session: title generation fails silently on free opencode models, leaving sessions permanently untitled #58411

Summary

Automatic session-title generation silently fails for sessions whose primary model is any free opencode/* model (e.g. opencode/glm-5.3-flash-free, opencode/deepseek-v4.1-flash-free). The session title stays empty forever and the TUI renders it as “New session — ”. There is no log line, no retry, and no fallback (e.g. to the first user message), so the failure is completely invisible — to the user and to debugging.

Environment

• opencode version: 2.1.3 (latest channel, Homebrew opencode-v2)
• OS: Darwin 27.0.0 arm64 (macOS)
• Terminal: VS Code integrated terminal (TERM_PROGRAM=vscode, xterm-256color, truecolor)
• Shell: /bin/zsh
• Install/channel: latest
• Active plugins: none (verified in global config)

Reproduction

1. Run a session whose primary model is a free opencode-zen model, without --title:

   opencode run -m opencode/glm-5.3-flash-free "Reply with two characters: OK"

2. Wait for the reply to finish, then wait 30+ seconds.

3. Check the session title (TUI session list, or GET /api/session): title is null/undefined — the TUI shows “New session — ”.

Results from 5 identical runs (same machine, same day, default log level) — reproducible consistently:

Also observed on ordinary TUI sessions opened with the default free model — every such session that day ended untitled, which is what surfaced this bug.

Additional observations:

• Not a timing problem: still untitled at t+30s and minutes later.
• The title LLM call persists no message record in either the failing or the successful arm, so nothing is observable session-side.
• Default-level service logs contain zero lines about title generation in either arm.

Expected Behavior

• If title generation fails, fall back — e.g. use the (truncated) first user message as the title — and/or retry. packages/core/test/session-title.test.ts already anticipates fallback (“falls back to the primary model when the small model fails”, “uses the primary model at its lowest variant without a small model”).
• Emit at least a warning log with provider/model and the upstream error, so failures are diagnosable.

Actual Behavior

Silent no-op: the LLM attempt failure is swallowed, nothing is logged, and the session remains untitled forever.

Additional Context

Source reading on dev (paths relative to repo root):

1. packages/core/src/session/runner/llm.ts (~L171): once the first user message is promoted, title.generate(sessionID) is started as a bare FiberMap fiber (onlyIfMissing: true) with no failure observer. In-file comment: “Title generation starts once input is visible and must not delay model execution.” — the title call runs concurrently with the first assistant turn.

2. packages/core/src/session/title.ts — every failure path returns silently:

   if (LLMEvent.is.providerError(event)) failed = true
   ...
   Effect.catchTag("AI.Error", () => Effect.sync(() => { failed = true }))
   ...
   if (failed) return   // attempt() → undefined
   ...
   if (!title) return   // generate() → silent exit

   The module contains no logging at all.

3. packages/core/src/model.ts — Model.small(providerID) only matches a hardcoded family list (gpt-luna, gemini-flash-lite, gemini-flash, claude-haiku). The opencode free catalog (glm-5.3-flash-free, deepseek-v4.1-flash-free, big-pickle, muse-spark-1.3-contributor-free, …) matches none, so selection falls through to the primary model itself — i.e. the very model already serving the first turn. On free models with tight concurrency, this extra concurrent request is the most plausible trigger for the hidden upstream error (unverified — no error surfaces anywhere to confirm it).

Probable impact: everyone running free opencode/* models (these are the provider's preferred defaults, see PREFERRED_FREE in model.ts) accumulates permanently untitled sessions.

Related but distinct open issues: #52004 (small-model selection prefers a metered model), #54185 (title agent produces conversational replies), #54401 (title request has no output limit). None of them cover the silent-failure / permanently-untitled symptom.

Suggested fixes (any of):

• log a warning including provider/model and the upstream error whenever a title attempt fails;
• schedule the title call after the first turn settles, or retry on later turns (the isUntitled re-check already exists);
• fall back to the truncated first user message as the title when generation fails.