@effect-agent/pr-review 0.0.1-beta.0 → 0.1.0-beta.8

package/dist/testing.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"testing.mjs","names":[],"sources":["../src/internal/scripted.ts","../src/internal/fan-out-scripted.ts","../src/internal/fixtures.ts"],"sourcesContent":["import { Effect, Layer, Ref, Schema, Stream } from \"effect\";\nimport { LanguageModel, Model, type Response } from \"effect/unstable/ai\";\n\nimport { CodeReview } from \"./review-agent.ts\";\n\n// ---------------------------------------------------------------------------\n// Deterministic offline model for the flat reviewer: a prompt-aware scripted\n// model that walks the real tool surface — list, diff, read — then returns\n// the scripted review as its terminal JSON. Decisions key on committed\n// history in the prompt, never on call order, so replays stay honest.\n// ---------------------------------------------------------------------------\n\nexport const OFFLINE_LIST_CALL_ID = \"list-1\";\nexport const OFFLINE_DIFF_CALL_ID = \"diff-1\";\nexport const OFFLINE_READ_CALL_ID = \"read-1\";\n\n/** Usage attached to EVERY scripted model turn; tests pin exact aggregates. */\nexport const SCRIPTED_TURN_USAGE = { inputTokens: 64, outputTokens: 48 } as const;\n\nconst scriptedUsage = {\n  inputTokens: { total: SCRIPTED_TURN_USAGE.inputTokens },\n  outputTokens: { total: SCRIPTED_TURN_USAGE.outputTokens },\n};\n\nexport const scriptedToolTurn = (\n  ...calls: ReadonlyArray<Response.StreamPartEncoded>\n): ReadonlyArray<Response.StreamPartEncoded> => [\n  ...calls,\n  { type: \"finish\", reason: \"tool-calls\", usage: scriptedUsage },\n];\n\nexport const scriptedFinalParts = (text: string): ReadonlyArray<Response.StreamPartEncoded> => [\n  { type: \"text-start\", id: \"code-review\" },\n  { type: \"text-delta\", id: \"code-review\", delta: text },\n  { type: \"text-end\", id: \"code-review\" },\n  { type: \"finish\", reason: \"stop\", usage: scriptedUsage },\n];\n\n/** A prompt-keyed scripted LanguageModel with call and prompt observability. */\nexport const makePromptKeyedModel = (\n  name: string,\n  decide: (promptJson: string) => ReadonlyArray<Response.StreamPartEncoded>,\n) =>\n  Effect.gen(function* () {\n    const calls = yield* Ref.make(0);\n    const prompts = yield* Ref.make<ReadonlyArray<string>>([]);\n    const model = Model.make(\n      \"scripted\",\n      name,\n      Layer.effect(\n        LanguageModel.LanguageModel,\n        LanguageModel.make({\n          generateText: () => Effect.succeed([]),\n          streamText: (request) =>\n            Stream.unwrap(\n              Effect.gen(function* () {\n                yield* Ref.update(calls, (value) => value + 1);\n                const promptJson = JSON.stringify(request.prompt);\n                yield* Ref.update(prompts, (previous) => [...previous, promptJson]);\n                return Stream.fromIterable(decide(promptJson));\n              }),\n            ),\n        }),\n      ),\n    );\n    return { model, calls: Ref.get(calls), prompts: Ref.get(prompts) };\n  });\n\n/**\n * Build the offline scripted reviewer model. Turn 1 lists the changeset,\n * Turn 2 reads one file diff, Turn 3 reads head context, Turn 4 returns the\n * scripted review JSON.\n */\nexport const makeOfflineReviewerModel = (script: {\n  readonly diffPath: string;\n  readonly readPath: string;\n  readonly review: CodeReview;\n}) =>\n  makePromptKeyedModel(\"pr-review-offline\", (promptJson) => {\n    if (promptJson.includes(OFFLINE_READ_CALL_ID)) {\n      return scriptedFinalParts(JSON.stringify(Schema.encodeSync(CodeReview)(script.review)));\n    }\n    if (promptJson.includes(OFFLINE_DIFF_CALL_ID)) {\n      return scriptedToolTurn({\n        type: \"tool-call\",\n        id: OFFLINE_READ_CALL_ID,\n        name: \"read_file\",\n        params: { path: script.readPath },\n        providerExecuted: false,\n      });\n    }\n    if (promptJson.includes(OFFLINE_LIST_CALL_ID)) {\n      return scriptedToolTurn({\n        type: \"tool-call\",\n        id: OFFLINE_DIFF_CALL_ID,\n        name: \"read_file_diff\",\n        params: { path: script.diffPath },\n        providerExecuted: false,\n      });\n    }\n    return scriptedToolTurn({\n      type: \"tool-call\",\n      id: OFFLINE_LIST_CALL_ID,\n      name: \"list_changed_files\",\n      params: { scope: \"all\" },\n      providerExecuted: false,\n    });\n  });\n","import { Effect, Layer, Ref, Schema, Stream } from \"effect\";\nimport { LanguageModel, Model, type Response } from \"effect/unstable/ai\";\n\nimport { FileReviewReport } from \"./fan-out.ts\";\nimport { CodeReview } from \"./review-agent.ts\";\nimport { makePromptKeyedModel, scriptedFinalParts, scriptedToolTurn } from \"./scripted.ts\";\n\n// ---------------------------------------------------------------------------\n// Deterministic offline models for the fan-out reviewer: prompt-keyed\n// scripted models for BOTH the coordinator and the file-reviewer children.\n// Both key every decision on committed history in the prompt (tool-call ids\n// and briefed unit ids), never on call order, so concurrent children and\n// replays stay honest.\n// ---------------------------------------------------------------------------\n\nexport const OFFLINE_UNITS_CALL_ID = \"units-1\";\n\n/** The delegation Tool Call id the scripted coordinator uses for one unit. */\nexport const offlineUnitCallId = (unitId: string): string => `delegate-${unitId}`;\n\n/** The diff Tool Call id the scripted child uses for one unit. */\nexport const offlineChildDiffCallId = (unitId: string): string => `fanout-diff-${unitId}`;\n\n/** One scripted delegation the offline coordinator declares. */\nexport interface OfflineUnitCall {\n  readonly unitId: string;\n  readonly paths: ReadonlyArray<string>;\n}\n\n/**\n * Build the offline scripted coordinator model. Turn 1 lists the review\n * units, Turn 2 declares one delegation Tool Call per scripted unit in one\n * batch, Turn 3 returns the scripted merged review JSON. Decisions key on\n * tool-call ids already committed to the prompt.\n */\nexport const makeOfflineFanOutCoordinatorModel = (script: {\n  readonly unitCalls: ReadonlyArray<OfflineUnitCall>;\n  readonly review: CodeReview;\n}) => {\n  const firstUnitCallId = offlineUnitCallId(script.unitCalls[0]?.unitId ?? \"unit-none\");\n  return makePromptKeyedModel(\"pr-fanout-coordinator-offline\", (promptJson) => {\n    if (promptJson.includes(firstUnitCallId)) {\n      return scriptedFinalParts(JSON.stringify(Schema.encodeSync(CodeReview)(script.review)));\n    }\n    if (promptJson.includes(OFFLINE_UNITS_CALL_ID)) {\n      return scriptedToolTurn(\n        ...script.unitCalls.map(\n          (unit): Response.StreamPartEncoded => ({\n            type: \"tool-call\",\n            id: offlineUnitCallId(unit.unitId),\n            name: \"delegate_file_review\",\n            params: { unitId: unit.unitId, paths: unit.paths },\n            providerExecuted: false,\n          }),\n        ),\n      );\n    }\n    return scriptedToolTurn({\n      type: \"tool-call\",\n      id: OFFLINE_UNITS_CALL_ID,\n      name: \"list_review_units\",\n      params: { scope: \"all\" },\n      providerExecuted: false,\n    });\n  });\n};\n\n/** How one scripted child behaves for its briefed unit. */\nexport type OfflineUnitOutcome =\n  /** Read one diff, then return the scripted report. */\n  | { readonly _tag: \"findings\"; readonly report: FileReviewReport }\n  /** Read one diff, then return non-JSON — the child fails typed (AgentOutputError). */\n  | { readonly _tag: \"malformed-output\" }\n  /**\n   * Declare more Tool Calls than the child's AgentPolicy allows in one turn —\n   * the child fails typed (AgentPolicyError \"tool-calls\") before any executes.\n   */\n  | { readonly _tag: \"budget-runaway\"; readonly declaredCalls: number };\n\nexport interface OfflineUnitScript {\n  readonly unitId: string;\n  /** The one file the scripted child reads the diff of. */\n  readonly diffPath: string;\n  readonly outcome: OfflineUnitOutcome;\n}\n\n/**\n * Build the offline scripted file-reviewer model shared by every delegated\n * child. Each child Run builds the Model Layer inside its own scope; the\n * script entry is selected by the briefed unitId present in the child's OWN\n * prompt, and the turn is selected by whether that unit's diff Tool Call id\n * is already committed there — content-keyed on both axes, so concurrent\n * children never interfere. First-turn child prompts are recorded for\n * context-isolation assertions.\n */\nexport const makeOfflineFileReviewerModel = (scripts: ReadonlyArray<OfflineUnitScript>) =>\n  Effect.gen(function* () {\n    const calls = yield* Ref.make(0);\n    const prompts = yield* Ref.make<ReadonlyArray<string>>([]);\n    const decide = (promptJson: string): ReadonlyArray<Response.StreamPartEncoded> | undefined => {\n      const script = scripts.find((candidate) => promptJson.includes(candidate.unitId));\n      if (script === undefined) return undefined;\n      switch (script.outcome._tag) {\n        case \"budget-runaway\": {\n          return scriptedToolTurn(\n            ...Array.from(\n              { length: script.outcome.declaredCalls },\n              (_, index): Response.StreamPartEncoded => ({\n                type: \"tool-call\",\n                id: `runaway-${script.unitId}-${index + 1}`,\n                name: \"read_file_diff\",\n                params: { path: script.diffPath },\n                providerExecuted: false,\n              }),\n            ),\n          );\n        }\n        case \"malformed-output\":\n        case \"findings\": {\n          if (promptJson.includes(offlineChildDiffCallId(script.unitId))) {\n            return scriptedFinalParts(\n              script.outcome._tag === \"findings\"\n                ? JSON.stringify(Schema.encodeSync(FileReviewReport)(script.outcome.report))\n                : \"this is not the JSON you are looking for\",\n            );\n          }\n          return scriptedToolTurn({\n            type: \"tool-call\",\n            id: offlineChildDiffCallId(script.unitId),\n            name: \"read_file_diff\",\n            params: { path: script.diffPath },\n            providerExecuted: false,\n          });\n        }\n      }\n    };\n    const model = Model.make(\n      \"scripted\",\n      \"pr-fanout-file-reviewer-offline\",\n      Layer.effect(\n        LanguageModel.LanguageModel,\n        LanguageModel.make({\n          generateText: () => Effect.succeed([]),\n          streamText: (request) =>\n            Stream.unwrap(\n              Effect.gen(function* () {\n                yield* Ref.update(calls, (value) => value + 1);\n                const promptJson = JSON.stringify(request.prompt);\n                yield* Ref.update(prompts, (previous) => [...previous, promptJson]);\n                const parts = decide(promptJson);\n                if (parts === undefined) {\n                  return yield* Effect.die(\n                    new Error(\"The child prompt names no scripted review unit\"),\n                  );\n                }\n                return Stream.fromIterable(parts);\n              }),\n            ),\n        }),\n      ),\n    );\n    return { model, calls: Ref.get(calls), prompts: Ref.get(prompts) };\n  });\n","import { Effect, Layer, Option, Ref, Schema } from \"effect\";\n\nimport { ChangedFile } from \"./diff.ts\";\nimport {\n  PriorReviewLookupFailure,\n  PriorReviews,\n  PublishedReview,\n  ReviewPublisher,\n} from \"./github.ts\";\nimport type { ReviewPublicationPlan } from \"./render.ts\";\nimport type { ReviewHeadComparison, ReviewState } from \"./review-state.ts\";\nimport {\n  MAX_CHANGED_FILES,\n  MAX_FILE_CHARS,\n  normalizeRepoRelativePath,\n  PullRequestMetadata,\n  PullRequestSource,\n  ReviewInputViolation,\n} from \"./source.ts\";\n\n// ---------------------------------------------------------------------------\n// Deterministic in-memory adapters for both ports: a fixture pull request\n// serving the PullRequestSource, and a collecting ReviewPublisher recording\n// every plan. Tests, dry runs, and live smokes run against these with no\n// network and no credentials.\n// ---------------------------------------------------------------------------\n\n/** One fixture file: its changeset entry plus optional head content. */\nexport class FixtureFile extends Schema.Class<FixtureFile>(\"@effect-agent/pr-review/FixtureFile\")({\n  file: ChangedFile,\n  headContent: Schema.optionalKey(Schema.String.check(Schema.isMaxLength(MAX_FILE_CHARS))),\n}) {}\n\n/** A complete in-memory pull request for tests, dry runs, and live smokes. */\nexport class FixturePullRequest extends Schema.Class<FixturePullRequest>(\n  \"@effect-agent/pr-review/FixturePullRequest\",\n)({\n  metadata: PullRequestMetadata,\n  files: Schema.Array(FixtureFile).check(Schema.isMaxLength(MAX_CHANGED_FILES)),\n}) {}\n\nconst requireChanged = (\n  fixture: FixturePullRequest,\n  path: string,\n): Effect.Effect<FixtureFile, ReviewInputViolation> => {\n  const entry = fixture.files.find((candidate) => candidate.file.path === path);\n  return entry === undefined\n    ? Effect.fail(\n        ReviewInputViolation.make({\n          input: path,\n          reason: \"Path is not part of this pull request's changeset.\",\n        }),\n      )\n    : Effect.succeed(entry);\n};\n\n/** Deterministic `PullRequestSource` over one fixture pull request. */\nexport const fixturePullRequestSourceLayer = (\n  fixture: FixturePullRequest,\n): Layer.Layer<PullRequestSource> =>\n  Layer.succeed(PullRequestSource)(\n    PullRequestSource.of({\n      metadata: Effect.succeed(fixture.metadata),\n      changedFiles: Effect.succeed(fixture.files.map((entry) => entry.file)),\n      anchorFiles: Effect.succeed(fixture.files.map((entry) => entry.file)),\n      readFile: (path) =>\n        Effect.gen(function* () {\n          const relative = yield* normalizeRepoRelativePath(path);\n          const entry = yield* requireChanged(fixture, relative);\n          if (entry.headContent === undefined) {\n            return yield* ReviewInputViolation.make({\n              input: relative,\n              reason: \"No head content is available for this file.\",\n            });\n          }\n          return entry.headContent;\n        }),\n    }),\n  );\n\n/** In-memory publisher: records every plan and mints a deterministic receipt. */\nexport const collectingReviewPublisherLayer = (\n  published: Ref.Ref<ReadonlyArray<ReviewPublicationPlan>>,\n): Layer.Layer<ReviewPublisher> =>\n  Layer.succeed(ReviewPublisher)(\n    ReviewPublisher.of({\n      publish: (plan) =>\n        Ref.update(published, (plans) => [...plans, plan]).pipe(\n          Effect.flatMap(() => Ref.get(published)),\n          Effect.map((plans) =>\n            PublishedReview.make({\n              reviewId: plans.length,\n              url: `memory://review/${plans.length}`,\n              event: plan.event,\n              inlineComments: plan.comments.length,\n            }),\n          ),\n        ),\n    }),\n  );\n\n/** Static `PriorReviews` service for tests: fixed history and comparisons. */\nexport const staticPriorReviews = (\n  fingerprint: Option.Option<string>,\n  options: {\n    readonly state?: Option.Option<ReviewState> | undefined;\n    readonly comparison?: ReviewHeadComparison | undefined;\n  } = {},\n): PriorReviews[\"Service\"] =>\n  PriorReviews.of({\n    latestFingerprint: Effect.succeed(fingerprint),\n    latestState: Effect.succeed(options.state ?? Option.none()),\n    compareHeads: () =>\n      options.comparison === undefined\n        ? Effect.fail(PriorReviewLookupFailure.make({ reason: \"no fixture comparison\" }))\n        : Effect.succeed(options.comparison),\n  });\n\n/** Layer form for consumers whose Effect explicitly requires `PriorReviews`. */\nexport const staticPriorReviewsLayer = (\n  fingerprint: Option.Option<string>,\n  options: {\n    readonly state?: Option.Option<ReviewState> | undefined;\n    readonly comparison?: ReviewHeadComparison | undefined;\n  } = {},\n): Layer.Layer<PriorReviews> =>\n  Layer.succeed(PriorReviews)(staticPriorReviews(fingerprint, options));\n"],"mappings":";;;;AAYA,MAAa,uBAAuB;AACpC,MAAa,uBAAuB;AACpC,MAAa,uBAAuB;;AAGpC,MAAa,sBAAsB;CAAE,aAAa;CAAI,cAAc;AAAG;AAEvE,MAAM,gBAAgB;CACpB,aAAa,EAAE,OAAO,oBAAoB,YAAY;CACtD,cAAc,EAAE,OAAO,oBAAoB,aAAa;AAC1D;AAEA,MAAa,oBACX,GAAG,UAC2C,CAC9C,GAAG,OACH;CAAE,MAAM;CAAU,QAAQ;CAAc,OAAO;AAAc,CAC/D;AAEA,MAAa,sBAAsB,SAA4D;CAC7F;EAAE,MAAM;EAAc,IAAI;CAAc;CACxC;EAAE,MAAM;EAAc,IAAI;EAAe,OAAO;CAAK;CACrD;EAAE,MAAM;EAAY,IAAI;CAAc;CACtC;EAAE,MAAM;EAAU,QAAQ;EAAQ,OAAO;CAAc;AACzD;;AAGA,MAAa,wBACX,MACA,WAEA,OAAO,IAAI,aAAa;CACtB,MAAM,QAAQ,OAAO,IAAI,KAAK,CAAC;CAC/B,MAAM,UAAU,OAAO,IAAI,KAA4B,CAAC,CAAC;CAoBzD,OAAO;EAAE,OAnBK,MAAM,KAClB,YACA,MACA,MAAM,OACJ,cAAc,eACd,cAAc,KAAK;GACjB,oBAAoB,OAAO,QAAQ,CAAC,CAAC;GACrC,aAAa,YACX,OAAO,OACL,OAAO,IAAI,aAAa;IACtB,OAAO,IAAI,OAAO,QAAQ,UAAU,QAAQ,CAAC;IAC7C,MAAM,aAAa,KAAK,UAAU,QAAQ,MAAM;IAChD,OAAO,IAAI,OAAO,UAAU,aAAa,CAAC,GAAG,UAAU,UAAU,CAAC;IAClE,OAAO,OAAO,aAAa,OAAO,UAAU,CAAC;GAC/C,CAAC,CACH;EACJ,CAAC,CACH,CAEW;EAAG,OAAO,IAAI,IAAI,KAAK;EAAG,SAAS,IAAI,IAAI,OAAO;CAAE;AACnE,CAAC;;;;;;AAOH,MAAa,4BAA4B,WAKvC,qBAAqB,sBAAsB,eAAe;CACxD,IAAI,WAAW,SAAA,QAA6B,GAC1C,OAAO,mBAAmB,KAAK,UAAU,OAAO,WAAW,UAAU,CAAC,CAAC,OAAO,MAAM,CAAC,CAAC;CAExF,IAAI,WAAW,SAAA,QAA6B,GAC1C,OAAO,iBAAiB;EACtB,MAAM;EACN,IAAI;EACJ,MAAM;EACN,QAAQ,EAAE,MAAM,OAAO,SAAS;EAChC,kBAAkB;CACpB,CAAC;CAEH,IAAI,WAAW,SAAA,QAA6B,GAC1C,OAAO,iBAAiB;EACtB,MAAM;EACN,IAAI;EACJ,MAAM;EACN,QAAQ,EAAE,MAAM,OAAO,SAAS;EAChC,kBAAkB;CACpB,CAAC;CAEH,OAAO,iBAAiB;EACtB,MAAM;EACN,IAAI;EACJ,MAAM;EACN,QAAQ,EAAE,OAAO,MAAM;EACvB,kBAAkB;CACpB,CAAC;AACH,CAAC;;;AC5FH,MAAa,wBAAwB;;AAGrC,MAAa,qBAAqB,WAA2B,YAAY;;AAGzE,MAAa,0BAA0B,WAA2B,eAAe;;;;;;;AAcjF,MAAa,qCAAqC,WAG5C;CACJ,MAAM,kBAAkB,kBAAkB,OAAO,UAAU,EAAE,EAAE,UAAU,WAAW;CACpF,OAAO,qBAAqB,kCAAkC,eAAe;EAC3E,IAAI,WAAW,SAAS,eAAe,GACrC,OAAO,mBAAmB,KAAK,UAAU,OAAO,WAAW,UAAU,CAAC,CAAC,OAAO,MAAM,CAAC,CAAC;EAExF,IAAI,WAAW,SAAA,SAA8B,GAC3C,OAAO,iBACL,GAAG,OAAO,UAAU,KACjB,UAAsC;GACrC,MAAM;GACN,IAAI,kBAAkB,KAAK,MAAM;GACjC,MAAM;GACN,QAAQ;IAAE,QAAQ,KAAK;IAAQ,OAAO,KAAK;GAAM;GACjD,kBAAkB;EACpB,EACF,CACF;EAEF,OAAO,iBAAiB;GACtB,MAAM;GACN,IAAI;GACJ,MAAM;GACN,QAAQ,EAAE,OAAO,MAAM;GACvB,kBAAkB;EACpB,CAAC;CACH,CAAC;AACH;;;;;;;;;;AA8BA,MAAa,gCAAgC,YAC3C,OAAO,IAAI,aAAa;CACtB,MAAM,QAAQ,OAAO,IAAI,KAAK,CAAC;CAC/B,MAAM,UAAU,OAAO,IAAI,KAA4B,CAAC,CAAC;CACzD,MAAM,UAAU,eAA8E;EAC5F,MAAM,SAAS,QAAQ,MAAM,cAAc,WAAW,SAAS,UAAU,MAAM,CAAC;EAChF,IAAI,WAAW,KAAA,GAAW,OAAO,KAAA;EACjC,QAAQ,OAAO,QAAQ,MAAvB;GACE,KAAK,kBACH,OAAO,iBACL,GAAG,MAAM,KACP,EAAE,QAAQ,OAAO,QAAQ,cAAc,IACtC,GAAG,WAAuC;IACzC,MAAM;IACN,IAAI,WAAW,OAAO,OAAO,GAAG,QAAQ;IACxC,MAAM;IACN,QAAQ,EAAE,MAAM,OAAO,SAAS;IAChC,kBAAkB;GACpB,EACF,CACF;GAEF,KAAK;GACL,KAAK;IACH,IAAI,WAAW,SAAS,uBAAuB,OAAO,MAAM,CAAC,GAC3D,OAAO,mBACL,OAAO,QAAQ,SAAS,aACpB,KAAK,UAAU,OAAO,WAAW,gBAAgB,CAAC,CAAC,OAAO,QAAQ,MAAM,CAAC,IACzE,0CACN;IAEF,OAAO,iBAAiB;KACtB,MAAM;KACN,IAAI,uBAAuB,OAAO,MAAM;KACxC,MAAM;KACN,QAAQ,EAAE,MAAM,OAAO,SAAS;KAChC,kBAAkB;IACpB,CAAC;EAEL;CACF;CA0BA,OAAO;EAAE,OAzBK,MAAM,KAClB,YACA,mCACA,MAAM,OACJ,cAAc,eACd,cAAc,KAAK;GACjB,oBAAoB,OAAO,QAAQ,CAAC,CAAC;GACrC,aAAa,YACX,OAAO,OACL,OAAO,IAAI,aAAa;IACtB,OAAO,IAAI,OAAO,QAAQ,UAAU,QAAQ,CAAC;IAC7C,MAAM,aAAa,KAAK,UAAU,QAAQ,MAAM;IAChD,OAAO,IAAI,OAAO,UAAU,aAAa,CAAC,GAAG,UAAU,UAAU,CAAC;IAClE,MAAM,QAAQ,OAAO,UAAU;IAC/B,IAAI,UAAU,KAAA,GACZ,OAAO,OAAO,OAAO,oBACnB,IAAI,MAAM,gDAAgD,CAC5D;IAEF,OAAO,OAAO,aAAa,KAAK;GAClC,CAAC,CACH;EACJ,CAAC,CACH,CAEW;EAAG,OAAO,IAAI,IAAI,KAAK;EAAG,SAAS,IAAI,IAAI,OAAO;CAAE;AACnE,CAAC;;;;ACtIH,IAAa,cAAb,cAAiC,OAAO,MAAmB,qCAAqC,CAAC,CAAC;CAChG,MAAM;CACN,aAAa,OAAO,YAAY,OAAO,OAAO,MAAM,OAAO,YAAY,cAAc,CAAC,CAAC;AACzF,CAAC,CAAC,CAAC,CAAC;;AAGJ,IAAa,qBAAb,cAAwC,OAAO,MAC7C,4CACF,CAAC,CAAC;CACA,UAAU;CACV,OAAO,OAAO,MAAM,WAAW,CAAC,CAAC,MAAM,OAAO,YAAA,GAA6B,CAAC;AAC9E,CAAC,CAAC,CAAC,CAAC;AAEJ,MAAM,kBACJ,SACA,SACqD;CACrD,MAAM,QAAQ,QAAQ,MAAM,MAAM,cAAc,UAAU,KAAK,SAAS,IAAI;CAC5E,OAAO,UAAU,KAAA,IACb,OAAO,KACL,qBAAqB,KAAK;EACxB,OAAO;EACP,QAAQ;CACV,CAAC,CACH,IACA,OAAO,QAAQ,KAAK;AAC1B;;AAGA,MAAa,iCACX,YAEA,MAAM,QAAQ,iBAAiB,CAAC,CAC9B,kBAAkB,GAAG;CACnB,UAAU,OAAO,QAAQ,QAAQ,QAAQ;CACzC,cAAc,OAAO,QAAQ,QAAQ,MAAM,KAAK,UAAU,MAAM,IAAI,CAAC;CACrE,aAAa,OAAO,QAAQ,QAAQ,MAAM,KAAK,UAAU,MAAM,IAAI,CAAC;CACpE,WAAW,SACT,OAAO,IAAI,aAAa;EACtB,MAAM,WAAW,OAAO,0BAA0B,IAAI;EACtD,MAAM,QAAQ,OAAO,eAAe,SAAS,QAAQ;EACrD,IAAI,MAAM,gBAAgB,KAAA,GACxB,OAAO,OAAO,qBAAqB,KAAK;GACtC,OAAO;GACP,QAAQ;EACV,CAAC;EAEH,OAAO,MAAM;CACf,CAAC;AACL,CAAC,CACH;;AAGF,MAAa,kCACX,cAEA,MAAM,QAAQ,eAAe,CAAC,CAC5B,gBAAgB,GAAG,EACjB,UAAU,SACR,IAAI,OAAO,YAAY,UAAU,CAAC,GAAG,OAAO,IAAI,CAAC,CAAC,CAAC,KACjD,OAAO,cAAc,IAAI,IAAI,SAAS,CAAC,GACvC,OAAO,KAAK,UACV,gBAAgB,KAAK;CACnB,UAAU,MAAM;CAChB,KAAK,mBAAmB,MAAM;CAC9B,OAAO,KAAK;CACZ,gBAAgB,KAAK,SAAS;AAChC,CAAC,CACH,CACF,EACJ,CAAC,CACH;;AAGF,MAAa,sBACX,aACA,UAGI,CAAC,MAEL,aAAa,GAAG;CACd,mBAAmB,OAAO,QAAQ,WAAW;CAC7C,aAAa,OAAO,QAAQ,QAAQ,SAAS,OAAO,KAAK,CAAC;CAC1D,oBACE,QAAQ,eAAe,KAAA,IACnB,OAAO,KAAK,yBAAyB,KAAK,EAAE,QAAQ,wBAAwB,CAAC,CAAC,IAC9E,OAAO,QAAQ,QAAQ,UAAU;AACzC,CAAC;;AAGH,MAAa,2BACX,aACA,UAGI,CAAC,MAEL,MAAM,QAAQ,YAAY,CAAC,CAAC,mBAAmB,aAAa,OAAO,CAAC"}
+{"version":3,"file":"testing.mjs","names":[],"sources":["../src/internal/scripted.ts","../src/internal/fan-out-scripted.ts","../src/internal/fixtures.ts"],"sourcesContent":["import { Effect, Layer, Ref, Schema, Stream } from \"effect\";\nimport { LanguageModel, Model, type Response } from \"effect/unstable/ai\";\n\nimport { CodeReview } from \"./review-agent.ts\";\n\n// ---------------------------------------------------------------------------\n// Deterministic offline model for the flat reviewer: a prompt-aware scripted\n// model that walks the real tool surface — list, diff, read — then returns\n// the scripted review as its terminal JSON. Decisions key on committed\n// history in the prompt, never on call order, so replays stay honest.\n// ---------------------------------------------------------------------------\n\nexport const OFFLINE_LIST_CALL_ID = \"list-1\";\nexport const OFFLINE_DIFF_CALL_ID = \"diff-1\";\nexport const OFFLINE_READ_CALL_ID = \"read-1\";\n\n/** Usage attached to EVERY scripted model turn; tests pin exact aggregates. */\nexport const SCRIPTED_TURN_USAGE = { inputTokens: 64, outputTokens: 48 } as const;\n\nconst scriptedUsage = {\n  inputTokens: { total: SCRIPTED_TURN_USAGE.inputTokens },\n  outputTokens: { total: SCRIPTED_TURN_USAGE.outputTokens },\n};\n\nexport const scriptedToolTurn = (\n  ...calls: ReadonlyArray<Response.StreamPartEncoded>\n): ReadonlyArray<Response.StreamPartEncoded> => [\n  ...calls,\n  { type: \"finish\", reason: \"tool-calls\", usage: scriptedUsage },\n];\n\nexport const scriptedFinalParts = (text: string): ReadonlyArray<Response.StreamPartEncoded> => [\n  { type: \"text-start\", id: \"code-review\" },\n  { type: \"text-delta\", id: \"code-review\", delta: text },\n  { type: \"text-end\", id: \"code-review\" },\n  { type: \"finish\", reason: \"stop\", usage: scriptedUsage },\n];\n\n/** A prompt-keyed scripted LanguageModel with call and prompt observability. */\nexport const makePromptKeyedModel = (\n  name: string,\n  decide: (promptJson: string) => ReadonlyArray<Response.StreamPartEncoded>,\n) =>\n  Effect.gen(function* () {\n    const calls = yield* Ref.make(0);\n    const prompts = yield* Ref.make<ReadonlyArray<string>>([]);\n    const model = Model.make(\n      \"scripted\",\n      name,\n      Layer.effect(\n        LanguageModel.LanguageModel,\n        LanguageModel.make({\n          generateText: () => Effect.succeed([]),\n          streamText: (request) =>\n            Stream.unwrap(\n              Effect.gen(function* () {\n                yield* Ref.update(calls, (value) => value + 1);\n                const promptJson = JSON.stringify(request.prompt);\n                yield* Ref.update(prompts, (previous) => [...previous, promptJson]);\n                return Stream.fromIterable(decide(promptJson));\n              }),\n            ),\n        }),\n      ),\n    );\n    return { model, calls: Ref.get(calls), prompts: Ref.get(prompts) };\n  });\n\n/**\n * Build the offline scripted reviewer model. Turn 1 lists the changeset,\n * Turn 2 reads one file diff, Turn 3 reads head context, Turn 4 returns the\n * scripted review JSON.\n */\nexport const makeOfflineReviewerModel = (script: {\n  readonly diffPath: string;\n  readonly readPath: string;\n  readonly review: CodeReview;\n}) =>\n  makePromptKeyedModel(\"pr-review-offline\", (promptJson) => {\n    if (promptJson.includes(OFFLINE_READ_CALL_ID)) {\n      return scriptedFinalParts(JSON.stringify(Schema.encodeSync(CodeReview)(script.review)));\n    }\n    if (promptJson.includes(OFFLINE_DIFF_CALL_ID)) {\n      return scriptedToolTurn({\n        type: \"tool-call\",\n        id: OFFLINE_READ_CALL_ID,\n        name: \"read_file\",\n        params: { path: script.readPath },\n        providerExecuted: false,\n      });\n    }\n    if (promptJson.includes(OFFLINE_LIST_CALL_ID)) {\n      return scriptedToolTurn({\n        type: \"tool-call\",\n        id: OFFLINE_DIFF_CALL_ID,\n        name: \"read_file_diff\",\n        params: { path: script.diffPath },\n        providerExecuted: false,\n      });\n    }\n    return scriptedToolTurn({\n      type: \"tool-call\",\n      id: OFFLINE_LIST_CALL_ID,\n      name: \"list_changed_files\",\n      params: { scope: \"all\" },\n      providerExecuted: false,\n    });\n  });\n","import { Effect, Layer, Ref, Schema, Stream } from \"effect\";\nimport { LanguageModel, Model, type Response } from \"effect/unstable/ai\";\n\nimport { FileReviewReport } from \"./fan-out.ts\";\nimport { CodeReview } from \"./review-agent.ts\";\nimport { makePromptKeyedModel, scriptedFinalParts, scriptedToolTurn } from \"./scripted.ts\";\n\n// ---------------------------------------------------------------------------\n// Deterministic offline models for the fan-out reviewer: prompt-keyed\n// scripted models for BOTH the coordinator and the file-reviewer children.\n// Both key every decision on committed history in the prompt (tool-call ids\n// and briefed unit ids), never on call order, so concurrent children and\n// replays stay honest.\n// ---------------------------------------------------------------------------\n\nexport const OFFLINE_UNITS_CALL_ID = \"units-1\";\n\n/** The delegation Tool Call id the scripted coordinator uses for one unit. */\nexport const offlineUnitCallId = (unitId: string): string => `delegate-${unitId}`;\n\n/** The diff Tool Call id the scripted child uses for one unit. */\nexport const offlineChildDiffCallId = (unitId: string): string => `fanout-diff-${unitId}`;\n\n/** One scripted delegation the offline coordinator declares. */\nexport interface OfflineUnitCall {\n  readonly unitId: string;\n  readonly paths: ReadonlyArray<string>;\n}\n\n/**\n * Build the offline scripted coordinator model. Turn 1 lists the review\n * units, Turn 2 declares one delegation Tool Call per scripted unit in one\n * batch, Turn 3 returns the scripted merged review JSON. Decisions key on\n * tool-call ids already committed to the prompt.\n */\nexport const makeOfflineFanOutCoordinatorModel = (script: {\n  readonly unitCalls: ReadonlyArray<OfflineUnitCall>;\n  readonly review: CodeReview;\n}) => {\n  const firstUnitCallId = offlineUnitCallId(script.unitCalls[0]?.unitId ?? \"unit-none\");\n  return makePromptKeyedModel(\"pr-fanout-coordinator-offline\", (promptJson) => {\n    if (promptJson.includes(firstUnitCallId)) {\n      return scriptedFinalParts(JSON.stringify(Schema.encodeSync(CodeReview)(script.review)));\n    }\n    if (promptJson.includes(OFFLINE_UNITS_CALL_ID)) {\n      return scriptedToolTurn(\n        ...script.unitCalls.map(\n          (unit): Response.StreamPartEncoded => ({\n            type: \"tool-call\",\n            id: offlineUnitCallId(unit.unitId),\n            name: \"delegate_file_review\",\n            params: { unitId: unit.unitId, paths: unit.paths },\n            providerExecuted: false,\n          }),\n        ),\n      );\n    }\n    return scriptedToolTurn({\n      type: \"tool-call\",\n      id: OFFLINE_UNITS_CALL_ID,\n      name: \"list_review_units\",\n      params: { scope: \"all\" },\n      providerExecuted: false,\n    });\n  });\n};\n\n/** How one scripted child behaves for its briefed unit. */\nexport type OfflineUnitOutcome =\n  /** Read one diff, then return the scripted report. */\n  | { readonly _tag: \"findings\"; readonly report: FileReviewReport }\n  /** Read one diff, then return non-JSON — the child fails typed (AgentOutputError). */\n  | { readonly _tag: \"malformed-output\" }\n  /**\n   * Declare more Tool Calls than the child's AgentPolicy allows in one turn —\n   * none executes and the child fails typed (AgentPolicyError \"tool-calls\",\n   * the reviewer's deliberate `onExhaustion: \"fail\"` pin).\n   */\n  | { readonly _tag: \"budget-runaway\"; readonly declaredCalls: number };\n\nexport interface OfflineUnitScript {\n  readonly unitId: string;\n  /** The one file the scripted child reads the diff of. */\n  readonly diffPath: string;\n  readonly outcome: OfflineUnitOutcome;\n}\n\n/**\n * Build the offline scripted file-reviewer model shared by every delegated\n * child. Each child Run builds the Model Layer inside its own scope; the\n * script entry is selected by the briefed unitId present in the child's OWN\n * prompt, and the turn is selected by whether that unit's diff Tool Call id\n * is already committed there — content-keyed on both axes, so concurrent\n * children never interfere. First-turn child prompts are recorded for\n * context-isolation assertions.\n */\nexport const makeOfflineFileReviewerModel = (scripts: ReadonlyArray<OfflineUnitScript>) =>\n  Effect.gen(function* () {\n    const calls = yield* Ref.make(0);\n    const prompts = yield* Ref.make<ReadonlyArray<string>>([]);\n    const decide = (promptJson: string): ReadonlyArray<Response.StreamPartEncoded> | undefined => {\n      const script = scripts.find((candidate) => promptJson.includes(candidate.unitId));\n      if (script === undefined) return undefined;\n      switch (script.outcome._tag) {\n        case \"budget-runaway\": {\n          return scriptedToolTurn(\n            ...Array.from(\n              { length: script.outcome.declaredCalls },\n              (_, index): Response.StreamPartEncoded => ({\n                type: \"tool-call\",\n                id: `runaway-${script.unitId}-${index + 1}`,\n                name: \"read_file_diff\",\n                params: { path: script.diffPath },\n                providerExecuted: false,\n              }),\n            ),\n          );\n        }\n        case \"malformed-output\":\n        case \"findings\": {\n          if (promptJson.includes(offlineChildDiffCallId(script.unitId))) {\n            return scriptedFinalParts(\n              script.outcome._tag === \"findings\"\n                ? JSON.stringify(Schema.encodeSync(FileReviewReport)(script.outcome.report))\n                : \"this is not the JSON you are looking for\",\n            );\n          }\n          return scriptedToolTurn({\n            type: \"tool-call\",\n            id: offlineChildDiffCallId(script.unitId),\n            name: \"read_file_diff\",\n            params: { path: script.diffPath },\n            providerExecuted: false,\n          });\n        }\n      }\n    };\n    const model = Model.make(\n      \"scripted\",\n      \"pr-fanout-file-reviewer-offline\",\n      Layer.effect(\n        LanguageModel.LanguageModel,\n        LanguageModel.make({\n          generateText: () => Effect.succeed([]),\n          streamText: (request) =>\n            Stream.unwrap(\n              Effect.gen(function* () {\n                yield* Ref.update(calls, (value) => value + 1);\n                const promptJson = JSON.stringify(request.prompt);\n                yield* Ref.update(prompts, (previous) => [...previous, promptJson]);\n                const parts = decide(promptJson);\n                if (parts === undefined) {\n                  return yield* Effect.die(\n                    new Error(\"The child prompt names no scripted review unit\"),\n                  );\n                }\n                return Stream.fromIterable(parts);\n              }),\n            ),\n        }),\n      ),\n    );\n    return { model, calls: Ref.get(calls), prompts: Ref.get(prompts) };\n  });\n","import { Effect, Layer, Option, Ref, Schema } from \"effect\";\n\nimport { ChangedFile } from \"./diff.ts\";\nimport {\n  PriorReviewLookupFailure,\n  PriorReviews,\n  PublishedReview,\n  ReviewPublisher,\n} from \"./github.ts\";\nimport type { ReviewPublicationPlan } from \"./render.ts\";\nimport type { ReviewHeadComparison, ReviewState } from \"./review-state.ts\";\nimport {\n  MAX_CHANGED_FILES,\n  MAX_FILE_CHARS,\n  normalizeRepoRelativePath,\n  PullRequestMetadata,\n  PullRequestSource,\n  ReviewInputViolation,\n} from \"./source.ts\";\n\n// ---------------------------------------------------------------------------\n// Deterministic in-memory adapters for both ports: a fixture pull request\n// serving the PullRequestSource, and a collecting ReviewPublisher recording\n// every plan. Tests, dry runs, and live smokes run against these with no\n// network and no credentials.\n// ---------------------------------------------------------------------------\n\n/** One fixture file: its changeset entry plus optional head content. */\nexport class FixtureFile extends Schema.Class<FixtureFile>(\"@effect-agent/pr-review/FixtureFile\")({\n  file: ChangedFile,\n  headContent: Schema.optionalKey(Schema.String.check(Schema.isMaxLength(MAX_FILE_CHARS))),\n}) {}\n\n/** A complete in-memory pull request for tests, dry runs, and live smokes. */\nexport class FixturePullRequest extends Schema.Class<FixturePullRequest>(\n  \"@effect-agent/pr-review/FixturePullRequest\",\n)({\n  metadata: PullRequestMetadata,\n  files: Schema.Array(FixtureFile).check(Schema.isMaxLength(MAX_CHANGED_FILES)),\n}) {}\n\nconst requireChanged = (\n  fixture: FixturePullRequest,\n  path: string,\n): Effect.Effect<FixtureFile, ReviewInputViolation> => {\n  const entry = fixture.files.find((candidate) => candidate.file.path === path);\n  return entry === undefined\n    ? Effect.fail(\n        ReviewInputViolation.make({\n          input: path,\n          reason: \"Path is not part of this pull request's changeset.\",\n        }),\n      )\n    : Effect.succeed(entry);\n};\n\n/** Deterministic `PullRequestSource` over one fixture pull request. */\nexport const fixturePullRequestSourceLayer = (\n  fixture: FixturePullRequest,\n): Layer.Layer<PullRequestSource> =>\n  Layer.succeed(PullRequestSource)(\n    PullRequestSource.of({\n      metadata: Effect.succeed(fixture.metadata),\n      changedFiles: Effect.succeed(fixture.files.map((entry) => entry.file)),\n      anchorFiles: Effect.succeed(fixture.files.map((entry) => entry.file)),\n      readFile: (path) =>\n        Effect.gen(function* () {\n          const relative = yield* normalizeRepoRelativePath(path);\n          const entry = yield* requireChanged(fixture, relative);\n          if (entry.headContent === undefined) {\n            return yield* ReviewInputViolation.make({\n              input: relative,\n              reason: \"No head content is available for this file.\",\n            });\n          }\n          return entry.headContent;\n        }),\n    }),\n  );\n\n/** In-memory publisher: records every plan and mints a deterministic receipt. */\nexport const collectingReviewPublisherLayer = (\n  published: Ref.Ref<ReadonlyArray<ReviewPublicationPlan>>,\n): Layer.Layer<ReviewPublisher> =>\n  Layer.succeed(ReviewPublisher)(\n    ReviewPublisher.of({\n      publish: (plan) =>\n        Ref.update(published, (plans) => [...plans, plan]).pipe(\n          Effect.flatMap(() => Ref.get(published)),\n          Effect.map((plans) =>\n            PublishedReview.make({\n              reviewId: plans.length,\n              url: `memory://review/${plans.length}`,\n              event: plan.event,\n              inlineComments: plan.comments.length,\n            }),\n          ),\n        ),\n    }),\n  );\n\n/** Static `PriorReviews` service for tests: fixed history and comparisons. */\nexport const staticPriorReviews = (\n  fingerprint: Option.Option<string>,\n  options: {\n    readonly state?: Option.Option<ReviewState> | undefined;\n    readonly comparison?: ReviewHeadComparison | undefined;\n  } = {},\n): PriorReviews[\"Service\"] =>\n  PriorReviews.of({\n    latestFingerprint: Effect.succeed(fingerprint),\n    latestState: Effect.succeed(options.state ?? Option.none()),\n    compareHeads: () =>\n      options.comparison === undefined\n        ? Effect.fail(PriorReviewLookupFailure.make({ reason: \"no fixture comparison\" }))\n        : Effect.succeed(options.comparison),\n  });\n\n/** Layer form for consumers whose Effect explicitly requires `PriorReviews`. */\nexport const staticPriorReviewsLayer = (\n  fingerprint: Option.Option<string>,\n  options: {\n    readonly state?: Option.Option<ReviewState> | undefined;\n    readonly comparison?: ReviewHeadComparison | undefined;\n  } = {},\n): Layer.Layer<PriorReviews> =>\n  Layer.succeed(PriorReviews)(staticPriorReviews(fingerprint, options));\n"],"mappings":";;;;AAYA,MAAa,uBAAuB;AACpC,MAAa,uBAAuB;AACpC,MAAa,uBAAuB;;AAGpC,MAAa,sBAAsB;CAAE,aAAa;CAAI,cAAc;AAAG;AAEvE,MAAM,gBAAgB;CACpB,aAAa,EAAE,OAAO,oBAAoB,YAAY;CACtD,cAAc,EAAE,OAAO,oBAAoB,aAAa;AAC1D;AAEA,MAAa,oBACX,GAAG,UAC2C,CAC9C,GAAG,OACH;CAAE,MAAM;CAAU,QAAQ;CAAc,OAAO;AAAc,CAC/D;AAEA,MAAa,sBAAsB,SAA4D;CAC7F;EAAE,MAAM;EAAc,IAAI;CAAc;CACxC;EAAE,MAAM;EAAc,IAAI;EAAe,OAAO;CAAK;CACrD;EAAE,MAAM;EAAY,IAAI;CAAc;CACtC;EAAE,MAAM;EAAU,QAAQ;EAAQ,OAAO;CAAc;AACzD;;AAGA,MAAa,wBACX,MACA,WAEA,OAAO,IAAI,aAAa;CACtB,MAAM,QAAQ,OAAO,IAAI,KAAK,CAAC;CAC/B,MAAM,UAAU,OAAO,IAAI,KAA4B,CAAC,CAAC;CAoBzD,OAAO;EAAE,OAnBK,MAAM,KAClB,YACA,MACA,MAAM,OACJ,cAAc,eACd,cAAc,KAAK;GACjB,oBAAoB,OAAO,QAAQ,CAAC,CAAC;GACrC,aAAa,YACX,OAAO,OACL,OAAO,IAAI,aAAa;IACtB,OAAO,IAAI,OAAO,QAAQ,UAAU,QAAQ,CAAC;IAC7C,MAAM,aAAa,KAAK,UAAU,QAAQ,MAAM;IAChD,OAAO,IAAI,OAAO,UAAU,aAAa,CAAC,GAAG,UAAU,UAAU,CAAC;IAClE,OAAO,OAAO,aAAa,OAAO,UAAU,CAAC;GAC/C,CAAC,CACH;EACJ,CAAC,CACH,CAEW;EAAG,OAAO,IAAI,IAAI,KAAK;EAAG,SAAS,IAAI,IAAI,OAAO;CAAE;AACnE,CAAC;;;;;;AAOH,MAAa,4BAA4B,WAKvC,qBAAqB,sBAAsB,eAAe;CACxD,IAAI,WAAW,SAAA,QAA6B,GAC1C,OAAO,mBAAmB,KAAK,UAAU,OAAO,WAAW,UAAU,CAAC,CAAC,OAAO,MAAM,CAAC,CAAC;CAExF,IAAI,WAAW,SAAA,QAA6B,GAC1C,OAAO,iBAAiB;EACtB,MAAM;EACN,IAAI;EACJ,MAAM;EACN,QAAQ,EAAE,MAAM,OAAO,SAAS;EAChC,kBAAkB;CACpB,CAAC;CAEH,IAAI,WAAW,SAAA,QAA6B,GAC1C,OAAO,iBAAiB;EACtB,MAAM;EACN,IAAI;EACJ,MAAM;EACN,QAAQ,EAAE,MAAM,OAAO,SAAS;EAChC,kBAAkB;CACpB,CAAC;CAEH,OAAO,iBAAiB;EACtB,MAAM;EACN,IAAI;EACJ,MAAM;EACN,QAAQ,EAAE,OAAO,MAAM;EACvB,kBAAkB;CACpB,CAAC;AACH,CAAC;;;AC5FH,MAAa,wBAAwB;;AAGrC,MAAa,qBAAqB,WAA2B,YAAY;;AAGzE,MAAa,0BAA0B,WAA2B,eAAe;;;;;;;AAcjF,MAAa,qCAAqC,WAG5C;CACJ,MAAM,kBAAkB,kBAAkB,OAAO,UAAU,EAAE,EAAE,UAAU,WAAW;CACpF,OAAO,qBAAqB,kCAAkC,eAAe;EAC3E,IAAI,WAAW,SAAS,eAAe,GACrC,OAAO,mBAAmB,KAAK,UAAU,OAAO,WAAW,UAAU,CAAC,CAAC,OAAO,MAAM,CAAC,CAAC;EAExF,IAAI,WAAW,SAAA,SAA8B,GAC3C,OAAO,iBACL,GAAG,OAAO,UAAU,KACjB,UAAsC;GACrC,MAAM;GACN,IAAI,kBAAkB,KAAK,MAAM;GACjC,MAAM;GACN,QAAQ;IAAE,QAAQ,KAAK;IAAQ,OAAO,KAAK;GAAM;GACjD,kBAAkB;EACpB,EACF,CACF;EAEF,OAAO,iBAAiB;GACtB,MAAM;GACN,IAAI;GACJ,MAAM;GACN,QAAQ,EAAE,OAAO,MAAM;GACvB,kBAAkB;EACpB,CAAC;CACH,CAAC;AACH;;;;;;;;;;AA+BA,MAAa,gCAAgC,YAC3C,OAAO,IAAI,aAAa;CACtB,MAAM,QAAQ,OAAO,IAAI,KAAK,CAAC;CAC/B,MAAM,UAAU,OAAO,IAAI,KAA4B,CAAC,CAAC;CACzD,MAAM,UAAU,eAA8E;EAC5F,MAAM,SAAS,QAAQ,MAAM,cAAc,WAAW,SAAS,UAAU,MAAM,CAAC;EAChF,IAAI,WAAW,KAAA,GAAW,OAAO,KAAA;EACjC,QAAQ,OAAO,QAAQ,MAAvB;GACE,KAAK,kBACH,OAAO,iBACL,GAAG,MAAM,KACP,EAAE,QAAQ,OAAO,QAAQ,cAAc,IACtC,GAAG,WAAuC;IACzC,MAAM;IACN,IAAI,WAAW,OAAO,OAAO,GAAG,QAAQ;IACxC,MAAM;IACN,QAAQ,EAAE,MAAM,OAAO,SAAS;IAChC,kBAAkB;GACpB,EACF,CACF;GAEF,KAAK;GACL,KAAK;IACH,IAAI,WAAW,SAAS,uBAAuB,OAAO,MAAM,CAAC,GAC3D,OAAO,mBACL,OAAO,QAAQ,SAAS,aACpB,KAAK,UAAU,OAAO,WAAW,gBAAgB,CAAC,CAAC,OAAO,QAAQ,MAAM,CAAC,IACzE,0CACN;IAEF,OAAO,iBAAiB;KACtB,MAAM;KACN,IAAI,uBAAuB,OAAO,MAAM;KACxC,MAAM;KACN,QAAQ,EAAE,MAAM,OAAO,SAAS;KAChC,kBAAkB;IACpB,CAAC;EAEL;CACF;CA0BA,OAAO;EAAE,OAzBK,MAAM,KAClB,YACA,mCACA,MAAM,OACJ,cAAc,eACd,cAAc,KAAK;GACjB,oBAAoB,OAAO,QAAQ,CAAC,CAAC;GACrC,aAAa,YACX,OAAO,OACL,OAAO,IAAI,aAAa;IACtB,OAAO,IAAI,OAAO,QAAQ,UAAU,QAAQ,CAAC;IAC7C,MAAM,aAAa,KAAK,UAAU,QAAQ,MAAM;IAChD,OAAO,IAAI,OAAO,UAAU,aAAa,CAAC,GAAG,UAAU,UAAU,CAAC;IAClE,MAAM,QAAQ,OAAO,UAAU;IAC/B,IAAI,UAAU,KAAA,GACZ,OAAO,OAAO,OAAO,oBACnB,IAAI,MAAM,gDAAgD,CAC5D;IAEF,OAAO,OAAO,aAAa,KAAK;GAClC,CAAC,CACH;EACJ,CAAC,CACH,CAEW;EAAG,OAAO,IAAI,IAAI,KAAK;EAAG,SAAS,IAAI,IAAI,OAAO;CAAE;AACnE,CAAC;;;;ACvIH,IAAa,cAAb,cAAiC,OAAO,MAAmB,qCAAqC,CAAC,CAAC;CAChG,MAAM;CACN,aAAa,OAAO,YAAY,OAAO,OAAO,MAAM,OAAO,YAAY,cAAc,CAAC,CAAC;AACzF,CAAC,CAAC,CAAC,CAAC;;AAGJ,IAAa,qBAAb,cAAwC,OAAO,MAC7C,4CACF,CAAC,CAAC;CACA,UAAU;CACV,OAAO,OAAO,MAAM,WAAW,CAAC,CAAC,MAAM,OAAO,YAAA,GAA6B,CAAC;AAC9E,CAAC,CAAC,CAAC,CAAC;AAEJ,MAAM,kBACJ,SACA,SACqD;CACrD,MAAM,QAAQ,QAAQ,MAAM,MAAM,cAAc,UAAU,KAAK,SAAS,IAAI;CAC5E,OAAO,UAAU,KAAA,IACb,OAAO,KACL,qBAAqB,KAAK;EACxB,OAAO;EACP,QAAQ;CACV,CAAC,CACH,IACA,OAAO,QAAQ,KAAK;AAC1B;;AAGA,MAAa,iCACX,YAEA,MAAM,QAAQ,iBAAiB,CAAC,CAC9B,kBAAkB,GAAG;CACnB,UAAU,OAAO,QAAQ,QAAQ,QAAQ;CACzC,cAAc,OAAO,QAAQ,QAAQ,MAAM,KAAK,UAAU,MAAM,IAAI,CAAC;CACrE,aAAa,OAAO,QAAQ,QAAQ,MAAM,KAAK,UAAU,MAAM,IAAI,CAAC;CACpE,WAAW,SACT,OAAO,IAAI,aAAa;EACtB,MAAM,WAAW,OAAO,0BAA0B,IAAI;EACtD,MAAM,QAAQ,OAAO,eAAe,SAAS,QAAQ;EACrD,IAAI,MAAM,gBAAgB,KAAA,GACxB,OAAO,OAAO,qBAAqB,KAAK;GACtC,OAAO;GACP,QAAQ;EACV,CAAC;EAEH,OAAO,MAAM;CACf,CAAC;AACL,CAAC,CACH;;AAGF,MAAa,kCACX,cAEA,MAAM,QAAQ,eAAe,CAAC,CAC5B,gBAAgB,GAAG,EACjB,UAAU,SACR,IAAI,OAAO,YAAY,UAAU,CAAC,GAAG,OAAO,IAAI,CAAC,CAAC,CAAC,KACjD,OAAO,cAAc,IAAI,IAAI,SAAS,CAAC,GACvC,OAAO,KAAK,UACV,gBAAgB,KAAK;CACnB,UAAU,MAAM;CAChB,KAAK,mBAAmB,MAAM;CAC9B,OAAO,KAAK;CACZ,gBAAgB,KAAK,SAAS;AAChC,CAAC,CACH,CACF,EACJ,CAAC,CACH;;AAGF,MAAa,sBACX,aACA,UAGI,CAAC,MAEL,aAAa,GAAG;CACd,mBAAmB,OAAO,QAAQ,WAAW;CAC7C,aAAa,OAAO,QAAQ,QAAQ,SAAS,OAAO,KAAK,CAAC;CAC1D,oBACE,QAAQ,eAAe,KAAA,IACnB,OAAO,KAAK,yBAAyB,KAAK,EAAE,QAAQ,wBAAwB,CAAC,CAAC,IAC9E,OAAO,QAAQ,QAAQ,UAAU;AACzC,CAAC;;AAGH,MAAa,2BACX,aACA,UAGI,CAAC,MAEL,MAAM,QAAQ,YAAY,CAAC,CAAC,mBAAmB,aAAa,OAAO,CAAC"}

package/package.json

@@ -1,18 +1,6 @@
 {
   "name": "@effect-agent/pr-review",
-  "version": "0.0.1-beta.0",
-  "description": "Bounded, fail-closed GitHub pull-request reviewer built on the effect-agent public surface: schema-first review contracts, read-only tools, GitHub adapters, a configuration factory, and CLI + GitHub Actions entrypoints.",
-  "license": "MIT",
-  "repository": {
-    "type": "git",
-    "url": "git+https://github.com/danieljvdm/effect-agent.git",
-    "directory": "packages/pr-review"
-  },
-  "files": [
-    "dist",
-    "src"
-  ],
-  "type": "module",
+  "version": "0.1.0-beta.8",
   "exports": {
     ".": {
       "types": "./dist/index.d.mts",
@@ -31,6 +19,25 @@
       "default": "./dist/cli.mjs"
     }
   },
+  "dependencies": {
+    "@effect/ai-anthropic": "4.0.0-beta.107",
+    "@effect/ai-openai": "4.0.0-beta.107",
+    "@effect/platform-node": "4.0.0-beta.107",
+    "effect": "4.0.0-beta.107",
+    "effect-agent": "0.1.0-beta.8"
+  },
+  "description": "Bounded, fail-closed GitHub pull-request reviewer built on the effect-agent public surface: schema-first review contracts, read-only tools, GitHub adapters, a configuration factory, and CLI + GitHub Actions entrypoints.",
+  "license": "MIT",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/danieljvdm/effect-agent.git",
+    "directory": "packages/pr-review"
+  },
+  "files": [
+    "dist",
+    "src"
+  ],
+  "type": "module",
   "publishConfig": {
     "access": "public"
   },
@@ -39,13 +46,6 @@
     "check": "tsc --noEmit -p tsconfig.json",
     "test": "vp test --passWithNoTests"
   },
-  "dependencies": {
-    "@effect/ai-anthropic": "4.0.0-beta.107",
-    "@effect/ai-openai": "4.0.0-beta.107",
-    "@effect/platform-node": "4.0.0-beta.107",
-    "effect": "4.0.0-beta.107",
-    "effect-agent": "0.0.1-beta.6"
-  },
   "devDependencies": {
     "@effect/vitest": "4.0.0-beta.107",
     "typescript": "7.0.2",

package/src/internal/fan-out-scripted.ts

@@ -73,7 +73,8 @@ export type OfflineUnitOutcome =
   | { readonly _tag: "malformed-output" }
   /**
    * Declare more Tool Calls than the child's AgentPolicy allows in one turn —
-   * the child fails typed (AgentPolicyError "tool-calls") before any executes.
+   * none executes and the child fails typed (AgentPolicyError "tool-calls",
+   * the reviewer's deliberate `onExhaustion: "fail"` pin).
    */
   | { readonly _tag: "budget-runaway"; readonly declaredCalls: number };
 

package/src/internal/fan-out.ts

@@ -2,19 +2,9 @@ import { Effect, Schema } from "effect";
 import {
   Agent,
   AgentPolicy,
-  AgentSpawner,
-  IdGenerator,
-  RunEventSink,
   Subagent,
-  SubagentBudgetExhausted,
-  SubagentDurability,
-  SubagentDurabilityError,
-  SubagentExecutionFailure,
   SubagentPolicy,
-  SubagentPrestartDenied,
-  SubagentProjectionFailure,
   SubagentRuntime,
-  ToolCallWaiting,
   ToolExecutionClass,
   type RuntimeBinding,
 } from "effect-agent";
@@ -59,6 +49,12 @@ export const MAX_CHILD_FINDINGS = 8;
 /** One child returns at most this many non-anchored concerns. */
 export const MAX_CHILD_CONCERNS = 3;
 
+/**
+ * One mandatory diff read plus one bounded context read for every path in a
+ * maximum-size unit. Keep the child and delegation reservation aligned.
+ */
+export const MAX_FILE_REVIEW_TOOL_CALLS = MAX_UNIT_FILES * 2;
+
 // ---------------------------------------------------------------------------
 // The child: a file reviewer over one unit. Its toolkit is intentionally
 // smaller than the flat reviewer's — diff and head-file reads only, no
@@ -138,10 +134,21 @@ export const fileReviewerInstructions = makeFileReviewerInstructions();
 /** The default per-unit child execution bounds. */
 export const defaultFileReviewerPolicy = AgentPolicy.make({
   maxTurns: 8,
-  maxToolCalls: 16,
+  maxToolCalls: MAX_FILE_REVIEW_TOOL_CALLS,
   maxDuration: "4 minutes",
   toolConcurrency: 2,
   tokenBudget: 200_000,
+  // Bound one live prompt independently from cumulative usage. The engine
+  // prunes old diff/file results before paying for a summary.
+  contextTokenLimit: 150_000,
+  // Typed exhaustion, deliberately NOT the final-answer soft landing: a
+  // review is a coverage claim, and a child whose reads were rejected could
+  // still emit schema-valid findings — laundering budget exhaustion into
+  // "reviewed". Until host-owned evidence proves every mandatory
+  // read_file_diff completed, an exhausted child fails typed and its unit
+  // stays honestly unreviewed (containment turns that into result data
+  // without failing the run).
+  onExhaustion: "fail",
 });
 
 // ---------------------------------------------------------------------------
@@ -193,7 +200,7 @@ export const fileReviewPolicy = SubagentPolicy.make({
   maxChildren: MAX_REVIEW_UNITS,
   maxConcurrency: 3,
   maxTurns: 8,
-  maxToolCalls: 16,
+  maxToolCalls: MAX_FILE_REVIEW_TOOL_CALLS,
   maxDuration: "4 minutes",
 });
 
@@ -213,45 +220,6 @@ export const mapFileReviewChildFailure = (failure: {
     message: (failure.message ?? "").slice(0, 400),
   });
 
-// ---------------------------------------------------------------------------
-// The parent-facing view of the delegation Tool.
-//
-// `Subagent.define` fixes `failureMode: "error"`, so a failed child would
-// fail the WHOLE parent Run typed — one unreviewable unit would abort the
-// entire fan-out review. This coordinator wants partial results with honest
-// reporting instead, so its Toolkit carries a same-name Tool value with the
-// identical parameter/success/failure Schemas but `failureMode: "return"`:
-// Effect AI resolves handlers by Tool NAME, so the real S1 delegation handler
-// from `SubagentRuntime.layer` still executes, and a typed unit failure
-// reaches the model as a failed tool result (bounded, encoded through the
-// declared failure union) instead of aborting the Run.
-// ---------------------------------------------------------------------------
-
-/** Exactly the failure union `Subagent.define` declares for this delegation. */
-export const FileReviewDelegationFailure = Schema.Union([
-  FileReviewUnitFailed,
-  SubagentPrestartDenied,
-  SubagentBudgetExhausted,
-  SubagentProjectionFailure,
-  SubagentExecutionFailure,
-  ToolCallWaiting,
-  SubagentDurabilityError,
-]);
-
-export const DelegateFileReview = Tool.make("delegate_file_review", {
-  description: delegationDescription,
-  parameters: FileReviewRequest,
-  success: FileReviewUnitResult,
-  failure: FileReviewDelegationFailure,
-  failureMode: "return",
-})
-  .addDependency(AgentSpawner)
-  .addDependency(RunEventSink)
-  .addDependency(SubagentDurability)
-  .addDependency(IdGenerator)
-  // The delegated child's whole tool surface is read-only.
-  .annotate(ToolExecutionClass, "readonly");
-
 // ---------------------------------------------------------------------------
 // The coordinator's own tool: the deterministic unit plan over the changeset.
 // Grouping is host code (review-units.ts), not model prose, so fan-out shape
@@ -293,8 +261,6 @@ export const FanOutCoordinatorToolkitLayer = FanOutCoordinatorToolkit.toLayer({
 // apply unchanged.
 // ---------------------------------------------------------------------------
 
-export const FanOutReviewToolkit = Toolkit.make(ListReviewUnits, DelegateFileReview);
-
 /**
  * Build the coordinator's instructions. The same consumer guidance the
  * children receive is injected between the mission framing and the procedure
@@ -331,11 +297,17 @@ export const defaultFanOutPolicy = AgentPolicy.make({
   maxToolCalls: 1 + MAX_REVIEW_UNITS,
   maxDuration: "15 minutes",
   toolConcurrency: 3,
-  // One declared batch may legitimately contain a failed result for every
-  // review unit. Leave the coordinator one turn to report all of them, while
-  // still stopping a model that declares another failed delegation.
-  repeatedFailureLimit: MAX_REVIEW_UNITS + 1,
+  // Contained unit failures (SUB-033) are ordinary successful Tool results,
+  // so they no longer fold into the repeated-failure counter; the default
+  // bound suffices.
+  repeatedFailureLimit: 3,
   tokenBudget: 300_000,
+  // Child reports can amplify the merge prompt; compact before the provider's
+  // 200k-class window becomes the failure boundary.
+  contextTokenLimit: 150_000,
+  // Budget soft landing (RUN-018): an exhausted coordinator merges what it
+  // has into one best-effort review instead of discarding every child report.
+  onExhaustion: "final-answer",
 });
 
 /** Everything one fan-out configuration is made of, built as one unit so the
@@ -363,18 +335,6 @@ export interface FanOutSuiteOptions extends FanOutInstructionOptions {
   readonly maxFindings?: number | undefined;
 }
 
-const makeFanOutReviewerDefinition = (options: FanOutSuiteOptions = {}) =>
-  Agent.define("pr-fanout-reviewer", {
-    input: ReviewMission,
-    output: CodeReview,
-    instructions: makeFanOutReviewInstructions(options),
-    toolkit: FanOutReviewToolkit,
-    policy: defaultFanOutPolicy,
-    description:
-      "Coordinate one pull-request review by fanning bounded per-unit file reviews out to delegated children and merging their findings into one structured review.",
-    metadata: { deploymentClass: "E", surface: "read-only", delegation: "S1-attached" },
-  });
-
 const makeFileReviewDelegation = (child: ReturnType<typeof makeFileReviewerDefinition>) =>
   Subagent.define("delegate_file_review", {
     description: delegationDescription,
@@ -382,6 +342,11 @@ const makeFileReviewDelegation = (child: ReturnType<typeof makeFileReviewerDefin
     parameters: FileReviewRequest,
     success: FileReviewUnitResult,
     failure: FileReviewUnitFailed,
+    // First-party containment (SUB-033): a failed unit is model-visible
+    // result data instead of a parent-Run-fatal error, so the coordinator
+    // reports it honestly and keeps reviewing the other units. This retires
+    // the former same-name shadow-Tool workaround (FRICTION #7).
+    failureMode: "return",
     prepareInput: (request) =>
       Effect.succeed(
         FileReviewBrief.make({
@@ -404,13 +369,38 @@ const makeFileReviewDelegation = (child: ReturnType<typeof makeFileReviewerDefin
     policy: fileReviewPolicy,
   });
 
+/**
+ * The coordinator-facing delegation Tool: the delegation's own first-party
+ * contained Tool plus the read-only execution class (the delegated child's
+ * whole tool surface is read-only). Effect AI resolves handlers by Tool name,
+ * so `SubagentRuntime.layer`'s handler serves this annotated copy unchanged.
+ */
+const delegationToolFor = (delegation: ReturnType<typeof makeFileReviewDelegation>) =>
+  delegation.tool.annotate(ToolExecutionClass, "readonly");
+
+const makeFanOutReviewerDefinition = (
+  options: FanOutSuiteOptions,
+  delegation: ReturnType<typeof makeFileReviewDelegation>,
+) =>
+  Agent.define("pr-fanout-reviewer", {
+    input: ReviewMission,
+    output: CodeReview,
+    instructions: makeFanOutReviewInstructions(options),
+    toolkit: Toolkit.make(ListReviewUnits, delegationToolFor(delegation)),
+    policy: defaultFanOutPolicy,
+    description:
+      "Coordinate one pull-request review by fanning bounded per-unit file reviews out to delegated children and merging their findings into one structured review.",
+    metadata: { deploymentClass: "E", surface: "read-only", delegation: "S1-attached" },
+  });
+
 /** Build one coherent fan-out suite: child, coordinator, and delegation. */
 export const makeFanOutReviewSuite = (options: FanOutSuiteOptions = {}): FanOutReviewSuite => {
   const child = makeFileReviewerDefinition({ guidance: options.guidance });
+  const delegation = makeFileReviewDelegation(child);
   return {
     child,
-    parent: makeFanOutReviewerDefinition(options),
-    delegation: makeFileReviewDelegation(child),
+    parent: makeFanOutReviewerDefinition(options, delegation),
+    delegation,
   };
 };
 
@@ -425,6 +415,19 @@ export const FanOutReviewer = defaultSuite.parent;
 /** The default delegation over the default child. */
 export const fileReviewDelegation = defaultSuite.delegation;
 
+/** The default coordinator-facing delegation Tool (first-party contained mode). */
+export const DelegateFileReview = delegationToolFor(fileReviewDelegation);
+
+/** The default coordinator Toolkit. */
+export const FanOutReviewToolkit = FanOutReviewer.toolkit;
+
+/**
+ * The contained failure family the delegation can surface as result data
+ * (SUB-033), derived from the delegation itself so the coverage decoder can
+ * never diverge from what the runtime actually contains.
+ */
+export const FileReviewDelegationFailure = fileReviewDelegation.containedFailure;
+
 /** Runtime wiring: one delegation plus one explicit child Binding. */
 export const fanOutHandlersLayerFor =
   (delegation: ReturnType<typeof makeFileReviewDelegation>) =>

package/src/internal/review-agent.ts

@@ -360,6 +360,12 @@ export const defaultReviewPolicy = AgentPolicy.make({
   maxDuration: "8 minutes",
   toolConcurrency: 2,
   tokenBudget: 300_000,
+  // Keep enough output/summary headroom for the 200k-class provider window;
+  // tool-heavy histories prune before the engine spends a summarization call.
+  contextTokenLimit: 150_000,
+  // Budget soft landing (RUN-018): an exhausted reviewer returns its partial
+  // review on one final tool-free turn instead of failing the whole run.
+  onExhaustion: "final-answer",
 });
 
 // ---------------------------------------------------------------------------