@skill-harness/core 0.19.0 → 0.20.0

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

@@ -1,10 +1,8 @@
 /**
- * Data contracts for the pi-native regression capture program.
+ * Persisted data contracts for Pi execution traces.
  *
- * Types only — no behavior. Phase 0 of
- * `docs/superpowers/plans/2026-08-07-pi-native-regression-capture-program.md`
- * fixes the shapes before any parser, capture UI, or gate evaluator is written,
- * because both shapes get persisted and a persisted shape is expensive to move.
+ * Types only — no behavior. Both shapes are versioned because persisted formats
+ * are expensive to move and readers must fail closed on unsupported versions.
  *
  * Both carry an explicit version. The pi event stream they derive from is not a
  * stable public contract (measured against pi 0.83.0; see

package/dist/capture-trace-types.js

@@ -1,10 +1,8 @@
 /**
- * Data contracts for the pi-native regression capture program.
+ * Persisted data contracts for Pi execution traces.
  *
- * Types only — no behavior. Phase 0 of
- * `docs/superpowers/plans/2026-08-07-pi-native-regression-capture-program.md`
- * fixes the shapes before any parser, capture UI, or gate evaluator is written,
- * because both shapes get persisted and a persisted shape is expensive to move.
+ * Types only — no behavior. Both shapes are versioned because persisted formats
+ * are expensive to move and readers must fail closed on unsupported versions.
  *
  * Both carry an explicit version. The pi event stream they derive from is not a
  * stable public contract (measured against pi 0.83.0; see

package/dist/defaults.d.ts

@@ -1,30 +1,35 @@
+import { type ModelRef } from "./adapters/types.js";
 /**
- * The judge used when nothing else says otherwise: Opus through the **`claude-code`
- * provider**, which authenticates with the user's Claude subscription (OAuth via
- * `claude -p`) rather than a metered API key.
+ * The judge used when nothing else says otherwise: a Pi `openai-codex` model,
+ * authenticated through the user's ChatGPT subscription rather than a metered API
+ * key. Subjects and judges now share the sole supported harness executable: `pi`.
  *
- * The model is deliberately the strongest available — judging is the one place
- * where a weak model silently corrupts every number in a scorecard. The *provider*
- * is what changed in 0.3.3: the default was `anthropic:claude-opus-4-8`, a metered
- * API, and it billed a corpus once by accident because nothing in the tool surface
- * distinguishes "the flag I forgot" from "the flag I meant". A default must not be
- * able to spend money that was not asked for. The metered path is still one flag
- * away (`--judge anthropic:claude-opus-4-8`) for anyone who wants it — an API key
- * scales past a subscription's rate limits, which matters for a large `--reps` run.
+ * A default must not be able to spend money that was not explicitly authorized.
+ * Direct API providers remain available only through the metered-judge opt-in.
  */
-export declare const BAKED_DEFAULT_JUDGE = "claude-code:claude-opus-4-8";
+export declare const BAKED_DEFAULT_JUDGE = "openai-codex:gpt-5.6-sol";
 /**
  * Resolve the default judge: `SKILL_HARNESS_JUDGE` if set, else the baked value.
  * An explicit `--judge` always wins over both — this is only the default.
  *
  * The env layer exists because this harness is built for someone who steers a
  * process rather than typing every flag: judge policy belongs to the repo or the
- * shell, set once. It is also how you opt *into* the metered API deliberately
- * (`SKILL_HARNESS_JUDGE=anthropic:claude-opus-4-8`) instead of by forgetting a
- * flag. Read through `readEnv`, so the pre-rename `SKILL_CHECK_JUDGE` keeps working
+ * shell, set once. It is also how you opt *into* a different judge deliberately
+ * instead of repeating a flag. Read through `readEnv`, so the pre-rename
+ * `SKILL_CHECK_JUDGE` keeps working
  * with the usual one-time notice.
  *
  * Resolved per call, not at module load: tests and long-lived processes (the pi
  * extension) must see an env change without a reload.
  */
 export declare function defaultJudge(): string;
+/**
+ * Choose a judge for an artifact-only operation when the caller supplied no
+ * explicit override. Retained results may name the removed direct Claude
+ * adapter; keep those files readable, but execute their next judge call through
+ * today's Pi default and let the rewrite record that current judge.
+ */
+export declare function recordedJudgeOrDefault(recorded?: ModelRef): {
+    judge: ModelRef;
+    migratedFrom?: ModelRef;
+};

package/dist/defaults.js

@@ -1,28 +1,23 @@
+import { parseModelRef } from "./adapters/types.js";
 import { readEnv } from "./util/env.js";
 /**
- * The judge used when nothing else says otherwise: Opus through the **`claude-code`
- * provider**, which authenticates with the user's Claude subscription (OAuth via
- * `claude -p`) rather than a metered API key.
+ * The judge used when nothing else says otherwise: a Pi `openai-codex` model,
+ * authenticated through the user's ChatGPT subscription rather than a metered API
+ * key. Subjects and judges now share the sole supported harness executable: `pi`.
  *
- * The model is deliberately the strongest available — judging is the one place
- * where a weak model silently corrupts every number in a scorecard. The *provider*
- * is what changed in 0.3.3: the default was `anthropic:claude-opus-4-8`, a metered
- * API, and it billed a corpus once by accident because nothing in the tool surface
- * distinguishes "the flag I forgot" from "the flag I meant". A default must not be
- * able to spend money that was not asked for. The metered path is still one flag
- * away (`--judge anthropic:claude-opus-4-8`) for anyone who wants it — an API key
- * scales past a subscription's rate limits, which matters for a large `--reps` run.
+ * A default must not be able to spend money that was not explicitly authorized.
+ * Direct API providers remain available only through the metered-judge opt-in.
  */
-export const BAKED_DEFAULT_JUDGE = "claude-code:claude-opus-4-8";
+export const BAKED_DEFAULT_JUDGE = "openai-codex:gpt-5.6-sol";
 /**
  * Resolve the default judge: `SKILL_HARNESS_JUDGE` if set, else the baked value.
  * An explicit `--judge` always wins over both — this is only the default.
  *
  * The env layer exists because this harness is built for someone who steers a
  * process rather than typing every flag: judge policy belongs to the repo or the
- * shell, set once. It is also how you opt *into* the metered API deliberately
- * (`SKILL_HARNESS_JUDGE=anthropic:claude-opus-4-8`) instead of by forgetting a
- * flag. Read through `readEnv`, so the pre-rename `SKILL_CHECK_JUDGE` keeps working
+ * shell, set once. It is also how you opt *into* a different judge deliberately
+ * instead of repeating a flag. Read through `readEnv`, so the pre-rename
+ * `SKILL_CHECK_JUDGE` keeps working
  * with the usual one-time notice.
  *
  * Resolved per call, not at module load: tests and long-lived processes (the pi
@@ -31,4 +26,17 @@ export const BAKED_DEFAULT_JUDGE = "claude-code:claude-opus-4-8";
 export function defaultJudge() {
     return readEnv("JUDGE") ?? BAKED_DEFAULT_JUDGE;
 }
+/**
+ * Choose a judge for an artifact-only operation when the caller supplied no
+ * explicit override. Retained results may name the removed direct Claude
+ * adapter; keep those files readable, but execute their next judge call through
+ * today's Pi default and let the rewrite record that current judge.
+ */
+export function recordedJudgeOrDefault(recorded) {
+    if (!recorded)
+        return { judge: parseModelRef(defaultJudge()) };
+    if (recorded.provider !== "claude-code")
+        return { judge: recorded };
+    return { judge: parseModelRef(defaultJudge()), migratedFrom: recorded };
+}
 //# sourceMappingURL=defaults.js.map

package/dist/execution-trace.js

@@ -5,9 +5,8 @@ import { redactArgs, redactText } from "./redaction.js";
 /**
  * Parse pi's `--mode json` event stream into an `ExecutionTraceV1`.
  *
- * Measured against pi 0.83.0; see `docs/pi-native-capture-design-2026-08-08.md`
- * and the fixtures in `packages/adapters/test/fixtures/pi-json/`. Three findings
- * from that spike are baked in here and each one is a bug if removed:
+ * Pinned by the fixtures in `packages/adapters/test/fixtures/pi-json/`. Three
+ * event-shape constraints are baked in here and each one is a bug if removed:
  *
  * 1. **Line-at-a-time, never buffered.** `message_update` re-sends the entire
  *    accumulated message on every delta, so the stream is quadratic in output

package/dist/grade.d.ts

@@ -54,6 +54,6 @@ export declare function gradeTranscript(adapter: HarnessAdapter, judge: ModelRef
 /**
  * Grade a transcript in a fresh, isolated, throwaway workspace — never the
  * subject's scenario dir — so the judge can't ingest repo context the subject
- * left behind (matters for CLI judges that read cwd, e.g. claude-code).
+ * left behind (matters for provider CLIs that read cwd, including Pi).
  */
 export declare function judgeInWorkspace(adapter: HarnessAdapter, judge: ModelRef, prompt: string, specDir: string): Promise<GradeResult>;

package/dist/grade.js

@@ -139,7 +139,7 @@ export async function gradeTranscript(adapter, judge, prompt, cwd) {
 /**
  * Grade a transcript in a fresh, isolated, throwaway workspace — never the
  * subject's scenario dir — so the judge can't ingest repo context the subject
- * left behind (matters for CLI judges that read cwd, e.g. claude-code).
+ * left behind (matters for provider CLIs that read cwd, including Pi).
  */
 export async function judgeInWorkspace(adapter, judge, prompt, specDir) {
     const ws = createWorkspace("none", { specDir });

package/dist/judge-policy.d.ts

@@ -31,8 +31,8 @@ export interface JudgeAllowOpts {
  *
  * That third path is latent rather than live in the corpus this was built against:
  * checked 2026-08-05, all ~140 committed `results.yaml` in `principal-pi-skills`
- * record `provider: claude-code`, because its owner always passed the subscription
- * judge explicitly. The old default was reachable, not taken. Worth stating
+ * recorded the then-supported subscription judge explicitly. The old default was
+ * reachable, not taken. Worth stating
  * precisely — "your whole archive bills on regrade" would have been a scarier claim
  * than the evidence supports.
  *

package/dist/judge-policy.js

@@ -3,9 +3,8 @@ import { BAKED_DEFAULT_JUDGE } from "./defaults.js";
 /**
  * Judge providers that cannot bill a per-token API.
  *
- * `claude-code` shells out to the `claude` CLI, authenticated with the user's
- * Claude subscription (OAuth). `openai-codex` stays in Pi and authenticates with
- * the user's ChatGPT subscription (OAuth). `ollama`/`lmstudio`/`llamacpp`/`local`
+ * `openai-codex` runs through Pi and authenticates with the user's ChatGPT
+ * subscription (OAuth). `ollama`/`lmstudio`/`llamacpp`/`local`
  * are local runtimes. Everything else — including direct `openai` — is assumed
  * to charge.
  *
@@ -14,7 +13,7 @@ import { BAKED_DEFAULT_JUDGE } from "./defaults.js";
  * being wrong in that direction produces a surprise invoice while being wrong in
  * this direction produces one extra flag.
  */
-const FREE_JUDGE_PROVIDERS = new Set(["claude-code", "openai-codex", "ollama", "lmstudio", "llamacpp", "local"]);
+const FREE_JUDGE_PROVIDERS = new Set(["openai-codex", "ollama", "lmstudio", "llamacpp", "local"]);
 /** Whether judging with this ref can charge a per-token API. */
 export function isMeteredJudge(judge) {
     return !FREE_JUDGE_PROVIDERS.has(judge.provider);
@@ -40,8 +39,8 @@ export function allowMeteredJudge() {
  *
  * That third path is latent rather than live in the corpus this was built against:
  * checked 2026-08-05, all ~140 committed `results.yaml` in `principal-pi-skills`
- * record `provider: claude-code`, because its owner always passed the subscription
- * judge explicitly. The old default was reachable, not taken. Worth stating
+ * recorded the then-supported subscription judge explicitly. The old default was
+ * reachable, not taken. Worth stating
  * precisely — "your whole archive bills on regrade" would have been a scarier claim
  * than the evidence supports.
  *
@@ -49,6 +48,9 @@ export function allowMeteredJudge() {
  * is what a run *is*.
  */
 export function assertJudgeAllowed(judge, opts) {
+    if (judge.provider === "claude-code") {
+        throw new Error("judge provider `claude-code` was removed; choose a provider configured in Pi, such as `openai-codex`");
+    }
     if (!isMeteredJudge(judge))
         return;
     if (opts.allowMetered || allowMeteredJudge())
@@ -56,7 +58,7 @@ export function assertJudgeAllowed(judge, opts) {
     const token = `${judge.provider}:${judge.model}`;
     throw new Error(`refusing to judge with ${token}: \`${judge.provider}\` bills a per-token API key, and it came from ${opts.source}.\n` +
         `  Judging is meant to cost nothing you did not ask for.\n` +
-        `  • judge on your Claude subscription instead:  --judge ${BAKED_DEFAULT_JUDGE}\n` +
+        `  • judge through Pi on your subscription:      --judge ${BAKED_DEFAULT_JUDGE}\n` +
         `  • allow the metered API for this command:     --allow-metered-judge\n` +
         `  • allow it for this repo or shell:            export SKILL_HARNESS_ALLOW_METERED_JUDGE=1`);
 }

package/dist/run.js

@@ -19,10 +19,8 @@ import { aggregateMetrics } from "./metrics.js";
 import { providerFailureFromTranscript } from "./provider-failure.js";
 /**
  * The pi-daddy arm's delivery proof, by convention: `arms.yaml` points
- * `PI_GRANTS_LEDGER` at exactly this filename inside the run dir (see
- * `../principal-pi-skills/tests/arms.yaml` and CODEX-ARMS-RUNBOOK.md §4.3), so
- * a fixed name here is what makes the count readable without parsing the arm's
- * own `env` map back out.
+ * `PI_GRANTS_LEDGER` at exactly this filename inside the run directory. A fixed
+ * name keeps the count readable without parsing the arm's `env` map back out.
  */
 const LEDGER_FILENAME = "pi-daddy.ledger.jsonl";
 /**

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@skill-harness/core",
-    "version": "0.19.0",
+    "version": "0.20.0",
     "description": "skill-harness engine — spec, discover, run, LLM-judge grade, score, results (internal API)",
     "type": "module",
     "license": "MIT",