feat: rebuild the Cursor adapter for OpenCode V2

.github/workflows/release.yml

@@ -39,6 +39,7 @@ jobs:
         shell: bash
         run: |
           set -euo pipefail
+          # Version bumps are reviewed separately; this workflow publishes package.json verbatim.
           version="$(node -p "require('./package.json').version")"
           if [[ ! "$version" =~ ^[0-9]+\.[0-9]+\.[0-9]+-beta\.[0-9]+$ ]]; then
             echo "Expected package version X.Y.Z-beta.N, got: $version" >&2

.gitignore

@@ -5,6 +5,7 @@ dist/
 
 # Local agent notes / environment-specific
 AGENTS.md
+.opencode/
 
 # Graphify knowledge graph output
 graphify-out/

CHANGELOG.md

@@ -1,80 +1,86 @@
 # Changelog
 
-All notable changes to this project are documented in this file.
-
-The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
-and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
-
-## [3.0.0-beta.2] - 2026-08-24
-
-### Added
-
-- OpenCode V2 `Plugin.define` entrypoint with native integration, catalog, and
-  `LanguageModelV3` registrations
-- Native Cursor AgentService transport for images, cancellation, parallel tool
-  calls, bounded H2 pooling, and live tool-result continuation
-- V2 loader, package, lifecycle, catalog, integration, and language-adapter tests
-- Preserved OpenCode V1 adapter at the `./v1` package export; npm `latest`
-  remains on the `2.x` release line while V2 publishes under npm `beta`
-
-### Changed
-
-- OpenCode V2 now owns transcript reconstruction, credential persistence,
-  permissions, tool execution, and compaction; the V2 path no longer uses the
-  localhost OpenAI-compatible proxy
-- Adapted from CasualDeveloper's V2 implementation in commit
-  `3af03f605243b58b33c2a9e1f9fa638280bca693`
-
-## [2.2.0] - 2026-08-05
-
-### Changed
-
-- Phase-aware stall budget for post-tool resumes: silent tool continuations recover in **90s** instead of 180s
-- H2 bridge workers pre-connect TLS/HTTP-2 at startup (faster first message after restart)
-- Tool-call debounce reduced from 500ms → 250ms
-- Title-gen model probe result persisted to disk (skip ~2.5s Zen probe after restart)
-
-### Added
-
-- `OPENCODE_CURSOR_POST_TOOL_PRE_OUTPUT_STALL_TIMEOUT_MS` (default 90s)
-- Regression coverage for silent post-tool stall recovery
-
-### Performance
-
-- First message (`gpt-5.4-nano`): **8.5s → 4.8s** (−44%)
-- First message (`cursor/default`): **6.1s → 3.7s** (−39%)
-
-## [2.1.0] - 2026-08-05
-
-### Changed
-
-- Phase-aware stall budgets: cold thinking gets **180s** so reasoning models are not discarded mid-thought
-- Recovery limits honor `MAX_STALL_RECOVERIES`
-- `proxyTelemetry` exported for observability and tests
-
-### Added
-
-- `OPENCODE_CURSOR_PRE_OUTPUT_STALL_TIMEOUT_MS` (default 180s)
-
-## [2.0.0] - 2026-08-04
-
-### Fixed
-
-- Root-cause restate loop: tool-result follow-ups no longer look like user interrupts that discard parked bridges
-- Frozen sessions: visible-text stalls surface an error in ~90s instead of holding a step for up to 12 minutes
-- Infinite recovery restarts capped by forward progress
-- Model discovery retries 3× on transient bridge failures
-
-### Removed
-
-- Accumulated loop detectors, loop-break notes, compaction re-framing, and auto-continue nudges (~1,450 lines)
-
-## [1.4.0] - 2026-08-03
-
-See [GitHub Releases](https://github.com/otto-assistant/opencode-cursor/releases) for earlier notes.
-
-[2.2.0]: https://github.com/otto-assistant/opencode-cursor/compare/v2.1.0...v2.2.0
-[3.0.0-beta.2]: https://github.com/otto-assistant/opencode-cursor/compare/v2.2.0...v3.0.0-beta.2
-[2.1.0]: https://github.com/otto-assistant/opencode-cursor/compare/v2.0.0...v2.1.0
-[2.0.0]: https://github.com/otto-assistant/opencode-cursor/compare/v1.4.0...v2.0.0
-[1.4.0]: https://github.com/otto-assistant/opencode-cursor/releases/tag/v1.4.0
+This channel targets OpenCode V2 exclusively.
+
+## Unreleased
+
+Live release validation remains pending for instruction precedence on Opus and
+image cases, opaque-reasoning replay, and continued work after automatic
+compaction, as recorded in
+the [acceptance report](docs/opencode-v2-release-acceptance.md).
+
+- Reconstruct authoritative host history through structured Cursor root blobs.
+- Observe host tool, permission, question, and session events while forwarding
+  results only from the next authoritative model invocation.
+- Use bounded, per-Run Node HTTP/2 workers with validated Connect completion.
+- Preserve late reasoning signatures in durable host metadata for fresh replay.
+- Report each host step's usage as Cursor's checkpoint occupancy, counted as
+  uncached input, plus that step's streamed output tokens, so OpenCode's context
+  meter and compaction follow the conversation size. Keep terminal input, output,
+  cache, and reasoning counters, which aggregate every model call in a Run, in
+  durable provider metadata with an explicit usage scope. Verify the contract
+  against the pinned host's automatic compaction behavior.
+- Preserve checkpoint-referenced opaque reasoning in durable host metadata and
+  restore its exact placement during replay. Reject ambiguous or changed new
+  roots rather than guessing their relationship to emitted output.
+- Match opaque roots to visible text and tools even when Cursor also streamed
+  thinking. Omit thinking-only roots and extra unmatched conversation roots
+  instead of failing a completed answer.
+- Treat Cursor `reasoning-effort` as a variant parameter so Gemini 3.8 Flash
+  High/Medium attach to one model, matching Gemini 3.7 Flash.
+- Treat any effort-valued parameter, including `reasoning_effort`, as variants
+  on one model. Keep Claude thinking on that same model: `none` is
+  non-thinking, and the effort variants stay thinking. Context and fast
+  listings stay separate.
+- Offer Cursor only its MCP tool family when tools are enabled, using the same
+  `x-cursor-agent-allowed-tools` list as the official SDK's `mcp` group, so models
+  see OpenCode's tools rather than Cursor's native ones. The list includes
+  `get_mcp_tools_tool_call`; allowing only `mcp_tool_call` stripped it and Cursor
+  failed at bootstrap. Map any native workspace tool execution requests
+  (`shellStreamArgs`, `shellArgs`, `grepArgs`, `readArgs`, `writeArgs`, `lsArgs`)
+  that still arrive to OpenCode host tools and format their native results.
+- Answer every Cursor exec frame. Fail unknown and unserved frames in band with
+  `throw` and `streamClose` instead of leaving Cursor waiting. Decline
+  Cursor-hosted web search, Exa search and fetch, web fetch, question, and
+  mode-switch queries so the Run continues with host tools, return an error for
+  plan creation, and fail the Run on VM setup and unrecognized queries.
+- Complete native shell streams with `start`, output, `exit`, the structured
+  `shellResult`, and `streamClose`. Without the result and close, Cursor kept the
+  turn pending after every native shell call, so resumed Runs stalled until the
+  watchdog fired.
+- Decline the read and write behind Cursor's native StrReplace edit so formatted
+  host reads are never written back as file content; the model edits through the
+  host `edit` tool instead.
+- Wait up to 180 seconds for a resumed Run's first output, as for fresh Runs.
+  Cursor needed 72 to 175 seconds at large contexts, so the previous 90-second
+  limit discarded continuations that would have answered.
+- Open a rebuilt Run that follows tool results by saying the results answer the
+  model's latest calls. The previous opener read as an interruption, and the
+  model re-checked state on every rebuilt Run instead of continuing.
+- Group the `minimal` effort tier into variants so models like Muse Spark 1.3
+  do not leak separate `-minimal` models.
+- Publish Cursor models under the stable catalog name, including newly available
+  Grok models, and send the exact usable wire ID, such as `cursor-grok-4.6-xhigh`,
+  as Cursor's model details id. New catalog entries appear without a plugin
+  code change. When OpenCode has loaded the models.dev Cursor provider, copy
+  that row's family, limits, modalities, and price. Do not copy lab caps, and
+  do not replace the live account catalog with the static list.
+- Add deterministic pinned-host acceptance and opt-in synthetic live probes.
+- Project host system instructions into ordered Cursor rules as well as history
+  roots. Add an explicit host contract that names those rules as the
+  application's system prompt, gives them priority over conflicting user and
+  tool text, and requires genuine tool outcomes. In live checks, Composer, Grok
+  4.7, and Auto followed a project instruction over a conflicting user request
+  in 18 of 18 trials, up from 6 of 9 with the earlier wording; Opus is untested.
+- Restrict this package to the native V2 plugin API and its runtime dependencies.
+- Migrate to released `@opencode/*` `2.0.6`, replacing the removed catalog API
+  with provider source registration and reload.
+- Refresh Cursor models on `credential.switched`, including sign-in and disconnect.
+- Preserve tool-error metadata in structured compaction and auxiliary requests.
+  Update host fixtures for directory loading and experimental session export APIs.
+
+## 3.0.0-beta.2
+
+- Native OpenCode V2 plugin, integration, catalog, and language-model APIs.
+- Account-discovered Cursor models with exact model/variant routing.
+- OpenCode-owned OAuth credentials, permissions, sessions, tools, and compaction.