Every Monday, a recap of the previous week’s public releases.
This changelog is generated by AI.
Productivity shows merged pull requests over time, including how many came through Tough Crowd. Filter by repository and switch between daily, weekly, and monthly views to see how coding-agent work contributes to your team’s output.
Bottom tabs now put Sessions, Issues, Productivity, and Me within reach, with quick actions for starting a session or creating an Issue. Detail screens have consistent back navigation, and switching tabs preserves your place, filters, and scroll position.
The installed app also supports pull-to-refresh and opens external links separately. Scrolling fixes keep session activity and detail screens usable on narrow displays.
If the attached pull request merges before the latest session changes reach it, publishing can now open a follow-up pull request for the remaining work. This also handles a merge that happens while publication is being reconciled, so new changes do not get stuck waiting on the already-merged pull request.
Submitting a new session no longer briefly flashes an empty composer while navigation finishes. New Issue text stays visible during creation, and repository names display correctly in the Issue composer.
Pull-request synchronization now respects GitHub rate-limit cooldowns, and transient repository-read failures get more accurate retry handling. Branch lookups also handle larger responses.
Sandbox cleanup no longer lets stale observations overwrite a destroyed state, and expected pauses or shutdowns produce fewer misleading evidence warnings.
Review Insights brings pull requests reviewed, total reviews, and findings into one workspace. Select repositories and a time range to see review activity over time, with findings broken down by severity.
Completed Reviews now have a Findings tab that shows the review verdict, the revisions reviewed, and when the review finished. Findings are grouped into blockers, warnings, and notices, with file locations and recommendations alongside each explanation. Longer findings can be expanded when more detail is needed.
The new Activity view is now available to everyone. Your prompts, the agent’s responses, and grouped work appear in a conversation, with clearer completion summaries and expandable verification details.
Follow-up prompts also move more smoothly from the queue into the activity history, reducing layout jumps as the agent starts the next task.
Coding sessions, Discussions, Reviews, and their sandboxes now use explicit lifecycle transitions to keep work, recovery, and cleanup in sync. Coding sessions can be canceled while queued, and Discussions can be canceled during initialization.
Recovery also handles external branch updates more reliably, and session diffs keep the correct scope after the target branch is merged into the working branch.
The mobile new-session composer now exposes Agent Profile, model, and reasoning-effort controls. Menus and session actions make better use of narrow screens, and long branch names wrap in the merge dialog.
Reasoning-effort controls are now available to everyone for supported models. Select an effort level alongside the model when configuring agent work, with the available choices matched to that model’s capabilities.
Canceling a Review no longer leaves later pull-request revisions attached to a session that cannot run. A new revision can start a fresh Review while preserving the canceled attempt as history.
Reviews also handle draft-state changes and moving commit targets more reliably, reducing stale work and GitHub checks left waiting without an active reviewer.
The app now retries interrupted reads within a short, bounded recovery window and coordinates connection feedback across affected requests. Brief network interruptions are less likely to replace the workspace with an error, while failures that persist still surface for recovery.
The installed app now checks for new versions on a daily schedule and applies eligible updates silently. This removes repeated update prompts from the session workflow.
Session activity now reports the total activity count, and duplicate agent-run events no longer leave stale thinking indicators behind. Provisioning retries are identified more accurately, while fixes to repair exhaustion and Review publication help interrupted work reach a settled state.
The setup checklist can now inspect connected repositories and prepare pull requests that add Tough Crowd’s automatic Review configuration. Each repository shows whether Reviews are ready, need configuration, or already have a setup pull request to review and merge.
People who cannot install the GitHub App themselves can send an installation link to an organization owner, then return to Tough Crowd and connect the approved installation without restarting setup.
Tough Crowd can now be added to the home screen on iPhone and Android as an online-first app. Mobile installation guidance appears from the Sessions screen and the installed app opens directly into the session workflow.
Connection status is now visible across the app. Draft follow-up prompts remain available when a connection drops, sending is held while offline, and an expired login returns the reviewer to the session and tab they were using.
A failed Review can now resume in its retained workspace when the agent still has work to finish, or republish a saved Review result without rerunning the agent when only publication failed. Reviewers can also start over against the pull request’s current revisions or end the Review.
Failed attempts now settle their GitHub checks with a truthful result, and Review publication retains enough durable state to recover from interruptions and terminal races.
Menus, tabs, fields, selection controls, loading placeholders, empty states, avatars, navigation, and common layout elements now share the same reusable interaction patterns across the product. This improves keyboard behavior, accessible labeling, and visual consistency while removing divergent copies of the same controls.
Browser page titles also identify the current workspace more clearly, including Sessions, Reviews, Issues, settings, and administration screens.
Sign-in, access approval, and password recovery now apply tighter authorization, redirect validation, and shared production rate limits. Sensitive recovery details are kept out of telemetry, while signed access decisions make approval links safer to use across the full request flow.
Queued follow-up work now continues correctly after the current agent is stopped, and attached pull requests recover more reliably when drafts are converted or branches move. Review admission and publication also preserve prepared findings before posting them, reducing the chance that an interruption leaves GitHub and Tough Crowd in conflicting states.
Configured repositories can now run Review sessions when pull requests open or change. Reviews publish their status and findings back to GitHub, track progress across repeated reviews, and keep the pull request linked from Tough Crowd.
Organizations can choose whose model credentials power reviews. Failed reviews can be retried or ended from the Review queue, and review state now recovers more cleanly when a pull request becomes a draft, changes its target branch, or a publication is interrupted.
Issues are now available to everyone as a durable record for bugs, features, and tasks. Issues can be prioritized, discussed, linked to coding or Review sessions, and followed through resolution without depending on an external issue provider.
The Issues list supports repository and status filters, bulk categorization, and a mobile-friendly detail workflow.
When starting a coding session, reviewers can now choose an existing branch instead of always creating a new one. Commits can go directly to that branch, and an associated pull request can remain the review and merge target.
Tough Crowd prevents two coding sessions from publishing to the same branch at once and preserves the branch attachment through recovery.
Session lists now use mobile cards with swipe actions and bottom-sheet filters. Session detail has focused drill-down views for activity, files, and other evidence, while the new-session composer better handles the on-screen keyboard, drafts, and compact submit controls.
Failed work now remains in the Review queue, where its recovery action is available instead of being separated from the rest of review-ready work.
Autofix status on pull requests now focuses on the current repair campaign, with earlier attempts summarized separately. Queued, running, and completed repair states are easier to distinguish, and the default repair allowance has increased to seven attempts.
Organization settings now present members and invitations in one clearer People workspace. Administrators can update roles, resend invitations, and manage review credential ownership with more reliable feedback and better compact-screen layouts.
The CLI can now add and display Issue comments, set Issue types and priorities, and save machine-local defaults for Agent Profile, model, and reasoning effort.
Tough Crowd now links to the public 1/500/0 Challenge from its home page. The scorecard, weekly changelog, and lessons are available to search engines, and lesson pages include generated social-preview images.
Session and Issue inputs now lock while submissions are in progress, reducing accidental duplicates. Terminal-session file records are retained before sandbox cleanup, while sandbox revival and Review publication paths have stronger retry and concurrency handling.
Preview request queues and terminal streams now have explicit bounds, and session continuation requests receive additional cross-site request protection.
Autofix now shows its repair allowance in the session action panel. When that allowance is exhausted, reviewers can choose to keep autofixing and give the session another bounded attempt.
Repair budgets now carry across replacement runs, so restarting a repair does not silently reset its limits.
Publication recovery now handles branches that move while work is finishing. Sandboxes are preserved while a Git push is still in progress, and paused work is less likely to be mistaken for a missing or failed sandbox.
Fixes to queue stalls, database contention, and progress detection also reduce cases where active work is incorrectly stopped or left waiting.
Older session activity loads as you scroll, with fixes to pagination so the history remains navigable.
New coding sessions can now include file attachments and pasted images alongside the prompt. Reviewers can give the agent supporting material at creation time and inspect those attachments from the session.
Discussion sessions are now available to everyone. Ask questions and explore a repository with an agent without opening a coding task or publishing a branch. Follow-up prompts keep the discussion in the same session.
Follow-up changes now wait for an explicit publish action before updating the remote branch. Reviewers can inspect the work in the session before making the next revision available on GitHub.
Tough Crowd can now record production deployment results and show a Deployed status for the associated work. This distinguishes a merged change from one that has actually reached production.
A forgot-password flow now helps users regain access to their account, with clearer authentication feedback along the way.
Sessions can recover when the sandbox provider pauses an active workspace. Paused previews are no longer incorrectly treated as deleted, and recent reviewer interaction resets the pause timer.
Publication retries and finalization were also tightened so interrupted work can settle without duplicate actions.
The setup section is now collapsed by default, keeping the agent’s work easier to find. Elapsed times include hours for longer sessions, and post-creation loading uses a consistent indicator.
Organization settings are now broadly available. Administrators can manage the organization’s name, members, and invitations from a shared settings screen.
The redesigned Sessions list is now the standard experience. Work is grouped into In progress, Review, Failed, and Ended views, with focused status and check filters for finding the sessions that need attention.
Session dates are shown in the reviewer’s local time, and each view presents the timestamp that matters for that stage of work.
Session detail now brings readiness evidence, pull-request context, and the appropriate next action into one consistent panel. Reviewers can see what is blocking a merge and act on recoverable, failed, completed, or archived work without translating internal lifecycle states.
Tough Crowd now verifies the target branch immediately before merging. If the branch has advanced since the change was reviewed, the merge stops instead of applying stale review evidence to a different base.
The Tough Crowd CLI can now cancel active work or abandon a session through the public API. These commands use the same durable lifecycle transitions as the browser, so cleanup and final state remain consistent.
Session state transitions now use a single auditable path, reducing contradictory status, sandbox, and usage records. Publication recovery and sandbox cleanup were also strengthened so interrupted workers are less likely to strand completed work or leave resources running.
Text contrast, focus behavior, navigation labels, and button states were tightened across the application. Long agent activity now stays within its column and can be copied without breaking the session layout.
Consecutive agent activity is grouped into readable units instead of appearing as a long stream of individual events. Expanded and collapsed groups keep their state while new work arrives, so a reviewer’s place is not lost during a live session.
Session lists now show created and updated timestamps, while elapsed time is displayed in hours, minutes, and seconds. This makes it easier to distinguish newly active work from sessions that have been waiting for review.
Branch publication now uses one authenticated path with stronger ownership and branch-head checks. Sessions that finish without changing files settle cleanly, and sessions whose pull-request branch advances externally are reconciled instead of being left in an ambiguous state.
Idle and review-ready sandboxes now use recent user interaction when deciding when to pause. Untouched ready sessions can pause immediately, while a reviewer who is actively inspecting the work does not lose the live workspace underneath them.
Verification and follow-up deadlocks were repaired, transient sandbox transport failures now receive bounded retries, and oversized live-diff artifacts can no longer overwhelm session monitoring. Artifact failures are surfaced explicitly instead of silently producing incomplete evidence.
Sessions that need input now pause their sandbox instead of consuming resources indefinitely. Reviewers can resume recoverable sessions, and a new “Verify and finish” path completes verification when the coding work is already done.
Several conditions that could strand sessions during GitHub authentication, sandbox recovery, artifact collection, or worker reconciliation were also fixed.
Pull-request changes are now calculated from the true merge base rather than the repository’s current default-branch head. This prevents unrelated upstream changes from appearing in the review diff.
Sandbox file and artifact failures are also surfaced explicitly instead of silently producing incomplete review evidence.
The Sessions screen now loads history incrementally. Reviewers can load additional sessions without fetching the entire history upfront.
New model options have been added to Agent Profiles. TypeScript and Python repositories also receive more reliable default toolchain setup, with setup activity visible in the session timeline.
Database contention was reduced, several sandbox lifecycle races were repaired, and deployments were changed to avoid brief API 5xx gaps while ECS tasks and load-balancer targets rotate.
A public system status page is now available at statuspage.incident.io/tough-crowd for production incidents and recovery updates.
The redesigned New Session composer is now the standard experience. Repository, branch, coding-agent profile, model, and reasoning effort can be selected in one place, and those choices are preserved through execution and retries.
Reviewers can send additional instructions while an agent is working. Prompts are held in a durable queue and can be edited, withdrawn, paused, and resumed without losing the session workspace.
Stopping an active task no longer discards queued follow-up work.
The pull-request Conversation tab is now available to everyone, bringing GitHub discussion into the session review workflow.
The merge experience now presents blocking and advisory checks separately, explains why a change is or is not ready, and supports the available merge strategies without hiding unresolved evidence.
Recoverable verification failures can be repaired and resumed instead of forcing the reviewer to abandon the session.
Images in repository and diff views can now be previewed without downloading them or leaving the review workflow.
The Files Changed view was also expanded to make larger diffs easier to inspect.
Session status and activity now update through the live event stream. Counts, verification progress, and lifecycle transitions converge without requiring repeated page refreshes.
Review-ready and inactive sessions automatically pause their sandboxes while preserving the workspace for later review or follow-up work. Cancellation and cleanup races that could leave orphaned sandboxes were also corrected.
The Tough Crowd website moved to a faster static architecture and is now deployed independently from the authenticated application.
The Tough Crowd CLI can authorize through the browser instead of asking users to manually transfer credentials. The authorization flow creates a revocable API key and returns control to the CLI after approval.
API keys can now be named, copied from a one-time reveal, inspected through safe masked previews, and revoked with explicit confirmation. The Settings screen also shows creation and last-used information without retaining recoverable key secrets.
Tough Crowd now checks whether a repository can support a coding session before committing agent time. Problems with configuration, credentials, branches, or the development environment are reported earlier and more clearly.
File changes and diff counts now appear as the agent works. Reviewers no longer need to reload the session to see newly created or modified files.
Verification failures now enter bounded repair workflows with clearer state transitions. Worker retries, session reconciliation, and CI outcome handling were strengthened so transient failures are less likely to strand otherwise valid work.
