jeo-code 0.6.28 → 0.6.30

package/CHANGELOG.md

@@ -6,6 +6,29 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 The README mirrors the latest 5 entries — regenerate with `bun run changelog:sync`.
 
+## [0.6.30] - 2026-06-19
+_gjc-style intermediate-judgment guard classification extracted from the engine loop, plus a re-verification that `jeo --tmux` does not leak bun memory or slow down._
+
+### Changed
+- **Loop intermediate-judgment guards extracted into a classified module (`src/agent/loop-guards.ts`).** The mid-run "continue / self-correct / stop" decisions that were inlined across `engine.ts`'s `while` loop as scattered booleans and message strings are now a named `GuardState` discriminated-union taxonomy — jeo's descendant of gjc's `ultragoal-guard` `UltragoalGuardState` pattern. A single frozen `GUARD_LIMITS` object is the source of truth for every threshold (`MAX_REPEAT`, `MAX_FAILURES`, `MAX_REFUSAL_RETRIES`, `MAX_INVALID_CALLS`, `MAX_PARSE_BOUNCES`, `CYCLE_WINDOW`), and pure classifiers (`isVerificationSignal`, `repeatHint`, `nearestToolName`, `classifyDoneGate`) are now independently testable. `engine.ts` still owns all control flow (history mutation, `step++`, `continue`, `return finish(...)`) — only the JUDGMENT moved, so behavior is unchanged (net −19 lines in `engine.ts`). Removed the now-unused `src/agent/tool-registry.ts`.
+
+### Verified
+- **`jeo --tmux` has no bun memory leak and does not slow down.** An in-process probe streaming 5,000,000 SGR mouse-report escapes through `queuePromptInputChunk` (10 × 500k, `Bun.gc(true)` between batches) holds RSS flat (133.9 → 135.2 MB, slope ≈0.13 MB/round) with zero prompt-queue accumulation; a real `jeo --tmux` session flooded with 60k live mouse reports via `tmux send-keys` plateaus in RSS (129,456 → 129,472 KB). `jeo --tmux -p` end-to-end creates the profiled session, runs the turn, and tears down cleanly.
+- **Full suite green:** `bun run typecheck` clean and `bun test` 1687 pass / 0 fail across 210 files (includes the new `test/loop-guards.test.ts`, 9 tests, and the signature-only Anthropic replay test).
+
+## [0.6.29] - 2026-06-19
+_Signature-only thinking-block replay (Anthropic opus-4-7/4-8), plus a tmux mouse-flood memory guard confirming `jeo --tmux` does not leak._
+
+### Fixed
+- **Anthropic thinking-block replay now covers signature-only artifacts.** Newer Opus models (opus-4-7/opus-4-8) think internally — tokens billed, a valid `signature` present — but return empty thinking text. The cross-turn replay required both `signature` AND `text`, so those models' reasoning was dropped between steps. Replay now sends a signed `thinking` block whenever a `signature` (or `redacted`) is present (text defaults to `""`), restoring multi-step reasoning continuity for signature-only models. API-key requests also send the `interleaved-thinking` + `prompt-caching-scope` betas so thinking+tools and scoped caching work outside OAuth.
+
+### Added
+- **`claude-opus-4-7` catalogued** (FULL thinking, 200k ctx) and a dynamic context-window fallback for uncatalogued ids (claude 200k / gpt-5 400k / gemini-3 1M).
+- **tmux mouse-report-flood memory guard** (`test/mouse-report-filter.test.ts`): 100k SGR mouse-move reports through `queuePromptInputChunk` leave the prompt queue at zero accumulation — the regression guard for the "`jeo --tmux` slows down over time" concern.
+
+### Verified
+- **`jeo --tmux` has no bun memory leak.** The in-process lifecycle probe (`scripts/mem-probe.ts`, 3000 turns) reports a per-turn heap slope of ≈0 (returns to baseline, exit-listeners flat); a real `jeo --tmux` process plateaus in RSS under sustained mouse/resize/keystroke churn instead of climbing; and mouse reports are filtered (not buffered) with `activityLog` bounded to a 200-entry per-turn ring.
+
 ## [0.6.28] - 2026-06-19
 _Signed thinking-block replay: native reasoning is now sent BACK to providers across steps/turns, restoring multi-step reasoning continuity (gajae parity)._
 

package/README.ja.md

@@ -200,11 +200,11 @@ CI は `.github/workflows/npm-publish.yml` で公開します — GitHub リリ
 ## 変更履歴 (Changelog)
 
 <!-- CHANGELOG:START (auto-generated from CHANGELOG.md — run `bun run changelog:sync`) -->
+- **[0.6.30]** (2026-06-19) — gjc-style intermediate-judgment guard classification extracted from the engine loop, plus a re-verification that `jeo --tmux` does not leak bun memory or slow down.
+- **[0.6.29]** (2026-06-19) — Signature-only thinking-block replay (Anthropic opus-4-7/4-8), plus a tmux mouse-flood memory guard confirming `jeo --tmux` does not leak.
 - **[0.6.28]** (2026-06-19) — Signed thinking-block replay: native reasoning is now sent BACK to providers across steps/turns, restoring multi-step reasoning continuity (gajae parity).
 - **[0.6.27]** (2026-06-19) — Ponytail pass on the reasoning-tier mapper, plus a real-tmux verification of `jeo --tmux`.
 - **[0.6.26]** (2026-06-19) — The forge emblem is redrawn again as the mascot crayfish, foregrounding its signature pincer claws (집게).
-- **[0.6.25]** (2026-06-19) — Reasoning works at every thinking level (gajae parity), and the forge emblem is redrawn as the neon-lens coding wizard.
-- **[0.6.24]** (2026-06-19) — `/provider` opens an interactive onboarding selector (OAuth vs API-compatible), and OpenAI-compatible backends gain per-vendor native-reasoning formats.
 
 See [CHANGELOG.md](CHANGELOG.md) for the full history.
 <!-- CHANGELOG:END -->

package/README.ko.md

@@ -200,11 +200,11 @@ CI는 `.github/workflows/npm-publish.yml`로 배포합니다 — GitHub 릴리
 ## 변경 이력 (Changelog)
 
 <!-- CHANGELOG:START (auto-generated from CHANGELOG.md — run `bun run changelog:sync`) -->
+- **[0.6.30]** (2026-06-19) — gjc-style intermediate-judgment guard classification extracted from the engine loop, plus a re-verification that `jeo --tmux` does not leak bun memory or slow down.
+- **[0.6.29]** (2026-06-19) — Signature-only thinking-block replay (Anthropic opus-4-7/4-8), plus a tmux mouse-flood memory guard confirming `jeo --tmux` does not leak.
 - **[0.6.28]** (2026-06-19) — Signed thinking-block replay: native reasoning is now sent BACK to providers across steps/turns, restoring multi-step reasoning continuity (gajae parity).
 - **[0.6.27]** (2026-06-19) — Ponytail pass on the reasoning-tier mapper, plus a real-tmux verification of `jeo --tmux`.
 - **[0.6.26]** (2026-06-19) — The forge emblem is redrawn again as the mascot crayfish, foregrounding its signature pincer claws (집게).
-- **[0.6.25]** (2026-06-19) — Reasoning works at every thinking level (gajae parity), and the forge emblem is redrawn as the neon-lens coding wizard.
-- **[0.6.24]** (2026-06-19) — `/provider` opens an interactive onboarding selector (OAuth vs API-compatible), and OpenAI-compatible backends gain per-vendor native-reasoning formats.
 
 See [CHANGELOG.md](CHANGELOG.md) for the full history.
 <!-- CHANGELOG:END -->

package/README.md

@@ -200,11 +200,11 @@ Required npm token permissions (repository secret `NPM_TOKEN`):
 ## Changelog
 
 <!-- CHANGELOG:START (auto-generated from CHANGELOG.md — run `bun run changelog:sync`) -->
+- **[0.6.30]** (2026-06-19) — gjc-style intermediate-judgment guard classification extracted from the engine loop, plus a re-verification that `jeo --tmux` does not leak bun memory or slow down.
+- **[0.6.29]** (2026-06-19) — Signature-only thinking-block replay (Anthropic opus-4-7/4-8), plus a tmux mouse-flood memory guard confirming `jeo --tmux` does not leak.
 - **[0.6.28]** (2026-06-19) — Signed thinking-block replay: native reasoning is now sent BACK to providers across steps/turns, restoring multi-step reasoning continuity (gajae parity).
 - **[0.6.27]** (2026-06-19) — Ponytail pass on the reasoning-tier mapper, plus a real-tmux verification of `jeo --tmux`.
 - **[0.6.26]** (2026-06-19) — The forge emblem is redrawn again as the mascot crayfish, foregrounding its signature pincer claws (집게).
-- **[0.6.25]** (2026-06-19) — Reasoning works at every thinking level (gajae parity), and the forge emblem is redrawn as the neon-lens coding wizard.
-- **[0.6.24]** (2026-06-19) — `/provider` opens an interactive onboarding selector (OAuth vs API-compatible), and OpenAI-compatible backends gain per-vendor native-reasoning formats.
 
 See [CHANGELOG.md](CHANGELOG.md) for the full history.
 <!-- CHANGELOG:END -->

package/README.zh.md

@@ -200,11 +200,11 @@ CI 通过 `.github/workflows/npm-publish.yml` 发布 — GitHub 发布 release
 ## 更新日志 (Changelog)
 
 <!-- CHANGELOG:START (auto-generated from CHANGELOG.md — run `bun run changelog:sync`) -->
+- **[0.6.30]** (2026-06-19) — gjc-style intermediate-judgment guard classification extracted from the engine loop, plus a re-verification that `jeo --tmux` does not leak bun memory or slow down.
+- **[0.6.29]** (2026-06-19) — Signature-only thinking-block replay (Anthropic opus-4-7/4-8), plus a tmux mouse-flood memory guard confirming `jeo --tmux` does not leak.
 - **[0.6.28]** (2026-06-19) — Signed thinking-block replay: native reasoning is now sent BACK to providers across steps/turns, restoring multi-step reasoning continuity (gajae parity).
 - **[0.6.27]** (2026-06-19) — Ponytail pass on the reasoning-tier mapper, plus a real-tmux verification of `jeo --tmux`.
 - **[0.6.26]** (2026-06-19) — The forge emblem is redrawn again as the mascot crayfish, foregrounding its signature pincer claws (집게).
-- **[0.6.25]** (2026-06-19) — Reasoning works at every thinking level (gajae parity), and the forge emblem is redrawn as the neon-lens coding wizard.
-- **[0.6.24]** (2026-06-19) — `/provider` opens an interactive onboarding selector (OAuth vs API-compatible), and OpenAI-compatible backends gain per-vendor native-reasoning formats.
 
 See [CHANGELOG.md](CHANGELOG.md) for the full history.
 <!-- CHANGELOG:END -->

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "jeo-code",
-  "version": "0.6.28",
+  "version": "0.6.30",
   "description": "Clean, highly optimized AI coding agent using spec-first loop",
   "type": "module",
   "main": "src/cli.ts",

package/src/agent/AGENTS.md

@@ -17,6 +17,7 @@ The core runtime loop, tool registry, session management, and state persistence
 | `hooks.ts` | Brief description of purpose |
 | `json.ts` | Brief description of purpose |
 | `loop.ts` | The primary execution loop orchestrating model calls and tool execution |
+| `loop-guards.ts` | Intermediate-judgment classification (gjc ultragoal-guard parity): named `GuardState` taxonomy, `GUARD_LIMITS` thresholds, and pure classifiers (`isVerificationSignal`, `repeatHint`, `classifyDoneGate`) consumed by `engine.ts` |
 | `memory.ts` | OKF concept-bundle memory: session distill, query-aware budget injection, legacy MEMORY.md migration (`migrateLegacyMemory`) + `JEO_MEMORY_LEGACY` rollback toggle |
 | `memory-okf.ts` | OKF v0.1 format layer: frontmatter parse/serialize, concept IDs, conformance validation |
 | `memory-graph.ts` | Concept cross-link graph: build/expand (1-hop search), broken-link-tolerant lint, optional graphify detection |
@@ -35,7 +36,6 @@ The core runtime loop, tool registry, session management, and state persistence
 | `todo-tool.ts` | Brief description of purpose |
 | `tokenizer.ts` | Brief description of purpose |
 | `tool-output.ts` | Brief description of purpose |
-| `tool-registry.ts` | Brief description of purpose |
 | `tools.ts` | Built-in tool definitions (bash, read, write, edit, etc.) |
 | `web-search.ts` | Brief description of purpose |
 

package/src/agent/engine.ts

@@ -22,6 +22,7 @@ export { TOOL_OUTPUT_MAX, READ_OUTPUT_MAX, TOOL_SPILL_THRESHOLD, MAX_TOOL_ARTIFA
 import { StepBudget, dynamicStepBudgetConfig, resolveStepBudgetConfig, hashSignature, type StepBudgetConfig } from "./step-budget";
 import { historyTokens, trimToolResultsInPlace } from "./compaction";
 import { jeoEnv } from "../util/env";
+import { GUARD_LIMITS, isVerificationSignal, repeatHint, classifyDoneGate } from "./loop-guards";
 
 
 async function invokeCallLlm(history: Message[], options: {
@@ -378,29 +379,15 @@ export async function runAgentLoop(history: Message[], opts: AgentLoopOptions):
     } catch { /* best-effort; fall through to the plain stop message */ }
     return finish({ done: false, steps: step, doneReason: `Stopped: ${stopReason}` });
   };
-  // Result-aware repeat nudge (A): tell the model WHY repeating won't help and what to
-  // try instead, tailored to the repeated tool and its last actual result.
-  const repeatHint = (tool: string, prev?: { success: boolean; output: string }): string => {
-    const out = prev?.output ?? "";
-    const empty = !prev || !prev.success || out.trim() === "" || /no match|0 match|no result|not found|no file/i.test(out);
-    if (tool === "search" || tool === "find" || tool === "ls") {
-      return empty
-        ? `That '${tool}' returned nothing useful and will again — BROADEN it (a looser pattern, a parent directory, or a different tool such as ${tool === "search" ? "find" : "search"}), or call done if this lookup isn't needed.`
-        : `That '${tool}' already returned results — open one of the hits with read, or move on; re-running it changes nothing.`;
-    }
-    if (tool === "read") return `You already read that and its content is unchanged — use what you read, or read a DIFFERENT file.`;
-    if (tool === "bash") return `That command already ran with the same output — change the command, or call done.`;
-    return `That call's result is unchanged — take a different action, or call done.`;
-  };
   // No-progress guard: weak/local models often repeat the same tool call without
   // ever emitting `done`. Two escalating corrections (B), then a consolidated stop.
-  const MAX_REPEAT = 4;
+  const MAX_REPEAT = GUARD_LIMITS.MAX_REPEAT;
   // Last executed step's per-call results — fed to repeatHint so a corrective bounce
   // can cite the repeated call's ACTUAL last outcome (A).
   let lastResults: { success: boolean; output: string; executed: boolean }[] = [];
   // Consecutive-failure guard: a model that keeps emitting *different* but failing
   // calls (bad edits, failing commands) would otherwise burn the whole step budget.
-  const MAX_FAILURES = 5;
+  const MAX_FAILURES = GUARD_LIMITS.MAX_FAILURES;
   let consecutiveFailures = 0;
   // done-verification guard (plan/gjc-inheritance.md B4, gjc ultragoal-guard 경량 계승):
   // a turn that MUTATED files but shows no verification signal gets ONE pushback on
@@ -424,16 +411,15 @@ export async function runAgentLoop(history: Message[], opts: AgentLoopOptions):
   // as-is, then once more with an explicit re-grounding note; only a third
   // refusal in the turn surfaces the (friendly) error. Bounded per turn so a
   // genuinely refused request can never burn billed calls in a loop.
-  const MAX_REFUSAL_RETRIES = 3;
+  const MAX_REFUSAL_RETRIES = GUARD_LIMITS.MAX_REFUSAL_RETRIES;
   let refusalRetries = 0;
-  const VERIFY_SIGNAL_RE = /\b(test|tests|tsc|typecheck|lint|build|check|spec|pytest|vitest|jest)\b/i;
   let lastSig = "";
   let repeatCount = 0;
   // Cycle guard (the A↔B ping-pong the exact-repeat guard cannot see): the recent
   // executed step signatures, as fixed-size digests. When a full window cycles
   // through ≤2 distinct calls, bounce ONCE with an explicit correction; a spin that
   // persists through the correction stops the turn.
-  const CYCLE_WINDOW = 6;
+  const CYCLE_WINDOW = GUARD_LIMITS.CYCLE_WINDOW;
   const recentStepSigs: string[] = [];
   let cycleBounceUsed = false;
   // Invalid-tool-call guard: a model that returns JSON without a usable `tool`
@@ -441,10 +427,10 @@ export async function runAgentLoop(history: Message[], opts: AgentLoopOptions):
   let invalidToolCalls = 0;
   // A JSON reply with no usable `tool` field can't drive the loop — stop sooner than the
   // repeat-spin guard (no escalating correction helps a model that isn't producing a call).
-  const MAX_INVALID_CALLS = 3;
+  const MAX_INVALID_CALLS = GUARD_LIMITS.MAX_INVALID_CALLS;
   // Prose-bounce guard: after this many invalid-JSON corrections, salvage the
   // model's text as the final answer instead of burning the whole step budget.
-  const MAX_PARSE_BOUNCES = 2;
+  const MAX_PARSE_BOUNCES = GUARD_LIMITS.MAX_PARSE_BOUNCES;
   let parseFailures = 0;
   while (true) {
     if (turnBudgetMs > 0 && Date.now() - turnStartedAt > turnBudgetMs) {
@@ -703,19 +689,14 @@ export async function runAgentLoop(history: Message[], opts: AgentLoopOptions):
     ev.onAssistant?.(responseText, toolCalls[0]);
 
     if (toolCalls.length === 1 && toolCalls[0].tool === "done") {
-      if (sawMutation && (!sawVerification || pendingHookFailure !== null) && !donePushbackUsed) {
+      // done-verification gate — jeo's descendant of gjc's ultragoal-guard completion
+      // state machine (plan/gjc-inheritance.md B4). The classifier owns the JUDGMENT
+      // (which named state, which message); the loop owns the once-pushback latch.
+      const doneGate = classifyDoneGate({ sawMutation, sawVerification, pendingHookFailure });
+      if (doneGate.block && !donePushbackUsed) {
         donePushbackUsed = true; // second done always passes — escape hatch
         pushAssistantTurn(history, responseText, reasonBuf, artifactBuf);
-        history.push({
-          role: "user",
-          content: pendingHookFailure !== null
-            ? `Your latest mutation left the post-turn hook "${pendingHookFailure}" FAILING (non-zero exit) — its diagnostics were shown in the tool result above. ` +
-              "Fix the reported problems (the hook re-runs on your next mutation), then call done. " +
-              "If the hook failure is a false positive, call done again and say why in the reason."
-            : "You modified files this turn but ran NO verification (no test/build/typecheck command succeeded). " +
-              "Run the narrowest command that proves your change works, then call done. " +
-              "If verification is genuinely not applicable (docs/config-only change), call done again and say why in the reason.",
-        });
+        history.push({ role: "user", content: doneGate.message });
         step++;
         continue;
       }
@@ -1039,7 +1020,7 @@ export async function runAgentLoop(history: Message[], opts: AgentLoopOptions):
       if (t === "write" || t === "edit") sawMutation = true;
       else if (t === "bash") {
         const cmd = String(toolCalls[i].arguments?.command ?? "");
-        if (VERIFY_SIGNAL_RE.test(cmd) || VERIFY_SIGNAL_RE.test(results[i].output.slice(0, 2000))) sawVerification = true;
+        if (isVerificationSignal(cmd, results[i].output)) sawVerification = true;
       }
     }
     // F6 (round 4 architect, Low): judge the step by its NON-TRIVIAL calls — a

package/src/agent/loop-guards.ts

@@ -0,0 +1,135 @@
+/**
+ * Intermediate-judgment guards for the agent loop — the mid-run "should this turn
+ * continue, correct itself, or stop" decisions that run between model calls.
+ *
+ * gjc keeps this concern in its own layer: `gjc-runtime/ultragoal-guard.ts` computes a
+ * named `UltragoalGuardState` discriminated union PURELY, and the runtime merely acts on
+ * the verdict. jeo previously inlined the same logic inside `engine.ts`'s `while` loop as
+ * scattered booleans and message strings. This module gives jeo the same classification:
+ * a named `GuardState` taxonomy plus pure, independently-testable classifier functions.
+ * `engine.ts` still owns the control flow (history mutation, `step++`, `continue`,
+ * `return finish(...)`) — only the JUDGMENT moves here, so behavior is unchanged.
+ */
+
+/**
+ * Named taxonomy of the loop's intermediate judgments — jeo's descendant of gjc's
+ * `UltragoalGuardState`. Each member names one decision the loop can reach mid-turn.
+ */
+export type GuardState =
+  | "ok" // proceed: emit / execute the tool call as-is
+  | "repeat_correct" // exact-repeat detected → ONE corrective bounce (skip execution)
+  | "repeat_stop" // exact-repeat survived the correction → consolidate-stop
+  | "cycle_correct" // A↔B alternation detected → ONE corrective bounce
+  | "cycle_stop" // cycle survived the correction → consolidate-stop
+  | "consecutive_failure_stop" // MAX_FAILURES different-but-failing steps → stop
+  | "invalid_tool_stop" // MAX_INVALID_CALLS replies with no usable tool field → stop
+  | "parse_salvage" // repeated non-JSON prose → salvage the text as the final answer
+  | "context_overflow_retry" // provider reported context overflow → ONE trim + retry
+  | "refusal_retry" // transient safety refusal → bounded resend ladder
+  | "done_unverified" // mutated files, no verification signal → pushback on done
+  | "done_hook_failing" // post-turn hook still failing → pushback on done
+  | "done_ok"; // done accepted — the turn is finished
+
+/**
+ * Bounded thresholds for every loop guard — the single, named source of truth.
+ * Kept in one frozen object so the limits are discoverable and testable instead of
+ * sprinkled as bare literals through the loop body.
+ */
+export const GUARD_LIMITS = Object.freeze({
+  /** Identical step repeats tolerated before a consolidated stop (with corrections en route). */
+  MAX_REPEAT: 4,
+  /** Consecutive different-but-failing steps before the turn stops. */
+  MAX_FAILURES: 5,
+  /** Safety-refusal resends per turn before surfacing the friendly error. */
+  MAX_REFUSAL_RETRIES: 3,
+  /** Replies with no usable `tool`/`tools` field before the turn stops. */
+  MAX_INVALID_CALLS: 3,
+  /** Consecutive non-JSON parse failures before the prose is salvaged as the answer. */
+  MAX_PARSE_BOUNCES: 2,
+  /** Recent-signature window scanned for an A↔B (≤2 distinct calls) cycle. */
+  CYCLE_WINDOW: 6,
+});
+
+/**
+ * Commands (or their output) that count as a verification signal: a test, build,
+ * typecheck, or lint invocation. The done-verification guard treats a turn that mutated
+ * files without any such signal as "unverified".
+ */
+export const VERIFY_SIGNAL_RE = /\b(test|tests|tsc|typecheck|lint|build|check|spec|pytest|vitest|jest)\b/i;
+
+/**
+ * True when a bash command (or the head of its output) proves the work was verified.
+ * Output is examined only up to the first 2000 chars — enough to catch a tool runner's
+ * banner without rescanning a megabyte of logs.
+ */
+export function isVerificationSignal(cmd: string, output = ""): boolean {
+  return VERIFY_SIGNAL_RE.test(cmd) || VERIFY_SIGNAL_RE.test(output.slice(0, 2000));
+}
+
+/**
+ * Result-aware repeat nudge: tells the model WHY repeating the call won't help and what
+ * to try instead, tailored to the repeated tool and its last actual result.
+ */
+export function repeatHint(tool: string, prev?: { success: boolean; output: string }): string {
+  const out = prev?.output ?? "";
+  const empty = !prev || !prev.success || out.trim() === "" || /no match|0 match|no result|not found|no file/i.test(out);
+  if (tool === "search" || tool === "find" || tool === "ls") {
+    return empty
+      ? `That '${tool}' returned nothing useful and will again — BROADEN it (a looser pattern, a parent directory, or a different tool such as ${tool === "search" ? "find" : "search"}), or call done if this lookup isn't needed.`
+      : `That '${tool}' already returned results — open one of the hits with read, or move on; re-running it changes nothing.`;
+  }
+  if (tool === "read") return `You already read that and its content is unchanged — use what you read, or read a DIFFERENT file.`;
+  if (tool === "bash") return `That command already ran with the same output — change the command, or call done.`;
+  return `That call's result is unchanged — take a different action, or call done.`;
+}
+
+/** Inputs for the done-verification gate (jeo's descendant of gjc's ultragoal-guard). */
+export interface DoneGateInput {
+  /** A write/edit succeeded this turn. */
+  sawMutation: boolean;
+  /** A test/build/typecheck/lint command succeeded this turn. */
+  sawVerification: boolean;
+  /** The run-command of the most recent still-failing post-turn hook, or null. */
+  pendingHookFailure: string | null;
+}
+
+/** Verdict from {@link classifyDoneGate}: whether to bounce `done`, and the message. */
+export interface DoneGateVerdict {
+  state: Extract<GuardState, "done_ok" | "done_unverified" | "done_hook_failing">;
+  /** When true, `done` should be bounced ONCE with `message` (the caller owns the once-gate). */
+  block: boolean;
+  /** Corrective message to push back on `done`; empty when `state === "done_ok"`. */
+  message: string;
+}
+
+/**
+ * Classify whether a `done` should be accepted or bounced — the direct descendant of
+ * gjc's `ultragoal-guard` completion gate (plan/gjc-inheritance.md B4).
+ *
+ * A turn that MUTATED files but has either NO verification signal or a still-failing
+ * post-turn hook is blocked ONCE. The caller owns the single-pushback latch; a second
+ * `done` always passes (the escape hatch for genuinely-unverifiable docs/config changes).
+ */
+export function classifyDoneGate(input: DoneGateInput): DoneGateVerdict {
+  const hookFailing = input.pendingHookFailure !== null;
+  const block = input.sawMutation && (!input.sawVerification || hookFailing);
+  if (!block) return { state: "done_ok", block: false, message: "" };
+  if (hookFailing) {
+    return {
+      state: "done_hook_failing",
+      block: true,
+      message:
+        `Your latest mutation left the post-turn hook "${input.pendingHookFailure}" FAILING (non-zero exit) — its diagnostics were shown in the tool result above. ` +
+        "Fix the reported problems (the hook re-runs on your next mutation), then call done. " +
+        "If the hook failure is a false positive, call done again and say why in the reason.",
+    };
+  }
+  return {
+    state: "done_unverified",
+    block: true,
+    message:
+      "You modified files this turn but ran NO verification (no test/build/typecheck command succeeded). " +
+      "Run the narrowest command that proves your change works, then call done. " +
+      "If verification is genuinely not applicable (docs/config-only change), call done again and say why in the reason.",
+  };
+}

package/src/ai/model-catalog.ts

@@ -37,6 +37,8 @@ const STD: ThinkLevel[] = ["minimal", "low", "medium", "high"];
 export const ANTIGRAVITY_MODELS = [
   "claude-opus-4-5-thinking",
   "claude-opus-4-6-thinking",
+  "claude-opus-4-7",
+  "claude-opus-4-7-thinking",
   "claude-opus-4-8",
   "claude-opus-4-8-thinking",
   "claude-sonnet-4-5",
@@ -52,6 +54,7 @@ export const ANTIGRAVITY_MODELS = [
   "gemini-3.1-pro-high",
   "gemini-3.1-pro-low",
   "gpt-oss-120b-medium",
+  "gpt-5.5",
 ] as const;
 
 /** A curated set of common public models with their documented capabilities. */
@@ -62,9 +65,13 @@ export const MODEL_CATALOG: readonly CatalogModel[] = [
   { canonical: "claude-sonnet-4-5", provider: "anthropic", providerModel: "claude-sonnet-4-5-20250929", contextTokens: 200_000, maxOutputTokens: 64_000, thinking: FULL, images: true },
   { canonical: "claude-opus-4-1", provider: "anthropic", providerModel: "claude-opus-4-1-20250805", contextTokens: 200_000, maxOutputTokens: 32_000, thinking: FULL, images: true },
   { canonical: "claude-opus-4-5", provider: "anthropic", providerModel: "claude-opus-4-5-20251101", contextTokens: 200_000, maxOutputTokens: 64_000, thinking: FULL, images: true },
-  // NOTE: confirm exact dated provider ids when these ship publicly; the family
-  // heuristic in `catalogMetadata` keeps reasoning working even before that.
+  // NOTE: opus-4-7 accepts extended thinking but currently returns 0 thinking tokens
+  // (model-internal, no visible thought). opus-4-8 thinks internally (tokens billed,
+  // signature present) but returns empty thinking text. Both are FULL-capable in the
+  // catalog so the budget is always sent — the nativizable path handles signature-only
+  // artifacts for cross-turn continuity.
   { canonical: "claude-opus-4-6", provider: "anthropic", providerModel: "claude-opus-4-6", contextTokens: 200_000, maxOutputTokens: 64_000, thinking: FULL, images: true },
+  { canonical: "claude-opus-4-7", provider: "anthropic", providerModel: "claude-opus-4-7", contextTokens: 200_000, maxOutputTokens: 64_000, thinking: FULL, images: true },
   { canonical: "claude-opus-4-8", provider: "anthropic", providerModel: "claude-opus-4-8", contextTokens: 200_000, maxOutputTokens: 64_000, thinking: FULL, images: true },
   // OpenAI
   { canonical: "gpt-4o", provider: "openai", providerModel: "gpt-4o", contextTokens: 128_000, maxOutputTokens: 16_384, thinking: [], images: true },
@@ -96,9 +103,9 @@ export const MODEL_CATALOG: readonly CatalogModel[] = [
     canonical: `antigravity/${id}`,
     provider: "antigravity",
     providerModel: id,
-    contextTokens: id.includes("claude") ? 200_000 : id.includes("gemini-3") ? 1_000_000 : 1_000_000,
-    maxOutputTokens: id.includes("claude") ? 64_000 : 65_536,
-    thinking: id.includes("thinking") || id.includes("-high") || id.includes("-low") || id.includes("gemini-3") ? FULL : STD,
+    contextTokens: id.includes("claude") ? 200_000 : id.startsWith("gpt-5") ? 400_000 : id.includes("gemini-3") ? 1_000_000 : 1_000_000,
+    maxOutputTokens: id.includes("claude") ? 64_000 : id.startsWith("gpt-5") ? 128_000 : 65_536,
+    thinking: id.includes("thinking") || id.includes("-high") || id.includes("-low") || id.includes("gemini-3") || id.startsWith("gpt-5") ? FULL : STD,
     images: !id.includes("gpt-oss"),
     company: id.includes("claude") ? "Anthropic via Antigravity" : id.includes("gpt") ? "OpenAI via Antigravity" : "Google Antigravity",
   })),

package/src/ai/providers/anthropic.ts

@@ -11,6 +11,12 @@ const DEPRECATED_TEMPERATURE = "`temperature` is deprecated for this model.";
 const CLAUDE_CODE_VERSION = "2.1.63";
 const CLAUDE_CODE_SYSTEM_INSTRUCTION = "You are a Claude agent, built on Anthropic's Claude Agent SDK.";
 const CLAUDE_BILLING_HEADER_PREFIX = "x-anthropic-billing-header:";
+/** Betas needed for API-key requests: interleaved-thinking enables thinking+tools,
+ *  prompt-caching-scope gives scoped cache breakpoints. */
+const ANTHROPIC_API_KEY_BETA = [
+  "interleaved-thinking-2025-05-14",
+  "prompt-caching-scope-2026-01-05",
+].join(",");
 const ANTHROPIC_OAUTH_BETA = [
   "claude-code-20250219",
   "oauth-2025-04-20",
@@ -99,7 +105,7 @@ type AnthropicMessage = { role: string; content: string | AnthropicContentBlock[
 export function anthropicNativizable(m: Message, model: string, thinkingEnabled: boolean): boolean {
   return thinkingEnabled
     && !!m.toolUse?.length
-    && !!m.reasoningArtifacts?.some(a => a.provider === "anthropic" && a.model === model && ((!!a.signature && !!a.text) || !!a.redacted));
+    && !!m.reasoningArtifacts?.some(a => a.provider === "anthropic" && a.model === model && (!!a.signature || !!a.redacted));
 }
 
 /** Build Anthropic wire messages, reconstructing native tool_use / tool_result / thinking
@@ -121,7 +127,7 @@ export function buildAnthropicMessages(messages: Message[], model: string, think
       const blocks: AnthropicContentBlock[] = [];
       for (const a of m.reasoningArtifacts!) {
         if (a.provider !== "anthropic" || a.model !== model) continue;
-        if (a.signature && a.text) blocks.push({ type: "thinking", thinking: a.text, signature: a.signature });
+        if (a.signature) blocks.push({ type: "thinking", thinking: a.text ?? "", signature: a.signature });
         else if (a.redacted) blocks.push({ type: "redacted_thinking", data: a.redacted });
       }
       for (const tu of m.toolUse!) blocks.push({ type: "tool_use", id: tu.id, name: tu.tool, input: tu.arguments });
@@ -454,6 +460,7 @@ function headersFor(credential: Credential, stream: boolean): Record<string, str
       "content-type": "application/json",
       "x-api-key": credential.token,
       "anthropic-version": "2023-06-01",
+      "anthropic-beta": ANTHROPIC_API_KEY_BETA,
     };
   }
   throw new Error("anthropic adapter requires a credential");

package/src/agent/tool-registry.ts

@@ -1,54 +0,0 @@
-import { readTool, writeTool, editTool, bashTool, findTool, searchTool, lsTool, type ToolResult } from "./tools";
-
-export type ToolHandler = (args: Record<string, any>, cwd: string) => Promise<ToolResult>;
-
-export const DEFAULT_TOOLS: Record<string, ToolHandler> = {
-  read: (a, cwd) => readTool(a.filePath ?? a.path, a.lineRange ?? a.range, cwd, !!a.raw),
-  write: (a, cwd) => writeTool(a.filePath ?? a.path, a.content ?? "", cwd),
-  edit: (a, cwd) => editTool(a.filePath ?? a.path, a.editBlock ?? a.edit ?? "", cwd),
-  bash: (a, cwd) => bashTool(a.command ?? a.cmd, cwd, typeof a.timeoutMs === "number" ? a.timeoutMs : undefined, typeof a.cwd === "string" ? a.cwd : (typeof a.subdir === "string" ? a.subdir : undefined), a.env && typeof a.env === "object" ? a.env : undefined),
-  find: (a, cwd) => findTool(a.globPattern ?? a.pattern, cwd),
-  search: (a, cwd) => searchTool(a.pattern, a.globPattern ?? "*", cwd, !!(a.ignoreCase ?? a.i), { before: a.before, after: a.after, context: a.context, maxMatches: a.maxMatches }),
-  ls: (a, cwd) => lsTool(a.dirPath ?? a.path ?? a.dir ?? ".", cwd),
-};
-
-export const TOOL_PROTOCOL = [
-  "You have these tools (call exactly ONE per step):",
-  "1. read   {filePath, lineRange?, raw?} — read a file",
-  "2. write  {filePath, content}         — create/overwrite a file",
-  "3. edit   {filePath, editBlock}       — replace/insert lines",
-  "4. bash   {command, timeoutMs?, cwd?, env?} — run a shell command",
-  "5. find   {globPattern}               — find files by name",
-  "6. search {pattern, globPattern?, ignoreCase?, context?, maxMatches?} — grep",
-  "7. ls     {dirPath}                   — list a directory",
-  "8. done   {reason?}                   — call when done",
-  "",
-  "Reply with STRICT JSON only:",
-  '{ "tool": "<name>", "arguments": { ... } }',
-].join("\n");
-
-export const READONLY_TOOL_PROTOCOL = [
-  "You have these READ-ONLY tools:",
-  "1. read   {filePath, lineRange?}      — read a file",
-  "2. find   {globPattern}               — find files by name",
-  "3. search {pattern, globPattern?, ignoreCase?} — grep",
-  "4. ls     {dirPath}                   — list a directory",
-  "5. done   {reason?}                   — call when complete",
-  "",
-  "Reply with STRICT JSON only:",
-  '{ "tool": "<name>", "arguments": { ... } }',
-].join("\n");
-
-export function nearestToolName(name: string, known: string[]): string | undefined {
-  const want = name.trim().toLowerCase();
-  if (!want) return undefined;
-  let best: string | undefined;
-  let bestD = Infinity;
-  for (const k of known) {
-    const kl = k.toLowerCase();
-    if (kl === want) return k;
-    const d = kl.startsWith(want) || want.startsWith(kl) ? 1 : 10;
-    if (d < bestD) { bestD = d; best = k; }
-  }
-  return bestD <= 2 ? best : undefined;
-}