Repository navigation
Dev Companion 0.5.0 — local help that follows the work
Pre-releaseDev Companion 0.5.0 — local help that follows the work
Free Linux x86_64 workflow preview; application source remains private. Git is required. A downloaded Ollama model is needed for reasoning; reminders, incidents and the work desk work without inference. Initial downloads and explicit updates need internet.
What changed
- Separately opted-in background investigation: new saved editor errors and captured failures can produce a local hypothesis with chat closed. One worker, quiet/cooldown controls, bounded source/token/time budgets, no automatic commands, checks or patch application.
- More Python/pytest, Node/TypeScript, Go and ESLint formats alongside Rust. Related import context supports Python src/ packages and separates reported errors from current disk source. Correction tasks return a concrete draft or ask for missing evidence; reviewable diffs retain context.
dot setup --checkreports readiness.dot setup --yes --presence --proactive --goal "Your task"combines explicit project memory, login presence and background investigation opt-in.dot update --checkanddot updateuse official GitHub HTTPS releases, verify archive checksum/layout/architecture/version and replace the executable atomically, keeping one previous copy. Memory, models and alias remain; a matching managed presence service restarts. Version 0.4 and earlier need the one-time archive bootstrap described in the guide.- Voluntary local trial sessions, helpful/incorrect/noisy ratings and private missed-error notes. Aggregate reports omit code, paths, identity and notes, and stay local until manually shared.
Evidence and limits
95 Rust tests passed (11 unit, 84 integration), formatting and Clippy passed, and a static binary was built. A real downloaded 4B model explained a billing failure automatically with chat closed. The maker requested, inspected and applied its one-line correction; all nine unchanged tests passed. A final returning turn correctly identified the current check/run/count.
A larger official Flask 3.1.2 checkout exposed wrong model patches, a missed empty-string case and invalid replacements. Review and unchanged checks exposed these errors. A billing recap also invented example details, despite identifying the right passing run. The recorded demo retains these failures. Evidence labels establish a real location, not a correct diagnosis. These are maker evaluations, not independent developer trials or a general reliability benchmark; v0.5 evaluations were not network-isolated.
Optional English voice uses local Whisper/Piper with an explicit recording turn. Generated audio passed in v0.4; the voice code and VS Code adapter are unchanged. Physical microphone/playback comfort remains untested. The optional VSIX requires separate trusted-workspace consent. Presence observes enabled adapters, not arbitrary screens or all terminals. No cloud fallback or default application telemetry.
Install and try
Download the archive and SHA256SUMS below; check the archive hash before extraction. The archive contains only the executable, preview license, third-party notices and quickstart. Checksums verify consistency; they are not independent signatures. Install into ~/.local/bin and use the dot alias. Existing chats should quit and reopen after updating.
Setup/update/privacy guide · Recorded workflow, including failures · First developer trial recipe · Feedback
Binary SHA256: a6e59468b0fbc145e59f21bae2c635824e42068be38045e20d383859af6e22b2
Archive and VSIX hashes are in SHA256SUMS. Please try one actual failure or reminder and tell us what helped, what was incorrect or noisy, and whether you returned. Downloads do not establish unique users or retention.