synergyspec-selfevolving 2.0.0 → 2.1.0

package/dist/commands/learn.js

@@ -179,12 +179,19 @@ export function registerLearnCommand(program, deps = {}) {
                     changeName: report.changeName,
                     report,
                 });
+                // Thread the loop-v2 reward judge-quality config (samples / noiseFloor /
+                // orderSwap / tamperCheck). Omitted ⇒ the orchestrator's single-sample,
+                // flag-only default (no extra spawns).
+                const episodeConfig = readProjectConfig(projectRoot);
                 episodeOutcome = await runEpisodeImpl({
                     repoRoot: projectRoot,
                     targetId: concreteEvolveTarget.targetId,
                     changeName: report.changeName,
                     changeDirPath: report.changeDir,
                     mainArm,
+                    ...(episodeConfig?.selfEvolution?.reward
+                        ? { reward: episodeConfig.selfEvolution.reward }
+                        : {}),
                 });
             }
             if (options.json) {

package/dist/commands/self-evolution-episode.js

@@ -9,6 +9,7 @@ lookupCanonicalTarget, listCanonicalTargets, DESIGN_ARTIFACT_TARGET_ID, } from '
 import { generateLearnReport } from '../core/learn.js';
 import { validateExplicitTrajectoryHandle } from '../core/learn/trajectory-discovery.js';
 import { validateChangeExists } from './workflow/shared.js';
+import { readProjectConfig } from '../core/project-config.js';
 /**
  * The 主智能体 MAIN AGENT arm is graded from a learn report exactly the way the
  * `learn` command grades it (the orchestrator REUSES that grading; it never
@@ -148,6 +149,7 @@ export async function runEpisodeCommand(args, opts) {
     // orchestrator's behavior when it is unaware of the flag.
     let outcome;
     try {
+        const episodeConfig = readProjectConfig(opts.repoRoot);
         const episodeOptions = {
             repoRoot: opts.repoRoot,
             targetId,
@@ -155,6 +157,12 @@ export async function runEpisodeCommand(args, opts) {
             changeDirPath,
             mainArm,
             ...(args.noBaseline ? { skipBaseline: true } : {}),
+            ...(episodeConfig?.selfEvolution?.reward
+                ? { reward: episodeConfig.selfEvolution.reward }
+                : {}),
+            ...(episodeConfig?.selfEvolution?.critic
+                ? { critic: episodeConfig.selfEvolution.critic }
+                : {}),
         };
         outcome = await runEpisode(episodeOptions);
     }

package/dist/core/fitness/test-failures.js

@@ -57,6 +57,14 @@ function findAssertion(lines, from) {
 function cleanToken(value) {
     return value.replace(/^[`'"]+|[`'"]+$/g, '');
 }
+/**
+ * POSIX-normalize a path so a Windows pytest path (`tests\test_x.py`) matches the
+ * already-POSIX-normalized file-edit paths in the action skeleton — the reward
+ * agent's renamed/edited-test caveat compares the two by exact string.
+ */
+function toPosix(p) {
+    return p.replace(/\\/g, '/');
+}
 /**
  * Extract failing test ids + assertion lines from observed runner output.
  * Returns `[]` when nothing is recognized. Deduplicates by testId, preserves
@@ -86,7 +94,7 @@ export function parseTestFailures(output) {
             const inline = pytest[2]?.trim();
             push({
                 testId,
-                file: testId.split('::')[0],
+                file: toPosix(testId.split('::')[0]),
                 ...(inline
                     ? { assertion: capAssertion(inline) }
                     : (() => {
@@ -98,7 +106,7 @@ export function parseTestFailures(output) {
         }
         const vitest = VITEST_FAIL_RE.exec(line);
         if (vitest) {
-            const file = cleanToken(vitest[1]);
+            const file = toPosix(cleanToken(vitest[1]));
             const rest = vitest[2]?.trim();
             const testId = rest ? `${file} > ${rest}` : file;
             const assertion = findAssertion(lines, i);

package/dist/core/project-config.d.ts

@@ -27,6 +27,23 @@ export declare const ProjectConfigSchema: z.ZodObject<{
         focus: z.ZodOptional<z.ZodBoolean>;
         advantageRollbackThreshold: z.ZodOptional<z.ZodNumber>;
         editBudget: z.ZodOptional<z.ZodNumber>;
+        reward: z.ZodOptional<z.ZodObject<{
+            samples: z.ZodOptional<z.ZodNumber>;
+            noiseFloor: z.ZodOptional<z.ZodNumber>;
+            orderSwap: z.ZodOptional<z.ZodBoolean>;
+            requireCorrectnessGate: z.ZodOptional<z.ZodBoolean>;
+            tamperCheck: z.ZodOptional<z.ZodEnum<{
+                off: "off";
+                flag: "flag";
+                block: "block";
+            }>>;
+        }, z.core.$strip>>;
+        critic: z.ZodOptional<z.ZodObject<{
+            baselineMode: z.ZodOptional<z.ZodEnum<{
+                "re-test": "re-test";
+                "re-do": "re-do";
+            }>>;
+        }, z.core.$strip>>;
     }, z.core.$strip>>;
     health: z.ZodOptional<z.ZodObject<{
         source: z.ZodDefault<z.ZodEnum<{

package/dist/core/project-config.js

@@ -60,6 +60,42 @@ export const ProjectConfigSchema = z.object({
         // 演进智能体 EVOLVING AGENT's ONE bounded edit may total. Default 40.
         // Optional/omitted ⇒ the agent's DEFAULT_EVOLVING_AGENT_EDIT_BUDGET applies.
         editBudget: z.number().optional(),
+        // Loop v2 — 奖励智能体 REWARD AGENT judge-quality knobs. ALL optional; omitted
+        // ⇒ the historical single-sample, flag-only behaviour (no extra LLM spawns).
+        reward: z
+            .object({
+            // ② How many judged duels per episode. Default 1 (single sample, no
+            //   extra spawns). >1 enables the A/A noise floor + SPRT + order-swap.
+            samples: z.number().optional(),
+            // ② Minimum |advantage| to trust; within the floor ⇒ insufficient-signal.
+            //   Omitted ⇒ measured from an A/A pair when samples>1, else unused.
+            noiseFloor: z.number().optional(),
+            // ③ Swap arm presentation order across samples to cancel position bias.
+            orderSwap: z.boolean().optional(),
+            // ① Enforce the correctness hard-gate inside the judge (default on).
+            requireCorrectnessGate: z.boolean().optional(),
+            // ④ Test-tamper handling: 'off' (no check), 'flag' (annotate only,
+            //   default), or 'block' (force insufficient-signal + reject-buffer).
+            tamperCheck: z.enum(['off', 'flag', 'block']).optional(),
+        })
+            .optional(),
+        // Loop v2 — CRITIC AGENT（基线智能体 baseline agent）baseline construction.
+        //   're-do' (default): the baseline arm RE-DOES the change under the prior
+        //     policy vN — it resets the change's GENERATED artifacts (design.md,
+        //     tasks.md), re-authors design under the installed vN template, then
+        //     re-implements → gen-test → run-test. So advantage ＝ reward(主臂) −
+        //     reward(基线臂) reflects the POLICY change, not re-run noise. Faithful
+        //     when the change's implementation is still uncommitted at episode time
+        //     (the workflow default) and an isolated git worktree can be created;
+        //     on a non-git copy fallback it degrades to a re-measure (documented).
+        //   're-test': the prior behaviour — re-run the EXISTING change's tests
+        //     under vN's template (cheaper, but an already-authored change does not
+        //     exercise the design template). Omitted ⇒ 're-do'.
+        critic: z
+            .object({
+            baselineMode: z.enum(['re-test', 're-do']).optional(),
+        })
+            .optional(),
     })
         .optional()
         .describe('Per-canonical-target self-evolution toggles'),
@@ -246,6 +282,38 @@ export function readProjectConfig(projectRoot) {
                 else if (rawSE.editBudget !== undefined) {
                     console.warn(`Invalid 'selfEvolution.editBudget' in config (must be a number), ignoring`);
                 }
+                // Loop v2 — 奖励智能体 REWARD AGENT knobs. Resilient: each sub-field is
+                // validated independently; a bad value is dropped with a warning (the
+                // judge/aggregator default applies). Omitted ⇒ undefined (single-sample,
+                // flag-only — byte-identical to configs that never set `reward`).
+                const rewardSchema = ProjectConfigSchema.shape.selfEvolution
+                    .unwrap()
+                    .shape.reward.unwrap();
+                const rewardResult = rewardSchema.safeParse(rawSE.reward);
+                if (rewardResult.success) {
+                    if (Object.keys(rewardResult.data).length > 0) {
+                        selfEvolution.reward = rewardResult.data;
+                    }
+                }
+                else if (rawSE.reward !== undefined) {
+                    console.warn(`Invalid 'selfEvolution.reward' in config (samples/noiseFloor numbers, ` +
+                        `orderSwap/requireCorrectnessGate booleans, tamperCheck off|flag|block), ignoring`);
+                }
+                // Loop v2 — CRITIC AGENT knobs. Resilient: a bad value is dropped with a
+                // warning (the critic default 're-do' then applies). Omitted ⇒ undefined
+                // (byte-identical to configs that never set `critic`).
+                const criticSchema = ProjectConfigSchema.shape.selfEvolution
+                    .unwrap()
+                    .shape.critic.unwrap();
+                const criticResult = criticSchema.safeParse(rawSE.critic);
+                if (criticResult.success) {
+                    if (Object.keys(criticResult.data).length > 0) {
+                        selfEvolution.critic = criticResult.data;
+                    }
+                }
+                else if (rawSE.critic !== undefined) {
+                    console.warn(`Invalid 'selfEvolution.critic' in config (baselineMode must be 're-test' or 're-do'), ignoring`);
+                }
                 config.selfEvolution = selfEvolution;
             }
             else {

package/dist/core/self-evolution/critic-agent.d.ts

@@ -4,23 +4,30 @@
  *
  * The CRITIC AGENT is an AGENT with the SAME input/output as the 主智能体 MAIN
  * AGENT (frozen actor; the user's host agent running the current 策略 policy
- * vN+1). It reruns LAST episode's 策略 policy vN on the SAME change in an
- * ISOLATED worktree, so the 奖励智能体 REWARD AGENT can later 算分 calculate
- * reward(主臂)＆reward(基线臂) and advantage ＝ reward(主臂) − reward(基线臂).
- * Only its baseline trajectory survives — 产物即弃 (worktree artifacts
- * discarded): the worktree is torn down in `finally`, and the single durable
- * output is the `baseline-arm/` capture in the episode store.
+ * vN+1). It produces the baseline arm for LAST episode's 策略 policy vN on the
+ * SAME change in an ISOLATED worktree (by default RE-DOING the change under vN —
+ * see {@link CriticBaselineMode}), so the 奖励智能体 REWARD AGENT can later 算分
+ * calculate reward(主臂)＆reward(基线臂) and advantage ＝ reward(主臂) −
+ * reward(基线臂). Only its baseline trajectory survives — 产物即弃 (worktree
+ * artifacts discarded): the worktree is torn down in `finally`, and the single
+ * durable output is the `baseline-arm/` capture in the episode store.
  *
  * This module orchestrates ONE baseline arm:
- *   1. create an isolated worktree OUTSIDE the repo (git worktree, else a
- *      recursive file copy fallback),
+ *   1. create an isolated worktree OUTSIDE the repo (git worktree at detached
+ *      HEAD — which excludes the change's still-uncommitted implementation — else
+ *      a recursive file copy fallback),
  *   2. make it runnable (node_modules junction/symlink + the untracked surfaces
  *      the rerun reads),
+ *   2b. ('re-do' mode, the default) reset the copied change dir to its INPUTS —
+ *      remove the GENERATED artifacts (design.md, tasks.md) + reports — so the
+ *      rerun re-authors them under the installed prior policy ({@link
+ *      resetChangeArtifactsForRedo}),
  *   3. INSTALL 策略 policy vN into the worktree from the byte-for-byte version
  *      snapshot, so the baseline arm reruns the PRIOR policy and not the live
  *      templates,
  *   4. rerun headlessly via the host-aware {@link runHeadlessAgent} with
- *      cwd = worktree, measurement only, never editing canonical files,
+ *      cwd = worktree ('re-do' regenerates design→tasks→impl→tests; 're-test'
+ *      re-runs the existing change's tests), never editing canonical files,
  *   5. persist the baseline arm (stdout always; the claude session transcript +
  *      action skeleton when discoverable; an `objective.json` shaped IDENTICALLY
  *      to the main arm's), and
@@ -32,6 +39,7 @@
  * strips every arm/candidate word.
  */
 import { spawn as nodeSpawn } from 'node:child_process';
+import type { ObservedTestFailure } from '../trajectory/facts.js';
 /** Error thrown when the worktree could not be created (git AND copy fallback failed). */
 export declare class CriticWorktreeError extends Error {
     constructor(message: string);
@@ -66,6 +74,18 @@ export interface ArmObjective {
     verified: boolean;
     observedStatus: 'success' | 'failure' | null;
     measuredAt: string;
+    /**
+     * Whether a real test-runner invocation was OBSERVED in the trajectory (vs a
+     * self-reported pass rate). Lets the 奖励智能体 REWARD AGENT calibrate
+     * confidence on the correctness anchor (P2). Omitted on older captures.
+     */
+    testRunObserved?: boolean;
+    /**
+     * Failing test ids (+ assertion lines) parsed from the OBSERVED runner output
+     * — the per-arm failure CONTENT the judge contrasts across arms (P1). Omitted
+     * when nothing was recognized (keeps JSON baselines stable).
+     */
+    observedFailures?: ObservedTestFailure[];
 }
 export interface ShouldRunCriticAgentOptions {
     repoRoot: string;
@@ -110,7 +130,22 @@ export declare function shouldRunCriticAgent(opts: ShouldRunCriticAgentOptions):
  * canonical files, and to print the runner summary line verbatim as its final
  * line.
  */
-export declare function assembleCriticPrompt(changeName: string): string;
+/**
+ * How the CRITIC AGENT builds the baseline arm:
+ *   - 're-do' (default): RE-DO the change from its inputs under the prior policy
+ *     vN — reset the GENERATED artifacts (design.md, tasks.md), re-author design
+ *     under the installed vN template, then re-implement → gen-test → run-test.
+ *     advantage ＝ reward(主臂) − reward(基线臂) then reflects the POLICY change,
+ *     because the design template is actually exercised (it shapes the freshly
+ *     authored design). This is faithful when the change's implementation is
+ *     still uncommitted at episode time (the workflow default — so the isolated
+ *     git worktree, checked out at detached HEAD, holds the PRE-change code).
+ *   - 're-test': re-run the EXISTING change's tests under vN's template. Cheaper,
+ *     but an already-authored change never exercises the design template, so the
+ *     baseline mostly re-measures the main arm's own work.
+ */
+export type CriticBaselineMode = 're-test' | 're-do';
+export declare function assembleCriticPrompt(changeName: string, mode?: CriticBaselineMode): string;
 export interface RunCriticAgentOptions {
     repoRoot: string;
     targetId: string;
@@ -118,6 +153,11 @@ export interface RunCriticAgentOptions {
     episodeId: string;
     /** LAST episode's policy version vN, from {@link shouldRunCriticAgent}. */
     baselineVersion: number;
+    /**
+     * How the baseline arm is built (see {@link CriticBaselineMode}). Default
+     * 're-do' (regenerate the change under vN so the policy is exercised).
+     */
+    baselineMode?: CriticBaselineMode;
     /** Injectable spawn seam for tests; defaults to node's spawn. */
     spawn?: typeof nodeSpawn;
     /** Hard timeout per agent run (ms). Default 600000 (10 min). */
@@ -138,6 +178,8 @@ export interface RunCriticAgentResult {
     worktreePath: string;
     /** How the worktree was created. */
     worktreeMode: 'git-worktree' | 'copy-fallback';
+    /** Which baseline construction ran (see {@link CriticBaselineMode}). */
+    baselineMode: CriticBaselineMode;
 }
 /**
  * Run the CRITIC AGENT（基线智能体 baseline agent）'s full baseline arm and

package/dist/core/self-evolution/critic-agent.js

@@ -4,23 +4,30 @@
  *
  * The CRITIC AGENT is an AGENT with the SAME input/output as the 主智能体 MAIN
  * AGENT (frozen actor; the user's host agent running the current 策略 policy
- * vN+1). It reruns LAST episode's 策略 policy vN on the SAME change in an
- * ISOLATED worktree, so the 奖励智能体 REWARD AGENT can later 算分 calculate
- * reward(主臂)＆reward(基线臂) and advantage ＝ reward(主臂) − reward(基线臂).
- * Only its baseline trajectory survives — 产物即弃 (worktree artifacts
- * discarded): the worktree is torn down in `finally`, and the single durable
- * output is the `baseline-arm/` capture in the episode store.
+ * vN+1). It produces the baseline arm for LAST episode's 策略 policy vN on the
+ * SAME change in an ISOLATED worktree (by default RE-DOING the change under vN —
+ * see {@link CriticBaselineMode}), so the 奖励智能体 REWARD AGENT can later 算分
+ * calculate reward(主臂)＆reward(基线臂) and advantage ＝ reward(主臂) −
+ * reward(基线臂). Only its baseline trajectory survives — 产物即弃 (worktree
+ * artifacts discarded): the worktree is torn down in `finally`, and the single
+ * durable output is the `baseline-arm/` capture in the episode store.
  *
  * This module orchestrates ONE baseline arm:
- *   1. create an isolated worktree OUTSIDE the repo (git worktree, else a
- *      recursive file copy fallback),
+ *   1. create an isolated worktree OUTSIDE the repo (git worktree at detached
+ *      HEAD — which excludes the change's still-uncommitted implementation — else
+ *      a recursive file copy fallback),
  *   2. make it runnable (node_modules junction/symlink + the untracked surfaces
  *      the rerun reads),
+ *   2b. ('re-do' mode, the default) reset the copied change dir to its INPUTS —
+ *      remove the GENERATED artifacts (design.md, tasks.md) + reports — so the
+ *      rerun re-authors them under the installed prior policy ({@link
+ *      resetChangeArtifactsForRedo}),
  *   3. INSTALL 策略 policy vN into the worktree from the byte-for-byte version
  *      snapshot, so the baseline arm reruns the PRIOR policy and not the live
  *      templates,
  *   4. rerun headlessly via the host-aware {@link runHeadlessAgent} with
- *      cwd = worktree, measurement only, never editing canonical files,
+ *      cwd = worktree ('re-do' regenerates design→tasks→impl→tests; 're-test'
+ *      re-runs the existing change's tests), never editing canonical files,
  *   5. persist the baseline arm (stdout always; the claude session transcript +
  *      action skeleton when discoverable; an `objective.json` shaped IDENTICALLY
  *      to the main arm's), and
@@ -105,29 +112,79 @@ export async function shouldRunCriticAgent(opts) {
         baselineVersion,
     };
 }
-/**
- * Assemble the CRITIC AGENT（基线智能体 baseline agent）rerun prompt. STRIPPED
- * of every arm/candidate word: the agent is simply told to re-run change
- * <changeName> end-to-end (apply → gen-test → run-test) under the templates
- * already installed in its working directory, measurement only, never editing
- * canonical files, and to print the runner summary line verbatim as its final
- * line.
- */
-export function assembleCriticPrompt(changeName) {
+export function assembleCriticPrompt(changeName, mode = 're-do') {
+    if (mode === 're-test') {
+        return [
+            `You are RE-RUNNING an existing SynergySpec change end-to-end to measure its`,
+            `test outcome under the artifact templates already installed in your working`,
+            `directory. This is a measurement run only — do NOT modify any canonical`,
+            `workflow prompt, artifact template, or schema, and do NOT edit the frozen`,
+            `gen-test/run-test oracle.`,
+            ``,
+            `Change name: ${changeName}`,
+            ``,
+            `Run the change's tests (apply → gen-test → run-test) and output the test`,
+            `runner's SUMMARY LINE verbatim as the final line of your response, e.g.`,
+            `"Tests  12 passed | 1 failed (13)" or "5 passed, 0 failed in 0.4s".`,
+        ].join('\n');
+    }
+    // 're-do' — regenerate the change end-to-end so the installed prior-policy
+    // design template is actually exercised.
     return [
-        `You are RE-RUNNING an existing SynergySpec change end-to-end to measure its`,
-        `test outcome under the artifact templates already installed in your working`,
-        `directory. This is a measurement run only — do NOT modify any canonical`,
-        `workflow prompt, artifact template, or schema, and do NOT edit the frozen`,
-        `gen-test/run-test oracle.`,
+        `You are RE-DOING an existing SynergySpec change from scratch under the`,
+        `artifact templates currently installed in your working directory, to measure`,
+        `the test outcome those templates produce. This is a measurement run.`,
         ``,
         `Change name: ${changeName}`,
         ``,
-        `Run the change's tests (apply → gen-test → run-test) and output the test`,
-        `runner's SUMMARY LINE verbatim as the final line of your response, e.g.`,
-        `"Tests  12 passed | 1 failed (13)" or "5 passed, 0 failed in 0.4s".`,
+        `The change's INPUT artifacts (proposal.md, usecases.md, specs/) are present.`,
+        `Its design.md and tasks.md have been intentionally REMOVED so you regenerate`,
+        `them under the installed templates. Re-create the change end-to-end:`,
+        ``,
+        `1. Regenerate the design — run`,
+        `     synergyspec-selfevolving instructions design --change "${changeName}" --json`,
+        `   read the returned template + dependency files (proposal.md, usecases.md),`,
+        `   and author design.md using that template as the structure.`,
+        `2. Regenerate the tasks the same way`,
+        `   (synergyspec-selfevolving instructions tasks --change "${changeName}" --json),`,
+        `   then apply them — implement the code each task requires.`,
+        `3. Generate the change's tests (gen-test), then run the test runner (run-test).`,
+        ``,
+        `Do NOT modify any canonical workflow prompt, artifact TEMPLATE, or schema, and`,
+        `do NOT edit the frozen gen-test/run-test oracle. Write ONLY the change's own`,
+        `artifacts (design.md, tasks.md in the change dir) and the implementation`,
+        `source the tasks require.`,
+        ``,
+        `Output the test runner's SUMMARY LINE verbatim as the final line of your`,
+        `response, e.g. "Tests  12 passed | 1 failed (13)" or "5 passed, 0 failed in 0.4s".`,
     ].join('\n');
 }
+/**
+ * Generated artifacts a 're-do' baseline removes from the copied change dir
+ * before the rerun, so the agent re-authors them under the installed prior
+ * policy. design.md + tasks.md are the policy-shaped chain; the report files are
+ * post-implementation residue that would otherwise read the change as already
+ * applied (status keys doneness off file existence). The INPUT artifacts
+ * (proposal.md, usecases.md, specs/) — which define "the same task" — are KEPT.
+ */
+const REDO_REGENERATED_ARTIFACTS = [
+    'design.md',
+    'tasks.md',
+    'test-report.md',
+    'test-plan.md',
+    'spec-tests.md',
+    'spec-blast-radius.md',
+    'verification-report.md',
+];
+/**
+ * Reset a copied change dir to its inputs for a 're-do' baseline (see
+ * {@link REDO_REGENERATED_ARTIFACTS}). Best-effort: a missing artifact is fine.
+ */
+async function resetChangeArtifactsForRedo(changeDir) {
+    for (const rel of REDO_REGENERATED_ARTIFACTS) {
+        await fs.rm(path.join(changeDir, rel), { force: true }).catch(() => { });
+    }
+}
 const NODE_MODULES = 'node_modules';
 const CONFIG_DIR = '.synergyspec-selfevolving';
 const SCHEMAS_REL = path.join('synergyspec-selfevolving', 'schemas');
@@ -143,6 +200,7 @@ export async function runCriticAgent(opts) {
     const spawnImpl = opts.spawn ?? nodeSpawn;
     const timeoutMs = opts.timeoutMs ?? 600000;
     const homeDir = opts.homeDir ?? os.homedir();
+    const baselineMode = opts.baselineMode ?? 're-do';
     if (!Number.isInteger(opts.baselineVersion) || opts.baselineVersion < 0) {
         throw new Error(`runCriticAgent requires a non-negative integer baselineVersion, got ${JSON.stringify(opts.baselineVersion)}`);
     }
@@ -155,13 +213,31 @@ export async function runCriticAgent(opts) {
     try {
         // 1) Isolated worktree OUTSIDE the repo (git worktree --detach, else copy).
         worktreeMode = await createIsolatedWorktree(repoRoot, worktreePath, spawnImpl);
+        // 're-do' fidelity needs the detached-HEAD tree (pre-change code). The copy
+        // fallback (non-git repo) brings the LIVE tree — including the change's
+        // uncommitted implementation — so it cannot reach the pre-change state and
+        // degrades to a re-measure. Surface that so a degraded baseline is not silent.
+        if (baselineMode === 're-do' && worktreeMode === 'copy-fallback') {
+            console.warn(`[critic] re-do baseline degraded for "${opts.changeName}": no git worktree ` +
+                `(copy fallback) — the change's implementation could not be isolated, so the ` +
+                `baseline re-measures rather than re-does. Use a git repo, or set ` +
+                `selfEvolution.critic.baselineMode: re-test to silence this.`);
+        }
         // 2) Make it runnable: node_modules junction/symlink + untracked surfaces.
         await makeWorktreeRunnable(repoRoot, worktreePath, opts.changeName);
+        // 2b) 're-do': reset the copied change dir to its inputs so the rerun
+        //     RE-AUTHORS design+tasks under the prior policy (and re-implements on
+        //     the pre-change code the detached-HEAD worktree already holds). The
+        //     fidelity over 're-test' is that the design TEMPLATE is actually
+        //     exercised, so advantage reflects the policy change, not re-run noise.
+        if (baselineMode === 're-do') {
+            await resetChangeArtifactsForRedo(path.join(worktreePath, 'synergyspec-selfevolving', 'changes', opts.changeName));
+        }
         // 3) INSTALL 策略 policy vN (byte-for-byte snapshot files) — the fidelity
         //    fix the old GA replay never performed.
         await installPolicyVersion(repoRoot, worktreePath, opts.targetId, opts.baselineVersion);
-        // 4) Rerun headlessly with cwd = worktree (measurement only).
-        const prompt = assembleCriticPrompt(opts.changeName);
+        // 4) Rerun headlessly with cwd = worktree (re-do: regenerate; re-test: measure).
+        const prompt = assembleCriticPrompt(opts.changeName, baselineMode);
         const run = await runHeadlessAgent(prompt, {
             cwd: worktreePath,
             spawn: spawnImpl,
@@ -212,6 +288,10 @@ export async function runCriticAgent(opts) {
             verified,
             observedStatus,
             measuredAt,
+            ...(facts ? { testRunObserved: facts.testRunObserved } : {}),
+            ...(facts?.observedFailures && facts.observedFailures.length > 0
+                ? { observedFailures: facts.observedFailures }
+                : {}),
         };
         // Transcript: the claude session `.jsonl` when discovered, else stdout.
         let transcriptDiscovered = false;
@@ -256,6 +336,7 @@ export async function runCriticAgent(opts) {
             transcriptDiscovered,
             worktreePath,
             worktreeMode,
+            baselineMode,
         };
     }
     finally {

package/dist/core/self-evolution/episode-orchestrator.d.ts

@@ -47,9 +47,10 @@
 import { spawn as nodeSpawn } from 'node:child_process';
 import type { LearnReport } from '../learn.js';
 import type { TrajectorySource } from '../trajectory/source.js';
-import { type PolicyResolveFiles } from './policy/policy-store.js';
+import { type PolicyResolveFiles, type PolicyLedgerEntry } from './policy/policy-store.js';
 import { type EpisodeStage } from './episode-store.js';
-import { type ArmObjective } from './critic-agent.js';
+import { type ArmObjective, type CriticBaselineMode } from './critic-agent.js';
+import { type RewardConfig } from './reward-aggregator.js';
 import { type RunEvolvingAgentResult } from './evolving-agent.js';
 /** The 主智能体 MAIN AGENT (policy vN+1) capture the orchestrator records. */
 export interface MainArmCapture {
@@ -107,6 +108,30 @@ export interface CaptureMainArmOptions {
 export declare function captureMainArm(opts: CaptureMainArmOptions): Promise<MainArmCapture>;
 /** The decision the orchestrator made on the main arm's edits. */
 export type EpisodeDecision = 'rolled-back' | 'kept' | 'abstained';
+/**
+ * Count the consecutive trailing rolled-back episodes in the 版本账本 ledger.
+ *
+ * A bad streak's ledger tail reads `…, evolve, rollback, evolve, rollback` — the
+ * 演进智能体 EVOLVING AGENT appends exactly one 'evolve' after each decision, so
+ * each counted rollback is reached by skipping the single 'evolve' that follows
+ * it. A 'kept' episode leaves a bare 'evolve' (no following rollback) which
+ * breaks the streak, as do 'init'/'refused'. Returns 0 when the head is not a
+ * rollback (the last episode kept). Pure.
+ */
+export declare function consecutiveRollbacks(ledger: readonly PolicyLedgerEntry[]): number;
+/**
+ * 步长 step-size schedule for the 演进智能体 EVOLVING AGENT's edit budget L.
+ *
+ * Backtracking-line-search / trust-region move (and SkillOpt's decaying edit
+ * budget): after an edit LOST ground and was rolled back, the next edit should
+ * be SMALLER — a smaller blast radius is cheaper to undo and its cause is more
+ * legible, and it keeps a struggling lineage from drifting via repeated
+ * full-size swings. HALVE the base budget once per consecutive rolled-back
+ * episode, never below `minBudget` (itself clamped to `base`, so a caller-shrunk
+ * base is never RAISED). A healthy lineage (no trailing rollback) keeps `base`.
+ * Pure.
+ */
+export declare function scheduledEditBudget(ledger: readonly PolicyLedgerEntry[], base: number, minBudget?: number): number;
 export interface RunEpisodeOptions {
     repoRoot: string;
     targetId: string;
@@ -123,6 +148,18 @@ export interface RunEpisodeOptions {
     advantageRollbackThreshold?: number;
     /** Edit budget L for the 演进智能体 EVOLVING AGENT. Default 40. */
     editBudget?: number;
+    /**
+     * 奖励智能体 REWARD AGENT judge-quality knobs (from `selfEvolution.reward`).
+     * Omitted ⇒ single sample, flag-only tamper (historical, zero extra spawns).
+     */
+    reward?: RewardConfig;
+    /**
+     * CRITIC AGENT（基线智能体）baseline construction (from `selfEvolution.critic`).
+     * Omitted ⇒ the critic's default 're-do' (regenerate the change under vN).
+     */
+    critic?: {
+        baselineMode?: CriticBaselineMode;
+    };
     /** Injectable spawn seam — threaded to ALL THREE agents. Defaults to node's spawn. */
     spawn?: typeof nodeSpawn;
     /** Injectable clock for the lock + episode id; defaults to `new Date()`. */