dsh-continual-evolve 0.1.1 → 0.3.0

package/lib/llm-text.js

@@ -0,0 +1,49 @@
+import { BlockAssembler, createUserMessage, ReasoningEffortId } from "@deepseek-ai/dsh-llm";
+/**
+ * Stream a single-turn text completion through `ctx.llm`. Forces
+ * `reasoningEffort: off` so the model spends its budget on the answer,
+ * not visible thinking (reasoning models otherwise produce zero text
+ * blocks — the exact failure recorded in FAQ #7).
+ *
+ * @returns The concatenated text blocks from the response.
+ * @throws On provider error, abort, max-token truncation, or empty output.
+ */
+export async function streamText(ctx, opts) {
+    const assembler = new BlockAssembler();
+    for await (const chunk of ctx.llm.stream({
+        provider: opts.provider,
+        model: opts.model,
+        system: opts.system,
+        messages: [
+            createUserMessage({
+                content: [{ type: "text", text: opts.prompt }],
+                source: { kind: "plugin", plugin: "dsh-continual-evolve" },
+            }),
+        ],
+        reasoningEffort: ReasoningEffortId("off"),
+        maxTokens: opts.maxTokens ?? 8000,
+        ...(opts.signal ? { signal: opts.signal } : {}),
+    })) {
+        assembler.push(chunk);
+    }
+    const finish = assembler.finish;
+    if (finish.kind === "error") {
+        throw new Error(`evolve: LLM call failed: ${finish.failure?.message ?? "unknown"}`);
+    }
+    if (finish.kind === "aborted") {
+        throw new Error("evolve: LLM call aborted");
+    }
+    if (finish.kind === "max-tokens") {
+        throw new Error("evolve: LLM 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: LLM produced no text output");
+    }
+    return text;
+}
+//# sourceMappingURL=llm-text.js.map

package/lib/mount-command.d.ts

@@ -0,0 +1,10 @@
+/**
+ * The `/evolve mount` and `/evolve unmount` subcommand handlers.
+ * Extracted from command.ts (P2-2).
+ */
+import type { Context } from "@deepseek-ai/cordis";
+import type { CommandInvocation, CommandResult } from "@deepseek-ai/dsh-commands";
+import type { EvolutionEngine } from "./service.js";
+export declare function executeMountCommand(ctx: Context, engine: EvolutionEngine, invocation: CommandInvocation, rest: string[]): Promise<CommandResult>;
+export declare function executeUnmountCommand(ctx: Context, engine: EvolutionEngine, rest: string[]): Promise<CommandResult>;
+//# sourceMappingURL=mount-command.d.ts.map

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/mount.js

@@ -124,6 +124,11 @@ export function renderParameters(entry) {
  * the ledger records the entry for the next boot.
  */
 export async function mountSkill(ctx, baseDir, entry) {
+    // Guidance skills are SKILL.md documents with no python reference — there
+    // is no function to mount as a tool; only executable skills can hot-mount.
+    if (entry.skill_kind === "guidance" || Object.keys(entry.reference ?? {}).length === 0) {
+        throw new Error(`skill ${entry.id} has no python reference (guidance skills cannot be mounted — load them with the skill tool instead)`);
+    }
     const dir = renderMountPackage(baseDir, entry);
     const entryId = `evolve-mount-${skillNameOf(entry.id)}`;
     const loader = ctx.get("loader");

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- 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      \"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;
@@ -22,6 +22,13 @@ export interface PlanOptions {
      * every planning call is grounded in what the user actually said.
      */
     trajectory?: string;
+    /**
+     * Skills root to read the skill-creator template facts from
+     * (`<root>/skill-creator/references/template.md`). When omitted or the
+     * skills are not installed, the builtin distilled quality guide is
+     * injected instead — the skill standard is always present.
+     */
+    skillsRoot?: string;
     global?: boolean;
     signal?: AbortSignal;
     maxOutputTokens?: number;

package/lib/planner.js

@@ -1,7 +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
@@ -14,7 +15,32 @@ Rules:
 - prompt = narrow behavioral policy addendums; memory = durable facts/preferences/failures;
   skill = repeatable procedures (must carry a python reference {type:"python", import, callable}
   and an arguments object); subagent = reusable delegation roles.
+- Skill entries are authored to the DSH skill quality standard
+  (skill-creator, distilled from the official deepseek-harness 11 skills;
+  the full facts are provided in the <skill_quality_standard> block below):
+  only for a REAL trigger scenario grounded in the trajectory
+  (who, in what real task, what signal) — never invent one to pad the store;
+  never duplicate the official 11 skills or existing entries; content is a
+  SKILL.md document (frontmatter routing with "use when / do not use when"
+  description, boundary declaration, prerequisites and exclusions, layered
+  information, verifiable completion criteria). Self-check every proposed
+  skill against the 7 structural features and state the result in its
+  reason field.
+- Repeated multi-step workflows (session start/end routines, recurring
+  wrap-up or handoff procedures) may be proposed as guidance skills:
+  kind=skill, skill_kind="guidance", content = a SKILL.md document (no
+  python reference — executable skills keep requiring reference +
+  arguments). Only propose with repeated evidence in the trajectory, never
+  for one-off flows. Guidance skills materialize as discoverable SKILL.md
+  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.
@@ -37,8 +63,10 @@ Rules:
       "path": "optional grouping path",
       "reference": {"type":"python","import":"pkg.mod","callable":"fn"} ,
       "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"
     }
   ]
 }`;
@@ -54,57 +82,30 @@ export async function planWithLlm(ctx, options) {
     // most recent direct user messages ("" when none qualify — the block is
     // then omitted entirely, keeping an empty trajectory zero-cost).
     const trajectory = options.trajectory ?? recentUserText(agent);
+    // The skill quality standard is always present: the skill-creator
+    // template facts when installed, the builtin distilled guide otherwise
+    // (~1KB — planning is low-frequency, and the standard keeps skill
+    // proposals from drifting off the quality bar).
+    const qualityGuide = skillQualityGuide(options.skillsRoot);
     const userPrompt = [
         `<current_harness_state>\n${formatHarnessStateForPrompt(state)}\n</current_harness_state>`,
         `<refinement_history>\n${historyForPrompt(history)}\n</refinement_history>`,
         `<scope_policy>\n${scopeInstruction}\n</scope_policy>`,
         trajectory ? `<session_trajectory>\n${trajectory}\n</session_trajectory>` : "",
+        `<skill_quality_standard>\n${qualityGuide.text}\n</skill_quality_standard>`,
         options.instructions ? `<user_instructions>\n${options.instructions}\n</user_instructions>` : "",
         "Return only JSON edits. If no useful edit is justified, return an empty edits array with a rationale.",
     ]
         .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

@@ -18,7 +18,8 @@ export function entryLine(entry, maxContentLength) {
         : "";
     const citationText = citationSuffix(entry);
     const archivedText = isArchived(entry) ? " [archived]" : "";
-    return `- [${entry.scope}:${entry.id}] ${entry.title} (${entry.path}, v${entry.version})${archivedText}${referenceText}${argumentsText}${citationText}: ${compactText(entry.content, maxContentLength)}`;
+    const formText = entry.kind === "skill" && entry.skill_kind === "guidance" ? " [guidance]" : "";
+    return `- [${entry.scope}:${entry.id}] ${entry.title} (${entry.path}, v${entry.version})${archivedText}${formText}${referenceText}${argumentsText}${citationText}: ${compactText(entry.content, maxContentLength)}`;
 }
 /** Trajectory citation suffix (` src=sessionId:1,2`), empty when uncited. */
 function citationSuffix(entry) {
@@ -76,8 +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;
@@ -27,8 +27,11 @@ 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\nReturn JSON only:\n{\n  \"shouldRefine\": true|false,\n  \"rationale\": \"short reason\",\n  \"instructions\": \"optional concise instructions for /evolve if shouldRefine is true\"\n}";
+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. */
 export declare function parseAutoRefineReview(text: string): AutoRefineReview;
 /** Serialize surface events to bounded role-prefixed text. */

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
@@ -22,6 +22,23 @@ refine target: approve with instructions naming the entry ids, and tell the
 planner to archive them (archive hides from injection, data stays restorable)
 rather than delete.
 
+Skill-related trajectories (the evidence concerns creating or improving a
+skill entry) are judged against the DSH skill quality standard (skill-audit
+dimensions: frontmatter routing, the 7 structural features, paragraph
+skeleton, no duplication of the official 11 skills or covered skills).
+Approve only when the trajectory shows a REAL trigger scenario and the
+resulting skill would meet the standard; otherwise decline and say in the
+rationale what must improve — drafting follows skill-creator, and the
+planner receives the standard as its <skill_quality_standard> block.
+
+Repeated multi-step workflows (session start/end routines, recurring
+wrap-up or handoff procedures) are a valid refine target: approve with
+instructions telling the planner to propose a guidance skill (kind=skill,
+skill_kind=guidance — a SKILL.md document, no python reference). Only
+propose when the same workflow recurs in the trajectory — never for
+one-off flows. Auto-created skills are always offered to the user for a
+decision before they land; the gate never writes a skill silently.
+
 Return JSON only:
 {
   "shouldRefine": true|false,
@@ -71,7 +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) {
@@ -84,44 +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

@@ -9,12 +9,30 @@ export interface AggregateOptions {
     passThreshold: number;
     /** Per-case regression tolerance: candidate may drop below reference by at most this much. */
     regressionTolerance: number;
+    /**
+     * Failure-cell protocol (gap A2): a round with more failed cells than
+     * this is rejected outright — failed cells are NEVER averaged in as
+     * zeros. Default 0 (any failure rejects the round).
+     */
+    maxFailedCells: number;
 }
 export declare const DEFAULT_AGGREGATE: AggregateOptions;
-/** Aggregate raw cells into code-owned per-case means + overall mean. */
-export declare function aggregate(cells: readonly CellScore[]): Record<string, number | null> & {
+export interface AggregateResult extends Record<string, number | null> {
     overall: number | null;
-};
+    /** Count of failed cells (excluded from every mean). */
+    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.
+ * Failure-cell protocol (gap A2): failed cells are EXCLUDED from means and
+ * counted separately — a crashed unit can never silently drag the mean down
+ * like a zero. A case whose cells all failed reports null (no mean).
+ */
+export declare function aggregate(cells: readonly CellScore[]): AggregateResult;
 export declare function entryFromCells(label: string, cells: readonly CellScore[], refinementId?: string): EvaluationEntry;
 export interface Decision {
     accepted: boolean;
@@ -22,10 +40,25 @@ 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
- * reference AND no case regresses by more than `regressionTolerance` points.
+ * reference, no case regresses by more than `regressionTolerance` points,
+ * and neither side has more failed cells than `maxFailedCells` (failure-cell
+ * protocol, gap A2 — a partial/invalid round is never accepted).
  */
 export declare function decide(reference: EvaluationEntry, candidate: EvaluationEntry, opts: AggregateOptions): Decision;
 //# sourceMappingURL=score.d.ts.map