aiti-codebuddy-agent

❝ Tencent CodeBuddy sessions, ready to chat anywhere. ❞

Tencent CodeBuddy Code (the official CLI carved out of the WorkBuddy desktop app; npm package @tencent-ai/codebuddy-code, command cbc) session backend: plugs into the All Sessions multi-agent frontend — a cbc history row in the panel (codebuddy:<sid>) revives into a live session with one click.

(verified on 2.148.0, 2026-09-10; stream long connection from 2.150.0).

Mechanism (two process tiers, feature codebuddy.connectionMode)

• stream (default, verified on 2.150.0): single process, long connection, multi-turn — stdin continuation + control_request three channels (interrupt / set_model / set_permission_mode effective immediately while running) + can_use_tool forwarded to the frontend authorization card (respondPermission ✅)
• exec (compat tier): cbc -p <prompt> --output-format stream-json [--model M] [--permission-mode X] [--resume sid], one shot per turn; model/permission switches take effect next turn, respondPermission:false (print mode has no approval surface; --permission-mode pre-allows)
• stream-json output is isomorphic to Claude Code: system/init (session_id/model/tools/slash_commands) + stream_event (text_delta increments) + assistant + result (is_error/usage/modelUsage.contextWindow)
• ⚠ error turns also exit with code 0 — success/failure is judged only by the result frame
• Lazy persistence: the first turn's init frame captures the real sid → session_backend_ready rebinds to codebuddy:<uuid>
• History replay: the revival path goes through the All Sessions service outlet (session-manager services.sessions; the only reader of cbc storage is its lib/agents/codebuddy.ts — the stored jsonl is not isomorphic to claude's: top-level role/content + numeric millisecond timestamps + function_call/function_call_result/reasoning as separate row types, tool calls folded into structured tool rows → viewer's native tool cards (input/output sections); reasoning stays off the timeline but provides the real model)
• CLI resolution: feature codebuddy.cliPath → cbc/codebuddy on PATH → npx fallback (process.execPath + npm npx-cli.js, immune to systemd PATH pitfalls; first use pays a package-download delay)

Login and credentials (both tiers complete)

API-key tier (P0)

• iam.md official env read directly: CODEBUDDY_API_KEY / CODEBUDDY_AUTH_TOKEN (higher priority) / CODEBUDDY_BASE_URL (third-party OpenAI-compatible gateway, bypassing Tencent credits)
• Credential storage: extension configDir codebuddy-credentials.json (0600); the three credential features managed by extman (apiKey/authToken/baseUrl) take priority over the credentials file
• API: GET /api/extensions/aiti-codebuddy-agent/login/status (badge data source) · POST /login/key {apiKey?,authToken?,baseUrl?} (empty string = clear that field) · POST /login/clear

Browser OAuth tier (server-side PTY driven)

• POST /login/tui {site?}: via aiti-cmd services.runner, starts a cbc interactive TUI session in the background (process ownership on the server side; the sid is shared with aiti-cmd's frontend xterm rendering), feeds /login programmatically (optional site index auto-selects the site)
• GET /login/tui/status: waiting|success|failed|timeout|idle + authorization URL (OSC8 restoration + SSO prompt filtering, sticky); completion is judged by fingerprint changes in ~/.codebuddy (settings.json + local_storage/ mtime); 330s fallback timeout
• POST /login/tui/input {text}: forwards manual keystrokes such as site selection (the process stays in the backend); POST /login/tui/cancel terminates
• Implements no OAuth protocol, touches no credential content (controlled-login standard philosophy)

Capability surface (fail-loud, routed per channel)

followUp / queueManage / setModel / setPermissionMode / listModels / getUsage / contextUsage (result.modelUsage) / getMessages / listTools / getCommands ✅; stream mode setModel/setPermissionMode effective immediately while running + respondPermission ✅ (can_use_tool → frontend approve/deny); exec compat tier switches take effect next turn, respondPermission ❌; steer / file attachments ❌

Session recycle

GET/POST /api/extensions/aiti-codebuddy-agent/session/recycle — upgrades existing handles to the current extension code (after hot reload, old-code sessions rebuild), same as aiti-zcode-agent.

Bridge tool injection (Adapter C)

• MCP_TARGETS already includes codebuddy (~/.codebuddy/.mcp.json): node dev-repo/bridge/bin/aiti-mcp-register.mjs --targets codebuddy registers the aiti shim entry
• Spawn env: the governance whitelist passes the service process's AITI_BRIDGE_URL through; the feature codebuddy.bridgeUrl provides an explicit override (takes priority). With neither, cbc sessions simply lack aiti_* tools and normal chat is unaffected

Model catalog (dynamic, D13)

The real list from cbc --help is probed by the shared resolver probeCbcCatalog + a cache layer (<cacheDir>/cbc-model-catalog.json, regenerable) — the dynamic probe wins; the documented list degrades to a fallback seed (default/fast/balanced/primary/deep tiers + glm-5.3 / glm-5.2 / kimi-k3 / kimi-k2.6 / minimax-m3 / gemini-3.5-flash / gpt-5.6-{sol,terra,luna} / gpt-5.5 / gpt-5.4 / gpt-5.3-codex… newer cbc adds hy4-preview / glm-5.3-flash / kimi-k3-1 / kimi-k2.8-preview etc., 16 tiers), merged with the pi catalog for display.