[ranxianglei]

What & why

Restores the /acp status command for OpenCode 2.x native mode.

Root cause: /acp was implemented only against the legacy .server() plugin API (src/agent/opencode.ts). The V2 migration makes opencode load plugins as {id, setup} and never call .server(), so /acp was orphaned — it never registered, and native mode looked "inactive".

This registers /acp in the shared createOpencodeV2Setup factory so BOTH the launcher entry and the native entry surface it.

API contract (verified against @opencode/cli v2.0.3 source)

• Commands: ctx.command.transform(editor.add({ name, description, execute })); execute(input) receives input.sessionID.
• Panel render: ctx.session.synthetic({ sessionID, text }) inserts a synthetic message that does NOT trigger a model turn (V2 replacement for the legacy ignored no-reply prompt).

/acp fetches <proxy>/__bili/plugin/status?conversationID=<sid>&fallback=latest and renders the panel (idle/error notice when no proxy). Inert-safe on surfaces without ctx.command; TUI-only.

Scope check: the legacy .server() registered ONLY command["acp"] (src/agent/opencode.ts:98/:104) — no other local slash commands exist in src/agent/, so /acp is the complete set of orphaned commands.

Files

• src/agent/opencode-v2.ts: add command seam to V2PluginContext + V2CommandEditor types; register /acp after tool registration.
• tests/opencode-v2.test.ts: fake ctx gains command.transform + session.synthetic; fake proxy serves /status; 2 new tests (panel render, no-proxy notice).

Verification

• typecheck clean
• full suite: 1654 pass / 0 fail / 2 skip (win32-only)
• build OK; dist markers confirmed in both agent entries
• clean single-commit delta vs feat/pi-native-plugin (merge-base 1d4f38c, +133/-1, 2 files)

Refs #809. (#809 intentionally left open per owner — baseline reference.)

Note: this is the /acp delta only. If you'd rather bundle it with the #842 coexistence gate (#840) before it lands on the umbrella, say so and I'll close this and re-cut from the updated branch.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review of the /acp V2 restore (commit 11a82d7) — verified, ready to land.

What I did: pulled the branch (opencode-native-plugin @ 11a82d7, identical to origin/2026-09-15_opencode-native-plugin), cross-checked the diff against the server contract, and re-ran the full verification gauntlet locally. This is a review of completed work — I did not redo it.

Diff integrity — matches your description exactly: single commit on baseline 1d4f38c, +133/-1 across 2 files (src/agent/opencode-v2.ts +71/-1, tests/opencode-v2.test.ts +63).

Contract cross-check (all correct):

• Handler sends conversationId=…&fallback=latest (src/agent/opencode-v2.ts:231); server reads params.get("conversationId") (src/server.ts:733). Names match. (Minor: the issue body writes conversationID with a capital ID — cosmetic typo in the writeup; the code itself is correct.)
• Response shape {ok, panel, error} matches handlePluginStatus (src/plugin.ts:521-537): {ok:true, panel,…} on success, {ok:false, error} on unknown/no-session.
• Faithful port of the legacy /acp (src/agent/opencode.ts:108/:116): same status URL, same panel/idle/error branching, same fetchProxyVersion fallback. V2 renders via ctx.session.synthetic (non-model-turn), TUI-only, inert-safe through optional chaining when ctx.command is absent.
• console.error in the agent file is fine — src/agent/* runs in-host (client-side), not in the proxy process, and every sibling agent module already uses console.error (the "no console.error" rule targets server-side modules).
• Scope claim holds: /acp is the only local slash command registered under src/agent/; nothing else was orphaned by the V2 migration.

Verification gauntlet (re-ran, all green, matches your numbers):

• typecheck: clean
• npm test: 1654 pass / 0 fail / 2 skip (win32-only)
• build: OK
• dist markers: Show ACP status present in both dist/agent/opencode.js (×2 = V2 setup() + legacy server()) and dist/agent/opencode-native.js (×1 = V2 only) — as expected.

Landing path (your actual question):
The umbrella feat/pi-native-plugin has advanced past 1d4f38c to ca4b1ab (it already absorbed the earlier part of this branch via PR #824). I verified:

• Neither file you touched was modified by any commit in 1d4f38c..ca4b1ab (empty diff-stat).
• Cherry-picking 11a82d7 onto the current tip ca4b1ab applies with zero conflicts, still +133/-1.
• The [coexistence] cross-host dual-mode (launcher vs native) isolation audit — per-seam verification + opencode spurious-warning wart #840 coexistence gate touches disjoint files (src/agent/opencode-native.ts + 2 overlap-test files) — zero overlap with this delta.

Recommendation: land /acp standalone as its own small PR against the current feat/pi-native-plugin tip. Rationale: it's a self-contained single commit, rebases cleanly with no conflicts, has zero file overlap with the #842/#840 gate, and is independently green — so there's no technical coupling forcing a bundle. Keeping it separate means the coexistence gate can't block it (and vice versa), and each lands with a smaller, easier review. Bundling only pays off if there's a shared-file conflict or ordering dependency, and there isn't one here. (Final call is yours — happy to re-cut if you'd rather bundle.)

I have not pushed or opened anything — confirm standalone vs. bundle and I'll cut the PR accordingly.

中文摘要：独立复核了 /acp V2 恢复提交（11a82d7）——diff 与描述一致（+133/-1、2 文件），契约核对无误，typecheck / 测试（1654 通过 / 0 失败 / 2 跳过）/ 构建 / dist 标记全部复现通过；该单提交可零冲突落到当前 umbrella 尖端、与 #840 共存门文件零重叠，建议作为独立小 PR 合入而非捆绑。

[ranxianglei]

继续