[defer] fix(sse): keep Cursor prompt cache with a session-stable conversation id

[QuangBlue]

Why

Cursor felt slow and the dashboard showed no cache reads. The cause is in the router, not in Cursor:
CursorExecutor sent a fresh random AgentRunRequest.conversation_id on every request. Cursor routes a
run to the backend that holds that conversation's prompt cache by this id, so every agentic turn landed on
a random backend and paid a cold prefill.

Live A/B with a real Cursor account (synthetic ~57k-token session, 8 growing turns):

conversation_id	cache hits	provider TTFT
random per request (before)	0/7 (~22% across all runs)	~3.9 s, spikes to 17 s
stable per session (after)	6/7 (~87% across all runs)	~1.2 s

Cursor usage-events for a busy account (3 h window): only 5% of large requests were ≥80% cached.

What

• open-sse/executors/cursor/conversationId.ts (new): wire conversation id =
  explicit body.conversation_id → client session identity (Claude Code / Codex / OpenCode headers or
  metadata) → conversation fingerprint → random. Derived ids are sha256-hashed with the connection id,
  so raw client identifiers never reach Cursor.
• CursorSessionManager keeps its per-request key, so parallel subagents of one session never share an
  h2 tool-resume session.
• Real usage: decode turn_ended usage (input/output/cache_read/cache_write/reasoning) and report it
  instead of the local estimate; cached_tokens / cache_creation_tokens now reach usage_history.
  Tool-resumed runs report cumulative minus what earlier segments already reported.
• One log line per turn: [CURSOR] <model> turn: server_first_token=… provider_ttft=… cache_read=…
  (counts only, no content).

Tests

• node --import tsx/esm --test tests/unit/cursor-conversation-cache.test.ts → 16/16 pass.
• tests/unit/*cursor*.test.ts → pass except the known timing flake in cursor-agent-models.test.ts
  (sigkillFollowupMs; passes alone 13/13).
• Live tool round-trip: segment usage sums to Cursor's metered turn_ended input (176939).

Observability

grep '\[CURSOR\] .* turn:' app.log and usage_history.tokens_cache_read for provider=cursor.

⚠️ base-red inherited: #14963

[QuangBlue]

Hi @diegosouzapw, I merged the latest release/v3.8.51 to resolve the conflicts with #14737 (thanks @MikeTuev for restoring tool calling). Both PRs decode Cursor's TurnEndedUpdate usage, so I reconciled them:

• this branch now uses fix(cursor): restore tool calling by aligning the agent protocol with Cursor's schema #14737's CursorTurnUsage type and decodeTurnUsage, and I dropped the duplicate decoder I had added;
• one difference I'd like to flag: in live runs input already includes the cache reads (e.g. input=56201, cache_read=56192 on a ~57k-token prompt). Adding cache_read/cache_write on top would roughly double prompt_tokens on a cache hit, so prompt_tokens is now input and cached_tokens is cache_read. I updated the expectation in cursor-streaming.test.ts (12/8/4 → prompt_tokens 12) with a comment. @MikeTuev, if you've seen a case where input excludes the cache, please let me know and I'll adjust;
• a malformed usage body no longer throws out of the frame, so turn_ended still ends the turn (covered by an existing test in this PR);
• the ttft_breakdown decoder moved to open-sse/utils/cursorAgentProtobuf/ttft.ts to keep cursorAgentProtobuf.ts near its ceiling (+2, with a _rebaseline note).

All Cursor suites pass locally (557/557), and typecheck, check:cycles and check:file-size are clean for these files. Happy to change anything you'd prefer done differently.

[diegosouzapw]

Re-homed to release/v3.8.52: v3.8.51 entered its release freeze, so the branch now belongs to the release captain and development continues on the next cycle. Nothing is wrong with this PR — it just needed a live base. No action needed from you; CI will re-run against the new base.

[diegosouzapw]

Thanks @QuangBlue — merged into release/v3.8.52; it ships in the next release.