变更记录一览。

Features

• mount Advanced fields + Per-repo overrides on the Default card

Bug Fixes

• merge the base into a fleet PR only on proof, never on a timer

Features

• box-side push paths are screened before a tool-using agent reads them

Features

• per-repo thinking depth, key, provider and task-tag pins on the Agents page

Features

• developerz.ai's own nightly self-improve checkup

Bug Fixes

• read.claude front matter the way Claude Code does

Features

• an operator-comped free account is served as solo at every capability gate

Features

• platform.health aggregate telemetry for the self-improve checkup

Features

• a babysit round waits while a person is pushing to its pull request

Bug Fixes

• a plan's headline skips the plan-files bookkeeping group

Features

• user-token lane runs end to end

Features

• triage answers whether each report is true, and an outsider's ask mints work only when it is

Bug Fixes

• refuse restarting a slice whose plan is done

Features

• reasoning HIGH by default where provider and model support it

Bug Fixes

• a setup the base is proven to fix is refreshed now, and the team lane blames a setup its branch broke

Features

• a box reports whether its unit carries the process-death reap, and the fleet names the one-time command

Bug Fixes

• /proof names one filer for the benchmark feature

Features

• working agents only receive screened third-party text on mothership push paths, with a guard

Features

• a repo declares its services' up and down; down runs first in every close, and a missing down is a readiness gap

Bug Fixes

• a summary line never acquits a red diff, and a base gate red on the boxes files one fix

Bug Fixes

• a stale PR head is asked its base's question before its setup failure is counted

Features

• team self-review before commit, repo-scoped editor checks, no whole-run retry on an identical red

Bug Fixes

• every run's scope is closed on every exit, and a dead run's by the next start

Bug Fixes

• the box formats before it gates, owns a shared red on its own paths, and a babysit round is graded before it lands

Features

• an isolated, tool-less LLM screen runs on every raw ask before triage

Bug Fixes

• an App's own review threads are ours; GraphQL spells its viewer <slug>[bot] and its comments <slug>

Bug Fixes

• waiting babysits no longer hold back the review-only reserve, so a capacity-1 key stops parking every review

Features

• every raw task records author trust at intake
• Agents page Advanced drawer + mode switches

Features

• triage sees open human PRs and can defer an ask citing one

Bug Fixes

• glob matching is polynomial and test-only, so a stacked-** glob cannot freeze the worker

Features

• a scheduled checkup files its findings with anchor, paths, metric and evidence, behind a gate-path floor

Bug Fixes

• correct the n8n signature claim and three competitor rows

Bug Fixes

• fragment links land with headings under the sticky nav when motion is on
• the Policy tab shows a broken snapshot as invalid-with-reason over the effective policy, not a 404

Features

• schedules: key and an hourly lane that fires a scout checkup
• a never-planned coarse task plans on an idle box without its declared area, and the drain counts plans

Bug Fixes

• OSS Free tagline names orchestration, not "the agent"

Features

• ModelPicker and model saves on the agent cards

Bug Fixes

• a signal-killed leg outranks a plain failure, so the verify reporter's exit no longer depends on step order

Features

• a coding run locks its PLANNED set before its editors write

Bug Fixes

• a re-pended run's usage folds on the task's tenancy, and every other owner hook still needs the assignee