jeo-code 0.6.37 → 0.6.39

package/CHANGELOG.md

@@ -6,6 +6,36 @@ 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`.
 
+## [0.6.39] - 2026-06-21
+_A long "thinking" phase no longer trips a false stream-idle retry: reasoning/thinking deltas now act as a stream heartbeat, so a model that streams thought tokens past the idle window before any visible text is no longer mistaken for a stalled stream and retried (which discarded the in-progress reasoning). Re-verified leak-free (`mem-probe`, 2000 turns, −525 bytes/turn slope) with a fresh `jeo --tmux` boot battery (6/6)._
+
+### Fixed
+- **A long "thinking" phase no longer trips a false `stream idle for <ms>ms (no chunk)` retry.** Reasoning/thinking deltas are routed to `onReasoning` and never yielded as a stream chunk, so a model that streams thought tokens for longer than the idle window (default 120s) before emitting visible text looked stalled and got retried — discarding the in-progress reasoning. The per-chunk idle watchdog now treats reasoning activity as a heartbeat (`lastActivityAt`) and re-arms while thinking is actively streaming; it aborts only a genuinely silent stream. The `JEO_STREAM_IDLE_MS` override still covers models that reason fully server-side and stream no thought tokens at all.
+
+### Verified
+- `bun run typecheck` clean; full suite **1752 pass / 0 fail** (216 files), including the new `retryableStream` reasoning-heartbeat tests in `test/round-b.test.ts` (a silent-but-thinking stream stays alive; the watchdog still fires once reasoning activity stops).
+- `scripts/mem-probe.ts` (2000 turns × 40 tools): no long-term leak (per-turn slope **−525 bytes/turn**, exit-listeners flat at 1, heap returns to baseline).
+- `scripts/tmux-verify.sh battery`: **6/6 PASSED** on a fresh `jeo --tmux` boot (boot, `/help`, unknown `$skill`, `/agents`, `$ultragoal`, unresolved `/command`).
+
+## [0.6.38] - 2026-06-21
+_The OKF concept-bundle memory no longer silently drops what it learns: a legacy single-doc `MEMORY.md` (or a text-only distill fallback) can no longer shadow the concept bundle, break OKF conformance, or lose a turn's learnings — its content is folded into the concept merge and the stale blob is archived. Re-verified leak-free (`mem-probe`, 2000 turns, −570 bytes/turn slope) with a fresh `jeo --tmux` boot battery (6/6)._
+
+### Fixed
+- **A text-only distill fallback no longer loses the turn's learnings.** When the model returns no extractable JSON, `distillSessionMemory` used to write the raw text to a single-doc `MEMORY.md` blob — but `memoryPromptSection` injects that blob ONLY when the concept bundle is empty (concepts always win), so once any concept existed the fallback's learnings were recorded yet never injected. Now, if a concept bundle already exists, the dead blob is suppressed and the prior concepts are kept as the durable memory (`skipped: "distill produced no JSON; kept existing concept bundle (legacy blob suppressed)"`).
+- **A lingering legacy `MEMORY.md` can no longer break OKF conformance.** A frontmatter-less single-doc `MEMORY.md` coexisting with the concept bundle made `validateBundle` non-conformant (`error: concept document missing YAML frontmatter block`) while the loader silently skipped it — an inconsistency that left the bundle permanently invalid (`migrateLegacyMemory` only runs on the explicit `jeo memory-migrate`). Distill now feeds the legacy doc's content into the concept-merge context so its learnings are absorbed, then archives it to `MEMORY.md.bak` on a successful JSON distill — off the active read path, with rollback preserved.
+- **Concept enumeration skips non-concept blobs.** `existingConcepts` now skips any `.md` without YAML frontmatter, so a stray `MEMORY.md` can never be mis-parsed as a concept during the merge.
+
+### Changed
+- **Distill prompt enforces concept granularity.** The merge prompt now forbids a catch-all "Project Memory Bundle" mega-concept and instructs one concept per distinct fact/command/gotcha/preference, so the bundle grows as discrete, cross-linkable concepts.
+- **Per-session injection budget raised** from 3,000 to 5,000 chars (`MEMORY_INJECT_MAX_CHARS`) so a healthy multi-concept bundle is injected in full.
+
+### Verified
+- `bun run typecheck` clean; full suite `1751 pass / 0 fail` (216 files).
+- Memory/OKF suites `57 pass / 0 fail` across `memory`, `memory-okf`, `memory-distill-okf`, `memory-migration-okf`, `memory-search-okf`, `memory-graph-okf`.
+- Live bundle proof: `validateBundle` flips `false (error: MEMORY.md)` → `true (no issues)` after the legacy archive, injection block intact (3,220 chars); bundle now holds discrete facts/commands/gotchas/preferences concepts + a valid `index.md`.
+- `scripts/mem-probe.ts` (2000 turns × 40 tools) reports no long-term leak (−570 bytes/turn, exit-listeners flat at 1); `scripts/tmux-verify.sh battery` passes 6/6 on a fresh `jeo --tmux` boot.
+
+
 ## [0.6.37] - 2026-06-20
 _Two dead-end fixes: the boxed prompt's ↑/↓ now recalls input history on a soft-wrapped one-liner (only a genuine multi-line draft gets in-box caret nav), and every terminating Spec-first stage (deep-interview, ralplan, team) now surfaces a user-visible answer instead of silently stalling — re-verified leak-free (`mem-probe`) with a fresh `jeo --tmux` boot._
 

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`) -->
+- **[0.6.39]** (2026-06-21) — A long "thinking" phase no longer trips a false stream-idle retry: reasoning/thinking deltas now act as a stream heartbeat, so a model that streams thought tokens past the idle window before any visible text is no longer mistaken for a stalled stream and retried (which discarded the in-progress reasoning). Re-verified leak-free (`mem-probe`, 2000 turns, −525 bytes/turn slope) with a fresh `jeo --tmux` boot battery (6/6).
+- **[0.6.38]** (2026-06-21) — The OKF concept-bundle memory no longer silently drops what it learns: a legacy single-doc `MEMORY.md` (or a text-only distill fallback) can no longer shadow the concept bundle, break OKF conformance, or lose a turn's learnings — its content is folded into the concept merge and the stale blob is archived. Re-verified leak-free (`mem-probe`, 2000 turns, −570 bytes/turn slope) with a fresh `jeo --tmux` boot battery (6/6).
 - **[0.6.37]** (2026-06-20) — Two dead-end fixes: the boxed prompt's ↑/↓ now recalls input history on a soft-wrapped one-liner (only a genuine multi-line draft gets in-box caret nav), and every terminating Spec-first stage (deep-interview, ralplan, team) now surfaces a user-visible answer instead of silently stalling — re-verified leak-free (`mem-probe`) with a fresh `jeo --tmux` boot.
 - **[0.6.36]** (2026-06-20) — When `jeo --tmux` flips the mouse on so you can drag-select, the drag now actually lands on the system clipboard — the in-session tmux profile sets `set-clipboard on` + a local `copy-command` on the CURRENT session only — plus `/help` documents the drag-to-copy and the Shift/Option-drag escape hatch, re-verified leak-free (`mem-probe`) with a fresh `jeo --tmux` boot.
 - **[0.6.35]** (2026-06-20) — The prompt's Ctrl+C now clears a non-empty input box on the first press and only exits on the next press of an empty box; plus app-driven system-clipboard copy (OSC 52 + local tool, tmux-aware), drag-and-drop image attachment, a Ctrl-L prompt re-anchor, and a SIGCONT resume repaint — verified leak-free (`mem-probe`) with a fresh `jeo --tmux` boot check.
-- **[0.6.34]** (2026-06-20) — Per-session model memory — each saved session now remembers the model it was last using and restores it on `/resume` — plus clearer `jeo --tmux` attach diagnostics, a tmux session-name double-dash fix, and a more robust no-leak probe gate.
-- **[0.6.33]** (2026-06-19) — A redesigned `jeo` forge mark — a hollow line-board crayfish/eyeglass emblem drawn as thick rounded-corner tubes (no letters, no DNA helix) — that now renders inside compact-scaled forge cards, plus a unified verification directive that adds gjc's test-quality contract, and a fresh `jeo --tmux` no-leak re-verification.
 
 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`) -->
+- **[0.6.39]** (2026-06-21) — A long "thinking" phase no longer trips a false stream-idle retry: reasoning/thinking deltas now act as a stream heartbeat, so a model that streams thought tokens past the idle window before any visible text is no longer mistaken for a stalled stream and retried (which discarded the in-progress reasoning). Re-verified leak-free (`mem-probe`, 2000 turns, −525 bytes/turn slope) with a fresh `jeo --tmux` boot battery (6/6).
+- **[0.6.38]** (2026-06-21) — The OKF concept-bundle memory no longer silently drops what it learns: a legacy single-doc `MEMORY.md` (or a text-only distill fallback) can no longer shadow the concept bundle, break OKF conformance, or lose a turn's learnings — its content is folded into the concept merge and the stale blob is archived. Re-verified leak-free (`mem-probe`, 2000 turns, −570 bytes/turn slope) with a fresh `jeo --tmux` boot battery (6/6).
 - **[0.6.37]** (2026-06-20) — Two dead-end fixes: the boxed prompt's ↑/↓ now recalls input history on a soft-wrapped one-liner (only a genuine multi-line draft gets in-box caret nav), and every terminating Spec-first stage (deep-interview, ralplan, team) now surfaces a user-visible answer instead of silently stalling — re-verified leak-free (`mem-probe`) with a fresh `jeo --tmux` boot.
 - **[0.6.36]** (2026-06-20) — When `jeo --tmux` flips the mouse on so you can drag-select, the drag now actually lands on the system clipboard — the in-session tmux profile sets `set-clipboard on` + a local `copy-command` on the CURRENT session only — plus `/help` documents the drag-to-copy and the Shift/Option-drag escape hatch, re-verified leak-free (`mem-probe`) with a fresh `jeo --tmux` boot.
 - **[0.6.35]** (2026-06-20) — The prompt's Ctrl+C now clears a non-empty input box on the first press and only exits on the next press of an empty box; plus app-driven system-clipboard copy (OSC 52 + local tool, tmux-aware), drag-and-drop image attachment, a Ctrl-L prompt re-anchor, and a SIGCONT resume repaint — verified leak-free (`mem-probe`) with a fresh `jeo --tmux` boot check.
-- **[0.6.34]** (2026-06-20) — Per-session model memory — each saved session now remembers the model it was last using and restores it on `/resume` — plus clearer `jeo --tmux` attach diagnostics, a tmux session-name double-dash fix, and a more robust no-leak probe gate.
-- **[0.6.33]** (2026-06-19) — A redesigned `jeo` forge mark — a hollow line-board crayfish/eyeglass emblem drawn as thick rounded-corner tubes (no letters, no DNA helix) — that now renders inside compact-scaled forge cards, plus a unified verification directive that adds gjc's test-quality contract, and a fresh `jeo --tmux` no-leak re-verification.
 
 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`) -->
+- **[0.6.39]** (2026-06-21) — A long "thinking" phase no longer trips a false stream-idle retry: reasoning/thinking deltas now act as a stream heartbeat, so a model that streams thought tokens past the idle window before any visible text is no longer mistaken for a stalled stream and retried (which discarded the in-progress reasoning). Re-verified leak-free (`mem-probe`, 2000 turns, −525 bytes/turn slope) with a fresh `jeo --tmux` boot battery (6/6).
+- **[0.6.38]** (2026-06-21) — The OKF concept-bundle memory no longer silently drops what it learns: a legacy single-doc `MEMORY.md` (or a text-only distill fallback) can no longer shadow the concept bundle, break OKF conformance, or lose a turn's learnings — its content is folded into the concept merge and the stale blob is archived. Re-verified leak-free (`mem-probe`, 2000 turns, −570 bytes/turn slope) with a fresh `jeo --tmux` boot battery (6/6).
 - **[0.6.37]** (2026-06-20) — Two dead-end fixes: the boxed prompt's ↑/↓ now recalls input history on a soft-wrapped one-liner (only a genuine multi-line draft gets in-box caret nav), and every terminating Spec-first stage (deep-interview, ralplan, team) now surfaces a user-visible answer instead of silently stalling — re-verified leak-free (`mem-probe`) with a fresh `jeo --tmux` boot.
 - **[0.6.36]** (2026-06-20) — When `jeo --tmux` flips the mouse on so you can drag-select, the drag now actually lands on the system clipboard — the in-session tmux profile sets `set-clipboard on` + a local `copy-command` on the CURRENT session only — plus `/help` documents the drag-to-copy and the Shift/Option-drag escape hatch, re-verified leak-free (`mem-probe`) with a fresh `jeo --tmux` boot.
 - **[0.6.35]** (2026-06-20) — The prompt's Ctrl+C now clears a non-empty input box on the first press and only exits on the next press of an empty box; plus app-driven system-clipboard copy (OSC 52 + local tool, tmux-aware), drag-and-drop image attachment, a Ctrl-L prompt re-anchor, and a SIGCONT resume repaint — verified leak-free (`mem-probe`) with a fresh `jeo --tmux` boot check.
-- **[0.6.34]** (2026-06-20) — Per-session model memory — each saved session now remembers the model it was last using and restores it on `/resume` — plus clearer `jeo --tmux` attach diagnostics, a tmux session-name double-dash fix, and a more robust no-leak probe gate.
-- **[0.6.33]** (2026-06-19) — A redesigned `jeo` forge mark — a hollow line-board crayfish/eyeglass emblem drawn as thick rounded-corner tubes (no letters, no DNA helix) — that now renders inside compact-scaled forge cards, plus a unified verification directive that adds gjc's test-quality contract, and a fresh `jeo --tmux` no-leak re-verification.
 
 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`) -->
+- **[0.6.39]** (2026-06-21) — A long "thinking" phase no longer trips a false stream-idle retry: reasoning/thinking deltas now act as a stream heartbeat, so a model that streams thought tokens past the idle window before any visible text is no longer mistaken for a stalled stream and retried (which discarded the in-progress reasoning). Re-verified leak-free (`mem-probe`, 2000 turns, −525 bytes/turn slope) with a fresh `jeo --tmux` boot battery (6/6).
+- **[0.6.38]** (2026-06-21) — The OKF concept-bundle memory no longer silently drops what it learns: a legacy single-doc `MEMORY.md` (or a text-only distill fallback) can no longer shadow the concept bundle, break OKF conformance, or lose a turn's learnings — its content is folded into the concept merge and the stale blob is archived. Re-verified leak-free (`mem-probe`, 2000 turns, −570 bytes/turn slope) with a fresh `jeo --tmux` boot battery (6/6).
 - **[0.6.37]** (2026-06-20) — Two dead-end fixes: the boxed prompt's ↑/↓ now recalls input history on a soft-wrapped one-liner (only a genuine multi-line draft gets in-box caret nav), and every terminating Spec-first stage (deep-interview, ralplan, team) now surfaces a user-visible answer instead of silently stalling — re-verified leak-free (`mem-probe`) with a fresh `jeo --tmux` boot.
 - **[0.6.36]** (2026-06-20) — When `jeo --tmux` flips the mouse on so you can drag-select, the drag now actually lands on the system clipboard — the in-session tmux profile sets `set-clipboard on` + a local `copy-command` on the CURRENT session only — plus `/help` documents the drag-to-copy and the Shift/Option-drag escape hatch, re-verified leak-free (`mem-probe`) with a fresh `jeo --tmux` boot.
 - **[0.6.35]** (2026-06-20) — The prompt's Ctrl+C now clears a non-empty input box on the first press and only exits on the next press of an empty box; plus app-driven system-clipboard copy (OSC 52 + local tool, tmux-aware), drag-and-drop image attachment, a Ctrl-L prompt re-anchor, and a SIGCONT resume repaint — verified leak-free (`mem-probe`) with a fresh `jeo --tmux` boot check.
-- **[0.6.34]** (2026-06-20) — Per-session model memory — each saved session now remembers the model it was last using and restores it on `/resume` — plus clearer `jeo --tmux` attach diagnostics, a tmux session-name double-dash fix, and a more robust no-leak probe gate.
-- **[0.6.33]** (2026-06-19) — A redesigned `jeo` forge mark — a hollow line-board crayfish/eyeglass emblem drawn as thick rounded-corner tubes (no letters, no DNA helix) — that now renders inside compact-scaled forge cards, plus a unified verification directive that adds gjc's test-quality contract, and a fresh `jeo --tmux` no-leak re-verification.
 
 See [CHANGELOG.md](CHANGELOG.md) for the full history.
 <!-- CHANGELOG:END -->

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "jeo-code",
-  "version": "0.6.37",
+  "version": "0.6.39",
   "description": "Clean, highly optimized AI coding agent using spec-first loop",
   "type": "module",
   "main": "src/cli.ts",

package/src/agent/memory.ts

@@ -22,7 +22,7 @@ import { buildConceptGraph, expandByGraph, lintConceptGraph, type GraphLintRepor
 /** On-disk document cap — the distill prompt instructs the model to stay under it. */
 export const MEMORY_MAX_CHARS = 6_000;
 /** Per-session prompt injection budget. */
-export const MEMORY_INJECT_MAX_CHARS = 3_000;
+export const MEMORY_INJECT_MAX_CHARS = 5_000;
 /** Transcript slice fed to the distill call. */
 const TRANSCRIPT_MAX_CHARS = 12_000;
 /** A session shorter than this has nothing durable to learn. */
@@ -560,6 +560,7 @@ export async function distillSessionMemory(
         try {
           const content = await fs.readFile(file, "utf-8");
           const parsed = parseConcept(content);
+          if (!parsed.hasFrontmatter) continue; // skip MEMORY.md and other legacy blobs
           existingConcepts.push({
             type: parsed.frontmatter.type,
             title: parsed.frontmatter.title || "",
@@ -573,6 +574,11 @@ export async function distillSessionMemory(
         } catch {}
       }
     } catch {}
+    // A lingering single-doc MEMORY.md (legacy, or a prior text-only bootstrap
+    // fallback) is NOT a concept document — loadConcepts ignores it and it breaks
+    // OKF conformance. Feed its content into the merge context so its learnings
+    // are absorbed into concepts, then archive it on a successful JSON distill.
+    const legacyDoc = await loadMemory(cwd);
 
     const prompt: Message[] = [
       {
@@ -583,12 +589,21 @@ export async function distillSessionMemory(
           "Keep ONLY what helps future sessions in THIS repository: repo facts (structure, conventions, key files), " +
           "commands that work (build/test/run), gotchas (failures and their fixes), and user preferences. " +
           "Drop session-specific noise (one-off tasks, transient errors, conversational detail). " +
+          "CRITICAL RULES for concept granularity:\n" +
+          "  1. Create ONE concept per distinct fact/command/gotcha/preference — never combine multiple learnings into a single concept.\n" +
+          "  2. NEVER create a catch-all 'Project Memory Bundle' or similar mega-concept that lists many things. Split it.\n" +
+          "  3. Each Command concept covers exactly one command or workflow (e.g., 'bun test', NOT 'All bun commands').\n" +
+          "  4. Each Gotcha covers exactly one failure mode and its fix.\n" +
+          "  5. Each UserPreference covers exactly one observable preference.\n" +
+          "  6. RepoFact concepts describe structure/conventions — split by area (e.g., 'CLI entrypoint', 'Test setup').\n" +
+          "  7. If an existing concept is a mega-concept (lists many things), REPLACE it with properly split granular concepts.\n" +
+          "  8. Keep each concept body under 400 chars.\n\n" +
           "You must output a JSON object with a single key \"concepts\", which is an array of concept objects. " +
           "Each concept object must have the following fields:\n" +
           "  - \"type\": one of \"RepoFact\", \"Command\", \"Gotcha\", \"UserPreference\"\n" +
           "  - \"title\": a short, descriptive title (e.g., \"Bun test runner\")\n" +
           "  - \"description\": a brief one-line summary of the concept\n" +
-          "  - \"body\": the detailed markdown content/body of the concept\n" +
+          "  - \"body\": the detailed markdown content/body of the concept (≤400 chars)\n" +
           "  - \"tags\": an array of string tags (optional)\n" +
           "  - \"confidence\": one of \"high\", \"medium\", \"low\" (optional)\n" +
           "  - \"links\": an array of other concept paths/IDs this concept links to (optional)\n\n" +
@@ -598,6 +613,7 @@ export async function distillSessionMemory(
         role: "user",
         content:
           `Existing concepts:\n${JSON.stringify(existingConcepts, null, 2)}\n\n` +
+          (legacyDoc ? `Legacy single-doc memory to fold into concepts:\n${legacyDoc}\n\n` : "") +
           `Session transcript (tail):\n${transcriptTail(history)}`
       }
     ];
@@ -695,9 +711,25 @@ export async function distillSessionMemory(
       if (updatedConcepts.length > 0) {
         await updateLog(bundleDir, updatedConcepts);
       }
+      // Concepts now own the durable memory and the legacy blob (if any) was
+      // folded into the merge above. Archive a lingering MEMORY.md off the active
+      // read path so it can never break OKF conformance or shadow the bundle.
+      if (legacyDoc) {
+        await fs.rename(memoryFilePath(cwd), `${memoryFilePath(cwd)}.bak`).catch(() => {});
+      }
       await cleanupStalePendingFiles(bundleDir);
       return { updated: true };
     } else {
+      // JSON extraction failed (text-only models often wrap or drop the JSON).
+      // The single-doc MEMORY.md is a BOOTSTRAP fallback only: it is injected by
+      // memoryPromptSection ONLY when the OKF concept bundle is empty. Writing it
+      // while concepts already exist would (a) be silently ignored at injection
+      // (concepts win) — losing the learning — and (b) break OKF conformance
+      // (MEMORY.md has no frontmatter). So when a bundle exists, do NOT clobber
+      // it with a dead blob; keep the prior concepts as the durable memory.
+      if (existingConcepts.length > 0) {
+        return { updated: false, skipped: "distill produced no JSON; kept existing concept bundle (legacy blob suppressed)" };
+      }
       const doc = distilled.trim().slice(0, MEMORY_MAX_CHARS);
       if (!doc) return { updated: false, skipped: "model returned an empty document" };
       const file = memoryFilePath(cwd);

package/src/ai/model-manager.ts

@@ -434,30 +434,49 @@ export interface StreamIdleOptions {
    *  Non-interactive contexts opt in (JEO_STREAM_MAX_MS) so a slow-drip stream
    *  (one token every idleMs-ε) cannot run unbounded. */
   deadlineAt?: number;
+  /** Epoch-ms of the most recent stream ACTIVITY, including reasoning/thinking deltas
+   *  that are routed to onReasoning and never yielded as a chunk. A long "thinking"
+   *  phase that streams thought tokens keeps bumping this, so the per-chunk idle
+   *  watchdog re-arms instead of falsely aborting an actively-reasoning stream. Absent
+   *  → only a yielded chunk counts as activity (legacy behaviour). */
+  lastActivityAt?: () => number;
   onIdle?: () => void;
 }
 
-/** `iter.next()`, racing the per-chunk idle timeout AND (when set) the overall deadline. */
+/** `iter.next()`, racing a per-chunk idle watchdog AND (when set) the overall deadline.
+ *  The watchdog re-arms while reasoning activity (idle.lastActivityAt) keeps advancing,
+ *  so a model that streams thinking tokens for longer than idleMs before emitting visible
+ *  text is NOT mistaken for a stalled stream — only a genuinely silent stream aborts. */
 async function nextMaybeIdle(iter: AsyncIterator<string>, idle?: StreamIdleOptions): Promise<IteratorResult<string>> {
   if (!idle) return iter.next();
-  const remaining = idle.deadlineAt !== undefined ? idle.deadlineAt - Date.now() : Infinity;
-  if (remaining <= 0) {
-    idle.onIdle?.();
-    throw new Error(`stream exceeded the overall deadline (JEO_STREAM_MAX_MS) — slow-drip stream aborted`);
-  }
-  const waitMs = Math.min(idle.idleMs, remaining);
-  const deadlineFires = remaining < idle.idleMs;
+  const next = iter.next();
+  // Baseline: this wait began now; reasoning deltas during it move lastActivityAt forward.
+  const startedAt = Date.now();
+  const lastActivity = () => Math.max(startedAt, idle.lastActivityAt?.() ?? 0);
   let timer: ReturnType<typeof setTimeout> | undefined;
-  const timeout = new Promise<never>((_, reject) => {
-    timer = setTimeout(() => {
-      idle.onIdle?.();
-      reject(new Error(deadlineFires
-        ? `stream exceeded the overall deadline (JEO_STREAM_MAX_MS) — slow-drip stream aborted`
-        : `stream idle for ${idle.idleMs}ms (no chunk) — provider sent no token within the idle window (load or long thinking); retrying. Raise JEO_STREAM_IDLE_MS or lower the thinking level if this persists.`));
-    }, waitMs);
+  const watchdog = new Promise<never>((_, reject) => {
+    const arm = () => {
+      const now = Date.now();
+      const overallRemaining = idle.deadlineAt !== undefined ? idle.deadlineAt - now : Infinity;
+      if (overallRemaining <= 0) {
+        idle.onIdle?.();
+        reject(new Error(`stream exceeded the overall deadline (JEO_STREAM_MAX_MS) — slow-drip stream aborted`));
+        return;
+      }
+      const idleRemaining = idle.idleMs - (now - lastActivity());
+      const wait = Math.min(idleRemaining, overallRemaining);
+      if (wait <= 0) {
+        // overallRemaining is > 0 here (pre-checked), so wait<=0 ⟺ the idle window lapsed.
+        idle.onIdle?.();
+        reject(new Error(`stream idle for ${idle.idleMs}ms (no chunk) — provider sent no token within the idle window (load or long thinking); retrying. Raise JEO_STREAM_IDLE_MS or lower the thinking level if this persists.`));
+        return;
+      }
+      timer = setTimeout(arm, wait);
+    };
+    arm();
   });
   try {
-    return await Promise.race([iter.next(), timeout]);
+    return await Promise.race([next, watchdog]);
   } finally {
     if (timer) clearTimeout(timer);
   }
@@ -512,15 +531,30 @@ export function createModelManager(): ModelManager {
         // JEO_STREAM_MAX_MS opts in to an OVERALL deadline (round-14): non-interactive
         // runs can bound a slow-drip stream the per-chunk idle alone never catches.
         let attempt: AbortController | null = null;
+        // Heartbeat: reasoning/thinking deltas are routed to onReasoning and NEVER yielded
+        // as a chunk, so a model that "thinks" (streams thought tokens) for longer than the
+        // idle window would otherwise look stalled and trip a false idle-stall retry. Wrap
+        // onReasoning/onReasoningStart to stamp lastActivityAt; the idle watchdog re-arms
+        // while thinking is actively streaming and only aborts a genuinely silent stream.
+        let lastActivityAt = Date.now();
+        const bump = () => { lastActivityAt = Date.now(); };
+        const userOnReasoning = callOptions.onReasoning;
+        const userOnReasoningStart = callOptions.onReasoningStart;
+        const liveOptions: CallOptions = {
+          ...callOptions,
+          onReasoning: (delta: string) => { bump(); userOnReasoning?.(delta); },
+          onReasoningStart: () => { bump(); userOnReasoningStart?.(); },
+        };
         const makeIter = () => {
           attempt = new AbortController();
           const signal = composeAbort(callOptions.signal, attempt.signal);
-          return streamFn(messages, { ...callOptions, signal }, credential)[Symbol.asyncIterator]();
+          return streamFn(messages, { ...liveOptions, signal }, credential)[Symbol.asyncIterator]();
         };
         const maxMs = streamMaxMs();
         yield* retryableStream(makeIter, retry, {
           idleMs: streamIdleMs(),
           ...(maxMs !== undefined ? { deadlineAt: Date.now() + maxMs } : {}),
+          lastActivityAt: () => lastActivityAt,
           onIdle: () => attempt?.abort(),
         });
       } else {