feat: opencode v1 — keep opencode-acp legacy sessions working under bili (#920)

[ranxianglei]

What

On OpenCode 1.x hosts, bili opencode now runs new sessions through the bili proxy while existing opencode-acp ("legacy") sessions keep working with their own in-process machinery — per the user-directed design in #920. Previously acp globally self-disabled on /bili/ baseURLs and its store (~/.local/share/opencode/storage/plugin/acp/<sessionID>.json) became unreachable: decompress on an old ref returned [Block … not found] and numbering restarted.

Bili-side absorption, zero opencode-acp changes:

• src/agent/acp-bridge.ts (new) — imports the installed opencode-acp package as a library (await import() around a temporarily unset BILLION_CONTEXT_PROXY, saved/restored in finally). Wraps every hook behind a per-session legacy gate = an acp store file exists for this session; non-legacy sessions are never touched, so acp's eager registry.getOrCreate adoption never fires. Its config hook is excluded entirely (that's where the global /bili/ self-disable lives). Tool executors are gated too (v1 tool registry is process-global/static — no per-request filtering hook exists). Any failure (package missing / import fails / version ≠ 1 / hooks shape mismatch) → null → exact pre-change behavior.
• src/agent/opencode.ts — V1 plugin merges the bridge hooks + acp's DCP tool map, registers /dcp alongside /acp, chains command handlers (acp first — its abort sentinel wins for legacy sessions), and adds a chat.headers hook stamping x-bili-plugin-bypass: 1 on legacy-session requests (verified the v1.14.46 binary routes plugin headers onto the real LLM request). V2 setup() untouched.
• src/server.ts — new bypass mode: x-bili-plugin-bypass: 1 → raw passthrough via the existing forward() ahead of JSON.parse / window resolution / session binding / injection / compression. Also flips the three wire-tool inject helpers from union to drop-same-named-then-inject: in proxy mode bili owns the compression tool names, so upstream sees exactly one definition per name (opencode ships acp's static DCP schemas in every body).
• src/launcher.ts — the temp-config clone strips opencode-acp entries (string / tuple / object forms; fork-safe matcher — my-opencode-acp-fork survives) so the host never loads acp armed; the first stripped spec rides along as BILI_OPENCODE_ACP_SPEC so the bridge imports exactly the copy the host would have loaded.

Both compression modes reasoned (§6)

• Bypass precedes the mode decision itself → neutral to both; it only ever fires for sessions the thin plugin has verified as legacy.
• The inject-drop change only affects proxy mode (the helpers are never called when pluginMode is set), so plugin-mode clients see no behavior change.

Graceful degradation (documented in README zh/en)

Package not found / import fails / not v1 → the plugin behaves exactly as before this change; legacy sessions degrade the way they did when opencode-acp self-disabled on /bili/ baseURLs. One once-per-process warning is logged.

Pre-flight results (rebased head 0ff133a, base = current master 0cfde7a)

• Rebased onto live master after ~35 PRs landed while this branch was open; two conflicts (launcher.ts opencode env block, server.ts import line) resolved as union of intent — master's [Suggestion] 建议 launcher 一并覆盖常见小写代理变量（http_proxy / https_proxy） #890 stripInheritedProxy env fix preserved alongside the opencode v1: keep opencode-acp legacy sessions working after switching to bili #920 strip-from-clone mechanism.
• npm run typecheck: clean
• npm test: 1751 pass, 0 fail, 2 skipped (pre-existing gated skips)
• npm run build: success; dist sanity checked (BILI_OPENCODE_ACP_SPEC in dist/agent/opencode.js, x-bili-plugin-bypass in dist/index.js)
• New tests: tests/acp-bridge.test.ts (probe order/authority, validation rejections, store-dir resolution matrix, env save/restore around acp init, full gating matrix, all degradation paths) · tests/fix-920-bypass-drop.test.ts (bypass byte-fidelity incl. unparseable bodies, proxy-mode same-name drop) · tests/launcher.test.ts (updated expectations + all-spec-forms strip)
• CI on the rebased head: all green — test (ubuntu-latest, 24) ✅ · test (windows-latest, 24) ✅ · real codex vs deterministic fake upstream (zero tokens) ✅

No config schema, persistence format, or cross-repo dependency changes. No version bump (content branch).

Closes #920

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Reviewing PR #923 (branch 2026-09-18_opencode-acp-legacy-bridge → master). Preliminary findings while I verify locally:

• Code review (src/agent/acp-bridge.ts, opencode.ts, server.ts, launcher.ts, util.ts): the design is sound. Legacy-gated hook wrapping, the config hook correctly excluded from wrapped (so the plugin's own config hook survives the Object.assign), env save/restore around the acp import, drop-same-named inject scoped to proxy mode, and full graceful degradation all check out. Both compression modes are reasoned correctly (§6).
• ⚠️ Blocker — stale base: the branch is 60 commits behind current master and GitHub reports mergeable_state: dirty — content conflicts in src/launcher.ts and src/server.ts (both hot files touched by concurrent PRs). It needs a rebase onto current master before merge.
• ⚠️ CI: the head commit has 0 check-runs (combined status pending) — CI appears not to have completed. The full E2E A/B also wasn't run from the author's env (no local sglang upstream); ci-e2e.yml dispatch is recommended.

Running local typecheck/test/build and a rebase dry-run now; will follow up with the full verdict.

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-18_opencode-acp-legacy-bridge (0ff133a)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-923

Each push to this PR publishes a new version under the pr-923 npm tag.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf billion-context-pr923.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Rebased onto current master (0cfde7a, head now 0ff133a) — ~35 PRs had landed while this branch was open. Two conflicts, both resolved as union of intent:

• src/launcher.ts opencode env block — kept master's [Suggestion] 建议 launcher 一并覆盖常见小写代理变量（http_proxy / https_proxy） #890 stripInheritedProxy(process.env) base-env fix alongside the opencode v1: keep opencode-acp legacy sessions working after switching to bili #920 strip-from-clone comment; behavior is the intersection of both (inherited lowercase proxy vars still stripped, acp entries still stripped from the clone).
• src/server.ts import line — merged both sides' additions (BILI_PLUGIN_BYPASS_HEADER + master's resolveOutputHeadroomCap / weak-overflow imports).

Evidence on the rebased head: typecheck clean · npm test 1751 pass / 0 fail / 2 skipped · build green · CI all green (test ubuntu-24 ✅, test windows-24 ✅, zero-token codex e2e ✅). PR body updated accordingly.

中文摘要：分支已 rebase 到最新 master（解决 launcher/server 两处冲突，均按双方意图合并，保留了 #890 的 env 剥离修复），rebase 后全量测试与 CI 全绿，可以合并。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 PR #923 review complete — independent verification done. Code is correct; the single red CI cell is an environmental flake, now re-running.

I re-ran the checks myself on the rebased head (0ff133a, base master 0cfde7a) rather than trusting earlier notes. Live state right now: PR open, not draft, mergeable=true, mergeable_state=blocked — blocked only by the still-running CI job, no conflict.

Local (Node v22.23.2):

• npm run typecheck → clean
• npm run build (tsup) → success
• npm test → 1751 pass / 0 fail / 2 skipped (the 2 skips are pre-existing gated tests)
• Bundle wiring confirmed in dist: BILI_OPENCODE_ACP_SPEC, x-bili-plugin-bypass, acp-bridge refs present; version unchanged at 0.1.118 (correct — content branch, no bump).

Code review (reasoned across both compression modes, §6): sound, no correctness issues. Highlights: acp-bridge.ts imports opencode-acp as a library and deliberately excludes the config hook from wrapped so the plugin's own config hook survives the Object.assign; every stateful hook gates on legacy-store existence before touching acp's handler; the proxy env is saved/restored around the import; full graceful degradation to null on any failure. In server.ts the three wire-tool inject helpers flip union→drop-same-named-then-inject, scoped to proxy mode only (they never run in plugin mode), so clients shipping their own ACP tool schemas see exactly one definition per name = bili's.

CI correction: the earlier "CI all green" note missed one cell. On head 0ff133a:

• ✅ PR Build Artifact · e2e-fake-codex · CI Windows codex overlay · test ubuntu-22 · test ubuntu-24 · test windows-24
• ❌ test (windows-latest, 22) — not a code regression. It's 2 assertions in tests/ca-combined.test.ts (#152 ensureRootCA writes combined-ca.pem… :36 and #152 combined bundle re-merges… :48, both assert … includes(mitmRoot)), and the same log shows [dump] SSE stream dump failed … ENOSPC: no space left on device. That Windows runner hit disk-full mid-run, so the CA-bundle write/read-back failed. This PR touches none of src/ca.ts / src/mitm.ts / tests/ca-combined.test.ts, and identical code passed on windows-24 / ubuntu-22 / ubuntu-24 → environmental flake.

I've already triggered a re-run of just that failed job (rerun-failed-jobs → HTTP 201 accepted); the CI run is in_progress again and should go green shortly.

Bottom line: no code changes needed (nothing to fix); I did not modify or merge anything. Once the re-running CI cell is green, the PR is ready for you to merge. Optional follow-up: dispatch ci-e2e.yml (hosted) for the full E2E A/B of the legacy-session path, which isn't exercised from this sandbox.

中文摘要：代码本身正确（本地 typecheck / build / 1751 通过、0 失败，全绿）；唯一红的 CI 是某个 Windows runner 磁盘写满（ENOSPC）导致未被本 PR 触及的 CA 证书测试断言失败——属环境抖动而非回归；我已触发该 job 重跑（现运行中），转绿后即可由你合并。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Re-run landed green — PR #923 is now fully verified. The re-run of test (windows-latest, 22) passed (fresh runner, no disk pressure), confirming the earlier failure was the ENOSPC environmental flake and not a code issue. The full CI run (35300590242) is now success — every cell green: version-guard · ubuntu-22 · ubuntu-24 · windows-22 · windows-24 (alongside the separate Build Artifact / e2e-fake-codex / codex-overlay runs, all already green).

Nothing else outstanding on my side: code reviewed across both compression modes, local typecheck/build/test clean, dist wiring confirmed, version untouched. The PR is ready for you to merge whenever — (reminder: merging stays human-only per repo rules; I have not merged or modified anything.) Optional: dispatch ci-e2e.yml (hosted) for the full E2E A/B of the legacy-session path.

中文摘要：重跑后 windows-22 通过，确认此前失败是磁盘写满（ENOSPC）的环境抖动而非代码问题；整条 CI 现已全绿，PR #923 验证完毕，可由你合并。