Symptom: when an AI request is proxied through nginx (or any gateway)
and the request body exceeds client_max_body_size, the proxy returns a
413 HTML error page. The Vercel AI SDK then fails to parse the HTML
as a chat completion and surfaces a cryptic Zod validation error like
"Expected 'id' to be a string." through the UI — users have no idea
what's wrong.
Root cause: classifyError only did light sanitization and returned the
raw SDK message. It also string-coerced the error before inspection, so
the structured statusCode / responseBody fields that APICallError
attaches were thrown away.
Fix: classifyError now accepts `unknown` and inspects the full error
shape. Adds explicit branches for:
- HTTP 413 (from statusCode, cause.statusCode, or message text) →
"Request too large — exceeded proxy size limit. Try shorter
message, fewer attachments, or raise client_max_body_size."
- HTTP 502/503/504 → retryable upstream-gateway message
- HTML response body (starts with <!DOCTYPE/<html> or contains such
tags anywhere) → "Server returned HTML error page, likely a proxy
intercept."
- Zod/schema parse shapes ("Expected 'X' to be …", "Invalid JSON
response", "Type validation failed") → "Response could not be
parsed; proxy may have replaced/truncated the body."
In every classified case the raw SDK text is still appended ("Raw: …")
so users can report the underlying error verbatim.
useAIChatStreaming.ts callers now pass the raw error to classifyError
instead of `.message`, so the new structured branches actually fire.
Also wired infrastructure/ai/*.test.ts into the npm test glob.
Closes#765
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Apply the same fix as #706 to the Claude Code agent. The `claude` CLI has
its own auth surface (`claude auth login/logout/status`) that manages
subscription-based logins (Claude Max / Pro via claude.ai) alongside
ANTHROPIC_API_KEY / settings-based configs. Silently forwarding a
netcatty-configured provider's API key to claude-agent-acp overrides that
login — the user's subscription gets bypassed and charges go to their API
balance without their knowledge.
Claude's settings card never surfaced the `claude auth status` so this
regression was more hidden than the Codex one, but the underlying coupling
is the same class of bug.
Changes:
- Stop forwarding any providerId for managed ACP agents from the renderer;
claude-agent-acp now resolves auth purely from its own CLI config / login
state / shell env.
- Remove ANTHROPIC_API_KEY and ANTHROPIC_BASE_URL injection at all three
codex-acp / claude-acp spawn sites in aiBridge.
- Drop Claude from the authFingerprint computation (it no longer has any
netcatty-side input to hash).
- Delete the now-unused `findManagedAgentProvider` helper and its
ProviderConfig import from managedAgents.ts.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Codex agent auth must be determined entirely by ~/.codex/auth.json or
~/.codex/config.toml. Before this change, if the user configured any
OpenAI-compatible API provider in netcatty settings (for Catty agent use),
useAIChatStreaming would silently hand that provider's apiKey to the Codex
agent too, causing aiBridge to spawn codex-acp with authMethodId
"codex-api-key" and completely override the user's ChatGPT login.
The regression was introduced in PR #702 (v1.0.89) when findManagedAgent
Provider started matching generic "custom" providers for Codex. Users who
logged into Codex via ChatGPT and also had a netcatty-configured custom
provider saw the UI flip to "API mode" on refresh and their ChatGPT
session get ignored.
Remove the codex branch from the agentProviderId resolver and from
findManagedAgentProvider itself. Also drop the now-meaningless
hasCompatibleProvider hint on the Codex settings card and its i18n copy.
Claude agent behavior is unchanged.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
A generic custom provider (OpenAI-compatible) could be selected for
Claude, passing wrong credentials. Now we prefer an explicit anthropic
provider and only fall back to a custom provider when it has a baseURL
configured (indicating intentional Anthropic-compatible gateway use).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The renderer only resolved OpenAI providers (for Codex) when passing
provider IDs to the main process. Claude agent was never matched, so
no API key was injected. Additionally, the main process only injected
CODEX_API_KEY — never ANTHROPIC_API_KEY or ANTHROPIC_BASE_URL.
Changes:
- Renderer now resolves anthropic/custom provider for Claude agent,
openai provider for Codex agent (via matchesManagedAgentConfig)
- Main process injects ANTHROPIC_API_KEY and ANTHROPIC_BASE_URL into
claude-agent-acp env when a provider is configured, across all three
ACP provider creation paths (list-models, stream, fallback)
This enables users who configure an Anthropic provider with a custom
base URL (e.g. CC Switch proxy) to use Claude Code without being
redirected to the official OAuth flow.
Closes#677
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: add deviceType field to Host model for network device support
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: pass deviceType through session metadata pipeline
Thread deviceType from Host model through AITerminalSessionInfo, IPC
types, and mcpServerBridge so AI agents can inspect device type per session.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat: route network device SSH sessions to raw PTY execution
When deviceType === 'network', handleExec now uses execViaRawPty
instead of execViaPty so vendor CLIs (Huawei VRP, Cisco IOS, etc.)
receive commands as-is without POSIX shell wrapping or markers.
The command blocklist is also skipped for network devices, consistent
with the existing serial session bypass. AI context description updated
to document the raw-execution behaviour for network device sessions.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat: add network device mode toggle to host settings UI
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: add network device awareness to Catty Agent system prompt
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: extend network device mode to Catty Agent exec path and host context
- Add network device detection and raw execution routing to aiBridge.cjs
(the primary Catty Agent command path), not just the MCP bridge
- Export getSessionMeta from mcpServerBridge for reuse in aiBridge
- Surface deviceType in Catty Agent system prompt host list so the AI
can identify which sessions are network devices
- Pass deviceType through buildSystemPrompt context
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: exempt network device sessions from client-side blocklist and update ACP context
- Add deviceType to ExecutorContext sessions type
- Skip renderer-side command blocklist for deviceType=network sessions
in shared toolExecutors.ts (not just main-process side)
- Update ACP agent context hint to mention network device sessions
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: only show network device mode toggle for SSH hosts
Telnet and local hosts don't support the network device execution path,
so hiding the toggle prevents users from enabling a broken configuration.
Serial hosts already use raw mode by default.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: exclude Mosh sessions from network device raw execution path
Mosh uses a shell-backed PTY and cannot connect to vendor CLIs, so
network device mode should only apply to SSH and serial sessions.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: prefer session.protocol over metadata for Mosh detection
Mosh tabs report protocol:"ssh" in renderer metadata but "mosh" in
the main-process session object. Prioritize session.protocol (runtime
truth) to correctly exclude Mosh from network device raw execution.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: suppress deviceType metadata for Mosh sessions
Mosh requires a shell-backed PTY and cannot connect to vendor CLIs,
so omit deviceType from AI-facing metadata when session is Mosh-backed.
This prevents the AI from being told to use vendor CLI syntax when the
actual execution path uses normal shell wrapping.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: use exit code 0 for network device sessions and hide toggle for Mosh
- Network device / serial sessions return exitCode: null from vendor
CLIs. Default to 0 instead of -1 so the AI doesn't misinterpret
successful commands as failures.
- Hide the network device mode toggle when Mosh is enabled, since
the setting is suppressed at runtime for Mosh sessions anyway.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: preserve null exit codes and restrict raw mode to SSH/serial
- Preserve exitCode: null for network device sessions instead of
coercing to 0, so the AI knows exit status is unavailable rather
than seeing a misleading success code.
- Explicitly whitelist SSH/serial protocols for network device mode
instead of just excluding mosh, preventing local/telnet sessions
from accidentally entering raw execution.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: use UTF-8 encoding for SSH network device raw execution
execViaRawPty hardcodes latin1 for serial port data decoding. Add an
encoding option (default: latin1) and pass utf8 from SSH network
device call sites so multi-byte characters aren't corrupted.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: use host charset for serial port decoding instead of hardcoded latin1
- Extract charsetToNodeEncoding() to module scope in terminalBridge
- Serial sessions now read options.charset (from Host.charset) for
both terminal display decoding and AI command output
- Store serialEncoding on session object so exec paths can use it
- Pass encoding through all execViaRawPty call sites
- Default encoding changed from latin1 to utf8 (matches most modern
network equipment and is the safer default for CJK environments)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: move serialEncoding declaration before session object creation
serialEncoding was referenced in the session object literal before its
const declaration, causing a TDZ ReferenceError that would crash every
serial connection.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: tighten isNetworkDevice logic and clean up edge cases
- Align toolExecutors isNetworkDevice check with bridge logic: require
explicit SSH/serial protocol match instead of trusting deviceType alone
- Remove empty-string protocol match from isSshOrSerial in both bridges
to prevent local/unknown sessions from being treated as network devices
- Widen exitCode return type to `number | null` to match actual behavior
- Clear deviceType when enabling Mosh (incompatible combination)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: update MCP server tool descriptions for network device sessions
The get_environment and terminal_execute tool descriptions only
mentioned serial/raw sessions for network devices. Updated to also
reference deviceType: network SSH sessions so external AI agents
(Claude, Codex) know about the new execution mode.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: include deviceType in get_environment and guard execViaChannel fallback
- Add deviceType to executeWorkspaceGetInfo session mapping and return
type so Catty Agent's get_environment tool matches MCP bridge output
- Guard both aiBridge and mcpServerBridge against falling through to
execViaChannel for network device sessions — network devices require
an interactive PTY and exec channels would produce broken behavior
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: add charset setting to serial host configuration UI
Serial hosts now have a charset input in the Advanced section,
defaulting to UTF-8. The value is saved to Host.charset and used
by the serial decoder in terminalBridge.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: add charset to serial quick-connect modal with full pipeline
- Add charset input to SerialConnectModal (Advanced section)
- Thread charset through onConnect callback → handleConnectSerial →
createSerialSession → TerminalSession.charset
- Add charset field to TerminalSession interface
- Include charset in fallback host builder for quick-connect sessions
so createTerminalSessionStarters can pass it to startSerialSession
- Saved hosts also store charset via onSaveHost
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: constrain serial connect modal height with scrollable content
Modal content could overflow the viewport when Advanced section was
expanded. Add max-h-[85vh] to DialogContent with flex layout so the
content area scrolls while header and footer buttons stay visible.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: propagate charset through all serial session creation paths
- Add charset to startSerialSession type in global.d.ts
- Copy host.charset to TerminalSession in connectToHost serial path
- Copy host.charset in createWorkspaceWithHosts serial path
- Propagate session.charset in splitSession (both workspace and standalone)
- Propagate session.charset in copySession
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: propagate charset in remaining session creation paths
Add host.charset to connectToHost (non-serial), createWorkspaceWithHosts
(non-serial), and runSnippet session creation for consistency.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: expose advanced AI model parameters in provider settings (#532)
Add collapsible "Advanced Parameters" section to provider config with
optional max_tokens, temperature, top_p, frequency_penalty, and
presence_penalty fields. Parameters are merged into streamText() calls
only when explicitly set, otherwise provider defaults apply.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: use maxOutputTokens instead of maxTokens for ai@6 SDK
The streamText CallSettings in ai@6 expects maxOutputTokens, not
maxTokens. Without this fix the user's max_tokens setting is silently
ignored.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: allow negative penalty input and clamp params on save
- Use raw string state for penalty fields so typing "-" is not
discarded before the digit is entered
- Clamp all parameters to valid ranges on save (temperature 0-2,
topP 0-1, penalties -2 to 2)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: use raw string state for all numeric advanced param inputs
Prevents intermediate text like "0." from being normalized to "0"
during keyboard entry of decimal values for temperature, topP, and
maxTokens fields.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: clamp max_tokens to minimum of 1 after rounding
Prevents Math.round(0.4) = 0 from being persisted and causing
streamText to reject with "maxOutputTokens must be >= 1".
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: reject non-finite max_tokens before persisting
Guard with Number.isFinite to prevent Infinity from being stored
and forwarded to streamText.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: inline approval gate for tool execution
Replace SDK-level needsApproval with Promise-based approval gate inside
tool execute functions. The SDK stream stays alive while the UI shows
inline approve/reject buttons on ToolCall blocks.
Changes:
- Add approvalGate.ts: Promise-based approval system with event listeners
- tools.ts: requestApproval() inside execute for confirm mode
- tool-call.tsx: inline approval buttons and keyboard shortcuts
- ChatMessageList.tsx: subscribe to approval events, render approval UI
- useAIChatStreaming.ts: remove old useToolApproval hook integration
- AIChatSidePanel.tsx: remove old approval hook, clean up unused destructuring
- systemPrompt.ts: update confirm mode to not ask for text confirmation
- preload.cjs: filter pager env var prefixes from terminal display
- mcpServerBridge.cjs: add approval gate for ACP/MCP write operations
- aiBridge.cjs: wire IPC for MCP approval response and main window getter
- preload.cjs: add onMcpApprovalRequest/respondMcpApproval APIs
* fix: scope approval gate by chatSessionId and replay for late subscribers
Address Codex PR review comments:
- Add chatSessionId to ApprovalRequest for session isolation
- Scope clearAllPendingApprovals(chatSessionId?) to only clear
approvals belonging to the target session
- Add replayPendingApprovals() so late-mounting ChatMessageList
picks up approvals that fired while unmounted
- Scope MCP clearPendingApprovals in aiBridge cancel handler to
effectiveChatSessionId instead of clearing all
- Pass chatSessionId through MCP approval IPC flow
* chore: remove old approval flow code
- Delete useToolApproval.ts (unused hook)
- Delete InlineApprovalCard.tsx (replaced by ToolCall inline buttons)
- Remove stale comments referencing old hook in AIChatSidePanel
- Remove unused ai.chat.toolApprovalTitle i18n key from en/zh-CN
* fix: session-scoped approval gate and MCP replay survival
- handleStop passes activeSessionId to clearAllPendingApprovals
- setupMcpApprovalBridge stores MCP approvals in pendingApprovals map
so they survive ChatMessageList unmount/remount cycles
- ChatMessageList accepts activeSessionId prop and filters standalone
MCP approval blocks to the current session only
- AIChatSidePanel passes activeSessionId to ChatMessageList
* fix: filter PTY exec marker echoes and exit code lines from terminal
Extend filterMcpMarkers in preload.cjs to strip all shell-visible
artifacts from AI command execution:
- Echoed printf start marker: printf '%s\n' '__NCMCP_..._S'
- Echoed exit code restoration: (exit $__nc)
- PowerShell: Write-Output, $global:LASTEXITCODE, $__nc assignment
- Fish: set __nc $status
- Cmd: echo __NCMCP_...
- Widen guard to also trigger on __nc and PAGER=cat strings
* fix: scope SDK approvals, deny MCP on no renderer, fix memo comparator
- createCattyTools accepts chatSessionId and passes it to
requestApproval so SDK approvals can be matched by
clearAllPendingApprovals(activeSessionId) on stop
- useAIChatStreaming passes sessionId to createCattyTools
- mcpServerBridge: deny (resolve false) when no renderer window is
available instead of auto-approving, preserving confirm mode safety
- ChatMessageList: add activeSessionId to React.memo comparator so
switching sessions triggers re-render for correct MCP approval filter
* fix: MCP listener lifecycle, approval timeout, and UI sync on stop
- Move setupMcpApprovalBridge from ChatMessageList to AIChatSidePanel
so the IPC listener survives tab/panel switches
- Add 5-minute auto-deny timeout to requestApproval to prevent
indefinite isStreaming hangs when user walks away
- Add onApprovalCleared listener system: clearAllPendingApprovals now
notifies UI subscribers so ChatMessageList removes stale cards
- ChatMessageList subscribes to onApprovalCleared to sync local state
* fix: main-process approval timeout and full tool args in payload
- Add 5-minute auto-deny timeout to requestApprovalFromRenderer
matching the renderer-side requestApproval behavior
- Forward all tool params (excluding chatSessionId) to approval UI
instead of cherry-picking command/input/path, so sftpRename
oldPath/newPath and other tool-specific args are visible
* fix: move MCP bridge to TerminalLayer, narrow terminal filter guard
- Move setupMcpApprovalBridge from AIChatSidePanel to TerminalLayer
so the IPC listener stays alive regardless of side panel tab.
AIChatSidePanel only mounts when activeSidePanelTab==='ai'.
- Narrow preload.cjs filter guard back to __NCMCP_ only, preventing
false-positive stripping of user scripts containing __nc or PAGER=cat
* fix: eliminate PTY wrapper echo leakage and duplicate prompts
- Posix wrapper now emits 2 lines instead of 4: start marker + command
on line 1 (joined with ;), end marker + exit on line 2. This
eliminates the duplicate prompt echo from the separate start marker.
- Rename __nc to __NCMCP_rc in all shell variants (posix/fish/powershell)
so every wrapper variable contains the __NCMCP_ prefix. The preload
guard `data.includes("__NCMCP_")` now reliably catches ALL wrapper
artifacts regardless of chunk boundaries.
- Update all filterMcpMarkers regex patterns to match the restructured
wrapper format and renamed variable.
* fix: sync main-process approval timeout with renderer UI cleanup
- When requestApprovalFromRenderer times out, send IPC event
netcatty:ai:mcp:approval-cleared to renderer so stale approval
cards are removed
- Add onMcpApprovalCleared preload bridge for the new IPC channel
- setupMcpApprovalBridge now subscribes to cleared events, removes
timed-out entries from pendingApprovals and notifies clearedListeners
so ChatMessageList drops the stale card
* fix: surface denied inline approvals as errors in UI
- Detect error or denial payloads ("error" string or "ok: false")
returned by tools when the user denies an execution
- Set isError: true on the tool-result message so the ToolCall UI
renders it as a failure (red/rejected) instead of a success (green)
* Add AI support for local terminal sessions
* Fix local AI session metadata and shell safety
* Fix local session cloning and multi-exec errors
* Refactor local shell detection helpers
* Fix local shell helper import path
* Fix CJS imports in renderer
* Use ESM local shell helpers in renderer
* Normalize local shell paths and platform metadata
* feat: support multimodal attachments (images, PDFs, files) in AI chat
Previously uploaded images were displayed in the UI but never sent to
the AI model, and non-image files (PDF, text) were silently rejected.
- Rename useImageUpload → useFileUpload; accept image/*, PDF, and text/*
- Rename ChatMessageImage → ChatMessageAttachment with filePath support
- Build multimodal SDK messages (ImagePart/FilePart) for Catty Agent
- Fix ACP agent path: images inline, non-image files via local path hint
so ACP agents (Claude Code, etc.) read them with native file access
- Use Electron webUtils.getPathForFile() for reliable file path capture
- Compact user message bubble padding
Closes#294 (AI file upload issues)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: show real tool names in AI chat instead of ACP wrapper names
- Unwrap ACP dynamic tool calls in serializeStreamChunk to extract
real tool name, args, and toolCallId from chunk.input
- Simplify MCP tool name prefixes (mcp__server__tool → tool)
- Pass toolCallId from ACP tool-call events to match tool results
- Prevent onToolResult from overwriting correct names with wrapper name
- Build toolCallNames map in ChatMessageList for tool result display
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: backward-compatible fallback for legacy `images` field in chat messages
Persisted sessions may still have `images` instead of `attachments`.
Add `?? m.images` fallback in SDK message builder and renderer so
historical image attachments are not silently dropped after upgrade.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: broaden file type support and handle pasted files without path
- Accept all file types except video/audio (instead of allowlist)
so .json, .yaml, .sh, etc. are not silently rejected
- For ACP agents, save pasted/virtual files (no filePath) to temp
directory so the agent can still read them
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: use managed temp dir for pasted ACP attachments
Use tempDirBridge.getTempFilePath() instead of manual os.tmpdir() path
so pasted file attachments are tracked by the app's cleanup system.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: show real error message instead of [object Object]
When an error object (not a string or Error instance) reaches the
error display path, String(obj) produces "[object Object]". Now
extract .message from error-like objects, or JSON.stringify as fallback.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: guard JSON.stringify fallback against undefined return
JSON.stringify(undefined) returns undefined (not a string), which would
crash classifyError().toLowerCase(). Add ?? 'Unknown error' fallback.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: use non-throwing fallback for error serialization
JSON.stringify can throw on circular objects or BigInt values. Wrap in
try-catch to avoid losing the original error and leaving the stream
stuck in a streaming state.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Remove multi_host_execute tool — the AI can call terminal_execute for
each host individually, which is simpler, more reliable, and avoids
the hang issue where parallel remote commands block the stream.
Fix AI_MissingToolResultsError that occurs after user stops a stream
mid-tool-execution: when building SDK messages, skip orphaned tool
calls that have no matching tool result instead of including them
(which causes the SDK to reject the next message).
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: add web search and URL fetch tools for AI agent
Add web_search and url_fetch tools to Catty Agent, allowing the AI to
search the internet for current information and fetch webpage content.
- Support 5 search providers: Tavily, Exa, Bocha, Zhipu, SearXNG
- Settings UI with provider selection, API key encryption, and config
- web_search is conditional on config; url_fetch is always available
- Both tools are read-only and work in all permission modes (incl. observer)
- aiFetch skipHostCheck for AI tool requests to arbitrary URLs
- System prompt guidelines for when to use search/fetch
- i18n support (en + zh-CN)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address code review findings (SSRF, key exposure, state race)
- P1: Restore SSRF protection when skipHostCheck is true — still block
localhost, RFC1918, link-local, and cloud metadata endpoints; only
skip the domain allowlist for public HTTPS hosts
- P2: Move web search API key decryption to main process via dedicated
IPC handler, matching the existing provider key security model
- P2: Use configRef to avoid stale closure in async settings callbacks
that could overwrite newer user changes
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address second review — DNS rebinding, url_fetch approval, maxResults
- P1: url_fetch now requires approval in confirm mode (outbound GET is
a side effect that could exfiltrate data via query strings)
- P1: Add DNS resolution check when skipHostCheck is set — resolve
hostname and reject if any IP is private/loopback/link-local, blocking
DNS rebinding attacks against internal services
- P2: Slice search results after provider call to enforce maxResults
consistently (Zhipu and SearXNG ignore the limit parameter)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address third review — localhost/IPv6 SSRF, API key blur race
- P1: Block localhost/loopback when skipHostCheck is enabled — restructure
isAllowedFetchUrl to check private hosts first in the skipHostCheck path,
preventing access to local services on allowlisted ports
- P1: Handle IPv6 private ranges (fc00::/7, fe80::/10, ::ffff: mapped),
strip brackets from URL.hostname, block [::1] and fd00:: addresses
- P2: Guard handleApiKeyBlur against provider change during async
encryption — skip stale write if provider switched while encrypting
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address fourth review — main-process key isolation, SearXNG compat
- P1: Replace aiWebSearchDecryptKey IPC with __WEB_SEARCH_KEY__ placeholder
pattern — renderer never sees plaintext keys; main process replaces
placeholder in headers before HTTP request, matching provider key flow
- P1: Search API requests use normal allowlist path (not skipHostCheck),
so SearXNG on localhost/HTTP/private networks works via aiSyncWebSearch;
only url_fetch uses skipHostCheck for arbitrary public HTTPS URLs
- P2: Remove needsApproval from url_fetch — treat as read-only like
sftp_read_file, consistent with observer mode allowlist
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address fifth review — private LAN providers, maxResults default
- P1: Allow private-IP hosts that are explicitly in the provider/search
allowlist (e.g. https://192.168.x.x model providers or SearXNG)
- P2: Remove .default(5) from web_search maxResults schema so the user's
configured maxResults setting is used when the model omits the param
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address sixth review — HTTPS scope, config gate, redirects
- P2: Scope HTTP exception to private/LAN IPs only — remote allowlisted
hosts still require HTTPS to protect API keys in transit
- P2: Gate web_search tool on complete config (API key for providers that
require it, apiHost for SearXNG) to avoid advertising a broken tool
- P2: Add redirect following (up to 5 hops) to aiFetch for url_fetch —
handles 301/302/307 for short links, www canonicalization, etc.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address seventh review — redirect SSRF, decrypt race, HTTPS-only
- P1: Revalidate each redirect hop against SSRF guards (allowlist check
+ DNS resolution) before following, preventing open-redirect SSRF
- P2: Add sequence counter to API key decryption effect — stale promise
results from a previous provider are discarded on provider switch
- P3: Restrict url_fetch to HTTPS-only URLs, matching the skipHostCheck
policy that already rejects HTTP in the bridge
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address eighth review — OS resolver, allowlisted HTTP hosts
- P1: Use dns.lookup (OS resolver) instead of dns.resolve4/6 for private
IP checks — matches what http.request actually connects to, respects
/etc/hosts, mDNS, and other local resolver sources
- P2: Allow HTTP for any explicitly allowlisted host (not just literal
private IPs), so self-hosted SearXNG at http://searxng.lan works
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address ninth review — HTTP scope, blur ordering, decrypt flag
- P1: Narrow HTTP exception to web search apiHost only — AI provider
endpoints remain HTTPS-only to protect credentials in transit
- P2: Add blur sequence counter to prevent out-of-order encryption
results from overwriting newer API key saves
- P2: Reset isDecrypting flag when cancelling decrypt on provider switch,
preventing permanently disabled API key input
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address tenth review — DNS pinning, prompt/tool alignment
- P1: Pin validated DNS result to the HTTP request via custom lookup
function, preventing TOCTOU/DNS-rebinding between validation and
actual connection
- P2: Extract isWebSearchReady() helper and use it consistently in
both tool registration and system prompt, so the model isn't told
web search is available when config is incomplete
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address eleventh review — single DNS lookup, redirect pinning, CGNAT
- P1: Combine DNS validation and pinning into a single lookup call,
eliminating the TOCTOU window between hasPrivateResolution and pinnedLookup
- P1: Pin DNS for redirect targets too — resolve/validate/pin in one step
before following each redirect hop
- P2: Add 100.64.0.0/10 (CGNAT) to private IP ranges for Tailscale and
similar CGNAT-addressed internal services
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address twelfth review — apiHost validation, sync on enable
- P2: Validate apiHost is a well-formed URL in isWebSearchReady(),
preventing tool exposure when user enters a malformed host
- P2: Add webSearchConfig.enabled to sync effect deps so the main
process gets updated immediately when the toggle changes
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: remove DNS-level SSRF checks that break fakedns/proxy environments
DNS resolution validation (dns.lookup + IP pinning) breaks in proxy
environments where fakedns resolves all domains to LAN addresses.
Revert to hostname-level checks only (blocking localhost, 127.0.0.1,
metadata endpoints, etc.) which are sufficient without false positives.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: resolve empty catch block lint warning
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When rebuilding SDK messages from conversation history, tool-result
messages had toolName hardcoded to an empty string. This works for
OpenAI/Claude APIs but Gemini requires functionResponse.name to be
non-empty, causing AI_APICallError on every follow-up message.
Now looks up the tool name from the matching assistant tool call
via toolCallId.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Security: API keys no longer transit IPC as plaintext; renderer sends
providerId and main process decrypts via safeStorage. Add ReDoS
protection for user-supplied blocklist regex. Sanitize error messages
to strip file paths and sensitive URLs before displaying in chat.
- Bug fixes: approval timeout now notifies user in chat instead of
silently aborting. statusText cleared consistently across all 7 code
paths. useAIState persistence race condition fixed with mountedRef
guard, storage sync validated with type checks.
- Accessibility: InlineApprovalCard uses role="alertdialog" with focus
on approve button. ChatInput menus have proper ARIA roles, labels,
and aria-expanded. ThinkingBlock toggle has aria-expanded/controls.
Model selector submenu supports keyboard navigation.
- UX: switching agents preserves old session and creates new one.
Approval card buttons disabled immediately after click.
- Performance: text-delta streaming batched via requestAnimationFrame.
- Refactor: extract useAIChatStreaming, useToolApproval, and
useConversationExport hooks from AIChatSidePanel (1514 → 751 lines).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>