[ranxianglei]

Fixes #957

omp is an opencode fork on Bun that keeps the full pi extension API (config hooks / chat hooks / fetch patch / tool registration) — verified by live probe in the issue thread. This aligns omp with the #964 (pi) / #941 (dsh) three-mode playbook.

1. Native mode (bili plugin install omp, no launcher)

• New src/agent/omp-native.ts mirroring pi-native.ts: at load it spawns its own proxy (or attaches to a healthy running one, parent-pid watchdog), rewrites model traffic to <proxy>/bili/<upstream> via the global fetch intercept, registers compress / decompress / acp_status as native client tools (plugin mode), and gates the plugin-mode header on tool readiness (round 1 rides wire mode). Sets BILLION_CONTEXT_NATIVE=omp synchronously at module evaluation so standalone extensions back off in-process. Opt-out: BILI_NATIVE_OMP=0.
• tsup.config.ts: new entry → self-contained dist/agent/omp-native.js (acp-kernel inlined). Thin dist/agent/omp.js stays for launcher mode.

2. Installer migration (src/plugin-install.ts)

• pluginInstall("omp") now targets dist/agent/omp-native.js. Pre-[native] omp: three-mode alignment (proxy / launcher / native plugin) + runtime-info #957 thin dist/agent/omp.js entries are recognized by the shared entry regex and replaced on re-install — old installs upgrade to native mode, exactly one bili owner per config, output notes (replaced …). Remove / status / launcher loader-check accept both forms; stale paths still report broken.

3. Runtime-info (#955 protocol, src/agent/pi.ts)

• omp has no per-request header event, so the "omp" flavor of createBiliPlugin reports via POST /__bili/plugin/runtime-info from before_provider_request — same shape as the other native hosts, deduped per model switch, soft-fail. Proxy-side window resolution unchanged (runtime-info table already exists).

Tests: 12 new — tests/omp-native.test.ts (bootstrap gate table incl. cross-host opt-out isolation, native-marker invariant, script-path resolution, installer fresh / idempotent / thin→native migration / dedupe / remove / status / launcher -e skip) and a real-HTTP fake-proxy end-to-end for the runtime-info flow in tests/plugin-agent.test.ts (silence before tool readiness, dedupe, model switch). Two pre-existing tests updated for the intentional migration semantics (#392 sub-test a, fake-HOME roundtrip). 1930 pass / 0 fail; typecheck + build green on the rebased head (current master ba6c9ad).

Docs: README en/zh — client table, Option-1 list, opt-out envs, runtime-info implementations list.

Not touched: wire protocol, config schema, persistence format, acp-kernel pin, src/update.ts.

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-19_omp-native-three-mode (5071a29)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-981

Each push to this PR publishes a new version under the pr-981 npm tag.

Option B — Download artifact

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

tar xzf billion-context-pr981.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Picked up — this is a completed-work PR, so I'm verifying rather than redoing: fetching 2026-09-19_omp-native-three-mode (f017055), reviewing the diff against master ba6c9ad, re-running typecheck/test/build on the head, and tracking CI (one windows check still in progress at pickup). Will report findings here shortly.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Verification complete — this PR is solid and ready for a human merge. I reviewed rather than redid (completed-work PR from the #957 session).

Branch & base

• Single commit f017055 directly on current master ba6c9ad — fresh base, zero drift.
• CI: all 8 checks green (ubuntu/windows × 22/24, real-codex e2e, build-artifact, version-guard); mergeable_state: clean.

Local pre-flight on the head (re-ran everything)

• npm run typecheck — green
• npm test — 1930 pass / 0 fail / 2 skipped (matches the PR claim exactly)
• npm run build — green; dist/agent/omp-native.js = 1.91 MB self-contained (acp-kernel inlined, same class as pi-native.js), thin dist/agent/omp.js still built for launcher mode
• Artifact smoke: importing the built dist/agent/omp-native.js sets BILLION_CONTEXT_NATIVE=omp synchronously at module evaluation (the coexistence-marker invariant holds in the shipped bundle, not just source) and exports an ExtensionFactory

Code review findings

• src/agent/omp-native.ts mirrors pi-native.ts faithfully but DRY-er than the PR body implies — it reuses the shared native-bootstrap.ts / native-intercept.ts plumbing instead of duplicating. Gate table verified: BILI_NATIVE_OMP=0 / BILLION_CONTEXT_PLUGIN=0 opt-outs, backoff when a bili launch already owns the proxy (BILLION_CONTEXT_PROXY / BILI_PROVIDER_REWRITES — so bili omp never double-spawns against a loadable installed entry), cross-host opt-out isolation (BILI_NATIVE_PI=0 doesn't affect omp), NODE_TEST_CONTEXT guard so tests never spawn a real proxy.
• Installer migration (src/plugin-install.ts): traced every edge — fresh / idempotent / thin→native replace-with-note / duplicate-collapse / stale-only→broken status / remove keeps foreign entries / launcher -e skip accepts both forms. OMP_ENTRY_RE now matches both omp.js and omp-native.js; "exactly one bili owner per config" holds. The intentional semantic change (any in-block entry matching our shape is now managed, even if stale) is covered by the updated Claude Opus 5 经中转被当 200K 用:models.dev 查找被 host 门槛旁路 + 面板硬编码 200K 兜底 + 变体后缀失配 + 自愈只降不升 #392 sub-test.
• Runtime-info (src/agent/pi.ts): omp reports via POST /__bili/plugin/runtime-info from before_provider_request, gated on toolsReady (round 1 rides wire mode, same as pi's header path), deduped per model switch through the shared reportRuntimeInfoOnChange, soft-fails when no proxy base — same shape as the other native hosts. Two-mode check (§6): plugin mode is the only mode omp native uses (carrier = the agent's own tool calls); proxy mode is untouched.
• Docs en/zh in sync (client table, Option-1 list, opt-out envs, runtime-info implementations list).
• Diff is clean: 8 files, every line on-purpose, no whitespace churn, no lock-file/generated churn. Nothing I would fix directly — no defects found.
• E2E regression suite not required here: no changes to server.ts, src/loop/, adapters, or preflight/compact paths.

⚠️ Sequencing note (hot-file contention with open PRs)
This PR shares files with three open PRs; whoever merges second must rebase + re-run the full gate:

• fix: register /acp-cache command in the pi/omp thin plugin (#800) #974 (/acp-cache command): touches src/agent/pi.ts (+56/-10) and tests/plugin-agent.test.ts — same regions feat: omp three-mode alignment (proxy / launcher / native plugin) + runtime-info #981 edits.
• feat: claude native posture — managed settings block + SessionStart hook #979 (claude native): touches src/plugin-install.ts (+207/-13), tests/plugin-agent.test.ts, and the same single-line entry array in tsup.config.ts — guaranteed textual conflict with feat: omp three-mode alignment (proxy / launcher / native plugin) + runtime-info #981's one-line entry addition.
• feat: unify the dsh install lane on dsh's own plugin channel (#966) #980 (dsh install lane): rewrites the dsh section of src/plugin-install.ts (-146) and touches src/update.ts (load-bearing — needs its own scrutiny either way).

Suggested order: merge whichever of #974/#979/#980 is furthest along first, then rebase #981 (trivial — its hunks are disjoint in intent; the tsup.config.ts array line is the only likely textual collision) and re-run typecheck+test+build. Merging remains human-only; this is a feat: with new capability surface, so it doesn't qualify for auto-merge regardless.

---

中文摘要：核验了 #957 的 omp 三模式对齐 PR（分支新、CI 全绿，本地重跑 typecheck/1930 测试/build 全过，产物冒烟确认原生标记同步生效）；代码与 pi/dsh 既有模式一致、迁移语义完整，未发现需修复的问题——可以合并，但注意与 #974/#979/#980 存在热文件竞争（尤其 tsup.config.ts 入口行必撞），建议先合其中一个再 rebase 本 PR。