Release v2.40.0

[atomantic]

Release v2.40.0

Released: 2026-08-08

Highlights

Coding agents

• Cursor's agent CLI ships as a first-class provider in both headless and interactive flavors, disabled by default and selectable anywhere a process provider is. "Refresh Models" pulls your account's real catalog from the binary rather than the shipped 27-model slice.
• Refreshing a provider's model list is now the server's decision instead of a client-side guess at it, so the button appears exactly where it works — and when a refresh fails, the toast finally says why.
• A provider whose display name mentions another vendor is no longer refreshed against that vendor's catalog; the binary it actually runs decides.
• Adding a new coding-agent vendor is one registry row instead of eight coordinated edits across five files, with a test that walks every shipped provider through each launch step so a missed one fails in CI rather than in a dead agent session.

Agent reliability

• A provider that goes down mid-queue no longer takes every waiting task with it — provider-attributable failures now sideline that provider for a cooldown scaled to the cause, so retries land on a fallback instead of repeating the same three-second death.
• Grok and Kimi agents actually start now, and an agent that fails to spawn reports the real reason immediately instead of sitting in "initializing" until a sweeper calls it a crash.
• Pull requests opened under a "merge when CI is green" policy get merged again; the watch no longer depends on an optional scheduled task that ships disabled.

Code review

• Every reviewer row can pin a reasoning-effort tier alongside its model, offering only the tiers that reviewer's CLI actually accepts.
• A configured reviewer whose CLI isn't installed on this machine is flagged before the run instead of after the PR is already open — and the Antigravity reviewer, which was being invoked under a command name that exists nowhere, no longer silently degrades into a self-review.

Brain & Comms

• Pasting a GitHub repo into Brain capture can hand the clone to an agent: opt in to a read-only malware scan, a clean-room "study this for ideas" run that files what it finds, or both. Queued scans are now visible on the link they're scanning.
• Comms gains an X page with visibility and reach diagnostics plus review-gated compose handoffs — no stored credentials, no automatic posting.
• Stacker News triage is newest-first, capped per sync, and every item carries explicit reply / zap / moderation / analysis actions behind the existing human review gate.

Fit and finish

• The claim-issue author gate gains a "Me + collaborators" setting between "Filed by me only" and "Any author".
• The series list is readable on a phone again, and LoRA training dataset image labels no longer wash out over bright photos or collide with each other.
• The Antigravity Usage card parses quota again after agy 1.1.x renamed its limit rows.

Code Review Defaults

• Each code reviewer can now be pinned to a reasoning-effort tier. The reviewer table already let you pick which model a reviewer runs; how hard that model thinks was stuck at whatever its CLI defaults to, so the only way to buy a more careful review (or a cheaper, faster one) was to change the model. Every reviewer row now has an Effort column alongside Model, on the AI Providers → Code Review Defaults panel, the task form, the CoS schedule config, and the /do:next run drawer. Each row offers only the tiers that reviewer actually accepts — Codex's minimal…ultra, Claude's low…max, Antigravity's low/medium/high, and low/medium/high for the local LM Studio / Ollama reviewers — so the picker can't offer a level the CLI rejects, and Copilot, Grok, and @username reviewers show no control at all because they have no such knob. Because slashdo's --review-with grammar has no effort suffix, the pinned tier travels on the invocation instead: as codex -c model_reasoning_effort=high / claude --effort high / agy --effort low when an agent drives the reviewer CLI, and as an effort field on the local review request when PortOS runs it. Leaving a row unset keeps that reviewer's own default, exactly as before.
• [issue-3606] A Review Loop reviewer whose CLI isn't installed is now flagged up front. Previously the only way to find out a configured reviewer (Claude, Antigravity, Codex, or Grok) couldn't actually run was to wait for the review loop itself to discover its command was missing — after PortOS had already opened the PR and spent an agent run. The Code Review Defaults panel, task form, schedule config, and /do:next run drawer now show a "not installed" badge next to any reviewer whose CLI binary isn't found on this machine. It's informational only — the reviewer stays selectable, since a federated peer machine (or a later install) may still be able to run it.

Cursor Agent provider

• [issue-3615] "Refresh Models" now works for Cursor, and pulls your account's real catalog. The Cursor providers shipped with a hand-curated 27-model slice, and the Refresh button was hidden for them because nothing on the server could answer — so the only way to reach a model Cursor offers you but PortOS didn't list was to type the id into the provider's model list yourself. Refresh now asks the binary directly (cursor-agent models) and stores everything it reports: 177 ids on a current account, including the -fast priority-compute twin of most tiers, and Cursor's own auto router at the top. Because that catalog varies by account, plan, and binary version, it's the only authoritative answer — which is also why a refresh that can't reach cursor-agent (not installed, or not on the server's PATH) reports the failure and leaves your existing list untouched, rather than quietly re-saving the shipped 27 and telling you it refreshed. The button appears for both the CLI and TUI providers, since --model applies to interactive sessions too.
• A failed "Refresh Models" now tells you what actually went wrong. Every refresh failure — any provider, not just Cursor — used to surface as "Provider not found or not an API type", which is simply false when the provider exists and its type is fine; the real reason (the binary isn't on the server's PATH, the endpoint timed out, that CLI has no catalog to fetch) only ever reached the server log. The toast now carries the actual message. A model list that comes back legitimately empty still saves as empty — an empty answer and a failed one stay distinguishable, which is the whole reason a failed probe refuses to fall back to the shipped list.
• A provider whose name mentions another vendor is no longer refreshed against the wrong catalog. Refresh picked its fetcher by checking the display name for vendor words before checking which binary the provider actually runs. So naming a Cursor provider "Cursor Claude Opus", or an Antigravity provider "Antigravity Claude Sonnet 4.6" — the natural name for a second provider pinned to that model, and Antigravity's own catalog does list Claude models — silently replaced its model list with Anthropic's, ids those tools reject. In the Antigravity case it also dropped the "use agy's own configured default" entry while the provider was still set to it, which blanked the model dropdown and read as though nothing was selected. The binary a provider runs now decides, and only falls back to the name when no binary matches.

Series Pipeline

• The series list is readable on a phone again. Each row put the series' logline and the row's controls — the sync badge, share, sync-to-peer, and delete — side by side in one line. Those controls won't compress below their own width, so on a narrow screen they took the row and left the logline a column about one word wide, turning a two-line premise into a thirty-line ribbon that pushed the next series off the screen entirely. The controls now drop to their own line beneath the text on small screens, so the logline gets the card's full width; on a tablet or desktop the row is unchanged.

Added

• Cursor's agent CLI ships as a provider, in both CLI and TUI flavors. cursor-agent is a harness-capable coding agent, but PortOS had no entry for it — using it meant hand-rolling a custom provider and guessing at its flags. It now ships as cursor-cli (headless one-shot) and cursor-tui (interactive PTY), mirroring the Codex/Antigravity/Grok/Kimi pairs. Both ship disabled, so nothing changes until you enable them, and once enabled they're selectable anywhere a process provider is — CoS agents, runners, pipeline stages, TUI sessions. The headless path runs cursor-agent --print with the prompt on stdin; the TUI path runs the interactive agent. Both inject --force, which is doing more work than the name suggests: besides auto-approving tool calls, it clears Cursor's workspace-trust gate, and without it cursor-agent prints "Workspace Trust Required" and exits — which in a fresh CoS worktree is every single run. The model list is a 27-entry slice of the account catalog (Composer, Claude, GPT, Gemini families), each id verified against the live binary; the default is Cursor's own auto router, so unlike Grok/Kimi/Antigravity no configured-default sentinel was needed. Cursor bakes reasoning effort into the model id (…-thinking-high, …-xhigh, …-max) rather than exposing a flag, so no Effort control is offered for it. Existing installs pick the pair up on restart via migration 231-cursor-providers. If your account offers a model outside the shipped 27 (the full catalog runs to 177 ids, most of them -fast twins), "Refresh Models" pulls the whole list from the binary — see above.

• X account diagnostics and review-gated compose handoffs. Comms now has an X page that stores bounded public snapshots, checks profile/search visibility and post reach through the managed browser, captures visible metrics, and routes drafts through Review Hub before opening X for a final manual submission. It does not store X credentials, post automatically, or claim that ordinary low reach proves a shadowban.

• Pasting a GitHub repo into Brain capture can now hand it to an agent. A bare repo URL captured in the Brain has always been saved as a link and cloned in the background, but nothing said so and nothing followed. Both capture boxes (the Quick Capture dashboard widget and the Brain → Inbox form) now recognize a repo URL, name the owner/repo that will be cloned, and offer two opt-in checkboxes: Scan for malware runs the same read-only static audit as the Links tab's Scan button, ending in a CLEAN / CAUTION / DANGEROUS report you open from the link; Study for PortOS ideas queues an agent that reads the clone for implementation ideas worth adopting and files the good ones as issues (or PLAN.md items, whichever tracker this install is configured for). Both preferences stick, so "always scan what I clone" is a one-time tick. The agents only start once the clone actually lands — a failed clone queues nothing — and the study run is clean-room by contract: it never executes anything from the clone, never edits it, reads its LICENSE before proposing anything, and describes techniques in its own words rather than copying code. Re-pasting a repo you already saved re-queues nothing, since there's no new clone to read.

• Stacker News triage is newest-first and action-oriented. Each community sync is capped to a configurable newest-item count (30 by default), preserves source timestamps for stable ordering, captures direct image posts for Ollama vision analysis, and gives every newest item explicit reply, zap-handoff, moderation-handoff, and analysis actions behind the existing human review gate.

• The claim-issue author gate has a "Me + collaborators" setting. Until now the only way to let an agent work on a teammate's issue was "Any author", which also opens the queue to drive-by issues from strangers — so a repo with any real collaborators had to choose between ignoring their issues and claiming everyone's. The new option trusts you plus every account the forge already grants access to (GitHub collaborators, GitLab project members including group-inherited ones), which is the same set that could push the code anyway. It sits between "Filed by me only" (still the default) and "Any author" in the CoS → Schedule author-filter dropdown, in the per-app override, and in the /do:next run drawer, and applies to claim-issue and claim-work alike. Neither CLI's --author flag accepts more than one account, so the gate is applied to the issue listing rather than pushed into the query — both the work detector and the agent prompt resolve the trusted set first, then filter. If that lookup fails (a token without push access can't read the collaborator list), PortOS reports it and retries rather than quietly narrowing back to your own issues or widening to everybody — and because that failure repeats identically on every retry, both the run drawer and the "Run now" toast name the filter and the way out instead of the generic "try again shortly".

Changed

• [issue-3618] A newly added coding-agent provider can no longer silently miss one of the steps that launch its interactive session. Adding a vendor means teaching several independent launch steps about it, and forgetting one is invisible from the outside: the provider appears in the picker and spawns, then either runs the wrong binary or comes up without the flags that let it work unattended and stalls on its first approval prompt with nobody there to answer. A test now walks every provider in the shipped catalog through those steps, so a vendor any of them doesn't recognize fails in CI instead of in a dead agent session. No shipped provider's behavior changed.

Documentation

• Updated documentation index and feature references. Added missing feature deep dive links for Quota-burn automation (QUOTA-BURN.md), Three.js procedural 3D models (THREEJS_MODELS.md), Stacker News stewardship (stacker-news.md), and the PortDeck native companion app API contract (COMPANION_APP_API.md) across README.md and docs/README.md. Improved inline JSDoc comments for client and server utility modules.

Fixed

• A provider that goes down mid-queue no longer takes every waiting task with it. When an agent run failed because the provider failed, PortOS sidelined that provider only if the failure was a usage or rate limit. Every other provider-side failure — an expired credential, a refused account, an Antigravity install still waiting on Google to verify eligibility — left it marked healthy, so the next task off the queue picked the same provider and died the same way, and the next, and the next. The most visible version: agy reporting "We're finishing verifying your account eligibility… please try again shortly" would burn down a whole queue of tasks in about a minute, each in three seconds, with nothing to show for it. A failure the run can attribute to the provider itself now sidelines it for a cooldown scaled to the cause (minutes for a transient block, longer for something needing a human), which is what makes the retries land on a fallback provider instead of repeating the failure; the sideline expires on its own, so the provider comes back without anyone visiting settings. Failures that describe the request rather than the provider — a bad model id, a content refusal, an off-shape response — still never sideline anything, since the provider's other models are fine. And a failure diagnosed only by keyword-matching the agent's own transcript can't sideline a provider either; that takes a structured signal from the provider itself. One shared cooldown policy now answers this for both agent runs and one-shot prompts, which previously disagreed about how long the same failure was worth.
• The Antigravity account-verification banner no longer blocks the task it interrupts. PortOS recognized the banner and failed the run promptly rather than waiting out the idle reaper, but filed it as something a human had to fix — which blocked the task outright, over a condition whose own error message says it clears itself. It now takes the ordinary retry path, which (with the provider sidelined per the entry above) lands on a fallback rather than re-reading the same banner three seconds later.