Native experience overhaul: reliability, streaming fidelity, transport (v0.5.0-next.0)

[justin-carper]

Summary

Overhaul making Cursor-as-provider in opencode feel, function, and perform as natively as possible. Targets the three pain areas identified: reliability, streaming fidelity, and model/feature parity. Ships as 0.5.0-next.0.

22 commits, +2 net test files. 338/338 unit tests green, typecheck clean, build emits DTS. Live E2E transport matrix 6/6 green (verified against Cursor API, cheapest model).

Reliability

• Typed error classification (error-classify.ts): name/status/code heuristics → agent-not-found | agent-busy | rate-limit | network | auth | config | unknown (never instanceof — sidecar errors cross a JSON boundary as plain objects). Sidecar now preserves status/code/isRetryable/helpUrl round-trip.
• Typed send recovery: agent-busy → single local.force resend; rate-limit/network → bounded backoff (same agent); auth/config → fail fast. Replaces prior retry-on-any-error.
• Self-heal gating: fresh-agent+full-replay only for the classified replayable set; auth/config rethrow.
• Stream watchdog (OPENCODE_CURSOR_STALL_MS, default 60s): pre-first-event stall → cancel + force-resend once; mid-stream stall → cancel wedged run + terminal error (no orphan runs).
• Idempotency keys on all send sites → safe retries via server-side dedupe.

Transport (native HTTP)

• Bun node:http2 breaks Cursor's connect-RPC streaming (upstream oven-sh/bun#31499, NOT_PLANNED, still failing). No released Bun fix.
• New transport resolution: http1 (Bun default, via Cursor.configure({local:{useHttp1ForAgent:true}})), http2-direct (Node default), sidecar (fallback). Configurable via transport option or OPENCODE_CURSOR_TRANSPORT. Legacy OPENCODE_CURSOR_SIDECAR=0 routes to http1 under Bun (was routing to broken http2-direct).
• Node sidecar is now fallback, not required — simpler default path, live-verified TTFT parity (http1 0.93× sidecar).

Streaming fidelity

• Tool-input streaming: partial-tool-call → tool-input-start/-delta/-end (opportunistic, SDK-dependent; edit stays diff-buffered). Live plan streaming for createPlan. Kill-switch OPENCODE_CURSOR_TOOL_INPUT_STREAM=0.
• Thinking duration + compaction count surfaced on finish providerMetadata.cursor.
• semSearch native adapter.
• Silent-replay usage aggregation: multi-message replay turns' token usage folded into the visible turn's finish (fixes undercount).

Model / feature parity

• SDK-authoritative model variants: prefer Cursor.models.list() variants (displayName/isDefault) over generated ones; deterministic slug with collision counter.
• autoReview option (classifier tool-gating; best-effort, not a security boundary).
• cursor_delegate multi-root workspaces (additionalCwds).

Hygiene

• Deps: @cursor/sdk ^1.0.24, @opencode-ai/plugin+sdk ^1.18.4. Node floor >=22.13 (SDK requirement — package.json previously claimed 22.0 while the SDK needed 22.13).

Test plan

• npm test → 338/338 (29 files)
• npm run typecheck → clean
• npm run build → ESM + DTS emitted
• E2E=1 npm run test:e2e → transport matrix 6/6 (tool-parity, resume, long-stream >64KB, recovery) against live Cursor API