@odla-ai/harness 0.10.1 → 0.10.2

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 (38) hide show
  1. package/dist/{chunk-U324RQ4N.js → chunk-CPV7ZVHH.js} +1 -1
  2. package/dist/{chunk-U324RQ4N.js.map → chunk-CPV7ZVHH.js.map} +1 -1
  3. package/dist/{chunk-OZBJNTML.js → chunk-ESLFS3FY.js} +4 -4
  4. package/dist/{chunk-5LRYJKUI.js → chunk-FUDTSAOZ.js} +3 -3
  5. package/dist/{chunk-VDY5V7ZG.js → chunk-JV45JAAW.js} +3 -3
  6. package/dist/chunk-JV45JAAW.js.map +1 -0
  7. package/dist/{chunk-FAN2R3GW.js → chunk-KXFI3WM2.js} +5 -1
  8. package/dist/chunk-KXFI3WM2.js.map +1 -0
  9. package/dist/{chunk-WM34GGTK.js → chunk-VIZALA6O.js} +101 -12
  10. package/dist/chunk-VIZALA6O.js.map +1 -0
  11. package/dist/cli.cjs +1 -1
  12. package/dist/cli.cjs.map +1 -1
  13. package/dist/cli.js +4 -4
  14. package/dist/code-runtime-cli.cjs +173 -78
  15. package/dist/code-runtime-cli.cjs.map +1 -1
  16. package/dist/code-runtime-cli.js +5 -5
  17. package/dist/index.cjs +5 -1
  18. package/dist/index.cjs.map +1 -1
  19. package/dist/index.d.cts +2 -2
  20. package/dist/index.d.ts +2 -2
  21. package/dist/index.js +3 -3
  22. package/dist/node.cjs +177 -82
  23. package/dist/node.cjs.map +1 -1
  24. package/dist/node.d.cts +2 -2
  25. package/dist/node.d.ts +2 -2
  26. package/dist/node.js +6 -6
  27. package/dist/testing.cjs.map +1 -1
  28. package/dist/testing.d.cts +1 -1
  29. package/dist/testing.d.ts +1 -1
  30. package/dist/testing.js +1 -1
  31. package/dist/{types-_8y8vBDI.d.ts → types-ocy70g_p.d.cts} +1 -1
  32. package/dist/{types-_8y8vBDI.d.cts → types-ocy70g_p.d.ts} +1 -1
  33. package/package.json +1 -1
  34. package/dist/chunk-FAN2R3GW.js.map +0 -1
  35. package/dist/chunk-VDY5V7ZG.js.map +0 -1
  36. package/dist/chunk-WM34GGTK.js.map +0 -1
  37. /package/dist/{chunk-OZBJNTML.js.map → chunk-ESLFS3FY.js.map} +0 -0
  38. /package/dist/{chunk-5LRYJKUI.js.map → chunk-FUDTSAOZ.js.map} +0 -0
package/dist/node.d.cts CHANGED
@@ -1,4 +1,4 @@
1
- import { v as HarnessTaskSpec, i as HarnessAgentOutput, h as HarnessAgentInput, H as HarnessControlPlane, f as HarnessToolBroker, a as HarnessLease, d as HarnessInferenceRequest, e as HarnessInferenceResponse, C as CodeSessionEventData, y as HarnessToolName, z as HarnessToolRequest } from './types-_8y8vBDI.cjs';
1
+ import { v as HarnessTaskSpec, i as HarnessAgentOutput, h as HarnessAgentInput, H as HarnessControlPlane, f as HarnessToolBroker, a as HarnessLease, d as HarnessInferenceRequest, e as HarnessInferenceResponse, C as CodeSessionEventData, y as HarnessToolName, z as HarnessToolRequest } from './types-ocy70g_p.cjs';
2
2
  import { CodePortableCheckpoint, CodeSourceFile, CodeVerificationReceipt, CodeCheckpointState } from '@odla-ai/camel/code';
3
3
  import { TaintLabel, ToolOutput, Skill, Inference, AgentRunBudget, AgentRun, CompactionPolicy } from '@odla-ai/ai';
4
4
  import { PolicyOutcome } from '@odla-ai/camel/policy';
@@ -428,7 +428,7 @@ declare const V2_SYSTEM_PROMPT = "You are the coding agent inside an odla Code h
428
428
  * three tools so the recorded baseline stays comparable. */
429
429
  type CodeSurface = "v1" | "v2" | "v3";
430
430
  /** v3 leads with orientation, because that is where the tokens went. */
431
- declare const V3_SYSTEM_PROMPT = "You are the coding agent inside an odla Code harness.\n\nOrient before you look. odla_overview gives the directory shape of the whole\nrepository in a few hundred lines; odla_where_is finds where a symbol is defined,\ndisambiguated by package; odla_who_imports finds what depends on a file; and\nodla_who_touches finds the code that reads and writes a table or database\nnamespace, which is how a bug report about wrong data becomes a file path.\nPrefer these over listing the tree \u2014 a full listing of a real repository is tens\nof thousands of tokens and you will carry it for the rest of the session.\n\nThen odla_search for a literal string, odla_read for a bounded range, and\nodla_apply_git_diff to change something. A patch must start with\n\"diff --git a/<path> b/<path>\", include matching \"---\" and \"+++\" headers and\nnumbered \"@@\" hunks with at least one line of surrounding context, and must never\nuse \"*** Begin Patch\" wrappers.\n\nThe workspace, model, and tool effects are controlled by the host broker.\nNever claim a build or test passed unless odla_run_recipe returned that result.";
431
+ declare const V3_SYSTEM_PROMPT = "You are the coding agent inside an odla Code harness.\n\nOrient before you look. odla_overview gives the directory shape of the whole\nrepository in a few hundred lines; odla_where_is finds where a symbol is defined,\ndisambiguated by package; odla_who_imports finds what depends on a file; and\nodla_who_touches finds the code that reads and writes a table or database\nnamespace, which is how a bug report about wrong data becomes a file path.\nPrefer these over listing the tree \u2014 a full listing of a real repository is tens\nof thousands of tokens and you will carry it for the rest of the session.\n\nThen odla_search for a literal string, odla_read for a bounded range, and\nodla_edit_file to change an existing file: give it the exact current text\n(copied verbatim from odla_read) and the text that replaces it; it must match\nonce. Use odla_apply_git_diff only to create or delete whole files. A patch must\nstart with \"diff --git a/<path> b/<path>\", include matching \"---\" and \"+++\"\nheaders and numbered \"@@\" hunks with at least one line of surrounding context,\nand must never use \"*** Begin Patch\" wrappers.\n\nThe workspace, model, and tool effects are controlled by the host broker.\nNever claim a build or test passed unless odla_run_recipe returned that result.";
432
432
  /**
433
433
  * The system prompt for each tool surface, keyed by version.
434
434
  *
package/dist/node.d.ts CHANGED
@@ -1,4 +1,4 @@
1
- import { v as HarnessTaskSpec, i as HarnessAgentOutput, h as HarnessAgentInput, H as HarnessControlPlane, f as HarnessToolBroker, a as HarnessLease, d as HarnessInferenceRequest, e as HarnessInferenceResponse, C as CodeSessionEventData, y as HarnessToolName, z as HarnessToolRequest } from './types-_8y8vBDI.js';
1
+ import { v as HarnessTaskSpec, i as HarnessAgentOutput, h as HarnessAgentInput, H as HarnessControlPlane, f as HarnessToolBroker, a as HarnessLease, d as HarnessInferenceRequest, e as HarnessInferenceResponse, C as CodeSessionEventData, y as HarnessToolName, z as HarnessToolRequest } from './types-ocy70g_p.js';
2
2
  import { CodePortableCheckpoint, CodeSourceFile, CodeVerificationReceipt, CodeCheckpointState } from '@odla-ai/camel/code';
3
3
  import { TaintLabel, ToolOutput, Skill, Inference, AgentRunBudget, AgentRun, CompactionPolicy } from '@odla-ai/ai';
4
4
  import { PolicyOutcome } from '@odla-ai/camel/policy';
@@ -428,7 +428,7 @@ declare const V2_SYSTEM_PROMPT = "You are the coding agent inside an odla Code h
428
428
  * three tools so the recorded baseline stays comparable. */
429
429
  type CodeSurface = "v1" | "v2" | "v3";
430
430
  /** v3 leads with orientation, because that is where the tokens went. */
431
- declare const V3_SYSTEM_PROMPT = "You are the coding agent inside an odla Code harness.\n\nOrient before you look. odla_overview gives the directory shape of the whole\nrepository in a few hundred lines; odla_where_is finds where a symbol is defined,\ndisambiguated by package; odla_who_imports finds what depends on a file; and\nodla_who_touches finds the code that reads and writes a table or database\nnamespace, which is how a bug report about wrong data becomes a file path.\nPrefer these over listing the tree \u2014 a full listing of a real repository is tens\nof thousands of tokens and you will carry it for the rest of the session.\n\nThen odla_search for a literal string, odla_read for a bounded range, and\nodla_apply_git_diff to change something. A patch must start with\n\"diff --git a/<path> b/<path>\", include matching \"---\" and \"+++\" headers and\nnumbered \"@@\" hunks with at least one line of surrounding context, and must never\nuse \"*** Begin Patch\" wrappers.\n\nThe workspace, model, and tool effects are controlled by the host broker.\nNever claim a build or test passed unless odla_run_recipe returned that result.";
431
+ declare const V3_SYSTEM_PROMPT = "You are the coding agent inside an odla Code harness.\n\nOrient before you look. odla_overview gives the directory shape of the whole\nrepository in a few hundred lines; odla_where_is finds where a symbol is defined,\ndisambiguated by package; odla_who_imports finds what depends on a file; and\nodla_who_touches finds the code that reads and writes a table or database\nnamespace, which is how a bug report about wrong data becomes a file path.\nPrefer these over listing the tree \u2014 a full listing of a real repository is tens\nof thousands of tokens and you will carry it for the rest of the session.\n\nThen odla_search for a literal string, odla_read for a bounded range, and\nodla_edit_file to change an existing file: give it the exact current text\n(copied verbatim from odla_read) and the text that replaces it; it must match\nonce. Use odla_apply_git_diff only to create or delete whole files. A patch must\nstart with \"diff --git a/<path> b/<path>\", include matching \"---\" and \"+++\"\nheaders and numbered \"@@\" hunks with at least one line of surrounding context,\nand must never use \"*** Begin Patch\" wrappers.\n\nThe workspace, model, and tool effects are controlled by the host broker.\nNever claim a build or test passed unless odla_run_recipe returned that result.";
432
432
  /**
433
433
  * The system prompt for each tool surface, keyed by version.
434
434
  *
package/dist/node.js CHANGED
@@ -1,7 +1,7 @@
1
1
  import {
2
2
  runHarnessRunner,
3
3
  runLeasedAttempt
4
- } from "./chunk-OZBJNTML.js";
4
+ } from "./chunk-ESLFS3FY.js";
5
5
  import {
6
6
  CODE_RUNTIME_PROTOCOL_VERSION,
7
7
  CodeRuntimeCheckpointManager,
@@ -49,8 +49,8 @@ import {
49
49
  validateMemory,
50
50
  validateRelativePath,
51
51
  verifyCodeCandidate
52
- } from "./chunk-WM34GGTK.js";
53
- import "./chunk-FAN2R3GW.js";
52
+ } from "./chunk-VIZALA6O.js";
53
+ import "./chunk-KXFI3WM2.js";
54
54
  import {
55
55
  assertPinnedImage,
56
56
  buildContainerRunArgs,
@@ -61,9 +61,9 @@ import {
61
61
  stageWorkspace,
62
62
  stageWorkspacePair,
63
63
  verifyContainerEngineBoundary
64
- } from "./chunk-5LRYJKUI.js";
65
- import "./chunk-VDY5V7ZG.js";
66
- import "./chunk-U324RQ4N.js";
64
+ } from "./chunk-FUDTSAOZ.js";
65
+ import "./chunk-JV45JAAW.js";
66
+ import "./chunk-CPV7ZVHH.js";
67
67
 
68
68
  // src/code-runtime-memory.ts
69
69
  async function mutationId(memory) {
@@ -1 +1 @@
1
- {"version":3,"sources":["../src/testing.ts","../src/types.ts"],"sourcesContent":["import type { OracleResponse } from \"@odla-ai/ai\";\nimport {\n HARNESS_PROTOCOL_VERSION,\n type HarnessCompletion,\n type HarnessControlPlane,\n type HarnessEventInput,\n type HarnessInferenceRequest,\n type HarnessInferenceResponse,\n type HarnessLease,\n} from \"./types\";\n\n/** Create a deterministic, valid lease for harness unit and integration tests. */\nexport function fixtureLease(overrides: Partial<HarnessLease> = {}): HarnessLease {\n return {\n protocolVersion: HARNESS_PROTOCOL_VERSION,\n leaseId: \"lease_fixture\",\n generation: 1,\n expiresAt: Date.now() + 60_000,\n task: {\n taskId: \"task_fixture\",\n attemptId: \"attempt_fixture\",\n title: \"Fixture task\",\n prompt: \"Create HARNESS_RESULT.md\",\n workspace: \"fixture\",\n aiRoute: \"coding\",\n policy: {\n network: \"none\",\n timeoutMs: 30_000,\n maxOutputBytes: 1_000_000,\n maxPatchBytes: 1_000_000,\n },\n },\n ...overrides,\n };\n}\n\n/** In-memory control plane that records runner interactions without network access. */\nexport class FakeHarnessControlPlane implements HarnessControlPlane {\n readonly leases: HarnessLease[] = [];\n readonly events: HarnessEventInput[] = [];\n readonly completions: HarnessCompletion[] = [];\n readonly inferenceRequests: HarnessInferenceRequest[] = [];\n cancelRequested = false;\n response: OracleResponse = {\n id: \"oracle_fixture\",\n model: \"fixture-model\",\n provider: \"openai\",\n role: \"assistant\",\n content: [{ type: \"text\", text: \"deterministic fixture response\" }],\n stopReason: \"end_turn\",\n usage: { inputTokens: 5, outputTokens: 4 },\n };\n\n constructor(...leases: HarnessLease[]) { this.leases.push(...leases); }\n\n async lease(workspaces: string[]): Promise<HarnessLease | null> {\n const index = this.leases.findIndex((candidate) => workspaces.includes(candidate.task.workspace));\n return index < 0 ? null : this.leases.splice(index, 1)[0]!;\n }\n async heartbeat(): Promise<{ cancelRequested: boolean; expiresAt: number }> {\n return { cancelRequested: this.cancelRequested, expiresAt: Date.now() + 60_000 };\n }\n async appendEvents(_attemptId: string, _leaseId: string, events: HarnessEventInput[]): Promise<void> {\n this.events.push(...events);\n }\n async infer(_attemptId: string, _leaseId: string, request: HarnessInferenceRequest): Promise<HarnessInferenceResponse> {\n this.inferenceRequests.push(request);\n return {\n requestId: request.requestId,\n response: this.response,\n receipt: {\n provider: this.response.provider,\n model: this.response.model,\n policyVersion: 1,\n inputTokens: this.response.usage.inputTokens ?? 0,\n outputTokens: this.response.usage.outputTokens ?? 0,\n },\n };\n }\n async complete(_attemptId: string, _leaseId: string, completion: HarnessCompletion): Promise<void> {\n this.completions.push(completion);\n }\n}\n","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":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACKO,IAAM,2BAA2B;;;ADOjC,SAAS,aAAa,YAAmC,CAAC,GAAiB;AAChF,SAAO;AAAA,IACL,iBAAiB;AAAA,IACjB,SAAS;AAAA,IACT,YAAY;AAAA,IACZ,WAAW,KAAK,IAAI,IAAI;AAAA,IACxB,MAAM;AAAA,MACJ,QAAQ;AAAA,MACR,WAAW;AAAA,MACX,OAAO;AAAA,MACP,QAAQ;AAAA,MACR,WAAW;AAAA,MACX,SAAS;AAAA,MACT,QAAQ;AAAA,QACN,SAAS;AAAA,QACT,WAAW;AAAA,QACX,gBAAgB;AAAA,QAChB,eAAe;AAAA,MACjB;AAAA,IACF;AAAA,IACA,GAAG;AAAA,EACL;AACF;AAGO,IAAM,0BAAN,MAA6D;AAAA,EACzD,SAAyB,CAAC;AAAA,EAC1B,SAA8B,CAAC;AAAA,EAC/B,cAAmC,CAAC;AAAA,EACpC,oBAA+C,CAAC;AAAA,EACzD,kBAAkB;AAAA,EAClB,WAA2B;AAAA,IACzB,IAAI;AAAA,IACJ,OAAO;AAAA,IACP,UAAU;AAAA,IACV,MAAM;AAAA,IACN,SAAS,CAAC,EAAE,MAAM,QAAQ,MAAM,iCAAiC,CAAC;AAAA,IAClE,YAAY;AAAA,IACZ,OAAO,EAAE,aAAa,GAAG,cAAc,EAAE;AAAA,EAC3C;AAAA,EAEA,eAAe,QAAwB;AAAE,SAAK,OAAO,KAAK,GAAG,MAAM;AAAA,EAAG;AAAA,EAEtE,MAAM,MAAM,YAAoD;AAC9D,UAAM,QAAQ,KAAK,OAAO,UAAU,CAAC,cAAc,WAAW,SAAS,UAAU,KAAK,SAAS,CAAC;AAChG,WAAO,QAAQ,IAAI,OAAO,KAAK,OAAO,OAAO,OAAO,CAAC,EAAE,CAAC;AAAA,EAC1D;AAAA,EACA,MAAM,YAAsE;AAC1E,WAAO,EAAE,iBAAiB,KAAK,iBAAiB,WAAW,KAAK,IAAI,IAAI,IAAO;AAAA,EACjF;AAAA,EACA,MAAM,aAAa,YAAoB,UAAkB,QAA4C;AACnG,SAAK,OAAO,KAAK,GAAG,MAAM;AAAA,EAC5B;AAAA,EACA,MAAM,MAAM,YAAoB,UAAkB,SAAqE;AACrH,SAAK,kBAAkB,KAAK,OAAO;AACnC,WAAO;AAAA,MACL,WAAW,QAAQ;AAAA,MACnB,UAAU,KAAK;AAAA,MACf,SAAS;AAAA,QACP,UAAU,KAAK,SAAS;AAAA,QACxB,OAAO,KAAK,SAAS;AAAA,QACrB,eAAe;AAAA,QACf,aAAa,KAAK,SAAS,MAAM,eAAe;AAAA,QAChD,cAAc,KAAK,SAAS,MAAM,gBAAgB;AAAA,MACpD;AAAA,IACF;AAAA,EACF;AAAA,EACA,MAAM,SAAS,YAAoB,UAAkB,YAA8C;AACjG,SAAK,YAAY,KAAK,UAAU;AAAA,EAClC;AACF;","names":[]}
1
+ {"version":3,"sources":["../src/testing.ts","../src/types.ts"],"sourcesContent":["import type { OracleResponse } from \"@odla-ai/ai\";\nimport {\n HARNESS_PROTOCOL_VERSION,\n type HarnessCompletion,\n type HarnessControlPlane,\n type HarnessEventInput,\n type HarnessInferenceRequest,\n type HarnessInferenceResponse,\n type HarnessLease,\n} from \"./types\";\n\n/** Create a deterministic, valid lease for harness unit and integration tests. */\nexport function fixtureLease(overrides: Partial<HarnessLease> = {}): HarnessLease {\n return {\n protocolVersion: HARNESS_PROTOCOL_VERSION,\n leaseId: \"lease_fixture\",\n generation: 1,\n expiresAt: Date.now() + 60_000,\n task: {\n taskId: \"task_fixture\",\n attemptId: \"attempt_fixture\",\n title: \"Fixture task\",\n prompt: \"Create HARNESS_RESULT.md\",\n workspace: \"fixture\",\n aiRoute: \"coding\",\n policy: {\n network: \"none\",\n timeoutMs: 30_000,\n maxOutputBytes: 1_000_000,\n maxPatchBytes: 1_000_000,\n },\n },\n ...overrides,\n };\n}\n\n/** In-memory control plane that records runner interactions without network access. */\nexport class FakeHarnessControlPlane implements HarnessControlPlane {\n readonly leases: HarnessLease[] = [];\n readonly events: HarnessEventInput[] = [];\n readonly completions: HarnessCompletion[] = [];\n readonly inferenceRequests: HarnessInferenceRequest[] = [];\n cancelRequested = false;\n response: OracleResponse = {\n id: \"oracle_fixture\",\n model: \"fixture-model\",\n provider: \"openai\",\n role: \"assistant\",\n content: [{ type: \"text\", text: \"deterministic fixture response\" }],\n stopReason: \"end_turn\",\n usage: { inputTokens: 5, outputTokens: 4 },\n };\n\n constructor(...leases: HarnessLease[]) { this.leases.push(...leases); }\n\n async lease(workspaces: string[]): Promise<HarnessLease | null> {\n const index = this.leases.findIndex((candidate) => workspaces.includes(candidate.task.workspace));\n return index < 0 ? null : this.leases.splice(index, 1)[0]!;\n }\n async heartbeat(): Promise<{ cancelRequested: boolean; expiresAt: number }> {\n return { cancelRequested: this.cancelRequested, expiresAt: Date.now() + 60_000 };\n }\n async appendEvents(_attemptId: string, _leaseId: string, events: HarnessEventInput[]): Promise<void> {\n this.events.push(...events);\n }\n async infer(_attemptId: string, _leaseId: string, request: HarnessInferenceRequest): Promise<HarnessInferenceResponse> {\n this.inferenceRequests.push(request);\n return {\n requestId: request.requestId,\n response: this.response,\n receipt: {\n provider: this.response.provider,\n model: this.response.model,\n policyVersion: 1,\n inputTokens: this.response.usage.inputTokens ?? 0,\n outputTokens: this.response.usage.outputTokens ?? 0,\n },\n };\n }\n async complete(_attemptId: string, _leaseId: string, completion: HarnessCompletion): Promise<void> {\n this.completions.push(completion);\n }\n}\n","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":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACKO,IAAM,2BAA2B;;;ADOjC,SAAS,aAAa,YAAmC,CAAC,GAAiB;AAChF,SAAO;AAAA,IACL,iBAAiB;AAAA,IACjB,SAAS;AAAA,IACT,YAAY;AAAA,IACZ,WAAW,KAAK,IAAI,IAAI;AAAA,IACxB,MAAM;AAAA,MACJ,QAAQ;AAAA,MACR,WAAW;AAAA,MACX,OAAO;AAAA,MACP,QAAQ;AAAA,MACR,WAAW;AAAA,MACX,SAAS;AAAA,MACT,QAAQ;AAAA,QACN,SAAS;AAAA,QACT,WAAW;AAAA,QACX,gBAAgB;AAAA,QAChB,eAAe;AAAA,MACjB;AAAA,IACF;AAAA,IACA,GAAG;AAAA,EACL;AACF;AAGO,IAAM,0BAAN,MAA6D;AAAA,EACzD,SAAyB,CAAC;AAAA,EAC1B,SAA8B,CAAC;AAAA,EAC/B,cAAmC,CAAC;AAAA,EACpC,oBAA+C,CAAC;AAAA,EACzD,kBAAkB;AAAA,EAClB,WAA2B;AAAA,IACzB,IAAI;AAAA,IACJ,OAAO;AAAA,IACP,UAAU;AAAA,IACV,MAAM;AAAA,IACN,SAAS,CAAC,EAAE,MAAM,QAAQ,MAAM,iCAAiC,CAAC;AAAA,IAClE,YAAY;AAAA,IACZ,OAAO,EAAE,aAAa,GAAG,cAAc,EAAE;AAAA,EAC3C;AAAA,EAEA,eAAe,QAAwB;AAAE,SAAK,OAAO,KAAK,GAAG,MAAM;AAAA,EAAG;AAAA,EAEtE,MAAM,MAAM,YAAoD;AAC9D,UAAM,QAAQ,KAAK,OAAO,UAAU,CAAC,cAAc,WAAW,SAAS,UAAU,KAAK,SAAS,CAAC;AAChG,WAAO,QAAQ,IAAI,OAAO,KAAK,OAAO,OAAO,OAAO,CAAC,EAAE,CAAC;AAAA,EAC1D;AAAA,EACA,MAAM,YAAsE;AAC1E,WAAO,EAAE,iBAAiB,KAAK,iBAAiB,WAAW,KAAK,IAAI,IAAI,IAAO;AAAA,EACjF;AAAA,EACA,MAAM,aAAa,YAAoB,UAAkB,QAA4C;AACnG,SAAK,OAAO,KAAK,GAAG,MAAM;AAAA,EAC5B;AAAA,EACA,MAAM,MAAM,YAAoB,UAAkB,SAAqE;AACrH,SAAK,kBAAkB,KAAK,OAAO;AACnC,WAAO;AAAA,MACL,WAAW,QAAQ;AAAA,MACnB,UAAU,KAAK;AAAA,MACf,SAAS;AAAA,QACP,UAAU,KAAK,SAAS;AAAA,QACxB,OAAO,KAAK,SAAS;AAAA,QACrB,eAAe;AAAA,QACf,aAAa,KAAK,SAAS,MAAM,eAAe;AAAA,QAChD,cAAc,KAAK,SAAS,MAAM,gBAAgB;AAAA,MACpD;AAAA,IACF;AAAA,EACF;AAAA,EACA,MAAM,SAAS,YAAoB,UAAkB,YAA8C;AACjG,SAAK,YAAY,KAAK,UAAU;AAAA,EAClC;AACF;","names":[]}
@@ -1,5 +1,5 @@
1
1
  import { OracleResponse } from '@odla-ai/ai';
2
- import { H as HarnessControlPlane, a as HarnessLease, b as HarnessEventInput, c as HarnessCompletion, d as HarnessInferenceRequest, e as HarnessInferenceResponse } from './types-_8y8vBDI.cjs';
2
+ import { H as HarnessControlPlane, a as HarnessLease, b as HarnessEventInput, c as HarnessCompletion, d as HarnessInferenceRequest, e as HarnessInferenceResponse } from './types-ocy70g_p.cjs';
3
3
 
4
4
  /** Create a deterministic, valid lease for harness unit and integration tests. */
5
5
  declare function fixtureLease(overrides?: Partial<HarnessLease>): HarnessLease;
package/dist/testing.d.ts CHANGED
@@ -1,5 +1,5 @@
1
1
  import { OracleResponse } from '@odla-ai/ai';
2
- import { H as HarnessControlPlane, a as HarnessLease, b as HarnessEventInput, c as HarnessCompletion, d as HarnessInferenceRequest, e as HarnessInferenceResponse } from './types-_8y8vBDI.js';
2
+ import { H as HarnessControlPlane, a as HarnessLease, b as HarnessEventInput, c as HarnessCompletion, d as HarnessInferenceRequest, e as HarnessInferenceResponse } from './types-ocy70g_p.js';
3
3
 
4
4
  /** Create a deterministic, valid lease for harness unit and integration tests. */
5
5
  declare function fixtureLease(overrides?: Partial<HarnessLease>): HarnessLease;
package/dist/testing.js CHANGED
@@ -1,6 +1,6 @@
1
1
  import {
2
2
  HARNESS_PROTOCOL_VERSION
3
- } from "./chunk-U324RQ4N.js";
3
+ } from "./chunk-CPV7ZVHH.js";
4
4
 
5
5
  // src/testing.ts
6
6
  function fixtureLease(overrides = {}) {
@@ -246,7 +246,7 @@ type CodeSessionEvent = CodeSessionEventData & {
246
246
  createdAt: number;
247
247
  };
248
248
  /** Closed set of effects an agent container may request from its trusted broker. */
249
- type HarnessToolName = "sandbox.read" | "sandbox.list" | "sandbox.search" | "sandbox.overview" | "sandbox.where_is" | "sandbox.who_imports" | "sandbox.who_touches" | "sandbox.apply_patch" | "sandbox.run_recipe";
249
+ type HarnessToolName = "sandbox.read" | "sandbox.list" | "sandbox.search" | "sandbox.overview" | "sandbox.where_is" | "sandbox.who_imports" | "sandbox.who_touches" | "sandbox.apply_patch" | "sandbox.edit" | "sandbox.run_recipe";
250
250
  /** Correlated, structured tool request emitted by an untrusted agent container. */
251
251
  interface HarnessToolRequest {
252
252
  requestId: string;
@@ -246,7 +246,7 @@ type CodeSessionEvent = CodeSessionEventData & {
246
246
  createdAt: number;
247
247
  };
248
248
  /** Closed set of effects an agent container may request from its trusted broker. */
249
- type HarnessToolName = "sandbox.read" | "sandbox.list" | "sandbox.search" | "sandbox.overview" | "sandbox.where_is" | "sandbox.who_imports" | "sandbox.who_touches" | "sandbox.apply_patch" | "sandbox.run_recipe";
249
+ type HarnessToolName = "sandbox.read" | "sandbox.list" | "sandbox.search" | "sandbox.overview" | "sandbox.where_is" | "sandbox.who_imports" | "sandbox.who_touches" | "sandbox.apply_patch" | "sandbox.edit" | "sandbox.run_recipe";
250
250
  /** Correlated, structured tool request emitted by an untrusted agent container. */
251
251
  interface HarnessToolRequest {
252
252
  requestId: string;
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@odla-ai/harness",
3
- "version": "0.10.1",
3
+ "version": "0.10.2",
4
4
  "description": "Safe, inspectable coding-task protocol and credentialless container runner for odla Studio.",
5
5
  "license": "MIT",
6
6
  "homepage": "https://odla.ai/docs/packages/harness",
@@ -1 +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 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,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;;;ACxHO,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":[]}
@@ -1 +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.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.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,oBAAoB,EAAE,SAAS,IAAI,GAAG;AAC3F,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":[]}