Repository navigation
AiFetchly 1.1.138
Since the last published release (1.1.118), this is the largest platform pass so far — the AI chat workspace becomes the default landing page with a durable, restart-safe message queue and steering, conversations name themselves automatically with a small-model title generator, the plugin/MCP stack ships a host-side artifact store, deadline-aware MCP execution, git-free GitHub installation and natural-language skill installation, workspace memory becomes portable files that follow the repo, all 50 inner pages converge on one design system, and a full security sweep (zod 4, vue-i18n 11, deepmerge hardening) plus worker/IPC/DB hardening and packaging-reliability work lands under it all.
This release is cumulative since 1.1.118: intermediate versions 1.1.119 – 1.1.137 were never published (they stayed as automated drafts), so everything built on master since v1.1.118 ships together here.
💬 AI chat workspace & reliability
- Workspace is now the default landing page: the AI chat workspace shell opens first, with header chips for goal/loop state, a background-agents entry in the header overflow menu, a report-conversation item in the chat overflow menu, and a docked inspector that became a real layout column (it pushes content instead of overlaying it)
- Durable message queue + steering: messages sent while a conversation is busy are queued in a new restart-safe pending-message store, drain FIFO exactly once after the run, and can steer a running turn at five safe boundaries — pending bubbles render with direction markers, a compact queue chips row sits above the composer, and queued content is capped and digest-deduplicated against transport retries
- No more empty assistant bubbles: blank placeholders are suppressed while waiting and a remote "working" indicator shows the assistant is processing, in both the workspace and classic chat; tool executions now render below the AI response consistently, and AI errors persist in chat history so they survive reload/reopen
- A bad server response no longer logs you out: AI server envelope errors and upstream provider failures are classified as recoverable chat errors instead of being mistaken for auth expiry
- Model list stays fresh: the model list refreshes when you open a new chat and when provider settings change, so newly configured models appear without a restart
- Retrieval budget loops stopped: repeated
RETRIEVAL_BUDGET_EXHAUSTEDretry loops in the query loop are detected and halted instead of spinning - Queue recovery UI: resuming a paused/queued conversation clears stale recovery badges, the error banner shows recovery attempt counts, keeps retry live, and maps stable error codes such as context-window-exceeded
- Scheduled/loop chat fixes: a
/loopstarted from an empty composer now gets a real conversation row (no more phantom conversation until restart), scheduled runs broadcast their bound workspace so sidebars/folders populate immediately, composer Stop actually aborts a live scheduled occurrence, and partial streams ending infinish_reason=errorare surfaced instead of silently accepted - Generated images, end to end: images fill their tile with icon actions (Save / Save all / edit) and readable hover tips, the index badge no longer blocks icon clicks, the assistant can edit previously generated images in a follow-up turn without a workspace, and batch image editing ships with confirmation, an ambiguity chooser, and retry/stop on batch rows; batch preview receipts are always externalized so batch result cards render correctly after reload or permission-resume
✨ Automatic conversation titles
- Chats name themselves: after the first meaningful exchange, a lightweight ("small") model generates a concise topic title from a bounded excerpt, replacing the raw first-message snippet — with a strict small-model policy that never falls back to an expensive default model, keeping naming cost near zero
- Rename anywhere: a shared rename dialog works from both the chat header and sidebar row context menus; user-set names are permanent and never overwritten by auto-titles (revisioned rename IPC contract)
- Settings section + privacy: automatic titles can be toggled off (with a rollback kill switch), and title generation sanitizes credentials, paths, and Bearer tokens so nothing sensitive leaks into a title; ships with six-language translations
🗂️ Workspace trust & lifecycle
- Trust that survives restarts: per-capability workspace trust (instructions/commands/agents/hooks/skills) is now persisted keyed by a SHA-256 of the normalized workspace root, is fail-closed on missing rows, and migration backfills already-approved workspaces — re-opening the app no longer silently re-trusts or re-asks everything
- Canonical workspace badge revoke: the workspace badge has a proper Remove affordance (screen-reader labeled, disabled during in-flight mutations) wired to the canonical revoke channel, and approving a new workspace now revokes the previous binding so the stale one cannot resurrect
- Project header in the workspace sidebar: a visible "Project" header with Refresh (re-runs the authoritative bootstrap) and Collapse/Expand; workspace summary broadcasts fire on standalone schedule runs so scheduled chats regroup out of "Other chats" in the same session, and the sidebar re-groups live as summary events arrive
- Workspace action counters: renderer-side report/image/voice actions and parity-fallback activations are counted into the workspace action counter (bounded labels)
🧠 Portable workspace memory
- Memory that follows the repo: workspace memory records become portable Markdown files under the workspace's
.aifetchly/tree (identity + index + per-record files with frontmatter), so project knowledge travels across clones, worktrees, and machines; per-record storage/sync badges show exactly where each memory lives - Stable workspace identity: clones and worktrees of the same repo resolve to one portable identity via
.aifetchly/workspace.json, so renaming/moving a folder doesn't orphan your memory - External-agent review flow: an instruction bridge (e.g. CLAUDE.md) lets external agents propose memory edits; external edits arrive as review-new imports with an import-review UI, an
expectedHasheditor guard, diagnostics dialog, and reveal-file — nothing is created or exported until you approve it - Conflict safety: real conflict detection on app writes, and an open conflict is held across background imports so watchers can't clobber it back to synced (race-safe resolution, transactional scope merges); every portable operation writes a sanitized, content-free audit trail with retention caps, plus settings with disable/enable round-trip — certified by nine E2E journeys including crash recovery and fault injection
🔌 Plugins, MCP & skills
- Host file-artifact store: plugin results land as immutable host-owned artifacts with realpath containment, symlink rejection, size quotas, crash-safe journals, streaming validation, and RFC-4180-aware bounded CSV previews; chat tool cards gain preview, native save-dialog export, and per-artifact delete — files never pass through uncontrolled renderer paths
- Plugin scrape pipeline hardened: per-execution staging directories, finalized
candidates.jsonpromotion into the artifact store, staging paths stripped from model payloads, and recoverable failures charged to a retry ledger with retry hints delivered back to the plugin - MCP rebuilt on the official SDK (
@modelcontextprotocol/sdk1.32.1): persistent session pool keyed by server identity, byte-buffered newline decoding (fixes CJK message split corruption), env allow/deny policy, shell-free argv validation, bounded message sizes with redacted stderr, and graceful close escalating to SIGKILL - Deadline-aware MCP execution: one deadline enforced across connect, execution, and cleanup (progress never extends it), the stdio child is killed when connect times out, and result normalization makes
isErrorwin over contradictory nested success so recoverable failures reach the next AI round - MCP launch-time ownership contract: host-managed MCP server rows must revalidate against the owning plugin and the current on-disk
servers.jsonbefore spawn — stale, forged, or modified entries fail closed - Browser capability for plugins: plugins receive an approved Chrome/Chromium descriptor through the trusted host context, with a typed
BROWSER_EXECUTABLE_MISSINGdiagnostic when no browser is available - Natural-language skill installation: "install this skill" in chat drives a deterministic installer with session state machine, secure credential store, install progress cards, and hardening against path traversal/symlink escapes/secret leaks — management is mounted on the Plugins settings page (backed by new skill-installation entities)
- Git-free GitHub plugin installation: install community plugins straight from GitHub releases/archives without local Git — bounded HTTPS transport, typed errors + cancellation, persisted GitHub provenance with resolved commit SHA, and temp dirs cleaned on every failure path (proven by packaged E2E on three OSes)
- Unified plugin page & hub policy: one page unifies plugin discovery and management over a shared community catalog; marketplace IPC is ungated so Free/non-hosted users can browse sources and plugins, and self-hosted deployments may run the plugin hub over HTTP when the host matches the login domain, is a private-network IP, or sets
AIFETCHLY_ALLOW_HTTP_HUB(bakedVITE_ALLOW_HTTP_HUBsupported in CI builds) - Plugin ecosystem compatibility: Claude Code-style agent tool declarations are accepted, plugin agent limits are enforced with optional-omission semantics and surfaced in the UI,
maxConcurrentPerConversationis enforced at job admission, and a failed plugin update restores the complete previous generation
📞 Contact verification
- New
verify_contact_infotool (free to use): validate scraped emails (DNS mail-route resolution with caching) and phones (per-block country evidence) from chat, with Zod-validated inputs, evidence-based results, and redaction - Verification is composed into extraction:
extract_contact_inforequires a follow-upverify_contact_infocall and attaches verification evidence; suspicious short local-parts are flagged conservatively, and audit logs stripsource_urlquery parameters
📧 Email replies
- Assisted send no longer dead-ends on inconclusive probes: when the sent-folder probe can't conclusively prove whether a reply went out, delivery now proceeds to the atomic send claim with threading safeguards and a
detection_inconclusive_proceededaudit row — IMAP servers with odd/hidden Sent folders stop blocking replies permanently - Review cards you can see, then send: saved email replies show as review cards inside AI chat, and a reply cannot be delivered unless its review card is actually visible in the workspace; the approve-and-send button shows real loading/disabled feedback
- Human confirmation override: a
reviewedparameter oncreate_email_reply_draftplushumanConfirmedexemption in the reply policy evaluator stops the "needs human review" loop for messages you already cleared, and reviewed claims are forbidden in scheduled/unattended guidance - CJK sent folders recognized: Chinese enterprise mailboxes (Aliyun qiye / QQ / NetEase) advertising
已发送/已发送邮件are fallback sent-folder names, with bounded probe-failure logging instead of silent swallows - Outbound tool schemas fixed for flash-tier models:
start_email_send_task/draft_outbound_email_batchschemas encode the four exact required parameter sets, and validation errors name the missing parameters — batch sends succeed on the first try instead of burning 6+ rounds
🎨 UI convergence, settings & voice
- All 50 customer-facing inner pages share one design: every page renders inside the persistent shell via
AppPageShellwith shared page templates (Collection/List, Create/Edit, Detail/Inspection, Results/Activity, Settings/Catalog), one token system, dual themes, and sharedui.*localization namespaces with a six-language parity test - Font size controls: a five-step app font scale in User Preferences (applies immediately, persists across launches) plus a separate chat font size for messages and the message box
- Settings polish: User Preferences is first in the left settings menu (applied at query time, no migration), the Language dropdown is fixed (was showing a raw stored value / "No data available", with six-language option seeding that self-heals existing DBs), and a startup crash from concurrent system-setting initialization (UNIQUE constraint) is fixed with single-flight init
- Speech-model install dialog: toggling spoken responses without a model installed opens a guided install dialog from both classic and workspace chat, the popup stays open with live progress and errors, and spoken responses auto-enable when the install completes; speech cancels on conversation switch/Stop and via composer Stop buttons
- WSL & GPU stability: WSLg sessions always run software rendering (fixes GPU-PV freezes),
AIFETCHLY_DISABLE_GPU=1is a supported escape hatch on any platform, Linux scrapers launch Chromium with--no-sandbox/--disable-gpuon Linux only, and dev builds always disable hardware acceleration
🔒 Security & dependencies
- Full Dependabot sweep: zod → 4.6 (with tool-envelope JSON schema
io:"input"restoration), vue-i18n → 11.4 (with linked-message@escape fixed in all six locales), nodemailer → 10.1+, node-forge → 1.4.3, fast-xml-parser → 5.12, minimatch → 11.0.8, qs → 6.16, joi → 17.13.9, follow-redirects → 1.16.2, serialize-javascript → 7.1.3, decode-uri-component → 0.5.1, glob → 14, basic-ftp → 6.2.6, hono → 4.13.18, @hono/node-server → 2.1.8, fast-uri → 3.1.9, ip-address → 10.7.4, undici → 7.30.1, happy-dom → 21, body-parser → 1.20.9 - deepmerge prototype-poisoning hardening: a postinstall patch guards
__proto__/constructor/prototypekeys in both merge loops of every deepmerge 4.x build (the upstream fix only guarded one), closing the poisoned-target rewrite while keeping benign merges byte-identical - CodeQL findings resolved: workflow permissions tightened (
contents: readin test.yml), unvalidated dynamic method call in the test server, and identity-replacement findings in the portable memory bridge - Dependency hygiene: puppeteer moved to the official 25.8.0 line (dropping the rebrowser fork) with puppeteer-core/cluster bumped, removed 11 confirmed-unused dependencies plus dead
crypto,sqlite3,http-proxy-agent,socks5-http-client,puppeteer-stream,puppeteer-video-recorder,protocol-registry, andvite-plugin-commonjs
🏗️ Platform hardening & build reliability
- Workers crash-safe by contract: all background workers validate inbound messages with Zod contracts, crashed workers restart under a bounded policy (exponential backoff + circuit breaker) with fatal drain/notify/exit, in-flight contact-extraction jobs are recovered after a crash, worker browsers close on shutdown, and a global WorkerCoordinator budgets browser-bearing workers so heavy scrapes no longer oversubscribe the machine
- IPC & renderer hardened: every IPC handler migrated to schema-validated registration, direct
ipcRendereraccess forbidden in the renderer (windowInvoke bridge only), navigation guards on the main window, and optional at-rest credential encryption via Electron safeStorage (opt-inAIFETCHLY_ENCRYPT_STORE) - Database layer modernized: legacy raw-SQL files replaced by typed async Models, hot foreign-key/status columns indexed for faster list queries, path-safe SQLite singleton, a centralized
WORKER_TYPEguard preventing workers from touching the database, and TypeORM CLI migrations wired (migration:generate/run/show) - Renderer architecture: Vuex → Pinia cut-over complete (dead Vuex/Vue2 code removed, legacy v1 AI chat retired), a shared
useApiCallcomposable standardizes loading/error states, and strict TypeScript discipline (noImplicitAny,strictPropertyInitialization,no-explicit-any, console→Logger codemod) plus module-layer constructor injection land - Packaging & CI reliability: a hang guard wraps packaging (hard timeouts, reliable termination), Vite production builds run isolated in child processes to prevent runner OOM, the Windows release build is split into guarded package + make phases, Ubuntu legs pinned/unpinned as needed, and worker bundling fixes (zod, js-yaml) fix worker crashes in installed builds — not just CI
📝 Notes
- Cumulative changes since
v1.1.118(2026-10-22 → 2026-11-17): 1,478 commits, 2,602 files (+815,782/−61,216) across the full tree including vendored tooling docs — of which ~1,170 code/docs commits touchsrc/test/docs(2,109 files, +376,878/−58,927): ≈ 393 feat / 484 fix / 247 docs / 138 test / 88 refactor / 22 chore / 14 ci - Additive local DB schema: new tables (
ai_chat_pending_messages,ai_workspace_memory_scopes,ai_workspace_memory_scope_paths,ai_workspace_memory_portable_states,ai_workspace_memory_sync_audits,aifetchly_workspace_trust,prompt_skill_invocations,skill_installations,skill_installation_sessions,skill_installation_events,skill_credential_bindings,skill_dependency_bindings) plus title/queue/sync columns on existing tables are created automatically on first launch (TypeORM synchronize) — no destructive migration, no data loss on upgrade or downgrade - New IPC families:
session-recording:*, ownership-scoped plugin artifact IPC (list/preview/export/delete), skill-installation lifecycle, revisioned chat-title rename, email reply review cards, renderer workspace action counters, and Zod-validated handler registration across existing channels - Flags & environment:
AIFETCHLY_ALLOW_HTTP_HUB/ bakedVITE_ALLOW_HTTP_HUB(self-hosted HTTP plugin hub),AIFETCHLY_DISABLE_GPU(any-platform GPU opt-out),AIFETCHLY_ENCRYPT_STORE(opt-in at-rest credential encryption) - UI & i18n: all new surfaces ship with six-language translations (en/zh/es/fr/de/ja) on the vue-i18n 11 engine
SHA-256 checksums
| File | SHA-256 |
|---|---|
AiFetchlySetup.exe |
f649af23cd0bfa868ff44b8ab152afc5e2028f73a473d90b08ab94051283b5f9 |
RELEASES |
55aeba9098180e0b41935fe826ddf2cc5ca37b8d1eccfe88a0a2623570d1694c |
aiFetchly-1.1.138-arm64.dmg |
1103f42099f80e151356c8dcbc2a076350bee158277482487209bb0e71773172 |
aiFetchly-darwin-arm64-1.1.138.zip |
4353dd90066f9aae0b14b106cd9fdc695e2deee26f2703fe6e5a7888161052c3 |
aiFetchly.msi |
f2ff665f4f441682ca9a972abab6add135a10ed25b296487ebb63ebdea207743 |
aiFetchly.msix |
38f27a266d31cdaa737b7ea5358609269d7a3728e8d5f8b36dfa0d89bebc0f18 |
aifetchly-1.1.138-full.nupkg |
c9dbaba549ad8627040289e04ce396f84ef1ac8bcaca0fe640b0a93e823ae78c |
Verify a download with sha256sum -c SHA256SUMS.txt (Linux/macOS) or Get-FileHash -Algorithm SHA256 (Windows).