TalkToAi Code 0.3.0 — Jobs, deliverables and your own OpenAI key

TalkToAi Code 0.3.0 preview — Jobs and deliverables

Windows preview for GitHub Releases and the public website. This unsigned EXE
is not a Microsoft Store submission package.

New capabilities

• Managed process jobs: the agent can start a native executable with literal
  arguments, poll incremental output and cancel the exact job. No command is
  silently rerun by polling. A new native Jobs panel exposes status, output,
  exit code and a Stop selected job action.
• Explicit lifecycle: up to three concurrent jobs and twenty per turn;
  deadlines up to thirty minutes, a 64,000-character retained output buffer and
  a 4 MiB output stop limit. All jobs are owned by their agent turn and stopped
  at turn end. Windows Job Objects bound descendant process lifetime.
• Batch file reads: inspect up to eight UTF-8 source files with one tool
  call. Bounded results include SHA-256 and continuation offsets. Changed files
  reject stale continuation hashes. One missing file does not hide other results.
• Deliverables in Evidence: register reports, build files and other existing
  project outputs with labels, paths, sizes and hashes. Registration never
  uploads or executes files. Double-click reveals registered files' folders.
• Practical task starters: build-and-collect and local-website testing prompts,
  alongside expanded in-app help. Tool sets are discoverable during a task.
• Bring-your-own OpenAI API: a dedicated preset, model metadata discovery,
  user-supplied session/environment/Windows-encrypted keys, explicit API selection
  and an optional API startup preference. Local/open-weight models stay the
  default. No developer/shared API key, account, subscription or credit is supplied.
• API controls and visibility: per-response token ceilings, first-party
  max_completion_tokens, usage reporting when supplied, and clearer HTTP errors.
  Token limits are not currency caps. Auto never falls back to a paid API.

Scope and limitations

Processes use the signed-in account's permissions; this is not a sandbox.
The new job interface has no interactive stdin, does not survive a turn/app
restart and is not a persistent hosting service. Completed job summaries are
saved with the conversation, not reattached to arbitrary system processes.
Exit code zero is not a claim of test coverage or game playability.

Output registration records bytes read at that moment, not an immutable backup.
Models and engines remain separate. No live-model quality/speed improvement is
claimed; the prior local-model timeout is not resolved by controller tests.

The three supplied DevSpace folders were evaluated for capabilities. Their
product licence did not permit code reuse; no DevSpace product code or assets
were copied. See docs/DEVSPACE-IDEAS-REVIEW.md.

The OpenAI integration follows current official Chat Completions, function-tool
and model-list documentation. Tests use dummy keys and a loopback HTTP fixture;
no live paid OpenAI request or real credential is used. Model access, pricing,
quota and inference/tool compatibility depend on the user's account and choice.

Validation

128 automated tests passed. Coverage includes real native command execution,
failure exit codes, cancellation, descendant cleanup after the parent exits,
output flooding, deadlines, turn cleanup, stale file-read hashes, output
registration, chat-scoped job events, own-key API storage, Windows encryption,
no automatic API fallback, metadata-only model discovery and a complete
OpenAI-style tool-call/result stream using a loopback fixture.

The final packaged executable's self-test exited 0: managed process execution,
batch reads, output registration, bundled example, browser interaction,
browser screenshot and image attachment all passed. Its native Qt preview
also launched and exited 0 with a separate temporary profile. Source UI
previews were inspected at 1440x900 and 1100x700, with the provider dialog
rendered in a smaller desktop layout. The packaged tree contained zero
config.json, providers.json, studio.json or *.dpapi files.

Windows x64 Inno Setup preview:

• File: TalkToAi-Code-0.3.0-Windows-Setup.exe
• Bytes: 66,138,598
• SHA-256: 02c67393cb762ce03479bc4610ce9f061b52435f761ad895e452a9310d43fd57
• Authenticode status: NotSigned (unsigned preview)
• Matching checksum file: SHA256SUMS-0.3.0.txt

The active user installation is not replaced. No in-place upgrade, clean-machine
installation, antivirus clearance or Microsoft Store certification is claimed.

TalkToAi Code 0.2.5 - Ask for the outcome

TalkToAi Code 0.2.5 — Ask for the outcome

This release improves agent follow-through for coding, websites and games. It builds on 0.2.4's recoverable chats, project memory, search, archive controls and bundled example game.

What changed

• On-demand tool sets: the agent can enable additional code-navigation, browser, game, desktop and SSH tools during a turn. Available capabilities are no longer restricted to keywords in the original prompt. Existing mode/access permissions remain in force.
• Automatic project overview: with automatic context enabled, the agent gets a bounded read-only overview of the project type, detected checks and available engines without spending a model turn requesting it.
• Saved task steps: a multi-step checklist appears in the new Steps tab and survives app restarts. An unfinished checklist prompts the agent to continue or report a real blocker. Checklist completion is explicitly labelled as agent-reported.
• Bounded automatic continuation: output-length limits can trigger up to two continuations within the normal turn budget. Incomplete tool batches are not executed. Stop remains available; the app does not run indefinitely.
• Check evidence beside the plan: runner-derived statuses distinguish passed, failed, unverified and stale checks. Ordinary shell commands do not count as verifying file edits. This tracks file-tool edits, not every external filesystem change.
• More resilient tool execution: invalid argument batches are rejected before any part runs, with bounded model retries. Repeated identical consecutive failures are not executed endlessly.
• Continue unfinished work: a button in Steps and an action in Ctrl+K resumes the saved conversation while preserving any existing draft.

Validation and limits

98 automated tests passed, including on-demand tool availability, permission boundaries, plan persistence, output-limit continuation, retention of the original request during continuation, incomplete-batch rejection, repeated-failure handling and checks derived from runner output rather than model claims. Existing browser/runtime and release tests remain included. The actual Qt layout was rendered using a labelled demonstration profile at two desktop sizes.

A separate read-only acceptance attempt against local qwen3.5:4b stopped at its 120-second deadline without producing a tool call. That is not a live-model automation pass. The configured localhost server-model endpoint was unavailable during that check. No paid API, credentials or personal project data were used. Quality and latency still depend on the selected model/runtime.

The Windows x64 installer compiled successfully using the existing standard Inno Setup configuration. The final packaged executable passed its runtime smoke check (browser interaction, browser screenshot, vision attachment and bundled example) with exit code 0. Its native Qt preview also launched and exited successfully in an isolated temporary profile. The user's existing installation is deliberately left untouched; an in-place 0.2.5 upgrade on that active installation is not claimed as tested.

Windows download verification

• File: TalkToAi-Code-0.2.5-Windows-Setup.exe
• Size: 66,094,889 bytes
• SHA-256: 277e50fed6a4dc10ae8c539a035a134823dd98ce395fad008c1453941dce7334
• A matching SHA256SUMS-0.2.5.txt is attached to this release.

Preview status

Unsigned preview. No claim of Codex parity, unrestricted autonomy, complete game playtesting or security certification. Loading tools never grants new permissions. Real credentials, first-time SSH host verification and other necessary user decisions still require the user's involvement. Models and game engines remain separate downloads; optional APIs retain their own prices and limits.

TalkToAi Code 0.2.4 — A clearer, more resilient workspace

TalkToAi Code 0.2.4 — A clearer, more resilient workspace

Build websites, software and games with your own local or self-hosted model, or an optional API provider. This release focuses on finding your work, keeping it recoverable and getting started with less setup.

New in this release

• Full-text chat search: find titles, projects, drafts and conversation text. Ctrl+Shift+F focuses chat search; Ctrl+F finds text in the current conversation.
• Archive and restore: Active, Archived and All chat views keep the sidebar tidy without deleting history. The chat menu and right-click menu offer rename, pin, branch, archive/restore, copy last reply and export report.
• Chat recovery: atomic saves retain a validated previous-save backup. At startup, damaged history is preserved under a unique filename and a valid backup is recovered with a visible notice. No model call is involved.
• Seven editable task starters: coding improvements, website features, game features, debugging, release review, SSH inspection and continuing unfinished work. Starters never execute automatically or overwrite an existing draft.
• A properly bundled example game: Game → Open example game creates a writable Score Arena copy in persistent app data. Existing example-project edits are preserved. Godot is a separate prerequisite.
• A slimmer sidebar: More groups project memory, API providers, ZeroThink linking, model choices and help. Settings, Connections and About & updates remain directly accessible.
• Reliability fixes: pinned chats no longer steal selection from new/branched conversations; chat-list refresh preserves unsaved file-editor text; local model choices persist in user settings.

Includes editable project memory, autosaved drafts, GitHub update checks and SHA-256 verified installer downloads, saved SSH alias discovery, task steering, file diffs/checkpoints, browser tools, desktop tools and Game Lab from earlier releases.

Install or update

Download TalkToAi-Code-0.2.4-Windows-Setup.exe below, or open About & updates in the app. This is a standard Windows x64 per-user Inno Setup installer with bundled Python runtime, Start Menu and optional Desktop shortcuts, and a Windows uninstaller. Existing chat history and settings remain in %LOCALAPPDATA%\TalkToAiCode during upgrades. Updates are user-initiated, not silently installed.

Validation and limits

The full 78-test suite passed, covering agent/tool protocols, project checks, memory, SSH discovery, provider integration, update verification, computer-tool mocks, real Edge interaction/screenshot fixtures, session corruption/recovery, bundled-example copying and desktop UI regressions. The actual Qt layout was rendered and reviewed at 1280×800 and 1600×960 using a disposable demo profile.

The final installer completed with exit code 0 on the build machine. Desktop and Start Menu shortcuts resolve to the installed executable; Windows registers version 0.2.4 and its uninstaller. The installed runtime passed its smoke check (exit 0): bundled example present, browser interaction successful, screenshot saved, and image attachment encoding working. This checks packaging and browser behavior, not model quality or full gameplay. Installer size: 66,078,972 bytes. SHA-256: 16e3c5ddd007d1afc18b47bf9ea567d67a0b9bf79fc28c5873006d74b7977899.

Windows preview; installer unsigned. Tests are not a security certification or proof of complete game playtesting. Chat recovery retains the previous successful save, not every historical version, and is not a substitute for project or off-device backups. Archived chats remain on your machine. Screenshots of the new layout use a clearly labelled UI demonstration, not a benchmark or live model result.

Models and game engines are separate downloads. CPU inference is supported, with speed/quality dependent on hardware and model. Optional APIs keep their own quotas and prices. Remote models receive the task context you send them; do not store passwords in chat or project memory. ZeroThink vault inference remains experimental. Linux/macOS remain source installations; this release does not claim native signed installers for them or parity with hosted assistants.

TalkToAi Code 0.2.3 — Project memory and workspace improvements

TalkToAi Code 0.2.3 adds editable project memory and everyday UI improvements.

• Project memory in the sidebar, or Ctrl+Shift+M: maintain project decisions, conventions, test commands and next steps. Notes live in .talktoai-code/PROJECT_MEMORY.md and are included in subsequent tasks for that project. Conflicting edits are detected before saving. Notes are capped at 12,000 characters to limit context cost.
• Draft autosave after a short typing pause, retaining unsent text across restarts.
• Ctrl+, opens Settings. Ctrl+K now also offers Project memory, About & updates, Open app data folder and Open project folder.
• Includes the GitHub update panel and saved SSH alias discovery from 0.2.1–0.2.2.

Validation: 34 tests covering project memory persistence/isolation/conflicting edits, update checks/download verification, SSH discovery, agent behavior and desktop controls. Packaged installation and runtime smoke checks are completed before publication.

Memory notes are included in requests to your selected model, including remote/API models; never put passwords or keys in them. This is editable project context, not automatic semantic recall. Windows x64 preview; installer unsigned. Local inference speed and quality depend on hardware/model. Model weights and third-party game engines are separate downloads.

TalkToAi Code 0.2.2 — GitHub updates and SSH discovery

TalkToAi Code 0.2.2 adds saved SSH connection discovery alongside the GitHub update panel introduced in 0.2.1.

• In Act mode, ask to connect to your server. The agent can inventory existing OpenSSH aliases (including Include files), select an existing alias, verify access and inspect a remote project before work.
• Explicit SSH connection requests make these tools available for that turn without requiring a previously selected Connections entry. Authentication uses your configured OpenSSH keys or agent.
• Saved app connection profiles remain supported. Unknown aliases are rejected. Password-only accounts and first-time host-key prompts still need interactive setup.
• About & updates checks GitHub, displays release notes and downloads SHA-256-verified installers without AI inference. Installation is initiated by the user after active work stops.
• Startup uses the tray flag, settings persist outside application binaries and Desktop shortcuts are selected by default.

Validation: 13 focused SSH, update and desktop tests passed. The actual SSH helper successfully connected to an existing configured server. The 0.2.1 installed runtime passed browser interaction, screenshot and vision-attachment checks; the 0.2.2 installer/runtime is checked again before publication.

Windows x64 preview. Unsigned installer. Local task history is supported; automatic cross-project semantic memory is not implemented. No model weights, keys, user profiles or server access are included.

TalkToAi Code 0.2.1 — GitHub updates and startup improvements

TalkToAi Code 0.2.1 adds About & updates to the sidebar.

• Check GitHub for a newer Windows release without using an AI model or API credits.
• Read release notes in the app, download the installer in the background, and verify its SHA-256 against GitHub asset metadata.
• Click Install update and quit when ready. Active agent work must finish or stop first. Updates are user initiated; nothing installs automatically.
• Start with Windows now launches into the tray when available.
• Saved settings now live alongside local task data, separate from installed program files.
• Desktop shortcut creation is selected by default in the Windows installer.

This remains a preview. The Windows installer is unsigned. Local model speed and coding quality depend on your model and hardware. Persistent task history is available; automatic cross-project semantic memory is not implemented.

Validation: update metadata/version tests, successful verified download, checksum mismatch cleanup, missing-digest and unexpected-download-location rejection, and existing desktop tests.

TalkToAi Code 0.2.0 - Standard Windows installer

Standard per-user Inno Setup installer. Installs the native Windows application, creates Start Menu/Desktop shortcuts and a normal Windows uninstaller. Existing local TalkToAi Code task data remains separate and is retained during upgrades. This replaces the earlier PyInstaller bootstrap installers.

TalkToAi Code 0.1.8 - Reinstall and upgrade fix

Fixes Windows error 183 when %LOCALAPPDATA%\TalkToAiCode already exists. Existing task history and configuration are retained; only the packaged runtime is refreshed and shortcuts are created.

TalkToAi Code 0.1.7 - Installer diagnostics and shortcut fix

Adds explicit COM initialization for Windows shortcut creation and visible error reporting. Installer failures now show a dialog and write %LOCALAPPDATA%\TalkToAiCode-installer.log instead of silently hanging.

TalkToAi Code 0.1.6 - Fixed Windows installer

Fixes the 0.1.5 installer hang during shortcut creation by using direct Windows Shell shortcut creation instead of a blocking PowerShell subprocess. Creates Desktop, Start Menu and uninstall shortcuts, and includes the opt-in Start with Windows setting.