feat: add Trae CLI client support (bili trae)

[ranxianglei]

Adds a bili trae launcher for Trae CLI (ByteDance's closed traecli agent). Trae CLI is a locally-signed-in agent that carries its own model quota (the issue's use case: a company only provides a traecli quota).

Why cert-MITM (not /bili/ rewrites)

Trae CLI is a stripped Go binary with no base-URL override — its model host is baked in, so the /bili/ path-rewrite mechanism (used by pi/omp/opencode) cannot reach it. The one route that works is cert-MITM: the Go runtime honors HTTPS_PROXY, so the launcher routes its model traffic through the bili proxy and injects the bili CA.

Verified by decompiling traecli v0.120.52 (see #655): honors HTTPS_PROXY (Go net/http.ProxyFromEnvironment); CA via the Go system trust store → SSL_CERT_FILE; model host overridable via TRAE_CLI_API_HOST; wire is the proprietary POST /api/ide/v2/llm_raw_chat (SSE, OpenAI-shaped body).

What changed

• src/client-config.ts — TraeConfig, resolveTraeHome (~/.trae), readTraeConfig (reads TRAE_CLI_API_HOST), TRAE_DEFAULT_MODEL_HOSTS (console.enterprise.trae.cn, www.trae.cn).
• src/launcher.ts — bili trae (proxy mode): buildTraeEnv sets HTTPS_PROXY + SSL_CERT_FILE (the combined CA bundle, since it replaces Go's system trust store — same shape as codex) + BILLION_CONTEXT_PROXY; whitelists the model host(s) for MITM (default hosts, or the TRAE_CLI_API_HOST override); resolves the traecli/trae-cli/trae binary. No budget/transport env (Trae manages its own context window internally).
• src/server.ts — recognizes the proprietary /api/ide/v2/llm_raw_chat wire as OpenAI (per decompilation the body is OpenAI-shaped). Safe: if the body is unparseable (non-JSON/encrypted), the existing gate falls through to verbatim passthrough — no compression, no breakage.
• src/discover.ts — trae hosts in MITM domain auto-discovery.
• src/cli.ts + README/AGENTS/CONFIGURATION (en + zh-CN) — help + docs.
• tests — launcher + discover coverage mirroring the qoder tests.

Residual risk (needs a real-machine capture)

Decompilation shows chatmodel.decryptMessage/decryptAK exist. The evidence (an OpenAI-response converter + a community bridge that reads the SSE directly) indicates the conversation body is plaintext, so decryptMessage most likely covers a signed-metadata field, not the whole body. If the body is encrypted, the proxy forwards it unchanged (no compression, no breakage) — so the worst case is "no value", not "broken". A real bili trae run against a company traecli confirms it.

The default MITM hosts are best-effort from the binary; if the company tenant is on a different host, set TRAE_CLI_API_HOST or --mitm-domain.

Verification

• npm run typecheck — clean
• npm test — 1242 pass (1 pre-existing sandbox failure in the codex/claude resolveClientCommand test, unrelated to this change — it hardcodes PATH=/usr/bin assuming codex is absent, but this sandbox has it; passes in CI)
• npm run build — success

Fixes #655

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-10_trae-client-support (6a28250)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-669

Each push to this PR publishes a new version under the pr-669 npm tag.

Option B — Download artifact

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

tar xzf billion-context-pr669.tgz
npm install -g package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Reviewing PR #669 (bili trae) now — status so far:

• Diff is clean and focused (12 files, +240/−17): launcher + client-config + server + discover + cli + docs + tests. No unrelated churn, no whitespace noise.
• Verified the key safety claim in src/server.ts: /api/ide/v2/llm_raw_chat is recognized as OpenAI, and when the body is unparseable (encrypted / non-JSON) parsed stays null → the compression gate (protocol && parsed && typeof parsed === "object") is skipped → prepared stays null → the request is forwarded verbatim (original bytes). Buffer bodies also skip the string-only compat.roles rewrite in forward(). So worst case is "no compression", never breakage — matches your description.
• Pre-flight: npm run typecheck clean, npm run build success, npm test 1242/1243 pass. The single failure is the pre-existing resolveClientCommand: codex/claude resolve to themselves sandbox failure you flagged (it hardcodes PATH assuming codex is absent; this environment has it, so it fails here but passes in CI) — unrelated to this change.

One small inconsistency I'm fixing directly: in discoverRoutes, the trae branch pushes modelApiHost verbatim, so a TRAE_CLI_API_HOST=host:port would whitelist host:port — but the MITM whitelist matches on the SNI hostname (port-less, isMitmHost compares h === d || h.endsWith("."+d)), so that entry would never match. Every other client branch uses url.hostname. Stripping the port for consistency.

Full review + the fix coming shortly.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review complete — full assessment of PR #669 (bili trae).

Verdict: ✅ Approve — ready to merge (merge stays human-only)

Diff hygiene

Clean and focused: 12 files, +240/−17. No unrelated churn, no whitespace/reformat noise, no lockfile drift. Every line serves the stated purpose.

Correctness — verified against the source, not just the description

1. /llm_raw_chat passthrough safety (the load-bearing claim). Confirmed in src/server.ts: the path is recognized as openai at protocol-detect; when the body is unparseable (encrypted / non-JSON) parsed stays null, so the compression gate (!opts.passthrough && hopMarker === undefined && protocol && parsed && typeof parsed === "object") is skipped, prepared stays null, and the request is forwarded verbatim via the original bytes. Buffer bodies also skip the string-only compat.roles rewrite in forward(). Worst case = "no compression", never breakage — exactly as described, and it's the same proven path Follow-up from #617: robust session identity + no-spurious-400 for header-less clients (e.g. ZCode) #619 uses for decode failures.
2. Launcher env shape. buildTraeEnv mirrors buildCodexEnv (Go binary → CA rides SSL_CERT_FILE, combined bundle since it replaces Go's trust store). Note: trae passes stripInheritedProxy(process.env) where codex passes raw process.env — that's more correct for trae (a user-set NO_PROXY can't bypass bili for the model host), not a bug.
3. No budget/transport env. Correct: collectModelWindows(config, "trae") returns {}; the ModelWindowScope type extension is precisely what makes collectModelWindows(config, base) type-valid for base === "trae". Matches "Trae manages its own context window internally."
4. MCP injection off. launcherInjectMcp returns false for trae (closed binary, no MCP surface). Correct.
5. Binary resolution. traecli → trae-cli → trae, falling back to bare traecli (fails loudly at spawn if absent). Reasonable.
6. CLI dispatch + docs. isLaunchClient picks up trae via LAUNCH_CLIENTS; help text + README/CONFIGURATION (en/zh-CN) all updated consistently.

Pre-flight (ran on this branch)

• npm run typecheck — clean
• npm run build — success
• npm test — 1242/1243 pass. The single failure is the pre-existing resolveClientCommand: codex/claude resolve to themselves sandbox failure you flagged (hardcodes PATH assuming codex is absent; this environment has it → resolves to a full path; passes in CI). Unrelated to this change.

One small inconsistency — fixed directly on the branch

Pushed commit 6965c22 on top of your head:

• Problem: discoverRoutes's trae branch pushed modelApiHost verbatim. With a TRAE_CLI_API_HOST=host:port, that whitelisted host:port — but isMitmHost matches on the port-less SNI hostname (h === d || h.endsWith("."+d)), so a host:port entry would never match and the host would silently fall through to a blind tunnel (no compression). Every other client branch reduces to url.hostname.
• Fix: strip the port (h.split(":", 2)[0]) so the entry actually matches, consistent with the rest of the file; added a test locking in host:port → host. Default hosts are unaffected (no port). Typecheck clean; launcher+discover tests green.

It was masked in practice by the server-side auto-discovery fallback (extractHttpsHosts already strips the port via url.hostname), so it wasn't a functional break — just dead weight in the launcher's explicit whitelist and an inconsistency worth closing.

Residual risk (unchanged — needs a real-machine capture)

Your decompilation-based assessment holds. The only outcome this PR produces on an encrypted body is "forward unchanged" (no breakage). A real bili trae run against a company traecli is the confirmation step; nothing here alters that.

---

中文摘要:修复了 discoverRoutes 里 Trae 分支把带端口的 modelApiHost(host:port)原样塞进 MITM 白名单、而 MITM 按无端口的 SNI 主机名匹配导致该条目永不命中(退化成盲隧道、失去压缩)的不一致——改为剥离端口(与其余客户端分支一致)并补了测试;核心的加密体透传安全性、env 形态、无预算 env、CLI/文档都已逐项核验通过,typecheck/build/test 全绿(仅 1 个与本改动无关的既有沙箱失败),可以合并。

[ranxianglei]

解决冲突

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

已解决冲突并推回 PR 分支。

冲突来源：自本 PR 分叉后 master 前进了约 40 个提交（含 acp-kernel 版本升级、Windows 启动/spawn 修复、turn-integrity、release v0.1.98–v0.1.102 等）。把 master 合入 PR 分支后，只有 src/launcher.ts 存在真实文本冲突，且集中在两处——client-config.js 的 import 行 和 re-export 块：master 新增了 resolveCodexHome，本 PR 新增了 trae 符号（resolveTraeHome、readTraeConfig、TRAE_DEFAULT_MODEL_HOSTS、type TraeConfig）。其余文件全部自动合并，无语义冲突。

解决方式：取两者并集——同时保留 master 的 resolveCodexHome 与本 PR 的全部 trae 符号（import 与 re-export 各一处，顺序对齐现有分组）。

合并后验证（在合并后的完整树上重跑，非仅冲突文件）：

• npm run typecheck：clean
• npm run build：成功
• npm test：1334 pass / 0 fail / 2 skipped。之前沙箱里那个 codex/claude resolveClientCommand 的预存失败这次也不再出现（现在 0 fail）。

推送：fast-forward 推到 PR 分支 2026-09-10_trae-client-support（6965c22..6a28250）。PR #669 现在 mergeable: true（内容冲突已清除）；mergeable_state: blocked 只是 CI / 分支保护在 gate「实际合并」这一步，属正常流程，合并仍由你手动触发（我不会代点）。

另外说明：未改动 version（合并后沿用 master 当前版本；本内容为功能 PR，不 bump 版本号）。

---

一句话：把 master 合入 PR 分支，解决了 src/launcher.ts 里 import/re-export 的符号冲突（取并集），合并树 typecheck / build / 1334 项测试全绿并已推回，PR 现已可合并，等你点合并按钮。