Reference · Changelog
Changelog
A hand-authored record of what shipped, newest first. It tracks releases, not commits — so it leans toward what changed for you, not every internal step.
v0.2 — current
The release that turns Agentry from a conductor-plus-memory core into a product you can see and steer: a multi-page docs & marketing site, a live Workbench, a run-state engine with a live channel bridge, and a self-eval harness with a dashboard.
Added
- The docs & marketing site — this multi-page docs area, covering Getting Started, Capabilities, and Reference.
- The Workbench (Agent Center V2) — a live graph of a run, a docs workspace, a comment/approve loop, and a Memory browser.
- The Flow MCP v1 — the run-state engine that records task lifecycle, append-only events, and review as plain files in
.agentry/work/. - The channel bridge — live comments, threaded replies, a permission relay, and human status changes that steer the run.
- The self-eval harness + dashboard — routing and decision-quality measures, controls, and a corrections log, with an embedded dashboard. Now includes a flow-compliance probe that grades a run on observable Flow signals.
Changed
- Agentry enforces its own process. Above the one-shot floor,
/agentry:gonow opens a Flow run and records the routing decision before work begins — so escalated work is resumable, gradeable, and visible in the Workbench, with a runtime guard that flags a bypass. The one-shot fast path is unchanged: trivial work still runs with zero ceremony. - Install & Channels docs now distinguish
--channels(official, allowlisted channels) from--dangerously-load-development-channels(custom/preview channels, like Agentry's own bridge today).
Fixed
- Task documents no longer render with a doubled header in the Workbench (the architect now authors a single, clean task frontmatter).
v0.1 → v0.1.2 — initial Agentry
The first releases established the core: a right-sized conductor, the SDLC specialist roster, durable memory, and the primer that warms a session.
Added
- The right-sized conductor — routes a task to the least process that wins (one-shot → spec-first → decompose + verify).
- The SDLC specialist roster — the explorer, architect, implementer, verifier, librarian, and the rest of the specialist agents.
- The memory MCP — durable facts and episodes the conductor can recall, so run #2 is warmer than run #1.
- The primer — a minimal session nudge that points the conductor at the repo's conventions and memory.