[BUG] Session title generation fails silently since v1.3.3 — effort parameter leaks into small model call

[CTHua]

Bug Description

Session title generation has been silently failing since v1.3.3. All new sessions retain their default "New session - <timestamp>" title instead of getting an LLM-generated title.

The root cause: when the user's selected model has a variant that includes an effort parameter (e.g. anthropic/claude-opus-4-6 with variant max), that parameter leaks into the LLM.stream call for the title agent. The title agent resolves to a small model (claude-haiku-4-5-20251001) which does not support the effort parameter, causing a 400 error from the Anthropic API. The error is swallowed by Effect.ignore on the fork, so the failure is completely silent.

Reproduction

1. Set your model to anthropic/claude-opus-4-6 with variant max (or any variant that maps to output_config.effort)
2. Start a new session and send a message
3. Observe the session title remains "New session - <timestamp>"

Evidence from logs

INFO  service=llm providerID=anthropic modelID=claude-haiku-4-5-20251001
      sessionID=ses_xxx small=true agent=title mode=primary stream

ERROR service=llm
      requestBodyValues: {
        "model": "claude-haiku-4-5-20251001",
        "output_config": { "effort": "high" },   ← should not be here
        ...
      }
      responseBody: {
        "type": "error",
        "error": {
          "type": "invalid_request_error",
          "message": "This model does not support the effort parameter."
        }
      }

ERROR service=session.prompt
      error=No output generated. Check the stream for errors.
      failed to generate title

Impact data (from local DB)

Version	Sessions with title	Sessions without title	Success rate
1.3.9	0	4	0%
1.3.5	0	3	0%
1.3.3	0	6	0%
1.3.2	29	2	93.5%
1.3.0	40	3	93.0%
1.2.27	42	2	95.5%

Title generation drops to 0% success starting from v1.3.3 for users with an effort-bearing variant.

Root cause

In packages/opencode/src/session/prompt.ts, the ensureTitle function passes user: firstInfo to LLM.stream:

const result = await LLM.stream({
  agent: ag,
  user: firstInfo,   // ← carries variant from user's main model
  // ...
  model: mdl,        // ← this is the small model (haiku)
})

The user message's variant (e.g. max → effort: high) is applied to the API call even though the resolved model (claude-haiku-4-5) doesn't support it.

Suggested fix

Strip or ignore the user variant when calling the title agent's small model. For example:

• Pass user: { ...firstInfo, variant: undefined } in the title generation call
• Or have LLM.stream skip unsupported parameters when small: true
• Or catch the specific error and retry without the variant

Workaround

Override the title agent's model in config to one that isn't affected:

// opencode.json agent config
"title": {
  "model": "google/gemini-3-flash"
}

Environment

• opencode: v1.3.3 → v1.3.9 (all affected)
• Provider: anthropic (confirmed), likely affects any provider where variant maps to unsupported model params
• OS: macOS (arm64)

[rekram1-node]

can u share opencode debug config, I cant replicate.

If request is "small" which title gen is, it shouldnt be using the options

[anas-asghar4831]

Reproducing this on the web UI with claude-opus-4-6 + max variant. All sessions stuck at "New session - " with no title generation. No errors visible in the web UI — completely silent failure.

Environment:

• OpenCode version: latest (web)
• Model: anthropic/claude-opus-4-6
• Variant: max
• Platform: mobile browser (Android, Chrome)

Switching variant away from max likely works around it, but the effort parameter leak into the title agent's small model call should be guarded regardless.

[smrr723]

Confirming title generation failures, though my data shows a slightly different pattern.

Environment:

• OpenCode version: v1.4.1 (macOS desktop app)
• Model: gitlab/duo-chat-opus-4-6 (GitLab provider, proxies to Anthropic)
• Platform: macOS (darwin, arm64)

Impact data (from local DB at ~/.local/share/opencode/opencode.db):

Version	With title	Without title	Success rate
1.2.24	37	0	100%
1.2.25	14	2	87.5%
1.2.27	16	1	94.1%
1.3.0	3	0	100%
1.3.2	23	3	88.5%
1.3.3	25	4	86.2%
1.3.7	9	4	69.2%
1.3.9	15	0	100%
1.3.10	23	2	92.0%
1.3.13	22	0	100%
1.3.17	25	3	89.3%
1.4.0	4	10	28.6%
1.4.1	1	0	100%

Total: 353 with title, 29 without (92.4% overall).

The biggest drop is at v1.4.0 (28.6%). Earlier versions show sporadic failures but not the complete 0% reported by others — likely because my provider (GitLab) may handle the effort parameter differently than direct Anthropic.