ollama-adapter: stop leaking raw chain-of-thought into chat replies

[TheGreatAxios]

Summary

CL-6654: an agent's visible reply leaked the model's raw reasoning/thinking
text and internal tool identifiers into the timeline (screenshots on the
ticket). Root cause: Ollama's OpenAI-compatible endpoint never populates
the reasoning_content field @intx/inference's OpenAI provider checks
for reasoning — gpt-oss/qwen instead wrap their chain-of-thought inline in
<think>...</think> inside the ordinary content stream, so it rode
through every hop as genuine reply text: harness → inferenceDoneBlocks
(correctly typed "text", since that's what it actually was) →
chat-orchestrator.ts's TextPart → timeline.tsx's plain bubble.

The fix is contained entirely in packages/ollama-adapter, which already
wraps @intx/inference's built-in OpenAI adapter for this provider: it now
also wraps parseResponse/parseJSONResponse to reclassify a <think>
span's tokens as inference.thinking.delta before anything downstream
ever sees them. Once correctly typed:

• The vendored harness (vendor/intx/hub-sessions/event-collector.ts)
  already turns a type: "thinking" content block into a "reasoning"
  part, and packages/agent-events's inferenceDoneBlocks already drops
  it from the finalized reply (see its existing "drops block kinds with no
  chat-part equivalent" test) — no visible chain-of-thought disclosure by
  default, per the product call on this ticket.
• The existing live TurnActivityStrip "Thinking…" indicator
  (packages/chat-ui/src/turn-activity.tsx) starts correctly firing for
  Ollama models for the first time — it already listens for
  inference.thinking.delta and previously never received one from this
  provider, so gpt-oss/qwen turns had no working-indicator at all.
• Tool calls the model narrates in its own reasoning ("I should call the
  Giphy tool...") are dropped along with the rest of the reasoning span,
  so no raw @corbits/... paths reach the transcript either.

This closes the leak without inventing a new UI surface — react-ui's
ReasoningPartView disclosure and tool-activity.ts's natural-language
tool rows already exist and are simply routed to correctly now.

Everything upstream (@intx/inference) is unmodified; the fix is a
wrapper in our own package, consistent with how it already wraps
buildRequest for request-body overrides.

Scope cut under timebox

A named, per-agent typing indicator ("Myra is typing…") and
connector-logo tool-activity rows were both raised as follow-on design
asks on this ticket. The generic "Thinking…" strip and text-only tool
rows already ship safely (no leak, no raw identifiers) as of this PR;
the identity/logo polish is filed as a follow-up rather than blocking this
fix.

Live browser verification (agent-browser against a running Ollama
model) was not completed in this pass — the fix is validated at the unit
level against the actual vendored harness's block-indexing behavior
(vendor/intx/hub-sessions) and @intx/inference's real event shapes.
Recommend a manual pass with a real gpt-oss/qwen model before merge.

Test plan

• packages/ollama-adapter/src/think-tags.test.ts — whole/split
  <think> spans reclassified, ordinary text and tool-call events
  pass through untouched, thinking events never collide with the
  text index the harness's block map keys on
• WORKBENCH_CHECK_SINCE=origin/main bun run test
• WORKBENCH_CHECK_SINCE=origin/main bun run typecheck
• bun run lint
• Manual: real Ollama gpt-oss/qwen session, confirm transcript shows
  only the polished reply, no <think> text, no raw tool paths