feat: structured agent frames + durable sessions for the AI bridge (0392) - #623
Merged
crs48 merged 8 commits intoJul 23, 2026
Merged
Conversation
…vity Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: xNet Test <test@xnet.dev>
Adds the ACP-aligned AgentFrame union and foldStreamJsonFrames, the frame-emitting counterpart of reduceStreamJsonLine. cliStreamingChatAgent gains streamTurnFrames (tool_call/tool_result/permission_request/cost/session frames) while streamTurn stays on the delta reducer so the OpenAI-compatible endpoint is byte-for-byte unchanged. Signed-off-by: xNet Test <test@xnet.dev>
Streams structured AgentFrames (session, delta, tool_call, tool_result, permission_request, cost, result) as SSE, with the same pairing-token, Host, and Origin discipline as /v1/chat/completions, which is unchanged. Framed agents forward their native frames; plain agents get synthesized delta+result frames. Session planning/recording is shared with the OpenAI endpoint. Signed-off-by: xNet Test <test@xnet.dev>
createBridgeSessionStore accepts a SessionPersistence; fileSessionPersistence writes the fingerprint to session-id map as JSON. The CLI wires it under the Claude bridge cwd so a restart continues --resume sessions instead of re-seeding every open conversation. In-memory remains the default. Signed-off-by: xNet Test <test@xnet.dev>
Adds fetchModelsDevCatalog/parseModelsDevCatalog with a vendored snapshot fallback so cloud-key/local model pickers stop drifting against hand-kept tables and never hang on a models.dev outage. OpenAICompatibleProvider now sends HTTP-Referer + X-Title app-attribution headers to OpenRouter only. Signed-off-by: xNet Test <test@xnet.dev>
Signed-off-by: xNet Test <test@xnet.dev>
Signed-off-by: xNet Test <test@xnet.dev>
Signed-off-by: xNet Test <test@xnet.dev>
Contributor
|
Preview removed for PR #623. |
crs48
deleted the
claude/0392-ai-harness-architectures-and-xnet-connectivity
branch
July 23, 2026 03:07
crs48
added a commit
that referenced
this pull request
Jul 29, 2026
…nt (0394 phase 2) (#656) Exploration 0394's deliberately-deferred phase 2, built with the design pass it asked for — and the unblocking of 0406's last implementation item. **The in-chat approval ceremony** (`ai-chat-ceremony.ts`): write tools ride the existing risk-tiered `AgentAuditRecorder` (0337). Reads bypass (no audit-node flood from searches); low-risk writes execute with an audit trail; **medium** parks on the one-time code — `APPROVE <code>` typed in chat rides the real nonce machinery (SHA-256 match, TTL, single-use); **high/critical refuse the chat path** structurally (no nonce exists) and need the deliberate in-app approval, stamped with the operator's DID, with the change reviewed before the button enables. Denial and expiry resolve the model's tool call honestly — the turn never hangs. 9 ceremony tests + 5 gating tests. **Doubly-gated write tools** (`ai-chat-write-tools.ts`): the tier must be `reliable` (`writeModeFor` = agentic) AND the operator must flip the new "Allow edits" toggle (default off). The badge escalates honestly: reads workspace → searches workspace → **edits with approval**. **The framed bridge client** (`ai-bridge-frames.ts`): the panel's bridge tier now speaks `/v1/agent/stream` — the endpoint #623 shipped and nothing consumed. Tool activity, per-turn cost, and permission asks arrive structured instead of flattened away. **Desktop MCP ceremony**: `agent-mcp-server.ts` now passes `agentAudit`, closing a real hole — a bridged agent could previously run `xnet_apply_page_markdown` with no gate but the system prompt. **Verified live on the desktop**: with writes armed, a chat turn through the bridged Claude Code created a real Page node in the desktop SQLite store (`Ceremony Test 0394`), and the panel showed "Last turn · $0.1581 · 495 output tokens" — cost frames only exist on the framed wire, so that line is proof of the new transport. With writes off, the agent still refuses (verified in the same session). Doc movement: 0394 → 9/10 impl, 3/7 val. **0406 → 23/23 impl, 9/10 val** (only the cold-open item remains, in flight separately). 🤖 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
1 participant
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.
Implements the backend protocol foundation from exploration 0392 — the "keystone" the doc says everything else hangs off. Scoped deliberately to the pieces that are fully unit-testable in-repo; the agent-integration and UI items ship as follow-ups (see below).
What landed (5/10 implementation, 4/7 validation)
AgentFramevocabulary + frame reducer (@xnetjs/devkit) — an ACP-aligned union (session,delta,tool_call,tool_result,permission_request,cost,result) andfoldStreamJsonFrames, the frame-emitting counterpart ofreduceStreamJsonLine. The Claude streaming agent gainsstreamTurnFrames; the delta-onlystreamTurnis untouched so the OpenAI endpoint is byte-for-byte unchanged.POST /v1/agent/stream— streams the frames as SSE with the same Host/Origin/pairing-token hardening as/v1/chat/completions(which is unchanged). Plain agents get synthesizeddelta+resultframes.createBridgeSessionStoreaccepts aSessionPersistence;fileSessionPersistencewrites the fingerprint→session-id map to disk, andxnet bridge serve --agent claudewires it under the bridge cwd so--resumesessions survive a daemon restart. In-memory stays the default.@xnetjs/plugins) —fetchModelsDevCatalogwith a vendored snapshot fallback (never hangs on an outage), plus OpenRouterHTTP-Referer/X-Titleapp-attribution headers on OpenRouter-bound requests only.Changeset: devkit minor, plugins minor, cli patch. All additive — no breaking API changes.
Verification
packages/devkit,packages/plugins/src/ai,packages/clibridge, electron bridge manager — 200 tests green.Deliberately deferred to follow-up PRs
These need a real vendor binary, the running React app, or are explicitly a separate PR in the exploration — they can't be responsibly built-and-merged here without integration verification:
codexAppServerChatAgent(JSON-RPC tocodex app-server) + its resume/interrupt validationgemini --experimental-acpagentpermission_request(and the "confirmed apply" / consent-ceiling validation that depends on it)generateWithTools+AiSurfaceServiceloop to emit the same framesbridge installbeyond launchd (marked "separate PR" in the doc)Exploration 0392 stays
[_]until the follow-ups land.🤖 Generated with Claude Code