SHIPPED · v1.81.0
Mention chips that show what they do, and an Activity tab for orchestration
New
Every mention chip shows its kind at a glance — in Agent Input, a chip that starts a new agent has a dashed outline and a ＋ on its logo, a linked open terminal has a solid outline and a live dot, a project shows its avatar and emoji under a folder tab, and a terminal on a paired computer carries that computer's name in a shaded segment with a green or amber presence dot
The chip's name stays at the size of the text around it, the brand colour lives only in the logo, flags such as --dangerously-skip-permissions read muted, and a chip looks the same whether its name is 4 or 40 characters long
A chip is written into the draft as @name, like a file mention; what the agent receives on Send is unchanged, and drafts saved before this release keep their chips
An idle composer that tells you the agent's state — with Agent Input closed, the pill over an AI terminal is now a compact capsule that reads Ready, Working or Draft, sits on the divider above the agent's own status line instead of covering it, and shrinks to fit its pane in 2×2 and 3×3 grids; its ⋯ menu can hide it in this pane, and the pane header's Agent Input button takes over
Activity tab — the Orchestration dashboard opens on Activity: what needs you at the top, then the timeline, with All, Needs you, Links, Teams and Runs filters; Mission Control shows the same view by default
Agents' link requests approve themselves — a new Auto-approve agent link requests switch, on by default, accepts a link the moment an agent asks for it, so a worker can report back to its lead without anyone clicking Approve link; a request that asks for read access still waits for you, and turning the switch on releases the requests that were only waiting for it
A link switched to Auto delivery now releases the messages already queued on it, instead of needing Submit now for each one
Logs & runs, restyled — find in the log, soft wrap, failure chips on runs, the error shown in the run drawer, and Export
Settings → AI → Orchestration is a page — Delegation, Models, Skill install, Links & approvals and Diagnosis & Autopilot on one sub-navigation, with the link policy as switches and pending requests listed the way Activity → Needs you lists them
A phone-app host behind a switch — Settings → Remote is split into Web UI and Phone; the Phone section holds the host switch (off by default), the addresses a phone can reach, a pairing code with Copy code and a fresh-code button, and the paired phones with rename and a Viewer, Approver, Operator or Admin role each
The phone reaches this computer over your local network or Tailscale; the Web UI keeps its tunnel
A phone pairs as Viewer and 1DevTool asks you which role it gets; an invitation ends after five failed attempts, one address may open twenty pairing attempts a minute, and up to 32 phones can be paired
Improved
Export diagnostics to send to support — the download button in the Resource Manager, or Settings → General → Diagnostics, saves one .zip with memory, CPU, open handles and threads for every 1DevTool process, recent stalls, versions and the updater log; the card lists what is inside and what is not before you export (no terminal text, prompts, files, project paths or keys), and 1DevTool never uploads it
New pairings start with every grant — a computer you have just confirmed by fingerprint on both screens can start terminals, orchestrate, share memories, run headless and install the skill right away; Settings → Machines still narrows any grant per computer, and existing pairings keep what they had
Faster git on Windows — every git read used to start three processes (Git for Windows' launcher, a console host, then the real git) at about 42 ms; 1DevTool now starts the real binary once, at about 20 ms, including Git for Windows 2.56, and git no longer rewrites the repository's index file on every status read
Hang reports name what actually blocked — a stall report in the Resource Manager now blames a call only when it held a real share of the stall; short reads that merely happened during it are listed as bystanders
Fixed
Fixed a prompt to a paired Windows computer failing once, then every retry refused as "grant revoked" — after one timed-out send the edge was treated as revoked and the chip ended on "no longer available"; a retry now reuses the same link, a timed-out message no longer marks the computer offline, and a terminal that was closed and reopened on the other computer is re-bound to the chip with one more Send
Fixed a Claude on a paired computer never becoming ready — Claude's "Quick safety check: Is this a project you created or one you trust?" on a folder it had not seen was not recognised, so a send waited 30 s and failed; the question is now recognised and answered like the other first-run prompts, and a send that still meets one fails at once naming the question instead of waiting
Fixed Open in Windows Terminal failing with error 0x80070002 — the Add Terminal option for a native Windows Terminal tab passed the launch command as one quoted token; the tab now opens, and its shell stays open after the agent exits
Fixed dragging a chip to move or swap it doing nothing — in Agent Input, dragging a chip to another spot or onto another chip, and dropping an agent from the palette into the text, had silently stopped working; both work again
Fixed Remote's phone pairing endpoints crashing the app or using 2 GB from one bad message — a malformed frame could quit the desktop app and an oversized one was buffered in full before its size was checked; both are rejected before anything is read, pairing invitations are single-use, and revoking a phone ends its live session
Agent Input draft with mention chips for a new OpenAI Codex agent, a linked terminal, a project, and a terminal on the paired computer nuc_16_pro 
A 2×2 grid of Claude, Codex, Gemini and Cursor terminals, each with a compact capsule reading Ready, Working 41s or Draft above the agent status line 
Agent Orchestration Activity tab with a Needs you box holding a follow-up and a read-access link request, above a timeline of auto-approved links and runs