feat(#640,#653,#655): client support batch — codebuddy + qoder + trae

[ranxianglei]

Fixes #640, #653, #655

Supersedes PRs #641, #654, #669 — all three client-support branches merged into one here so the shared-array conflicts (LAUNCH_CLIENTS / BaseClientName / ModelWindowScope / imports / help text) are resolved exactly once instead of thrice.

What's in this batch

codebuddy (#640) — CODEBUDDY_BASE_URL /bili/ rewrite (claude-style), default https://tencent.sso.codebuddy.cn/v2, OpenAI chat-completions wire; budget aligned via CODEBUDDY_AUTO_COMPACT_WINDOW; real ~/.codebuddy untouched. Per-model urls in a codebuddy models.json still bypass the proxy (documented v1 limitation).

qoder (#653) — model endpoint is hardcoded https (no base-URL override), so cert-MITM via HTTPS_PROXY + NODE_EXTRA_CA_CERTS; static host map whitelisted (api2-v2.qoder.sh, api1.qoder.sh, api2.qoder.sh, api3.qoder.sh, CN gateway gateway.qoder.com.cn), replaced entirely by explicit QODER_MODEL_SERVER_HOST (qoder's own env > static-map order); forces QODER(_CN)_MODEL_TRANSPORT=http; budget via QODER_AUTOCOMPACT_WINDOW. CN-site detection: QODERCLI_SITE=cn > QODERCN_* env > only-~/.qoder-cn-exists > intl.

trae (#655) — closed Go binary, no base-URL override → cert-MITM via HTTPS_PROXY + combined bundle at SSL_CERT_FILE + BILLION_CONTEXT_PROXY; host from TRAE_CLI_API_HOST or defaults (console.enterprise.trae.cn, www.trae.cn), port stripped for SNI matching; inherited proxy vars stripped from the child env. Server side: /llm_raw_chat added to the OpenAI-wire sniffing (logUnrecognizedPath / protocol sniff / inferWireProtocol).

Docs unified (was inconsistent across the three PRs)

pr#641/#654 touched the English README only, pr#669 touched zh-CN too. This branch brings all four docs to parity:

• README.zh-CN.md: 方式 1 heading + launcher usage block now include codebuddy/qoder/trae
• CONFIGURATION.md + CONFIGURATION.zh-CN.md: CLI Reference table gains codebuddy + qoder rows (both languages)
• AGENTS.md module map lists all clients
• cosmetic normalization: pr#654/feat: add Trae CLI client support (bili trae) #669 had re-indented two existing help lines in src/cli.ts/README.md by +1 space (visible misalignment in rendered help) — restored, and the master-side malformed wrap in the Launcher paragraph fixed while we're there

Conflict resolution notes

Every conflict was a mechanical adjacent-line union (three clients appended next to each other in shared arrays/imports/if-chains); no semantic clashes. Resolutions verified lossless: git diff <pr-tip> HEAD minus-lines are exclusively master-side additions since 3a603b8 plus the codebuddy/qoder union edits; all exported symbols of the three PR tips present (TRAE_/resolveTraeHome/readTraeConfig/buildTraeEnv/TraeConfig, QODER_/qoderIsCnSite, Codebuddy* set); test files took ours + pure-insertion hunks, applied insertion totals match each PR's own stat (+245/+6 for qoder, +140/+6 for trae).

Verification (all on this branch, local)

• npm run typecheck — clean
• full node --import tsx --test tests/*.test.ts — failure set identical to current master (only the two pre-existing sandbox-env failures in tests/plugin-agent.test.ts: before_provider_headers stays silent when the manifest fetch keeps failing, pi extension is inert without a proxy; both fail on master too). New coverage runs green: ~7 trae launcher/discover tests + qoder/codebuddy suites.
• npm run build — clean, dist/index.js 2.87MB (self-contained)

Residual risk carried from #655 review: trae chatmodel may encrypt the request body (Go-binary internals unknown) — worst case is passthrough (no compression, no breakage). Real-machine packet capture is the follow-up; not blocking this merge.

---

Fix (post-review): windows-latest/node22 hang — the three new runLaunch tests used extensionless fake client bins; after master merged planClientSpawn, win32 routes extensionless targets through comspec so the stub spawn never matched and the test tail hung (fail 0, cancelled N). Same class as #679; applied the established .exe workaround to all three clients in 755c863. All six jobs now green.

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-11_client-batch-641-654-669 (755c863)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-693

Each push to this PR publishes a new version under the pr-693 npm tag.

Option B — Download artifact

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

tar xzf billion-context-pr693.tgz
npm install -g package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 On it — reviewing PR #693 (client-support batch: codebuddy #640 + qoder #653 + trae #655, superseding #641/#654/#669).

Head db066a9 on 2026-09-11_client-batch-641-654-669, 12 files (+1296/−17), 10 commits, base master@9aa72ac. GitHub reports mergeable_state: behind, so I'll verify whether master has moved past this branch's base and re-check the three-way merge before judging cleanliness.

Review plan (per AGENTS.md): (1) duplicate-screen vs the three superseded PRs + confirm the union is lossless; (2) diff-cleanliness pass on every shared-array/import/help-text conflict resolution; (3) typecheck + full test suite + build locally; (4) both-compression-modes reasoning for any wire-touching change (trae /llm_raw_chat OpenAI-wire sniff). Will follow up with findings.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review complete — LGTM, no blocking issues. Safe to merge.

This was a read-only review on a local checkout of head db066a9 (I did not push, merge, or modify master).

Duplicate screening

Not a duplicate — #693 is the intended consolidation of #641/#654/#669 (union of their content). Keep open; those three should be closed as superseded once this lands.

What I verified (with evidence)

1. Merge safety vs current master — GitHub shows mergeable_state: behind (branch point 1d09d5b predates master's 5 newer commits), but git merge-tree origin/master pr693 returns a clean tree, zero conflicts. The PR touches none of the files master's new commits changed (package.json acp-kernel 0.0.61→0.0.63, version 0.1.101→0.1.102, turn-integrity-#684 src), so the three-way merge takes master's values automatically. No rebase needed to merge.

2. Losslessness vs the three tips — scripted per-line presence check of every line each tip added (vs its own base): the only deltas are the shared-union lines (LAUNCH_CLIENTS, BaseClientName, ModelWindowScope, the base=== chain, discover.ts imports, CLI help), and in pr693 each is the correct superset containing all three clients (codebuddy | qoder | trae). The large per-client blocks are present verbatim: launcher.test.ts qoder +245 / trae +140 / codebuddy +324 lines → zero misses. Named exports (resolveTraeHome/readTraeConfig/buildTraeEnv/TraeConfig, QODER_*/qoderIsCnSite, Codebuddy*) all present.

3. Local pre-flight on the PR head:

• npm run typecheck — clean
• npm test — 1358 tests: 1356 pass / 2 skip / 0 fail (better than the stated baseline; the two plugin-agent sandbox failures you saw did not reproduce here)
• npm run build — clean, dist/index.js 2.74 MB self-contained

4. Diff cleanliness — normal vs whitespace-insensitive numstat are identical for every file except tests/discover.test.ts (one whitespace-only line). So effectively zero cosmetic churn; the "+1 space" help misalignment from pr#654/#669 is gone (rendered bili --help aligns correctly). All source diffs are additive/surgical; AGENTS.md module map updated (also fixes the previously-missing dsh).

5. Wire change / both compression modes — server.ts adds /llm_raw_chat to OpenAI-wire recognition in exactly 3 places (unrecognized-path log, protocol sniff, inferWireProtocol). Trae is a closed Go binary that can never carry x-bili-plugin, so this is proxy-mode-only by construction — no plugin-mode cross-hazard. The carried residual risk (encrypted body → passthrough, no breakage) is reasonable and already documented.

6. Docs parity — all four docs (README.md/.zh-CN.md, CONFIGURATION.md/.zh-CN.md) get the three clients at EN/zh parity. Every functional env var is wired into src (CODEBUDDY_BASE_URL, CODEBUDDY_AUTO_COMPACT_WINDOW, QODER_MODEL_SERVER_HOST, QODER_AUTOCOMPACT_WINDOW, TRAE_CLI_API_HOST, SSL_CERT_FILE, NODE_EXTRA_CA_CERTS, BILLION_CONTEXT_PROXY). Also confirmed discover.ts's push() filters undefined/non-https hosts, so a missing codebuddyBaseUrl can't poison the MITM whitelist.

Minor nits (non-blocking, report-only)

• tests/discover.test.ts: one whitespace-only line change.
• codebuddy's test fake-binary uses plain "fake-codebuddy" while 10 other fake-client consts use the win32 .exe conditional ([bug] Windows: a space in the client or -e path truncates the launch (spawn shell:true, unquoted) — and silently kills compression when only the command is quoted #679 convention). Functionally irrelevant here — the test injects its own spawnImpl matching the literal BILI_CLIENT_BIN string, so no real win32 spawn planning occurs and it passes on all platforms. Pure consistency nit.

Happy to fold either nit in if you want it, but neither warrants holding the merge.

---

中文摘要：本 PR 是把 #641/#654/#669 三个客户端支持合并成一个的整合分支；我逐行核对确认对三个原 PR 无损(仅共享数组/导入/help 行为三客户并集,且为正确超集),merge-tree 显示与当前 master 无冲突可干净合并,本地 typecheck 全绿、测试 1358 项 0 失败、构建自包含成功,diff 除一处测试空格外零无关改动,/llm_raw_chat 改动属纯 proxy 模式、无跨模式风险——可以合并。