Product history

PAI is moving fast.

The changelog tracks major product milestones rather than every internal patch. The desktop app should surface the newest update first and keep the full history available here.

Current developer release

PAI 6.9.0-alpha.7

Native PAI-Coder, automatic first-run model/runtime setup, CUDA acceleration, neural voice, and desktop fixes.

Checking current release…

Download current build
Current generation

6.9 — Distribution, Accounts & Tester Delivery

  • Completed the native Windows installer and updater path so PAI can be distributed as a normal desktop application.
  • Added Developer, Beta, and Stable update-channel architecture with verified packages, staged installation, startup health checks, rollback foundations, and repair support.
  • Moved PAI-Coder toward zero-touch first-run setup with native local runtime acquisition, CUDA acceleration when supported, and CPU fallback.
  • Added visible first-run PAI-Coder download percentage and downloaded/total size progress.
  • Replaced the normal Local Mode path with Supabase-backed PAI account sign-in and secure saved-session restoration.
  • Added Free, Founding Supporter, and Pro entitlement mapping to the desktop application.
  • Hardened CI around account, entitlement, installer, dependency, frozen-executable, and release changes.

6.8 — Conversational Voice & Continuity

  • Added provider-independent duplex voice-session architecture with natural turn endpointing.
  • Added barge-in/interruption handling, streaming speech chunking, and echo-suppression guards.
  • Added bounded persistent voice-turn continuity while keeping push-to-talk available as a fallback.
  • Connected conversational voice to the semantic intent and continuity systems built in the previous generation.

6.7 — Fluent Intent & Natural-Language Control

  • Introduced a model-independent semantic-provider interface with strict structured validation.
  • Added an authoritative requirement ledger for hard requirements, preferences, preservation rules, rejected ideas, replacements, assumptions, and revisions.
  • Improved conversational revisions and ambiguity handling so PAI asks for clarification only when execution meaning is genuinely unresolved.
  • Added natural-language control over automation tasks while preserving deterministic safety.

6.6 — Persistent Automation & Background Autonomy

  • Added a durable project-aware automation queue instead of relying only on ad-hoc background folders.
  • Introduced read-only, safe-review, low-risk auto-apply, and checkpointed-autonomous task levels.
  • Added external task intake, source preservation, task priority, duplicate/cooldown protection, and persistent logs.
  • Added pause, resume, cancel, restart recovery, and unattended sequential work while keeping completion authority with verification evidence.

6.5 — Verifier-Owned Execution & Causal Completion

  • Formalized the model-independent Input → Understand → Plan → Code → Test → Diagnose → Repair → Verify → Finished pipeline.
  • Separated technical project health from proof that the user's actual request was satisfied.
  • Added objective requirement contracts and before/after acceptance evidence to block false completion and unrelated edits.
  • Strengthened bounded repair so failures become targeted repair evidence instead of permission for blind rewrites.
  • Validated the system against a synthetic 1,783-file multi-language project with multiple deliberate repair failures before final acceptance.

6.4 — Application Completion & Desktop Simplification

  • Added persistent application state, recent projects, project sessions, startup-health tracking, cancellable background tasks, friendly error handling, and runtime diagnostics.
  • Simplified visible branding to PAI and consolidated navigation around the primary Chat, Code/Workspace, and Activity/Logs surfaces.
  • Moved secondary controls into Settings and changed project search into an editor-style find workflow.
  • Added local model management, pause/resume/cancel controls, rollback handoff, window-mode persistence, and stronger Windows release gates.

6.3 — Universal Coding Core

  • Added language-agnostic project scanning and a persistent project-map API.
  • Added mixed-language, framework, build-system, manifest, entry-point, and dependency detection.
  • Introduced adapters for Python, C, C++, C#, Java, JavaScript, TypeScript, Lua, GLua, HTML/CSS, and config/data files.
  • Kept specialized GLua semantics isolated behind an adapter instead of making Garry's Mod assumptions universal.

6.0–6.2 — Native Application & Windows Distribution Foundation

  • Separated replaceable application/runtime files from mutable user data so updates could not overwrite projects and settings.
  • Built the native Windows packaging path for PAI.exe, PAIUpdater.exe, PAISetup.exe, update ZIPs, hashes, and release manifests.
  • Reduced UI contention during large builds with batched live-code updates and lower-frequency background polling.
  • Added updater states, package verification, Windows-path hardening, and model/runtime guards that later releases built on.

5.0 — Persistent Desktop Workbench

  • Preserved the proven autonomous engine while adding persistent workspace, session, task, and recent-project state.
  • Expanded the editor-first desktop with project tree, tabs, breadcrumbs, split editing, Problems, Search, Output, Terminal, References, Git, Build, and Preview surfaces.
  • Added durable queued/working/review/done/blocked/cancelled task states and structured activity history.
  • Added conservative cross-module engineering intelligence while retaining semantic verification, checkpoints, background autonomy, and forensic logs.
  • Redesigned the desktop identity around a clearer Code/Assistant/Activity workflow and laid the foundation for voice and runtime identity.

4.0 — Autonomous Project Builder

  • Moved project requests from approval-heavy execution to automatic multi-phase autonomous builds.
  • Added filesystem truth gates so PAI could not report success when planned files were missing or empty.
  • Added validated native ZIP artifacts and rejected empty or invalid packages.
  • Introduced Background Autopilot with isolated working copies, source preservation, completed/failed queues, and automatic verification.
  • Hardened completion into FILE VERIFIED → PHASE VERIFIED → REQUIREMENTS VERIFIED → PROJECT VERIFIED → PACKAGE VERIFIED → DONE.
  • Added requirement ledgers, artifact truth, dependency graphs, adversarial review, and stronger project-level integrity checks.

3.x — Reliable Coding Agent

  • Established bounded, targeted project editing instead of unsafe blind whole-file rewrites.
  • Added persistent edit receipts, diff-grounded follow-up explanations, and authoritative verification-driven terminal states.
  • Introduced assessment-first improvement work with EDIT, KEEP, and ADVISE outcomes instead of forcing every review into a code change.
  • Added no-op detection, malformed-edit recovery, canonical symbol identity recovery, and evidence-preserving repair restricted to failed targets.
  • Expanded runtime-aware semantic verification, project-defined API provenance, GLua realm/network ownership, and cross-file contract checks.
  • Consolidated the reliable-agent loop around UNDERSTAND → DECIDE SCOPE → MODIFY/CREATE → VERIFY → REPAIR → REVERIFY.