Skip to content

Node-based terminal manager for AI coding agents — tmux-backed terminals and parallel agent sessions as draggable nodes on an infinite pan/zoom canvas. macOS, Linux, and a browser Server Edition.

fix(grok): usar session_summary como título mientras grok aún no ha nombrado la sesión

Xavi Guardia1268d189fbfc
diff --git a/docs/grok-agent.md b/docs/grok-agent.md
index 09c90ecdc..670e788d3 100644
--- a/docs/grok-agent.md
+++ b/docs/grok-agent.md
@@ -275,7 +275,15 @@ resolve **nothing** there (`GROK_ENCODED_CWD_MAX_BYTES`); a session id must matc
The session **name** (what `/resume` shows and what a node title with `titleAuto` adopts) is read by
`core/grok-session.ts` → `pickGrokSessionMeta` over `summary.json`, in preference order
-`TITLE_KEYS = ['title', 'generated_title']`, plus `current_model_id` as the model. Reads are capped
+`TITLE_KEYS = ['title', 'generated_title', 'session_summary']`, plus `current_model_id` as the model.
+`generated_title` is the real key (`/rename` rewrites it in place). `session_summary` is accepted
+only as a fallback, and only when it is ≤ 80 characters — the longest observed title in a 49-file
+1.0.13 corpus is 49 characters; a longer value is **refused, not truncated**, because this result
+is adopted as the node title with no other length cap (header, sidebar, kanban card, window title,
+`project.json`). 20 of those 49 files have `session_summary` and no `generated_title` yet (young
+sessions); without the fallback they read as nameless. `'title'` stays first as forward
+compatibility: it is absent from every real file, costs one failed lookup, and would win the day
+grok splits a manual title from a generated one. Reads are capped
at 256 KB and answer `null` — never a throw — for an absent, oversized or unparseable file.
Resolution is a **direct open** of the directory a hook told us about: `rememberGrokSessionDir` /
`grokSessionDirFor` / `forgetGrokSession` keep a bounded (512-entry, least-recently-seen-evicted)
diff --git a/src/core/grok-session.test.ts b/src/core/grok-session.test.ts
index add3bda6e..acc5897b2 100644
--- a/src/core/grok-session.test.ts
+++ b/src/core/grok-session.test.ts
@@ -52,6 +52,42 @@ describe('pickGrokSessionMeta', () => {
expect(pickGrokSessionMeta(JSON.stringify({ generated_title: 'auto' }))?.title).toBe('auto')
})
+ it('falls back to session_summary when grok has not generated a title yet', () => {
+ // Measured: 20 of 49 local summary.json files (1.0.13) carry session_summary and no
+ // generated_title — young sessions. Those used to read as nameless.
+ expect(
+ pickGrokSessionMeta(JSON.stringify({ session_summary: 'Release checklist review' }))?.title
+ ).toBe('Release checklist review')
+ })
+
+ it('prefers generated_title over session_summary', () => {
+ expect(
+ pickGrokSessionMeta(
+ JSON.stringify({ generated_title: 'Plan the release', session_summary: 'Something else' })
+ )?.title
+ ).toBe('Plan the release')
+ })
+
+ it('refuses a session_summary longer than the cap rather than truncating it', () => {
+ // The cap exists because pickGrokSessionMeta feeds the node title with no other length
+ // bound, so a sentence adopted here reaches the header, the sidebar, the kanban card,
+ // the window title and project.json. A truncated sentence is "degrades to something
+ // wrong"; null is the node keeping its own name.
+ const long = 'x'.repeat(81)
+ expect(pickGrokSessionMeta(JSON.stringify({ session_summary: long }))?.title).toBeNull()
+ expect(pickGrokSessionMeta(JSON.stringify({ session_summary: 'x'.repeat(80) }))?.title).toBe(
+ 'x'.repeat(80)
+ )
+ })
+
+ it('does not let an over-long session_summary hide a generated title', () => {
+ expect(
+ pickGrokSessionMeta(
+ JSON.stringify({ generated_title: 'Plan the release', session_summary: 'x'.repeat(200) })
+ )?.title
+ ).toBe('Plan the release')
+ })
+
it('returns a null TITLE (not a null meta) when the session has no name yet', () => {
// A session with a model but no title is normal early in its life — the node keeps its own
// title and the poll simply finds nothing to adopt.
diff --git a/src/core/grok-session.ts b/src/core/grok-session.ts
index 31d7b7fa0..3f438a7f0 100644
--- a/src/core/grok-session.ts
+++ b/src/core/grok-session.ts
@@ -49,19 +49,26 @@ export interface GrokSessionMeta {
/**
* Title keys in PREFERENCE order.
*
- * MEASURED (grok 1.0.13, 2026-09-02): there is only ONE key. `/rename <name>` typed into a live
- * session rewrites **`generated_title`** in place — the same field the model-generated name uses —
- * and no `title` key ever appears in `summary.json`. Grok does not distinguish a manual title from a
- * generated one the way claude's `custom-title` / `ai-title` pair does.
+ * MEASURED (grok 1.0.13, 2026-09-02): `/rename <name>` typed into a live session rewrites
+ * **`generated_title`** in place — the same field the model-generated name uses — and no `title`
+ * key ever appears in `summary.json`. Grok does not distinguish a manual title from a generated one
+ * the way claude's `custom-title` / `ai-title` pair does.
*
- * `'title'` was here as an explicitly-marked guess at the manual key, kept first so a real manual
- * title would win the moment someone confirmed it. Someone did, and the answer is that the key does
- * not exist. It stays in the list ONLY as a harmless forward compatibility: if a future grok ever
- * splits the two, a manual title would win immediately. A key that is absent from every file costs
- * one failed lookup and can never produce a wrong name — which is why removing it would buy nothing
- * and lose that.
+ * `'title'` stays first as forward compatibility: a key absent from every file costs one failed
+ * lookup and can never produce a wrong name. If a future grok splits the two, a manual title wins.
+ *
+ * `session_summary` is the fallback for a session too young to have `generated_title` yet. Measured
+ * across 49 local summary.json files (1.0.13): 20 of them carry `session_summary` and no
+ * `generated_title`; where both exist they are identical in 28 of 29; the longest value in the
+ * corpus is 49 characters. It is accepted only under SESSION_SUMMARY_TITLE_MAX: a longer value is
+ * **refused, not truncated**, because this result is adopted as the node title with no other length
+ * cap on the path (header, sidebar, kanban card, window title, project.json). Null is the node
+ * keeping its own name; a truncated sentence is a wrong name.
*/
-const TITLE_KEYS = ['title', 'generated_title'] as const
+const TITLE_KEYS = ['title', 'generated_title', 'session_summary'] as const
+
+/** Longest observed title in the 1.0.13 corpus is 49 characters. 80 is "a title, generously". */
+const SESSION_SUMMARY_TITLE_MAX = 80
/** Pure: the meta from a summary.json body. null when this is not a summary object at all. */
export function pickGrokSessionMeta(summaryJson: string): GrokSessionMeta | null {
@@ -76,10 +83,12 @@ export function pickGrokSessionMeta(summaryJson: string): GrokSessionMeta | null
let title: string | null = null
for (const k of TITLE_KEYS) {
const v = rec[k]
- if (typeof v === 'string' && v.trim()) {
- title = v.trim()
- break
- }
+ if (typeof v !== 'string') continue
+ const trimmed = v.trim()
+ if (!trimmed) continue
+ if (k === 'session_summary' && trimmed.length > SESSION_SUMMARY_TITLE_MAX) continue
+ title = trimmed
+ break
}
const model = typeof rec.current_model_id === 'string' ? rec.current_model_id : null
return { title, model }