feat: agent bridge daemon — drive Claude Code/Codex from XNet (0194, Phase 0) - #148
Merged
Conversation
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Serves the loopback endpoint XNet's chat panel already probes at :31416 (the 'Local bridge' connector tier, preference #1) — the missing 'thin shell' that makes Claude Code / Codex / any agent usable from XNet's UI, leveraging the user's own subscription (the daemon spawns the user's CLI; it never sees a token). @xnetjs/devkit: - ChatAgent port + cliChatAgent (spawns claude -p / codex exec, returns reply) + fakeChatAgent; flattenChat. ToS-safe BYO-agent model. - createBridgeServer: hardened loopback HTTP daemon serving GET /health (bridgeHealth, so the panel detects it) and POST /v1/chat/completions — OpenAI-compatible, streaming (SSE, which the panel's provider requests) and one-shot. Loopback-only bind, OPTIONS preflight, Origin allowlist (loopback + configured; never wildcard), Access-Control-Allow-Private-Network. Mirrors the MCP HTTP transport's hardening. @xnetjs/cli: - builds a cliChatAgent and starts the daemon. Makes the bridge tier work on the web deployment (browser → local daemon) as well as Electron. apps/electron: - agent-bridge-manager starts the daemon on boot (fire-and-forget), but only advertises it when the agent CLI is actually runnable (a --version probe), so the panel never shows an 'available' bridge that errors on first message. Stops on quit; status/start/stop IPC + preload (window.xnetAgentBridge). Tests: devkit 44 (chat-agent 6, bridge-server 7 incl real ephemeral server + SSE + origin/PNA gates), cli 35 (bridge health + injected-agent chat). devkit + cli typecheck, electron main tsc, eslint --max-warnings 0, prettier clean. Deferred (in the exploration): ACP transport + agent registry (any agent), mapping permissions to the approval flow, and wiring the MCP tool surface into the spawned agent so it edits the workspace (Phase 2-4). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Contributor
|
Preview removed for PR #148. |
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
| } | ||
|
|
||
| /** Start the bridge if the chosen agent CLI is installed; otherwise record why. */ | ||
| export async function startAgentBridge( |
crs48
added a commit
that referenced
this pull request
Jun 17, 2026
## What & why Exploration [0194](docs/explorations/0194_[_]_AGENT_BRIDGE_CLAUDE_CODE_CODEX_AND_ANY_AGENT_IN_XNET.md), Phase 1 (real). Phase 0 ([#148](#148)) made the bridge **chat** through the user's Claude Code/Codex CLI. This makes it **act on the workspace**: the spawned agent is handed XNet's MCP tool server, so it can search/read and create/update pages, databases, and canvases via the `xnet_*` tools. ## Changes **`@xnetjs/devkit`** — `agent-launch.ts`: - `buildAgentArgs(agent, { mcpConfigPath, allowedTools })` maps a known agent CLI to its headless argv. Claude Code gets `--mcp-config` + `--allowedTools "mcp__xnet__*"` (the documented way to auto-approve MCP tools in print mode; `acceptEdits` does not cover MCP). Codex uses `exec` (MCP via its global config). - `mcpConfigFor(serverSpec)` builds the `mcpServers` JSON. Pure + tested. **`@xnetjs/cli`** — `xnet bridge serve --mcp [--mcp-api-url]` writes a **self-referential** MCP config (`node <this-cli> mcp serve --api-url :31415`, so it resolves without the `xnet` binary on PATH) and launches the agent with workspace tools. **`apps/electron`** — `agent-bridge-manager` wires MCP opt-in via env (`XNET_BRIDGE_MCP` + `XNET_BRIDGE_MCP_CLI`; default chat-only) so a packaged app can enable workspace tools once the CLI entry path is known. **Safety:** `xnet_*` writes flow through the existing server-side `McpWriteGuardrail` — low/medium-risk auto-apply (creating pages/rows), high-risk/outward-facing require `confirm:true`. The user's own loopback agent; XNet never sees an auth token. ## Tests / validation - **devkit 50** (agent-launch 6: claude with/without MCP, custom allowedTools, codex, mcpConfigFor) + **CLI 36** (bridge MCP-args wiring), green under the repo's vitest projects. - `typecheck` (devkit + cli), electron main `tsc` (only the pre-existing unrelated `social-import-ipc.ts` error), `eslint --max-warnings 0`, `prettier --check`: clean. - Not in CI: end-to-end with a real `claude` + `xnet` CLI editing a live workspace. ## Deferred (in the exploration) ACP transport + agent registry ("any agent": Gemini/OpenCode-Kimi); surfacing the approval flow in the panel UI; auto-resolving the CLI path so Electron MCP is on by default; plugin authoring via `runAgentTask`. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
crs48
added a commit
that referenced
this pull request
Jun 17, 2026
## What & why Exploration [0194](docs/explorations/0194_[_]_AGENT_BRIDGE_CLAUDE_CODE_CODEX_AND_ANY_AGENT_IN_XNET.md), Phase 4 (code/plugins). The bridge lets an agent edit your **workspace** (PRs #148, #152). This adds the other axis the prompt asked for — **authoring/editing code & plugins** — by exposing the proven `@xnetjs/devkit` dev-loop as a command. `xnet code "<intent>"` isolates in a git worktree → runs your own coding agent (Claude Code/Codex/aider) → runs the validation gate (typecheck → lint → test → fallow) → **checkpoints on pass / rolls back on fail** → optionally opens a PR. ## Changes - **`@xnetjs/cli`** — new `code` command. Thin wiring over devkit (`runAgentTask` + `defaultXnetGate` + `cliAgentRunner` + `Git` + `openPullRequest`), with pure, tested helpers: - `resolveCodeRunConfig(prompt, options, repoRoot, now?)` — maps CLI options → dev-loop config (id/branch/base/worktree/gate/keepWorktree); clock injected for deterministic ids. - `summarizeAgentTaskResult(result)` — result → one-liner (passed / no-changes / gate-failed-at-X / failed). - Flags: `--agent`, `--id`, `--branch`, `--base`, `--worktree`, `--pr`, `--keep`, `--repo`. ## Tests / validation - **CLI 43** (code 7: config derivation, explicit overrides, keep-on-PR, all four summary states). The heavy worktree→gate→checkpoint loop is already covered by devkit's real-temp-git e2e. - `typecheck`, `eslint --max-warnings 0`, `prettier --check`: clean. No new deps (cli already depends on `@xnetjs/devkit`). - Not in CI: end-to-end with a live coding agent. ## Deferred Wiring "create/edit a plugin" into the in-app UI (e.g. the bridge daemon's `/run` endpoint over devkit `handleBridgeRun`) + combining with the plugin scaffolder; and the ACP transport / in-panel approvals from earlier phases. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
crs48
added a commit
that referenced
this pull request
Jun 18, 2026
…193/0194) (#201) The `changelog-section` required check was only added in #164, so the 0192/0193/0194 feature batch merged before it never got changelog entries. This backfills the 12 user-facing features that were missing, each dated to its merge day with its PR number and contributors: | PR | Entry | |----|-------| | #138 | A safer foundation for plugins | | #145 | One trust model across plugins and Labs | | #146 | Reliability you can see for managed hubs | | #148 | Drive your own coding agent from xNet | | #149 | AI that can act on your workspace | | #150 | Your Labs become AI tools | | #152 | Your agent can use your workspace | | #154 | AI edits, right inside the editor | | #155 | An agentic dev loop in your terminal | | #158 | Review AI edits before they apply | | #159 | Kick off agentic code tasks from xNet | | #162 | Plugins run on the Labs runtime | Skipped: PRs already covered by umbrella entries (#142 plugin ecosystem, #144 extensibility fabric, #147 automated changelog, #163 agent panel, #180 changelog gallery), internal-only changes (#139 schema authz — zero user-facing effect), and meta/test/docs PRs. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What & why
Exploration 0194, Phase 0. XNet's chat panel already probes for a "Local bridge" daemon at
http://127.0.0.1:31416(the top-preference connector tier) — but nothing served it. This adds that missing daemon, so the panel can drive the user's own Claude Code / Codex CLI, leveraging their subscription. The daemon spawns the user's installed CLI and never sees an auth token (the ToS-safe BYO-agent model).Changes
@xnetjs/devkitChatAgentport +cliChatAgent(spawnsclaude -p/codex exec, returns the reply) +fakeChatAgent;flattenChat.createBridgeServer— a hardened loopback HTTP daemon:GET /health→bridgeHealth()so the panel'sbridgetier detects it.POST /v1/chat/completions→ OpenAI-compatible, streaming (SSE, which the panel's provider requests) and one-shot, backed by aChatAgent.OPTIONSpreflight, Origin allowlist (loopback + configured; never wildcard),Access-Control-Allow-Private-Network— mirrors the MCP HTTP transport's hardening.Because the existing
bridgetier already maps toopenai-compatible, this works with zero panel changes — the facade is the "quick win" from the exploration.@xnetjs/clixnet bridge serve [--agent claude|codex] [--port] [--allow-origin] [--cwd]— builds acliChatAgentand starts the daemon. Makes the bridge work on the web deployment (browser → local daemon) as well as Electron.apps/electronagent-bridge-managerstarts the daemon on boot (fire-and-forget) but only advertises it when the agent CLI is actually runnable (a--versionprobe), so the panel never shows an "available" bridge that errors on first message. Stops on quit;status/start/stopIPC +window.xnetAgentBridgepreload.Tests / validation
typecheck(devkit + cli), electron maintsc(only a pre-existing unrelated error insocial-import-ipc.ts),eslint --max-warnings 0,prettier --check: clean.claudeinstalled. By design, no subscription token is read.Deferred (in the exploration)
ACP transport + agent registry ("any agent": Gemini/OpenCode-Kimi), mapping agent permission requests to XNet's approval flow, and wiring XNet's MCP tool surface into the spawned agent so it edits the workspace (pages/db/canvas) and authors plugins (Phases 1–4).
🤖 Generated with Claude Code