Paperclip v2026.720.0

Released: 2026-07-20

Highlights

• Skill Studio & skill organization — A three-pane skill IDE with sandboxed test runs, so you can author, edit, and try out skills without leaving Paperclip. Skills now organize into nested folders with a dedicated "My Skills" view, you can import skills straight from a project, and company skill policy is open-by-default with core UX for tightening it. Company skill forks run prechecks before they land, and markdown edits get proper dirty-tracking and save. (#9241, #9235, #9356, #9633, #9620, #9564)
• Attention queue & Decisions — A new attention queue and Decisions surface bring everything that needs your input into one place, with faster scrolling and readable, mobile-friendly decision rows. (#9380, #9468, #9472)
• Better search — Search gains filters, sorting, and operators, with command-palette parity so you can find issues and entities the same way everywhere, plus a new bulk extract endpoint. (#9327, #9507)
• Tougher, self-healing runs — Run restart recovery, workspace self-heal, quota-aware retries, and failed-run metrics mean your instance tries harder before it involves you. Recovery is now routed by failure cause, waits for provider quota resets, throttles serial repeats, and reports observability metrics with a rate alert. Plan-approval continuations and external-wait liveness are durable across failed wakes, and resolved blockers reliably wake the tasks that depend on them. (#9183, #9331, #9373, #9229, #9634, #9635, #9651, #9644)

Improvements

• Sandbox execution — ACP sessions can now run in sandbox execution targets, with sandbox support for the Grok local adapter, custom-image snapshots applied across config tests and saves, and a sandbox wall-clock backstop raised to four hours for long-running work. (#9390, #9338, #9385, #9232)
• Issue monitors are visible everywhere — External-service and issue monitors now surface across task surfaces and in issue properties, so you can see what a task is waiting on at a glance. (#9783)
• Reworked secrets dialog with in-sheet agent access — The secrets dialog was reworked and you can now grant per-agent secret access directly in the sheet. (#9797)
• Agents can safely tidy inboxes — Agents can now archive and tidy user inboxes under a governed inbox-archive access policy. (#9724, #9658, #9654)
• Active PR gardening workflow — A new PR-gardening workflow keeps open pull requests moving. (#9510)
• Stable workspace service controls — The workspace service control bar was stabilized so start/stop/restart controls behave predictably. (#9705)
• Responsible-user attribution on activity logs — Activity logs now stamp the responsible user, and issue attribution is explicit. (#9731)
• Clearer blocked-issue messaging — When a message won't reopen a blocked issue, the UI now explains why (Rule C). (#9417)
• Interruptible queued runs — You can now interrupt queued issue runs instead of waiting them out. (#9725)
• Clearer task-level model overrides — Issue properties now clarify task-level model overrides. (#9710)
• Design-system convergence — Single-source design tokens, a visual regression suite, and a retuned theme, plus broad Card/Badge adoption, a multiplicative radius ladder, and unified list surfaces. Task status glyphs now use Lucide icons. (#9134, #9240, #9395)
• Activity-gated routines — Scheduled runs can be gated on external activity, so routines only fire when there's real work to do. (#9436)
• ACP is the default engine for local adapters — Local coding adapters now run through ACP by default, with local coding processes confined for safety, and the Claude ACP client was updated. (#9238, #9504, #9484)
• Spend & cost telemetry — ACP-lane usage and cost now flow into spend telemetry, unpriced CLI usage is recorded, and credential-health retention is documented. (#9471, #9505, #9248, @nickyleach)
• Durable run logs — Run logs are now mirrored to a durable object-storage-backed store. (#8984, @stubbi)
• Codex & model updates — Updated Codex adapter GPT-5.6 defaults and added gpt-5.4-mini to Codex/OpenCode selection (plus openai/gpt-5.5 to OpenCode). (#9352, #4357, @supertaz)
• Ship ripgrep in the agent runtime image — rg is now available out of the box in the base agent image. (#8976, @stubbi)
• Fewer redundant wakes — Redundant issue re-wakes are throttled and the execution contract is injected once per fresh heartbeat, cutting duplicate token spend. (#9470, #9469)
• Show source SHA for unreleased builds — Builds that aren't on a formal release now surface their source SHA, preserved even without Git metadata. (#9508, #9638)
• Prose editor for agent instructions — Markdown agent instructions now use the prose editor. (#9332)
• Health-gated dev services — Paperclip dev services now require health readiness before they're considered up. (#9269)
• Harder environment deletion — Environment deletion is hardened and now shows its delete blast radius. (#9250, @nickyleach)
• Faster release verification — The release verify workflow is parallelized and npm registry version queries are batched. (#9168, #9202)

Experimental

These features ship gated behind experimental settings — enable them under experimental settings to try them out.

• MCP Tool Gateway & Apps — A first-class, governed way to connect Model Context Protocol tools. A named MCP gateway brokers every tool call, governed access contracts and a tool-access policy decide which agents and profiles may use which tools, and the new Tools, Profiles, and Apps surfaces let you wire up, install, and smoke-test connections from the UI. (#9556, #9557, #9558, #9559, #9560, #9561, #9562, #9563)
• Built-in Summarizer & summary slots — A built-in summarizer agent and summary slots surface concise, always-current summaries right in the UI. (#9713)
• Decision Training — A new decision-training library and inspector let you capture decision snapshots and review how your agents decide over time, backed by a snapshot foundation and image galleries for decisions with visual context. (#9702, #9779, #9532)
• Cases — A new first-class Case object gives you a structured, document-backed record for things like releases and social threads. (#9198)

Fixes

• Cross-tenant existence oracle closed — The API now returns 404 instead of 403 so it no longer leaks whether another tenant's resource exists. (#3967, @stubbi)
• Stronger public invites — Invite-token entropy is widened and public invite endpoints are rate-limited. (#8979, @stubbi)
• apiCompression no longer corrupts auth — Fixed gzip clients getting corrupted or dropped Better Auth responses. (#9381)
• React performance-track memory leak — Reaped React 19.2 performance-track measures that were leaking memory in long-lived tabs. (#9827)
• No more duplicate tasks or recovery loops — Duplicate task creation and recovery loops are prevented, repeated creates are deduplicated, and cancelled productivity-review loops are stopped. (#9648, #9650, #5210, @YanisThePie)
• Reliable hot restart — Hot-restart run adoption is restored and a hot-restart shutdown deadlock is fixed. (#9647, #9670)
• Durable continuations — Accepted-interaction continuation recovery is bounded, stranded recovery continuations are restored, and stale handoff alarms are suppressed during live continuation. (#9656, #9630, #9695)
• Codex auth hardening — Refresh auth failures are classified, a host-unusable Codex auth merge fails closed, Codex models are configured at startup, the output-inactivity timeout is raised to 30 minutes, raw child output counts as activity, and shadowed sandbox auth is flagged. Codex auth-merge scripts moved into the adapter with an outbound copy-back home-asset restore path. (#9598, #9276, #9700, #9699, #9632, #9259, #9785, #9787, #9788, @nickyleach)
• Explain AWS secret failures — AWS secret creation failures are now explained in the UI instead of failing silently. (#9645)
• Routine catch-up coalescing — Sub-hourly catch-up runs are coalesced so a backlog doesn't stampede. (#9649)
• Worktree port conflicts — Recurring worktree port conflicts are prevented, and adapter env is forwarded to local agents while runtime env stays authoritative. (#9642, #9617)
• Search & PR-gardening hardening — Extract match limits are honored and PR-gardening candidate discovery is hardened. (#9652)
• Company skills honor responsible-user grants — Company skill access now honors responsible-user grants. (#9571)
• Decision cards survive machine comments — Machine-authored comments no longer supersede decision cards. (#9015, @nsollazzo)
• PID persistence for adapters — onSpawn is forwarded to the Hermes and process adapters so PIDs persist. (#8722, @machjesusmoto)
• Reliable process teardown — runChildProcess now escalates to SIGKILL on liveness rather than trusting child.killed. (#8598, @justinltodd)
• System comments show as "You" — System-authored comments now display as "You" instead of "Paperclip". (#6330, @BorClaw)
• Correct heartbeat activity attribution — Heartbeat invoke/resume now uses run.id for the activity log. (#3424, @tmartin2113)
• Prefix-aware export/import links — Company export/import links are prefix-aware again behind a path prefix. (#6648, @qbamca)
• Tolerate empty profile names — Empty-string user names in profile/session parsing no longer break. (#8986, @stubbi)
• Worktree repair — Dirty and foreign-branch execution worktrees are repaired, target attestation is required before repair, and worktree execution only starts after activation. (#9297, #9414, #9374)
• Workspace branch reconciliation — Added auto-forward execution-workspace branch reconciliation and its route, and avoided freezing an accepted-plan workspace branch before child realization. (#9172, #9170, #9233, @nickyleach)
• Request-storm and polling fixes — Fixed a request-storm polling loop, issue-list coalescing, bounded the shared polling cache, event-sourced the company live-runs list, and cut live-update churn that inflated tab memory. (#9190, #9406, #9627, #9701, #9624)
• Live-run readability & memory — Live run streaming text is more readable, live agent-run transcript buffers are capped to bound tab memory, and markdown mention observers are no longer recreated on every render. (#9330, #9569, #3809, @LeonSGP43)
• Threads stay put — Issue threads no longer jump to the latest comment while you're reading. (#9354)
• Markdown rendering fixes — Rendered markdown list markers stay visible and numbered lists are no longer cut off on the left. (#9359)
• Inbox & list polish — The inbox unread badge no longer indents the row, unread badge alignment is fixed, and inbox hover stays in sync with j/k keyboard selection across list reshapes. Agent names stay visible on the mobile agents index and inbox/task list nesting, hover perf, and routine detail were tidied. (#9767, #9685, #9680, #9236, #9317, #9383)
• Environment edit is a routed page — Editing an environment is now its own route, with a detailed unsaved-changes banner that guards against draft loss. (#9386, #9391)
• Experiments auto-recovery dialog — Enabling an experiment no longer leaves the UI dimmed and locked. (#9513)
• Custom image setup — Custom-image setup gets proper company context, and custom-image snapshots stay applied to probes and saves. (#9028, #9385)
• Reflection Coach & built-in assets — Startup no longer crashes when Reflection Coach assets are missing, and built-in agent assets are preserved in server builds. (#9351, #9339)
• Skip user-secret resolution for skills routes — Skills routes no longer attempt to resolve user secrets.
• Timeouts self-describe — acpx_local timeouts are now self-describing. (#9232)
• Sandbox network resilience — Stalled Daytona Git network commands now fail fast.

Upgrade Guide

• This release adds a large batch of additive database migrations for the new MCP Tool Gateway, tool access policy, OAuth state, connection installs, Smoke Lab, decision-training snapshots, summary slots, inbox-archive policies, and audit/activity attribution. Migrations run automatically on startup — no manual action is required.
• The MCP Tool Gateway, Cases, Decision Training, the built-in Summarizer, and Sandbox execution are gated behind experimental settings. Enable them under experimental settings to try them out.

Contributors

This release has 195 commits from 16 contributors. Thank you to everyone who contributed to this release!

@bluzername, @BorClaw, @justinltodd, @LeonSGP43, @machjesusmoto, @nickyleach, @nsollazzo, @qbamca, @RobinALG87, @stubbi, @supertaz, @tmartin2113, @YanisThePie