Paperclip turns production feedback into routing and policy contracts

paperclipai/paperclip issue #11446 is the inspected primary source: “Production deployment feedback: 10 feature requests (interaction routing, policy gates, API contracts).” The related records below were inspected as supporting context rather than independent confirmation.

The facts

- paperclipai/paperclip issue #11446 was created 2026-08-15T20:01:42Z and was open at inspection; its title is “Production deployment feedback: 10 feature requests (interaction routing, policy gates, API contracts).” - paperclipai/paperclip PR #11445 was created 2026-08-15T17:28:48Z and was open at inspection; its title is “fix(codex): isolate host capabilities per agent.” - paperclipai/paperclip PR #11454 was created 2026-08-15T20:24:14Z and was open at inspection; its title is “fix: ACP run lifecycle corrections — failure settlement, workspace sync-back, lease cleanup.” - The repository reported 78315 stars, 5081 open issues, default branch master, and last push 2026-08-16T00:35:47Z. - The latest tagged-release baseline checked was v2026.722.0 published 2026-07-22T23:05:41Z.

What changed

paperclipai/paperclip issue #11446 reports: We run Paperclip in production as a multi-agent software company (release 2026.722.0, self-hosted). Over two weeks of continuous operation we collected the following feature requests / defect reports. Each states what we observed, what we'd like natively, and our interim workaround. Happy to split any of these into separate issues, provide repro detail, or test previews. 1. Audience-class gate on CEO/founder-facing interaction cards (P0.3f) Observed:** any agent can create an interaction card targeting the human owner; no audience-class validation exists in the interaction routes. Two consecutive days of card floods; 27 classifier denials in one day. Requested:** an interaction targeting the company owner must declare a closed-list class (production merge · gates · credentials · money · screens · novel product) or creation is rejected with a routing hint to the delegated owner. Rejected-attempt counts exposed per agent. Interim:** instruction-file rule + audit counting out-of-class cards (defect count, not gate). 2. Superseded-card UX Observed:** a tap on a superseded card fails with a generic rejection; the replacement card is not pushed. The owner approved a stale card and the approval was lost (observed 2026-08-14/15). Requested:** stale-card responses return "replaced — see new card" with the live card id, and supersession pushes the replacement immediately. 3. Runtime-service reaping on ticket closure

paperclipai/paperclip PR #11445 reports: Thinking Path - Paperclip is the open source app that people use to manage AI agents for work. - The Codex adapter starts agents on a user-owned host. - Codex loads user skills from $HOME/.agents/skills even when CODEX_HOME is separate. - The old shared home also copied host configuration into agent runs. - An agent could therefore receive skills, plugins, instructions, and MCP servers that the operator did not assign. - This pull request gives each agent a managed Codex home and disables unselected host skills. - The benefit is a clear capability boundary with no change to the normal host Codex setup. Linked Issues or Issue Description No public issue exists. A search of open and closed Paperclip issues and pull requests found no matching result. What happened?** A managed Codex agent could load user-wide skills from $HOME/.agents/skills. A shared managed home could also copy host Codex configuration into the run. Expected behavior** A managed Codex agent must receive only its selected Paperclip skills and managed MCP gateways. The host setup must stay unchanged. Steps to reproduce** 1. Add many skills under $HOME/.agents/skills. 2. Configure a codex_local agent with desiredSkills: []. 3. Start the agent. 4. Observe that Codex discovers the host skills without an explicit deny list. Paperclip

paperclipai/paperclip PR #11454 reports: Thinking Path - Paperclip is the open source app people use to manage AI agents for work. - Agent adapters run ACP sessions and manage runtime, workspace, and lease resources. - Several failure paths left runtime bridges, staged workspaces, or environment leases active after an error. - These leaks reduce run reliability and can leave later runs without clean resources. - This pull request closes the failure paths, applies one teardown policy, and adds regression tests. - The benefit is consistent failure settlement and safer reuse of agent workspaces and leases. Linked Issues or Issue Description What happened?** ACP runs could leave runtime bridges, staged workspaces, or environment leases active after failures. Claude and Gemini ACP runs did not restore the sandbox workspace on teardown. Lease release stopped when one lease returned an error. Expected behavior** Each ACP failure must return an error result and settle its resources. Teardown must run each step, release leases independently, and restore the host workspace when the sandbox ends. Pending cleanup leases must receive bounded retry attempts. Steps to reproduce** 1. Run an ACP session that fails after runtime creation or during turn preparation. 2. Run an ACP session that fails during a warm hit or staged runtime handoff. 3. Run lease cleanup with

Why it matters

Deployment feedback is most useful when it changes authority boundaries rather than merely adding buttons. Per-agent host capabilities and explicit run settlement are the substrate for trustworthy policy. The operator test is whether the system remains bounded and its receipts still describe the action after failure, retry or restart.

Current

The primary record was open when captured on 2026-08-16. The tagged-release baseline was v2026.722.0 published 2026-07-22T23:05:41Z. Repository metadata, full source bodies, current pull requests and issues, releases, Google News, Hacker News, Lobsters, Metamesh and the rendered ClawCharts row were inspected. ClawCharts selected the subject; it did not prove the claim.

Evidence

The primary URL is https://github.com/paperclipai/paperclip/issues/11446. Supporting records are https://github.com/paperclipai/paperclip/pull/11445 and https://github.com/paperclipai/paperclip/pull/11454. Source bodies, timestamps and states are preserved in the daily evidence bundle. Test counts and reproductions remain attributed to their authors unless identified as independently rerun.

Source boundary

Open work is described as open, closed work as closed, and operator reports as reports. A pull request is evidence of proposed or reviewed direction, not proof of a shipped release. Search residue, package mirrors and historically published source spines were excluded.

Operator take

Convert each production request into an owned contract test, keep host capability grants agent-scoped, and make failed ACP runs settle leases and workspace state visibly. Preserve a before-state receipt, make the smallest reversible change, and verify the original failure independently.

Caveat

Public project records are mutable. Status, scope and evidence can change after publication. This brief records the inspected state and does not authorize changes to a reader’s deployment.

Source inspected; source state, environment and release boundary remain explicit.