Read the title events, not the surface.
Inspect whether the latest session/title is fallback, provider, or user.
Read provider, model, message seqs, and cap from session/title-llm-request.
Separate max-tokens, timeout, empty text, and tool-call output.
Keep its response budget separate from the title request.
One symptom, four branches.
max-tokensshared reasoning/text capRun one bounded larger-cap test.
SESSION_TITLE_TIMEOUTlatency or transportMeasure the auxiliary route before extending its deadline.
Compare the durable fold with the UI update.
The rc.8 title-provider schema has no title-specific reasoning field. Raise only the existing auxiliary row for a bounded test, or explicitly route titles to an approved lightweight model.
Prove a bounded workaround.
- id: session-title-llm
config:
targetWords: 5
targetCjkCharacters: 10
maxInputBytes: 4096
maxOutputTokens: 1024
timeoutMs: 60000Restart the Host and use a fresh Session. Require a new request event with the changed cap, a stop finish, and a later provider-sourced title event that survives reload. The larger cap creates room; it does not disable reasoning or guarantee success.
Primary evidence.
- Official report #3468
- rc.8 auxiliary request policy
- rc.8 fallback and provider lifecycle
- rc.8 shipped 64-token configuration
Keep the evidence with the title.
The complete runbook includes exact error signatures, event-level proof, inherited and explicit route behavior, privacy boundaries, alternate failure branches, twelve acceptance gates, and an incident bundle.
Read the complete guide