Previous changes
Release notes for the three most recent releases live in their own pages. Everything older is summarised here, newest first, with a link to the full GitHub release for each.
0.19.1 — a model newer than the catalogue
Section titled “0.19.1 — a model newer than the catalogue”A patch release that fixed the endpoint used for a model id newer than the bundled catalogue. An uncatalogued
id is built for the OpenAI-compatible adapter, and its base URL needed a version segment the provider’s own
catalogue base does not: the call became a 404 instead of answering. The provider registry also never carried
the endpoint, so bazilion provider test failed closed for exactly the models that endpoint was meant to
admit. Verified against a real model newer than the catalogue, through both the verification and review
loops. No schema change.
0.19.0 — a teammate can review a fixed revision
Section titled “0.19.0 — a teammate can review a fixed revision”A coder can ask a teammate to read one captured change. The reviewer’s entire capability is four read-only
tools — no execution, no editing, no browsing, no publication — findings are append-only and resolution needs
explicit proof, and an export is a publication held until policy releases it. The same release added
request_verification, so a coding Agent can ask for specialist verification at all, and closed three gaps in
specialist verification: a check now runs with the posture its receipt claims, declared output paths are
checked rather than trusted, and a check can no longer write the Team’s shared memory. Schema change.
0.18.0 — hand a captured change to a specialist
Section titled “0.18.0 — hand a captured change to a specialist”A coder hands one captured change to an existing member of its Team and gets back evidence naming the code, the commands, the environment and the outcomes. The specialist’s capability is two tools — read the request, and run one declared check once — and it cannot widen them. Results are tied to the exact captured revision. Schema change.
0.17.0 — coding work becomes reviewable
Section titled “0.17.0 — coding work becomes reviewable”Live command progress, retained diagnostics, and read-only Git change review with source-bound evidence.
- Watch a command while it runs, with output that arrives as it happens rather than when it finishes.
- Reopen what a command printed after the container is gone — retained, bounded, redacted evidence with an explicit expiry, and a truthful “not shared” state rather than an empty box.
- Review what changed: a read-only Git review of the working tree against a pinned baseline, with source snapshots that tie a result to the exact revision it tested.
- Schema change: 0.17.0 could not be upgraded in place from 0.16.x.
0.16.0 — coding work becomes agent-led
Section titled “0.16.0 — coding work becomes agent-led”Repository context, scoped commands and prepared execution during ordinary Agent tasks — no check catalogue or readiness form first.
- Ask for the work, not a checklist. At the start of a coding turn the Agent receives the applicable repository instructions and can resolve a deeper scope itself.
- Run a finite command in the turn’s admitted environment, with the outcome observed rather than asserted.
- Hand work to a teammate, with workspace preparation and recovery that survives interruption.
- Schema change: 0.16.0 could not be upgraded in place.
0.15.0 — retain, follow up and inspect work
Section titled “0.15.0 — retain, follow up and inspect work”Saved results, retained conversations, durable follow-ups, live questions and opt-in Attention notifications.
- Keep files and conversations: deliverables are published into a Team results library instead of scrolling away, and conversation threads are retained and named.
- Follow up while an Agent works: a bounded queue for steering a running turn, and questions the Agent can ask you without ending its work.
- Opt into operational notices: Attention notifications are off by default and forbidden from leaking any message payload into notification metadata.
- Schema change: 0.15.0 could not be upgraded in place.
0.14.x — GPT-6 Astra, and a safer first run
Section titled “0.14.x — GPT-6 Astra, and a safer first run”- 0.14.2 upgraded the bundled engine from Pi 0.84.3 to Pi 0.85.1 and added GPT-6 Astra selection through OpenAI API keys and ChatGPT OAuth.
- 0.14.1 improved recovery and failures: clearer recovery guidance when a home is incomplete, and failures that name what was wrong instead of failing silently.
- Earlier 0.14 work covered first-run setup, provider configuration at small screen sizes, mobile pairing with trustworthy streamed chat, protected drafts with explicit consequences, and gateway reliability.
0.13 — the security-focused release
Section titled “0.13 — the security-focused release”Credential-minimal protected execution, a private Tailscale gateway, encrypted recovery, hardened Telegram ownership, and Pi 0.84.3.
- Credential-minimal protected execution: Agent turns run with only the selected provider credential they need, never the daemon’s ambient environment.
- Private HTTPS through Tailscale Serve, with a strict production gateway profile and QR pairing over the same origin.
- Encrypted backup and safer recovery, with validated restore and rollback.
- Telegram owner pairing, so only the paired owner drives an Agent.
Older releases
Section titled “Older releases”Everything before 0.13 is covered by the commit history and the tagged releases.