Files
byre/TODO.md
T

3.4 KiB
Raw Blame History

Bugs

  • Codex 0.163 notifications byre doesn't know The service's codex is 0.163.0, and its app-server sends thread/readState/changed, which internal/codex rejects as unrecognised method (a protocol_errors entry in /api/admin/status). Regenerate the types with go generate ./internal/codex against the installed CLI and handle or ignore the new methods.

M1 — One agent end to end

  • Cancel a queued message A × on the transcript's "queued…" bubble withdraws the task_inputs row with a new DELETE, refused once the message has been handed over, and puts its text and images back in the composer. Claude Code, Zed and Devin all let a queued follow-up be taken back.

M2 — Seeing the work

  • Comment on diff lines in the Changes view Clicking a diff line opens a comment box under it; pending comments live in Drafts.swift beside the composer draft, and "Send N comments" posts them as one ordinary message, each as path:line, the quoted line and the comment. No new API. Conductor, Claude Code desktop, the Codex app, Warp, Factory and Superset all send batched line comments to the agent this way.
  • Screenshots from implementers Give implementer sessions the reviewer's playwrightMCP entry, so a turn that changes something visible can embed screenshots it took in its summary, as Cursor's artifacts, Devin's test recordings and Claude Code's autoVerify do.

M3 — Second vendor and review

  • Continue on another model or vendor with Compact Compact takes an optional model; the handoff is written by the old session as now, and the continuation session starts on the chosen model on either vendor, reusing handoffPrompt and handoffPreamble. Superset's "Continue with another agent" and Factory's /handoff do the same. It can't rescue a vendor that is rate-limited, since the old session writes the handoff.

M4 — Merging and history

M5 — Workflows

  • Scheduled tasks A schedules table (project, cron, prompt or workflow) and a timer that creates the task the way the lead's create_task does, managed with lead tools and no UI, like workflows. Example: every Monday, run go vet ./... and gofmt -l and check docs/operations.md against the code. Claude Code Routines, Codex Automations, Copilot cloud automations, Devin, Factory and Conductor all schedule agents.

M6 — Fan-out

  • Agents talking to agents A more collaborative version of forks and workflow handoff: agents in related tasks (a workflow run's stages, a task and its forks or subtasks) talk to each other while they work, instead of one stage passing a single output to the next. Claude Code agent teams (shared task list, per-agent mailboxes) and Claude Code desktop's cross-session messaging are the nearest designs.

M7 — Inbox triage

  • Close stale notifications once the item is dealt with Notifications.swift removes each delivered notification by its identifier when inbox.updated leaves the item no longer open, or summary.read covers its turn-<seq>, so a question answered on another Mac or by the lead leaves Notification Center. Factory clears notifications when the session is focused.

M8 — macOS client

  • Phone app A native client for the phone over the same API, sharing the macOS client's SwiftUI code where it can. Every surveyed tool has one (Claude, Codex, Cursor and Conductor apps, GitHub Mobile, Superset for iPhone); it needs its own layout, since two half-width panes don't fit a phone.