@tangle-network/agent-eval 0.174.0 → 0.176.0

package/dist/supervisor-run/index.d.ts

@@ -1,5 +1,5 @@
 import { y as RolloutSplit } from "../schema-BzWDXhOR.js";
-import { A as unavailable, C as SupervisorRunTreeGap, D as WorkerLogSource, E as WallDistribution, O as isUnavailable, S as SupervisorRunTree, T as Unavailable, _ as SupervisorRunNodeRole, a as OrchestrationMetrics, b as SupervisorRunRollup, c as PerWorkerRow, d as SUPERVISOR_RUN_ROLLUP_SCHEMA, f as SUPERVISOR_RUN_SCHEMA, g as SteerBreakdown, h as SpendMeasurements, i as NO_SOURCE_LIMITS, k as showMeasured, l as RoleSpend, m as SpendMeasurement, n as EconomicsMetrics, o as OutcomeMetrics, p as SourceLimits, r as Measured, s as PatchStats, t as DecisionMetrics, u as RollupCellRow, v as SupervisorRunReader, w as SupervisorRunTreeGapCode, x as SupervisorRunSources, y as SupervisorRunReport } from "../types-CoPUTiXb.js";
+import { A as isUnavailable, C as SupervisorRunTreeGap, D as Unavailable, E as TerminalFailure, M as unavailable, O as WallDistribution, S as SupervisorRunTree, T as SupervisorStatusSource, _ as SupervisorRunNodeRole, a as OrchestrationMetrics, b as SupervisorRunRollup, c as PerWorkerRow, d as SUPERVISOR_RUN_ROLLUP_SCHEMA, f as SUPERVISOR_RUN_SCHEMA, g as SteerBreakdown, h as SpendMeasurements, i as NO_SOURCE_LIMITS, j as showMeasured, k as WorkerLogSource, l as RoleSpend, m as SpendMeasurement, n as EconomicsMetrics, o as OutcomeMetrics, p as SourceLimits, r as Measured, s as PatchStats, t as DecisionMetrics, u as RollupCellRow, v as SupervisorRunReader, w as SupervisorRunTreeGapCode, x as SupervisorRunSources, y as SupervisorRunReport } from "../types-vUdAx2Cj.js";
 //#region src/supervisor-run/source-facts.d.ts
 interface Tokens {
   input: number;
@@ -366,6 +366,29 @@ declare function renderSupervisorRunHeadline(r: SupervisorRunReport): string;
 declare function renderSupervisorRunMarkdown(r: SupervisorRunReport): string;
 declare function renderSupervisorRollupMarkdown(rollup: SupervisorRunRollup, title?: string): string;
 //#endregion
+//#region src/supervisor-run/report-command.d.ts
+/**
+ * `agent-eval supervisor-run report <runDir>` takes one run directory and
+ * prints the report the module already renders. The command reads through
+ * `analyzeSupervisorRun`, so a Runtime run dir and a loops run dir take the
+ * same path and the status comes from the record `terminal-record.ts` names.
+ *
+ * Exit codes follow the other self-parsing subcommands: 2 for a usage error,
+ * 1 when the directory cannot be read (missing, not a directory, or a corrupt
+ * journal or terminal record), 0 after the report was written to stdout.
+ */
+type SupervisorRunReportFormat = 'headline' | 'markdown' | 'json';
+/** Output sinks, injectable so tests read the exact bytes a terminal would. */
+interface SupervisorRunCommandIo {
+  readonly stdout: (text: string) => void;
+  readonly stderr: (text: string) => void;
+}
+/**
+ * CLI driver for `agent-eval supervisor-run`. `argv` is everything after the
+ * command name, so `['report', runDir, ...flags]`. Returns the process exit code.
+ */
+declare function runSupervisorRunCommand(argv: readonly string[], io?: SupervisorRunCommandIo): Promise<number>;
+//#endregion
 //#region src/supervisor-run/rollout-nodes.d.ts
 interface SupervisorRolloutOptions {
   /** Benchmark/suite id for `task.suite`. Defaults to `'supervisor-run'`. */
@@ -415,9 +438,12 @@ interface RuntimeReaderOptions {
  * A run dir without `spawn-journal.jsonl` returns the same absent-shaped
  * sources `readLoopsSupervisorRun` returns for a missing store: `journal` and
  * `workers` null, each with its reason, so every dependent metric reads
- * `unavailable` — never 0 and never a throw. Pass `strict: true` to throw on
- * the missing journal instead. A journal that exists but cannot be parsed
- * always throws: a corrupt journal is a defect, not an absence.
+ * `unavailable` — never 0 and never a throw. Its `result.json` and
+ * `failure.json` are still read, so a run that died before its first spawn
+ * reports the failure Runtime recorded. Pass `strict: true` to throw on the
+ * missing journal instead. A journal, result, or failure document that exists
+ * but cannot be parsed always throws: a corrupt record is a defect, not an
+ * absence.
  *
  * The reader translates storage envelopes only. It does not assign research
  * roles, interpret artifacts, or turn process completion into a quality
@@ -426,8 +452,47 @@ interface RuntimeReaderOptions {
 declare function readRuntimeSupervisorRun(runDir: string, opts?: RuntimeReaderOptions): Promise<SupervisorRunSources>;
 /** The agent-runtime file-backed layout as a `SupervisorRunReader`. */
 declare function runtimeSupervisorRunReader(runDir: string, opts?: RuntimeReaderOptions): SupervisorRunReader;
-/** True when a directory contains Runtime's canonical file-backed journal. */
+/**
+ * True when a directory holds any file only Runtime's durable layer writes:
+ * the spawn journal, the observer journal, or the failure record. A run that
+ * threw before its first spawn has no journal yet and is still a Runtime run.
+ */
 declare function isRuntimeSupervisorRunDir(runDir: string): Promise<boolean>;
 //#endregion
-export { type ClaudeCodeReaderOptions, type CloseRow, type DecisionMetrics, type EconomicsMetrics, type LoopsReaderOptions, type Measured, NO_SOURCE_LIMITS, type OrchestrationMetrics, type OutcomeMetrics, type PatchStats, type PerWorkerRow, type RoleSpend, type RollupCellRow, type RuntimeReaderOptions, SUPERVISOR_RUN_INTEGRITY_SCHEMA, SUPERVISOR_RUN_ROLLUP_SCHEMA, SUPERVISOR_RUN_SCHEMA, type SourceLimits, type SpawnRow, type SpendMeasurement, type SpendMeasurements, type SteerAcknowledgementFact, type SteerBreakdown, type SteerRequestFact, type SupervisorJournalDialect, type SupervisorRolloutOptions, type SupervisorRunIntegrityEvidence, type SupervisorRunIntegrityIssue, type SupervisorRunIntegrityIssueCode, type SupervisorRunIntegrityOptions, type SupervisorRunIntegrityReport, type SupervisorRunIntegritySeverity, type SupervisorRunNodeRole, type SupervisorRunReader, type SupervisorRunReport, type SupervisorRunRollup, type SupervisorRunSourceOnlyCheckCode, type SupervisorRunSources, type SupervisorRunTree, type SupervisorRunTreeGap, type SupervisorRunTreeGapCode, type SupervisorTreeFacts, type Unavailable, type WallDistribution, type WorkerLogFacts, type WorkerLogSource, type WriteSupervisorRunOptions, analyzeSupervisorRun, analyzeSupervisorRunIntegrity, analyzeSupervisorRunSources, claudeCodeSupervisorRunReader, findSupervisorRunDirIn, findSupervisorRunDirs, isRuntimeSupervisorRunDir, isUnavailable, loopsSupervisorRunReader, parsePatch, parseSupervisorTree, readClaudeCodeSupervisorRun, readLoopsSupervisorRun, readRuntimeSupervisorRun, renderSupervisorRollupMarkdown, renderSupervisorRunHeadline, renderSupervisorRunMarkdown, reportSupervisorRound, rollupSupervisorRuns, runtimeSupervisorRunReader, showMeasured, supervisorRunRolloutLines, unavailable, writeSupervisorRunReport, writeSupervisorRunReportSafe };
+//#region src/supervisor-run/terminal-record.d.ts
+/** The reason every terminal field reads `unavailable` when no store wrote one. */
+declare const NO_TERMINAL_RECORD = "no terminal record: Runtime result.json kind, Runtime failure.json, or legacy state.json / result.json status";
+/** The status reported for a run whose directory holds Runtime's failure record and no result. */
+declare const RUNTIME_FAILED_STATUS = "failed";
+interface TerminalRecordInput {
+  /** Legacy loops `state.json`, parsed; null when absent. */
+  readonly state: Record<string, unknown> | null;
+  /** `result.json`, parsed; Runtime's `SupervisedResult` or the legacy loops result. */
+  readonly result: Record<string, unknown> | null;
+  /**
+   * Runtime's `failure.json`, parsed. `null` when the store was read and holds
+   * none; `undefined` when the store has no such document type.
+   */
+  readonly failure: Record<string, unknown> | null | undefined;
+}
+interface TerminalRecord {
+  readonly supStatus: Measured<string>;
+  readonly supStatusSource: Measured<SupervisorStatusSource>;
+  /**
+   * Runtime's `reason` on a `no-winner` result. `null` when the record carries
+   * none (a winner, a failure record, or a legacy document).
+   */
+  readonly supReason: Measured<string | null>;
+  /** The recorded error. `null` when the run recorded a result and no error. */
+  readonly failure: Measured<TerminalFailure | null>;
+  /**
+   * True when the run recorded a delivered result (Runtime `winner`, legacy
+   * `completed`); false on every other recorded terminal state; null when no
+   * terminal record exists.
+   */
+  readonly completed: boolean | null;
+}
+declare function readTerminalRecord(input: TerminalRecordInput): TerminalRecord;
+//#endregion
+export { type ClaudeCodeReaderOptions, type CloseRow, type DecisionMetrics, type EconomicsMetrics, type LoopsReaderOptions, type Measured, NO_SOURCE_LIMITS, NO_TERMINAL_RECORD, type OrchestrationMetrics, type OutcomeMetrics, type PatchStats, type PerWorkerRow, RUNTIME_FAILED_STATUS, type RoleSpend, type RollupCellRow, type RuntimeReaderOptions, SUPERVISOR_RUN_INTEGRITY_SCHEMA, SUPERVISOR_RUN_ROLLUP_SCHEMA, SUPERVISOR_RUN_SCHEMA, type SourceLimits, type SpawnRow, type SpendMeasurement, type SpendMeasurements, type SteerAcknowledgementFact, type SteerBreakdown, type SteerRequestFact, type SupervisorJournalDialect, type SupervisorRolloutOptions, type SupervisorRunCommandIo, type SupervisorRunIntegrityEvidence, type SupervisorRunIntegrityIssue, type SupervisorRunIntegrityIssueCode, type SupervisorRunIntegrityOptions, type SupervisorRunIntegrityReport, type SupervisorRunIntegritySeverity, type SupervisorRunNodeRole, type SupervisorRunReader, type SupervisorRunReport, type SupervisorRunReportFormat, type SupervisorRunRollup, type SupervisorRunSourceOnlyCheckCode, type SupervisorRunSources, type SupervisorRunTree, type SupervisorRunTreeGap, type SupervisorRunTreeGapCode, type SupervisorStatusSource, type SupervisorTreeFacts, type TerminalFailure, type TerminalRecord, type TerminalRecordInput, type Unavailable, type WallDistribution, type WorkerLogFacts, type WorkerLogSource, type WriteSupervisorRunOptions, analyzeSupervisorRun, analyzeSupervisorRunIntegrity, analyzeSupervisorRunSources, claudeCodeSupervisorRunReader, findSupervisorRunDirIn, findSupervisorRunDirs, isRuntimeSupervisorRunDir, isUnavailable, loopsSupervisorRunReader, parsePatch, parseSupervisorTree, readClaudeCodeSupervisorRun, readLoopsSupervisorRun, readRuntimeSupervisorRun, readTerminalRecord, renderSupervisorRollupMarkdown, renderSupervisorRunHeadline, renderSupervisorRunMarkdown, reportSupervisorRound, rollupSupervisorRuns, runSupervisorRunCommand, runtimeSupervisorRunReader, showMeasured, supervisorRunRolloutLines, unavailable, writeSupervisorRunReport, writeSupervisorRunReportSafe };
 //# sourceMappingURL=index.d.ts.map

package/dist/supervisor-run/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/supervisor-run/source-facts.ts","../../src/supervisor-run/analyze.ts","../../src/supervisor-run/claude-code-reader.ts","../../src/supervisor-run/integrity-types.ts","../../src/supervisor-run/integrity.ts","../../src/supervisor-run/loops-reader.ts","../../src/supervisor-run/render.ts","../../src/supervisor-run/rollout-nodes.ts","../../src/supervisor-run/runtime-reader.ts"],"mappings":";;;UAwBU;EACR;EACA;EACA;EACA;;EAEA;;;;;;EAMA;;UAGe;EACf,QAAQ;;;;;;EAMR;EACA;;;;;;EAMA;;UAmFe;;EAEf;EACA;EACA;EACA;;EAEA,MAAM;;EAEN;;EAEA;EACA;;EAEA;EACA;;UAGe;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA,OAAO;;EAEP;;UAGe;;EAEf;EACA;EACA;;EAEA;EACA;EACA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;;EAEA;EACA;;EAEA;;EAEA;EACA;EACA;EACA;EACA,eAAe;EACf,uBAAuB;;UAGR;;EAEf;EACA;;UAGe;;EAEf;;EAEA;EACA;;;;;;UAOe;WACN;WACA,iBAAiB;WACjB,iBAAiB;WACjB,uBAAuB;WACvB,uBAAuB;WACvB;;;;;;;IAOP;IACA;IACA;IACA;;IAEA;;IAEA;;IAEA;IACA;;IAEA;;IAEA;IACA;;IAEA;;WAEO,YAAY,oBAAoB;;WAEhC,wBAAwB;;WAExB;;;;;;;;;WASA;;WAEA;;;;;;WAMA,0BAA0B,SAAS;;;;;;;WAOnC;;WAEA,gBAAgB;;WAEhB,OAAO;WACP;WACA;;;;;;WAMA;WACA;;;;;;;;;;;;;;;;;;KAmDC;iBAiFI,oBAAoB,KAAK,uBAAuB;;;;;;;iBC9WhD,4BACd,KAAK,sBACL,qBACC;;iBAo2Ba,WAAW,eAAe;;;;;;iBA+C1B,qBAAqB,kBAAkB,wBAAwB;;;UCp4B9D;;WAEN;;;;;WAKA;WACA;WACA;WACA;WACA;WACA;WACA;;;;;;iBAmOW,4BACpB,MAAM,0BACL,QAAQ;;iBAiUK,8BAA8B,MAAM,0BAA0B;;;cCnoBjE;KAED;KAEA;KAKA;UAmCK;WACN;WACA;;UAGM;WACN,MAAM;WACN;WACA,UAAU;WACV;WACA;WACA;WACA,mBAAmB;WACnB;WACA,WAAW,SAAS;;UAGd;WACN,eAAe;WACf;WACA;WACA,MAAM;WACN,iBAAiB;;UAGX;;WAEN;;;;;;;;iBC5BK,8BACd,OAAO,uBAAuB,mBAC9B,UAAS,gCACR;;;;;;;iBCDmB,uBAAuB,aAAa;UAUzC;;WAEN;;WAEA;;WAEA;;WAEA;;;;;;;iBAQW,uBACpB,gBACA,OAAM,qBACL,QAAQ;;iBAmKK,yBACd,gBACA,OAAM,qBACL;;;;;;iBAkCmB,qBACpB,gBAAgB,sBAAsB,sBACtC,OAAM,qBACL,QAAQ;UAgBM,kCAAkC;;WAExC;;WAEA;;;;;;WAMA;;;;;;;;iBASW,yBACpB,gBACA,OAAM,4BACL,QAAQ;;;;;iBA0CW,6BACpB,gBACA,OAAM,4BACL,QAAQ;;;;;;iBAgBW,sBACpB,gBACA,OAAM;EAA8B;IACnC,QAAQ;;;;;;;iBA+BW,sBAAsB,eAAe;;;;;;;iBCzY3C,4BAA4B,GAAG;iBAiC/B,4BAA4B,GAAG;iBA+K/B,+BACd,QAAQ,qBACR;;;UC9Le;;WAEN;;WAEA,QAAQ;;WAER;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;WACA;WACA;WACA;WACA;;WAEA;;;;;;;iBAwBK,0BACd,KAAK,sBACL,OAAM,2BACL;;;UCiYc;;;;;;WAMN;;;;;;;;;;;;;;;;iBAsDW,yBACpB,gBACA,OAAM,uBACL,QAAQ;;iBAkEK,2BACd,gBACA,OAAM,uBACL;;iBAKmB,0BAA0B,iBAAiB"}
+{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/supervisor-run/source-facts.ts","../../src/supervisor-run/analyze.ts","../../src/supervisor-run/claude-code-reader.ts","../../src/supervisor-run/integrity-types.ts","../../src/supervisor-run/integrity.ts","../../src/supervisor-run/loops-reader.ts","../../src/supervisor-run/render.ts","../../src/supervisor-run/report-command.ts","../../src/supervisor-run/rollout-nodes.ts","../../src/supervisor-run/runtime-reader.ts","../../src/supervisor-run/terminal-record.ts"],"mappings":";;;UAwBU;EACR;EACA;EACA;EACA;;EAEA;;;;;;EAMA;;UAGe;EACf,QAAQ;;;;;;EAMR;EACA;;;;;;EAMA;;UAmFe;;EAEf;EACA;EACA;EACA;;EAEA,MAAM;;EAEN;;EAEA;EACA;;EAEA;EACA;;UAGe;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA,OAAO;;EAEP;;UAGe;;EAEf;EACA;EACA;;EAEA;EACA;EACA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;;EAEA;EACA;;EAEA;;EAEA;EACA;EACA;EACA;EACA,eAAe;EACf,uBAAuB;;UAGR;;EAEf;EACA;;UAGe;;EAEf;;EAEA;EACA;;;;;;UAOe;WACN;WACA,iBAAiB;WACjB,iBAAiB;WACjB,uBAAuB;WACvB,uBAAuB;WACvB;;;;;;;IAOP;IACA;IACA;IACA;;IAEA;;IAEA;;IAEA;IACA;;IAEA;;IAEA;IACA;;IAEA;;WAEO,YAAY,oBAAoB;;WAEhC,wBAAwB;;WAExB;;;;;;;;;WASA;;WAEA;;;;;;WAMA,0BAA0B,SAAS;;;;;;;WAOnC;;WAEA,gBAAgB;;WAEhB,OAAO;WACP;WACA;;;;;;WAMA;WACA;;;;;;;;;;;;;;;;;;KAmDC;iBAiFI,oBAAoB,KAAK,uBAAuB;;;;;;;iBC7WhD,4BACd,KAAK,sBACL,qBACC;;iBA62Ba,WAAW,eAAe;;;;;;iBA+C1B,qBAAqB,kBAAkB,wBAAwB;;;UC94B9D;;WAEN;;;;;WAKA;WACA;WACA;WACA;WACA;WACA;WACA;;;;;;iBAmOW,4BACpB,MAAM,0BACL,QAAQ;;iBAiUK,8BAA8B,MAAM,0BAA0B;;;cCnoBjE;KAED;KAEA;KAKA;UAmCK;WACN;WACA;;UAGM;WACN,MAAM;WACN;WACA,UAAU;WACV;WACA;WACA;WACA,mBAAmB;WACnB;WACA,WAAW,SAAS;;UAGd;WACN,eAAe;WACf;WACA;WACA,MAAM;WACN,iBAAiB;;UAGX;;WAEN;;;;;;;;iBC5BK,8BACd,OAAO,uBAAuB,mBAC9B,UAAS,gCACR;;;;;;;iBCDmB,uBAAuB,aAAa;UAUzC;;WAEN;;WAEA;;WAEA;;WAEA;;;;;;;iBAQW,uBACpB,gBACA,OAAM,qBACL,QAAQ;;iBAmKK,yBACd,gBACA,OAAM,qBACL;;;;;;iBAkCmB,qBACpB,gBAAgB,sBAAsB,sBACtC,OAAM,qBACL,QAAQ;UAgBM,kCAAkC;;WAExC;;WAEA;;;;;;WAMA;;;;;;;;iBASW,yBACpB,gBACA,OAAM,4BACL,QAAQ;;;;;iBA0CW,6BACpB,gBACA,OAAM,4BACL,QAAQ;;;;;;iBAgBW,sBACpB,gBACA,OAAM;EAA8B;IACnC,QAAQ;;;;;;;iBA+BW,sBAAsB,eAAe;;;;;;;iBCxY3C,4BAA4B,GAAG;iBAiD/B,4BAA4B,GAAG;iBAkL/B,+BACd,QAAQ,qBACR;;;;;;;;;;;;;KCzOU;;UAQK;WACN,SAAS;WACT,SAAS;;;;;;iBA4DE,wBACpB,yBACA,KAAI,yBACH;;;UChDc;;WAEN;;WAEA,QAAQ;;WAER;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;WACA;WACA;WACA;WACA;;WAEA;;;;;;;iBAwBK,0BACd,KAAK,sBACL,OAAM,2BACL;;;UC6Yc;;;;;;WAMN;;;;;;;;;;;;;;;;;;;iBAyEW,yBACpB,gBACA,OAAM,uBACL,QAAQ;;iBAmEK,2BACd,gBACA,OAAM,uBACL;;;;;;iBASmB,0BAA0B,iBAAiB;;;;cCpnBpD;;cAIA;UAEI;;WAEN,OAAO;;WAEP,QAAQ;;;;;WAKR,SAAS;;UAGH;WACN,WAAW;WACX,iBAAiB,SAAS;;;;;WAK1B,WAAW;;WAEX,SAAS,SAAS;;;;;;WAMlB;;iBAuBK,mBAAmB,OAAO,sBAAsB"}