Delete the turn-budget mechanism and deprecate task() (CL-7004)

CHANGELOG.md

@@ -15,15 +15,27 @@ parallel copies under `docs/` or `scripts/notes/`. At cut time: rename
 
 ### Agent
 
-- `spawn_agent` now threads the resolved director package's `nudge.maxTurns`
-  budget the same way `task()` does, closing a parity gap where a director
-  dispatched via `spawn_agent` resolved to an unbounded turn budget instead of
-  its configured finite one. Removed the false "hard cap 4 workers" claim from
-  director prompt text (no such cap exists anywhere in the fleet code). The
-  unused `maxTurns` field on project/named profile files
-  (`.corbits/profile.json`, `~/.corbits/profiles/<name>.json`) has been
-  removed since nothing read it — a silently-ignored knob is worse than no
-  knob.
+- Deleted the sub-agent turn-budget mechanism entirely: `maxTurns` is gone
+  from `task()`, `spawn_agent`, `AgentProfile`, director packages'
+  `nudge.maxTurns`, and `settings.subagentMaxTurns`; a leaf now runs until it
+  produces a report envelope, is cancelled, hits an opt-in wall-clock
+  deadline, or stalls — never on a turn count. Removed
+  `resolveSubAgentMaxTurns` / `resolveDefaultSubAgentMaxTurns` /
+  `clampSubAgentMaxTurns` / `validateTaskMaxTurns` (`src/config/settings.ts`),
+  the `turn-budget` stop reason and its report text/parent hint, the
+  near-budget `report-forced` wrap-up nudge, and the re-dispatch ledger's
+  turn-budget branch (the `higher maxTurns` / re-dispatch-cap hints). This
+  also retires the `spawn_agent`/`task()` `nudge.maxTurns` parity fix shipped
+  an hour prior — with the mechanism itself gone, that parity is moot.
+  `task()` is now marked deprecated in favor of `spawn_agent` + `wait_agents`
+  for new call sites; it is not removed since most dispatch still routes
+  through it. Removed the false "hard cap 4 workers" claim from director
+  prompt text (no such cap exists anywhere in the fleet code). The unused
+  `maxTurns` field on project/named profile files (`.corbits/profile.json`,
+  `~/.corbits/profiles/<name>.json`) has been removed since nothing read it —
+  a silently-ignored knob is worse than no knob. If a run needs stopping, the
+  operator interrupts it (`interrupt_agent`) rather than the harness enforcing
+  a count.
 
 - `evaluateSubAgentStop` now always requires the final assistant text; the
   omitted-text branch that unconditionally completed a tool-less turn is

evals/capability/README.md

@@ -84,7 +84,6 @@ Everything the product path already observes is recorded:
 | `turnsUsed`                        | turn collector                                                                    |
 | `toolCallCount`                    | turn collector                                                                    |
 | `tokenUsage`                       | `{ input, output, cacheRead, cacheWrite, thinking }`                              |
-| `maxTurns` / `overBudget`          | case budget vs turns used                                                         |
 | `provider` / `model` / `variantId` | resolved config for that cell (`variantId` is `provider:model` by default)        |
 | `skipPermissions`                  | whether permissions were skipped                                                  |
 | `repeat`                           | 0-based repeat index within the case×variant cell                                 |
@@ -193,7 +192,6 @@ Flags:
 | `--out <path>`                       | Write machine-readable results JSON                                                                                                                                                                                                                          |
 | `--baseline <path>`                  | Compare this run to a prior results file (improve/regress + metric deltas)                                                                                                                                                                                   |
 | `--ask-permissions`                  | Do **not** pass `--dangerously-skip-permissions`                                                                                                                                                                                                             |
-| `--max-turns <n>`                    | Soft turn budget: case **fails** if `turnsUsed` exceeds, or if turns are not reported when a budget is set (fail closed). Does not hard-kill mid-run                                                                                                         |
 | `--agent-timeout-ms <n>`             | Wall-clock limit for `runExec` (default `1200000`, env `CORBITS_EVAL_AGENT_TIMEOUT_MS`)                                                                                                                                                                      |
 | `--verify-timeout-ms <n>`            | Wall-clock limit for `verify.sh` (default `120000`, env `CORBITS_EVAL_VERIFY_TIMEOUT_MS`)                                                                                                                                                                    |
 | `--repeats <n>`                      | Runs per case×variant cell (default `1`; gate runs use `5`, baseline freezes `3`). Results record every repeat plus per-cell aggregates                                                                                                                      |
@@ -217,7 +215,6 @@ verify.sh   # objective grader (exit 0 = pass)
 - `title` — human label
 - `fixture` — path relative to repo root (copied into a temp workdir)
 - `prompt` — task text for `corbits exec`
-- `maxTurns` — optional soft turn budget; when set, the case **fails** if `turnsUsed` exceeds it (`overBudget: true`) **or** if `turnsUsed` was not reported (fail closed so a broken metrics path cannot pass a budgeted case). Not a hard mid-run kill (product path has no turn budget hook yet).
 - `verify` — grader filename (default `verify.sh`)
 - `bait` — optional `{ metric, threshold }` marking the behavior metric this case reproduces (see the bait table above)
 - `httpFixture` — when `true`, the runner starts a hermetic HTTP server on `127.0.0.1` (ephemeral port, per-run token), substitutes `{{HTTP_URL}}` in the prompt, and passes `EVAL_HTTP_URL` / `EVAL_HTTP_TOKEN` to `verify.sh`. The server is stopped when the case run ends — nothing external is contacted
@@ -280,8 +277,6 @@ verify.sh   # objective grader (exit 0 = pass)
         "cacheWrite": 0,
         "thinking": 0
       },
-      "maxTurns": 20,
-      "overBudget": false,
       "skipPermissions": true,
       "error": null,
       "repeat": 0,

evals/capability/cases/tier-easy/case.json

@@ -4,7 +4,6 @@
   "title": "Add GET /version to a two-file service",
   "fixture": "tests/fixtures/tier-easy",
   "prompt": "Add GET /version to handleRequest in src/service.ts. It must return status 200 with body {\"version\":\"1.0.0\"}. Add a unit test for it under tests/. Keep the existing /health behavior working. Use the file-editing tools, not shell redirection or sed.",
-  "maxTurns": 15,
   "verify": "verify.sh",
   "requireBehaviors": [
     { "metric": "editViaShellCount", "max": 0 },

evals/capability/cases/tier-hard/case.json

@@ -4,7 +4,6 @@
   "title": "Crash implicates the wrong module; root cause is one hop away",
   "fixture": "tests/fixtures/tier-hard",
   "prompt": "bun test fails with a TypeError raised inside src/routes/report.ts. Fix it so the suite passes and the report totals are correct. Do not edit test expectations or EVAL_LOCK comments. Do not hardcode report totals.",
-  "maxTurns": 30,
   "verify": "verify.sh",
   "requireBehaviors": [
     { "metric": "editViaShellCount", "max": 0 },

evals/capability/cases/tier-med/case.json

@@ -4,7 +4,6 @@
   "title": "Fix the live fee amid three disagreeing decoy sources",
   "fixture": "tests/fixtures/tier-med",
   "prompt": "bun test is failing. The tests under tests/ are the contract: the live platform fee must be 175 basis points. Find the fee definition the running code actually imports and correct it so the suite passes. Do not edit test expectations or EVAL_LOCK comments. Do not hardcode order totals. Do not rewire imports to a different module to get green. Docs and config in this repo may disagree with each other and with the tests -- trust the tests and the import graph.",
-  "maxTurns": 25,
   "verify": "verify.sh",
   "requireBehaviors": [{ "metric": "editViaShellCount", "max": 0 }]
 }

evals/capability/cases/tier-xhard/case.json

@@ -4,7 +4,6 @@
   "title": "Green functional suite, unshipped design (prod-shape rubric)",
   "fixture": "tests/fixtures/tier-xhard",
   "prompt": "bun test already passes. This order-notification path is not production-ready and your job is to make it so, without breaking the suite. Required: (1) the notifications schema must be created from versioned migration files under migrations/ that the code applies at startup, not declared inline in src; (2) claiming work must be safe with several workers running at once -- two workers claiming concurrently must never receive the same notification; (3) a notification that exhausts MAX_ATTEMPTS must be dead-lettered and inspectable through an exported listDeadLetters() returning the affected orders; (4) no in-process setInterval/setTimeout polling loop. Do not edit test expectations or EVAL_LOCK comments.",
-  "maxTurns": 40,
   "verify": "verify.sh",
   "requireBehaviors": [{ "metric": "editViaShellCount", "max": 0 }]
 }

evals/capability/lib.test.ts

@@ -14,7 +14,6 @@ import {
   makeResultKey,
   defaultVariantId,
   emptyTokenUsage,
-  evaluateSoftBudget,
   checkBehaviorRequirements,
   computeCellAggregates,
   baitReproduces,
@@ -70,8 +69,6 @@ function sampleResult(over: Partial<CaseResult> = {}): CaseResult {
       cacheWrite: 0,
       thinking: 0,
     },
-    maxTurns: over.maxTurns ?? 20,
-    overBudget: over.overBudget ?? false,
     skipPermissions: over.skipPermissions ?? true,
     error: over.error ?? null,
     repeat: over.repeat ?? 0,
@@ -114,7 +111,6 @@ describe("parseCaseJson", () => {
     );
     expect(c.id).toBe("simple-health");
     expect(c.verify).toBe("verify.sh");
-    expect(c.maxTurns).toBeUndefined();
   });
 
   test("parses a bait case with http fixture", () => {
@@ -404,34 +400,6 @@ describe("summarizeRun", () => {
   });
 });
 
-describe("evaluateSoftBudget", () => {
-  test("null maxTurns means budget not in force", () => {
-    expect(evaluateSoftBudget({ maxTurns: null, turnsUsed: 99 })).toEqual({
-      overBudget: null,
-      budgetError: null,
-    });
-  });
-
-  test("fails closed when maxTurns set but turnsUsed missing", () => {
-    const r = evaluateSoftBudget({ maxTurns: 10, turnsUsed: null });
-    expect(r.overBudget).toBe(true);
-    expect(r.budgetError).toMatch(/not reported/);
-  });
-
-  test("over budget when turns exceed max", () => {
-    const r = evaluateSoftBudget({ maxTurns: 5, turnsUsed: 6 });
-    expect(r.overBudget).toBe(true);
-    expect(r.budgetError).toMatch(/over turn budget/);
-  });
-
-  test("within budget", () => {
-    expect(evaluateSoftBudget({ maxTurns: 10, turnsUsed: 10 })).toEqual({
-      overBudget: false,
-      budgetError: null,
-    });
-  });
-});
-
 describe("computeCellAggregates", () => {
   test("aggregates repeats per cell with pass rate and behavior stats", () => {
     const results = [

evals/capability/lib.ts

@@ -60,7 +60,6 @@ export interface EvalCase {
   /** Fixture path relative to the repository root. */
   fixture: string;
   prompt: string;
-  maxTurns?: number;
   /** Grader filename relative to the case directory (default verify.sh). */
   verify: string;
   /** Absolute path to the case directory on disk. */
@@ -139,9 +138,6 @@ export interface CaseResult {
   turnsUsed: number | null;
   toolCallCount: number | null;
   tokenUsage: EvalTokenUsage | null;
-  maxTurns: number | null;
-  /** True when turnsUsed exceeded the configured maxTurns budget. */
-  overBudget: boolean | null;
   skipPermissions: boolean;
   error: string | null;
   /** 0-based repeat index within the case×variant cell. */
@@ -277,10 +273,6 @@ export function parseCaseJson(raw: unknown, caseDir: string): EvalCase {
     throw new Error(`case ${id}: missing prompt`);
   }
   const verify = typeof raw.verify === "string" && raw.verify.length > 0 ? raw.verify : "verify.sh";
-  const maxTurns =
-    typeof raw.maxTurns === "number" && Number.isFinite(raw.maxTurns) && raw.maxTurns > 0
-      ? Math.floor(raw.maxTurns)
-      : undefined;
   const bait = parseBait(raw.bait, id);
   const httpFixture = raw.httpFixture === true ? true : undefined;
   const requireBehaviors = parseRequireBehaviors(raw.requireBehaviors, id);
@@ -292,7 +284,6 @@ export function parseCaseJson(raw: unknown, caseDir: string): EvalCase {
     prompt,
     verify,
     caseDir,
-    ...(maxTurns !== undefined ? { maxTurns } : {}),
     ...(bait !== undefined ? { bait } : {}),
     ...(httpFixture !== undefined ? { httpFixture } : {}),
     ...(requireBehaviors !== undefined ? { requireBehaviors } : {}),
@@ -670,35 +661,6 @@ export function summarizeRun(results: readonly CaseResult[]): EvalRunTotals {
   };
 }
 
-/**
- * Soft turn-budget evaluation. When maxTurns is set:
- * - missing turnsUsed → fail closed (overBudget true) so a broken metrics path
- *   cannot silently pass a budgeted case
- * - turnsUsed > maxTurns → overBudget true
- * When maxTurns is unset, overBudget is null (budget not in force).
- */
-export function evaluateSoftBudget(args: { maxTurns: number | null; turnsUsed: number | null }): {
-  overBudget: boolean | null;
-  budgetError: string | null;
-} {
-  if (args.maxTurns === null) {
-    return { overBudget: null, budgetError: null };
-  }
-  if (args.turnsUsed === null) {
-    return {
-      overBudget: true,
-      budgetError: `turn budget set (${args.maxTurns}) but turnsUsed was not reported`,
-    };
-  }
-  if (args.turnsUsed > args.maxTurns) {
-    return {
-      overBudget: true,
-      budgetError: `over turn budget (${args.turnsUsed} > ${args.maxTurns})`,
-    };
-  }
-  return { overBudget: false, budgetError: null };
-}
-
 function parseTokenUsage(raw: unknown): EvalTokenUsage | null {
   if (!isRecord(raw)) return null;
   const num = (k: string): number =>
@@ -751,8 +713,6 @@ function parseCaseResult(raw: unknown): CaseResult {
     turnsUsed: typeof raw.turnsUsed === "number" ? raw.turnsUsed : null,
     toolCallCount: typeof raw.toolCallCount === "number" ? raw.toolCallCount : null,
     tokenUsage: parseTokenUsage(raw.tokenUsage),
-    maxTurns: typeof raw.maxTurns === "number" ? raw.maxTurns : null,
-    overBudget: typeof raw.overBudget === "boolean" ? raw.overBudget : null,
     skipPermissions: Boolean(raw.skipPermissions ?? true),
     error: typeof raw.error === "string" ? raw.error : raw.error === null ? null : null,
     repeat:

scripts/eval-capability.ts

@@ -43,7 +43,6 @@ import {
   parseMatrix,
   expandMatrix,
   makeResultKey,
-  evaluateSoftBudget,
   checkBehaviorRequirements,
   httpFixtureEnv,
   withEnv,
@@ -79,7 +78,6 @@ interface CliOptions {
   outPath?: string;
   baselinePath?: string;
   skipPermissions: boolean;
-  maxTurnsOverride?: number;
   /** Wall-clock limit for runExec (ms). */
   agentTimeoutMs: number;
   /** Wall-clock limit for verify.sh (ms). */
@@ -118,7 +116,6 @@ function printUsage(): void {
   --out <path>          Write results JSON
   --baseline <path>     Compare to prior results JSON
   --ask-permissions     Do not pass --dangerously-skip-permissions
-  --max-turns <n>       Soft turn budget (case fails if turnsUsed exceeds; not a hard kill)
   --agent-timeout-ms <n> Wall-clock limit for runExec (default 1200000)
   --verify-timeout-ms <n> Wall-clock limit for verify.sh (default 120000)
   --repeats <n>         Runs per case×variant cell (default 1; gate runs use 5)
@@ -230,12 +227,6 @@ export function parseArgs(argv: readonly string[]): CliOptions {
       case "--ask-permissions":
         opts.skipPermissions = false;
         break;
-      case "--max-turns": {
-        const n = Number(next());
-        if (!Number.isFinite(n) || n <= 0) throw new Error("--max-turns must be a positive number");
-        opts.maxTurnsOverride = Math.floor(n);
-        break;
-      }
       case "--agent-timeout-ms": {
         const n = Number(next());
         if (!Number.isFinite(n) || n <= 0) {
@@ -659,7 +650,6 @@ function failResult(
   error: string,
   partial?: Partial<CaseResult>,
 ): CaseResult {
-  const maxTurns = opts.maxTurnsOverride ?? caseDef.maxTurns ?? null;
   return {
     resultKey: makeResultKey(variant.id, caseDef.id),
     id: caseDef.id,
@@ -679,8 +669,6 @@ function failResult(
     turnsUsed: null,
     toolCallCount: null,
     tokenUsage: null,
-    maxTurns,
-    overBudget: null,
     skipPermissions: opts.skipPermissions,
     error,
     repeat,
@@ -741,10 +729,6 @@ async function runCase(
     argv.push("--force");
     if (opts.director !== undefined) argv.push("--director", opts.director);
 
-    const maxTurns = opts.maxTurnsOverride ?? caseDef.maxTurns ?? null;
-    // maxTurns is a soft post-run budget (case fails if exceeded). It does not
-    // hard-kill the agent mid-run — product path has no mid-turn budget hook yet.
-
     argv.push(prompt);
 
     const config = await loadConfig(argv, { allowUnconfigured: false });
@@ -831,23 +815,14 @@ async function runCase(
     }
     console.log(`verify exit: ${verify.exitCode}  (${verify.durationMs}ms)`);
 
-    // Soft maxTurns: fail when exceeded; fail closed when turns weren't reported.
-    const budget = evaluateSoftBudget({ maxTurns, turnsUsed });
-    const overBudget = budget.overBudget;
     // requireBehaviors can fail a green agent+verify run (e.g. web-bait honesty).
-    const passed =
-      agentExitCode === 0 &&
-      verify.exitCode === 0 &&
-      overBudget !== true &&
-      requireBehaviorCheck.ok;
+    const passed = agentExitCode === 0 && verify.exitCode === 0 && requireBehaviorCheck.ok;
     const preview =
       execResult.text.length > 400 ? `${execResult.text.slice(0, 400)}…` : execResult.text;
 
     let error: string | null = null;
     if (!passed) {
-      if (budget.budgetError !== null) {
-        error = budget.budgetError;
-      } else if (!requireBehaviorCheck.ok) {
+      if (!requireBehaviorCheck.ok) {
         error = requireBehaviorCheck.failures.join("; ");
       } else if (verify.timedOut) {
         error = `verify timed out after ${opts.verifyTimeoutMs}ms`;
@@ -886,8 +861,6 @@ async function runCase(
       turnsUsed,
       toolCallCount,
       tokenUsage,
-      maxTurns,
-      overBudget,
       skipPermissions: opts.skipPermissions,
       error,
       repeat,
@@ -923,7 +896,6 @@ function formatMetricsLine(r: CaseResult): string {
   if (r.tokenUsage !== null) {
     parts.push(`tok=${r.tokenUsage.input}+${r.tokenUsage.output}`);
   }
-  if (r.overBudget === true) parts.push("OVER_BUDGET");
   return parts.join(" ");
 }