Skip to content

v0.4.1

Choose a tag to compare

@kesteva kesteva released this 16 Sep 17:09
· 978 commits to main since this release

[0.4.1] — 2026-09-16

Added

  • Custom views. A customize mode for the project surfaces: a draft lifecycle with a header
    switcher, editable blocks, a widget library with a built-in catalog, per-widget settings, and
    save/manage dialogs. Widgets render in a sandboxed frame served from a loopback document server
    and read the backlog through a read-only query executor with a data cache and circuit breaker.
    The assistant can author widgets too — db_schema, widget_preview and widget_save tools, a
    placeholder slot with a live draft preview, publish/discard, and a composer kickoff carrying the
    context hint; a session-less widget can request publish into the library. (Migration 133.)
  • Two-way approve-design gate. The Planner, Ship and Launch flows' approve-design gate now
    declares a loopback: Revise re-runs the design pass with the reviewer's note threaded into every
    re-run step, Approve continues. The gate body is built from the adversarial review and every
    human gate shows its pending-findings count; the review-queue buttons are labelled as
    continue-and-log vs rerun-planning. A gate side-effects singleton binds designs on
    approve-ideas/approve-design, stamps thoroughness, files accepted-risk findings and reconciles at
    settle. approved_designs accepts flow-sourced rows (migration 134).
  • Solution thoroughness. Launch stamps a per-project thoroughness level (prototype → efficient,
    v1 → standard, production → thorough; migration 135), the wizard defaults its tuning level from
    it, and the step-prompt contracts carry thoroughness budgets, Design spec folds and design
    surfaces. An approved design opens from task/epic cards, task detail and sprint lanes.
  • Adversarial review artifact. A machine-parseable adversarial-review artifact type
    (migration 136) with its own parser, rendered as a tab on the artifacts view.
  • Onboarding checks for git. A machine without git — or with git but no user.name /
    user.email — used to fail 30 s into the first session with a generic quick-session timeout.
    The tour now probes git at boot (over tRPC) and blocks on a prerequisite card with per-platform
    install lines and a "Check again" that re-resolves PATH; the worktree's initial commit also gets
    a fallback identity when git has none, and session-creation job errors surface immediately
    instead of after the 30 s timeout.
  • Assistant composer image attachments, with a cap applied against the live list when pastes
    overlap; the composer shows which model is running the assistant, and the rail header's gear
    deep-links to Settings → Assistant.

Changed

  • Quick-session briefing rides the system prompt. The Claude PTY lane spent every quick
    session's first turn acknowledging its own briefing, sent as the argv prompt. It now goes out on
    --append-system-prompt (its own sessionBriefing option — systemPromptAppend remains the
    workflow channel), the prompt slot carries only a real user turn, idle spawns rest at the
    turn-end status, and resume/respawn receive the briefing too.
  • Transcript discovery is pinned to a minted --session-id instead of binding the first new
    *.jsonl in a directory shared by every process with the same cwd — which could adopt a
    stranger's conversation. Discovery arms on the first turn, and a deferred deadline re-defers
    instead of latching.

Fixed

  • Blocked ≠ failed in the programmatic plane. A lane whose prerequisite failed was written
    failed, so blocking spread transitively ("55 failed" for 4 real failures). Lanes now carry a
    separate blocked state, the partial-sprint gate lists never-started lanes with the prerequisite
    they wait on, and the built-in Sprint/Ship definitions self-loop implement once so a first-step
    failure costs one attempt, not the lane.
  • Systemic errors park the fan-out instead of failing lanes. Verb-first session/usage-limit
    errors ("hit your session limit") classify as systemic; three lanes failing with byte-identical
    error text corroborate into the park-and-resume path; lane-triage consults serialize behind the
    systemic latch; deferred failures persist on cancel; and a fan-out whose triage itself dies on a
    systemic error is parked rather than failed.
  • Run cost no longer overcounts resumed sessions. total_cost_usd is cumulative per SDK
    process, so summing every result inflated any resumed run ($1,976.97 shown for ~$125 real). Cost
    is now laddered per (run, session) segment; migration 132 drops the stale run_usage rows so
    the boot backfill recomputes them. An empty or partial modelUsage is treated as not comparable.
  • Cyboflow's git excludes live in .git/info/exclude, written by one shared helper, instead of
    being appended to the project's tracked .gitignore (which left an untracked file that blocked
    the in-app fast-forward merge). Merge failures now lead with git's own output.
  • The Sprint batch cap is enforced inside the lane store, so the assistant's launch-run seam can no
    longer bypass it (ship_batch_too_large).
  • The custom-views draft Save widget is a primary CTA; the customize dialogs lose their doubled
    close button; the draft preview polls the draft spec for authoring.
  • The dl.cyboflow.com redirector serves the Windows installer (its allowlist was per-extension
    and only knew .dmg).