Filter sidecar Haiku flows so session title generation stops leaking into the transcript

[Juliusolsson05]

Summary

Claude Code v2.1.x makes a small auxiliary /v1/messages call at session start to generate a 3-7 word session title via queryHaiku() (vendor utils/sessionTitle.ts). The proxy adapter currently treats it as a normal assistant turn — turn_started + text_delta + block_completed all fire on the SemanticChannel — and the renderer mints a ghost entry for the JSON-shaped output. Because Claude Code never writes auxiliary calls to the JSONL rollout, that ghost is never superseded, orphans after its 3 s TTL, and gets rendered as a phantom {"title": "..."} message at the bottom of the transcript.

The smoking gun, decoded straight from a captured proxy log:

POST /v1/messages?beta=true
event: message_start
data: {"type":"message_start","message":{"model":"claude-haiku-4-5-20251001","id":"msg_013xNUQa..."}}

— a Haiku sidecar inside an Opus session, exactly matching generateSessionTitle() in the vendor source.

The actual fix lives in claude-code-headless (submodule SHA bumped from 7984e29 → 35bb19c, branch fix/sidecar-haiku-leak, PR Juliusolsson05/claude-code-headless#3). The adapter now demotes a flow to the existing 'secondary' attribution state at message_start when the flow's model matches /haiku/i AND a caller-supplied getSessionModel() returns a model that does not. The rule reads "filter Haiku flows on non-Haiku sessions" — catches every known auxiliary Haiku call (title gen, compaction summaries, hook agents, teleport title-and-branch) without false-positiving Haiku-as-primary users.

This PR's cc-shell side is the wiring: pass-through of the new getSessionModel and sidecarModelPattern options through ClaudeCodeHeadlessOptions.proxy, and in claudeSession.ts an opt-in that returns a non-Haiku constant by default with a CC_SHELL_PRIMARY_MODEL env-var escape hatch for users running Haiku as their primary model.

Dependency

This PR's submodule pointer references 35bb19c on the headless repo. That commit lives on fix/sidecar-haiku-leak and is being reviewed in Juliusolsson05/claude-code-headless#3. The headless PR should land first so the submodule SHA is reachable on main when this PR merges (otherwise the submodule pointer dangles on a feature branch).

What's NOT in this PR

• Screen-header parsing of the actual primary model. Today getSessionModel() returns the constant 'claude-opus-4-7' because the model only matters for the /haiku/i pattern test, never as an authoritative model id. A TODO in claudeSession.ts marks where to drop in a real parser later.
• Bundling the .gitignore change as a separate commit. The .worktrees/ ignore rule is the second commit in the diff (9fa530b); it's small infra for the development workflow that produced this fix. Happy to split if you'd rather it be a standalone PR.

Test plan

• Headless package: npx tsx packages/claude-code-headless/src/testing/verify-sidecar-haiku.ts — 13/13 cases green
  • Haiku title-gen on Sonnet session → filtered (no turn_started, no text_delta, no block_completed, flow_ignored published)
  • Real Sonnet flow on Sonnet session → unaffected
  • Haiku flow on Haiku-as-primary session → unaffected (false-positive guard)
  • Bundle-replay (Haiku title flow → Opus real flow) → only the Opus flow surfaces
  • No getSessionModel callback → filter is inert (preserves pre-fix behaviour)
• tsc --noEmit for claudeSession.ts, ClaudeProxyAdapter.ts, ClaudeCodeHeadless.ts, and the new test — clean
• Manual: npm run dev against this branch on a fresh Opus/Sonnet session, confirm the {"title": "..."} ghost no longer appears at the bottom of the transcript
• Manual: same on a Haiku-as-primary session with CC_SHELL_PRIMARY_MODEL=claude-haiku-4-5-20251001 set, confirm Haiku conversation turns still render

🤖 Generated with Claude Code