Changelog

All notable changes to Aefos AI are documented here.

The format is based on Keep a Changelog, and this project aims to follow Semantic Versioning. Dates are in YYYY-MM-DD.

[1.6.0] - 2026-09-05

Added

A second doorway to the IDE's MCP server, for machines where the named pipe is not an option. Some corporate images block or restrict named pipes, and on those the agent could see the IDE's tools listed and never reach them. Aefos now also publishes its MCP server over a loopback HTTP listener on 127.0.0.1 — no relay executable, no bridge script, nothing off-box — and the local CLI connects straight to it. The named pipe stays the default and is untouched; the same server answers either door.

The port is bound to 127.0.0.1 only. Eight connections are served concurrently, each on its own thread, so one idle client can no longer hold every other request behind it; a ninth is refused rather than queued.

Two diagnostics, both off unless you switch them on. AEFOS_HTTP_TRACE=1 writes every step of a request — accept, read, dispatch, reply — with microsecond timings, and AEFOS_GATEWAY_TRACE=1 writes the PID of every addon MCP server Aefos starts, the same number Task Manager shows.

Fixed

Chat sessions were never saved on Delphi 10 Seattle and 10.1 Berlin. The upsert reached SQLite malformed through the older FireDAC, so every session vanished with the IDE. History is kept again on those versions.

An installed MCP addon was started twice per question. Aefos already re-exposes installed addon servers to every agent, and the configuration handed to the CLI listed them as well, so each turn started a second copy of every addon server — two AefosDesktopMcp.exe, two analyzer servers — and the agent saw the same tools twice, under two names. One server per addon now starts once per IDE session and is reused by every turn. AEFOS_CLI_ADDON_MCP=1 restores the previous behaviour.

The IDE could be held open at shutdown by the new transport. Shutting down now closes live connections first and waits with a five-second ceiling rather than forever.

A failed connection no longer costs the whole transport. An error while accepting could end the accept loop for the rest of the session, and every request after it would wait with nothing in any log.

The provisioning bridge script embedded in the plugin could ship out of date with the one in the repository. It is now compiled from source on every build.

[1.5.2] - 2026-09-03

Fixed

With a project open, every Chat message failed instantly. The turn never reached the CLI — it died in the spawn with CreateProcess failed for "…\claude.exe" (last error: 206). Closing the project made the same question work, which made the bug look like it was about the project and not about the message.

Error 206 is ERROR_FILENAME_EXCED_RANGE, and despite the name it is not a path problem: CreateProcess returns it when the command line passes 32767 characters. The prompt was the last token of that command line, and the prompt carries the rendered project context — whose active-unit body alone is capped at 32 KB (ProjectContextBuilder.ACTIVE_UNIT_MAX_BYTES). That cap is already one byte past what a command line can hold, before the executable, the flags and the quoting are added. Hence the exact correlation: with no project open the context degrades to a single short line and the same prompt fits.

The prompt now travels on stdin for Claude Code, which reads it there and which the dispatcher has always opened a pipe for (it just closed it empty). A dedicated writer thread feeds it, so a payload larger than the pipe buffer cannot deadlock against a child that writes output before draining its input. Off the command line, the 32 KB limit no longer applies at all.

GitHub Copilot was hit by the same 206, and is fixed the same way. The CLI documents its own second channel — with an empty stdin and no -p it answers "No prompt provided. Run in an interactive terminal or provide a prompt with -p or via standard in." — and it was measured live, in the exact shape the driver builds (copilot --session-id=<uuid> --allow-all-tools -s, prompt piped in): it answers and exits 0. The trailing -p left the driver's arguments in the same breath, because a flag whose value never arrives swallows whatever follows it.

Each driver declares its own answer (IExecutorProfile.PromptViaStdin), because where a CLI accepts its prompt is that CLI's dialect: Codex, Gemini and the local agent CLI keep the command line, unchanged, until their stdin form can be verified against a real binary. For those, an over-long command line is now refused up front with a message that says what actually happened, instead of surfacing as a bare last error: 206.

1.5.1 - 2026-08-10

Two things the addon store and the MCP list disagreed about. Both were found on a machine running Delphi 10 Seattle and Delphi 13 side by side, one day after 1.5.0 reached it.

Fixed

On Seattle and Berlin, an installed MCP addon was invisible — and the agent could not reach it. The Desktop MCP was missing from /MCP on Delphi 10 while the same user profile's Delphi 13 listed it, reading the very same file.

The cause is in the old RTL: on 10 Seattle (17.0) and 10.1 Berlin (18.0), TFile.ReadAllText(path, TEncoding.UTF8) discards the first three bytes of a file that carries no BOM — it charges the preamble whether or not the file paid it. 10.2 Tokyo (19.0) and up return the file whole (measured on all six). The addon aggregate is written UTF-8 without a BOM on purpose, so on those two IDEs it arrived with its opening brace gone, failed to parse, and degraded to "no servers" — in the modal and in the .mcp.json handed to the CLI, which is how the agent lost the addon's tools there.

It now reads the bytes and strips a BOM only when one is present — the method the in-IDE gateway already used, which is why the tools inside the IDE kept working while the config leg went silent.

The store said "Update" on a current addon, forever. The installer seeds the Desktop MCP offline from a copy this repository builds, and the catalogue decided the button by comparing sha256 with the gallery. Two honest builds of the same version have two different sha256, so every freshly installed machine was told to update an addon that was already the newest release.

The button now answers the question a user actually reads in it — is there a newer release? — by version. aefos update <slug> still compares sha256, so a re-published bundle can be pulled deliberately; only the catalogue stops calling a rebuild an upgrade.

1.5.0 - 2026-08-10

Every RAD Studio from Delphi 10 Seattle up, and an addon store inside the IDE. The installer now carries eight payloads (17.0 Seattle through 37.0), each built against its own compiler, and the chat grew an /addons window that browses and installs bundles without leaving the IDE.

Added

Delphi 10 Seattle, 10.1 Berlin, 10.2 Tokyo, 10.3 Rio and 10.4 Sydney (BDS 17.0 to 21.0). Reaching back a decade cost less than it looks: nothing in the RTL Aefos uses is newer than XE8, the WebView2 layer is our own COM import rather than Vcl.Edge, and the source has been free of inline var by house rule. What did have to change was measured, never guessed — static SQLite is now detected in the project instead of assumed from the version, an IDE with no C++ personality is named as such instead of failing on a missing stdint.h, DFM character escapes are written in decimal (Seattle rejects hex), and the BPL staging hook moved to _PostCompileTargets, inside the Base group, where an old IDE's own build engine still reads it.

Installed and run on a Delphi 10 Seattle machine, not merely compiled for it.

The addon store, in the chat. /addons opens a window that lists what an Aefos store offers, installs a bundle, and updates one already installed — including an MCP addon whose server is running at that moment. There is one official store nobody can move, and the user's own stores beside it: add one by picking a folder, and a folder that is an addon is read as a store of one. A Format tab documents the bundle format where the addons are, and an installed addon's tools reach every executor, namespaced <slug>__<tool>.

/install and /store still answer, but the command the picker offers is /addons — the word the product already uses for the store, for the contract and for ~/.aefos/addons.

One bundle format: addon.json v1. A bundle may carry several commands and several skills. The four foreign-plugin readers are gone: reading someone else's format made Aefos responsible for four moving contracts it does not own.

Fixed

Two RAD Studios were answering for each other's packages. Every BPL carried the same file name on every version, so whichever install came first on PATH served them all — a Seattle IDE could load Sydney's Core and die as an access violation with no visible parent. Each BPL now carries the IDE's own suffix (Aefos.MCP.Core230.bpl … 370), each IDE's search path points at its own folder, and an unsuffixed BPL is refused from inside the IDE too, where no build script can reach. The suffix is gated on VERxxx, because CompilerVersion is not declared inside a .dpk and an {$IF} on it is silently always true.

A link in a chat message navigated the chat away, taking the transcript with it. A link now opens the file.

The store window opened blank, or not at all. It is shown modeless — a WebView2 control never completes its initialization under ShowModal — and it receives its own user-data folder and its configuration before anything can allocate its handle.

Installing from a folder store copied the bundle instead of registering it, so a change at the source never reached the install.

The addon manager (aefos.exe) did not build below Delphi 12. Two structurally identical reference to procedure declarations are two different types, which Athens forgives and Delphi 11 does not; the addon stack now declares one log type and aliases it everywhere else.

A published Python tool installed where the loader did not look.

1.4.0 - 2026-08-05

Delphi 13's 64-bit IDE is supported. Aefos now installs into the 64-bit IDE as well as the classic 32-bit one, with its own set of packages.

Added

Delphi 13 (BDS 37.0), 64-bit IDE. Delphi 13 ships two IDE executables, and a design-time package is loaded into the IDE process — so a 32-bit package is invisible to the 64-bit IDE and always was. The installer now carries a second set of packages built for Win64, installs them into Bpl\Win64, and registers them under the Known Packages x64 key the 64-bit IDE reads. Confirmed running: chat panel, WebView2 rendering and the MCP server, all inside the 64-bit process.

Nothing changes for the 32-bit IDE, and the two sets coexist: different files, different folders, different registry keys.

Fixed

The 64-bit build reported version 1.0.0. Delphi writes a default version block for every new platform configuration, and the version bump deliberately skipped untagged blocks — correct while only the 32-bit build shipped, wrong the day the 64-bit one did. The 64-bit blocks now carry the same keys as the 32-bit build.

1.3.0 - 2026-08-05

Delphi 11 Alexandria is supported. The installer now carries a BDS 22.0 payload beside 23.0 and 37.0, built and proven in a Delphi 11 VM from this repository.

Added

Delphi 11 Alexandria (BDS 22.0), with one feature withheld: inline completion (ghost text) needs Delphi 12 or newer. It is built on ToolsAPI.Editor.INTACodeEditorEvents.PaintLine, and that unit does not exist before Athens. Everything else — Chat, Terminal, the MCP server, the agent tools — is present.

Fixed

The build scripts did not parse under Windows PowerShell 5.1, which is what a stock Windows has. build-packages.ps1 used && in argument position, which 5.1 rejects at parse time: sixteen errors, nothing runs. Nine scripts also carried non-ASCII characters with no BOM, and 5.1 reads a BOM-less .ps1 as CP1252 — so the third byte of a UTF-8 em dash became a typographic quote, which 5.1 accepts as a string delimiter, closing the string mid-sentence and turning the rest of the words into code. Anyone who cloned this repository without PowerShell 7 could not build it.

String literals over 255 characters stopped the compiler on Delphi 11 and anything older (E2056). There were 108, 107 of them in the generated assets, up to 900 characters each; all are re-chunked to 250, with the payload verified byte-identical.

The libvterm objects would not link on Delphi 11. Its bcc32c reaches the standard streams through the C runtime FILE table rather than the helper the unit already stubs; an empty table satisfies the linker and is never read, since every stdio call here is stubbed.

The CLI build scripts gave up on a machine with only Delphi 11, looking for 37.0 then 23.0 and stopping. They now fall through to 22.0.

The Lazarus installer shipped a README into the user's Lazarus folder along with the 64-bit runtime DLLs. It ships only the DLLs now.

1.2.0 - 2026-08-05

Aefos AI is free software. The source is published, the licence gate is gone, and every feature ships enabled. Delphi 12 Athens (BDS 23.0) and Delphi 13 (BDS 37.0); the installer detects what you have and lets you pick.

Changed

The installer no longer bundles an AI CLI. 1.1.0 shipped Codex and Gemini beside the agent so a fresh install was productive immediately. With the source public that trade stopped being worth it: redistributing third-party binaries adds licence surface and ships versions that go stale between releases, and the installer drops from 109 MB to about 5 MB. Aefos is, and is documented as, bring your own CLI — install the one you already use and point Aefos at it, or leave the path empty and let it be found on PATH.

Nothing is taken away from anyone: the bundled binaries were installed with uninsneveruninstall, so a 1.1.0 user keeps the copy already in %APPDATA%\Aefos\bin. New installs simply do not get one.

Removed

The licence gate is gone. There is no key, no activation, no trial and no editions: every feature ships enabled, including the Terminal, the MCP server and the agent tools.

With the source public, the gate had stopped being a lock and become a liability: under the GPL any fork could remove it in five minutes, while the client kept registering a machine fingerprint in the maintainer's own database — which would have meant every build by every stranger phoning home to it.

So the whole feature was removed rather than disabled: the Aefos.License package and its four units, the activation dialog, the menu items, the trial badge, the "(Pro)" tags, and all 44 call sites across both the RAD Studio and Lazarus editions. The installer now ships 10 runtime packages instead of 11.

The licence code remains in this repository's HISTORY, in the commit that first published the source. That is a fact of git, not an oversight.

Licence

Aefos AI became free software on 2026-08-05. The software is now licensed under the GNU General Public License version 3, with two additional permissions granted under GPLv3 section 7 (ADDITIONAL-PERMISSIONS.md):

• the code Aefos generates for you is yours, under any licence you choose, including a proprietary one — using Aefos places no obligation on what you build with it;
• linking with the Embarcadero RAD Studio libraries is expressly permitted, without which a design-time package could not be distributed at all.

Until this date the software was proprietary, governed by an EULA. That agreement is kept in the repository as EULA-historical.md — the record of what the licence used to be, not a document that still governs the software.

Releases continue to be published from this repository.

Fixed

Publication left references behind, and they are corrected here: the chat package listed two test units under a directory that is not published, so a DEBUG build could not resolve them; the Lazarus register still declared a License menu and described it in four comments; the third-party notices pointed at vendor directories for components that ship embedded in a .pas; and CONTRIBUTING.md still told visitors the source was private, on the page GitHub links from every new issue and pull request.

The READMEs advertised Delphi 11 support, which had never been built. They now say 12 and 13, which this release ships and was built from this tree.

1.1.0 - 2026-07-15

Aefos AI now works out of the box. This release bundles a ready-to-use AI CLI right in the installer, so a fresh install is productive immediately — no separate CLI download, no PATH setup. Delphi 12 Athens (BDS 23.0) and Delphi 13 (BDS 37.0); the installer detects your installed versions and lets you pick.

New — a ready-to-use AI CLI is bundled

• The installer ships Codex (and Gemini) beside the Aefos agent, installed to your user profile. After setup, open Tools → Options → Aefos → AI Chat and Codex is already selected and working — the executor path can stay empty; Aefos finds the bundled binary.
• The binaries are refreshed to the current upstream version at each Aefos release (evergreen), not frozen — you always get a recent build, and updating your own CLI keeps working exactly as before.
• Prefer your own install? A CLI you already have on PATH still takes precedence over the bundled one, so nothing changes for existing setups.

How the CLIs are provided

• Codex and Gemini are bundled (both Apache-2.0). Gemini has no official Windows binary, so Aefos compiles a standalone one from the published package.
• GitHub Copilot can be downloaded during setup — an optional checkbox fetches its official release directly, with an integrated progress page (Aefos never repackages it; Copilot needs its own subscription).
• Claude Code is user-supplied: Aefos detects it on PATH, and the one-click Download CLI action in Options opens the official page.

Under the hood

• New resolver step: a bundled CLI in your Aefos user folder resolves with zero configuration, after PATH so your own install always wins.
• The AI Chat settings (executor, model, path) are confirmed IDE-wide per user profile — editable on the Welcome page with no project open.

Carried forward from 1.0.0: the in-house clean-room WebView2 engine (Chat + MCP Tool Inspector), installed-addon MCP merge, the Reasoning Effort selector, the 17 live debugger tools, local models (Ollama), the live Form Designer tools, RULE #1 intent-guard, and the change-review gutter.

1.0.0 - 2026-07-15

Aefos AI leaves beta — the first stable release.

Added

• Our own WebView2 engine. The Chat and the MCP Tool Inspector now render on an in-house, clean-room WebView2 integration built directly on Microsoft's public COM API — decoupled from any specific RTL edition, opening the path to older Delphi releases (and, ahead, Lazarus).
• Installed addons show up everywhere. aefos install <addon> now reaches both the Chat and the raw terminal: the addon's MCP servers are merged into the project configuration automatically, and a read-only MCP Servers panel lists what the agent can reach.
• Reasoning Effort selector. For CLIs that expose it, pick the reasoning effort in the chat footer next to the model pill; remembered per executor.
• Download CLI link. A subtle link under the Executor path field opens the official download page for the selected CLI — the plugin still bundles none and owns no credentials.

Fixed

• Typed model labels are normalized to real model ids before reaching the CLI, and the
• Agent mode now works with CLIs that sandbox MCP tool calls: some CLIs auto-deny MCP calls in non-interactive turns ("user cancelled MCP tool call"); the agent dispatch now grants the CLI the access it needs — safety stays in the Aefos guards. suggested model list was re-seeded to currently valid entries.
• The Gemini / Ollama provider box in the AI Chat options no longer clips its bottom link.
• aefos --version now reports the real product version.

Changed

• A broad internal code-quality overhaul across the entire codebase — no behavior change, a cleaner foundation for what comes next.

0.30.0-beta - 2026-07-12

The biggest release since the plugin launched: the agent can now drive the Delphi debugger, and it can run entirely on local models — no cloud, no key.

Added

• The agent debugs your code — 17 debug tools, live in the IDE. It sets and clears breakpoints (conditional ones too), runs and stops the project, walks the code with step-over / step-into / step-out, reads the call stack and local variables, and evaluates expressions in the current frame — the same loop you would do by hand, driven from chat or the terminal. Ask it to "find why this returns nil" and it actually breaks, inspects state and tells you, instead of guessing from the source.
• Local models (Ollama). Pick "Local models (Ollama)" in the AI Flow options and the chat runs against your own machine. Test CLI pings your Ollama and fills the model list with the models you actually have installed. Free tier — no API key, and nothing leaves your computer.

Fixed

• The Form Designer is no longer bypassable. SetDFMContent could be used to author UI as raw DFM text — controls that never sprout in the Designer, fields Delphi never declares. The tool now refuses it and steers the agent to the proper Design-mode tools, so a form the agent builds is a form the Designer really owns.
• The terminal no longer freezes on Ctrl+D. The bundled claude profile ran the CLI as the root of the terminal session; exiting it with Ctrl+D left the session alive but dead inside, and the terminal hung. The profile is gone — open your AI CLI inside a normal shell instead, where exiting it just returns you to the prompt. An existing profile is removed for you.
• The chat no longer gets stuck on "working…" after switching provider, and Stop is guarded so it cannot strand the composer.
• Local models never spawn the wrong CLI — a stale shared executor path could launch a foreign CLI for an Ollama run.

Changed

• The bundled agent CLI got a leaner tool catalog, better orientation and conversation compaction, so long sessions stay coherent and cheaper.

⬇️ Free edition: https://www.pubpascal.dev · 💎 Subscription plans (Pro): https://isaquepinheiro.com.br/

[0.29.0-beta] - 2026-07-07

Fixed (refresh 2026-07-08 — installer assets updated in place)

• MCP now connects on default-policy Windows. The stdio↔pipe bridge is launched with -ExecutionPolicy Bypass; a Restricted policy (the Windows client default) used to kill it at startup, so every CLI reported "handshaking with MCP server failed: connection closed".
• "Test MCP" is now truthful for Codex. It performs a live MCP handshake through the same per-run wiring the chat uses, instead of probing codex mcp list (which can never see it) and always reporting "aefos not found in this project".
• The chat header honors your saved executor on the Welcome page. The model list showed the default executor's models and a picked model never took until the first message; the header now resolves the executor exactly like sending does.
• Fixed an IDE crash from the Terminal "Test MCP" button (External exception C0150014, then "Unknown Hard Error" on IDE close). The MCP host is no longer restarted on a no-op apply, and its shutdown can no longer deadlock against an in-flight tool call.
• The Login button now shows "Logged in" when the Codex CLI reports an authenticated session.
• Codex model list refreshed to the current generation (gpt-5.5, gpt-5.4, gpt-5.4-mini) — the backend retired the old gpt-5-codex/gpt-5 slugs. An untouched old list migrates automatically; a customized list is never modified.
• "Test CLI" now discovers the models your account actually supports — for Codex it reads the CLI's own per-account catalog and fills the model list ("Connected … - N model(s) on your account"), the same way the Ollama Test fills the list from your local models. CLIs that publish no list keep the curated suggestions plus the "+" button.

⬇️ Free edition: https://www.pubpascal.dev · 💎 Subscription plans (Pro): https://isaquepinheiro.com.br/

Fixed

• Sending from the Welcome page now works. With no project open, the chat used to echo your message and silently do nothing (the report behind most 0.28.0 complaints). The agent now answers from the Welcome page and can even open or create the project for you.
• A stuck "working..." never bricks the chat again. If the completion signal is ever lost, the composer detects the stale state and heals itself; your typed text is preserved.
• Blank assistant bubbles auto-recover. A corrupted chat page (the cause of empty responses) is detected and reloaded automatically, replaying the conversation.
• Garbled characters fixed (â€" instead of —) in the consent dialog title, the audit log and the tool descriptions the agent reads.

Added

• Send while the agent works. Type and press Enter mid-run: the message shows as a bubble at once, a "⏳ N messages queued" line tracks what's waiting, and each queued message dispatches automatically as its own turn when the current one finishes — the agent keeps the full conversation context. Stop cancels only the current turn; queued messages survive and run next. Starting a new session (or switching Chat↔Agent) discards the leftover queue.
• .pas ↔ .dfm desync guard. Bulk writes that would orphan a component (form declares it, code doesn't) are refused with a machine-actionable reason, steering the agent to the proper Design-mode tools.

Changed

• Faster first reaction on screenshot prompts. The agent is instructed to start building immediately (project first, then component by component) instead of planning in silence.

[0.28.0-beta] - 2026-07-04

Added

• Screenshot the form (CaptureForm). The agent can capture the live form designer to see what it built.
• Design↔Code guidance. The agent receives advisory next-prompts that conduct it smoothly through the Design/Code flow.
• SetComponentProperty accepts named constants. Set a TColor/TCursor by name (e.g. clSkyBlue, crHandPoint), not just a raw value.

Changed

• Change-review gutter (the Cursor-style ✓/✗ diff). Real unified-diff +/- markers; per-unit review state (opening another file no longer disturbs a review in progress); "Wait for my approval" is honoured (a save never silently auto-accepts); InsertCodeAtCursor and ReplaceEditorSelection now route through the same review gate; a pending change re-anchors when you edit the lines above it.
• Smarter Save-All guard. Aefos refuses a Save-All while a .dfm event still has no handler (no more "handler does not exist — remove the reference?" popup); the guard also covers Save-Active-File.
• AddComponent requires a Parent — components land where you intend, first time.

Fixed

• Audit-remediation robustness pass: UTF-8 no longer truncated mid-character on large frames; the whole edit buffer is read (256 KB cap removed); _RefreshModuleFromDisk no longer clobbers an unsaved sibling buffer; the pipe send timeout scales to the frame size; consent prompts are interpreter-aware and tool descriptions are honest about what each tool does.

[0.27.0-beta] - 2026-06-29

Fixed

• IDE version reported correctly. GetIDEVersion no longer hard-codes "13.0"; it reports the actual running IDE (Delphi 12 / BDS 23.0 → 12.0; Delphi 13 / BDS 37.0 → 13.0).
• Terminal find-bar teardown — fixed a use-after-free (the pnlFind window-proc subclass was never restored before the panel was freed).

Changed

• The MCP workspace facade (an 8,200-line god-object) was decomposed into 24 focused SOLID services — a pure delegation shell (−86%), every step validated and build-clean on D12 + D13.
• New WithLiveSource harness seam: all six code-write tools flip the IDE to Code and edit the live buffer through one battle-tested transaction.
• The terminal dock form (4,400 lines) was decomposed into 5 focused helpers (−26%).
• Debug breadcrumbs no longer ship in release builds; dead code removed.

[0.26.0-beta] - 2026-06-24

Added

• Manage your Python tools from the IDE (Pro). A new Aefos PyTools item in the IDE View menu opens a manager to create, edit and delete the Python tools the agent can use (in both Chat and Terminal) — no more hand-editing files. Each tool stays a folder under %APPDATA%\Aefos\pytools (tool.json + main.py); the manager writes them for you. New tools load on the next MCP session.

[0.25.1-beta] - 2026-06-24

Changed

• Issue reporting is now opt-in. A new Issue reporting toggle in Tools → Options → Aefos → AI Flow controls whether the agent's bug/suggestion dialog (ProposeAefosIssue) can open. Off by default — turn it on to send feedback; otherwise the agent can never pop the dialog.

[0.25.0-beta] - 2026-06-24

Added

• Built-in project templates. The agent (and the chat New Project picker) scaffolds a ready-to-build project from a template: Console, VCL, FMX, Library, Package, DUnitX. Each renders the name you choose, ships a clean .dproj, and gets a fresh project GUID so two projects from the same template never collide. Installed to %APPDATA%\Aefos\templates — drop your own folder there to add one.
• DUnitX projects come ready to run. A new DUnitX test project includes the canonical runner (console + NUnit-XML loggers, RTTI discovery) and a sample [TestFixture], so it builds and runs a test out of the box.
• Report an Aefos issue from the agent. When it hits a genuine defect the agent can propose a bug/suggestion: an editable window opens (your text pre-filled, IDE/Aefos/OS versions attached) and Send opens a pre-filled GitHub issue for you to review and submit. The agent never files anything itself.

[0.24.1-beta] - 2026-06-24

Added

• Auto-approve tool permissions. A new "Tool permissions" mode in Tools → Options → Aefos → AI Flow: Ask every time (default), Auto-approve edits, ask before destructive, or Auto-approve everything. Works for both Chat and Terminal; auto-approvals are still recorded in the audit log.

Changed

• Cleaner AI Flow options page — reorganized into clear sections (Tool permissions, Agent edits, IDE behavior, Diagnostics).

[0.23.0-beta] - 2026-06-22

Fixed

• Black Chat panel, attacked at the root. The Microsoft WebView2 loader (WebView2Loader.dll) is now shipped next to the plugin and loaded by full path, so it's found even when it isn't beside the IDE executable (a cause of the black panel on some Windows 10 machines).

Changed

• The installer now checks the WebView2 Runtime up front. If it's missing, setup stops with a clear message and the download link instead of installing into a panel that can't render — and it tells you explicitly when the machine is offline (so you know it couldn't download or verify it). Install the Runtime, then re-run setup.

[0.22.0-beta] - 2026-06-22

Added

• WebView2 diagnostic trace toggle in Tools → Options → Aefos → AI Flow. If the Chat panel ever shows a blank/black screen, turn this on to capture a diagnostic log (%TEMP%\aefos_comp.log) to send with a bug report — no need to touch environment variables. Leave it off for normal use.

[0.21.0-beta] - 2026-06-21

Added

• The Chat welcome shortcuts now do real work. Explain, Refactor, Test, Document, Find and Optimize each run a focused, built-in command — Refactor / Test / Document / Optimize apply their changes through the editor so they show up in the change-review for you to accept or reject; Explain and Find answer read-only. They're also available by typing / in the message box.
• Polished slash-command picker. Typing / shows a tidy panel — header, scrollable list and a footer with the keyboard hints (↑↓ navigate · Enter run · Esc close).

Changed

• "Thinking…" indicator now shows an animated robot while the assistant works, with cleaner spacing.
• Clicking a welcome shortcut now fills the message box directly (it previously popped an empty bar).

[0.20.0-beta] - 2026-06-21

Added

• See what an AI edit changes before you accept it. When the assistant edits your code, the change now shows inline as a before/after diff — the original lines struck out in red, stacked above the new lines in green — so you can tell at a glance what was replaced (works for wide and multi-line edits).
• Per-change Approve / Reject / Annotate, right in the gutter. Each pending change gets ✓ (keep), ✗ (undo) and ✎ (leave a note) controls, plus an Approve All / Reject All pill. Multiple changes accumulate without blocking the assistant.
• Your notes reach the assistant. A note you attach to a change is delivered back to the AI, tagged as approved or rejected, so it learns your feedback on each edit.
• Available in both Chat and Terminal. The change-review works the same in the AI Chat and the AI Terminal.

Changed

• Saving accepts pending changes. Saving the file (or running) keeps the new text and clears the review markers — it never writes a half-reviewed, duplicated file.

0.19.1-beta - 2026-06-20

Fixed

• No more crash on a click in the Chat right after it opens. A mouse event arriving while the WebView host was still initialising could raise an error inside the IDE; mouse input is now ignored until the host is ready (and never fatal).

0.19.0-beta - 2026-06-20

Fixed

• Chat no longer shows a black panel on machines where it failed to render. GPU compositing is now off by default (it was the cause of the black screen on some GPUs/drivers/VMs — the Chat is text, so there's no visible cost), and if the WebView2 host still can't start, the Chat falls back to plain text with a clear "install WebView2" notice instead of a black panel.
• Docked Chat/Terminal reliably comes back. If the pane disappears after a desktop/layout switch, clicking its menu entry now brings it forward again (it used to stay hidden until you switched profiles or restarted the IDE).
• About dialog no longer clips its last line (Chat and Terminal) — it sizes to its content.
• Version is consistent across Chat and Terminal (the splash/About showed mismatched versions before).

Changed

• Chat/executor settings are now editable without an open project (on the Welcome page) — they are machine-global, so you can pick your AI CLI, path, model and log in before opening any project.
• Smoother Chat first-open: a dark "Loading…" placeholder replaces the brief black flash while the WebView2 engine warms up; the panel background paints dark from the first frame.
• More tools show the reviewable diff: inserting a method now routes through the same inline red/green diff (accept/reject, with an optional reason) as editing.
• Installer only asks you to close RAD Studio when it's actually running (no more unconditional prompt).

Notes

• Advanced: set AEFOS_WEBVIEW_ENABLE_GPU=1 to opt back into GPU compositing if your machine has a good GPU.

0.18.0-beta - 2026-06-20

Added

• Reject with a reason. When you reject an agent's inline edit, a themed dialog (matching the IDE's dark/light theme) asks for an optional note — and that note is sent back to the agent, so it learns why the change was refused instead of just that it was.
• Reveal on code insertion. Tools that write code (AddEventHandler, InsertMethod, insert-at-cursor) now flip the IDE to Code view and scroll to the freshly written method, so what the agent wrote is on-screen — not buried below the form's component declarations.

Changed

• Inline diff scrolls to the change. When a diff appears it now jumps to the first changed line, so the red/green block and its Accept/Reject buttons are always visible (no more hunting for them).
• Cleaner generated code. Event handlers added by the agent are now separated by a single blank line (the IDE convention), not two.

Fixed

• Editor crash on diff preview. Fixed an access violation (use-after-free in the editor's diff painting) that could crash the IDE while previewing an inline edit.

0.17.0-beta - 2026-06-19

Added

• Composition-hosted WebView2 (TAefosWebView) for the docked Chat — renders smoothly while docked, fixing the blank/flicker on focus changes.
• Provider drivers for four AI CLIs: Claude Code, Codex, GitHub Copilot CLI, and Gemini, each with its own MCP wiring and remembered model.
• Design ↔ Code (intent → view) harness: design-mutation tools end in the Form Designer; code-mutation tools end in the editor.
• Atomic AddEventHandler — creates the handler in the correct section and wires the .dfm event in one step.
• Node-locked single-seat licensing (Community / Pro / Enterprise) with offline grace and a built-in trial.

Changed

• Inline diff routing broadened (edit unit, replace-in-editor, full-buffer rewrite) with accept/reject (Tab/Esc).
• License terms moved to a proprietary EULA with a free Community edition (including internal business use).

Security / supply chain

• Published a machine-readable SBOM (CycloneDX 1.5) and a security disclosure policy (coordinated vulnerability disclosure).