Hydrate v0.10.0: the memory now works on itself.
Your agents already share memory, and since v0.9.0 they can talk to each other over Peernet. v0.10.0 turns the memory inward. A background pass now mines what your sessions left behind: contradictions between what you decided last month and what you did yesterday, and patterns that no single session can see. This is a feature release, a minor-version bump. The memory-hook path you use every day is unchanged, and everything here is additive.
The dream cycle: memory that works while you don't
Run /hydrate-dream, or let the daemon trigger it in the
background, and Hydrate reads the project's own record for things a live
session cannot. It scans genuine conflicts between pinned canon, session
facts and recent work, matched on prose facts rather than just short
scalars, and it batch-confirms them so the cycle degrades gracefully
instead of failing closed.
It also reads the .hydrate/handovers/ archive, every distilled
session note, and mines it for cross-session patterns: decision reversals,
recurring blockers, the way an approach evolved over weeks. Observations
have to cite the handover they came from; uncited ones are dropped. You
recall the results with /hydrate-dream-recall, the command
formerly known as /hydrate-week.
Background brains that fail over
All of that needs an LLM running in the daemon. It now resolves one through
an ordered fallback cascade, subscription CLI first, then a local model,
then an API key, with bounded timeouts and a breaker on each rung. A hung
claude binary or a dead local endpoint degrades to the next
rung instead of silently switching off dreams, distills and extraction.
Background intelligence that quietly dies is worse than no background
intelligence; this closes that gap.
Orchestration: author a fleet in the browser
Fleets no longer have to be described in text. From the dashboard you can add agents, set each one's model and project directory, draw the dependencies between them, and hit Run, with per-agent critic overrides. All four modes roll up into one unified authoring canvas, so the fleet you draw is the fleet that executes.
Producer roles now fail over down a per-role model ladder instead of dying with their body. The implementer and the Studio builder default to Claude Sonnet 5, and Fable 5 is restored as the design specialist now that Anthropic brought it back, so Studio designs and design-critic rounds bind Fable first again. The audit gate stops wedging: cross-family writable failover, a hard deadline on the audit worker, and reaping of orphaned fleets left working by a daemon restart. Worker token and cost capture is honest now, with per-orchestration spend and truthful timers on the dashboard.
A dashboard that works on your phone
A responsive foundation across all fourteen routes, a redesigned mobile KPI hero with a swipeable rail, and a full 390px and 768px overflow sweep. Two long-standing symptoms, the blank home page and the live-feed pill stuck on "connecting", turned out to be the same thing: one uncached fifteen-second transcript scan on a hot path. It is now a single-flight cache, and both are instant.
Quality of life
/hydrate-yagni and /hydrate-yagni-spec are
user-invokable concision directives for build and spec work.
hydrate handover diff and hydrate skill-try are
new, hydrate dev gc reaps terminal-session worktrees, and
/hydrate-release-notes drafts a changelog from your own
history. Distill gained a zero-LLM pre-extraction draft and batched rails,
faster handovers with the same fact fidelity. Windows got consistent
.exe hook wiring and an .exe-tolerant doctor.
Upgrading
brew upgrade hydrate, then hydrate doctor to
verify hooks, daemon and runtime wiring. There are no schema migration
steps: the daemon applies the additive dream-synthesis tables on first
start. If you never run a dream cycle or a fleet, this release changes
nothing you touch, and that is by design.