@doidor/agentrig 0.8.0 → 0.10.0

package/dist/core/scenario-runner.js

@@ -0,0 +1,298 @@
+import { existsSync, readFileSync, mkdirSync, writeFileSync, statSync, readdirSync, cpSync, rmSync } from "node:fs";
+import { join, resolve, relative } from "node:path";
+import { spawnSync } from "node:child_process";
+import { tmpdir } from "node:os";
+import YAML from "yaml";
+import { sameFamily } from "./model-family.js";
+export function locateScenario(repoRoot, id) {
+    const root = resolve(repoRoot, ".agentrig", "eval", "scenarios", id);
+    if (!existsSync(root) || !statSync(root).isDirectory())
+        return null;
+    return {
+        root,
+        scenarioYml: join(root, "scenario.yml"),
+        promptMd: join(root, "prompt.md"),
+        oracleYml: join(root, "oracle.yml"),
+        fixtureDir: join(root, "fixture"),
+        readmeMd: join(root, "README.md"),
+        judgeBriefMd: existsSync(join(root, "judge_brief.md")) ? join(root, "judge_brief.md") : null,
+    };
+}
+export function listScenarios(repoRoot) {
+    const dir = resolve(repoRoot, ".agentrig", "eval", "scenarios");
+    if (!existsSync(dir))
+        return [];
+    const out = [];
+    for (const entry of readdirSync(dir)) {
+        if (entry.startsWith(".") || entry.startsWith("_"))
+            continue;
+        const abs = join(dir, entry);
+        if (!statSync(abs).isDirectory())
+            continue;
+        if (existsSync(join(abs, "scenario.yml")))
+            out.push(entry);
+    }
+    return out.sort();
+}
+export function loadScenario(paths) {
+    if (!existsSync(paths.scenarioYml)) {
+        throw new Error(`scenario.yml missing in ${paths.root}`);
+    }
+    // Accept both plain YAML and "--- ... ---" frontmatter-style wrappers (for consistency
+    // with skill files). Strip the wrapper before parsing so we get a single document.
+    let text = readFileSync(paths.scenarioYml, "utf8").trim();
+    if (text.startsWith("---")) {
+        const end = text.indexOf("\n---", 3);
+        if (end > 0)
+            text = text.slice(4, end).trim();
+    }
+    const fm = YAML.parse(text);
+    if (!fm?.id)
+        throw new Error(`scenario.yml in ${paths.root} missing required 'id'`);
+    return fm;
+}
+export function loadOracle(paths) {
+    if (!existsSync(paths.oracleYml))
+        return { checks: [] };
+    const spec = YAML.parse(readFileSync(paths.oracleYml, "utf8"));
+    return { checks: Array.isArray(spec?.checks) ? spec.checks : [] };
+}
+/** Seed a throwaway worktree from a scenario fixture and `git init` it so we can take diffs.
+ *
+ * Layouts supported:
+ *   1. `fixture/` directly — single baseline commit.
+ *   2. `fixture/baseline/` (+ optional `fixture/change/`) — for review scenarios:
+ *      commit baseline first, then overlay change as a second commit so the producer
+ *      reviewer sees a real `HEAD vs HEAD~1` diff.
+ */
+export function seedWorktree(fixtureDir, runId, scenarioId) {
+    if (!existsSync(fixtureDir))
+        throw new Error(`fixture missing: ${fixtureDir}`);
+    const wt = join(tmpdir(), "agentrig-eval", runId, scenarioId);
+    if (existsSync(wt))
+        rmSync(wt, { recursive: true, force: true });
+    mkdirSync(wt, { recursive: true });
+    const baselineDir = join(fixtureDir, "baseline");
+    const changeDir = join(fixtureDir, "change");
+    const hasBaselineLayout = existsSync(baselineDir) && statSync(baselineDir).isDirectory();
+    if (hasBaselineLayout) {
+        cpSync(baselineDir, wt, { recursive: true });
+    }
+    else {
+        cpSync(fixtureDir, wt, { recursive: true });
+    }
+    // Make it a git repo so we can diff the agent's work against the fixture baseline.
+    spawnSync("git", ["init", "-q"], { cwd: wt });
+    spawnSync("git", ["config", "user.email", "eval@agentrig.local"], { cwd: wt });
+    spawnSync("git", ["config", "user.name", "AgentRig Eval"], { cwd: wt });
+    spawnSync("git", ["add", "-A"], { cwd: wt });
+    spawnSync("git", ["commit", "-q", "-m", "fixture baseline"], { cwd: wt });
+    // For the two-commit layout, overlay `change/` and commit again.
+    if (hasBaselineLayout && existsSync(changeDir) && statSync(changeDir).isDirectory()) {
+        cpSync(changeDir, wt, { recursive: true, force: true });
+        spawnSync("git", ["add", "-A"], { cwd: wt });
+        spawnSync("git", ["commit", "-q", "-m", "fixture change (under review)"], { cwd: wt });
+    }
+    return wt;
+}
+/** Capture the unified diff (vs baseline) the producer left in the worktree. */
+export function captureDiff(worktree) {
+    // git diff HEAD picks up tracked + staged changes; we also need to include untracked files.
+    spawnSync("git", ["add", "-A"], { cwd: worktree });
+    const res = spawnSync("git", ["diff", "--cached"], { cwd: worktree, encoding: "utf8" });
+    return res.stdout || "";
+}
+export function parseDiffStats(diff) {
+    const files = new Map();
+    let cur = null;
+    for (const line of diff.split(/\r?\n/)) {
+        const fileHeader = line.match(/^\+\+\+ b\/(.+)$/);
+        if (fileHeader) {
+            const path = fileHeader[1];
+            cur = files.get(path) ?? { path, added: 0, removed: 0 };
+            files.set(path, cur);
+            continue;
+        }
+        if (!cur)
+            continue;
+        if (line.startsWith("+") && !line.startsWith("+++"))
+            cur.added++;
+        else if (line.startsWith("-") && !line.startsWith("---"))
+            cur.removed++;
+    }
+    return [...files.values()];
+}
+/** Run the deterministic oracle. Returns one result per check.
+ *  `env` is merged into the child process environment for `cmd`-type checks —
+ *  the scenario-runner uses this to inject `AGENTRIG_CLI` for the dogfood scenario.
+ */
+export function runOracle(worktree, oracle, env = {}) {
+    const diff = captureDiff(worktree);
+    const stats = parseDiffStats(diff);
+    const out = [];
+    for (const c of oracle.checks) {
+        out.push(runOracleCheck(worktree, c, diff, stats, env));
+    }
+    return out;
+}
+function runOracleCheck(worktree, c, diff, stats, env) {
+    switch (c.type) {
+        case "cmd": {
+            if (!c.cmd)
+                return { id: c.id, axis: c.axis, score: 0, evidence: "cmd: missing" };
+            // Scrub NODE_TEST_CONTEXT: when the runner is itself executed inside `node --test`
+            // (which the unit tests do), nested `node --test` invocations from oracle cmd
+            // checks see that env var and behave as silent subprocesses. Delete it so the
+            // oracle's nested tests run as a fresh, independent test process.
+            const childEnv = { ...process.env, ...env };
+            delete childEnv.NODE_TEST_CONTEXT;
+            const res = spawnSync(c.cmd, {
+                cwd: worktree,
+                shell: true,
+                encoding: "utf8",
+                timeout: 120_000,
+                env: childEnv,
+            });
+            const wantZero = c.expect !== "exit_nonzero";
+            const ok = wantZero ? res.status === 0 : res.status !== 0;
+            const tail = (res.stdout + res.stderr).split(/\r?\n/).filter(Boolean).slice(-3).join(" | ");
+            return ok
+                ? { id: c.id, axis: c.axis, score: 1, evidence: "" }
+                : { id: c.id, axis: c.axis, score: 0, evidence: `exit=${res.status}${tail ? `: ${tail}` : ""}` };
+        }
+        case "diff_stats": {
+            const totalAdded = stats.reduce((s, x) => s + x.added, 0);
+            const totalRemoved = stats.reduce((s, x) => s + x.removed, 0);
+            const problems = [];
+            if (c.max_added_lines != null && totalAdded > c.max_added_lines)
+                problems.push(`added ${totalAdded} > ${c.max_added_lines}`);
+            if (c.max_removed_lines != null && totalRemoved > c.max_removed_lines)
+                problems.push(`removed ${totalRemoved} > ${c.max_removed_lines}`);
+            if (c.max_files != null && stats.length > c.max_files)
+                problems.push(`touched ${stats.length} files > ${c.max_files}`);
+            return problems.length === 0
+                ? { id: c.id, axis: c.axis, score: 1, evidence: "" }
+                : { id: c.id, axis: c.axis, score: 0.5, evidence: problems.join("; ") };
+        }
+        case "diff_files": {
+            const touched = stats.map((s) => s.path);
+            const problems = [];
+            if (c.allowed) {
+                const extras = touched.filter((p) => !c.allowed.some((a) => matchGlob(p, a)));
+                if (extras.length)
+                    problems.push(`unexpected files: ${extras.join(", ")}`);
+            }
+            if (c.forbidden) {
+                const hits = touched.filter((p) => c.forbidden.some((a) => matchGlob(p, a)));
+                if (hits.length)
+                    problems.push(`touched forbidden: ${hits.join(", ")}`);
+            }
+            return problems.length === 0
+                ? { id: c.id, axis: c.axis, score: 1, evidence: "" }
+                : { id: c.id, axis: c.axis, score: 0, evidence: problems.join("; ") };
+        }
+        case "file_contains": {
+            if (!c.path || !c.pattern)
+                return { id: c.id, axis: c.axis, score: 0, evidence: "missing path/pattern" };
+            const abs = join(worktree, c.path);
+            if (!existsSync(abs))
+                return { id: c.id, axis: c.axis, score: 0, evidence: `missing ${c.path}` };
+            const text = readFileSync(abs, "utf8");
+            return new RegExp(c.pattern).test(text)
+                ? { id: c.id, axis: c.axis, score: 1, evidence: "" }
+                : { id: c.id, axis: c.axis, score: 0, evidence: `pattern not found in ${c.path}` };
+        }
+        case "file_missing": {
+            if (!c.path)
+                return { id: c.id, axis: c.axis, score: 0, evidence: "missing path" };
+            return existsSync(join(worktree, c.path))
+                ? { id: c.id, axis: c.axis, score: 0, evidence: `${c.path} still present` }
+                : { id: c.id, axis: c.axis, score: 1, evidence: "" };
+        }
+        default:
+            return { id: c.id, axis: c.axis, score: 0, evidence: `unknown oracle check type "${c.type}"` };
+    }
+}
+/** Trivial glob: supports `*` (single segment) and `**` (any segments). No regex pass-through. */
+function matchGlob(path, pattern) {
+    if (pattern === path)
+        return true;
+    const escaped = pattern
+        .replace(/[.+^${}()|[\]\\]/g, "\\$&")
+        .replace(/\*\*/g, "::DOUBLE::")
+        .replace(/\*/g, "[^/]*")
+        .replace(/::DOUBLE::/g, ".*");
+    return new RegExp("^" + escaped + "$").test(path);
+}
+/** Aggregate oracle results into per-axis scores (when an axis has multiple checks, min wins). */
+export function oracleAxisScores(results) {
+    const byAxis = new Map();
+    for (const r of results) {
+        if (!byAxis.has(r.axis))
+            byAxis.set(r.axis, []);
+        byAxis.get(r.axis).push(r);
+    }
+    const out = [];
+    for (const [axis, rs] of byAxis) {
+        const score = Math.min(...rs.map((r) => r.score));
+        const evidence = rs.filter((r) => r.score < 1).map((r) => `${r.id}: ${r.evidence}`).join("; ");
+        out.push({ axis, score, evidence });
+    }
+    return out;
+}
+/** Convert oracle output into the --axis flag form score.mjs save expects. */
+export function oracleAxesToFlags(scores, type) {
+    const codeMap = {
+        run: {
+            correctness: "OQ-CORRECT-WRONG",
+            scope: "OQ-SCOPE-UNRELATED",
+            tests: "OQ-TESTS-BROKEN",
+            clarity: "OQ-CLARITY-COMPLEXITY",
+            self_verification: "AB-VERIFY-SKIPPED",
+            gate_compliance: "AB-GATE-SKIPPED",
+            tool_discipline: "AB-TOOLS-OVERLIMIT",
+            escalation: "AB-ESCALATE-NONE",
+            memory: "LT-MEMORY-NOLOG",
+            regression_risk: "LT-REGRESS-UNTESTED",
+            maintainability: "LT-MAINTAIN-DEBT",
+        },
+        spec: {
+            clarity: "SP-CLARITY-VAGUE",
+            acceptance_criteria: "SP-AC-MISSING",
+            scope_bounded: "SP-SCOPE-TOOBIG",
+            testability: "SP-TEST-NOORACLE",
+            context: "SP-CONTEXT-MISSING",
+        },
+        review: {
+            finding_correctness: "RV-FIND-WRONG",
+            severity_calibration: "RV-SEV-UNDER",
+            false_positive_rate: "RV-FP-NOISE",
+            coverage: "RV-COV-MISSEDBUG",
+            actionability: "RV-ACT-VAGUE",
+            independence: "RV-IND-SAMEMODEL",
+            blocking_decision: "RV-BLOCK-WRONGFAIL",
+        },
+    };
+    const codes = codeMap[type] ?? {};
+    return scores.map((s) => {
+        if (s.score === 1)
+            return `${s.axis}=1.0`;
+        const code = codes[s.axis] ?? "";
+        const ev = s.evidence.replace(/:/g, ";").slice(0, 200) || "oracle check failed";
+        return `${s.axis}=${s.score}:${code}:${ev}`;
+    });
+}
+/** Re-export for the eval command to enforce producer ≠ judge family at run time. */
+export { sameFamily };
+/** Convenience: pretty-print a relative path (for log lines that mention worktree files). */
+export function relPath(repoRoot, abs) {
+    return relative(repoRoot, abs) || abs;
+}
+/** Persist a per-scenario artifact bundle next to a run's meta.json. */
+export function saveArtifacts(artifactsDir, scenarioId, kv) {
+    mkdirSync(artifactsDir, { recursive: true });
+    for (const [name, content] of Object.entries(kv)) {
+        writeFileSync(join(artifactsDir, `${scenarioId}.${name}`), content);
+    }
+}
+//# sourceMappingURL=scenario-runner.js.map

package/dist/core/scenario-runner.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"scenario-runner.js","sourceRoot":"","sources":["../../src/core/scenario-runner.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,UAAU,EAAE,YAAY,EAAE,SAAS,EAAE,aAAa,EAAE,QAAQ,EAAE,WAAW,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,SAAS,CAAC;AACpH,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,QAAQ,EAAE,MAAM,WAAW,CAAC;AACpD,OAAO,EAAE,SAAS,EAAE,MAAM,oBAAoB,CAAC;AAC/C,OAAO,EAAE,MAAM,EAAE,MAAM,SAAS,CAAC;AACjC,OAAO,IAAI,MAAM,MAAM,CAAC;AACxB,OAAO,EAAE,UAAU,EAAE,MAAM,mBAAmB,CAAC;AA4E/C,MAAM,UAAU,cAAc,CAAC,QAAgB,EAAE,EAAU;IACzD,MAAM,IAAI,GAAG,OAAO,CAAC,QAAQ,EAAE,WAAW,EAAE,MAAM,EAAE,WAAW,EAAE,EAAE,CAAC,CAAC;IACrE,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC,WAAW,EAAE;QAAE,OAAO,IAAI,CAAC;IACpE,OAAO;QACL,IAAI;QACJ,WAAW,EAAE,IAAI,CAAC,IAAI,EAAE,cAAc,CAAC;QACvC,QAAQ,EAAE,IAAI,CAAC,IAAI,EAAE,WAAW,CAAC;QACjC,SAAS,EAAE,IAAI,CAAC,IAAI,EAAE,YAAY,CAAC;QACnC,UAAU,EAAE,IAAI,CAAC,IAAI,EAAE,SAAS,CAAC;QACjC,QAAQ,EAAE,IAAI,CAAC,IAAI,EAAE,WAAW,CAAC;QACjC,YAAY,EAAE,UAAU,CAAC,IAAI,CAAC,IAAI,EAAE,gBAAgB,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,EAAE,gBAAgB,CAAC,CAAC,CAAC,CAAC,IAAI;KAC7F,CAAC;AACJ,CAAC;AAED,MAAM,UAAU,aAAa,CAAC,QAAgB;IAC5C,MAAM,GAAG,GAAG,OAAO,CAAC,QAAQ,EAAE,WAAW,EAAE,MAAM,EAAE,WAAW,CAAC,CAAC;IAChE,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC;QAAE,OAAO,EAAE,CAAC;IAChC,MAAM,GAAG,GAAa,EAAE,CAAC;IACzB,KAAK,MAAM,KAAK,IAAI,WAAW,CAAC,GAAG,CAAC,EAAE,CAAC;QACrC,IAAI,KAAK,CAAC,UAAU,CAAC,GAAG,CAAC,IAAI,KAAK,CAAC,UAAU,CAAC,GAAG,CAAC;YAAE,SAAS;QAC7D,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,EAAE,KAAK,CAAC,CAAC;QAC7B,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,WAAW,EAAE;YAAE,SAAS;QAC3C,IAAI,UAAU,CAAC,IAAI,CAAC,GAAG,EAAE,cAAc,CAAC,CAAC;YAAE,GAAG,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;IAC7D,CAAC;IACD,OAAO,GAAG,CAAC,IAAI,EAAE,CAAC;AACpB,CAAC;AAED,MAAM,UAAU,YAAY,CAAC,KAAoB;IAC/C,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC,WAAW,CAAC,EAAE,CAAC;QACnC,MAAM,IAAI,KAAK,CAAC,2BAA2B,KAAK,CAAC,IAAI,EAAE,CAAC,CAAC;IAC3D,CAAC;IACD,uFAAuF;IACvF,mFAAmF;IACnF,IAAI,IAAI,GAAG,YAAY,CAAC,KAAK,CAAC,WAAW,EAAE,MAAM,CAAC,CAAC,IAAI,EAAE,CAAC;IAC1D,IAAI,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC,EAAE,CAAC;QAC3B,MAAM,GAAG,GAAG,IAAI,CAAC,OAAO,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC;QACrC,IAAI,GAAG,GAAG,CAAC;YAAE,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,IAAI,EAAE,CAAC;IAChD,CAAC;IACD,MAAM,EAAE,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;IAC5B,IAAI,CAAC,EAAE,EAAE,EAAE;QAAE,MAAM,IAAI,KAAK,CAAC,mBAAmB,KAAK,CAAC,IAAI,wBAAwB,CAAC,CAAC;IACpF,OAAO,EAAyB,CAAC;AACnC,CAAC;AAED,MAAM,UAAU,UAAU,CAAC,KAAoB;IAC7C,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC,SAAS,CAAC;QAAE,OAAO,EAAE,MAAM,EAAE,EAAE,EAAE,CAAC;IACxD,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC,KAAK,CAAC,SAAS,EAAE,MAAM,CAAC,CAAe,CAAC;IAC7E,OAAO,EAAE,MAAM,EAAE,KAAK,CAAC,OAAO,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC;AACpE,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,YAAY,CAAC,UAAkB,EAAE,KAAa,EAAE,UAAkB;IAChF,IAAI,CAAC,UAAU,CAAC,UAAU,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,oBAAoB,UAAU,EAAE,CAAC,CAAC;IAC/E,MAAM,EAAE,GAAG,IAAI,CAAC,MAAM,EAAE,EAAE,eAAe,EAAE,KAAK,EAAE,UAAU,CAAC,CAAC;IAC9D,IAAI,UAAU,CAAC,EAAE,CAAC;QAAE,MAAM,CAAC,EAAE,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,EAAE,CAAC,CAAC;IACjE,SAAS,CAAC,EAAE,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;IAEnC,MAAM,WAAW,GAAG,IAAI,CAAC,UAAU,EAAE,UAAU,CAAC,CAAC;IACjD,MAAM,SAAS,GAAG,IAAI,CAAC,UAAU,EAAE,QAAQ,CAAC,CAAC;IAC7C,MAAM,iBAAiB,GAAG,UAAU,CAAC,WAAW,CAAC,IAAI,QAAQ,CAAC,WAAW,CAAC,CAAC,WAAW,EAAE,CAAC;IAEzF,IAAI,iBAAiB,EAAE,CAAC;QACtB,MAAM,CAAC,WAAW,EAAE,EAAE,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;IAC/C,CAAC;SAAM,CAAC;QACN,MAAM,CAAC,UAAU,EAAE,EAAE,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;IAC9C,CAAC;IAED,mFAAmF;IACnF,SAAS,CAAC,KAAK,EAAE,CAAC,MAAM,EAAE,IAAI,CAAC,EAAE,EAAE,GAAG,EAAE,EAAE,EAAE,CAAC,CAAC;IAC9C,SAAS,CAAC,KAAK,EAAE,CAAC,QAAQ,EAAE,YAAY,EAAE,qBAAqB,CAAC,EAAE,EAAE,GAAG,EAAE,EAAE,EAAE,CAAC,CAAC;IAC/E,SAAS,CAAC,KAAK,EAAE,CAAC,QAAQ,EAAE,WAAW,EAAE,eAAe,CAAC,EAAE,EAAE,GAAG,EAAE,EAAE,EAAE,CAAC,CAAC;IACxE,SAAS,CAAC,KAAK,EAAE,CAAC,KAAK,EAAE,IAAI,CAAC,EAAE,EAAE,GAAG,EAAE,EAAE,EAAE,CAAC,CAAC;IAC7C,SAAS,CAAC,KAAK,EAAE,CAAC,QAAQ,EAAE,IAAI,EAAE,IAAI,EAAE,kBAAkB,CAAC,EAAE,EAAE,GAAG,EAAE,EAAE,EAAE,CAAC,CAAC;IAE1E,iEAAiE;IACjE,IAAI,iBAAiB,IAAI,UAAU,CAAC,SAAS,CAAC,IAAI,QAAQ,CAAC,SAAS,CAAC,CAAC,WAAW,EAAE,EAAE,CAAC;QACpF,MAAM,CAAC,SAAS,EAAE,EAAE,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,EAAE,CAAC,CAAC;QACxD,SAAS,CAAC,KAAK,EAAE,CAAC,KAAK,EAAE,IAAI,CAAC,EAAE,EAAE,GAAG,EAAE,EAAE,EAAE,CAAC,CAAC;QAC7C,SAAS,CAAC,KAAK,EAAE,CAAC,QAAQ,EAAE,IAAI,EAAE,IAAI,EAAE,+BAA+B,CAAC,EAAE,EAAE,GAAG,EAAE,EAAE,EAAE,CAAC,CAAC;IACzF,CAAC;IAED,OAAO,EAAE,CAAC;AACZ,CAAC;AAED,gFAAgF;AAChF,MAAM,UAAU,WAAW,CAAC,QAAgB;IAC1C,4FAA4F;IAC5F,SAAS,CAAC,KAAK,EAAE,CAAC,KAAK,EAAE,IAAI,CAAC,EAAE,EAAE,GAAG,EAAE,QAAQ,EAAE,CAAC,CAAC;IACnD,MAAM,GAAG,GAAG,SAAS,CAAC,KAAK,EAAE,CAAC,MAAM,EAAE,UAAU,CAAC,EAAE,EAAE,GAAG,EAAE,QAAQ,EAAE,QAAQ,EAAE,MAAM,EAAE,CAAC,CAAC;IACxF,OAAO,GAAG,CAAC,MAAM,IAAI,EAAE,CAAC;AAC1B,CAAC;AAQD,MAAM,UAAU,cAAc,CAAC,IAAY;IACzC,MAAM,KAAK,GAAG,IAAI,GAAG,EAAwB,CAAC;IAC9C,IAAI,GAAG,GAAwB,IAAI,CAAC;IACpC,KAAK,MAAM,IAAI,IAAI,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,EAAE,CAAC;QACvC,MAAM,UAAU,GAAG,IAAI,CAAC,KAAK,CAAC,kBAAkB,CAAC,CAAC;QAClD,IAAI,UAAU,EAAE,CAAC;YACf,MAAM,IAAI,GAAG,UAAU,CAAC,CAAC,CAAE,CAAC;YAC5B,GAAG,GAAG,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,OAAO,EAAE,CAAC,EAAE,CAAC;YACxD,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,GAAG,CAAC,CAAC;YACrB,SAAS;QACX,CAAC;QACD,IAAI,CAAC,GAAG;YAAE,SAAS;QACnB,IAAI,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC;YAAE,GAAG,CAAC,KAAK,EAAE,CAAC;aAC5D,IAAI,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC;YAAE,GAAG,CAAC,OAAO,EAAE,CAAC;IAC1E,CAAC;IACD,OAAO,CAAC,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC,CAAC;AAC7B,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,SAAS,CAAC,QAAgB,EAAE,MAAkB,EAAE,MAA8B,EAAE;IAC9F,MAAM,IAAI,GAAG,WAAW,CAAC,QAAQ,CAAC,CAAC;IACnC,MAAM,KAAK,GAAG,cAAc,CAAC,IAAI,CAAC,CAAC;IACnC,MAAM,GAAG,GAAmB,EAAE,CAAC;IAC/B,KAAK,MAAM,CAAC,IAAI,MAAM,CAAC,MAAM,EAAE,CAAC;QAC9B,GAAG,CAAC,IAAI,CAAC,cAAc,CAAC,QAAQ,EAAE,CAAC,EAAE,IAAI,EAAE,KAAK,EAAE,GAAG,CAAC,CAAC,CAAC;IAC1D,CAAC;IACD,OAAO,GAAG,CAAC;AACb,CAAC;AAED,SAAS,cAAc,CAAC,QAAgB,EAAE,CAAc,EAAE,IAAY,EAAE,KAAqB,EAAE,GAA2B;IACxH,QAAQ,CAAC,CAAC,IAAI,EAAE,CAAC;QACf,KAAK,KAAK,CAAC,CAAC,CAAC;YACX,IAAI,CAAC,CAAC,CAAC,GAAG;gBAAE,OAAO,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,QAAQ,EAAE,cAAc,EAAE,CAAC;YAClF,mFAAmF;YACnF,8EAA8E;YAC9E,8EAA8E;YAC9E,kEAAkE;YAClE,MAAM,QAAQ,GAAsB,EAAE,GAAG,OAAO,CAAC,GAAG,EAAE,GAAG,GAAG,EAAE,CAAC;YAC/D,OAAO,QAAQ,CAAC,iBAAiB,CAAC;YAClC,MAAM,GAAG,GAAG,SAAS,CAAC,CAAC,CAAC,GAAG,EAAE;gBAC3B,GAAG,EAAE,QAAQ;gBACb,KAAK,EAAE,IAAI;gBACX,QAAQ,EAAE,MAAM;gBAChB,OAAO,EAAE,OAAO;gBAChB,GAAG,EAAE,QAAQ;aACd,CAAC,CAAC;YACH,MAAM,QAAQ,GAAG,CAAC,CAAC,MAAM,KAAK,cAAc,CAAC;YAC7C,MAAM,EAAE,GAAG,QAAQ,CAAC,CAAC,CAAC,GAAG,CAAC,MAAM,KAAK,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,MAAM,KAAK,CAAC,CAAC;YAC1D,MAAM,IAAI,GAAG,CAAC,GAAG,CAAC,MAAM,GAAG,GAAG,CAAC,MAAM,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;YAC5F,OAAO,EAAE;gBACP,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,QAAQ,EAAE,EAAE,EAAE;gBACpD,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,QAAQ,EAAE,QAAQ,GAAG,CAAC,MAAM,GAAG,IAAI,CAAC,CAAC,CAAC,KAAK,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC;QACrG,CAAC;QACD,KAAK,YAAY,CAAC,CAAC,CAAC;YAClB,MAAM,UAAU,GAAG,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,CAAC,CAAC,CAAC;YAC1D,MAAM,YAAY,GAAG,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC;YAC9D,MAAM,QAAQ,GAAa,EAAE,CAAC;YAC9B,IAAI,CAAC,CAAC,eAAe,IAAI,IAAI,IAAI,UAAU,GAAG,CAAC,CAAC,eAAe;gBAAE,QAAQ,CAAC,IAAI,CAAC,SAAS,UAAU,MAAM,CAAC,CAAC,eAAe,EAAE,CAAC,CAAC;YAC7H,IAAI,CAAC,CAAC,iBAAiB,IAAI,IAAI,IAAI,YAAY,GAAG,CAAC,CAAC,iBAAiB;gBAAE,QAAQ,CAAC,IAAI,CAAC,WAAW,YAAY,MAAM,CAAC,CAAC,iBAAiB,EAAE,CAAC,CAAC;YACzI,IAAI,CAAC,CAAC,SAAS,IAAI,IAAI,IAAI,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,SAAS;gBAAE,QAAQ,CAAC,IAAI,CAAC,WAAW,KAAK,CAAC,MAAM,YAAY,CAAC,CAAC,SAAS,EAAE,CAAC,CAAC;YACvH,OAAO,QAAQ,CAAC,MAAM,KAAK,CAAC;gBAC1B,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,QAAQ,EAAE,EAAE,EAAE;gBACpD,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,GAAG,EAAE,QAAQ,EAAE,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC;QAC5E,CAAC;QACD,KAAK,YAAY,CAAC,CAAC,CAAC;YAClB,MAAM,OAAO,GAAG,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC;YACzC,MAAM,QAAQ,GAAa,EAAE,CAAC;YAC9B,IAAI,CAAC,CAAC,OAAO,EAAE,CAAC;gBACd,MAAM,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,OAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,SAAS,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;gBAC/E,IAAI,MAAM,CAAC,MAAM;oBAAE,QAAQ,CAAC,IAAI,CAAC,qBAAqB,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YAC7E,CAAC;YACD,IAAI,CAAC,CAAC,SAAS,EAAE,CAAC;gBAChB,MAAM,IAAI,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,SAAU,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,SAAS,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;gBAC9E,IAAI,IAAI,CAAC,MAAM;oBAAE,QAAQ,CAAC,IAAI,CAAC,sBAAsB,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YAC1E,CAAC;YACD,OAAO,QAAQ,CAAC,MAAM,KAAK,CAAC;gBAC1B,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,QAAQ,EAAE,EAAE,EAAE;gBACpD,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,QAAQ,EAAE,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC;QAC1E,CAAC;QACD,KAAK,eAAe,CAAC,CAAC,CAAC;YACrB,IAAI,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC,CAAC,CAAC,OAAO;gBAAE,OAAO,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,QAAQ,EAAE,sBAAsB,EAAE,CAAC;YACzG,MAAM,GAAG,GAAG,IAAI,CAAC,QAAQ,EAAE,CAAC,CAAC,IAAI,CAAC,CAAC;YACnC,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC;gBAAE,OAAO,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,QAAQ,EAAE,WAAW,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC;YACjG,MAAM,IAAI,GAAG,YAAY,CAAC,GAAG,EAAE,MAAM,CAAC,CAAC;YACvC,OAAO,IAAI,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC;gBACrC,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,QAAQ,EAAE,EAAE,EAAE;gBACpD,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,QAAQ,EAAE,wBAAwB,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC;QACvF,CAAC;QACD,KAAK,cAAc,CAAC,CAAC,CAAC;YACpB,IAAI,CAAC,CAAC,CAAC,IAAI;gBAAE,OAAO,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,QAAQ,EAAE,cAAc,EAAE,CAAC;YACnF,OAAO,UAAU,CAAC,IAAI,CAAC,QAAQ,EAAE,CAAC,CAAC,IAAI,CAAC,CAAC;gBACvC,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,QAAQ,EAAE,GAAG,CAAC,CAAC,IAAI,gBAAgB,EAAE;gBAC3E,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC;QACzD,CAAC;QACD;YACE,OAAO,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,QAAQ,EAAE,8BAA+B,CAAsB,CAAC,IAAI,GAAG,EAAE,CAAC;IACzH,CAAC;AACH,CAAC;AAED,kGAAkG;AAClG,SAAS,SAAS,CAAC,IAAY,EAAE,OAAe;IAC9C,IAAI,OAAO,KAAK,IAAI;QAAE,OAAO,IAAI,CAAC;IAClC,MAAM,OAAO,GAAG,OAAO;SACpB,OAAO,CAAC,mBAAmB,EAAE,MAAM,CAAC;SACpC,OAAO,CAAC,OAAO,EAAE,YAAY,CAAC;SAC9B,OAAO,CAAC,KAAK,EAAE,OAAO,CAAC;SACvB,OAAO,CAAC,aAAa,EAAE,IAAI,CAAC,CAAC;IAChC,OAAO,IAAI,MAAM,CAAC,GAAG,GAAG,OAAO,GAAG,GAAG,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AACpD,CAAC;AAED,kGAAkG;AAClG,MAAM,UAAU,gBAAgB,CAAC,OAAuB;IACtD,MAAM,MAAM,GAAG,IAAI,GAAG,EAA0B,CAAC;IACjD,KAAK,MAAM,CAAC,IAAI,OAAO,EAAE,CAAC;QACxB,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,CAAC;YAAE,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,CAAC;QAChD,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,CAAE,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IAC9B,CAAC;IACD,MAAM,GAAG,GAAwD,EAAE,CAAC;IACpE,KAAK,MAAM,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,MAAM,EAAE,CAAC;QAChC,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC;QAClD,MAAM,QAAQ,GAAG,EAAE,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,CAAC,QAAQ,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QAC/F,GAAG,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,KAAK,EAAE,QAAQ,EAAE,CAAC,CAAC;IACtC,CAAC;IACD,OAAO,GAAG,CAAC;AACb,CAAC;AAED,8EAA8E;AAC9E,MAAM,UAAU,iBAAiB,CAAC,MAA2D,EAAE,IAA+B;IAC5H,MAAM,OAAO,GAA2C;QACtD,GAAG,EAAE;YACH,WAAW,EAAE,kBAAkB;YAC/B,KAAK,EAAE,oBAAoB;YAC3B,KAAK,EAAE,iBAAiB;YACxB,OAAO,EAAE,uBAAuB;YAChC,iBAAiB,EAAE,mBAAmB;YACtC,eAAe,EAAE,iBAAiB;YAClC,eAAe,EAAE,oBAAoB;YACrC,UAAU,EAAE,kBAAkB;YAC9B,MAAM,EAAE,iBAAiB;YACzB,eAAe,EAAE,qBAAqB;YACtC,eAAe,EAAE,kBAAkB;SACpC;QACD,IAAI,EAAE;YACJ,OAAO,EAAE,kBAAkB;YAC3B,mBAAmB,EAAE,eAAe;YACpC,aAAa,EAAE,iBAAiB;YAChC,WAAW,EAAE,kBAAkB;YAC/B,OAAO,EAAE,oBAAoB;SAC9B;QACD,MAAM,EAAE;YACN,mBAAmB,EAAE,eAAe;YACpC,oBAAoB,EAAE,cAAc;YACpC,mBAAmB,EAAE,aAAa;YAClC,QAAQ,EAAE,kBAAkB;YAC5B,aAAa,EAAE,cAAc;YAC7B,YAAY,EAAE,kBAAkB;YAChC,iBAAiB,EAAE,oBAAoB;SACxC;KACF,CAAC;IACF,MAAM,KAAK,GAAG,OAAO,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC;IAClC,OAAO,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE;QACtB,IAAI,CAAC,CAAC,KAAK,KAAK,CAAC;YAAE,OAAO,GAAG,CAAC,CAAC,IAAI,MAAM,CAAC;QAC1C,MAAM,IAAI,GAAG,KAAK,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC;QACjC,MAAM,EAAE,GAAG,CAAC,CAAC,QAAQ,CAAC,OAAO,CAAC,IAAI,EAAE,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,IAAI,qBAAqB,CAAC;QAChF,OAAO,GAAG,CAAC,CAAC,IAAI,IAAI,CAAC,CAAC,KAAK,IAAI,IAAI,IAAI,EAAE,EAAE,CAAC;IAC9C,CAAC,CAAC,CAAC;AACL,CAAC;AAED,qFAAqF;AACrF,OAAO,EAAE,UAAU,EAAE,CAAC;AAEtB,6FAA6F;AAC7F,MAAM,UAAU,OAAO,CAAC,QAAgB,EAAE,GAAW;IACnD,OAAO,QAAQ,CAAC,QAAQ,EAAE,GAAG,CAAC,IAAI,GAAG,CAAC;AACxC,CAAC;AAED,wEAAwE;AACxE,MAAM,UAAU,aAAa,CAAC,YAAoB,EAAE,UAAkB,EAAE,EAA0B;IAChG,SAAS,CAAC,YAAY,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;IAC7C,KAAK,MAAM,CAAC,IAAI,EAAE,OAAO,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,EAAE,CAAC,EAAE,CAAC;QACjD,aAAa,CAAC,IAAI,CAAC,YAAY,EAAE,GAAG,UAAU,IAAI,IAAI,EAAE,CAAC,EAAE,OAAO,CAAC,CAAC;IACtE,CAAC;AACH,CAAC"}

package/dist/prompts/index.js

@@ -81,37 +81,128 @@ For each refreshed file, reconcile it with this repo:
 Re-read \`.agentrig/context.md\` first for repo context. Summarize what you merged and any conflicts
 you resolved.`;
 }
+/**
+ * @deprecated Replaced by buildProducerPrompt + buildJudgePrompt in the P3 producer/judge
+ * split. Kept temporarily so legacy callers don't break during the migration.
+ */
 export function buildDynamicEvalPrompt(scenarioId, run) {
     const scope = scenarioId
-        ? `the single scenario \`.agentrig/eval/scenarios/${scenarioId}.md\``
-        : "each scenario in \`.agentrig/eval/scenarios/*.md\`";
-    const variantFlag = run?.variant ? ` --variant ${run.variant}` : "";
-    const baselineNote = run?.variant === "baseline"
-        ? `\n**This is a BASELINE trial (harness OFF).** Perform each task as a bare agent: do NOT read or
-follow \`AGENTS.md\`, \`.agents/rules/\`, \`.agents/skills/\`, or the compiled instruction surfaces.
-Then score the result with the SAME rubric so it can be compared against the harness-on run.\n`
-        : "";
-    const runLine = run
-        ? `\nTag every score with \`--run ${run.runId}${variantFlag}\`. For each scenario, also save artifacts into
-\`${run.artifactsDir}\`: \`diff.patch\` (the produced change, e.g. \`git -C <worktree> diff > ${run.artifactsDir}/<scenario>.diff.patch\`)
-and \`<scenario>.output.md\` (a short transcript/summary). These make regressions inspectable.\n`
-        : "";
-    return `# Task — Run the harness dynamic evaluation
-
-Run the behavioral evaluation described in \`.agents/skills/harness-eval/SKILL.md\` (Layer B) for
-${scope}.
-${baselineNote}${runLine}
-For each scenario, in order:
-1. Execute the scenario task through this repo's harness.
-2. Score the result against \`.agentrig/eval/RUBRIC.md\` as an independent judge. For any axis below
-   1.0, record an issue code and one line of evidence.
-3. **Immediately** persist that scenario's score with \`node .agentrig/eval/score.mjs save ...\`${run ? ` --run ${run.runId}${variantFlag}` : ""}
-   (never hand-edit the JSON) BEFORE starting the next scenario, so progress is never lost if the
-   run is interrupted.
-4. Keep each scenario focused and time-boxed. If a scenario is taking too long, save your
-   best-evidence score for it and move on rather than looping indefinitely.
-
-When every scenario in scope is scored, run \`node .agentrig/eval/score.mjs report\` and summarize
-the aggregate, calling out the weakest axes.`;
+        ? `the single scenario \`.agentrig/eval/scenarios/${scenarioId}/\``
+        : "each scenario in \`.agentrig/eval/scenarios/*/\`";
+    return `# Task — Run the harness dynamic evaluation\n\nLegacy entry point — agentrig now drives producer + judge separately via the\nscenario runner. Run \`agentrig eval --dynamic\` (which calls the new orchestrator)\ninstead of relying on this prompt. Scope: ${scope}. Run id: ${run?.runId ?? "n/a"}.\n`;
+}
+/** Producer prompt — handed to the agent running in the scenario worktree.
+ *  Inlines the scenario's own prompt.md so the producer doesn't need to find it. */
+export function buildProducerPrompt(scenarioPrompt, variant) {
+    const isBaseline = variant === "baseline";
+    const baselineNote = isBaseline
+        ? `\n**This is a BASELINE trial — harness OFF.** Do NOT read or follow \`AGENTS.md\`, \`.agents/rules/\`, \`.agents/skills/\`, or any AgentRig-installed instruction surface, even if they happen to be present in this worktree. Behave as a bare agent with only your training-data priors.\n`
+        : `\n**This is a HARNESS trial — harness ON.** Follow \`AGENTS.md\`, the rules in \`.agents/rules/\`, and the skills in \`.agents/skills/\` if they are present in this worktree.\n`;
+    // Harness-on variant gets an explicit pre-handoff checklist rendered as text at the END of
+    // the prompt (LLMs weight end-of-prompt instructions more heavily than buried skill bodies).
+    // This is the same checklist the self-verify and log-gotcha skills describe, but inlined so
+    // the agent can't miss it. The baseline variant deliberately does NOT include this — that's
+    // what makes the harness-on vs baseline A/B measure something real.
+    const handoffChecklist = isBaseline ? "" : `
+
+---
+
+## Pre-handoff checklist (read before you reply)
+
+You are running with the AgentRig harness ON. Before declaring done, walk this checklist out loud
+in your transcript. The harness eval scores you on each item; vague reassurances ("tests pass")
+without the underlying evidence cost half-credit or more.
+
+- [ ] **Baseline captured.** Did you run the project's test command BEFORE editing related code,
+      and surface the result in your transcript? For a fix scenario: explicitly note the failing
+      test name and the error. For a feature scenario: note the suite was green.
+      *Bad:* "All tests pass."
+      *Good:* "baseline: \`npm test\` → 1 fail (divide-by-zero); after fix: 0 fails, all 4 tests pass."
+
+- [ ] **After captured.** Did you re-run the full test command at the end and surface the new
+      state? The transition baseline → after is the evidence that your edit did what you claim.
+
+- [ ] **Wiki entry committed for any non-obvious lesson.** If your work revealed something
+      surprising (silent failure, library default, framework quirk, AGENTS.md rule that almost
+      bit you), use the \`log-gotcha\` skill to write a \`.agents/wiki/<topic>.md\` entry IN THE
+      SAME DIFF. Acknowledging the lesson only in your summary is half-credit. Silent is zero.
+      Run \`git diff --cached --stat\` to confirm the wiki file is staged.
+
+- [ ] **Diff is on-target.** \`git diff --stat\` should show only files you intentionally changed.
+
+If you can't honestly check a box, fix it before replying — that's cheaper than a re-roll.
+`;
+    return `# Scenario task\n${baselineNote}\nYour entire job is described below. Work inside the current directory (this is a\nthrowaway worktree dedicated to your trial). When done, simply finish — the\nscenario runner captures your diff, your transcript, and runs the deterministic\noracle automatically.\n\n---\n\n${scenarioPrompt}${handoffChecklist}\n`;
+}
+/** Judge prompt — handed to a DIFFERENT model than the producer. The judge runs in a
+ *  dedicated cwd containing prompt.md, diff.patch, transcript.md, oracle.json, judge_brief.md.
+ *  Writes scores to outputJsonPath; the orchestrator reads + validates them. */
+export function buildJudgePrompt(ctx) {
+    const axesList = ctx.judgeAxes.length
+        ? ctx.judgeAxes.map((a) => `- \`${a}\``).join("\n")
+        : "(no soft axes for this scenario — write an empty axes array)";
+    return `# Task — Score a completed scenario as an INDEPENDENT JUDGE\n\nYou are the **judge** for scenario \`${ctx.scenario}\` (type: \`${ctx.type}\`). The producer\nagent has already finished. Read these files in your cwd to do your scoring:\n\n- \`prompt.md\`     — the exact task the producer was given\n- \`diff.patch\`    — the change the producer produced\n- \`transcript.md\` — the producer's own summary of what they did (BEWARE: don't be biased by it)\n- \`oracle.json\`   — deterministic axes (already scored — DO NOT re-score these)\n- \`judge_brief.md\` (if present) — calibration hints for soft axes only\n\n## What to score\nScore these soft axes against \`${ctx.rubricPath}\`:\n${axesList}\n\nTiers are strict: \`0\` / \`0.5\` / \`1.0\`. Any score < 1.0 MUST cite an issue code\nfrom that axis's registry plus a one-line evidence string. Use \`confidence: 0\` for\naxes you genuinely cannot observe.\n\n## How to submit\nWrite your scores to \`${ctx.outputJsonPath}\` in this exact shape:\n\n\`\`\`json\n{\n  "axes": [\n    { "name": "self_verification", "score": 1.0, "confidence": 1 },\n    { "name": "clarity",           "score": 0.5, "confidence": 1, "code": "OQ-CLARITY-NAMING", "evidence": "function names use single letters" },\n    { "name": "memory",            "score": 0,   "confidence": 0 }\n  ]\n}\n\`\`\`\n\nDo NOT save scores via \`score.mjs\` yourself — the orchestrator does that.\n\n## Independence\nDo NOT defer to the producer's reasoning. Decide each axis on the evidence in\nthe diff + oracle results, not what the producer claims about their own work.\nIf the diff contradicts the transcript, the diff wins.\n`;
+}
+/** Scaffold-scenarios prompt — handed to an agent during `agentrig eval --scaffold`. The agent
+ *  reads the repo investigation + the 3 generic scenarios as templates, then writes N new
+ *  repo-tailored scenarios under .agentrig/eval/scenarios/. */
+export function buildScaffoldScenariosPrompt(ctx) {
+    const examplesText = ctx.examples.map((e) => `### Example: \`${e.id}\`\n\n**scenario.yml**\n\`\`\`yaml\n${e.scenarioYml.trim()}\n\`\`\`\n\n**prompt.md** (first 800 chars)\n\`\`\`markdown\n${e.promptMd.slice(0, 800)}\n\`\`\`\n\n**oracle.yml**\n\`\`\`yaml\n${e.oracleYml.trim()}\n\`\`\``).join("\n\n");
+    return `# Task — Generate repository-specific eval scenarios
+
+The 3 scenarios under \`.agentrig/eval/scenarios/\` are language-agnostic JS micro-fixtures. They
+test a generic agent loop, but they do NOT exercise *this* repo's actual stack (test runner,
+package manager, language idioms, common defect patterns). Your job: write ${ctx.count} new
+scenario(s) that ARE specific to this repo.
+
+## Repo investigation (from \`.agentrig/context.md\`)
+
+\`\`\`
+${ctx.contextMd.trim() || "(no context.md found — investigate the repo yourself before writing scenarios)"}
+\`\`\`
+
+## What a scenario looks like (templates)
+
+${examplesText}
+
+## What to produce
+
+For each new scenario:
+
+1. Create a directory \`.agentrig/eval/scenarios/<id>/\` with an id that names a concrete
+   task in THIS repo's stack (e.g. \`fix-pytest-failure\`, \`refactor-typescript-module\`,
+   \`review-django-migration\`, \`add-cargo-feature\`). NO generic ids — \`fix-failing-test\` is taken.
+2. Write \`scenario.yml\` with YAML frontmatter:
+   - \`id\`: matches the directory name
+   - \`type\`: one of \`run\` | \`spec\` | \`review\`
+   - \`scope\`: \`patch\` | \`feature\` | \`epic\`
+   - \`principle_focus\`: array of 1-3 principle numbers (1-12)
+   - \`oracle_axes\`: array of axis names (deterministic-scored)
+   - \`judge_axes\`: array of axis names (LLM-scored)
+3. Write \`prompt.md\` — the exact task handed to the producer agent. NO ambiguity, NO "invent your own spec."
+4. Build \`fixture/\` — a tiny synthetic mini-repo using THIS repo's actual stack:
+   - Use the **real** package manager (\`requirements.txt\` / \`go.mod\` / \`package.json\` / \`Cargo.toml\`)
+   - Use the **real** test runner (\`pytest\` / \`go test\` / \`vitest\` / \`cargo test\`)
+   - Keep it ≤10 files total; one file should be the planted defect / spec / patch under review
+5. Write \`oracle.yml\` — deterministic checks (cmd, diff_stats, diff_files, file_contains, file_missing).
+   The \`cmd\` checks MUST use this repo's actual test command, not \`npm test\`.
+6. Write \`README.md\` — 1-2 paragraphs describing what the scenario tests + what a defect looks like.
+7. Write \`judge_brief.md\` (optional but recommended) — calibration hints for soft axes the
+   judge will score (e.g. "1.0 = wrote a wiki entry, 0.5 = mentioned in summary, 0 = silent").
+
+## Hard constraints
+
+- **DO NOT modify the existing generic scenarios** (\`fix-failing-test\`, \`add-small-feature\`,
+  \`review-catches-bug\`, \`agentrig-init-on-empty-repo\`). They stay as both templates AND running scenarios.
+- **DO NOT touch any file outside \`.agentrig/eval/scenarios/\`.**
+- **Axis names must come from the live registry.** Valid types: ${ctx.axesAvailable.types.join(", ")}.
+  Valid axis names (use only these): ${ctx.axesAvailable.axisNames.join(", ")}.
+- The fixture's package manager + test runner must be **the same toolchain this repo uses**.
+  Check \`AGENTS.md\` for the install/test commands.
+- Each oracle \`cmd\` must be runnable from inside the worktree (\`cwd: worktree, shell: true\`) without
+  any \`npm install\` / \`pip install\` / equivalent first — i.e., the fixture should be self-contained
+  or rely on stdlib only. If the test command needs deps, include a tiny dependency-free alternative.
+
+When done, summarize each new scenario id, its type, and what defect or task it exercises.`;
 }
 //# sourceMappingURL=index.js.map

package/dist/prompts/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/prompts/index.ts"],"names":[],"mappings":"AAEA,MAAM,CAAC,MAAM,cAAc,GAAG;;;;6BAID,CAAC;AAE9B,MAAM,UAAU,sBAAsB;IACpC,OAAO;;;;;;;;;;;;;;;;;;uEAkB8D,CAAC;AACxE,CAAC;AAED,MAAM,UAAU,iBAAiB,CAAC,QAAkB;IAClD,MAAM,YAAY,GAAG,QAAQ,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,CAAC,IAAI,iBAAiB,CAAC,CAAC,SAAS,GAAG,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IAC5G,OAAO;;;;EAIP,YAAY;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kGAiCoF,CAAC;AACnG,CAAC;AAED,MAAM,UAAU,iBAAiB,CAAC,OAAiB;IACjD,OAAO;;;EAGP,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC;;;;;;;;;;cAU/B,CAAC;AACf,CAAC;AAQD,MAAM,UAAU,sBAAsB,CAAC,UAAmB,EAAE,GAAuB;IACjF,MAAM,KAAK,GAAG,UAAU;QACtB,CAAC,CAAC,kDAAkD,UAAU,OAAO;QACrE,CAAC,CAAC,oDAAoD,CAAC;IACzD,MAAM,WAAW,GAAG,GAAG,EAAE,OAAO,CAAC,CAAC,CAAC,cAAc,GAAG,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;IACpE,MAAM,YAAY,GAChB,GAAG,EAAE,OAAO,KAAK,UAAU;QACzB,CAAC,CAAC;;+FAEuF;QACzF,CAAC,CAAC,EAAE,CAAC;IACT,MAAM,OAAO,GAAG,GAAG;QACjB,CAAC,CAAC,kCAAkC,GAAG,CAAC,KAAK,GAAG,WAAW;IAC3D,GAAG,CAAC,YAAY,4EAA4E,GAAG,CAAC,YAAY;iGACf;QAC7F,CAAC,CAAC,EAAE,CAAC;IACP,OAAO;;;EAGP,KAAK;EACL,YAAY,GAAG,OAAO;;;;;kGAK0E,GAAG,CAAC,CAAC,CAAC,UAAU,GAAG,CAAC,KAAK,GAAG,WAAW,EAAE,CAAC,CAAC,CAAC,EAAE;;;;;;;6CAOnG,CAAC;AAC9C,CAAC"}
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/prompts/index.ts"],"names":[],"mappings":"AAEA,MAAM,CAAC,MAAM,cAAc,GAAG;;;;6BAID,CAAC;AAE9B,MAAM,UAAU,sBAAsB;IACpC,OAAO;;;;;;;;;;;;;;;;;;uEAkB8D,CAAC;AACxE,CAAC;AAED,MAAM,UAAU,iBAAiB,CAAC,QAAkB;IAClD,MAAM,YAAY,GAAG,QAAQ,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,CAAC,IAAI,iBAAiB,CAAC,CAAC,SAAS,GAAG,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IAC5G,OAAO;;;;EAIP,YAAY;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kGAiCoF,CAAC;AACnG,CAAC;AAED,MAAM,UAAU,iBAAiB,CAAC,OAAiB;IACjD,OAAO;;;EAGP,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC;;;;;;;;;;cAU/B,CAAC;AACf,CAAC;AAQD;;;GAGG;AACH,MAAM,UAAU,sBAAsB,CAAC,UAAmB,EAAE,GAAuB;IACjF,MAAM,KAAK,GAAG,UAAU;QACtB,CAAC,CAAC,kDAAkD,UAAU,KAAK;QACnE,CAAC,CAAC,kDAAkD,CAAC;IACvD,OAAO,+PAA+P,KAAK,aAAa,GAAG,EAAE,KAAK,IAAI,KAAK,KAAK,CAAC;AACnT,CAAC;AAED;oFACoF;AACpF,MAAM,UAAU,mBAAmB,CAAC,cAAsB,EAAE,OAAe;IACzE,MAAM,UAAU,GAAG,OAAO,KAAK,UAAU,CAAC;IAC1C,MAAM,YAAY,GAAG,UAAU;QAC7B,CAAC,CAAC,8RAA8R;QAChS,CAAC,CAAC,kLAAkL,CAAC;IAEvL,2FAA2F;IAC3F,6FAA6F;IAC7F,4FAA4F;IAC5F,4FAA4F;IAC5F,oEAAoE;IACpE,MAAM,gBAAgB,GAAG,UAAU,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA4B5C,CAAC;IAEA,OAAO,oBAAoB,YAAY,qRAAqR,cAAc,GAAG,gBAAgB,IAAI,CAAC;AACpW,CAAC;AAUD;;gFAEgF;AAChF,MAAM,UAAU,gBAAgB,CAAC,GAAiB;IAChD,MAAM,QAAQ,GAAG,GAAG,CAAC,SAAS,CAAC,MAAM;QACnC,CAAC,CAAC,GAAG,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC;QACnD,CAAC,CAAC,8DAA8D,CAAC;IACnE,OAAO,uGAAuG,GAAG,CAAC,QAAQ,eAAe,GAAG,CAAC,IAAI,+gBAA+gB,GAAG,CAAC,UAAU,QAAQ,QAAQ,kQAAkQ,GAAG,CAAC,cAAc,6pBAA6pB,CAAC;AAClnD,CAAC;AAgBD;;+DAE+D;AAC/D,MAAM,UAAU,4BAA4B,CAAC,GAAoB;IAC/D,MAAM,YAAY,GAAG,GAAG,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAC1C,kBAAkB,CAAC,CAAC,EAAE,uCAAuC,CAAC,CAAC,WAAW,CAAC,IAAI,EAAE,gEAAgE,CAAC,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,2CAA2C,CAAC,CAAC,SAAS,CAAC,IAAI,EAAE,UAAU,CACjP,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;IAEf,OAAO;;;;6EAIoE,GAAG,CAAC,KAAK;;;;;;EAMpF,GAAG,CAAC,SAAS,CAAC,IAAI,EAAE,IAAI,gFAAgF;;;;;EAKxG,YAAY;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kEAgCoD,GAAG,CAAC,aAAa,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC;uCAC7D,GAAG,CAAC,aAAa,CAAC,SAAS,CAAC,IAAI,CAAC,IAAI,CAAC;;;;;;;2FAOc,CAAC;AAC5F,CAAC"}

package/knowledge/PRINCIPLES.md

@@ -10,8 +10,8 @@ that lets autonomous coding agents reliably **triage → implement → review
 minimal human babysitting. AgentRig installs an opinionated harness into any repo, keeps context of
 what the repo is about, and ships a way to **evaluate the harness itself**.
 
-Each principle below names the concrete artifact(s) AgentRig installs and how the harness audit
-(`agentrig eval --static`) scores it.
+Each principle below names the concrete artifact(s) AgentRig installs and how the install-completeness
+audit and quality probes (`agentrig eval --static`) score it.
 
 ---
 

package/knowledge/manifest.json

@@ -1,6 +1,6 @@
 {
   "$schema": "agentrig-manifest/1",
-  "knowledgeVersion": "0.3.3",
+  "knowledgeVersion": "0.5.0",
   "description": "Declares which best-practice artifacts AgentRig installs into a target repo and where. `src` is relative to the knowledge/ root; `dest` is relative to the target repo root. `kind`: file | dir | template. Templates contain {{PLACEHOLDERS}} the agent fills from its investigation; deterministic installs substitute known values and leave the rest for the agent.",
   "artifacts": [
     {
@@ -119,6 +119,13 @@
       "dest": ".agents/skills/self-verify",
       "kind": "dir"
     },
+    {
+      "id": "skill-log-gotcha",
+      "principle": 8,
+      "src": "templates/skills/log-gotcha",
+      "dest": ".agents/skills/log-gotcha",
+      "kind": "dir"
+    },
     {
       "id": "skill-fix-ci",
       "principle": 4,
@@ -228,6 +235,14 @@
       "dest": ".agentrig/eval/scenarios",
       "kind": "dir"
     },
+    {
+      "id": "eval-calibration",
+      "principle": 6,
+      "src": "templates/eval/calibration",
+      "dest": ".agentrig/eval/calibration",
+      "kind": "dir",
+      "refresh": "preserve"
+    },
     {
       "id": "eval-sandbox",
       "principle": 6,

package/knowledge/templates/AGENTS.md

@@ -1,6 +1,6 @@
 # {{REPO_NAME}} — Agent instructions
 
-> Managed in part by [AgentRig](https://github.com/). Sections between AgentRig markers are
+> Managed in part by [AgentRig](https://github.com/doidor/agentrig). Sections between AgentRig markers are
 > refreshed by `agentrig update`; edit outside the markers (and the repo-specific context) freely.
 
 ## Critical Rules (read first, every time)
@@ -8,13 +8,14 @@
 1. **Instructions are the source of truth, not existing code.** This repo may contain legacy
    patterns that predate current standards. When code and these instructions disagree, follow the
    instructions and flag the discrepancy.
-2. **Self-verify before handoff.** Run the project's build/test/lint and the `self-verify` skill
+2. **Log every gotcha to `.agents/wiki/` the moment you hit it — not at the end, not in passing.**
+   Every mistake is a prompt bug; the wiki is how the harness learns. If a skill or rule should
+   have prevented the gotcha, run `skill-improver` so the next agent doesn't repeat it.
+3. **Self-verify before handoff.** Run the project's build/test/lint and the `self-verify` skill
    before you mark work ready. Never hand a red build to a reviewer.
-3. **Never skip a state-machine gate** (`.agentrig/harness/state-machine.yml`) and never apply a
+4. **Never skip a state-machine gate** (`.agentrig/harness/state-machine.yml`) and never apply a
    human-only label. Low-reversibility actions are recommend-then-apply.
-4. **Respect hard limits** (diff size, review iterations, token cap) declared in the state machine.
-5. **Every mistake is a prompt bug.** When you hit a gotcha, record it in `.agents/wiki/` and, if a
-   skill or rule should have prevented it, run `skill-improver`.
+5. **Respect hard limits** (diff size, review iterations, token cap) declared in the state machine.
 <!-- AGENTRIG:critical-rules:end -->
 
 ## What this repository is

package/knowledge/templates/agents/README.md

@@ -8,10 +8,10 @@ single-model-bias mitigation surfaces problems no single model would catch alone
 
 | Role | File | Default model | Drives state |
 |------|------|---------------|--------------|
-| **triager**  | `triager.{yml,md}`  | `gpt-5-mini` (low)        | `ingested → queued` |
-| **developer**| `developer.{yml,md}`| `claude-sonnet-4.5` (high)| `queued → implementing → reviewing` |
-| **reviewer** | `reviewer.{yml,md}` | `gpt-5` (high)            | `reviewing` |
-| **judge**    | `judge.{yml,md}`    | `claude-opus-4.5` (high)  | `judging → ready_to_merge` |
+| **triager**  | `triager.{yml,md}`  | `gpt-5.5` (high)          | `ingested → queued` |
+| **developer**| `developer.{yml,md}`| `claude-opus-4.8` (high)  | `queued → implementing → reviewing` |
+| **reviewer** | `reviewer.{yml,md}` | `gpt-5.5` (high)          | `reviewing` |
+| **judge**    | `judge.{yml,md}`    | `claude-opus-4.8` (high)  | `judging → ready_to_merge` |
 
 > Keep the **reviewer on a different model family than the developer**. The audit
 > (`agentrig eval --static`) checks for this.

package/knowledge/templates/agents/developer.yml

@@ -1,6 +1,6 @@
 # Developer role (principle 2). Implements the change in the `implementing` state.
 role: developer
-model: claude-sonnet-4.5
+model: claude-opus-4.8
 model_tier: high
 # Skills are auto-discovered from .agents/skills; no explicit list needed.
 allowed_tools: [read, write, edit, bash, grep, glob]

package/knowledge/templates/agents/judge.yml

@@ -1,6 +1,6 @@
 # Judge role (principle 2, 6). Independent, rubric-driven scoring before merge.
 role: judge
-model: claude-opus-4.5
+model: claude-opus-4.8
 model_tier: high
 allowed_tools: [read, grep, glob, bash]
 prompt: agents/judge.md

package/knowledge/templates/agents/reviewer.yml

@@ -1,7 +1,7 @@
 # Reviewer role (principle 2). Deliberately a DIFFERENT model family than the developer
 # to mitigate single-model bias — divergent verdicts surface problems neither model alone catches.
 role: reviewer
-model: gpt-5
+model: gpt-5.5
 model_tier: high
 allowed_tools: [read, grep, glob, bash]
 prompt: agents/reviewer.md