BUG: All new sessions get default title "New session - []" — chat history is indiscernible

[jeonghun-jj-lee]

Summary

Every new session receives the static title "New session - [timestamp]" and it never updates to reflect the conversation content. The project chat history becomes a wall of identical titles, making it impossible to find or distinguish past sessions.

Context

• Extension: harmoniqs.amicode v0.2.1
• Engine: v1.18.10-amicode.4 (base: v1.18.10 of sst/opencode)
• OS: Darwin 25.5.0 arm64

Root Cause Analysis

Title generation in packages/opencode/src/session/prompt.ts fires exactly once, gated on step === 1 (line 1285). Multiple code paths silently prevent this from ever succeeding:

1. Silent-turn guard pre-increments step (line 1208). If the model's first response is reasoning-only (no visible text or tool calls — common with Claude Opus on Bedrock), the guard does step++; continue before the title trigger at line 1284 is reached. By the time we get there, step is already ≥1 and the post-increment makes step === 1 false.

2. Prose-question guard (line 1265) has the same effect — another step++; continue that races the title trigger.

3. Single-message guard inside title() (line 208). The function bails if userMessages.length !== 1, meaning that if the trigger fires late (after a second user message), it's permanently dead.

4. No retry. Title generation is attempted at most once. Effect.ignore swallows all errors — empty LLM responses, transient Bedrock failures, concurrency limits — with no recovery path.

Approach

Decouple title generation from the step counter and add bounded retries.

Approaches Considered

Approach	Pros	Cons
Fix step counter only — don't increment step in the guards	Minimal diff	Fragile: future guards could re-break it; doesn't address silent LLM failures
Decouple only — own boolean flag, fire once	Clean separation of concerns	Still one-shot: a single transient failure = permanent default title
Decouple + bounded retry (chosen)	Handles all failure modes; converges quickly	Slightly more code; up to 3 forked LLM calls per session (capped)

Scope

Two changes in packages/opencode/src/session/prompt.ts, no other files affected.

Acceptance Criteria

• Sessions receive a meaningful auto-generated title after the first user message
• Title generation succeeds even when silent-turn or prose-question guards fire before the main loop body
• A transient LLM failure on the first attempt does not permanently prevent title generation
• Title generation stops retrying after success (no wasted API calls)
• Maximum 3 title-generation attempts per session

Key Decisions

1. Replace step === 1 with a dedicated titleAttempts counter

Location: line ~1091 (initialization) and lines 1284–1291 (trigger)

// At initialization (alongside `let step = 0`)
let titleAttempts = 0

// At the trigger point (replacing `if (step === 1)`)
if (Session.isDefaultTitle(session.title) && titleAttempts < 3) {
    titleAttempts++
    yield* title({
        session,
        modelID: lastUser.model.modelID,
        providerID: lastUser.model.providerID,
        history: msgs,
    }).pipe(Effect.ignore, Effect.forkIn(scope))
}

The counter is checked every loop pass (including tool-call cycles), so retries fire immediately without waiting for a new user message. The isDefaultTitle check prevents wasted calls after success.

2. Relax the single-message guard in title()

Location: line 208

// Before:
if (input.history.filter(real).length !== 1) return

// After:
if (input.history.filter(real).length < 1) return

This allows retries to succeed even after the second user message arrives. Double-titling is already prevented by the isDefaultTitle check on line 202.

Constraints & Invariants

• Effect.ignore stays — title generation remains non-critical and silent on failure
• Model selection logic unchanged (getSmallModel → fallback to main model)
• Title generation remains forked (never blocks the main response)
• Maximum 3 attempts is a hard cap (not configurable)
• The <think> tag stripping and 100-char truncation logic are untouched

Prior Art

The upstream sst/opencode has the same step === 1 pattern and is likely affected by the same bug when used with reasoning models that produce silent turns. This fix is fork-local but could be upstreamed.