pikiclaw 0.3.82 → 0.3.84

package/dist/agent/drivers/claude-tui.js

@@ -40,11 +40,11 @@ import fs from 'node:fs';
 import path from 'node:path';
 import { randomUUID } from 'node:crypto';
 import { tmpdir } from 'node:os';
-import { Q, agentLog, agentWarn, buildStreamPreviewMeta, computeContext, joinErrorMessages, emitSessionIdUpdate, normalizeClaudeModelId, pushRecentActivity, summarizeClaudeToolUse, summarizeClaudeToolResult, previewToolCallInput, previewToolCallResult, detectClaudeApiError, } from '../utils.js';
+import { Q, agentLog, agentWarn, buildStreamPreviewMeta, computeContext, joinErrorMessages, emitSessionIdUpdate, normalizeClaudeModelId, pushRecentActivity, summarizeClaudeToolUse, summarizeClaudeToolResult, previewToolCallInput, previewToolCallResult, detectClaudeApiError, detectClaudeModelError, claudeModelErrorMessage, } from '../utils.js';
 import { encodePathAsDirName, getHome, whichSync } from '../../core/platform.js';
 import { createRetainedLogSink } from '../../core/logging.js';
 import { stripAnsiEscapes } from '../../core/utils.js';
-import { AGENT_STREAM_HARD_KILL_GRACE_MS, CLAUDE_TUI_STALL_QUIET_MS, CLAUDE_TUI_STALL_PENDING_TOOL_MS, CLAUDE_TUI_STALL_PTY_DEAD_MS, CLAUDE_TUI_STOP_HOLD_QUIET_TTL_MS, } from '../../core/constants.js';
+import { AGENT_STREAM_HARD_KILL_GRACE_MS, CLAUDE_TUI_STALL_QUIET_MS, CLAUDE_TUI_STALL_PENDING_TOOL_MS, CLAUDE_TUI_STALL_PTY_DEAD_MS, CLAUDE_TUI_STOP_HOLD_QUIET_TTL_MS, CLAUDE_TUI_MODEL_ERROR_SETTLE_MS, } from '../../core/constants.js';
 import { claudeParse, createClaudeStreamState, claudeContextWindowFromModel, claudeEffectiveContextWindow, registerClaudeBackgroundAgentLaunch, pendingClaudeBackgroundAgentCount, registerClaudeBackgroundBashLaunch, pendingClaudeBackgroundBashCount, extractClaudeBackgroundTaskId, extractClaudeWorkflowRunId, claudeEffortAndWorkflowArgs, scrubClaudeSessionContextEnv, } from './claude.js';
 // ---------------------------------------------------------------------------
 // Stall diagnostics (capture-only)
@@ -1080,6 +1080,7 @@ export async function doClaudeTuiStream(opts) {
     let exitSignal = null;
     let terminalLimitNotice = null;
     let terminalLimitNoticeAt = 0;
+    let terminalModelError = null;
     let proc;
     const emit = () => {
         try {
@@ -1116,6 +1117,36 @@ export async function doClaudeTuiStream(opts) {
         s.activity = s.recentActivity.join('\n');
         emit();
     };
+    // Selected-model-unavailable notice (404 model_not_found). Unlike the limit
+    // banner this is terminal AND invisible to every structured signal: the TUI
+    // paints it to the PTY screen, writes nothing to the JSONL, and fires no Stop
+    // hook — so the turn would otherwise idle at the REPL until the 3–10 min stall
+    // watchdog kills it with a misleading "CLI freeze" message. We surface the
+    // real reason and end the turn now. The banner is still EVIDENCE, not a bare
+    // verdict: a short settle confirms nothing substantive followed (cross-
+    // validating the screen scrape) before we kill — never granting a lone text
+    // match the authority to shoot a healthy turn.
+    const noteTerminalModelError = (notice) => {
+        if (terminalModelError)
+            return;
+        terminalModelError = notice;
+        agentWarn(`[claude-tui] model unavailable observed (settling before terminate): ${notice}`);
+        pushRecentActivity(s.recentActivity, notice);
+        s.activity = s.recentActivity.join('\n');
+        emit();
+        setTimeout(() => {
+            if (processExited || interrupted)
+                return;
+            const hadOutput = !!s.text.trim()
+                || lastAssistantEventAt > 0 || lastSidecarEventAt > 0 || lastToolEventAt > start;
+            if (hadOutput) {
+                agentWarn('[claude-tui] model-unavailable banner was followed by real output — not terminating');
+                return;
+            }
+            agentWarn('[claude-tui] model unavailable confirmed (no JSONL/tool/Stop activity) — terminating turn');
+            killProc('SIGTERM');
+        }, CLAUDE_TUI_MODEL_ERROR_SETTLE_MS);
+    };
     // Simulated streaming. See TuiStreamBuffer / applyAssistantStreaming above.
     const streamBuf = makeTuiStreamBuffer();
     const scheduleStreamTick = () => {
@@ -1359,6 +1390,13 @@ export async function doClaudeTuiStream(opts) {
             if (notice)
                 noteTerminalLimitNotice(notice);
         }
+        // Selected-model-unavailable notice — see noteTerminalModelError. The TUI
+        // only paints this to the screen (no JSONL, no Stop hook), so the live
+        // screen tail is the sole signal. detectClaudeModelError is whitespace-
+        // insensitive so it survives the TUI's char-by-char paint.
+        if (!terminalModelError && detectClaudeModelError(screenTail)) {
+            noteTerminalModelError(claudeModelErrorMessage(s.model || opts.claudeModel || null));
+        }
     });
     // 7. Abort handling.
     const abortStream = () => {
@@ -1841,28 +1879,39 @@ export async function doClaudeTuiStream(opts) {
                     screenSample: stallScreen.sample,
                 });
                 if (!s.errors) {
-                    // Limit-notice arbitration first: a turn that showed a limit banner
-                    // and then produced nothing substantive didn't freeze — the limit
-                    // ate it. Label it rate_limit with the banner's own text (which
-                    // carries the reset time) so the user gets the real reason, and so
-                    // doClaudeWithRetry doesn't auto-resume into the same wall.
-                    const limitOutcome = resolveClaudeTuiLimitOutcome({
-                        noticeText: terminalLimitNotice,
-                        noticeAt: terminalLimitNoticeAt,
-                        lastSubstantiveEventAt: Math.max(lastAssistantEventAt, lastToolEventAt, lastSidecarEventAt),
-                        hasOutputText: !!s.text.trim(),
-                    });
-                    if (limitOutcome === 'fatal') {
-                        s.stopReason = 'rate_limit';
-                        s.errors = [terminalLimitNotice];
+                    if (terminalModelError && !s.text.trim()) {
+                        // Model-unavailable arbitration first: if the stall watchdog beat the
+                        // settle-timer (noteTerminalModelError) to the kill, the turn didn't
+                        // freeze — the selected model is disabled / no-access. Surface the real
+                        // reason; stopReason 'model_error' is non-retryable so doClaudeWithRetry
+                        // won't auto-resume into the same dead model.
+                        s.stopReason = 'model_error';
+                        s.errors = [terminalModelError];
                     }
                     else {
-                        // Be honest about which kind of stall this is. looksLikePrompt here
-                        // means the auto-answer (detectClaudeProceedPrompt) did NOT clear an
-                        // interactive prompt — so it's a blocking dialog, not the CLI freeze.
-                        s.errors = [stallScreen.looksLikePrompt
-                                ? `Claude blocked mid-turn on an interactive prompt (PTY quiet ${ptyQuietS}s) that auto-answer couldn't clear. Terminated for auto-resume.`
-                                : `Claude process went silent mid-turn for ${quietMin}m (no JSONL, hook, or sub-agent events; PTY quiet ${ptyQuietS}s) — known claude CLI freeze. Terminated for auto-resume.`];
+                        // Limit-notice arbitration: a turn that showed a limit banner and then
+                        // produced nothing substantive didn't freeze — the limit ate it. Label
+                        // it rate_limit with the banner's own text (which carries the reset
+                        // time) so the user gets the real reason, and so doClaudeWithRetry
+                        // doesn't auto-resume into the same wall.
+                        const limitOutcome = resolveClaudeTuiLimitOutcome({
+                            noticeText: terminalLimitNotice,
+                            noticeAt: terminalLimitNoticeAt,
+                            lastSubstantiveEventAt: Math.max(lastAssistantEventAt, lastToolEventAt, lastSidecarEventAt),
+                            hasOutputText: !!s.text.trim(),
+                        });
+                        if (limitOutcome === 'fatal') {
+                            s.stopReason = 'rate_limit';
+                            s.errors = [terminalLimitNotice];
+                        }
+                        else {
+                            // Be honest about which kind of stall this is. looksLikePrompt here
+                            // means the auto-answer (detectClaudeProceedPrompt) did NOT clear an
+                            // interactive prompt — so it's a blocking dialog, not the CLI freeze.
+                            s.errors = [stallScreen.looksLikePrompt
+                                    ? `Claude blocked mid-turn on an interactive prompt (PTY quiet ${ptyQuietS}s) that auto-answer couldn't clear. Terminated for auto-resume.`
+                                    : `Claude process went silent mid-turn for ${quietMin}m (no JSONL, hook, or sub-agent events; PTY quiet ${ptyQuietS}s) — known claude CLI freeze. Terminated for auto-resume.`];
+                        }
                     }
                 }
                 agentWarn(`[claude-tui] stall detected: no progress for ${quietMin}m (pendingTools=${pendingHookToolIds.size}, ptyQuiet=${ptyQuietS}s) — terminating TUI pid=${proc.pid}${s.stopReason === 'rate_limit' ? ' (usage limit)' : ' for auto-resume'}`);
@@ -1942,6 +1991,16 @@ export async function doClaudeTuiStream(opts) {
         if (!s.errors)
             s.errors = [`Anthropic API error: ${apiErrorReason}`];
     }
+    // Model-unavailable arbitration: the TUI painted the "selected model is
+    // unavailable" banner (noteTerminalModelError) and the turn produced nothing
+    // — no JSONL, no Stop hook. The early settle-timer's SIGTERM (or any exit)
+    // brought us here; surface the real reason instead of a bare "(no textual
+    // response)". stopReason 'model_error' is non-retryable (doClaudeWithRetry
+    // only auto-resumes 'stalled'), so we never loop on the same dead model.
+    if (!interrupted && !s.errors && terminalModelError && !s.text.trim()) {
+        s.stopReason = 'model_error';
+        s.errors = [terminalModelError];
+    }
     // Limit-notice arbitration (see resolveClaudeTuiLimitOutcome). Covers the
     // paths the stall watchdog never reaches: the TUI painted a limit banner,
     // then Stop fired on an empty turn or the process exited — nothing

package/dist/agent/drivers/claude.js

@@ -8,7 +8,7 @@ import { createInterface } from 'node:readline';
 import { registerDriver } from '../driver.js';
 import { 
 // shared helpers
-Q, run, agentError, agentLog, agentWarn, buildStreamPreviewMeta, computeContext, pushRecentActivity, summarizeClaudeToolUse, summarizeClaudeToolResult, joinErrorMessages, parseTodoWriteAsPlan, previewToolCallInput, previewToolCallResult, detectClaudeApiError, isRetryableClaudeApiError, emitSessionIdUpdate, IMAGE_EXTS, mimeForExt, listPikiclawSessions, mergeManagedAndNativeSessions, readTailLines, stripInjectedPrompts, sanitizeSessionUserPreviewText, SESSION_PREVIEW_IMAGE_PLACEHOLDER_RE, CLAUDE_AT_MENTION_IMAGE_RE, extractClaudeAtMentionImagePaths, attachAgentImage, applyTurnWindow, shortValue, roundPercent, modelFamily, normalizeClaudeModelId, emptyUsage, normalizeUsageStatus, collapseSkillPrompt, } from '../index.js';
+Q, run, agentError, agentLog, agentWarn, buildStreamPreviewMeta, computeContext, pushRecentActivity, summarizeClaudeToolUse, summarizeClaudeToolResult, joinErrorMessages, parseTodoWriteAsPlan, previewToolCallInput, previewToolCallResult, detectClaudeApiError, isRetryableClaudeApiError, detectClaudeModelError, claudeModelErrorMessage, emitSessionIdUpdate, IMAGE_EXTS, mimeForExt, listPikiclawSessions, mergeManagedAndNativeSessions, readTailLines, stripInjectedPrompts, sanitizeSessionUserPreviewText, SESSION_PREVIEW_IMAGE_PLACEHOLDER_RE, CLAUDE_AT_MENTION_IMAGE_RE, extractClaudeAtMentionImagePaths, attachAgentImage, applyTurnWindow, shortValue, roundPercent, modelFamily, normalizeClaudeModelId, emptyUsage, normalizeUsageStatus, collapseSkillPrompt, } from '../index.js';
 import { AGENT_STREAM_HARD_KILL_GRACE_MS, AGENT_GRACEFUL_ABORT_GRACE_MS, SESSION_RUNNING_THRESHOLD_MS } from '../../core/constants.js';
 import { terminateProcessTree } from '../../core/process-control.js';
 import { getHome, IS_MAC, encodePathAsDirName } from '../../core/platform.js';
@@ -652,8 +652,22 @@ export function claudeParse(ev, s) {
         // model error, …), not real Claude output. The historical jsonl reader
         // converts them into `system_notice` blocks; on the live stream we just
         // drop them so they don't pollute s.text / s.thinking.
-        if (msg.model === '<synthetic>')
+        if (msg.model === '<synthetic>') {
+            // …except the "selected model is unavailable" notice (404 model_not_found):
+            // a hard, non-retryable failure. The result event's text fallback below
+            // also catches it, but recording s.errors here upgrades the turn from a
+            // bare "(no textual response)" reply to a clear error + non-retryable
+            // stopReason (so doClaudeWithRetry won't loop on the same dead model).
+            if (!s.errors) {
+                const synthText = (msg.content || [])
+                    .filter((b) => b?.type === 'text').map((b) => b.text || '').join(' ');
+                if (ev.error === 'model_not_found' || detectClaudeModelError(synthText)) {
+                    s.stopReason = 'model_error';
+                    s.errors = [claudeModelErrorMessage(s.model)];
+                }
+            }
             return;
+        }
         const contents = msg.content || [];
         const th = contents.filter((b) => b?.type === 'thinking').map((b) => b.thinking || '').join('\n\n');
         const tx = contents.filter((b) => b?.type === 'text').map((b) => b.text || '').join('\n\n');
@@ -867,7 +881,11 @@ export function claudeParse(ev, s) {
             s.errors = ev.errors;
         if (ev.result && !s.text.trim())
             s.text = ev.result;
-        s.stopReason = ev.stop_reason ?? s.stopReason;
+        // A model-unavailable turn carries a normal stop_reason on its result event
+        // (e.g. 'stop_sequence') that would otherwise clobber the 'model_error' the
+        // synthetic handler set — preserve it so the failure stays diagnosable.
+        if (s.stopReason !== 'model_error')
+            s.stopReason = ev.stop_reason ?? s.stopReason;
         const u = ev.usage;
         if (u) {
             // Per-call semantics: the last message_start/message_delta snapshot is
@@ -2171,6 +2189,16 @@ const CLAUDE_MODELS = [
 // ---------------------------------------------------------------------------
 // Usage
 // ---------------------------------------------------------------------------
+// The account-usage query below hits api.anthropic.com/api/oauth/usage, which is
+// itself rate-limited. The dashboard rebuilds agent status ~every 30s (plus a
+// forced refresh on usage-ring hover), and querying that often trips the
+// endpoint's 429 — which (since we treat a query error as "unknown") blanks the
+// header usage ring entirely. Quota windows (5h/7d) move slowly, so we query at
+// most once per this interval and serve the last good result in between
+// (including across transient 429s), decoupling usage cadence from how often
+// agent status is rebuilt.
+const CLAUDE_USAGE_QUERY_TTL_MS = 5 * 60_000;
+const claudeUsageCache = { lastGood: null, lastAttemptAt: 0 };
 function getClaudeOAuthToken() {
     // `security` is macOS-only; other platforms store Claude creds differently
     // (DPAPI on Windows, libsecret on Linux) and Claude Code manages those itself.
@@ -2573,9 +2601,25 @@ class ClaudeDriver {
         const home = getHome();
         if (!home)
             return emptyUsage('claude', 'HOME is not set.');
-        return getClaudeUsageFromOAuth()
-            || getClaudeUsageFromTelemetry(home, opts.model)
+        const telemetry = () => getClaudeUsageFromTelemetry(home, opts.model)
             || emptyUsage('claude', 'No recent Claude usage data found.');
+        // Throttle the rate-limited OAuth usage query (see CLAUDE_USAGE_QUERY_TTL_MS).
+        // Within the window we reuse the last good result rather than re-querying on
+        // every agent-status rebuild, so a transient query-API 429 can't blank the
+        // ring between successful polls.
+        const now = Date.now();
+        if (now - claudeUsageCache.lastAttemptAt < CLAUDE_USAGE_QUERY_TTL_MS) {
+            return claudeUsageCache.lastGood ?? telemetry();
+        }
+        claudeUsageCache.lastAttemptAt = now;
+        const oauth = getClaudeUsageFromOAuth();
+        if (oauth) {
+            claudeUsageCache.lastGood = oauth;
+            return oauth;
+        }
+        // OAuth unavailable (non-mac, no token, or transient 429): keep showing the
+        // last good windows if we have any; otherwise fall back to telemetry.
+        return claudeUsageCache.lastGood ?? telemetry();
     }
     async deleteNativeSession(workdir, sessionId) {
         const file = claudeSessionTranscriptPath(workdir, sessionId);

package/dist/agent/index.js

@@ -21,7 +21,7 @@ export { IMAGE_EXTS } from './types.js';
 // ── Re-export: image pipeline ──────────────────────────────────────────────
 export { attachAgentImage, attachInlineImage, materializeImage, rewriteImageBlocksForTransport, resolveAllowedAttachmentPath, allowedAttachmentRoots, decodeAttachmentPathParam, sessionAttachmentsDir, codexHome, } from './images.js';
 // ── Re-export: utilities ────────────────────────────────────────────────────
-export { Q, agentLog, agentWarn, agentError, dedupeStrings, numberOrNull, normalizeStreamPreviewPlan, parseTodoWriteAsPlan, normalizeActivityLine, pushRecentActivity, detectClaudeApiError, isRetryableClaudeApiError, firstNonEmptyLine, shortValue, normalizeErrorMessage, joinErrorMessages, appendSystemPrompt, mimeForExt, computeContext, buildStreamPreviewMeta, summarizeClaudeToolUse, summarizeClaudeToolResult, previewToolCallInput, previewToolCallResult, roundPercent, toIsoFromEpochSeconds, normalizeUsageStatus, labelFromWindowMinutes, usageWindowFromRateLimit, parseJsonTail, modelFamily, normalizeClaudeModelId, emptyUsage, readTailLines, stripInjectedPrompts, sanitizeSessionUserPreviewText, SESSION_PREVIEW_IMAGE_PLACEHOLDER_RE, CLAUDE_AT_MENTION_IMAGE_RE, extractClaudeAtMentionImagePaths, stripClaudeAtMentionImages, isPendingSessionId, emitSessionIdUpdate, sessionListDisplayTitle, } from './utils.js';
+export { Q, agentLog, agentWarn, agentError, dedupeStrings, numberOrNull, normalizeStreamPreviewPlan, parseTodoWriteAsPlan, normalizeActivityLine, pushRecentActivity, detectClaudeApiError, isRetryableClaudeApiError, detectClaudeModelError, claudeModelErrorMessage, firstNonEmptyLine, shortValue, normalizeErrorMessage, joinErrorMessages, appendSystemPrompt, mimeForExt, computeContext, buildStreamPreviewMeta, summarizeClaudeToolUse, summarizeClaudeToolResult, previewToolCallInput, previewToolCallResult, roundPercent, toIsoFromEpochSeconds, normalizeUsageStatus, labelFromWindowMinutes, usageWindowFromRateLimit, parseJsonTail, modelFamily, normalizeClaudeModelId, emptyUsage, readTailLines, stripInjectedPrompts, sanitizeSessionUserPreviewText, SESSION_PREVIEW_IMAGE_PLACEHOLDER_RE, CLAUDE_AT_MENTION_IMAGE_RE, extractClaudeAtMentionImagePaths, stripClaudeAtMentionImages, isPendingSessionId, emitSessionIdUpdate, sessionListDisplayTitle, } from './utils.js';
 // ── Re-export: session management ───────────────────────────────────────────
 export { updateSessionMeta, promoteSessionId, recordFork, listPikiclawSessions, findPikiclawSession, getSessionStoredConfig, ensureManagedSession, findManagedThreadSession, stageSessionFiles, mergeManagedAndNativeSessions, getSessions, getSessionTail, getSessionMessages, applyTurnWindow, applyTurnFilter, classifySession, deriveUserStatus, exportSession, importSession, deleteAgentSession, isProcessAlive, isRunningSessionStale, reconcileOrphanedRunningSessions, } from './session.js';
 // ── Re-export: stream & detection ───────────────────────────────────────────

package/dist/agent/utils.js

@@ -204,6 +204,37 @@ export function isRetryableClaudeApiError(reason) {
         return false;
     return /overloaded|overload|timeout|timed out|500|502|503|504|529|temporar|gateway|connection|network|internal (server )?error/i.test(r);
 }
+/**
+ * Detect Claude Code's "selected model is unavailable" notice — emitted when
+ * the requested `--model` id is disabled / not provisioned for the account (a
+ * 404 model_not_found). Its delivery differs by mode:
+ *   - `-p`/stream-json: a `<synthetic>` assistant event carrying
+ *     `error:"model_not_found"` plus the banner text, and a `result` event with
+ *     `is_error` — both reach the parser.
+ *   - TUI: the banner is *only* painted to the PTY screen. It is never written
+ *     to the transcript JSONL and fires no Stop hook (verified 2026-06-13), so
+ *     the screen scrape is the sole signal and the turn would otherwise hang
+ *     until the stall watchdog (3–10 min).
+ *
+ * Matching is whitespace-insensitive on purpose: the TUI renders the banner
+ * character-by-character with cursor positioning, so after ANSI stripping the
+ * words lose their spaces and wrap arbitrarily ("issuewiththeselectedmodel").
+ * Collapsing whitespace on both sides makes the match survive that rendering.
+ * Callers compose the user-facing message via `claudeModelErrorMessage` with
+ * the concrete model id they hold.
+ */
+export function detectClaudeModelError(text) {
+    if (!text)
+        return false;
+    const collapsed = text.replace(/\s+/g, '').toLowerCase();
+    return collapsed.includes('issuewiththeselectedmodel')
+        || collapsed.includes('maynotexistoryoumaynothaveaccess');
+}
+/** User-facing message for an unavailable / no-access model (see {@link detectClaudeModelError}). */
+export function claudeModelErrorMessage(model) {
+    const id = (model || '').trim();
+    return `The selected model${id ? ` (${id})` : ''} is unavailable — it may not exist, or this account doesn't have access to it. Switch to a different model in pikiclaw settings.`;
+}
 export function appendSystemPrompt(base, extra) {
     const lhs = String(base || '').trim();
     const rhs = String(extra || '').trim();

package/dist/core/constants.js

@@ -318,6 +318,15 @@ export const CLAUDE_TUI_STALL_PENDING_TOOL_MS = 30 * 60_000;
  * refreshes the PTY signal and defers this path to the slow thresholds.
  */
 export const CLAUDE_TUI_STALL_PTY_DEAD_MS = 3 * 60_000;
+/**
+ * Settle window after the TUI paints the "selected model is unavailable" banner
+ * (a 404 model_not_found). The notice is terminal — claude paints it then idles
+ * at the REPL forever: no JSONL is written, no Stop hook fires. We wait this
+ * brief window to cross-validate that nothing substantive followed (the banner
+ * alone is evidence, not a verdict — same discipline as resolveClaudeTuiLimitOutcome)
+ * before ending the turn, instead of waiting out the 3–10 minute stall watchdog.
+ */
+export const CLAUDE_TUI_MODEL_ERROR_SETTLE_MS = 2_500;
 /**
  * TTL for the post-Stop `hold-background` path. The hold protects
  * run_in_background agents living inside the claude process — but a live

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pikiclaw",
-  "version": "0.3.82",
+  "version": "0.3.84",
   "description": "Put the world's smartest AI agents in your pocket. Command local Claude & Gemini via IM. | 让最好用的 IM 变成你电脑上的顶级 Agent 控制台",
   "type": "module",
   "bin": {