dsh-continual-evolve 0.2.0 → 0.3.0

package/lib/mount-command.js

@@ -0,0 +1,48 @@
+import { loadLedger, mountSkill, unmountSkill } from "./mount.js";
+import { stripAngleBrackets } from "./command.js";
+function success(text) {
+    return { kind: "success", text };
+}
+function error(text) {
+    return { kind: "error", text };
+}
+export async function executeMountCommand(ctx, engine, invocation, rest) {
+    const sub = rest[0] ?? "";
+    if (sub === "list") {
+        const ledger = loadLedger(engine.baseDir);
+        if (ledger.mounted.length === 0) {
+            return success("(no hot-mounted plugins — /evolve mount <skillId>)");
+        }
+        return success(ledger.mounted.map((m) => `- ${m.id} (${m.entryId}, v${m.version}, ${m.mountedAt})`).join("\n"));
+    }
+    const skillId = stripAngleBrackets(sub);
+    if (!skillId) {
+        return error(`mount requires a skill entry id.\nUsage: /evolve mount <skillId> | /evolve mount list`);
+    }
+    const sessionId = invocation.agent.id;
+    const local = engine.load("local", sessionId);
+    const globalState = engine.load("global", undefined);
+    const entry = local.entries.skill[skillId] ??
+        globalState.entries.skill[skillId] ??
+        Object.values(local.entries.skill).find((e) => e.id === skillId) ??
+        Object.values(globalState.entries.skill).find((e) => e.id === skillId);
+    if (!entry) {
+        return error(`skill entry ${skillId} not found in local or global store`);
+    }
+    try {
+        const record = await mountSkill(ctx, engine.baseDir, entry);
+        return success(`mounted ${record.id} as ${record.entryId} (v${record.version}) — tool: skill_${record.id.replace(/_/g, "-")}`);
+    }
+    catch (cause) {
+        return error(cause instanceof Error ? cause.message : String(cause));
+    }
+}
+export async function executeUnmountCommand(ctx, engine, rest) {
+    const id = stripAngleBrackets(rest[0] ?? "");
+    if (!id) {
+        return error(`unmount requires a mount id (see /evolve mount list).`);
+    }
+    const record = await unmountSkill(ctx, engine.baseDir, id);
+    return record ? success(`unmounted ${record.id} (${record.entryId})`) : error(`no mount found for ${id}`);
+}
+//# sourceMappingURL=mount-command.js.map

package/lib/plan.js

@@ -99,6 +99,11 @@ export function parseProposal(text) {
             assignIfString(built, "content", edit["content"]);
             assignIfString(built, "path", edit["path"]);
             assignIfString(built, "reason", edit["reason"]);
+            // Gap C2: validate blastRadius values.
+            const br = asString(edit["blastRadius"]);
+            if (br === "general" || br === "project" || br === "session") {
+                built.blastRadius = br;
+            }
             const reference = asRecord(edit["reference"]);
             if (reference)
                 built.reference = reference;

package/lib/planner.d.ts

@@ -9,7 +9,7 @@
 import type { Context } from "@deepseek-ai/cordis";
 import type { Agent } from "@deepseek-ai/dsh-agent";
 import type { HarnessState, RefinementProposal, RefinementResult } from "./types.js";
-export declare const PLANNER_SYSTEM_PROMPT = "You are the /evolve continual harness subsystem.\n\nYour job is to improve the editable continual harness state. Instead of\nsummarizing the conversation you emit precise Create, Update, or Delete edits\nto reusable state: prompt notes, memories, skills, and subagent specs.\n\nRules:\n- The base system prompt is immutable and MUST NOT be rewritten (never edit id \"base_system_prompt\").\n- Prefer small evidence-backed edits. If no useful edit is justified, return an empty edits array.\n- prompt = narrow behavioral policy addendums; memory = durable facts/preferences/failures;\n  skill = repeatable procedures (must carry a python reference {type:\"python\", import, callable}\n  and an arguments object); subagent = reusable delegation roles.\n- Skill entries are authored to the DSH skill quality standard\n  (skill-creator, distilled from the official deepseek-harness 11 skills;\n  the full facts are provided in the <skill_quality_standard> block below):\n  only for a REAL trigger scenario grounded in the trajectory\n  (who, in what real task, what signal) \u2014 never invent one to pad the store;\n  never duplicate the official 11 skills or existing entries; content is a\n  SKILL.md document (frontmatter routing with \"use when / do not use when\"\n  description, boundary declaration, prerequisites and exclusions, layered\n  information, verifiable completion criteria). Self-check every proposed\n  skill against the 7 structural features and state the result in its\n  reason field.\n- Repeated multi-step workflows (session start/end routines, recurring\n  wrap-up or handoff procedures) may be proposed as guidance skills:\n  kind=skill, skill_kind=\"guidance\", content = a SKILL.md document (no\n  python reference \u2014 executable skills keep requiring reference +\n  arguments). Only propose with repeated evidence in the trajectory, never\n  for one-off flows. Guidance skills materialize as discoverable SKILL.md\n  files under <skillsRoot>/<kebab-name>/SKILL.md and are always offered to\n  the user for a decision before they land.\n- Local edits are session-scoped; global edits persist across sessions.\n- Ground every edit in evidence: the session trajectory (recent direct user\n  messages) is provided when available; prefer edits backed by it over\n  speculation, and never invent preferences the user did not express.\n- Stale entries (superseded by newer ones, never referenced in recent\n  trajectories, obsolete facts): propose action \"archive\" instead of\n  \"delete\" \u2014 archive hides the entry from injection while keeping its data\n  restorable; it requires only kind + id.\n- Output JSON only, exactly this shape:\n{\n  \"summary\": \"one sentence\",\n  \"rationale\": \"why these edits are justified by the evidence\",\n  \"expectedOutcome\": \"what should improve and how to validate it\",\n  \"edits\": [\n    {\n      \"action\": \"create|update|delete\",\n      \"kind\": \"prompt|memory|skill|subagent\",\n      \"id\": \"stable id for update/delete, optional for create\",\n      \"title\": \"required for create/update except delete\",\n      \"content\": \"required for create/update except delete\",\n      \"path\": \"optional grouping path\",\n      \"reference\": {\"type\":\"python\",\"import\":\"pkg.mod\",\"callable\":\"fn\"} ,\n      \"arguments\": {\"name\": {\"type\":\"string\",\"required\":true,\"description\":\"...\"}},\n      \"skill_kind\": \"executable|guidance (optional; skill kind only \u2014 guidance = SKILL.md document without reference)\",\n      \"metadata\": {},\n      \"reason\": \"why this edit is useful\"\n    }\n  ]\n}";
+export declare const PLANNER_SYSTEM_PROMPT = "You are the /evolve continual harness subsystem.\n\nYour job is to improve the editable continual harness state. Instead of\nsummarizing the conversation you emit precise Create, Update, or Delete edits\nto reusable state: prompt notes, memories, skills, and subagent specs.\n\nRules:\n- The base system prompt is immutable and MUST NOT be rewritten (never edit id \"base_system_prompt\").\n- Prefer small evidence-backed edits. If no useful edit is justified, return an empty edits array.\n- prompt = narrow behavioral policy addendums; memory = durable facts/preferences/failures;\n  skill = repeatable procedures (must carry a python reference {type:\"python\", import, callable}\n  and an arguments object); subagent = reusable delegation roles.\n- Skill entries are authored to the DSH skill quality standard\n  (skill-creator, distilled from the official deepseek-harness 11 skills;\n  the full facts are provided in the <skill_quality_standard> block below):\n  only for a REAL trigger scenario grounded in the trajectory\n  (who, in what real task, what signal) \u2014 never invent one to pad the store;\n  never duplicate the official 11 skills or existing entries; content is a\n  SKILL.md document (frontmatter routing with \"use when / do not use when\"\n  description, boundary declaration, prerequisites and exclusions, layered\n  information, verifiable completion criteria). Self-check every proposed\n  skill against the 7 structural features and state the result in its\n  reason field.\n- Repeated multi-step workflows (session start/end routines, recurring\n  wrap-up or handoff procedures) may be proposed as guidance skills:\n  kind=skill, skill_kind=\"guidance\", content = a SKILL.md document (no\n  python reference \u2014 executable skills keep requiring reference +\n  arguments). Only propose with repeated evidence in the trajectory, never\n  for one-off flows. Guidance skills materialize as discoverable SKILL.md\n  files under <skillsRoot>/<kebab-name>/SKILL.md and are always offered to\n  the user for a decision before they land.\n- Local edits are session-scoped; global edits persist across sessions.\n- Gap C2: EVERY edit MUST include a \"blastRadius\" field indicating how broadly\n  the edit applies: \"general\" (cross-project tactical rule, applies everywhere),\n  \"project\" (valid for the current project/repo), or \"session\" (one-off,\n  specific to this session's context). The review gate validates that\n  local-scope edits use \"session\" or \"project\", and global-scope edits use\n  \"general\" or \"project\". When in doubt, prefer narrower blast radius.\n- Ground every edit in evidence: the session trajectory (recent direct user\n  messages) is provided when available; prefer edits backed by it over\n  speculation, and never invent preferences the user did not express.\n- Stale entries (superseded by newer ones, never referenced in recent\n  trajectories, obsolete facts): propose action \"archive\" instead of\n  \"delete\" \u2014 archive hides the entry from injection while keeping its data\n  restorable; it requires only kind + id.\n- Output JSON only, exactly this shape:\n{\n  \"summary\": \"one sentence\",\n  \"rationale\": \"why these edits are justified by the evidence\",\n  \"expectedOutcome\": \"what should improve and how to validate it\",\n  \"edits\": [\n    {\n      \"action\": \"create|update|delete\",\n      \"kind\": \"prompt|memory|skill|subagent\",\n      \"id\": \"stable id for update/delete, optional for create\",\n      \"title\": \"required for create/update except delete\",\n      \"content\": \"required for create/update except delete\",\n      \"path\": \"optional grouping path\",\n      \"reference\": {\"type\":\"python\",\"import\":\"pkg.mod\",\"callable\":\"fn\"} ,\n      \"arguments\": {\"name\": {\"type\":\"string\",\"required\":true,\"description\":\"...\"}},\n      \"skill_kind\": \"executable|guidance (optional; skill kind only \u2014 guidance = SKILL.md document without reference)\",\n      \"metadata\": {},\n      \"reason\": \"why this edit is useful\",\n      \"blastRadius\": \"general|project|session\"\n    }\n  ]\n}";
 export interface PlanOptions {
     agent: Agent;
     state: HarnessState;

package/lib/planner.js

@@ -1,8 +1,8 @@
-import { BlockAssembler, createUserMessage, ReasoningEffortId } from "@deepseek-ai/dsh-llm";
 import { parseProposal } from "./plan.js";
 import { formatHarnessStateForPrompt, historyForPrompt } from "./render.js";
 import { recentUserText } from "./inject.js";
 import { skillQualityGuide } from "./skillquality.js";
+import { streamText } from "./llm-text.js";
 export const PLANNER_SYSTEM_PROMPT = `You are the /evolve continual harness subsystem.
 
 Your job is to improve the editable continual harness state. Instead of
@@ -35,6 +35,12 @@ Rules:
   files under <skillsRoot>/<kebab-name>/SKILL.md and are always offered to
   the user for a decision before they land.
 - Local edits are session-scoped; global edits persist across sessions.
+- Gap C2: EVERY edit MUST include a "blastRadius" field indicating how broadly
+  the edit applies: "general" (cross-project tactical rule, applies everywhere),
+  "project" (valid for the current project/repo), or "session" (one-off,
+  specific to this session's context). The review gate validates that
+  local-scope edits use "session" or "project", and global-scope edits use
+  "general" or "project". When in doubt, prefer narrower blast radius.
 - Ground every edit in evidence: the session trajectory (recent direct user
   messages) is provided when available; prefer edits backed by it over
   speculation, and never invent preferences the user did not express.
@@ -59,7 +65,8 @@ Rules:
       "arguments": {"name": {"type":"string","required":true,"description":"..."}},
       "skill_kind": "executable|guidance (optional; skill kind only — guidance = SKILL.md document without reference)",
       "metadata": {},
-      "reason": "why this edit is useful"
+      "reason": "why this edit is useful",
+      "blastRadius": "general|project|session"
     }
   ]
 }`;
@@ -91,47 +98,14 @@ export async function planWithLlm(ctx, options) {
     ]
         .filter(Boolean)
         .join("\n\n");
-    const assembler = new BlockAssembler();
-    for await (const chunk of ctx.llm.stream({
+    const text = await streamText(ctx, {
         provider: agent.options.provider,
         model: agent.options.model,
         system: PLANNER_SYSTEM_PROMPT,
-        messages: [
-            createUserMessage({
-                content: [{ type: "text", text: userPrompt }],
-                source: { kind: "plugin", plugin: "dsh-continual-evolve" },
-            }),
-        ],
-        // Force non-reasoning output: the proposal must be pure JSON text.
-        reasoningEffort: ReasoningEffortId("off"),
+        prompt: userPrompt,
         maxTokens: options.maxOutputTokens ?? 8000,
-        ...(options.signal ? { signal: options.signal } : {}),
-    })) {
-        assembler.push(chunk);
-    }
-    throwOnFinishError(assembler.finish);
-    const blocks = assembler.blocks();
-    const text = blocks
-        .filter((block) => block.type === "text")
-        .map((block) => block.text)
-        .join("\n");
-    if (text.length === 0) {
-        throw new Error("evolve: planner produced no text output");
-    }
+        signal: options.signal,
+    });
     return parseProposal(text);
 }
-/** Surface terminal stream states as errors so the caller never sees a silent partial plan. */
-function throwOnFinishError(finish) {
-    switch (finish.kind) {
-        case "stop":
-        case "tool-calls":
-            return;
-        case "max-tokens":
-            throw new Error("evolve: planner output budget exhausted (max-tokens)");
-        case "aborted":
-            throw new Error("evolve: planner call aborted");
-        case "error":
-            throw new Error(`evolve: planner call failed: ${finish.failure?.message ?? "unknown error"}`);
-    }
-}
 //# sourceMappingURL=planner.js.map

package/lib/render.d.ts

@@ -3,13 +3,11 @@
  * refinement history. Every cap exists to keep token cost predictable no
  * matter how large the store grows.
  */
-import type { HarnessEntry, HarnessRefinementEvent, HarnessState, RefinementResult } from "./types.js";
+import type { HarnessEntry, HarnessState, RefinementResult } from "./types.js";
 export declare function compactText(text: string, maxLength: number): string;
 export declare function entryLine(entry: HarnessEntry, maxContentLength: number): string;
 /** Render the full merged state as a bounded overview for the system prompt. */
 export declare function formatHarnessStateForPrompt(state: HarnessState): string;
 /** Render recent refinement results for the planner. */
 export declare function historyForPrompt(history: readonly RefinementResult[]): string;
-/** Serialize a refinement event for persistence (lightweight). */
-export declare function eventToLine(event: HarnessRefinementEvent): string;
 //# sourceMappingURL=render.d.ts.map

package/lib/render.js

@@ -77,8 +77,4 @@ export function historyForPrompt(history) {
     })
         .join("\n\n");
 }
-/** Serialize a refinement event for persistence (lightweight). */
-export function eventToLine(event) {
-    return JSON.stringify(event);
-}
 //# sourceMappingURL=render.js.map

package/lib/review.d.ts

@@ -13,7 +13,7 @@ export interface AutoRefineReview {
     rationale: string;
     instructions?: string;
 }
-export type AutoRefineReason = "turn_interval" | "compact";
+export type AutoRefineReason = "turn_interval" | "compact" | "goal_blocked";
 export interface AutoRefineReviewContext {
     reason: AutoRefineReason;
     turnsSinceLastReview: number;
@@ -27,6 +27,9 @@ export interface ReviewOptions {
     trajectory?: string;
     signal?: AbortSignal;
     budgetTokens?: number;
+    /** Gap C1: optional provider/model override for the review gate (cheaper model). */
+    overrideProvider?: string;
+    overrideModel?: string;
 }
 export declare const AUTO_REVIEW_SYSTEM_PROMPT = "You are the automatic /evolve review gate.\n\nDecide whether this checkpoint should run /evolve. Auto /evolve writes local\nharness state by default, so approve when the trajectory contains evidence\nuseful to this session's future turns: a repeated failure, a reusable tactic,\na repeated delegation role, a durable fact or preference, a user correction\nthat should persist, or a narrow behavioral policy.\n\nThe current harness state below includes GLOBAL entries (scope=global) plus\nthis session's local entries (scope=local). When a topic is already covered\nby a global entry, do NOT approve a local duplicate of it \u2014 decline and say\nin the rationale that the topic is already covered globally.\n\nReject one-off noise, unsupported hypotheses, transient tool outputs, and\nrequests that carry no reusable content.\n\nStale local entries (superseded, long-unused, obsolete facts) are a valid\nrefine target: approve with instructions naming the entry ids, and tell the\nplanner to archive them (archive hides from injection, data stays restorable)\nrather than delete.\n\nSkill-related trajectories (the evidence concerns creating or improving a\nskill entry) are judged against the DSH skill quality standard (skill-audit\ndimensions: frontmatter routing, the 7 structural features, paragraph\nskeleton, no duplication of the official 11 skills or covered skills).\nApprove only when the trajectory shows a REAL trigger scenario and the\nresulting skill would meet the standard; otherwise decline and say in the\nrationale what must improve \u2014 drafting follows skill-creator, and the\nplanner receives the standard as its <skill_quality_standard> block.\n\nRepeated multi-step workflows (session start/end routines, recurring\nwrap-up or handoff procedures) are a valid refine target: approve with\ninstructions telling the planner to propose a guidance skill (kind=skill,\nskill_kind=guidance \u2014 a SKILL.md document, no python reference). Only\npropose when the same workflow recurs in the trajectory \u2014 never for\none-off flows. Auto-created skills are always offered to the user for a\ndecision before they land; the gate never writes a skill silently.\n\nReturn JSON only:\n{\n  \"shouldRefine\": true|false,\n  \"rationale\": \"short reason\",\n  \"instructions\": \"optional concise instructions for /evolve if shouldRefine is true\"\n}";
 /** Parse the gate's JSON reply. */

package/lib/review.js

@@ -1,6 +1,6 @@
-import { BlockAssembler, createUserMessage, ReasoningEffortId } from "@deepseek-ai/dsh-llm";
 import { extractJsonObject } from "./plan.js";
 import { formatHarnessStateForPrompt, historyForPrompt } from "./render.js";
+import { streamText } from "./llm-text.js";
 export const AUTO_REVIEW_SYSTEM_PROMPT = `You are the automatic /evolve review gate.
 
 Decide whether this checkpoint should run /evolve. Auto /evolve writes local
@@ -88,7 +88,9 @@ export function serializeSurface(events, maxChars) {
 }
 export async function reviewAutoRefine(ctx, options) {
     const { agent, state, history } = options;
-    if (!agent.options.provider || !agent.options.model) {
+    const provider = options.overrideProvider ?? agent.options.provider;
+    const model = options.overrideModel ?? agent.options.model;
+    if (!provider || !model) {
         throw new Error("evolve: no provider/model route for the review gate");
     }
     if (!options.trajectory || options.trajectory.length === 0) {
@@ -101,44 +103,14 @@ export async function reviewAutoRefine(ctx, options) {
         `<conversation>\n${options.trajectory}\n</conversation>`,
         "Return shouldRefine=true when the trajectory contains evidence useful to this session's future turns. Prefer local edits; do not ask for global refinement here.",
     ].join("\n\n");
-    const assembler = new BlockAssembler();
-    for await (const chunk of ctx.llm.stream({
-        provider: agent.options.provider,
-        model: agent.options.model,
+    const text = await streamText(ctx, {
+        provider,
+        model,
         system: AUTO_REVIEW_SYSTEM_PROMPT,
-        messages: [
-            createUserMessage({
-                content: [{ type: "text", text: userPrompt }],
-                source: { kind: "plugin", plugin: "dsh-continual-evolve" },
-            }),
-        ],
-        // Force non-reasoning output so the model spends its budget on the JSON
-        // answer, not on visible thinking (reasoning models otherwise produce
-        // zero text blocks — the exact failure recorded in reviews.jsonl).
-        reasoningEffort: ReasoningEffortId("off"),
+        prompt: userPrompt,
         maxTokens: options.budgetTokens ?? 8000,
-        ...(options.signal ? { signal: options.signal } : {}),
-    })) {
-        assembler.push(chunk);
-    }
-    const finish = assembler.finish;
-    if (finish.kind === "error") {
-        throw new Error(`evolve: review gate call failed: ${finish.failure?.message ?? "unknown"}`);
-    }
-    if (finish.kind === "aborted") {
-        throw new Error("evolve: review gate call aborted");
-    }
-    if (finish.kind === "max-tokens") {
-        throw new Error("evolve: review gate output budget exhausted (max-tokens)");
-    }
-    const text = assembler
-        .blocks()
-        .filter((block) => block.type === "text")
-        .map((block) => block.text)
-        .join("\n");
-    if (text.length === 0) {
-        throw new Error("evolve: review gate produced no text");
-    }
+        signal: options.signal,
+    });
     return parseAutoRefineReview(text);
 }
 //# sourceMappingURL=review.js.map

package/lib/rollback.d.ts

@@ -3,9 +3,7 @@
  * result, in reverse order. Rollback is pure data transformation — no LLM
  * is asked to "guess" the previous state.
  */
-import type { HarnessEntry, RefinementProposal, RefinementResult } from "./types.js";
+import type { RefinementProposal, RefinementResult } from "./types.js";
 /** Build the inverse proposal for an applied refinement. */
 export declare function rollbackProposal(target: RefinementResult): RefinementProposal;
-/** Recreate an entry from a prior snapshot (used when an inverse edit is an update). */
-export declare function restoreEntry(prior: HarnessEntry): HarnessEntry;
 //# sourceMappingURL=rollback.d.ts.map

package/lib/rollback.js

@@ -58,12 +58,4 @@ function inverseEdit(edit, refinementId) {
     }
     return undefined;
 }
-/** Recreate an entry from a prior snapshot (used when an inverse edit is an update). */
-export function restoreEntry(prior) {
-    return {
-        ...prior,
-        updated_at: new Date().toISOString(),
-        version: prior.version + 1,
-    };
-}
 //# sourceMappingURL=rollback.js.map

package/lib/score.d.ts

@@ -23,6 +23,8 @@ export interface AggregateResult extends Record<string, number | null> {
     failed: number;
     /** Total cells considered (ok + failed). */
     total: number;
+    /** Gap C3: total wall-clock duration of all cells in milliseconds. */
+    totalDurationMs: number;
 }
 /**
  * Aggregate raw cells into code-owned per-case means + overall mean.
@@ -38,6 +40,19 @@ export interface Decision {
 }
 /** Human-readable decision report with per-case before → after deltas. */
 export declare function decisionReport(reference: EvaluationEntry, candidate: EvaluationEntry, decision: Decision): string[];
+/**
+ * Gap A3 (version_changed semantics): detect cells whose case material
+ * changed between the reference and candidate evaluation runs. A candidate
+ * cell whose `caseHash` differs from the reference cell of the SAME case
+ * means the statement/rubric was edited between the two runs — its score is
+ * not comparable to the baseline and must not count toward the decision.
+ *
+ * The check is conservative: cells without a hash on either side (pre-A3
+ * data) and cells already failed are left untouched. Mismatched cells are
+ * returned re-marked as `failed` with a reason in notes, so aggregation
+ * excludes them and the acceptance rule can reject the round.
+ */
+export declare function flagMaterialDrift(reference: EvaluationEntry, candidateCells: readonly CellScore[]): CellScore[];
 /**
  * Non-regressive acceptance rule (Self-Harness style):
  * the candidate is accepted iff its overall mean is STRICTLY higher than the

package/lib/score.js

@@ -26,11 +26,18 @@ export function aggregate(cells) {
         perCase[caseId] = mean(scores);
     }
     const all = [...byCase.values()].flat();
+    let totalDurationMs = 0;
+    for (const cell of cells) {
+        if (cell.durationMs !== undefined && cell.durationMs >= 0) {
+            totalDurationMs += cell.durationMs;
+        }
+    }
     return {
         ...perCase,
         overall: all.length > 0 ? mean(all) : null,
         failed,
         total: cells.length,
+        totalDurationMs,
     };
 }
 export function entryFromCells(label, cells, refinementId) {
@@ -47,8 +54,12 @@ export function entryFromCells(label, cells, refinementId) {
 /** Human-readable decision report with per-case before → after deltas. */
 export function decisionReport(reference, candidate, decision) {
     const lines = [`overall: ${reference.overall ?? "?"} → ${candidate.overall ?? "?"}`];
-    for (const [caseId, refScore] of Object.entries(reference.aggregate)) {
-        if (caseId === "overall" || caseId === "failed" || caseId === "total" || refScore === null)
+    // Per-case deltas over the actual evaluated cases only (aggregate() mixes
+    // case means with metadata keys like totalDurationMs — never render those
+    // as cases).
+    for (const caseId of new Set(reference.cells.map((cell) => cell.caseId))) {
+        const refScore = reference.aggregate[caseId];
+        if (refScore === null || refScore === undefined)
             continue;
         const candScore = candidate.aggregate[caseId];
         const failedMark = isCaseFailed(reference, caseId) || isCaseFailed(candidate, caseId) ? " (failed)" : "";
@@ -59,11 +70,55 @@ export function decisionReport(reference, candidate, decision) {
     if (refFailed > 0 || candFailed > 0) {
         lines.push(`failed cells: reference ${refFailed}/${reference.aggregate.total ?? 0} · candidate ${candFailed}/${candidate.aggregate.total ?? 0}`);
     }
+    // Gap C3: show duration summary when available.
+    const refDuration = reference.aggregate.totalDurationMs ?? 0;
+    const candDuration = candidate.aggregate.totalDurationMs ?? 0;
+    if (refDuration > 0 || candDuration > 0) {
+        lines.push(`duration: ${formatDuration(refDuration)} → ${formatDuration(candDuration)}`);
+    }
     lines.push(decision.accepted
         ? "DECISION: ACCEPTED — overall improved, no regression"
         : `DECISION: REJECTED — ${decision.reasons.join("; ")}`);
     return lines;
 }
+/**
+ * Gap A3 (version_changed semantics): detect cells whose case material
+ * changed between the reference and candidate evaluation runs. A candidate
+ * cell whose `caseHash` differs from the reference cell of the SAME case
+ * means the statement/rubric was edited between the two runs — its score is
+ * not comparable to the baseline and must not count toward the decision.
+ *
+ * The check is conservative: cells without a hash on either side (pre-A3
+ * data) and cells already failed are left untouched. Mismatched cells are
+ * returned re-marked as `failed` with a reason in notes, so aggregation
+ * excludes them and the acceptance rule can reject the round.
+ */
+export function flagMaterialDrift(reference, candidateCells) {
+    const referenceHashes = new Map();
+    for (const cell of reference.cells) {
+        if (cell.status !== "failed" && cell.caseHash !== undefined && !referenceHashes.has(cell.caseId)) {
+            referenceHashes.set(cell.caseId, cell.caseHash);
+        }
+    }
+    if (referenceHashes.size === 0) {
+        return [...candidateCells];
+    }
+    return candidateCells.map((cell) => {
+        if (cell.status === "failed" || cell.caseHash === undefined) {
+            return cell;
+        }
+        const refHash = referenceHashes.get(cell.caseId);
+        if (refHash !== undefined && refHash !== cell.caseHash) {
+            return {
+                ...cell,
+                status: "failed",
+                passed: false,
+                notes: `materials changed: case ${cell.caseId} hash ${cell.caseHash} ≠ reference ${refHash} (re-run the reference or fix the material)`,
+            };
+        }
+        return cell;
+    });
+}
 function isCaseFailed(entry, caseId) {
     return entry.cells.some((cell) => cell.caseId === caseId && cell.status === "failed");
 }
@@ -93,9 +148,17 @@ export function decide(reference, candidate, opts) {
     if (candidate.overall <= reference.overall) {
         reasons.push(`overall not improved: ${candidate.overall} <= ${reference.overall}`);
     }
-    for (const [caseId, refScore] of Object.entries(reference.aggregate)) {
-        if (caseId === "overall" || caseId === "failed" || caseId === "total" || refScore === null)
-            continue;
+    // Per-case regression is judged ONLY over the cases that actually exist in
+    // the evaluation — aggregate() mixes per-case means with metadata keys
+    // (failed/total/totalDurationMs), and iterating raw keys would (and did)
+    // treat totalDurationMs as a case score, rejecting a candidate whose run
+    // merely took longer. Derived from reference.cells, never from the
+    // aggregate key set.
+    for (const caseId of new Set(reference.cells.map((cell) => cell.caseId))) {
+        const refScore = reference.aggregate[caseId];
+        if (refScore === null || refScore === undefined) {
+            continue; // case had no comparable mean (all cells failed) — nothing to regress
+        }
         const candScore = candidate.aggregate[caseId];
         if (candScore === null || candScore === undefined) {
             reasons.push(`candidate missing case ${caseId}`);
@@ -119,4 +182,10 @@ function clampScore(score) {
 function round2(value) {
     return Math.round(value * 100) / 100;
 }
+/** Gap C3: human-readable duration (ms → "1.2s" or "340ms"). */
+function formatDuration(ms) {
+    if (ms < 1000)
+        return `${Math.round(ms)}ms`;
+    return `${(ms / 1000).toFixed(1)}s`;
+}
 //# sourceMappingURL=score.js.map

package/lib/service.d.ts

@@ -5,7 +5,6 @@
  */
 import type { EntrySource, HarnessScope, RefinementProposal, RefinementResult } from "./types.js";
 import { applyRefinementProposal } from "./apply.js";
-import { storePaths } from "./store.js";
 export interface ApplyContext {
     scope: HarnessScope;
     sessionId?: string;
@@ -13,6 +12,8 @@ export interface ApplyContext {
     baselineState?: Parameters<typeof applyRefinementProposal>[0];
     /** Trajectory citation stamped into newly created entries (see apply.ts). */
     source?: EntrySource | undefined;
+    /** Marks the resulting refinement as the deterministic rollback of another (audit chain). */
+    rollbackOf?: string | undefined;
 }
 export interface EvolutionHooks {
     /** Called after every applied refinement (side-effect boundary: skills sync, etc.). */
@@ -26,5 +27,4 @@ export declare function createEvolutionEngine(baseDir: string, hooks?: Evolution
     baseDir: string;
 };
 export type EvolutionEngine = ReturnType<typeof createEvolutionEngine>;
-export { storePaths };
 //# sourceMappingURL=service.d.ts.map

package/lib/service.js

@@ -17,6 +17,7 @@ export function createEvolutionEngine(baseDir, hooks = {}) {
             scope,
             ...(context?.source ? { source: context.source } : {}),
             ...(context?.baselineState ? { baselineState: context.baselineState } : {}),
+            ...(context?.rollbackOf ? { rollbackOf: context.rollbackOf } : {}),
         });
         saveHarnessState(paths.stateDir, state);
         appendResult(paths, result);
@@ -31,12 +32,14 @@ export function createEvolutionEngine(baseDir, hooks = {}) {
             throw new Error(`Refinement ${refinementId} not found in ${scope} history`);
         }
         const proposal = rollbackProposal(target);
-        return apply(scope, sessionId, proposal);
+        // The rollback refinement carries rollbackOf so the audit chain links
+        // the inverse operation back to its origin (previously the rollback
+        // record only echoed "Rollback refinement <id>" in its summary text).
+        return apply(scope, sessionId, proposal, { scope, rollbackOf: refinementId });
     }
     function history(scope, sessionId) {
         return loadResults(storePaths(baseDir, scope, sessionId));
     }
     return { load, apply, rollback, history, baseDir };
 }
-export { storePaths };
 //# sourceMappingURL=service.js.map

package/lib/skill-render.d.ts

@@ -0,0 +1,15 @@
+/**
+ * Skill rendering: pure functions that convert harness entries into
+ * SKILL.md documents. Extracted from skill.ts to break the circular
+ * dependency between skill.ts ↔ skillquality.ts.
+ *
+ * Both skill.ts (materializer) and skillquality.ts (validator) need
+ * these rendering functions; importing from this shared leaf module
+ * keeps the dependency graph acyclic.
+ */
+import type { HarnessEntry } from "./types.js";
+/** Convert a harness entry id (underscore slug) to a kebab-case skill name. */
+export declare function skillNameOf(id: string): string;
+/** Render a harness skill entry as a discoverable SKILL.md document. */
+export declare function renderSkillMarkdown(entry: HarnessEntry): string;
+//# sourceMappingURL=skill-render.d.ts.map

package/lib/skill-render.js

@@ -0,0 +1,30 @@
+/** Convert a harness entry id (underscore slug) to a kebab-case skill name. */
+export function skillNameOf(id) {
+    return id.toLowerCase().replace(/_/g, "-");
+}
+/** Render a harness skill entry as a discoverable SKILL.md document. */
+export function renderSkillMarkdown(entry) {
+    const lines = [
+        "---",
+        `name: ${skillNameOf(entry.id)}`,
+        `description: ${oneLine(entry.title)}`,
+        "---",
+        "",
+        entry.content.trim(),
+    ];
+    const reference = entry.reference;
+    if (reference && typeof reference === "object" && Object.keys(reference).length > 0) {
+        lines.push("", "## Invocation");
+        for (const [key, value] of Object.entries(reference)) {
+            lines.push(`- ${key}: ${JSON.stringify(value)}`);
+        }
+    }
+    if (Object.keys(entry.arguments).length > 0) {
+        lines.push("", "## Arguments", "```json", JSON.stringify(entry.arguments, null, 2), "```");
+    }
+    return `${lines.join("\n").trimEnd()}\n`;
+}
+function oneLine(text) {
+    return text.replace(/\s+/g, " ").trim();
+}
+//# sourceMappingURL=skill-render.js.map

package/lib/skill.d.ts

@@ -1,10 +1,7 @@
-import type { HarnessEntry, RefinementResult } from "./types.js";
-/** Convert a harness entry id (underscore slug) to a kebab-case skill name. */
-export declare function skillNameOf(id: string): string;
+import type { RefinementResult } from "./types.js";
+export { renderSkillMarkdown, skillNameOf } from "./skill-render.js";
 /** Resolve and defend the skill directory for an entry id. */
 export declare function skillDir(skillsRoot: string, id: string): string;
-/** Render a harness skill entry as a discoverable SKILL.md document. */
-export declare function renderSkillMarkdown(entry: HarnessEntry): string;
 /**
  * Apply the skill-kind edits of an applied refinement to the skills root.
  * Returns materialization warnings (rendered-SKILL.md mechanical problems

package/lib/skill.js

@@ -9,11 +9,9 @@
  */
 import { existsSync, mkdirSync, renameSync, rmSync, writeFileSync } from "node:fs";
 import { join, resolve, sep } from "node:path";
+import { renderSkillMarkdown, skillNameOf } from "./skill-render.js";
 import { skillResourceRefs, validateRenderedSkill } from "./skillquality.js";
-/** Convert a harness entry id (underscore slug) to a kebab-case skill name. */
-export function skillNameOf(id) {
-    return id.toLowerCase().replace(/_/g, "-");
-}
+export { renderSkillMarkdown, skillNameOf } from "./skill-render.js";
 /** Resolve and defend the skill directory for an entry id. */
 export function skillDir(skillsRoot, id) {
     const root = resolve(skillsRoot);
@@ -23,28 +21,6 @@ export function skillDir(skillsRoot, id) {
     }
     return dir;
 }
-/** Render a harness skill entry as a discoverable SKILL.md document. */
-export function renderSkillMarkdown(entry) {
-    const lines = [
-        "---",
-        `name: ${skillNameOf(entry.id)}`,
-        `description: ${oneLine(entry.title)}`,
-        "---",
-        "",
-        entry.content.trim(),
-    ];
-    const reference = entry.reference;
-    if (reference && typeof reference === "object" && Object.keys(reference).length > 0) {
-        lines.push("", "## Invocation");
-        for (const [key, value] of Object.entries(reference)) {
-            lines.push(`- ${key}: ${JSON.stringify(value)}`);
-        }
-    }
-    if (Object.keys(entry.arguments).length > 0) {
-        lines.push("", "## Arguments", "```json", JSON.stringify(entry.arguments, null, 2), "```");
-    }
-    return `${lines.join("\n").trimEnd()}\n`;
-}
 /**
  * Apply the skill-kind edits of an applied refinement to the skills root.
  * Returns materialization warnings (rendered-SKILL.md mechanical problems
@@ -101,7 +77,4 @@ function removeSkill(skillsRoot, id) {
         rmSync(dir, { recursive: true, force: true });
     }
 }
-function oneLine(text) {
-    return text.replace(/\s+/g, " ").trim();
-}
 //# sourceMappingURL=skill.js.map

package/lib/skillquality.d.ts

@@ -1,5 +1,4 @@
 import type { HarnessEntry } from "./types.js";
-import { skillNameOf } from "./skill.js";
 /** Relative location of the skill-creator template facts. */
 export declare const SKILL_CREATOR_TEMPLATE_REL: string;
 /**
@@ -77,5 +76,5 @@ export declare function validateRenderedSkill(entry: HarnessEntry): string[];
  */
 export declare function skillResourceRefs(content: string): string[];
 /** Kebab-case name under which the entry materializes (exported for diagnostics). */
-export { skillNameOf };
+export { skillNameOf } from "./skill-render.js";
 //# sourceMappingURL=skillquality.d.ts.map