Warp on X

People praise the terminal itself — command prediction and shortcuts (typing "c" for a codebase, "n" for npm run dev), input they can edit like a normal text field rather than terminal-style, pinning and colour-grouping chats; one calls it "finally the terminal experience" they wanted. A contributor reports a merged PR adding Muse Code support and another in progress importing model names from custom inference endpoints, and one user credits Warp with pinpointing a damaged charger in about five minutes.

The 7 people praising Warp Factories name the timeline UI, artifacts and benchmarks, and one repeats the claim of "40% more prs at 67% lower cost"; all 7 of those posts arrived in the past week.

The 6 people drifting away are fewest: one retired Warp's sessions and agent orchestrator for orca, one switched to t3code, and others keep it open from habit, mainly to edit environment variables or view files.

The bug reports are everyday friction rather than outages: text manipulation going erratic mid-session in one long-time user's terminal (who says it doesn't happen in other terminals), Cmd+Click on files in the Claude or Codex CLIs no longer working after an update, a broken opencode v2 plugin, and a regression someone flags for a fix.

Plan trouble shows up as errors, not bills: a subscription reported ended mid-task, cutting the agent's work short; a paid OpenAI plan that Warp says has no tokens left while the user says their plan is nearly untouched; and a Business workspace connection that starts and then fails for lack of a token-sharing scope.

Smaller groups: 2 people describe Ctrl+Q closing Warp and stopping running agent sessions with no confirm, one saying it "doesnt stop nor warn me"; 1 person has pending updates kill every session and break the sandbox's network and file system; two people want Windows support; two say recordings and screenshots don't prove a change is right; and 1 person found the self-improvement workflows in Warp's open repo failing every run.

Ghostty is used alongside Warp rather than instead of it by one person — agents in "ghostty + herdr", Warp kept for editing or viewing files — while another asks why anyone leaves Warp at all: "i haven't found anything better/faster than warp". Orca draws people moving their agent work over; one retired Warp's sessions and agent orchestrator and is "going all in to orca", and another says Warp stays open from habit and is now used only to edit environment variables now and then. One person landed on t3code after considering ghostex, which another user recommended to them.

Other tools are named by one person each, and several are agent tools: Conductor's multi-git worktree workflow with both a CLI and a UI, Devin from someone who says they haven't used Warp in a long time, herdr, and open harnesses such as Claude Code and Codex instead of a proprietary single-model harness. One person weighed Warp against Cognition for a coding factory, and PowerShell appears as a default people swapped for Warp rather than a rival. cmux, which one person left for Warp in September, has had no posts since.

The verification questions follow Warp's own posts about condensed recordings and benchmarks. People want the mechanics: how click and key presses became overlays in the videos, and whether the self-improvement loops re-benchmark on a fixed cadence or only when agent quality drops, since a model that wins on PRs merged can lose on cost per PR once retries and review churn land. Two ask what screenshots and recordings actually prove — "what catches the PR that looks correct and breaks the flow next to it?" — and whether agents diff screenshots against a baseline.

Owning a plan is the second group: whether Warp takes ChatGPT subscriptions or API access, why a paid OpenAI plan reads as having no tokens left, whether a Business workspace can connect when it lacks token-sharing scope, and whether someone can "bring my own harnesses".

Smaller questions: which to pick for a coding factory, Warp or Cognition, or why anyone would leave Warp; and how Warp is sold — whether it will be open sourced with a paid hosted version, and whether a one-person business may sign up for the Factories Business plan.

The proof requests come from people reacting to Warp's condensed recordings, which they like: "recordings are so much nicer than reading a diff". What they want on top is a cheap automated check first — a Playwright screenshot test against a baseline in the repo, or the agent writing a cypress test for its own implementation that passes, with screenshots attached as proof of its work — because 15 clips on every PR is a lot to watch.

Three single requests follow. One wants reviewer feedback fed directly back into the agent for another verification pass. One wants the rubric's insufficient_evidence label charted apart from clean passes, so a rise in unreadable diffs doesn't look like fewer redundant tests. One, arguing in a thread about quitting, wants Warp to stop and warn before closing while processes are running.