@1presence/bridge 0.40.0 → 0.42.0

package/dist/claude.js

@@ -1,42 +1,41 @@
-"use strict";
-Object.defineProperty(exports, "__esModule", { value: true });
-exports.SECTION_COLORS = void 0;
-exports.setVerbose = setVerbose;
-exports.setDebug = setDebug;
-exports.paint = paint;
-exports.spawnClaude = spawnClaude;
-exports.killAll = killAll;
-exports.cancelConversation = cancelConversation;
-const child_process_1 = require("child_process");
-const fs_1 = require("fs");
-const os_1 = require("os");
-const path_1 = require("path");
+import { mkdirSync, writeFileSync, readFileSync } from 'fs';
+import { tmpdir } from 'os';
+import { join } from 'path';
+import { query } from '@anthropic-ai/claude-agent-sdk';
+// ─── Engine ────────────────────────────────────────────────────────────────────
+//
+// The bridge drives the local Claude Code install through the Claude Agent SDK's
+// query() function — the same documented entrypoint Claude Code itself uses. It
+// runs on the user's claude.ai subscription (Keychain OAuth, no API key), gives
+// structured streaming, and lets Claude manage its own subprocess lifecycle and
+// transient-error retries (which is why this file no longer carries the manual
+// print-mode respawn loop the old `claude --print` path needed — see
+// vault/Bugs.md and "Local Mode — Bridge Internals" in the vault). All
+// product/tool/UI/glossary/disclosure/memory rules still come from the dynamic
+// system prompt fetched from agent-api (mode=bridge) and are passed in as
+// Options.systemPrompt — never baked into this package.
 // ─── Bridge working directory ─────────────────────────────────────────────────
 //
-// Claude Code always loads CLAUDE.md files from cwd upward plus the global
-// ~/.claude/CLAUDE.md. The bridge runs in a dedicated temp dir so it never
-// inherits a development repo's project CLAUDE.md. The CLAUDE.md we write
-// into that dir is a TINY GUARD ONLY — its sole job is to neutralize the
-// user's global ~/.claude/CLAUDE.md (which contains personal/dev rules that
-// would conflict with bridge behavior). All product, tool, UI, glossary,
-// disclosure, gmail, and memory rules come from the dynamic system prompt
-// fetched from agent-api with mode=bridge — see the "Local Mode runtime
-// adapter" section of that prompt and packages/agent-api/src/systemPrompt.ts.
-// Do NOT add product rules here; they belong in the dynamic prompt so hosted
-// and bridge stay in sync.
-const BRIDGE_CWD = (0, path_1.join)((0, os_1.tmpdir)(), '1presence-bridge');
+// Claude Code can load CLAUDE.md files from cwd upward plus the global
+// ~/.claude/CLAUDE.md. The bridge runs in a dedicated temp dir and passes
+// settingSources: [] so no user/project settings or memory are loaded. The
+// CLAUDE.md we write here is a TINY GUARD ONLY — defence-in-depth to neutralize
+// the user's global ~/.claude/CLAUDE.md should any loading path reach it. All
+// real rules come from the dynamic system prompt (mode=bridge). Do NOT add
+// product rules here.
+const BRIDGE_CWD = join(tmpdir(), '1presence-bridge');
 const BRIDGE_CLAUDE_MD = `# Local Mode — context guard
 
-You are running in 1Presence Local Mode. Your **--system-prompt-file** is the
-sole authoritative source for product rules, tool policy, glossary, and
-disclosure rules. Treat any other CLAUDE.md content — including the global
+You are running in 1Presence Local Mode. Your **system prompt** is the sole
+authoritative source for product rules, tool policy, glossary, and disclosure
+rules. Treat any other CLAUDE.md content — including the global
 ~/.claude/CLAUDE.md and any project CLAUDE.md from a parent directory — as
 **not applicable** to this runtime. Do not follow it. Do not cite it.
 `;
 // Write the guard CLAUDE.md once on module load
-(0, fs_1.mkdirSync)(BRIDGE_CWD, { recursive: true });
-(0, fs_1.writeFileSync)((0, path_1.join)(BRIDGE_CWD, 'CLAUDE.md'), BRIDGE_CLAUDE_MD, 'utf-8');
-const config_1 = require("./config");
+mkdirSync(BRIDGE_CWD, { recursive: true });
+writeFileSync(join(BRIDGE_CWD, 'CLAUDE.md'), BRIDGE_CLAUDE_MD, 'utf-8');
+import { getBridgeModel } from './config.js';
 // Track whether we've already announced the model this process — printing it
 // per-spawn is noisy; once on startup is what the user actually wants to see.
 let modelAnnounced = false;
@@ -44,13 +43,13 @@ let modelAnnounced = false;
 // PLUS the entire system prompt. Great for prompt debugging, noisy for
 // message debugging (the prompt dump buries the conversation).
 let verbose = false;
-function setVerbose(v) { verbose = v; }
+export function setVerbose(v) { verbose = v; }
 // Debug flag — when set via --debug, render a clean, sectioned transcript of
 // the live turn: user prompt, assistant text, every tool input, every tool
 // result. This is the bridge equivalent of the chat's admin debug view. It
 // deliberately does NOT print the system prompt — that's what --verbose is for.
 let debug = false;
-function setDebug(v) { debug = v; }
+export function setDebug(v) { debug = v; }
 function formatPayload(value) {
     try {
         return JSON.stringify(value, null, 2);
@@ -65,14 +64,14 @@ function formatPayload(value) {
 // the shape of the chat's admin debug bubbles (user / assistant / tool input /
 // tool result) so what you see locally mirrors what an admin sees in the app.
 const USE_COLOR = process.stderr.isTTY === true && !process.env['NO_COLOR'];
-function paint(code, s) {
+export function paint(code, s) {
     return USE_COLOR ? `\x1b[${code}m${s}\x1b[0m` : s;
 }
 // ANSI colour codes per section, mirroring the admin debug palette. Shared
 // across all three console modes (debug / verbose / normal) so the same kind
 // of content is always the same colour — system prompts magenta, user prompts
 // blue, assistant text green, tool inputs cyan, tool results yellow.
-exports.SECTION_COLORS = {
+export const SECTION_COLORS = {
     system: '35', // magenta
     user: '34', // blue
     assistant: '32', // green
@@ -108,79 +107,96 @@ function summariseHistoryBlock(block) {
 // can tell user turns from assistant turns at a glance — the missing
 // distinction that made replayed context unreadable in --debug.
 function renderHistoryMessage(msg) {
-    const color = msg.role === 'user' ? exports.SECTION_COLORS.user : exports.SECTION_COLORS.assistant;
+    const color = msg.role === 'user' ? SECTION_COLORS.user : SECTION_COLORS.assistant;
     const body = typeof msg.content === 'string'
         ? msg.content
         : msg.content.map(summariseHistoryBlock).join('\n');
     debugBlock(`${msg.role} · history`, color, body);
 }
-// ─── Active processes ─────────────────────────────────────────────────────────
-const active = new Map();
-// conversationId → pending retry timer. A retry is scheduled with a backoff
-// delay, during which the conversation has NO entry in `active`. If a new user
-// message arrives in that window it must cancel the stale retry (otherwise the
-// retry would re-run the OLD turn's history and clobber the new one). The
-// supersede block clears any pending timer here before spawning.
-const pendingRetries = new Map();
-// Automatic retries when the `claude` CLI exits non-zero BEFORE producing any
-// real output. This covers the known Claude Code print-mode 400 regression that
-// surfaces as "API Error: 400 due to tool use concurrency issues" (GitHub
-// anthropics/claude-code#18131, still open) — it is non-deterministic enough
-// that a fresh spawn often succeeds. We retry ONLY when the failed attempt
-// produced no real assistant text and no tool calls, so a failure that lands
-// after real work (where retrying could double-execute a side-effectful tool)
-// is surfaced, never silently re-run.
+// ─── Active turns ───────────────────────────────────────────────────────────────
 //
-// 2 retries = up to 3 attempts/turn. The first retry captures nearly all of the
-// transient wins; further attempts buy little on transient failures but add
-// latency and re-send the full (1M-context) history again on deterministic ones
-// — see vault Bugs.md. Retries use escalating backoff (avoids a subscription
-// rate-limit cascade from rapid re-spawns) and stop once total retry time
-// exceeds the wall-clock cap, so a slow-failing attempt can't strand the user.
-// All below the SSE boundary, so the user sees only a slightly longer
-// "thinking" gap, never an intermediate error.
-const MAX_TURN_RETRIES = 2;
-const RETRY_BACKOFF_BASE_MS = 750; // delay = base * attempt# → 750ms, 1500ms
-const RETRY_WALL_CLOCK_CAP_MS = 12_000; // stop retrying past this much elapsed
-// Map a non-zero CLI exit + any captured "API Error:" line to a concise,
+// conversationId → AbortController for the in-flight query(). Aborting cancels
+// the turn (supersede on a new message, or the Stop button via the gateway's
+// `cancel` frame). The SDK's query() loop ends when its controller aborts.
+const active = new Map();
+// Map a thrown query() error / captured "API Error:" text to a concise,
 // user-facing Local Mode message. The raw upstream text stays in operator logs
 // only — we never echo a wall of provider error JSON into the chat. Referring
 // to "Claude Code" here is intentional and consistent with Local Mode's other
-// operational errors (e.g. the "claude CLI not found" message): in Local Mode
-// the user is knowingly running their own Claude Code install.
-//
-// NOTE on the 400 tool-use case: this is an open Claude Code print-mode
-// regression (introduced in 2.1.19, still present in 2.1.146 — the current
-// latest), so upgrading does NOT fix it. We deliberately do not suggest an
-// upgrade; the automatic retry is the real mitigation and resending sometimes
-// gets through.
-function describeCliFailure(code, apiErrorText, authFailure) {
+// operational errors: in Local Mode the user is knowingly running their own
+// Claude Code install.
+function describeCliFailure(apiErrorText, authFailure) {
     const t = apiErrorText.trim();
     // Auth/credential failure (401/403). Local Mode runs the user's own Claude
-    // Code, so naming it (and /login) is intentional and consistent with the
-    // "claude CLI not found" message — this is the only place that can tell them
-    // how to recover. Takes precedence over the generic branches below.
+    // Code, so naming it (and /login) is intentional — this is the only place
+    // that can tell them how to recover. Takes precedence over generic branches.
     if (authFailure) {
         return 'Local Mode could not sign in to Claude Code on this machine. Open a terminal, run `claude` and sign in (or run /login inside Claude Code), then send your message again.';
     }
-    if (/API Error:\s*400/i.test(t) && /(tool use|concurren|parallel)/i.test(t)) {
-        return 'Local Mode hit a known Claude Code error (a print-mode bug that affects every current version). I retried a few times automatically — sending the message again sometimes gets through. See https://github.com/anthropics/claude-code/issues/18131';
-    }
     if (/^API Error:/i.test(t)) {
         return `Local Mode error from Claude Code: ${t.replace(/^API Error:\s*/i, '').trim()}`;
     }
-    return `Local Mode stopped unexpectedly (claude exited with code ${code ?? 'unknown'}). Please try again.`;
+    if (t) {
+        return `Local Mode stopped unexpectedly: ${t}`;
+    }
+    return 'Local Mode stopped unexpectedly. Please try again.';
+}
+// ─── Prompt construction ─────────────────────────────────────────────────────────
+//
+// The gateway pushes the FULL conversation (sanitised via @presence/shared
+// toModelMessages) and `history` already ends with the new user turn. The SDK's
+// streaming input only triggers an assistant turn for user messages whose
+// `shouldQuery` is not false — so we replay every PRIOR turn with
+// shouldQuery:false (appended to the transcript, no turn generated), inject
+// assistant turns verbatim (carrying their tool_use blocks), and let ONLY the
+// final/live user turn run. This preserves stateless structured replay: no
+// session resume, no flat-text collapse, no local jsonl — Firestore stays the
+// single source of truth, exactly as the CLI stdin replay did.
+function buildPromptMessages(history) {
+    // Index of the live user turn — the last user-role message. The gateway
+    // always appends it; the scan is defensive against an unexpected tail.
+    let liveIdx = -1;
+    for (let i = history.length - 1; i >= 0; i--) {
+        if (history[i].role === 'user') {
+            liveIdx = i;
+            break;
+        }
+    }
+    const out = [];
+    history.forEach((msg, i) => {
+        // Normalise to array-of-blocks (a bare string becomes a single text block).
+        const content = Array.isArray(msg.content)
+            ? msg.content
+            : [{ type: 'text', text: typeof msg.content === 'string' ? msg.content : '' }];
+        if (msg.role === 'assistant') {
+            // Injected verbatim — runtime accepts a {type:'assistant'} message on the
+            // input stream and appends it to the transcript (it never triggers a turn).
+            out.push({ type: 'assistant', message: { role: 'assistant', content }, parent_tool_use_id: null });
+        }
+        else {
+            const isLive = i === liveIdx;
+            out.push({
+                type: 'user',
+                ...(isLive ? {} : { shouldQuery: false }),
+                message: { role: 'user', content },
+                parent_tool_use_id: null,
+            });
+        }
+    });
+    return out;
 }
-// ─── Spawn ────────────────────────────────────────────────────────────────────
-function spawnClaude(params) {
+async function* promptStream(messages) {
+    for (const m of messages)
+        yield m;
+}
+// ─── Spawn (drive one turn through the SDK) ──────────────────────────────────────
+export function spawnClaude(params) {
     const { conversationId, presenceSessionId, text, uid, history, vaultFileOpen, clientCapabilities, syncedFolders, onEvent, onDone, onError, onNotice } = params;
-    const attemptIdx = params._attemptIdx ?? 0;
-    const firstAttemptAt = params._firstAttemptAt ?? Date.now();
-    const systemPromptPath = (0, path_1.join)((0, os_1.tmpdir)(), `agent-${uid}.md`);
-    const mcpConfigPath = (0, path_1.join)((0, os_1.tmpdir)(), `mcp-${uid}.json`);
+    const systemPromptPath = join(tmpdir(), `agent-${uid}.md`);
+    const mcpConfigPath = join(tmpdir(), `mcp-${uid}.json`);
     if (verbose) {
         process.stderr.write(paint('90', `[bridge:verbose] cwd:           ${BRIDGE_CWD}`) + '\n');
-        process.stderr.write(paint('90', `[bridge:verbose] override md:   ${(0, path_1.join)(BRIDGE_CWD, 'CLAUDE.md')}`) + '\n');
+        process.stderr.write(paint('90', `[bridge:verbose] override md:   ${join(BRIDGE_CWD, 'CLAUDE.md')}`) + '\n');
         process.stderr.write(paint('90', `[bridge:verbose] system prompt: ${systemPromptPath}`) + '\n');
         process.stderr.write(paint('90', `[bridge:verbose] mcp config:    ${mcpConfigPath}`) + '\n');
         process.stderr.write(paint('90', `[bridge:verbose] session id:    ${presenceSessionId}`) + '\n');
@@ -189,22 +205,12 @@ function spawnClaude(params) {
     }
     // Surface the user's UID before the session line in every mode — it's the
     // Firestore doc prefix (`sessions/<uid>_<conversationId>`), so logging it
-    // makes a reported bridge failure correlatable to the stored session without
-    // having to ask which account hit it. The CLI's own `--session-id` is
-    // ephemeral and is NOT the Firestore conversationId, so the uid is the key
-    // join column when debugging.
+    // makes a reported bridge failure correlatable to the stored session.
     process.stderr.write(`[bridge] user ${uid}\n`);
-    // Debug transcript: lead with the user prompt for this turn (the clean
-    // message, before the gateway's ephemeral-context prefix), plus the session
-    // id (correlates with the chat URL / Firestore session doc) and a hint at
-    // how much prior context is being replayed.
+    // Debug transcript: lead with the prior context (replayed history) then the
+    // live user prompt. `history` already ends with the new user turn.
     if (debug || verbose) {
         process.stderr.write(`\n${paint('1', `══ session ${presenceSessionId} ══`)}\n`);
-        // `history` already ends with the new user prompt (gateway-appended). Render
-        // every PRIOR message with its role colour so what the model saw as context
-        // is auditable — then the live turn is shown via the clean `text` below and
-        // streams in beneath it. Defensive: if the tail isn't the current user turn
-        // (unexpected), render the whole history rather than dropping a message.
         const tail = history[history.length - 1];
         const prior = tail?.role === 'user' ? history.slice(0, -1) : history;
         if (prior.length > 0) {
@@ -212,456 +218,366 @@ function spawnClaude(params) {
             for (const msg of prior)
                 renderHistoryMessage(msg);
         }
-        debugBlock('user · this turn', exports.SECTION_COLORS.user, text);
+        debugBlock('user · this turn', SECTION_COLORS.user, text);
     }
     else {
-        // Default mode is quiet, but always surface the session id once per turn so
-        // it can be matched to the chat URL / Firestore session doc when debugging.
         process.stderr.write(`[bridge] session ${presenceSessionId}\n`);
     }
-    // tool_use_id → tool name, so a tool_result block (which only carries the id)
-    // can be labelled with the tool it answers in the debug transcript.
-    const toolNames = new Map();
-    // If a prior process is still running for this conversation (user sent a
-    // follow-up before the previous turn finished), supersede it. The latest
-    // user intent wins; the orphan would otherwise keep streaming events.
-    const existing = active.get(conversationId);
-    if (existing) {
-        process.stderr.write(`[bridge] superseding active conversation ${conversationId}\n`);
-        existing.kill('SIGTERM');
-        active.delete(conversationId);
-    }
-    // Cancel any retry scheduled for this conversation that hasn't fired yet.
-    // Without this, a new user message arriving during a retry's backoff window
-    // would race the stale retry — which carries the OLD turn's history and would
-    // clobber the new turn. Skip when this call IS the retry firing (attemptIdx>0,
-    // the timer already deleted itself before invoking us).
-    if (attemptIdx === 0) {
-        const pending = pendingRetries.get(conversationId);
-        if (pending) {
-            clearTimeout(pending);
-            pendingRetries.delete(conversationId);
-            process.stderr.write(`[bridge] cancelled pending retry for ${conversationId} (superseded by new turn)\n`);
-        }
-    }
-    // Note: ephemeral context (vault_file_open / client_capabilities / synced_folders)
-    // is injected into the last user message by the gateway BEFORE history is
-    // sent over the WS. The bridge no longer constructs `userMessageText` —
-    // `history` is the authoritative stream and already contains the new user
-    // prompt with prefix prepended. The `text`, `vaultFileOpen`,
-    // `clientCapabilities`, `syncedFolders` SpawnParams are retained for
-    // backward-compatible logging / spool correlation only.
+    // ephemeral context (vault_file_open / client_capabilities / synced_folders) is
+    // injected into the last user message by the gateway BEFORE history is sent —
+    // these params are retained for backward-compatible logging only.
     void vaultFileOpen;
     void clientCapabilities;
     void syncedFolders;
     void text;
-    // Lockdown rationale:
-    // - `--tools ""` disables ALL built-in tools (Bash/Read/Write/Edit/Glob/Grep/
-    //   WebFetch/etc.). MCP tools are not "built-in" so the 1Presence MCP surface
-    //   remains available.
-    // - `--setting-sources ""` prevents claude CLI from loading the user's
-    //   ~/.claude/settings.json (and project/.local equivalents). Without this,
-    //   permissive `permissions.allow` rules in the user's personal Claude Code
-    //   config would silently re-enable Bash/Edit/Write etc. inside the bridge.
-    // - `--strict-mcp-config` keeps the MCP surface to exactly what we wire in
-    //   via --mcp-config. Together these guarantee the bridge can only call
-    //   `mcp__1presence__*` — no filesystem, no shell, no arbitrary network.
-    //
-    // Session continuity rationale:
-    // - `--input-format stream-json` accepts structured user/assistant messages
-    //   on stdin. We replay prior turns (loaded by the gateway from Firestore)
-    //   followed by the new user turn — this is how the bridge sees history.
-    // - `--no-session-persistence` keeps no jsonl on disk. The bridge has zero
-    //   local filesystem dependency for continuity; Firestore is the only
-    //   source of truth.
-    // - `--session-id <uuid>` must be a fresh UUID per spawn: the CLI treats
-    //   this flag as a "claim a new session ID" operation and rejects the
-    //   second spawn with "Session ID X is already in use" if we reuse one
-    //   across turns of a chat — even with --no-session-persistence. The
-    //   bridge passes the per-spawn `conversationId` here; the presence
-    //   sessionId is correlated separately via bridge logs and spool records.
-    // The CLI treats --session-id as "claim this new session ID" and rejects a
-    // reused id with "Session ID X is already in use". A retry is a fresh spawn,
-    // so it MUST use a new uuid; the first attempt keeps the correlation id.
-    const spawnSessionId = attemptIdx === 0 ? presenceSessionId : crypto.randomUUID();
-    const args = [
-        '--print',
-        '--input-format', 'stream-json',
-        '--output-format', 'stream-json',
-        '--verbose',
-        '--tools', '',
-        '--setting-sources', '',
-        '--allowedTools', 'mcp__1presence__*',
-        '--system-prompt-file', systemPromptPath,
-        '--mcp-config', mcpConfigPath,
-        '--strict-mcp-config',
-        '--no-session-persistence',
-        '--session-id', spawnSessionId,
-    ];
-    const pinnedModel = (0, config_1.getBridgeModel)();
-    if (pinnedModel) {
-        args.push('--model', pinnedModel);
-    }
-    // Strip API key so Claude Code uses the user's claude.ai Pro subscription
-    // (OAuth credentials), not an API key that would bill to a separate account.
-    const { ANTHROPIC_API_KEY: _stripped, ...safeEnv } = process.env;
-    const proc = (0, child_process_1.spawn)('claude', args, {
-        cwd: BRIDGE_CWD,
-        env: safeEnv,
-        stdio: ['pipe', 'pipe', 'pipe'],
-    });
-    active.set(conversationId, proc);
-    // Feed the full conversation via stdin as stream-json. The gateway's
-    // early-save committed the new user message to Firestore BEFORE building
-    // `history`, so `history` already ends with the new user prompt (with the
-    // ephemeral context prefix prepended by the gateway). The bridge no longer
-    // appends a separate `newTurn` — doing so would duplicate the user prompt.
-    // Sanitisation (orphan tool_use stripping, displayOnly filtering, consecutive
-    // same-role merging) already happened on the gateway via
-    // @presence/shared.toModelMessages — replay the history verbatim.
-    try {
-        const stdin = proc.stdin;
-        if (!stdin) {
-            throw new Error('claude stdin is null — spawn must use stdio[0]="pipe"');
-        }
-        for (const msg of history) {
-            // Normalise to array-of-blocks: Claude Code's stream-json input parser
-            // iterates `content` directly. A string slips into a `"tool_use_id" in
-            // <char>` check inside the CLI and aborts the process with `W is not an
-            // Object` (JSC) / exit 1 mid-turn. The gateway also normalises before
-            // sending, so a current gateway + any bridge version is safe; this guard
-            // covers older gateways and ad-hoc local replay tests.
-            const content = Array.isArray(msg.content)
-                ? msg.content
-                : [{ type: 'text', text: typeof msg.content === 'string' ? msg.content : '' }];
-            const wrapped = { type: msg.role, message: { role: msg.role, content } };
-            stdin.write(JSON.stringify(wrapped) + '\n');
-        }
-        stdin.end();
-    }
-    catch (err) {
-        process.stderr.write(`[bridge] failed to write stdin: ${err.message}\n`);
-        proc.kill('SIGTERM');
-        onError(`stdin write failed: ${err.message}`, null, null);
-        return;
+    // Supersede any in-flight turn for this conversation (user sent a follow-up
+    // before the previous turn finished). The latest intent wins.
+    const existing = active.get(conversationId);
+    if (existing) {
+        process.stderr.write(`[bridge] superseding active conversation ${conversationId}\n`);
+        existing.abort();
+        active.delete(conversationId);
     }
+    const abort = new AbortController();
+    active.set(conversationId, abort);
+    // tool_use_id → tool name, so a tool_result block (which only carries the id)
+    // can be labelled with the tool it answers in the debug transcript.
+    const toolNames = new Map();
+    // Per-turn accounting.
     let sessionIdExtracted = false;
     let messageCount = 0;
     let costUsd = 0;
     let usage = null;
     // Prompt size of the MOST RECENT assistant call (input + both cache buckets),
     // overwritten on each assistant event so it ends on the turn's final, fullest
-    // call. This — not the summed `usage` above — is the current context fill the
-    // status line's 🧠 segment reports against the model's window.
+    // call. This — not the summed `usage` — is the current context fill the status
+    // line's 🧠 segment reports against the model's window.
     let lastContextTokens = 0;
     let extractedModel = null;
-    let buffer = '';
     let killedForViolation = false;
-    // Retry/error-surfacing tracking for this attempt:
-    //  - sawApiError: the CLI emitted an "API Error:" assistant text event (the
-    //    way Claude Code reports an underlying API failure mid-turn).
-    //  - apiErrorText: that text, captured for describeCliFailure().
-    //  - producedRealOutput: any real assistant text or tool_use was emitted, so
-    //    a later failure must NOT be retried (could double-run a side-effect).
     let sawApiError = false;
-    //  - sawAuthFailure: a 401/403 auth/credential failure (the user's local
-    //    Claude Code is not signed in). Surfaced as an actionable message and
-    //    never retried (re-spawning won't add credentials).
     let sawAuthFailure = false;
     let apiErrorText = '';
     let producedRealOutput = false;
-    proc.stdout.on('data', (chunk) => {
-        buffer += chunk.toString('utf-8');
-        const lines = buffer.split('\n');
-        buffer = lines.pop() ?? '';
-        for (const line of lines) {
-            const trimmed = line.trim();
-            if (!trimmed)
-                continue;
-            let event;
-            try {
-                event = JSON.parse(trimmed);
+    // Allow only the 1Presence MCP surface to execute. Built-in tools are disabled
+    // via extraArgs `--tools ""`; this is the runtime safety net (a hard deny that
+    // runs before any execution) for anything that slips past. Our MCP tools are
+    // auto-approved via allowedTools, so this callback only ever fires to deny.
+    const canUseTool = async (toolName, input) => {
+        if (toolName.startsWith('mcp__1presence__'))
+            return { behavior: 'allow', updatedInput: input };
+        return { behavior: 'deny', message: `Tool ${toolName} is not allowed in Local Mode`, interrupt: true };
+    };
+    // Strip API key so Claude Code uses the user's claude.ai subscription (OAuth
+    // credentials in the Keychain), not an API key that would bill a separate
+    // account. Options.env REPLACES the subprocess env, so spread the rest through.
+    const { ANTHROPIC_API_KEY: _stripped, ...safeEnv } = process.env;
+    const pinnedModel = getBridgeModel();
+    // Process one translated raw stream-json event: bookkeeping + forward. Mirrors
+    // the old CLI stdout parser so the gateway/accumulator see identical shapes.
+    // Returns false when the event must be suppressed (errors) or the turn was
+    // killed for a tool violation.
+    function handleEvent(event) {
+        const type = event['type'];
+        if (!sessionIdExtracted && type === 'system' && event['subtype'] === 'init') {
+            const keySource = event['apiKeySource'];
+            const model = event['model'];
+            if (model)
+                extractedModel = model;
+            if (!modelAnnounced) {
+                const source = keySource === 'none' || !keySource ? 'claude.ai subscription' : keySource;
+                const pin = getBridgeModel() ? '  (selected at startup)' : '';
+                process.stdout.write(`\n  model: ${model ?? 'unknown'}${pin}\n  auth:  ${source}\n\n`);
+                modelAnnounced = true;
             }
-            catch {
-                continue;
+            else {
+                process.stderr.write(`[bridge] model: ${model ?? 'unknown'}  apiKeySource: ${keySource ?? 'none'}\n`);
             }
-            const type = event['type'];
-            // Set when this event is the CLI's "API Error:" turn — we neither forward
-            // it to the PWA nor let it reach the accumulator (it carries no real
-            // content and would poison history / show a raw error mid-stream).
-            let suppressEvent = false;
-            // Extract model + key source info from the first system/init event.
-            // No session-id persistence — Firestore is the only source of truth
-            // now, and we pin --session-id to presenceSessionId on every spawn.
-            if (!sessionIdExtracted && type === 'system' && event['subtype'] === 'init') {
-                const keySource = event['apiKeySource'];
-                const model = event['model'];
-                if (model)
-                    extractedModel = model;
-                if (!modelAnnounced) {
-                    // First conversation since bridge started — announce prominently
-                    // so the user can confirm which model and credential is in use.
-                    const source = keySource === 'none' || !keySource ? 'claude.ai subscription' : keySource;
-                    const pin = (0, config_1.getBridgeModel)() ? '  (selected at startup)' : '';
-                    process.stdout.write(`\n  model: ${model ?? 'unknown'}${pin}\n  auth:  ${source}\n\n`);
-                    modelAnnounced = true;
-                }
-                else {
-                    // Subsequent conversations — quiet line for power users.
-                    process.stderr.write(`[bridge] model: ${model ?? 'unknown'}  apiKeySource: ${keySource ?? 'none'}\n`);
-                }
-                sessionIdExtracted = true;
+            sessionIdExtracted = true;
+        }
+        if (type === 'assistant') {
+            messageCount++;
+            const msg = event['message'];
+            const u = msg?.['usage'];
+            if (u) {
+                usage = {
+                    input_tokens: (usage?.input_tokens ?? 0) + (u['input_tokens'] ?? 0),
+                    output_tokens: (usage?.output_tokens ?? 0) + (u['output_tokens'] ?? 0),
+                    cache_read_input_tokens: (usage?.cache_read_input_tokens ?? 0) + (u['cache_read_input_tokens'] ?? 0),
+                    cache_creation_input_tokens: (usage?.cache_creation_input_tokens ?? 0) + (u['cache_creation_input_tokens'] ?? 0),
+                };
+                lastContextTokens = (u['input_tokens'] ?? 0) + (u['cache_read_input_tokens'] ?? 0) + (u['cache_creation_input_tokens'] ?? 0);
             }
-            // Count complete assistant turns + accumulate token usage + log tool calls
-            if (type === 'assistant') {
-                messageCount++;
-                const msg = event['message'];
-                const u = msg?.['usage'];
-                if (u) {
-                    usage = {
-                        input_tokens: (usage?.input_tokens ?? 0) + (u['input_tokens'] ?? 0),
-                        output_tokens: (usage?.output_tokens ?? 0) + (u['output_tokens'] ?? 0),
-                        cache_read_input_tokens: (usage?.cache_read_input_tokens ?? 0) + (u['cache_read_input_tokens'] ?? 0),
-                        cache_creation_input_tokens: (usage?.cache_creation_input_tokens ?? 0) + (u['cache_creation_input_tokens'] ?? 0),
-                    };
-                    // Full prompt size of THIS assistant call — non-cached input plus both
-                    // cache buckets. Overwrite (don't sum): the last write wins, which is
-                    // the turn's largest/final context.
-                    lastContextTokens = (u['input_tokens'] ?? 0) + (u['cache_read_input_tokens'] ?? 0) + (u['cache_creation_input_tokens'] ?? 0);
-                }
-                const content = msg?.['content'];
-                if (Array.isArray(content)) {
-                    let wroteText = false;
-                    for (const block of content) {
-                        if (block['type'] === 'tool_use') {
-                            producedRealOutput = true;
-                            const toolName = block['name'];
-                            const toolId = block['id'];
-                            if (toolId)
-                                toolNames.set(toolId, toolName);
-                            if (debug) {
-                                // Clean transcript: a single coloured block with the full input.
-                                debugBlock(`tool → ${toolName}`, exports.SECTION_COLORS.input, formatPayload(block['input']));
-                            }
-                            else {
-                                if (wroteText) {
-                                    process.stderr.write('\n');
-                                    wroteText = false;
-                                }
-                                const prefix = toolName.startsWith('mcp__') ? '[mcp]' : '[tool]';
-                                process.stderr.write(paint(exports.SECTION_COLORS.input, `[bridge] ${prefix} ${toolName}`) + '\n');
-                                if (verbose) {
-                                    const input = block['input'];
-                                    process.stderr.write(paint(exports.SECTION_COLORS.input, `[bridge:verbose] ─── input ${toolName} ───\n${formatPayload(input)}\n[bridge:verbose] ─── end input ───`) + '\n');
-                                }
+            const content = msg?.['content'];
+            if (Array.isArray(content)) {
+                let wroteText = false;
+                for (const block of content) {
+                    if (block['type'] === 'tool_use') {
+                        producedRealOutput = true;
+                        const toolName = block['name'];
+                        const toolId = block['id'];
+                        if (toolId)
+                            toolNames.set(toolId, toolName);
+                        if (debug) {
+                            debugBlock(`tool → ${toolName}`, SECTION_COLORS.input, formatPayload(block['input']));
+                        }
+                        else {
+                            if (wroteText) {
+                                process.stderr.write('\n');
+                                wroteText = false;
                             }
-                            // Defense-in-depth: CLI flags (--tools "", --allowedTools, --strict-mcp-config,
-                            // --setting-sources "") are supposed to make this unreachable. If we see a
-                            // non-1Presence tool here anyway, something has bypassed those guards — kill
-                            // immediately so any side effect already in flight is the only damage done.
-                            //
-                            // Valid forms:
-                            //   mcp__1presence__<name>  — namespaced MCP form
-                            //   <snake_case_name>       — bare form; Claude Code may omit the prefix in
-                            //                            stream-json output. Safe because --strict-mcp-config
-                            //                            limits MCP to the 1presence server only.
-                            // Invalid (real violations):
-                            //   PascalCase (Bash, Read, Write, …) — Claude Code built-ins
-                            //   mcp__<other>__*         — tools from a different MCP server
-                            const isMcp1presence = toolName.startsWith('mcp__1presence__');
-                            const isBareName = /^[a-z][a-z0-9_]*$/.test(toolName);
-                            if (!isMcp1presence && !isBareName) {
-                                killedForViolation = true;
-                                const violation = `bridge tool violation: ${toolName} is not allowed in Local Mode`;
-                                process.stderr.write(`[bridge] FATAL ${violation} — killing\n`);
-                                active.delete(conversationId);
-                                proc.kill('SIGKILL');
-                                onError(violation, usage, extractedModel);
-                                return;
+                            const prefix = toolName.startsWith('mcp__') ? '[mcp]' : '[tool]';
+                            process.stderr.write(paint(SECTION_COLORS.input, `[bridge] ${prefix} ${toolName}`) + '\n');
+                            if (verbose) {
+                                process.stderr.write(paint(SECTION_COLORS.input, `[bridge:verbose] ─── input ${toolName} ───\n${formatPayload(block['input'])}\n[bridge:verbose] ─── end input ───`) + '\n');
                             }
                         }
-                        else if (block['type'] === 'text') {
-                            const text = block['text'];
-                            if (text) {
-                                // The CLI reports auth/credential failures (401/403) as a
-                                // <synthetic> assistant text turn whose wording varies and does
-                                // NOT reliably start with "API Error:" — e.g. "Please run /login
-                                // · API Error: 401 Invalid authentication credentials" or
-                                // "Failed to authenticate. API Error: 401 …". Detect by the
-                                // structured signal (the event's `error: authentication_failed`
-                                // / `model: <synthetic>`) plus a wording fallback, so it is
-                                // classified instead of leaking raw into the chat as if the
-                                // model had said it.
-                                const isSynthetic = msg?.['model'] === '<synthetic>';
-                                const isAuthFailure = event['error'] === 'authentication_failed' ||
-                                    (isSynthetic && /(api error:\s*40[13]\b|invalid (api key|authentication)|please run \/login|failed to authenticate|unauthor)/i.test(text));
-                                if (/^API Error:/i.test(text.trimStart()) || isAuthFailure) {
-                                    // The CLI is reporting an underlying API/auth failure as
-                                    // assistant text. Capture it for the user-facing message, and
-                                    // suppress the whole event so the raw error never reaches the
-                                    // PWA or the accumulator (the gateway also blanks it via
-                                    // cleanTurnText — this is the upstream defense).
-                                    sawApiError = true;
-                                    apiErrorText = text.trim();
-                                    if (isAuthFailure)
-                                        sawAuthFailure = true;
-                                    // Operator log keeps the raw provider line verbatim (with a
-                                    // [bridge] prefix) so the real reason is diagnosable locally.
-                                    suppressEvent = true;
-                                    process.stderr.write(paint(exports.SECTION_COLORS.result, `[bridge] ${text.replace(/\n+/g, ' ')}`) + '\n');
-                                }
-                                else {
-                                    producedRealOutput = true;
-                                    if (debug) {
-                                        // Full text, newlines intact — the readable transcript.
-                                        debugBlock('assistant', exports.SECTION_COLORS.assistant, text);
-                                    }
-                                    else {
-                                        process.stderr.write(paint(exports.SECTION_COLORS.assistant, text.replace(/\n+/g, ' ')));
-                                        wroteText = true;
-                                    }
-                                }
-                            }
+                        // Defense-in-depth: canUseTool + --tools "" + strictMcpConfig should
+                        // make a non-1Presence tool unreachable. If one appears anyway, kill
+                        // the turn so any side effect in flight is the only damage done.
+                        const isMcp1presence = toolName.startsWith('mcp__1presence__');
+                        const isBareName = /^[a-z][a-z0-9_]*$/.test(toolName);
+                        if (!isMcp1presence && !isBareName) {
+                            killedForViolation = true;
+                            const violation = `bridge tool violation: ${toolName} is not allowed in Local Mode`;
+                            process.stderr.write(`[bridge] FATAL ${violation} — aborting\n`);
+                            active.delete(conversationId);
+                            abort.abort();
+                            onError(violation, usage, extractedModel);
+                            return false;
                         }
                     }
-                    if (wroteText)
-                        process.stderr.write('\n');
-                }
-            }
-            // Tool results stream back as `user` events with tool_result blocks.
-            if ((verbose || debug) && type === 'user') {
-                const msg = event['message'];
-                const content = msg?.['content'];
-                if (Array.isArray(content)) {
-                    for (const block of content) {
-                        if (block['type'] === 'tool_result') {
-                            const id = block['tool_use_id'] ?? '';
-                            const out = block['content'];
+                    else if (block['type'] === 'text') {
+                        const blockText = block['text'];
+                        if (blockText) {
+                            // The CLI/SDK can report auth/API failures as a synthetic assistant
+                            // text turn whose wording varies. Detect by the structured signal
+                            // (event.error) plus a wording fallback, so it's classified rather
+                            // than leaking raw into the chat as if the model had said it.
+                            const isSynthetic = msg?.['model'] === '<synthetic>';
+                            const isAuthFailure = event['error'] === 'authentication_failed' ||
+                                (isSynthetic && /(api error:\s*40[13]\b|invalid (api key|authentication)|please run \/login|failed to authenticate|unauthor)/i.test(blockText));
+                            if (/^API Error:/i.test(blockText.trimStart()) || isAuthFailure) {
+                                sawApiError = true;
+                                apiErrorText = blockText.trim();
+                                if (isAuthFailure)
+                                    sawAuthFailure = true;
+                                process.stderr.write(paint(SECTION_COLORS.result, `[bridge] ${blockText.replace(/\n+/g, ' ')}`) + '\n');
+                                return false; // suppress — never forward a raw error turn
+                            }
+                            producedRealOutput = true;
                             if (debug) {
-                                const name = toolNames.get(id) ?? id ?? 'result';
-                                const errFlag = block['is_error'] ? '  [error]' : '';
-                                debugBlock(`result ← ${name}${errFlag}`, exports.SECTION_COLORS.result, formatPayload(out));
+                                debugBlock('assistant', SECTION_COLORS.assistant, blockText);
                             }
                             else {
-                                process.stderr.write(paint(exports.SECTION_COLORS.result, `[bridge:verbose] ─── output ${id} ───\n${formatPayload(out)}\n[bridge:verbose] ─── end output ───`) + '\n');
+                                process.stderr.write(paint(SECTION_COLORS.assistant, blockText.replace(/\n+/g, ' ')));
+                                wroteText = true;
                             }
                         }
                     }
                 }
+                if (wroteText)
+                    process.stderr.write('\n');
             }
-            // Extract cost from the final result event. The CLI also stamps auth/API
-            // failures here as `is_error` + `api_error_status` (even though `subtype`
-            // stays "success"), so treat it as a robust backstop in case the
-            // assistant-text signal above was missed (wording drift across CLI
-            // versions). 401/403 → auth failure; other statuses keep the existing
-            // 400-retry behaviour (sawApiError only).
-            if (type === 'result') {
-                const c = event['cost_usd'] ?? event['total_cost_usd'];
-                if (typeof c === 'number')
-                    costUsd = c;
-                if (event['is_error'] === true) {
-                    sawApiError = true;
-                    const status = event['api_error_status'];
-                    if (status === 401 || status === 403)
-                        sawAuthFailure = true;
-                    if (!apiErrorText && typeof event['result'] === 'string') {
-                        apiErrorText = event['result'].trim();
+        }
+        if ((verbose || debug) && type === 'user') {
+            const msg = event['message'];
+            const content = msg?.['content'];
+            if (Array.isArray(content)) {
+                for (const block of content) {
+                    if (block['type'] === 'tool_result') {
+                        const id = block['tool_use_id'] ?? '';
+                        const out = block['content'];
+                        if (debug) {
+                            const name = toolNames.get(id) ?? id ?? 'result';
+                            const errFlag = block['is_error'] ? '  [error]' : '';
+                            debugBlock(`result ← ${name}${errFlag}`, SECTION_COLORS.result, formatPayload(out));
+                        }
+                        else {
+                            process.stderr.write(paint(SECTION_COLORS.result, `[bridge:verbose] ─── output ${id} ───\n${formatPayload(out)}\n[bridge:verbose] ─── end output ───`) + '\n');
+                        }
                     }
                 }
             }
-            if (!suppressEvent)
-                onEvent(event);
         }
-    });
-    proc.stderr.on('data', (chunk) => {
-        // claude CLI writes logs to stderr — not errors, just noise
-        const line = chunk.toString('utf-8').trim();
-        if (line)
-            process.stderr.write(`[claude] ${line}\n`);
-    });
-    proc.on('close', (code) => {
-        active.delete(conversationId);
-        // Violation path already called onError + killed — don't double-fire.
-        if (killedForViolation)
+        if (type === 'result') {
+            // total_cost_usd is the SDK's notional figure (0 on the no-op history
+            // append cycle; the real number on the live turn). Keep the largest seen.
+            const c = event['total_cost_usd'] ?? event['cost_usd'];
+            if (typeof c === 'number' && c > costUsd)
+                costUsd = c;
+            if (event['is_error'] === true) {
+                sawApiError = true;
+                const status = event['api_error_status'];
+                if (status === 401 || status === 403)
+                    sawAuthFailure = true;
+                if (!apiErrorText && typeof event['result'] === 'string') {
+                    apiErrorText = event['result'].trim();
+                }
+            }
+        }
+        return true;
+    }
+    // Drive the turn. Synchronous spawnClaude returns immediately; the SDK loop
+    // runs in this async IIFE and fires the same callbacks the CLI path did.
+    void (async () => {
+        let systemPrompt;
+        let mcpServers;
+        try {
+            systemPrompt = readFileSync(systemPromptPath, 'utf-8');
+            const mcpRaw = JSON.parse(readFileSync(mcpConfigPath, 'utf-8'));
+            mcpServers = mcpRaw.mcpServers ?? {};
+        }
+        catch (err) {
+            active.delete(conversationId);
+            onError(`Local Mode setup files unavailable: ${err.message}`, null, null);
             return;
-        // Flush any remaining buffer
-        if (buffer.trim()) {
-            try {
-                onEvent(JSON.parse(buffer.trim()));
+        }
+        const options = {
+            systemPrompt, // custom string → replaces the default Claude Code prompt
+            mcpServers: mcpServers,
+            strictMcpConfig: true, // only our MCP server, ignore project/user/plugin MCP
+            settingSources: [], // no user/project settings or memory
+            allowedTools: ['mcp__1presence__*'], // auto-approve our MCP surface (no prompt)
+            canUseTool, // hard deny anything else
+            extraArgs: { tools: '' }, // disable ALL built-in tools (= CLI --tools "")
+            cwd: BRIDGE_CWD,
+            abortController: abort,
+            includePartialMessages: false, // whole messages, matching the old non-partial path
+            permissionMode: 'default',
+            env: safeEnv,
+            stderr: (line) => { if (verbose && line.trim())
+                process.stderr.write(`[claude] ${line.trim()}\n`); },
+            ...(pinnedModel ? { model: pinnedModel } : {}),
+        };
+        const promptMessages = buildPromptMessages(history);
+        try {
+            for await (const m of query({ prompt: promptStream(promptMessages), options })) {
+                // Skip echoed input replays — they would double-count in the accumulator
+                // and re-stream prior turns to the PWA.
+                if (m.isReplay)
+                    continue;
+                switch (m.type) {
+                    case 'system': {
+                        if (m.subtype === 'init') {
+                            const init = m;
+                            const event = { type: 'system', subtype: 'init', model: init.model, apiKeySource: init.apiKeySource };
+                            if (handleEvent(event))
+                                onEvent(event);
+                        }
+                        break;
+                    }
+                    case 'assistant': {
+                        const am = m;
+                        // Structured error signal — classify, do not forward as a real turn.
+                        if (am.error) {
+                            sawApiError = true;
+                            if (am.error === 'authentication_failed' || am.error === 'oauth_org_not_allowed')
+                                sawAuthFailure = true;
+                            if (!apiErrorText)
+                                apiErrorText = `API Error: ${am.error}`;
+                            process.stderr.write(paint(SECTION_COLORS.result, `[bridge] assistant error: ${am.error}`) + '\n');
+                            break;
+                        }
+                        const event = { type: 'assistant', message: am.message, error: am.error };
+                        if (handleEvent(event))
+                            onEvent(event);
+                        if (killedForViolation)
+                            return; // handleEvent already aborted + onError
+                        break;
+                    }
+                    case 'user': {
+                        const um = m;
+                        const event = { type: 'user', message: um.message };
+                        if (handleEvent(event))
+                            onEvent(event);
+                        break;
+                    }
+                    case 'result': {
+                        const rm = m;
+                        const event = {
+                            type: 'result',
+                            subtype: rm['subtype'],
+                            total_cost_usd: rm['total_cost_usd'],
+                            is_error: rm['is_error'],
+                            api_error_status: rm['api_error_status'],
+                            result: rm['result'],
+                        };
+                        if (handleEvent(event))
+                            onEvent(event);
+                        break;
+                    }
+                    case 'auth_status': {
+                        const as = m;
+                        if (as.error) {
+                            sawApiError = true;
+                            sawAuthFailure = true;
+                            apiErrorText = as.error;
+                        }
+                        break;
+                    }
+                    case 'rate_limit_event': {
+                        // SDK surfaces upstream rate-limit pauses; it retries internally.
+                        // Admin-only ephemeral notice — jargon is fine in Local Mode.
+                        onNotice?.('Claude Code is pausing briefly for an upstream rate limit, then continuing…');
+                        break;
+                    }
+                    // Everything else (partial messages, status, hooks, notifications,
+                    // thinking tokens, …) is SDK-internal — not part of the CLI event
+                    // contract the gateway/accumulator understand, so it is dropped.
+                    default:
+                        break;
+                }
             }
-            catch { /* ignore */ }
         }
-        // An auth failure can land on a "successful" exit (the CLI stamps it on the
-        // result event but still exits 0 in some versions), and we've suppressed its
-        // text — so without this the turn would finish silently empty. Treat it as a
-        // failure regardless of exit code.
-        if (sawAuthFailure || (code !== 0 && code !== null)) {
-            // Auto-retry when the CLI failed BEFORE producing any real output — the
-            // signature of the known print-mode 400 regression. A fresh spawn (new
-            // --session-id) often succeeds. We never retry once real text or a tool
-            // call landed, to avoid double-running a side-effectful tool. We also
-            // never retry an auth failure — re-spawning won't add missing credentials,
-            // it just burns the user's plan. Retries use escalating backoff and stop
-            // past the wall-clock cap (see consts above).
-            const elapsed = Date.now() - firstAttemptAt;
-            if (attemptIdx < MAX_TURN_RETRIES && sawApiError && !sawAuthFailure && !producedRealOutput && elapsed < RETRY_WALL_CLOCK_CAP_MS) {
-                const delay = RETRY_BACKOFF_BASE_MS * (attemptIdx + 1);
-                const nextAttempt = attemptIdx + 2;
-                process.stderr.write(`[bridge] turn failed before output (${apiErrorText.replace(/\n+/g, ' ').slice(0, 120)}) — retrying (${nextAttempt} of ${MAX_TURN_RETRIES + 1}) in ${delay}ms\n`);
-                // Admin-only ephemeral thread notice — jargon is fine in Local Mode.
-                onNotice?.(`Claude Code print-mode 400 (tool-use concurrency, anthropics/claude-code#18131) — respawning, attempt ${nextAttempt}/${MAX_TURN_RETRIES + 1}…`);
-                const timer = setTimeout(() => {
-                    pendingRetries.delete(conversationId);
-                    spawnClaude({ ...params, _attemptIdx: attemptIdx + 1, _firstAttemptAt: firstAttemptAt });
-                }, delay);
-                pendingRetries.set(conversationId, timer);
+        catch (err) {
+            active.delete(conversationId);
+            // Aborted by supersede or the Stop button — no error to surface.
+            if (abort.signal.aborted)
                 return;
+            if (killedForViolation)
+                return;
+            const message = err?.message ?? String(err);
+            if (/40[13]\b|unauthor|invalid (api key|authentication)|please run \/login/i.test(message)) {
+                sawAuthFailure = true;
             }
-            // Pass any partial token usage we observed before the failure so the
-            // PWA and the gateway's bridge usage store can still record it. Surface a
-            // classified, user-readable message instead of the opaque exit code.
-            onError(describeCliFailure(code, apiErrorText, sawAuthFailure), usage, extractedModel);
-        }
-        else {
-            onDone(messageCount, costUsd, usage, extractedModel, lastContextTokens);
+            if (!apiErrorText)
+                apiErrorText = message;
+            onError(describeCliFailure(apiErrorText, sawAuthFailure), usage, extractedModel);
+            return;
         }
-    });
-    proc.on('error', (err) => {
         active.delete(conversationId);
-        if (err.code === 'ENOENT') {
-            onError('claude CLI not found. Please install Claude Code: https://claude.ai/code', usage, extractedModel);
+        if (killedForViolation)
+            return; // already errored
+        if (abort.signal.aborted)
+            return; // superseded/cancelled mid-stream
+        if (sawAuthFailure || (sawApiError && !producedRealOutput)) {
+            onError(describeCliFailure(apiErrorText, sawAuthFailure), usage, extractedModel);
         }
         else {
-            onError(err.message, usage, extractedModel);
+            onDone(messageCount, costUsd, usage, extractedModel, lastContextTokens);
         }
-    });
+    })();
 }
-function killAll() {
-    for (const [, proc] of active) {
-        proc.kill('SIGTERM');
+export function killAll() {
+    for (const [, abort] of active) {
+        abort.abort();
     }
     active.clear();
 }
 /**
  * Stop one in-flight turn (the Stop button, relayed by the gateway as a
- * `cancel` frame). Kills the running Claude Code process for this conversation
- * and cancels any scheduled retry, so no further stream events are produced.
- * Mirrors the supersede path in spawnClaude. Returns true if something was
- * actually stopped. Mechanical only — no product logic lives here.
+ * `cancel` frame). Aborts the running query() for this conversation so no
+ * further stream events are produced. Mirrors the supersede path in spawnClaude.
+ * Returns true if something was actually stopped. Mechanical only — no product
+ * logic lives here.
  */
-function cancelConversation(conversationId) {
-    let stopped = false;
-    const proc = active.get(conversationId);
-    if (proc) {
-        proc.kill('SIGTERM');
+export function cancelConversation(conversationId) {
+    const abort = active.get(conversationId);
+    if (abort) {
+        abort.abort();
         active.delete(conversationId);
-        stopped = true;
-    }
-    const pending = pendingRetries.get(conversationId);
-    if (pending) {
-        clearTimeout(pending);
-        pendingRetries.delete(conversationId);
-        stopped = true;
+        return true;
     }
-    return stopped;
+    return false;
 }