIP_WORKSPACE_DIRS,
   assertPinnedImage,
   stageWorkspace,
   stageWorkspacePair,
   verifyContainerEngineBoundary
-} from "./chunk-5LRYJKUI.js";
+} from "./chunk-FUDTSAOZ.js";
 import {
   HARNESS_PROTOCOL_VERSION
-} from "./chunk-U324RQ4N.js";
+} from "./chunk-CPV7ZVHH.js";
 
 // src/workspace-digest.ts
 import { createHash } from "crypto";
@@ -1801,10 +1801,12 @@ Prefer these over listing the tree \u2014 a full listing of a real repository is
 of thousands of tokens and you will carry it for the rest of the session.
 
 Then odla_search for a literal string, odla_read for a bounded range, and
-odla_apply_git_diff to change something. A patch must start with
-"diff --git a/<path> b/<path>", include matching "---" and "+++" headers and
-numbered "@@" hunks with at least one line of surrounding context, and must never
-use "*** Begin Patch" wrappers.
+odla_edit_file to change an existing file: give it the exact current text
+(copied verbatim from odla_read) and the text that replaces it; it must match
+once. Use odla_apply_git_diff only to create or delete whole files. A patch must
+start with "diff --git a/<path> b/<path>", include matching "---" and "+++"
+headers and numbered "@@" hunks with at least one line of surrounding context,
+and must never use "*** Begin Patch" wrappers.
 
 The workspace, model, and tool effects are controlled by the host broker.
 Never claim a build or test passed unless odla_run_recipe returned that result.`;
@@ -1864,6 +1866,21 @@ function codeSkill(opts) {
     },
     handler: (input, ctx) => call("sandbox.apply_patch", input, ctx.signal)
   };
+  const editFile = {
+    name: "odla_edit_file",
+    description: "Change an existing file by replacing one exact stretch of its current text with new text. oldText must match the file exactly once (copy it verbatim from odla_read, including indentation and blank lines; widen it if it would match more than once). Prefer this over odla_apply_git_diff for every change to an existing file; use odla_apply_git_diff only to create or delete whole files.",
+    inputSchema: {
+      type: "object",
+      required: ["path", "oldText", "newText"],
+      properties: {
+        path: { type: "string", minLength: 1, maxLength: 1024 },
+        oldText: { type: "string", minLength: 1, maxLength: 262144 },
+        newText: { type: "string", maxLength: 262144 }
+      },
+      additionalProperties: false
+    },
+    handler: (input, ctx) => call("sandbox.edit", input, ctx.signal)
+  };
   const runRecipe = {
     name: "odla_run_recipe",
     description: `Run one app-registered build or test recipe through CaMeL policy.${opts.recipeIds?.length ? ` Available recipes: ${opts.recipeIds.join(", ")}.` : ""}`,
@@ -1950,8 +1967,9 @@ function codeSkill(opts) {
       true
     )
   ];
-  const effects = opts.readOnly ? [] : [applyPatch, runRecipe];
-  const tools = opts.surface === "v3" ? [...orientation, searchFiles, read2, ...effects] : opts.surface === "v2" ? [listFiles, searchFiles, read2, ...effects] : [read2, ...effects];
+  const runsRecipes = opts.recipes ?? !opts.readOnly;
+  const effects = [...opts.readOnly ? [] : [applyPatch], ...runsRecipes ? [runRecipe] : []];
+  const tools = opts.surface === "v3" ? [...orientation, searchFiles, read2, ...opts.readOnly ? [] : [editFile], ...effects] : opts.surface === "v2" ? [listFiles, searchFiles, read2, ...effects] : [read2, ...effects];
   return { name: "code", tools };
 }
 
@@ -2013,6 +2031,7 @@ async function runCodeAgentAttempt(options) {
       model: "brokered",
       surface,
       ...options.readOnly === void 0 ? {} : { readOnly: options.readOnly },
+      ...options.recipes === void 0 ? {} : { recipes: options.recipes },
       ...options.recipeIds ? { recipeIds: options.recipeIds } : {},
       ...options.extraSkills ? { extraSkills: options.extraSkills } : {},
       ...options.maxSteps === void 0 ? {} : { maxSteps: options.maxSteps },
@@ -2397,6 +2416,11 @@ var PATCH = descriptor("sandbox.apply_patch", "reversible_mutation", {
   authority: "authority",
   patch: "payload"
 });
+var EDIT = descriptor("sandbox.edit", "reversible_mutation", {
+  workspace: "destination",
+  authority: "authority",
+  patch: "payload"
+});
 var RECIPE = descriptor("sandbox.run_recipe", "code_execution", {
   workspace: "destination",
   authority: "authority",
@@ -2464,6 +2488,14 @@ function createCodePolicyGate(options) {
         patch: { role: "payload", value: patch2 }
       }, []);
     },
+    edit: async (input) => {
+      const base = await environment(input, options, "sandbox.edit");
+      const patch2 = unsafe(base, input.patch, "patch");
+      return authorize(input, options, base, EDIT, {
+        ...base.fixedArgs,
+        patch: { role: "payload", value: patch2 }
+      }, []);
+    },
     recipe: async (input) => {
       const base = await environment(input, options, "sandbox.run_recipe");
       const conversions = await conversionRegistry([
@@ -2608,8 +2640,8 @@ function response(request, ok, content, details) {
   return { requestId: request.requestId, ok, content, details: { ...details, failureReason } };
 }
 
-// src/code-tool-reads.ts
-import { readFile as readFile4, stat } from "fs/promises";
+// src/code-tool-edit.ts
+import { readFile as readFile4, writeFile as writeFile3 } from "fs/promises";
 
 // src/code-tool-graph.ts
 import { readFile as readFile3 } from "fs/promises";
@@ -2679,7 +2711,65 @@ function renderWhoTouches(graphs, query) {
   }).join("\n\n");
 }
 
+// src/code-tool-edit.ts
+var MAX_TEXT_BYTES = 256 * 1024;
+function editAsDiff(path, oldText, newText) {
+  const removed = oldText.split("\n").map((line) => `-${line}`);
+  const added = newText.split("\n").map((line) => `+${line}`);
+  return `diff --git a/${path} b/${path}
+--- a/${path}
++++ b/${path}
+@@ -1,${removed.length} +1,${added.length} @@
+${[...removed, ...added].join("\n")}
+`;
+}
+function occurrences(haystack, needle) {
+  let count = 0;
+  for (let at = haystack.indexOf(needle); at !== -1; at = haystack.indexOf(needle, at + needle.length)) count += 1;
+  return count;
+}
+async function editCodeFile(workspaceDir, path, oldText, newText) {
+  if (!oldText) throw new TypeError("oldText must not be empty; to create a file use sandbox.apply_patch with a new-file diff");
+  if (Buffer.byteLength(oldText) > MAX_TEXT_BYTES || Buffer.byteLength(newText) > MAX_TEXT_BYTES) {
+    throw new TypeError(`oldText and newText must each stay under ${MAX_TEXT_BYTES} bytes; split the change`);
+  }
+  if (oldText.includes("\0") || newText.includes("\0")) throw new TypeError("edit text contains NUL bytes; use plain text");
+  const target = resolveCodePath(workspaceDir, path);
+  const current = await readFile4(target, "utf8");
+  const found = occurrences(current, oldText);
+  if (found === 0) {
+    throw new TypeError(`oldText was not found in "${path}"; sandbox.read the current lines and copy them exactly, including indentation and blank lines`);
+  }
+  if (found > 1) {
+    throw new TypeError(`oldText occurs ${found} times in "${path}"; include more of the surrounding lines so it matches exactly once`);
+  }
+  const at = current.indexOf(oldText);
+  await writeFile3(target, `${current.slice(0, at)}${newText}${current.slice(at + oldText.length)}`, "utf8");
+  return { deletions: oldText.split("\n").length, additions: newText.split("\n").length };
+}
+async function edit(context, request, options, policy, registry) {
+  exactKeys(request.input, ["path", "oldText", "newText"]);
+  const path = stringField(request.input, "path");
+  const oldText = stringField(request.input, "oldText");
+  const newText = typeof request.input.newText === "string" ? request.input.newText : stringField(request.input, "newText");
+  const paths = await registry.files(context.workspaceDir);
+  if (!paths.includes(path)) {
+    throw new TypeError(`no such file in the staged workspace: "${path}". Use sandbox.overview, sandbox.where_is or sandbox.search to find the correct path.`);
+  }
+  if (options.readOnlyPrefixes?.some((prefix) => path === prefix || path.startsWith(`${prefix}/`))) {
+    throw new TypeError("edit targets a read-only reference source");
+  }
+  const patch2 = editAsDiff(path, oldText, newText);
+  const allowed = await policy.edit(policyContext(context, request, options, { patch: patch2 }));
+  if (!allowed) return response(request, false, "tool denied by CaMeL policy");
+  const changed = await editCodeFile(context.workspaceDir, path, oldText, newText);
+  registry.invalidate(context.workspaceDir);
+  forgetWorkspaceGraphs(context.workspaceDir);
+  return response(request, true, `Edited ${path}: -${changed.deletions} +${changed.additions} line(s).`, { paths: [path], ...changed });
+}
+
 // src/code-tool-reads.ts
+import { readFile as readFile5, stat } from "fs/promises";
 var GRAPH_TOOLS = /* @__PURE__ */ new Set([
   "sandbox.overview",
   "sandbox.where_is",
@@ -2705,7 +2795,7 @@ async function read(context, request, options, policy, registry) {
   if (!info.isFile() || info.size > Math.max(options.maxReadBytes ?? 128 * 1024, 2 * 1024 * 1024)) {
     throw new TypeError("file is not a bounded regular source file");
   }
-  const source = await readFile4(target);
+  const source = await readFile5(target);
   if (source.includes(0)) throw new TypeError("binary files are not readable through this tool");
   const lines = source.toString("utf8").split("\n");
   const content = lines.slice(startLine - 1, endLine).join("\n");
@@ -2820,6 +2910,7 @@ async function route(context, request, options, recipes, policy, registry) {
     if (request.tool === "sandbox.search") return await search(context, request, options, policy, registry);
     if (GRAPH_TOOLS.has(request.tool)) return await graphQuery(context, request, options, policy, registry);
     if (request.tool === "sandbox.apply_patch") return await patch(context, request, options, policy, registry);
+    if (request.tool === "sandbox.edit") return await edit(context, request, options, policy, registry);
     return await recipe(context, request, options, recipes, policy);
   } catch (reason) {
     return response(request, false, toolFailureMessage(reason));
@@ -3529,6 +3620,9 @@ var TheseusRuntimeEngine = class {
       prompt: metadata.prompt,
       signal: active.abort.signal,
       readOnly: metadata.readOnly,
+      // A review-role session is read-only and still runs the proof; a planner
+      // marked read-only by its payload runs nothing.
+      recipes: metadata.role === "review" || !metadata.readOnly,
       recipeIds: this.options.recipes.map((recipe2) => recipe2.id),
       ...extraSkills.length ? { extraSkills } : {}
     });
@@ -3641,4 +3735,4 @@ export {
   runGoal,
   TheseusRuntimeEngine
 };
-//# sourceMappingURL=chunk-WM34GGTK.js.map
+//# sourceMappingURL=chunk-TMBYA5JX.js.map