Claude Code. Any Model. The most powerful AI coding agent now speaks every language.
feat(mcp)!: stop sending notifications/claude/channel
Jack Rudenko753a3c3e87ba
diff --git a/packages/cli/src/mcp-server.ts b/packages/cli/src/mcp-server.tsindex 3cdf9995..ce91b4a1 100644--- a/packages/cli/src/mcp-server.ts+++ b/packages/cli/src/mcp-server.ts@@ -4,7 +4,7 @@* Claudish MCP Server** Exposes all claudish models (OpenRouter, Kimi, GLM, Qwen, MiniMax, Gemini, OpenAI,- * local models, etc.) and channel sessions as MCP tools for Claude Code.+ * local models, etc.) and sessions as MCP tools for Claude Code.* Routes through the same proxy engine as the CLI — same auto-routing, fallback chains,* custom routing rules, and provider transports.*@@ -21,10 +21,8 @@ import { CallToolRequestSchema, ListToolsRequestSchema } from "@modelcontextprotimport { config } from "dotenv";import { searchCatalogModels } from "./adapters/model-catalog.js";import { prehydrateCredentialsForSpawn } from "./auth/credentials/prehydrate.js";-import { installWireTap, watchNotificationResult, wrapStateChange } from "./channel/diagnostics.js";import { SessionManager } from "./channel/index.js";import { parentSessionForCall } from "./channel/parent-session.js";-import type { ChannelEventType } from "./channel/types.js";import { isSubscriptionProvider } from "./handlers/shared/remote-provider-types.js";import {type HeartbeatHandle,@@ -122,39 +120,32 @@ const CLAUDISH_CACHE_DIR = join(homedir(), ".claudish");const OPENROUTER_MODELS_CACHE_PATH = join(CLAUDISH_CACHE_DIR, "openrouter-models.json");const CACHE_MAX_AGE_DAYS = 2;-/** Instructions added to Claude's system prompt when channel mode is active. */-const INSTRUCTIONS = `Claudish MCP server provides access to external AI models (OpenRouter, Ollama, LM Studio, etc.) for coding tasks.+/** The server's `instructions`, which the host adds to its system prompt. */+const INSTRUCTIONS = `Claudish MCP server gives you access to external AI models (OpenRouter, Ollama, LM Studio and every claudish provider), and to Claude models on the user's own Claude Code login.-## Channel Mode — External Model Sessions+## Which tool-When channel mode is active, you receive <channel source="claudish" ...> notifications about running external model sessions. Each session is an interactive Claude Code running the model in its own headless terminal pane.+- run_prompt: one quick prompt to one model's API. One request, one answer; no tools, no files, no session.+- create_session: one model doing a task as a full Claude Code session (tools, files, several turns).+- team: several models on the same task, each a full Claude Code session, with optional blind judging.-### Events+## Sessions-- running: The session's turn is running (its prompt was accepted). Note the session_id for future calls.-- tool_executing: The model is using a tool (Read, Write, Bash, etc.). Carries tool and tool_count.-- input_required (event waiting_for_input): The session waits for send_input: an interactive session between turns, or a session stopped on a question (meta activity "AskUserQuestion"). A send declines a question and becomes the next prompt.-- input_required (event awaiting_permission): A permission or plan-approval dialog is open. A send declines it and becomes the next prompt.-- completed: The session finished successfully. Call get_output to retrieve the full output.-- failed: The session failed (boot, prompt not accepted or not read, child exited, API error, pane lost). Call get_diagnostics for the cause.-- timeout: The session hit its timeout_seconds; the pane was closed. Call get_diagnostics to see how far it got.-- cancelled: The session was cancelled via cancel_session.+Each session is an interactive Claude Code running the model in its own headless terminal pane. Nothing is pushed to you while it runs: poll.-### Workflow+1. create_session returns {session_id, state: "STARTING"} at once while Claude Code boots.+2. Poll list_sessions (include_completed: true), about once every 10 seconds. Each row has state, activity, idle_seconds and accounting.+ - STARTING: booting, or the first prompt is not yet accepted.+ - RUNNING: a turn is in progress (activity names the tool in use).+ - AWAITING_INPUT: waiting for send_input — an interactive session between turns, or a turn stopped on a question (activity "AskUserQuestion"). A send declines a question and becomes the next prompt.+ - AWAITING_PERMISSION: a permission or plan-approval dialog is open. A send declines it and becomes the next prompt.+ - COMPLETED, FAILED, TIMEOUT, CANCELLED, EMPTY: finished.+3. send_input may be called at any time; it is queued until the session is idle.+4. COMPLETED: call get_output for the answer.+5. FAILED or TIMEOUT — or COMPLETED with an empty or surprising answer — call get_diagnostics. It returns the child's final screen, the upstream error bodies, the recent state records, the resolved model chain and the paths to the full records. It needs no re-run and no debug flag.+6. capture_session reads a session's current screen; cancel_session stops it. Nothing stops a session for being idle: a child inside a long build is silent and working.-1. Call create_session with a model and prompt to start an async session (state STARTING while Claude Code boots).-2. Watch for <channel> notifications — they arrive automatically.-3. send_input may be called at any time; it is queued until the session is idle. On input_required, call send_input with the answer.-4. On completed: call get_output to get the full response.-5. On failed or timeout — or on a completed session whose output is empty or- surprising — call get_diagnostics. It returns the child's final screen, the- upstream error bodies, the recent state records, the resolved model chain and the- paths to the full records. It needs no re-run and no debug flag.-6. Use list_sessions to see sessions: each row has state, activity, idle_seconds and accounting.-7. Use capture_session to read a session's current terminal screen.-8. Use cancel_session to stop a running session.--The session_id in the channel tag's meta attributes is the key for all tool calls.`;+The session_id from create_session is the key for every session tool.`;// ─── Types ───────────────────────────────────────────────────────────────────@@ -852,37 +843,7 @@ const sanitize = sanitizeForReport;// ─── Tool Definitions ────────────────────────────────────────────────────────-/**- * Pushes a live progress frame to the client, or is a no-op when the channel- * capability is not enabled. `team` uses this to report per-model token/cost- * stats while a multi-minute run is still in flight.- *- * Channels are the ONLY measured-working push mechanism FOR RENDERING:- * `notifications/progress` renders nowhere on Claude Code 2.1.220 — verified empty- * in both the agent's context and the interactive terminal UI. See- * ai-docs/sessions/dev-arch-20260729-171308-1dad34b5/capability-findings.md.- *- * That says nothing about KEEPALIVE, which is the opposite result: a progress frame- * resets the client's MCP idle timer and a channel frame does not (measured- * 2026-08-14 on 2.1.231 — a real `team` run died at exactly 1800s while emitting- * channel frames throughout). See- * ai-docs/sessions/mcp-progress-idle-timeout-20260814-000000-b7e21f4a/findings.md.- * The two mechanisms are complementary and both stay: channel is the visible- * surface, `notifications/progress` (mcp/progress-heartbeat.ts) is the keepalive.- */-type ChannelNotifier = (params: {- content: string;- sessionId: string;- event: string;- model: string;- elapsedSeconds: number;- createdAt: string;-}) => void;--function defineTools(- sessionManager: SessionManager,- notifyChannel: ChannelNotifier-): ToolDefinition[] {+function defineTools(sessionManager: SessionManager): ToolDefinition[] {const tools: ToolDefinition[] = [];// ── Low-Level Tools ──────────────────────────────────────────────────@@ -1609,9 +1570,8 @@ function defineTools(},group: "agentic",// The tool the whole keepalive exists for: a real run was aborted at exactly- // 1800s of channel-frame-only silence. NOT gated on `channelEnabled` — that- // failure happened in a session with channels unregistered, so gating the- // keepalive behind the channel group would reproduce the bug exactly.+ // 1800s with no `notifications/progress` on the wire. A progress frame is the+ // only notification measured to reset the client's MCP idle timer.//// Still needed even though `run` no longer blocks on answers: `run-and-judge` is// the pipeline mode and holds the call open for the whole run, and `run` itself@@ -1654,36 +1614,15 @@ function defineTools(const resolved = validateSessionPath(path);- // Live per-model token/cost progress for the run modes. The session- // basename is a stable, human-recognisable id for the channel frames.- const teamSessionId = resolved.split("/").filter(Boolean).pop() ?? "team";- const teamCreatedAt = new Date().toISOString();const runOpts = {requirePattern,minOutputBytes,claudeFlags: childFlags,- onProgress: (u: {- rendered: string;- phase: "running" | "settled";- allFailed: boolean;- }) => {- // Two mechanisms, deliberately both: the channel frame below is what- // the agent and the human SEE, and this one is what keeps the client- // from aborting the call. A state change is real news, so refresh the- // idle timer at that moment as well as on the periodic tick. Short- // phase string only — the rendered grid stays on the channel path.+ onProgress: (u: { phase: "running" | "settled" }) => {+ // A state change is real news, so refresh the client's idle timer at+ // that moment as well as on the periodic tick. Short phase string only:+ // the run's progress is read with `team(mode:"status")`, never pushed.ctx.reportProgress(`team: ${u.phase}`);- notifyChannel({- content: u.rendered,- sessionId: teamSessionId,- // A settled run must emit a TERMINAL event. Emitting "running"- // for the final frame leaves any status-tracking consumer- // believing the run never closed.- event: u.phase === "settled" ? (u.allFailed ? "failed" : "completed") : "running",- model: "team",- elapsedSeconds: (Date.now() - Date.parse(teamCreatedAt)) / 1000,- createdAt: teamCreatedAt,- });},};@@ -1710,10 +1649,10 @@ function defineTools(// The run's record in the sessions directory, for observers outside// this process (the magus claudish plugin monitor): the run's own- // status.json lives wherever the caller pointed `path`, and its only- // completion push is a channel frame Claude Code drops without- // `--channels`. Written BEFORE any pane exists; a throw here fails- // the call with nothing started.+ // status.json lives wherever the caller pointed `path`, and the record+ // is how an observer outside this process learns the run ended.+ // Written BEFORE any pane exists; a throw here fails the call with+ // nothing started.const parentClaudeSessionId = await parentSessionForCall({hostPid: sessionManager.hostPid,});@@ -1987,8 +1926,9 @@ function defineTools(name: "create_session",description:"Create a new claudish proxy session for an external model: an interactive Claude Code running " +- "it in its own headless magmux pane. Returns at once with state STARTING; the session produces " +- "channel notifications as it runs. With a prompt it is one-shot (it ends when that turn " ++ "it in its own headless magmux pane. Returns at once with state STARTING; nothing is " ++ "pushed while it runs: poll list_sessions (about once every 10 s) for its state, then " ++ "call get_output once it is COMPLETED. With a prompt it is one-shot (it ends when that turn " +"settles); without one it is interactive and waits for send_input. Needs magmux >= 0.14.0 " +"(darwin and linux).",inputSchema: {@@ -2149,7 +2089,8 @@ function defineTools(name: "get_output",description:"Get a session's answer prose (each settled turn's assistant text, from its transcript). " +- "Call after the 'completed' notification to get the full response.",+ "Call once list_sessions shows the session COMPLETED (or between turns of an " ++ "interactive session).",inputSchema: {type: "object",properties: {@@ -2244,7 +2185,7 @@ function defineTools(// Diagnostics as API.//- // Two channel sessions once ran 900 seconds of genuine billed work — 241+ // Two sessions once ran 900 seconds of genuine billed work — 241// assistant messages, ~94k output tokens, 150 tool calls — and reported// success with an empty output log. The cause was one line, written to// `~/.claudish/sessions/<id>/stderr.log` and left there:@@ -2360,152 +2301,34 @@ function resolveToolGroups(mode: string): Set<ToolGroup> {}}-// ─── SEP-1686 status mapping ─────────────────────────────────────────────────-// Maps Claudish's 9-value `event` enum to SEP-1686's 5-value `TaskStatus`.-// See: https://github.com/modelcontextprotocol/modelcontextprotocol/pull/1732-// Migration plan: ai-docs/sessions/.../sep-1686-migration-schema.md-type TaskStatus = "working" | "input_required" | "completed" | "failed" | "cancelled";--// A `Record` over `ChannelEventType`, so an event type without a projection is a-// compile error rather than a silent `"working"` from the fall-through below.-const TASK_STATUS_BY_EVENT: Record<ChannelEventType, TaskStatus> = {- starting: "working",- running: "working",- tool_executing: "working",- waiting_for_input: "input_required",- // A permission dialog is open and only the caller's `send_input` moves the- // session on. It had no key, so it fell through to "working": a SEP-1686- // consumer polled a session that was waiting on it as if it were busy.- awaiting_permission: "input_required",- completed: "completed",- failed: "failed",- cancelled: "cancelled",- // The key whose ABSENCE forced the channel wire to lie. `mapEventToTaskStatus`- // falls through to `?? "working"`, so a session killed by its own timeout was- // reported to a SEP-1686 consumer as still working — which is why the timeout- // path once emitted `"failed"` on the wire while the record said `"timeout"`.- // With this key present the timeout emits its own event; SEP-1686 has no- // `timeout` member, and `failed` is the only honest projection of it.- timeout: "failed",-};--const EVENT_TO_TASK_STATUS = new Map<string, TaskStatus>(Object.entries(TASK_STATUS_BY_EVENT));--// Exported ONLY so the regression guard can call it instead of grepping this-// file's source text. The behaviour worth guarding is the fall-through below: a-// key missing from EVENT_TO_TASK_STATUS does not throw, it silently yields-// "working", so a dead session (e.g. one killed by its own timeout) reports to a-// SEP-1686 consumer as still alive. Only a real call can catch that; a source-// regex stays green if the entry moves to a dead map, is shadowed, or ends up in-// a comment.-export function mapEventToTaskStatus(event: string): TaskStatus {- return EVENT_TO_TASK_STATUS.get(event) ?? "working";-}-// ─── Server Setup ────────────────────────────────────────────────────────────async function main() {const toolMode = (process.env.CLAUDISH_MCP_TOOLS || "all").toLowerCase();const enabledGroups = resolveToolGroups(toolMode);- // Create server with channel capability+ // Tools only, for every tool mode. No `experimental: {"claude/channel": {}}`: the+ // server pushes nothing; a host polls `list_sessions` and reads the records.const server = new Server(// `version.ts` is generated from packages/cli/package.json, so serverInfo names// the release the host is actually talking to.{ name: "claudish", version: VERSION },{- capabilities: {- ...(enabledGroups.has("session") ? { experimental: { "claude/channel": {} } } : {}),- tools: {},- },+ capabilities: { tools: {} },instructions: INSTRUCTIONS,});- // Create session manager with channel notification bridge.- // The bridge translates SessionManager events into MCP notifications.- // When CLAUDISH_CHANNEL_TRACE=1 is set, the diagnostics module wraps the- // callback to log each step of the producer→bridge→wire pipeline.- //- // Wire format aligns with SEP-1686 (Tasks) for forward-compat: each- // notification carries both Claudish-specific fields (`session_id`, `event`)- // and SEP-1686-shaped fields (`task_id`, `status`, `created_at`,- // `last_updated_at`). When Claude Code ships notifications/tasks/status- // receiver support, the migration is a method-name swap + payload restructure- // (no semantic changes).- // See: ai-docs/sessions/.../sep-1686-migration-schema.md- const sessionManager = new SessionManager({- onStateChange: wrapStateChange((sessionId, event) => {- // `timeout` is here because it used to arrive AS `failed` — the wire had- // no timeout event until EVENT_TO_TASK_STATUS learned to project one — and- // dropping the hint when the event split in two would have been a silent- // regression. Both are terminal failures worth reporting, and both are now- // explainable without a re-run via get_diagnostics.- const notificationContent =- event.type === "failed" || event.type === "timeout"- ? `${event.content}\n\nCall get_diagnostics with session_id: "${sessionId}" for the final screen, the upstream error bodies and the transcript path. To report it, use the report_error tool with error_type: "provider_failure" and model: "${event.model}".`- : event.content;- const result = server.notification({- method: "notifications/claude/channel",- params: {- content: notificationContent,- meta: {- session_id: sessionId,- event: event.type,- model: event.model,- elapsed_seconds: String(event.elapsedSeconds),- // SEP-1686 forward-compat fields (additive, do not break consumers- // of the existing fields above):- task_id: sessionId,- status: mapEventToTaskStatus(event.type),- created_at: event.createdAt,- last_updated_at: new Date().toISOString(),- ...event.extraMeta,- },- },- });- watchNotificationResult(result, { sessionId, eventType: event.type });- }),- });-- // Live-progress push for long-running tools. Gated on the channel capability- // actually being declared — emitting a channel frame the client never- // registered for is silently dropped, and `team` still writes status.txt- // regardless, so there is always a visible path.- const channelEnabled = enabledGroups.has("session");- const notifyChannel: ChannelNotifier = (p) => {- if (!channelEnabled) return;- try {- const result = server.notification({- method: "notifications/claude/channel",- params: {- content: p.content,- meta: {- // meta keys must match [a-zA-Z0-9_]+ — Claude Code silently drops- // keys containing hyphens or other characters.- session_id: p.sessionId,- event: p.event,- model: p.model,- elapsed_seconds: String(Math.round(p.elapsedSeconds)),- task_id: p.sessionId,- status: mapEventToTaskStatus(p.event),- created_at: p.createdAt,- last_updated_at: new Date().toISOString(),- },- },- });- watchNotificationResult(result, { sessionId: p.sessionId, eventType: p.event });- } catch {- // Progress reporting must never fail the tool call it is describing.- }- };+ // Sessions are observed by polling (`list_sessions`, `get_output`) and through+ // their records on disk; nothing is pushed to the host.+ const sessionManager = new SessionManager();// Keepalive cadence for `heartbeat: true` tools. Resolved ONCE per server rather// than per call, so a long-lived session cannot change its behaviour mid-flight.const progressIntervalMs = resolveProgressIntervalMs();// Build tool registry- const allTools = defineTools(sessionManager, notifyChannel);+ const allTools = defineTools(sessionManager);const enabledTools = allTools.filter((t) => enabledGroups.has(t.group));const toolMap = new Map(enabledTools.map((t) => [t.name, t]));@@ -2528,7 +2351,7 @@ async function main() {// keepalive for as long as it runs: the client aborts a call that emits nothing// for its idle window (30 min on stdio by default — the exact 1800s a real// `team` run died at), and a progress frame is the ONLY notification measured to- // reset that timer. Channel frames do not.+ // reset that timer.//// `extra.sendNotification` is preferred over `server.notification` because it is// the request-scoped sender; on non-stdio transports it is what correlates the@@ -2581,11 +2404,8 @@ async function main() {}});- // Connect via stdio transport. installWireTap() is a no-op unless- // CLAUDISH_CHANNEL_TRACE=1 is set; when active it mirrors outbound channel- // notification frames to stderr so we can confirm wire-level delivery.+ // Connect via stdio transport.const transport = new StdioServerTransport();- installWireTap();await server.connect(transport);installMcpShutdown(sessionManager);// The startup sweep: panes whose owner died (and whose watcher died too) are reaped@@ -2596,7 +2416,7 @@ async function main() {/*** Shutdown (D15 layer 3, §20.3 item 8): stdin EOF, transport close, SIGINT, SIGTERM and- * SIGHUP settle every team run and channel session CANCELLED — so their records end+ * SIGHUP settle every team run and session CANCELLED — so their records end* (`meta.json`, the closing wait line) — then reap every pane in parallel and exit.* Installed before any pane exists, so `startPaneSession`'s own default hooks never* replace these options. A SIGKILL of this process ends no record; the per-pane