github paperclipai/paperclip v2026.1117.0

41 minutes ago

Paperclip v2026.1117.0

Released: 2026-11-17

Paperclip v2026.1117.0 carries 1,586 commits, promoted from 2026.1110.0-beta.0. The headline is a new default workspace: the Coordinator layout and Agent Chat are now standard, and task chat shows what a Runner agent actually did — its plan, inline diffs, tool calls, subagents, and a plan you can approve. Around that: agents that can set up connections within delegated authority, a full YAML editor for agent configuration, context-window and usage-limit readouts with a one-click Compact, proactive chat messages and automatic mid-turn steering for Codex and Pi, Codex image generation, and 1Password, Supabase, Vercel, and boxd in the catalog. Underneath, Runner runs continue through restarts, crashes, and container redeploys in far more cases, local Codex and OpenCode agents stay warm for much faster follow-ups, sandbox workspaces can survive a server replacement on S3-backed installs, the board keeps working through dropped connections, and run logs and skill packages move out of PostgreSQL into your configured storage.

Breaking Changes

  • The built-in Google Workspace connector is off by default. While Paperclip's Google app verification is pending, the unified Google Workspace connection introduced in v2026.1106.0 is gated behind a new experimental setting: with it off, the connector is hidden from Apps and its setup, discovery, and tool execution are blocked — including connections saved on v2026.1106.0 — and the Composio and Arcade Google apps appear instead. To keep a Google Workspace connection working, turn on Settings → Experimental → Google Workspace connector (Paperclip) (enableGoogleWorkspaceConnector). (#19247)
  • Every subtask gets its own execution workspace. Child tasks no longer share the parent's checkout, branch, or unmerged changes, on every creation path (API, create_task, chat suggestions, plugins, and the New sub-task dialog); inheritExecutionWorkspaceFromIssueId now copies only the project and workspace strategy, and an ancestor's executionWorkspaceId passed on create is dropped. A child an older release bound to its parent's workspace moves to its own workspace on its next run. Work that relied on a child editing the parent's checkout must merge or push through Git; unrelated tasks can still share a workspace through an explicit existing selection. (#19860)
  • Agent-created task trees stop at parent → child. When an agent asks for a grandchild, the server creates the task beside the requester under the requester's parent and returns parentAdjusted: true instead of nesting deeper; board users can still nest to any depth and existing trees stay valid. Scripts that expect parentId to match the request should read parentAdjusted. (#19312)
  • Native agents report task status through paperclip_finish/paperclip_block. While a native run holds its task, the assigned agent's own PATCH /api/issues/{id} status change returns 409 native_run_status_reported_by_finish, and the same applies to the recovery-action resolve route; custom instructions or skills that told Runner agents to PATCH their own status should switch to the finish tools. (#18411, #18489)
  • Local Codex and OpenCode Runner agents stay warm by default. When the Runner lifecycle is unset, local Codex and OpenCode agents keep their runner for a 5-minute idle window after a run; Claude, remote targets, and other harnesses stay per-turn. Migration 0443 clears the stored "Turn by turn" value from existing native Codex, OpenCode, and Claude agents unless their configuration history shows someone explicitly chose it, logging an agent.updated activity entry per agent. To keep a runner from staying resident, set Runner lifecycle back to Turn by turn. (#19888)
  • Skill packages move out of PostgreSQL, and this release becomes the rollback floor. After migration 0415, repository and managed skill package contents convert in the background to your configured file or object storage after a read-back check. Once that runs, an older binary needs the database contents rehydrated; set PAPERCLIP_SKILL_PACKAGE_STORAGE_ENABLED=false on every process during a mixed-version rolling upgrade, and keep that storage durable and backed up. Skill version list responses now return file names and kinds without file bodies. (#18082)
  • Enter submits task replies and new tasks. Shift+Enter inserts a new line; to keep Cmd/Ctrl+Enter submission, turn on the experimental preference Command/Ctrl+Enter to Submit (enableLegacyComposerSubmit). (#19272)
  • pi_local processes no longer inherit the server's PAPERCLIP_* environment. Only the runtime API URL and listen host/port, the run-scoped variables, and the adapter's configured env pass through; a pi_local command that relied on an inherited server variable must set it in the agent's adapter env. (#14503)

Highlights

  • The Coordinator workspace and Agent Chat are standard — Everyone now lands in a workspace built around one conversation per agent: task proposals turn a chat request into an assigned task linked back to its conversation, tasks and decisions show as compact cards, documents open in contextual panels, and an Inbox queue gathers what needs you with search, filters, and history, while task access, assignment authorization, and review gates still apply (#17983, #19360). Navigation groups Chat, Tasks, and Inbox with a New task button, Tasks opens All tasks and remembers your view, and old Inbox links redirect into Tasks (#18703, #18831, #19363, #19393); the Chat directory lists every agent with avatar, role, search, and unread counts, and agents can be pinned alongside pages (#19685, #19627). Each person can return to the previous layout with Classic workspace in Experimental settings.
  • Task chat shows what the agent actually did — Runner turns render the agent's own plan as a docked task list (#19252); file edits carry inline +N −M diffs and turns that changed files end with a "Changed in this turn" card (#19269); tool rows name the verb, subject, exit code, and duration, and say when policy refused a tool (#19265); Claude subagents appear as collapsible cards with their prompt, progress, nested tools, and result (#19264); hook decisions and Codex auto-review verdicts, web searches, model switches, waits, and verification checks get their own rows (#19266, #19262); Codex reasoning summaries show as headline bullets (#19271); and harness banners such as "Compacting…" show as chips instead of leaking into the answer (#19267). Planning runs put Codex and Claude into their native plan mode, and the proposed plan becomes the task's Plan document behind the confirmation gate, so Approve resumes the agent and Request changes sends your reason back (#19291). The final chat reply is now the agent's complete answer (up to 32,000 characters) instead of a 4,000-character preview (#19971).
  • Runner runs continue through restarts and redeploys — After a graceful restart, deploy, crash, or container replacement, a native run continues on its own in far more cases: a run caught during bootstrap or sandbox and workspace preparation restarts its bootstrap on the same run, and a turn interrupted mid-flight continues once in the same conversation on Claude and OpenCode as well as Codex (#19524, #19616, #19694, #20153, #19501, #19699, #20118). Remote Runner agents are re-adopted after a server restart from encrypted controller checkpoints, keeping the sandbox lease, session, and active turn (#17439). Each server boot now holds a fenced database lease, so work stranded by a replaced Docker, Kubernetes, or Fly container is recovered instead of staying stuck, and held records are released as soon as their owner is proven gone rather than after an age timeout (#20054, #19953, #19950, #20004, #20006); accepted results finalize within seconds of a restart after a hard kill (#19713, #19591); queued runs start within about a second of the server listening (#19431); and restarts no longer use up a run's recovery budget (#19948, #19625). For a record no server can judge, the board gets a Confirm the previous server stopped action (#20067, #20084).
  • Much faster follow-ups — With local Codex and OpenCode agents warm by default, measured follow-up time to first token at p50 fell from 12.0 s to 4.7 s for Codex and from 21.7 s to 4.2 s for OpenCode (#19888); follow-ups claim a parked runner while still preparing (#20195), Claude follow-ups reach the provider about 1.5–2 s sooner (#20145), and starts reuse verified executables, provider packs, and startup context (#18390, #17749, #18405).
  • Sandbox workspaces that survive a server replacement — On installs with external PostgreSQL, S3 storage, and a stable secrets master key, new task workspaces for Runner agents on sandbox environments stay in the sandbox and each run's files are saved as encrypted, verified snapshots in object storage; a replacement server reconnects to the recorded sandbox or restores the last saved version into a new one, without the old host or copying files back to the web server's disk. The Files panel labels Live workspace or Saved copy, existing workspaces can be imported explicitly with server/scripts/migrate-remote-workspaces.ts (dry run first, then --apply), and other setups keep their current path. Edits made since the last save still depend on the live sandbox. (#18403, #18826)
  • Agents can set up connections — Within their own grants (tools:manage_connections, tools:manage_profiles, and related) and their responsible person's current authority, agents can configure app and tool connections, chat and email channels, AI accounts and pools, access profiles, and MCP gateways through the same APIs the app uses; steps that need a person — sign-in, consent, credential entry — are saved as handoffs in the usual setup screens and the agent continues when you finish them. Agents still cannot approve their own governed requests or edit their own permission grants; a change a user explicitly asked for may loosen the acting agent's own tool controls without a separate approval. Existing agents' permissions are not backfilled. (#17737, #17739)
  • Edit an agent as one YAML document — A new Advanced tab edits the whole configuration — profile, harness and runtime, instruction files, skill pins, tools, channels, permissions and trust, secrets, and API keys — as a versioned paperclip/v1 Agent document. Saves go through each resource's own API and validation with a change preview, stale writes are rejected, partial saves keep the draft, and credentials stay redacted; omitted settings are left unchanged and explicit empty values clear them. (#18036, #18168)
  • Context and usage limits in view — Task properties show the assignee's context window ("10% of 258k") with a breakdown of tokens, cached input, cost, and compactions for Codex and Claude Runner agents, and the chat marks each compaction (#19273); Compact now runs the harness's own compaction between runs, board-only and unavailable mid-run (#19378); a Usage limit line shows the provider's rate-limit window ("5h 78% · resets 15:40") (#19270); and an agent that hits its provider limit waits as "Waiting for usage limit reset" instead of burning runs and retries, while manual runs and conversation turns still start (#19246).
  • Chat that reaches you, and agents you can steer — Agents can post a result to their standing chat with their responsible person through a new send_chat_message tool, with task-access and privacy checks, a link to the source task, and per-person unread state (#18374). A message queued while a Codex or Pi Runner turn is running is steered into the live turn after a 5-second "Sending in Xs" window with Edit and Cancel; other harnesses keep the queue, and Steer and Interrupt stay available (#19970).
  • A board that rides out outages — One connection banner replaces per-page errors, loaded content stays visible during an outage and refreshes on recovery (#15754, #18822); task edits, documents, decisions, new tasks, and messages keep their draft through a disconnect and are verified or offered for resend instead of being lost or duplicated (#17960, #18412); a small set of reads is cached per user and company for 24 hours so a reload during an outage still shows them, cleared on sign-out (#18825); and a stale tab after a deploy reloads itself when a page chunk fails (#17349, #19361).
  • Codex image generation and skill discovery — Codex Runner agents keep Codex's built-in image tool on their existing sign-in, and agents on a direct OpenAI API-key connection get a Paperclip generate_image tool that calls gpt-image-2 server-side, keeps the key out of the agent's shell, and counts the spend against the agent's budget (#19516, #19654). New search_skills and load_skill tools let an agent find and read company skills it was not assigned, under the run's existing company and project access checks and without changing its assignments (#19676).
  • Storage that leaves PostgreSQL — On S3-backed instances, new run logs and opted-in provider traces are written as immutable segments in the existing bucket with no new settings, raw traces encrypted with the existing vault key, and existing local history is copied in the background with verified digests (#18327). Skill packages move to the configured file or object storage (#18082). Extra project repositories under .paperclip-repositories/ become worktrees of one shared clone per project and repository instead of full per-task clones, with opt-in sweeps that move existing workspaces over (#17382, #17369, #17407).
  • More in the catalog — 1Password as a service-account connection with eight governed vault and item actions (#17249); Supabase gains a Management API connection with eleven reviewed project, Auth, Data API, and Edge Function actions, and Vercel becomes a team-scoped REST connection replacing its unavailable MCP entry (#17217); and a new boxd sandbox provider plugin runs each sandbox as a KVM microVM that suspends between runs of the same task with disk, memory, and processes kept (#17284). Kimi Code joins the Runner as an experimental harness while the Native runner experimental setting is on (#18028).

Fixes

  • Runs, Stop, and Retry — Retry after exhausted or blocked recovery continues in the task's own worktree with uncommitted files, branch, and .git intact (#19666, #19693); a plain task Retry gets the normal bounded automatic retry on a transient setup failure (#19384); Retry on a chat-connected task under lock contention is saved or returns a retryable 409 instead of a 500 (#18895, #18896); holds left by cancelled retries, duplicate no-replay holds, and full-disk restarts no longer block later messages and Retry (#19276, #17055, #18838, #19181); stopping a run while its native session is starting returns 409 native_stop_pending instead of a 500 (#19189), Stop retries a PostgreSQL deadlock instead of failing (#17895), and a run whose runner connection goes silent for 90 seconds is ended and recovered instead of showing "Working" forever (#19407); an Interrupt settles like a Stop followed by your message instead of a failed-attempt banner (#19633, #19149); a sandbox provider's temporary 5xx, 429, or timeout at lease time retries automatically (#19412); releasing a subtree pause restarts the work it interrupted, once (#19394, #19421); a blocked task whose dependency finished during its assignee's active turn continues when that turn ends (#19457); successful routine runs no longer trigger a handoff wake that ends blocked (#9129); the Runner stops every process a provider or tool started when a run ends, including nohup/setsid jobs (#19727); a Retry that waits on an unfinished blocked subtask shows as waiting and links that subtask (#20109); a parent task is always woken when its children finish, even if the wake was missed while its run was ending (#20180); the Resume dialog no longer promises to wake blocked tasks or former assignees, and releasing a pause rechecks the task's current state (#17521); a native run that fails at session.open, before any provider work, is retried automatically instead of blocking the task (#18854); queued turns resume when a maintenance drain ends, and the task view shows the wait (#17967, #18065); Retry resumes saved chat input still queued after a successful run (#18621); closing a task from the board stops its active native run cleanly (#18616); and Check now works on monitors kept on blocked tasks, and the UI says automatic checks are paused (#18061).
  • Messages and comments — A comment is no longer lost when a restart lands between saving it and writing its wake, when its run is cancelled as stale, or when it was added to a queued run (#19992, #20066, #18913); a comment sent while the agent or company is paused is saved and delivered once when the pause lifts (#19956); a steered message is no longer lost to a restart before the model reads it (#19962); messages sent while a run prepares go into its first turn instead of interrupting it, and messages sent while it finalizes are delivered once afterward (#20168, #19183); Send on a saved follow-up after the earlier run finished the task reopens and delivers it once (#19600); an answered card's receipt leaves the composer queue (#19376); a run whose task was reassigned away can no longer close the new owner's task (#19186); withdrawing a card or cancelling a question no longer makes the agent's next native run fail before it starts (#19163); an accepted final reply is published reliably, so a task cannot show done with no visible answer (#19200); a redirect message sent while a stopped run is still settling is saved instead of dropped (#19122); a person's explicit approval in the conversation of an agent's human-only plan or proposal is recorded again (#19243); and queued follow-up messages kept after Stop get a Send control and are delivered (#17959).
  • Costs and budgets — One missing usage receipt no longer freezes a company with a hard budget behind "completed runs await accounting": once recovery ends and the provider is verifiably stopped, known usage settles (priced from the rate card, or recorded as unpriced usage under your policy), a run without closure evidence holds new work for at most 10 minutes, and admission refuses only for a reached hard stop, an explicit pause, or an inactive company (#19497, #19673, #19606, #19944); comments, Retries, and new assignments made during a temporary accounting hold run exactly once when it clears (#19741, #19702, #19829); a budget or agent pause that stops a native run no longer leaves the task Blocked with no control (#20143); Raise budget on a hard stop now files one Retry for each task run the budget stopped, without charging an attempt (#20144); usage reported before a Codex run's first turn is no longer billed to that run (#19783); the budget overview no longer fails on a policy that references a deleted agent or project (#18389); and Grok/xAI OAuth token refreshes no longer add a duplicate subscription card to the Costs page on every run (#16512).
  • Permissions and access — PATCH /api/agents/{id} with a status field no longer gets around the board-only Pause and Terminate actions (#19955, #20164); turning off "Can create agents" on a CEO agent now stops hire_agent and POST /agent-hires (#20041); the default unscoped tools:use grant no longer turns on actions a connection's tool profile left off (#20160), and saving an installation list no longer re-attaches the broad profile to agents you had narrowed (#17906); plugin logs, job runs, and dashboard data return only the caller's companies, with instance admins keeping the instance-wide view (#17435); the credentialValues envelope is redacted from HTTP logs (#13742); secret redaction runs in linear time, so a large base64 payload no longer freezes the server (#16999); agents read only their own commentary unless granted the new commentary:view permission, which is off by default (#19842); a goal can no longer be given a parent goal from another company, and a plugin patch can no longer move a goal into another company (#19203); when a project access denial stops a low-trust agent's task, the task says so, and authorized operators can add the project from agent settings and retry (#17509, #17552); AWS Secrets Manager access denials and missing credentials show as access problems with guidance instead of a generic provider error (#10946); and OpenCode on OpenRouter gets a temporary local credential instead of the reusable upstream key in its environment (#17949).
  • Workspaces and sandboxes — Paperclip records a sandbox's owner before asking the provider to create it, so a failed write or server replacement no longer leaves an untracked sandbox running (#18357), and a restart during acquire or teardown reclaims it at the next startup sweep (#19696, #19749); orphaned sandboxes are stopped and kept for their files, then destroyed once unreferenced (#19964, #19976); a consented environment delete also destroys sandboxes kept for failed runs' file recovery (#20085, #20099); Daytona auto-stop always fires at least 5 minutes after the warm idle window, a sandbox in Daytona's error state is recovered, and exports stream around a full sandbox disk (#19199, #19178, #18172); a leftover .paperclip-upload-* file or <name>.clone-* staging directory no longer breaks every later copy-back or run (#19730, #19374); stopping a run stuck retrying copy-back saves its files and frees the workspace (#19385); two tasks starting at once on one repository no longer collide on .git/config (#20039); plain non-Git directories can be closed (#19553); child tasks sharing a busy workspace wait as one parked retry in arrival order (#19143); a fresh checkout with no author settings commits as Paperclip <noreply@paperclip.ing> instead of failing (#18094); after an SSH environment run, git status no longer shows the run's new files as staged deletions (#16644); the per-company workspace seed cache expires seeds unused for 7 days and evicts the least recently used, instead of filling to its 8 GiB cap (#17357); Git cache expiration reclaims more closed worktrees, restore points save ignored files and stale remote-tracking commits from nested checkouts, and a closed parent workspace can expire while its children stay open (#17290, #17289, #17464); a Daytona sandbox that auto-stopped restarts cleanly, and a resume refused during auto-stop is retried within the bootstrap budget (#19050, #19195); a task repository requested at a ref that does not exist can be corrected instead of failing setup on every attempt (#19072); and a sandbox run no longer fails when another restore writes into the same workspace at the same time (#18426).
  • Adapters — Sandbox Claude runs in approve-all mode no longer stall 30 minutes on a permission request that never arrives (#19430); "Claude - Subscription" sign-in works again with Claude CLI 2.1.291 (#15550); native Claude agents see request_human_input again and re-apply the chosen effort after a cold resume (#18046, #19708); legacy Codex shell commands can use the run's managed git and gh again (#17605) and a saved search-off setting is kept on Runner launches (#17744); OpenCode sessions over 4 MiB resume again, a reply cut off at the output limit is never settled as the answer, and tool results over 64 KiB are delivered whole up to 1 MiB (#18218, #19242, #17825); failed runs show the provider's sanitized error, provider, phase, and HTTP status, and a model the provider rejects shows as "Model unavailable" and fails at once instead of retrying (#18622, #18087, #19548); native failover backups no longer carry per-launch homes, plugin node_modules, or pnpm stores, and earlier leftovers are swept (#17618); Claude ACP session saves on Windows no longer fail with EPERM (#16925); in sandboxes, a Bash tool that runs out of memory is killed before the Claude or OpenCode agent process, so the agent reports a failed tool instead of a lost connection (#16844, #17398); Upgrade to Runner works on Daytona computers that are not yet qualified and for company owners with agent configuration permission, and saved Claude turn limits are ignored with a warning instead of failing the conversion (#17361, #17448); native OpenCode agents with a large assigned-tool catalog no longer hit provider grammar timeouts (#17740); configured environment bindings and company secrets reach OpenCode shell commands again (#19319); OpenCode agents hired by another OpenCode agent inherit its OPENROUTER_API_KEY reference instead of failing with harness_auth_failed (#19977); shell commands in native Claude runs can use the managed GitHub launcher again (#18850); a rejected Codex API key goes to sign-in recovery instead of being retried (#19146); Cursor runs start again on local source builds (#19255); remote Codex Runner agents receive their selected skills again (#18080); and Claude Runner sessions keep their allowed tools, instructions, and settings after an ACP reconnect or resume (#18135).
  • Chat and task UI — The editor stores snake_case, a=b, and x => y as typed instead of adding escapes that reached agents verbatim (#19350, #19540); pasting over a selection replaces it instead of linking it (#18796); sending a message while a confirmation is open keeps the request to answer later (#19496); a run waiting for a busy workspace, AI connection, or runner update no longer shows "Couldn't finish this attempt" (#19369, #19507); approving the newest confirmation no longer brings an older replaced card back to your inbox (#19520); creating a task already done or cancelled no longer wakes its assignee (#19961); answering the onboarding opening question starts the agent immediately (#19886); a reply in a Slack DM thread continues that thread's task even after it is Done (#19486); iPhone task chat no longer jumps while typing and streaming (#17232); New task opens its form immediately, without the loading delay added in v2026.1106.0 (#17041); confirmation prompts render as Markdown, so review links are clickable (#18851); a cached document row with no key no longer crashes the task page (#18870); a deep link to an orphaned document annotation thread shows that thread and its history (#18950); activity toasts show the task identifier instead of a short UUID (#18802); a new comment from someone else no longer marks the task as read for its assignee (#18670); agents whose setup is unfinished show Finish setup instead of a Run button that fails (#17976); and first-agent onboarding no longer hangs on "Testing connection…", a refresh resumes unfinished setup, a failed or expired Claude/OpenAI sign-in can be retried, Back cancels a subscription sign-in, and finishing with an API key after going back uses that key (#19336, #19505, #19557, #19610).
  • Connections, GitHub, skills, and plugins — A task reassigned to another agent keeps its MCP tools (#17340); fresh personal app setup after an archived connection creates a new connection instead of a 403 (#18902); a later distinct "Ask first" write opens a new review instead of returning an old decline (#18742); GitHub 403 Resource not accessible by integration shows a Manage GitHub access card (#19315, #19655), transient repository-listing errors retry with backoff (#19690), a webhook mid-refresh no longer drops repository access (#18805), and background Git jobs fetch private repositories with the task's managed identity (#17515); a crashed skill-cache build no longer blocks later runs, and slow preparation survives databases with an idle-transaction timeout (#19849, #18996); plugin secret-ref settings fields save (#20181) and a retired plugin worker restarts on a fresh process (#19245); purging a terminated agent with cost history returns 409 agent_cost_history_retained instead of a 500 (#19022); GitHub create_repository and fork_repository default to the only organization where the GitHub App is installed (#20141); with a managed GitHub credential, gh repo create, gh repo new, and gh repo fork are refused and point to the connection's tools, so repository creation keeps tool switches, approvals, and audit (#20142); a suspended GitHub App installation no longer counts as healthy access (#17343); the background refresh of GitHub PR and issue status uses a real credential, and a rate limit is no longer reported as unreachable (#17792); adding a GitHub external contributor by username uses the App installation instead of an anonymous request (#17902); Slack shows the full context of questions, publishes every kind of review card, and an answer given in Slack continues the task in the same thread (#17255); managed Slack setup no longer fails on a saved draft, keeps Open Slack available, offers a real Reconnect on an authorization error, and Done opens settings (#18707, #18722, #18837, #18878); runs no longer stop with ai_connection_busy when several runs share one AI connection (#18778); a plugin gets agent_runtime events only for agents it manages, and an environment plugin creates lifecycle resources only for agents on that environment (#19170, #19173); a project stays on the storage it was created with when a default storage plugin is installed (#19036); opening Edit on a skill file with YAML frontmatter no longer shows "Unsaved changes" (#19241); status cards (experimental) compile their query with the built-in Summarizer (#19554); and an MCP server that fails to start shows a warning line at the top of the run (#19289).
  • Server and CLI — Task file-resource routes accept identifiers such as ABC-230 (#16352); approving a tool review whose target changed returns the gateway's 409 instead of a 500 (#19158); malformed UUID filters on issue lists return 422 (#17389); deleting a company with a chat endpoint no longer fails (#17561); PGMAX_PIPELINE=0 no longer breaks transactions (#17831); leftover paperclip-run-* scratch directories are swept (#19671); paperclipai service install writes KillMode=mixed so embedded PostgreSQL shuts down cleanly (#18598); paperclipai doctor no longer warns about key-file permissions on Windows (#14308); automatic database backups from the built-in JavaScript engine keep binary (bytea) columns intact (#18650); a child process that closes its stdin early no longer crashes the server (#14057); company skill routes accept a skill slug or key as skillId instead of failing with a UUID error (#19808); S3 reads that time out with HTTP 408 are retried instead of returning a 500 (#19039); and an environment test whose plugin cannot describe its Runner reports a failed check instead of a 500 (#19172).

Improvements

  • A new agents:read permission allows redacted configuration, skill, and instruction reads with no write authority, and scoped grants now work on individual configuration routes; every agent can read its own configuration (#18766). Agents hitting a board-only route get code: "board_access_required" with remediation that links the board page holding the control (#20151, #20182, #20207).
  • The agent page is regrouped into Agent and Audit sections with name and title editable in the header and Activity, Runs, Costs, and Budgets opening inside the agent (#18378); Identity files get a save footer (#19535); the Secrets page shows the agent's environment bindings, masked (#18121); and AI connection choices show sign-in method, scope, and account so two accounts named "OpenAI" can be told apart (#17738).
  • Routines show a Responsible user you can transfer or adopt; automatic tasks run as that person, and changes need tasks:assign or an instance admin (#19128). An agent-owned task waiting only on a saved monitor shows a Monitoring status (#18624), shows In Progress only while a task-linked run is queued or running (#19453), and a parent is woken once when a child stalls on a board-only recovery action (#20098).
  • The task feed groups adjacent images into a thumbnail grid (#19626), shows agent-created subtasks as linked cards (#18320), renders .patch/.diff attachments with colors (#19478), labels GitHub links as owner/repo#123 (#19476), and task documents open images in a shareable gallery (#18157). Retry shows its current recovery step (#19260), and tool approval cards show server-redacted arguments (#18734).
  • Faster tasks, chats, and inbox: cold task opens show the instruction and recent conversation first and load older history on demand (#17593, #18092, #18461); standing chats open on their 20 most recent messages and reopen from cache (#18358, #19452); Tasks and Inbox load a small index first (#17756, #18492); and the Apps Permissions page no longer downloads every profile (#18687).
  • More background work waits for committed writes and saved deadlines instead of polling the database — issue monitors and routines, automatic backups, native finalization and recovery, GitHub webhooks and the review bot, chat channels, tool health and reviews, approvals and join requests, plugin jobs, status cards, and agent timers (#17863, #17840, #17836, #17801, #17666, #17653, #17554, #17510, #17518, #18401, #17876, #18387); recovery sweeps and UI polls read only the fields they need, for example about 2.5 MB down to under 2 KB per queued-comment poll (#17650, #17463, #17460).
  • A public GET/HEAD /api/ready reports deployment readiness from memory and returns 503 with Retry-After while starting, recovering, or draining (#18518); the company issue list sets X-Paperclip-Has-More and next-page headers when capped (#19972); comment_on_task gains an optional resume flag (#20028); and the API catalog agents search through call_api gains bodies and guidance for company import/export and skill-source import (#19945, #18757).
  • Agent observability exports native OpenCode turns, tool outcomes, and usage to Braintrust or OTLP (#18098); traces gain spans for Daytona lease acquisition, turn preparation, and skill preparation (#18626, #18213, #17918); and with PAPERCLIP_SENTRY_RUN_FAILURES on, task-level "Couldn't finish this attempt" failures are reported and a Codex capacity failure with a scheduled retry is not (#18541, #18295).
  • Daytona sandboxes gain an optional ttlMinutes wall-clock lifetime and gzip uploads on images without zstd (#17980, #20104); S3 multipart and workspace snapshot uploads run parts in parallel (#18659, #19917); Linux provider packs are smaller (linux-x64 from 1,131 MB to 876 MB) (#17387); and Grok Build agents get Upgrade runner on qualified hosts (#17922).
  • Plugins can show task-sidebar panels driven by recorded runtime activity (#18129, #17796), environment drivers can declare managed skills and require an agent (#17774, #17761), a project writer can authorize a workspace-provider plugin to use their GitHub grant for repository maintenance (#17798), and workspace cleanup recognizes squash-merged GitHub PRs (#18493).
  • The Paperclip Runner gains an experimental Cursor Cloud provider; the legacy Cursor Cloud adapter stays the default (#18068, #18069).
  • Model catalogs and sandbox CLI pins are refreshed — new OpenCode model IDs, Grok CLI 1.0.70, Gemini CLI 0.65.0, a newer Cursor CLI, and GitHub CLI 2.103.3 — and the environment Test names the Codex models retired for ChatGPT sign-in (#17809).
  • The Connectors view marks your default personal AI account for each provider and changes it from a menu (#19845); connector sign-in opens in its own window so you keep your place (#18716); and first-agent onboarding can connect a custom Responses or Messages AI gateway (#17253).
  • Files and images can be dropped anywhere on the task page or the New task dialog (#18849); you can copy your own task messages as Markdown (#19442); the skill slash menu shows skill descriptions (#19502); task lists show assignee avatars (#18607); and agents ask independent questions together in one card (#19634).
  • A task created by a routine links back to it under Relationships, and routines link to a task list filtered by a new Routine tasks option (#16471, #18317).
  • Adding the board to a phone home screen installs it as a standalone app without the address bar (#13461).
  • A new read-only endpoint returns a line diff between an installed tracked-source skill and its pending upstream update (#20155).

Upgrade Guide

  • Migrations 0375–0451 run on upgrade; none drops a table or column. Additive tables: 0394 native_runner_checkpoints; 0404 and 0433 per-user workspace preferences (Coordinator by default, classic_workspace_enabled off); 0405 interaction idempotency keys; 0415 and 0417 skill package storage, deletions, and mutation receipts; 0418 sandbox_creation_operations; 0421 and 0427 execution_workspace_snapshots; 0422 agent_file_roots; 0426 and 0428 project execution admissions (no route writes them yet); 0429 native_harness_backups; 0430 budget evaluation work; 0438 native provider artifacts; 0449 controller_instances; 0451 budget_resume_claims. Additive columns: 0379 project Git authorization, 0380 gateway token responsible user, 0396 inbound voice-call cleanup tracking (with a batched backfill), 0420, 0424, 0425, 0431, 0434 issues.origin_issue_id, 0439, 0441, 0442, 0446–0448. Triggers and functions: 0378 environment and plugin runtime hooks, 0444 GitHub refresh bookkeeping no longer advances agent runtimes. Row updates: 0376 and 0377 clear settled native recovery state and process identities; 0419 backfills Agent Chat read states; 0423 grants agents:read; 0443 clears stored per-turn lifecycles. Indexes: 0375 replaces the restore-point uniqueness index so stale pending restore points no longer block retries (#17344); the index-only migrations in 0381–0414, 0432, 0435, 0436, 0447, and 0450 are built CONCURRENTLY outside a transaction and are safe to rerun if interrupted. 0437 and 0440 add structures self-hosted installs do not use and are harmless. Expect 0377, 0395, 0412, 0414, 0419, and 0432 to take longer on instances with large run, activity, or comment histories; 0416 rebuilds a heartbeat_runs index inside a transaction and blocks writes to that table while it runs, 0445 likewise builds a native_run_finalizations index inside a transaction and blocks writes to that table while it runs, and 0434 briefly blocks issues writes.
  • New environment variables, none required. Recovery: PAPERCLIP_CONTROLLER_INSTANCE_LEASE (on unless exactly off) — a server that loses its instance lease while holding work, for example because the database was unreachable for more than 60 seconds, logs a fatal error and exits with code 1 so a newer server can recover the work, so run Paperclip under a supervisor that restarts it (#20054). Storage: PAPERCLIP_SKILL_PACKAGE_STORAGE_ENABLED (on unless exactly false) (#18082). Cleanup: PAPERCLIP_NATIVE_WORKSPACE_SYNC_RETENTION_ENABLED (on unless false) with _DRY_RUN (off unless true) and _DAYS (7) expires settled runs' native workspace sync folders (#17358); PAPERCLIP_NATIVE_RUNNER_EVIDENCE_RETENTION_DAYS (7) expires Runner continuity archives and quarantined sessions (#17616); PAPERCLIP_TASK_DIRECTORY_RETENTION_ENABLED (default false) with _DRY_RUN (on unless false) and PAPERCLIP_TASK_DIRECTORY_ORPHAN_GRACE_DAYS (7) saves finished plain task directories to restore points and removes them, and needs PAPERCLIP_GIT_RESTORE_POINTS_ENABLED=true (#17472); PAPERCLIP_ISOLATED_SOURCE_REHOME_ENABLED and PAPERCLIP_NESTED_CLONE_MIGRATION_ENABLED (default false, and unlike the other sweeps their _DRY_RUN switches are also off by default) move existing workspaces onto the shared clones, with rehome also requiring PAPERCLIP_GIT_CACHE_EXPIRATION_ENABLED=true (#17369, #17407). Other: BOXD_API_KEY is the host fallback key for the boxd provider (#17284), and PAPERCLIP_DB_RESPONSE_CAPTURE_WINDOW opts into one bounded database response-size diagnostic window (#17504). An explicit DATABASE_PREPARED_STATEMENTS=true now also prepares ORM queries; behind a pooler without named-statement support, use false (#17673).
  • Remote workspace storage is on by default where it applies (#18403): with S3 storage, a postgres:// DATABASE_URL, and a secrets master key, new task workspaces of Runner agents on sandbox environments enroll automatically. Every replacement server must use the same master key; snapshot expiry needs s3:ListBucketVersions and s3:DeleteObjectVersion on the snapshot prefix; and once a workspace is enrolled, do not roll back to an older binary. To opt out before upgrading, store remoteWorkspaceStorageEnabled: false in the experimental instance settings.
  • New and changed experimental settings: enableGoogleWorkspaceConnector (default off, see Breaking Changes) (#19247); enableLegacyComposerSubmit (default off) (#19272); remoteWorkspaceStorageEnabled (default on); enableAgentChat is retired and Agent Chat is always on, ignoring a stored off value (#19360); and Artifacts V1 is on for every instance with its toggle removed (#19164).
  • Default views change: every person starts on the Coordinator workspace, and Classic workspace in Experimental settings brings back the previous layout (#19360). The runner dropdown, including the "Legacy runner" override under Advanced, is removed from agent setup and configuration; new agents get the automatically selected runner, existing agents keep theirs, and the API and CLI still accept runner options (#19517). New Codex agents default to gpt-6.1-sol; agents with a saved model keep it (#19623). First-agent onboarding offers only Claude and Codex; other providers remain available when you add more agents (#18309).
  • Permissions (#18766): migration 0423 grants company-wide agents:read to existing standard agents that are active and have an active membership, even where an operator had removed their configure grant; agents in setup, managed built-ins, low-trust or trust-boundary agents, and agents that already hold a scoped configure or suggest-changes grant are skipped. Each grant is logged as agent.permissions_updated; remove it from any agent that should not read peer configuration, and the removal persists. Separately, a CEO hiring denial is enforced only once it is saved on this release, so turn Can create agents off again on any CEO that should not hire (#20041).
  • API clients: native-run self status PATCH returns 409 native_run_status_reported_by_finish (#18411); agent-created child tasks may return parentAdjusted (#19312); inheritExecutionWorkspaceFromIssueId copies only project and strategy (#19860); workspace detail responses no longer run Git and return delivery state unknown, so read close-readiness for Git safety (#18410); skill version lists omit file bodies (#18082); GET /api/companies/{companyId}/user-directory now returns each member's membershipRole to anyone with company access (#15183); and new routes include GET /api/issues/{id}/context-usage, POST /api/issues/{id}/context/compact, POST /api/agents/me/chat/messages, POST /api/issues/{id}/recovery-actions/confirm-owner-gone, and the public GET /api/ready.
  • systemd users: re-run paperclipai service install to pick up KillMode=mixed; existing unit files are not rewritten (#18598).
  • Telemetry: when telemetry is enabled, server.started now carries only the OS family, and connection.created, connection.updated, and connection.invoked events carrying only the connector key and status are sent; PAPERCLIP_TELEMETRY_DISABLED=1 and DO_NOT_TRACK=1 still turn telemetry off (#18897).
  • Harness pins are unchanged from v2026.1106.0: Codex 0.160.0, Claude Agent SDK 0.3.286, OpenCode 1.18.34, and GitHub Copilot CLI 1.0.88. Kimi Code ships at 2.3.3 (#18028).
  • No other configuration or API changes require action.

Contributors

This release has 1,586 commits from 40 contributors. Thank you to everyone who contributed to this release!

@afurm, @AlexFrontegg, @ARTeMobAgents, @caniko, @clempat, @conreo, @d4kine, @Ethereal49, @fengwuqingchen, @genius-0963, @greegorij, @harjothkhara, @HviorForgeFlow, @jamiebuchanan, @LKL-ZREO, @lorenzozanee, @lyuboslavmedarov, @mdc2122, @MichielMAnalytics, @MindSyncHub, @MrBlackTongue, @naeemakhtar110, @phant0um, @Podwarden, @Rockiez, @stefanriegel, @sudheerDev, @tiangao88, @timafen, @xiehuanyi, @yzxcj797

Don't miss a new paperclip release

NewReleases is sending notifications on new releases.