IA OlympusAsistente de la web
# Limitations and Non-Certified Scope - Linux 1.0.5 ## 1.0.5 status - TypeScript, 102 Vitest tests, 4 Playwright tests and the bilingual tutorial validator pass. - The 1.0.5 AppImage and DEB were rebuilt. The AppImage launched through the capture flow and the DEB installed and exposed a live application target in disposable Debian Bookworm containers. - The language selector now consists of two accessible buttons. The install review closes when installation starts so the progress panel is visible. - The tutorial is a static local site. Its screenshots use a local HTTPS mock and a fake tutorial key; they do not prove a real conversational session. - Live conversations were certified against the 1.0.2 code base. The seven-agent live matrix was not re-run for 1.0.5 because the agent, network, conversation and storage paths were not changed. - The 1.0.4 notes below remain historical and still describe the inherited limitations. ## 1.0.4 status (historical) - TypeScript, 101 Vitest tests and 3 Playwright tests pass in the Linux container. - The 1.0.4 AppImage and DEB were rebuilt and launched; older 1.0.3 hashes are historical. - Disk-space detection uses `fs.statfs` on the application install volume. Filesystems where `statfs` is unavailable block installation rather than reporting a false positive. - The check covers the application install volume, not arbitrary project disks or separate mounts a user may fill later with generated code. - The UI explains that Node.js is never installed, updated, reinstalled or uninstalled by the app. - The operating-system dropdown was removed. It never installed Windows software from Linux and did not use Docker. - Live conversations were certified against the 1.0.2 code base. The seven-agent live matrix was not re-run for 1.0.4 because the agent, network and conversation paths were not changed. This document states exactly what was and was not proven. It is intentionally narrower than a general “works on Linux” claim. ## Certification boundary The tested runtime was Debian GNU/Linux 12 (bookworm) amd64 in Docker, with Xvfb 1280x800x24. The 1.0.4 AppImage launched and the 1.0.4 DEB was installed and launched in disposable containers. There was no physical display, DBus session or system keyring. The following flows passed in that environment on the 1.0.2 base for Codex, Claude Code, OpenCode, OpenClaw, Hermes, Reasonix and the official CodeBuddy CLI: 1. Install or detect the agent. 2. Write and verify a IA Olympus-only configuration. 3. Validate the stored key against `GET /v1/models`. 4. Start a one-turn, non-interactive conversation in an empty directory. 5. Receive `OK` and identify the assistant as IA Olympus. The 1.0.4 delta adds real disk-space checks and clearer Node.js/filter UI; the agent and configuration paths are unchanged. The 1.0.4 pass revalidated build, UI, packaging, AppImage launch and DEB installation/launch. It did not repeat the seven-agent conversation campaign. This does not certify every Linux desktop or every future release of a third-party agent. ## Not certified ### WorkBuddy / CodeBuddy desktop The desktop application remains a guided installation. The official CodeBuddy CLI 2.176.0 was installed, configured, validated and conversation-tested, but the desktop application's effective model loading was not tested. The public label therefore remains `guided`. ### Qoder Qoder is intentionally unavailable because its current assistant does not support a safe custom base URL. The app refuses both install and configuration requests and creates no Qoder files. ### Physical desktop environments Not tested: - Ubuntu or Fedora outside Docker. - GNOME, KDE or another desktop shell. - Wayland. - X11 outside Xvfb. - FUSE-mounted AppImage execution. The AppImage was launched with `APPIMAGE_EXTRACT_AND_RUN=1` because the test container had no desktop/FUSE setup. - Notification services, tray integration and desktop file association behaviour outside a minimal XDG environment. - Opening `https://iaolympus.com/` in a real desktop browser. The packaged UI link and the allowlisted IPC call were verified, but the container had no default browser. ### OS keyring No real DBus/keyring backend was available. The tested storage backend was the documented AES-256-GCM fallback. Its file permissions were `0600` and its directory `0700`, but it is not equivalent to a hardware-backed or OS-managed keyring. The `safeStorage` path must still be exercised on a real desktop. ### Linux signing and distribution - The AppImage and DEB are unsigned. - No GPG detached signature or repository metadata was produced. - No update feed, repository hosting or auto-update exists in this version. - Nothing was published. ### Windows Windows is outside this Linux certification. Its source and artifact are separate, and the Windows 1.0.3 candidate is pending validation in a Windows VM. The Windows result must not be inferred from this Linux evidence, and Linux 1.0.5 does not certify any Windows artifact. ## Agent-specific limitations - **Codex:** tested with `-s read-only`; no tool calls were observed. The agent remains capable of changing files when the user later runs it with broader permissions. - **Claude Code:** it reported that it does not recognize `iaolympus-low` and assumes a 200k context window. Conversation still passed. The message is cosmetic/context-related in this environment but should be revisited if the model's real context window differs. - **OpenCode:** it has no formal switch equivalent to `--tools ""`. No tool events appeared in its JSON output and the empty-directory probe passed, but a formally tool-disabled sandbox was not imposed by the harness. - **OpenClaw:** it catalogued 40 available tools behind its prompt surface, but the stable JSON envelope reported `bridgeCalls.search/describe/call = 0` and the default IA Olympus provider was used without `--model`. - **Hermes:** toolsets were disabled for the probe and no tools were used. Hermes was tested on native Linux; WSL2 on Windows is a separate path. - **Reasonix:** it warned that its bash sandbox was unavailable and refused to run unconfined bash. This is safer than silently disabling the sandbox, but it means bash-dependent workflows were not exercised. Its private `.env` is a documented exception to the environment-variable-only key rule. - **CodeBuddy CLI:** `--tools "" --permission-mode plan` was used. The desktop app was not tested. ## Configuration and repair limitations - Generated configurations are verified structurally and by reading them back. Agent configuration parsers can still change between agent releases; the catalog must be rechecked against official documentation for each release. - Repair preserves keys and tables but normalizes TOML formatting. A harmless comment used as a sentinel was not preserved, although the unrelated table and its value were. Do not rely on repair to preserve comments or exact whitespace. - Backups are created before modification, but backup retention and restoration are the user's responsibility. - Reasonix requires writing the key to its private `.env`; if that file is copied, backed up insecurely or synchronised, the key may leak. The app applies `0600` and removes the key on uninstall/clear when the file contains only the managed setting. ## Test-harness limitations - The process-runner tests use simulated processes; they verify arguments, timeouts, shell avoidance and error classification without installing real packages on every run. - Playwright runs against the renderer with a simulated IPC bridge. The live pass additionally exercised the real packaged Electron process through the Chrome DevTools Protocol, but automated CI does not yet do that. - The live conversation probes are deliberately minimal. They prove connection, model routing and a real response, not the quality of generated code or every agent tool. - The key was exposed in chat during the test cycle and must be rotated. No key is stored in the repository or evidence. - The local Docker image `iaolympus-linux-test:ready2` contains a `IAOLYMPUS_API_KEY` environment entry from the test campaign. It is test infrastructure, not part of the release. Rotate that credential, do not export the image and remove it from the local Docker store when it is no longer needed. ## Before a public Linux release 1. Rotate the exposed IA Olympus key and repeat only the minimal validation needed with the replacement. 2. Test AppImage and DEB on at least one clean Ubuntu and one clean Fedora desktop, including Wayland and X11. 3. Test the real OS keyring and the fallback warning on a desktop without a keyring. 4. Sign the artifacts and publish `SHA256SUMS` plus a detached signature if available. 5. Re-run the live matrix against the exact versions intended for release. 6. Complete the guided WorkBuddy/CodeBuddy desktop validation or keep it explicitly uncertified. 7. Keep Qoder disabled until a supported custom endpoint mechanism exists.