@skill-harness/core 0.16.0 → 0.18.0

package/README.md

@@ -1,8 +1,7 @@
 # @skill-harness/core
 
 The engine behind [skill-harness](https://github.com/mojomanyana/skill-harness):
-spec parsing, discovery, run orchestration, LLM-judge grading, scoring, results,
-and the separately versioned durable qualification-runner-v1 primitives.
+spec parsing, discovery, run orchestration, LLM-judge grading, scoring, and results.
 
 Internal API — use the `skill-harness` CLI. See the repo for docs.
 

package/dist/adapters/types.d.ts

@@ -1,6 +1,4 @@
 import type { ExecutionTraceV1 } from "../capture-trace-types.js";
-import type { TrajectoryEventSource } from "../spec.js";
-import type { TrajectoryEventV1 } from "../trajectory-gates.js";
 export type RunMode = "red" | "green" | "force";
 /** A provider+model pair, e.g. { provider: "fireworks", model: "accounts/.../deepseek-v4-pro" }. */
 export interface ModelRef {
@@ -57,16 +55,12 @@ export interface RunReq {
      * extension the developer happens to have installed cannot join the test.
      */
     extensions?: string[];
-    /** Workspace-local native ledgers to normalize after the subject finishes. */
-    eventSources?: TrajectoryEventSource[];
     /**
      * Extra env for the subject process, from the arm. Merged over `process.env` by
      * the adapter — an arm must be able to add `PI_GRANTS_*` without the harness
      * inheriting a caller's copy of them.
      */
     armEnv?: Record<string, string>;
-    /** Adapter-produced observation callback. Callers receive evidence; they never supply delivery status. */
-    onPromptObservation?: (observation: PromptProvenance) => void;
 }
 /** A judge request: single prompt, no skills, no session. */
 export interface JudgeReq {
@@ -79,10 +73,6 @@ export interface StructuredRun {
     transcript: string;
     /** One trace per turn — each `pi` invocation emits an independent event stream. */
     traces: ExecutionTraceV1[];
-    /** Adapter-neutral workflow/tool events. Additive; absent on older adapters. */
-    events?: TrajectoryEventV1[];
-    /** Native sources that were required but missing/malformed. Never treated as an empty success. */
-    eventErrors?: string[];
     /**
      * Set when pi failed provider-side (auth, transport) rather than the model
      * answering badly. `run.ts` turns this into ERROR — never a model verdict.
@@ -91,8 +81,6 @@ export interface StructuredRun {
 }
 export interface HarnessAdapter {
     name: string;
-    /** True only when the adapter computes final provider-payload observations. */
-    observesPrompts?: boolean;
     available(): Promise<boolean>;
     run(req: RunReq): Promise<string>;
     /**

package/dist/capture-trace-types.d.ts

@@ -151,99 +151,3 @@ export interface ExecutionTraceV1 {
     /** SHA-256 over the deterministic serialization of this trace, minus this field. */
     trace_sha256?: string;
 }
-export declare const CAPTURE_SCHEMA_VERSION = 1;
-/** Why a conversation was captured. */
-export type CaptureClassification = "failure" | "good_example";
-/** Where a capture is in the human review pipeline. */
-export type CapturePromotionStatus = "pending" | "promoted" | "rejected";
-/** What the human confirmed as responsible for the behavior. */
-export interface CaptureTarget {
-    kind: "skill" | "subagent";
-    /** Path relative to the skills root or repo root — never absolute. */
-    path: string;
-    /** SHA-256 of the target's content at capture time, so drift is detectable. */
-    content_sha256: string;
-}
-/**
- * Provenance for a capture.
- *
- * Hashed, not absolute: an absolute session path identifies a machine and a
- * user, and the capture file is meant to be committed. The hash is enough to
- * recognize the same session again locally, which is all provenance needs to do.
- */
-export interface CaptureProvenance {
-    session_sha256: string;
-    /** Indices of the selected turn range within the active branch, inclusive. */
-    turn_range: {
-        start: number;
-        end: number;
-    };
-    subject?: ModelRef;
-    git_commit?: string;
-    /** True when the working tree was dirty at capture time. */
-    git_dirty?: boolean;
-}
-/**
- * A reviewed conversation awaiting promotion into a scenario.
- *
- * Lives under `<skill>/tests/captures/`, NOT in `specification.yaml`. A pending
- * capture is not a test: putting it in the spec would drag it into ship-bar
- * totals, staleness, lift, and stability, and the alternative — a `draft: true`
- * flag every runner, scorer and linter has to honor — is a state that only has
- * to be forgotten once to either corrupt a grade or silently drop a real
- * scenario from a release run.
- */
-export interface CaptureCaseV1 {
-    capture_schema: typeof CAPTURE_SCHEMA_VERSION;
-    id: string;
-    created: string;
-    classification: CaptureClassification;
-    /** Sanitized user turns only. Assistant prose is evidence, never an oracle. */
-    turns: string[];
-    /** Human-written, in their own words. Required — a capture with no expectation is not reviewable. */
-    expected_behavior: string;
-    /** Editable draft checklist. Offline-derived by default; LLM drafting is opt-in and costs tokens. */
-    checklist: string[];
-    target: CaptureTarget;
-    provenance: CaptureProvenance;
-    status: CapturePromotionStatus;
-    /** Set once promoted: the scenario id appended to `specification.yaml`. */
-    scenario_id?: string;
-    /**
-     * `covers` refs for this pending case, so `coverage` can park it against the
-     * instructions it is about before anyone promotes it.
-     *
-     * Derived from `target` — the author has already been made to choose which
-     * instructions are responsible, and asking again in different words would be
-     * asking the same question twice. File granularity, not `#section`: the target
-     * choice attributes a file, and inventing a section would be a guess the
-     * session cannot support.
-     *
-     * Written relative to the SPEC dir (`<skill>/tests`), matching how a scenario's
-     * own `covers` resolve. `coverage` read this field from the start; nothing ever
-     * wrote it, so every pending case parked against nothing and the feature was
-     * inert.
-     */
-    covers?: string[];
-}
-/**
- * Local-only evidence sidecar for a capture.
- *
- * Written to `<skill>/tests/captures/.local/` and git-ignored by default. It may
- * hold a sanitized assistant excerpt and tool-call summaries to help a human
- * review the capture, and must NEVER hold hidden thinking, complete tool-result
- * bodies, or the effective system prompt.
- *
- * pi's stream carries thinking in `message_end`, `turn_end` AND `agent_end`, so
- * dropping it is an explicit filter at every one of those, not a side effect of
- * reading the convenient field.
- */
-export interface CaptureEvidenceV1 {
-    capture_id: string;
-    /** Sanitized, truncated excerpt of the assistant text of every selected turn, joined. */
-    assistant_excerpt: string;
-    /** Tool name, error state and redacted arguments — no bodies. */
-    tool_calls: Array<Pick<TraceToolCall, "name" | "isError"> & {
-        args: SanitizedArgs;
-    }>;
-}

package/dist/capture-trace-types.js

@@ -16,8 +16,4 @@
 // Execution trace
 // ---------------------------------------------------------------------------
 export const EXECUTION_TRACE_VERSION = 2;
-// ---------------------------------------------------------------------------
-// Capture case
-// ---------------------------------------------------------------------------
-export const CAPTURE_SCHEMA_VERSION = 1;
 //# sourceMappingURL=capture-trace-types.js.map

package/dist/command-cost.d.ts

@@ -4,6 +4,6 @@
  * Keep this list narrow: commands that can conditionally spend (notably `regate`)
  * do not belong here even when their common path only replays saved artifacts.
  */
-export declare const FREE_OFFLINE_COMMANDS: readonly ["affected", "archive", "coverage", "init", "judge-agreement", "learning", "lint", "list", "rescore", "restamp", "screen", "stability"];
+export declare const FREE_OFFLINE_COMMANDS: readonly ["coverage", "init", "lint", "list", "rescore", "restamp", "stability"];
 export type FreeOfflineCommand = typeof FREE_OFFLINE_COMMANDS[number];
 export declare function isFreeOfflineCommand(command: string): command is FreeOfflineCommand;

package/dist/command-cost.js

@@ -5,17 +5,12 @@
  * do not belong here even when their common path only replays saved artifacts.
  */
 export const FREE_OFFLINE_COMMANDS = [
-    "affected",
-    "archive",
     "coverage",
     "init",
-    "judge-agreement",
-    "learning",
     "lint",
     "list",
     "rescore",
     "restamp",
-    "screen",
     "stability",
 ];
 export function isFreeOfflineCommand(command) {

package/dist/execution-trace.js

@@ -1,7 +1,7 @@
 import { createHash } from "node:crypto";
 import { normalizeSubagentCall } from "./trace-gates.js";
 import { EXECUTION_TRACE_VERSION } from "./capture-trace-types.js";
-import { redactArgs, redactText } from "./capture.js";
+import { redactArgs, redactText } from "./redaction.js";
 /**
  * Parse pi's `--mode json` event stream into an `ExecutionTraceV1`.
  *

package/dist/index.d.ts

@@ -1,4 +1,3 @@
-export * from "./trusted-host-supervision.js";
 export * from "./spec.js";
 export * from "./discover.js";
 export * from "./run.js";
@@ -13,6 +12,7 @@ export * from "./rescore.js";
 export * from "./workspace.js";
 export * from "./seeded.js";
 export * from "./report.js";
+export * from "./metrics.js";
 export * from "./trends.js";
 export * from "./lint.js";
 export * from "./restamp.js";
@@ -26,39 +26,19 @@ export * from "./version.js";
 export * from "./defaults.js";
 export * from "./command-cost.js";
 export * from "./judge-policy.js";
-export * from "./judge-agreement.js";
 export * from "./regate.js";
-export * from "./screen.js";
 export * from "./prompt-normalization.js";
 export * from "./downgrade.js";
 export * from "./canary.js";
 export * from "./stability.js";
 export * from "./capture-trace-types.js";
-export * from "./work-capture.js";
-export * from "./work-signals.js";
-export * from "./work-capture-schema.js";
-export * from "./factory-calibration.js";
-export * from "./adoption.js";
-export * from "./investigation.js";
-export * from "./intervention.js";
-export * from "./intervention-results.js";
 export * from "./spec-write.js";
-export * from "./capture.js";
+export * from "./redaction.js";
 export * from "./execution-trace.js";
 export * from "./trace-gates.js";
-export * from "./trajectory-gates.js";
+export * from "./trajectory-events.js";
 export * from "./instruction-coverage.js";
-export * from "./affected.js";
-export * from "./adjudication.js";
 export * from "./vote-panel.js";
-export * from "./qualification-panels.js";
-export * from "./qualification-panel-store.js";
-export * from "./qualification-validation.js";
-export * from "./comparison.js";
 export * from "./sandbox.js";
 export * from "./provider-failure.js";
 export * from "./arms.js";
-export * from "./qualification-config.js";
-export * from "./qualification-oauth-directory.js";
-export * from "./qualification-store.js";
-export * from "./qualification-runner.js";

package/dist/index.js

@@ -1,4 +1,3 @@
-export * from "./trusted-host-supervision.js";
 export * from "./spec.js";
 export * from "./discover.js";
 export * from "./run.js";
@@ -13,6 +12,7 @@ export * from "./rescore.js";
 export * from "./workspace.js";
 export * from "./seeded.js";
 export * from "./report.js";
+export * from "./metrics.js";
 export * from "./trends.js";
 export * from "./lint.js";
 export * from "./restamp.js";
@@ -26,40 +26,20 @@ export * from "./version.js";
 export * from "./defaults.js";
 export * from "./command-cost.js";
 export * from "./judge-policy.js";
-export * from "./judge-agreement.js";
 export * from "./regate.js";
-export * from "./screen.js";
 export * from "./prompt-normalization.js";
 export * from "./downgrade.js";
 export * from "./canary.js";
 export * from "./stability.js";
 export * from "./capture-trace-types.js";
-export * from "./work-capture.js";
-export * from "./work-signals.js";
-export * from "./work-capture-schema.js";
-export * from "./factory-calibration.js";
-export * from "./adoption.js";
-export * from "./investigation.js";
-export * from "./intervention.js";
-export * from "./intervention-results.js";
 export * from "./spec-write.js";
-export * from "./capture.js";
+export * from "./redaction.js";
 export * from "./execution-trace.js";
 export * from "./trace-gates.js";
-export * from "./trajectory-gates.js";
+export * from "./trajectory-events.js";
 export * from "./instruction-coverage.js";
-export * from "./affected.js";
-export * from "./adjudication.js";
 export * from "./vote-panel.js";
-export * from "./qualification-panels.js";
-export * from "./qualification-panel-store.js";
-export * from "./qualification-validation.js";
-export * from "./comparison.js";
 export * from "./sandbox.js";
 export * from "./provider-failure.js";
 export * from "./arms.js";
-export * from "./qualification-config.js";
-export * from "./qualification-oauth-directory.js";
-export * from "./qualification-store.js";
-export * from "./qualification-runner.js";
 //# sourceMappingURL=index.js.map

package/dist/instruction-coverage.d.ts

@@ -61,8 +61,6 @@ export interface SectionCoverage {
     section: Section;
     /** Scenario ids that declare a reference to this section. */
     scenarios: string[];
-    /** Capture ids parked against this section but not yet promoted. */
-    pendingCaptures: string[];
 }
 export interface BrokenRef {
     scenarioId: string;
@@ -87,11 +85,6 @@ export interface CoverageOptions {
     scenarios: Scenario[];
     /** Instruction files to report on even if nothing references them. */
     baseFiles?: string[];
-    /** capture id → covers refs, for parking a pending case against a section. */
-    pendingCaptures?: {
-        id: string;
-        covers: string[];
-    }[];
 }
 /**
  * Build the coverage report.

package/dist/instruction-coverage.js

@@ -130,7 +130,7 @@ export function computeCoverage(opts) {
         const k = key(file, section.slug);
         let entry = bySection.get(k);
         if (!entry) {
-            entry = { file, section, scenarios: [], pendingCaptures: [] };
+            entry = { file, section, scenarios: [] };
             bySection.set(k, entry);
         }
         return entry;
@@ -142,7 +142,7 @@ export function computeCoverage(opts) {
             ensure(file, s);
     const broken = [];
     const unmapped = [];
-    const attach = (id, refs, into) => {
+    const attach = (id, refs) => {
         for (const raw of refs) {
             const ref = parseCoversRef(raw);
             const sections = readSections(ref.file);
@@ -155,7 +155,7 @@ export function computeCoverage(opts) {
             if (ref.slug === undefined) {
                 // A whole-file reference covers every section in it.
                 for (const s of sections)
-                    ensure(ref.file, s)[into].push(id);
+                    ensure(ref.file, s).scenarios.push(id);
                 continue;
             }
             const match = sections.find((s) => s.slug === ref.slug);
@@ -168,7 +168,7 @@ export function computeCoverage(opts) {
                 });
                 continue;
             }
-            ensure(ref.file, match)[into].push(id);
+            ensure(ref.file, match).scenarios.push(id);
         }
     };
     for (const s of opts.scenarios) {
@@ -176,10 +176,8 @@ export function computeCoverage(opts) {
             unmapped.push(s.id);
             continue;
         }
-        attach(s.id, s.covers, "scenarios");
+        attach(s.id, s.covers);
     }
-    for (const c of opts.pendingCaptures ?? [])
-        attach(c.id, c.covers, "pendingCaptures");
     const sections = [...bySection.values()].sort((a, b) => a.file.localeCompare(b.file) || a.section.startLine - b.section.startLine);
     const covered = sections.filter((s) => s.scenarios.length > 0);
     const uncovered = sections.filter((s) => s.scenarios.length === 0);

package/dist/journal.d.ts

@@ -63,14 +63,6 @@ export type JournalEvent = {
     ok: boolean;
     detail: string;
     rep?: number;
-}
-/** One adjudication pass: which cells were re-judged, what it cost in CALLS, what stayed unresolved. */
- | {
-    event: "adjudication";
-    ts: string;
-    triggered: string[];
-    judge_calls: number;
-    unresolved: string[];
 } | {
     event: "judge-verdict";
     ts: string;

package/dist/metrics.d.ts

@@ -0,0 +1,18 @@
+import type { ScenarioResult } from "./results.js";
+export interface AggregateMetrics {
+    wall_time_ms: number;
+    judge_calls: number;
+    judge_rejudge_calls: number;
+    subject_metrics_reps: number;
+    total_reps: number;
+    input_tokens: number | null;
+    output_tokens: number | null;
+    cache_read_tokens: number | null;
+    cache_write_tokens: number | null;
+    subject_cost_usd: number | null;
+    cost_source: "provider-reported" | "subscription" | "unreported" | null;
+    tool_calls: number | null;
+    delegated_children: number | null;
+    max_concurrency: number | null;
+}
+export declare function aggregateMetrics(scenarios: ScenarioResult[]): AggregateMetrics;

package/dist/metrics.js

@@ -0,0 +1,28 @@
+export function aggregateMetrics(scenarios) {
+    const metrics = scenarios.map((scenario) => scenario.metrics).filter((value) => value !== undefined);
+    const sumOptional = (field) => {
+        const values = metrics.map((value) => value[field]).filter((value) => typeof value === "number");
+        return values.length ? values.reduce((sum, value) => sum + value, 0) : null;
+    };
+    const maxValues = metrics.map((value) => value.max_concurrency).filter((value) => typeof value === "number");
+    return {
+        wall_time_ms: metrics.reduce((sum, value) => sum + value.wall_time_ms, 0),
+        judge_calls: metrics.reduce((sum, value) => sum + value.judge_calls, 0),
+        judge_rejudge_calls: metrics.reduce((sum, value) => sum + value.judge_rejudge_calls, 0),
+        subject_metrics_reps: metrics.reduce((sum, value) => sum + value.subject_metrics_reps, 0),
+        total_reps: metrics.reduce((sum, value) => sum + value.total_reps, 0),
+        input_tokens: sumOptional("input_tokens"),
+        output_tokens: sumOptional("output_tokens"),
+        cache_read_tokens: sumOptional("cache_read_tokens"),
+        cache_write_tokens: sumOptional("cache_write_tokens"),
+        subject_cost_usd: sumOptional("subject_cost_usd"),
+        cost_source: (() => {
+            const sources = scenarios.map((scenario) => scenario.metrics?.cost_source).filter((source) => source !== undefined);
+            return sources.length === 0 ? null : sources.every((source) => source === sources[0]) ? sources[0] : "unreported";
+        })(),
+        tool_calls: sumOptional("tool_calls"),
+        delegated_children: sumOptional("delegated_children"),
+        max_concurrency: maxValues.length ? Math.max(...maxValues) : null,
+    };
+}
+//# sourceMappingURL=metrics.js.map

package/dist/redaction.d.ts

@@ -0,0 +1,6 @@
+import type { SanitizedArgs } from "./capture-trace-types.js";
+/** Values longer than this are truncated rather than persisted whole. */
+export declare const MAX_VALUE_CHARS = 2000;
+export declare function redactText(input: string, homeDir?: string): string;
+export declare function truncate(input: string, max?: number): string;
+export declare function redactArgs(args: unknown, homeDir?: string, depth?: number): SanitizedArgs;

package/dist/redaction.js

@@ -0,0 +1,55 @@
+/** Values longer than this are truncated rather than persisted whole. */
+export const MAX_VALUE_CHARS = 2000;
+const REDACTED = "[redacted]";
+const SECRET_KEY = /^(.*[-_])?(password|passwd|secret|token|api[-_]?key|apikey|auth|authorization|credential|private[-_]?key|access[-_]?key|session[-_]?key)([-_].*)?$/i;
+const SECRET_VALUE = [
+    /-----BEGIN [A-Z ]*PRIVATE KEY-----[\s\S]*?-----END [A-Z ]*PRIVATE KEY-----/g,
+    /\bBearer\s+[A-Za-z0-9._~+/-]{16,}=*/g,
+    /\bsk-[A-Za-z0-9]{16,}\b/g,
+    /\bgh[pousr]_[A-Za-z0-9]{16,}\b/g,
+    /\bxox[abposr]-[A-Za-z0-9-]{10,}\b/g,
+    /\bAKIA[0-9A-Z]{16}\b/g,
+    /\bey[A-Za-z0-9_-]{10,}\.[A-Za-z0-9_-]{10,}\.[A-Za-z0-9_-]{10,}\b/g,
+    /\b([a-z][a-z0-9+.-]*:\/\/)[^/\s:@]+:[^/\s@]+@/gi,
+];
+export function redactText(input, homeDir) {
+    let out = input;
+    out = out.replace(SECRET_VALUE[SECRET_VALUE.length - 1], `$1${REDACTED}@`);
+    for (const re of SECRET_VALUE.slice(0, -1))
+        out = out.replace(re, REDACTED);
+    if (homeDir && homeDir.length > 1)
+        out = out.split(homeDir).join("~");
+    return out;
+}
+export function truncate(input, max = MAX_VALUE_CHARS) {
+    if (input.length <= max)
+        return input;
+    return `${input.slice(0, max)}… [truncated ${input.length - max} chars]`;
+}
+export function redactArgs(args, homeDir, depth = 0) {
+    if (args === null || typeof args !== "object" || Array.isArray(args))
+        return {};
+    const out = {};
+    for (const [key, value] of Object.entries(args)) {
+        if (SECRET_KEY.test(key)) {
+            out[key] = REDACTED;
+            continue;
+        }
+        out[key] = redactValue(value, homeDir, depth);
+    }
+    return out;
+}
+function redactValue(value, homeDir, depth) {
+    if (typeof value === "string")
+        return truncate(redactText(value, homeDir));
+    if (typeof value === "number" || typeof value === "boolean" || value === null)
+        return value;
+    if (depth >= 3)
+        return "[nested]";
+    if (Array.isArray(value))
+        return value.slice(0, 20).map((v) => redactValue(v, homeDir, depth + 1));
+    if (typeof value === "object")
+        return redactArgs(value, homeDir, depth + 1);
+    return String(value);
+}
+//# sourceMappingURL=redaction.js.map

package/dist/regate.js

@@ -4,9 +4,8 @@ import { parseVerdict, detectMisfire } from "./grade.js";
 import { evaluateNeedleGates, hasNeedleGates } from "./seeded.js";
 import { evaluateTraceGates } from "./trace-gates.js";
 import { mergeTraces, deserializeTrace } from "./execution-trace.js";
-import { deserializeTrajectoryEvents, evaluateTrajectoryGates } from "./trajectory-gates.js";
 import { judgeOneRep } from "./regrade.js";
-import { readResults, writeResults, diffPath, transcriptPath, judgeRawPath, repIndexOf, findDiffFiles, findTraceFiles, findTrajectoryFiles, tracePath, trajectoryPath, effectiveThreshold, scoreContextFor, rebuildScenarioResult, mergeScenarioMetrics, deliveryStatusForObservations, } from "./results.js";
+import { readResults, writeResults, diffPath, transcriptPath, judgeRawPath, repIndexOf, findDiffFiles, findTraceFiles, tracePath, effectiveThreshold, scoreContextFor, rebuildScenarioResult, mergeScenarioMetrics, deliveryStatusForObservations, } from "./results.js";
 import { outcomesToResult } from "./reps.js";
 import { appendJournal } from "./journal.js";
 import { gatesDigest, GATES_PREFIX } from "./sources.js";
@@ -89,8 +88,7 @@ export async function regateRun(opts) {
         const s = specById.get(rec.id);
         const needles = hasNeedleGates(s ?? {});
         const traceGated = Boolean(s?.traceAssert);
-        const trajectoryGated = Boolean(s?.trajectoryAssert);
-        if (!s || (!needles && !traceGated && !trajectoryGated))
+        if (!s || (!needles && !traceGated))
             continue; // nothing for regate to re-decide
         if (s.assert?.vitest || s.assert?.post_test) {
             blocked.push(`${s.id}: declares ${s.assert.vitest ? "assert.vitest" : "assert.post_test"}, which needs the workspace — ` +
@@ -112,16 +110,11 @@ export async function regateRun(opts) {
                 `(traces are gitignored, or capture failed) — it needs a re-run`);
             continue;
         }
-        if (trajectoryGated && !complete(findTrajectoryFiles(opts.runDir, s.id, mode), expectedSuffixes.map((rep) => trajectoryPath(opts.runDir, s.id, mode, rep)))) {
-            blocked.push(`${s.id}: assert.trajectory artifacts are incomplete for ${expectedReps} recorded rep(s) ` +
-                `(events are gitignored, or required native evidence was missing) — it needs a re-run`);
-            continue;
-        }
         targets.push(s);
     }
     if (targets.length === 0) {
         throw new Error(`nothing to regate in ${opts.runDir}` +
-            (blocked.length > 0 ? `:\n  ${blocked.join("\n  ")}` : " — no scenario declares diff_contains/diff_excludes, assert.trace, or assert.trajectory"));
+            (blocked.length > 0 ? `:\n  ${blocked.join("\n  ")}` : " — no scenario declares diff_contains/diff_excludes or assert.trace"));
     }
     const changes = [];
     let judgeCalls = 0;
@@ -134,15 +127,9 @@ export async function regateRun(opts) {
         }
         const diffFiles = findDiffFiles(opts.runDir, scenario.id, mode);
         const traceFiles = findTraceFiles(opts.runDir, scenario.id, mode);
-        const eventFiles = findTrajectoryFiles(opts.runDir, scenario.id, mode);
-        // Reps come from whichever artifact this scenario actually has. A trajectory-only
-        // scenario has neither diff nor trace, so falling back only once would regate
-        // nothing and report success.
         const repKeys = diffFiles.length > 0
             ? diffFiles.map((f) => ({ rep: repIndexOf(f) ?? undefined, diffFile: f }))
-            : traceFiles.length > 0
-                ? traceFiles.map((f) => ({ rep: repIndexOf(f) ?? undefined, diffFile: undefined }))
-                : eventFiles.map((f) => ({ rep: repIndexOf(f) ?? undefined, diffFile: undefined }));
+            : traceFiles.map((f) => ({ rep: repIndexOf(f) ?? undefined, diffFile: undefined }));
         const outcomes = [];
         // Per scenario, not run-wide: with several regated scenarios, a global counter
         // would report every change as "re-judged" because some other scenario was.
@@ -192,61 +179,9 @@ export async function regateRun(opts) {
                     }
                 }
             }
-            // Adapter-neutral trajectory gate, replayed from the saved normalized
-            // event artifact. Native ledgers are deliberately not re-read here: the
-            // artifact is the immutable evidence this run actually captured.
-            let trajectoryFailure = null;
-            if (scenario.trajectoryAssert) {
-                const ep = trajectoryPath(opts.runDir, scenario.id, mode, rep);
-                const events = existsSync(ep) ? deserializeTrajectoryEvents(readFileSync(ep, "utf8")) : null;
-                // The normalized artifact contains only events that were successfully captured.
-                // Source-read/normalization failures are persisted on the original objective;
-                // dropping them here would let a partial ledger become PASS on replay.
-                const priorEvidenceErrors = (rec.objective?.assertions ?? [])
-                    .filter((result) => result.kind === "trajectory_evidence" && result.status === "ERROR");
-                if (!events) {
-                    trajectoryFailure = "objective: saved normalized events are missing, malformed, or from an unsupported version";
-                    objective = {
-                        ...(objective ?? { assertions: [] }),
-                        status: "ERROR",
-                        assertions: [...(objective?.assertions ?? []), { kind: "trajectory_evidence", status: "ERROR", detail: trajectoryFailure }],
-                    };
-                }
-                else {
-                    const g = evaluateTrajectoryGates(scenario.trajectoryAssert, events);
-                    const expectedHash = rec.objective?.rep_events_sha256?.[rep ?? 0] ?? (rec.reps === undefined ? rec.objective?.events_sha256 : undefined);
-                    const digestMismatch = expectedHash !== undefined && expectedHash !== g.events_sha256;
-                    const integrityAssertion = digestMismatch
-                        ? [{ kind: "trajectory_evidence", status: "ERROR", detail: "saved normalized events no longer match the hash recorded by the run" }]
-                        : [];
-                    const priorStatus = objective?.status ?? "PASS";
-                    const status = priorStatus === "ERROR" || g.status === "ERROR" || priorEvidenceErrors.length || digestMismatch
-                        ? "ERROR"
-                        : priorStatus === "NOT-MEASURED"
-                            ? "NOT-MEASURED"
-                            : priorStatus === "FAIL" || g.status === "FAIL"
-                                ? "FAIL"
-                                : "PASS";
-                    objective = {
-                        ...(objective ?? {}),
-                        status,
-                        trajectory_version: g.event_version,
-                        events_sha256: g.events_sha256,
-                        assertions: [...(objective?.assertions ?? []), ...priorEvidenceErrors, ...integrityAssertion, ...g.assertions],
-                    };
-                    if (g.status !== "PASS" || priorEvidenceErrors.length || digestMismatch) {
-                        const details = [
-                            ...priorEvidenceErrors.map((result) => result.detail),
-                            ...integrityAssertion.map((result) => result.detail),
-                            ...g.assertions.filter((result) => result.status === g.status && g.status !== "PASS").map((result) => result.detail),
-                        ];
-                        trajectoryFailure = `objective: ${details.join("; ")}`;
-                    }
-                }
-            }
             // Prompt delivery was observed during the subject invocation and cannot be
             // replayed by regate. Carry its objective finding while refreshing only
-            // trace/trajectory/needle assertions.
+            // trace/needle assertions.
             const retainedDelivery = rec.objective?.assertions.find(assertion => assertion.kind === "skill_delivered");
             const repObservations = prev.subject_invocations?.filter(observation => observation.scenario_id === scenario.id && observation.repetition === (rep ?? 0)) ?? [];
             const delivery = prev.schema === 3
@@ -259,7 +194,7 @@ export async function regateRun(opts) {
                 if (delivery.status !== "PASS")
                     traceFailure = `objective: ${delivery.detail}`;
             }
-            const gate = { lines: needleGate.lines, failure: needleGate.failure ?? traceFailure ?? trajectoryFailure };
+            const gate = { lines: needleGate.lines, failure: needleGate.failure ?? traceFailure };
             const tPath = transcriptPath(opts.runDir, scenario.id, mode, rep);
             const before = existsSync(tPath) ? readFileSync(tPath, "utf8") : "";
             // Two sources, because the two gate kinds record their prior state

package/dist/regrade.d.ts

@@ -1,6 +1,6 @@
 import type { Spec, Scenario } from "./spec.js";
 import type { HarnessAdapter, ModelRef } from "./adapters/types.js";
-import { type Judgment, type ScenarioResult, type ResultsFile } from "./results.js";
+import { type ScenarioResult, type ResultsFile } from "./results.js";
 import { type RepOutcome } from "./reps.js";
 /**
  * Carry a run's recorded hashes forward, refreshing only the `rubric:` keys this
@@ -84,6 +84,4 @@ export interface RegradeRunOptions {
  * results.yaml, emits the `score` journal event, and returns the new
  * ResultsFile. Shared by `cmdGrade` and the pi-extension's `judge` command.
  */
-/** Preserve prior panel votes and append a full-cell grade without writing invalid one-vote history. */
-export declare function appendJudgeHistory(prior: ScenarioResult | undefined, priorJudge: ModelRef | undefined, fresh: Omit<Judgment, "ordinal">): Judgment[] | undefined;
 export declare function regradeRun(opts: RegradeRunOptions): Promise<ResultsFile>;