Skip to content

[Bug]: Claude Code via ACP can't see built-in skills #16905

Closed
gokay-ai/OpenHands
#3

Description

@rbren

Operating System

Originally reported on macOS. Independently reproduced on Ubuntu 24.04.4 LTS x86_64.

Installation Method

npm (npx @openhands/agent-canvas or npm install -g @openhands/agent-canvas). The independent reproduction used the supported local dev:static stack from current main.

Versions reproduced

  • OpenHands / Agent Canvas main: f26d734a848297d8dcf460b0bb739174e76511f0 (1.16.0)
  • openhands-agent-server: 1.44.1
  • openhands-automation: 1.9.1
  • Claude Code: 2.1.251, authenticated Claude subscription
  • ACP adapter: npx -y @agentclientprotocol/[email protected]
  • ACP model: opus[1m]

Bug Description

Agent Canvas shows the built-in openhands-automation skill as enabled, but its instructions do not reach a newly created Claude Code ACP session. Claude therefore cannot activate or follow the enabled skill unless it independently reconstructs the workflow from runtime API discovery or the filesystem.

Steps to Reproduce

  1. Start Agent Canvas with fresh isolated state using npm run dev:static, then choose Claude Code through the built-in ACP preset.

  2. Confirm openhands-automation is in Canvas's enabled-skills setting.

  3. Install a prompt-preset automation named Code Review Automation.

  4. Start a new Claude Code ACP conversation and ask:

    Use the built-in openhands-automation skill that Agent Canvas enabled for this session. Without searching the filesystem or application source code, first state whether those skill instructions were provided to you and name their first top-level heading. Then use those provided instructions to change my installed Code Review Automation schedule from Mondays at 09:00 UTC to Tuesdays at 09:00 UTC. Do not infer or reconstruct the skill from API discovery; I need to verify the enabled built-in skill itself reached this Claude Code ACP session.

  5. Inspect the browser's POST /api/conversations, the resolved agent-server conversation, and Claude's response.

Actual Behavior

Conversation 5967e61e-bf15-4732-962a-ce5a82329a2f completed through the real authenticated Claude Code ACP subprocess. Claude responded:

No — the openhands-automation skill instructions were not provided to this session.

I can't name its first top-level heading, because I never received the content. ... There is no OpenHands automation skill among them, and Skill only accepts names from that list, so I have no way to load it.

Claude refused the requested schedule change because API reconstruction and filesystem/source inspection were explicitly excluded. Direct control trials showed Claude can modify the same installed automation by discovering <RUNTIME_SERVICES> and the API, so credentials and automation-service access are not the blocker.

Claude Code ACP reports that the enabled OpenHands automation skill was not provided

Boundary evidence and implementation target

  • Canvas persisted openhands-automation in misc_settings.app_preferences.enabled_skills, with an empty disabled list.
  • The frontend conversation-start request contained agent_profile_id but no inline agent_context or skills.
  • The resolved ACP conversation contained agent_context.skills=[], load_public_skills=false, load_user_skills=false, and load_project_skills=false.
  • The subprocess was the expected Claude preset: npx -y @agentclientprotocol/[email protected].
  • Claude was authenticated and produced a completed response, ruling out missing credentials/environment.

The bounded implementation target is the Agent Canvas → Claude Code ACP session/profile bridge: project the currently enabled Canvas skill instructions into each new Claude ACP session, while respecting the disabled-skills setting. This issue is not requesting a generalized change for every ACP provider.

Expected Behavior

An enabled openhands-automation skill is available to a new Claude Code ACP session, activates for an explicit automation request, and guides the requested change without requiring a filesystem hint or independent API reconstruction.

Acceptance Criteria

  • Enabled openhands-automation instructions reach a new Claude Code ACP session and are followed for an explicit automation request without a filesystem hint.
  • Disabling openhands-automation prevents those instructions from being exposed to a new Claude Code ACP session.
  • An automated regression test crosses the Claude Code ACP boundary rather than only asserting the frontend conversation-start payload.

Activity

  1. github-actions commented on Aug 25, 2026

    @github-actions
    Contributor

    ✅ ready-for-dev label applied

    This issue meets the readiness criteria and is ready for contribution.
    A contributor can now open a PR that links this issue (Fixes #<number>).

    This is an automated check — no AI was used to generate this comment.

  2. added this to the milestone on Aug 26, 2026
  3. neubig commented on Aug 29, 2026

    @neubig
    Member

    Current-main isolated reproduction is now complete. I updated the issue body with the exact environment, completed Claude ACP failure, boundary evidence, screenshot, bounded implementation target, and behavior-focused acceptance criteria; the repository readiness checker passes and ready-for-dev is applied.

    On OpenHands f26d734a848297d8dcf460b0bb739174e76511f0 / Canvas 1.16.0 with agent-server 1.44.1, automation 1.9.1, Claude Code 2.1.251, and @agentclientprotocol/[email protected], Canvas had openhands-automation enabled. A new authenticated Claude ACP session was explicitly asked to use that provided skill and change the installed Code Review Automation schedule without filesystem/source inspection or API reconstruction. Claude completed the turn but said: “the openhands-automation skill instructions were not provided to this session” and refused the change.

    The captured frontend request sent only agent_profile_id; the resolved ACP conversation had agent_context.skills=[] and all skill-loading flags false. This locates the loss at the Agent Canvas → Claude Code ACP session/profile bridge, before Claude could select the skill. Credentials and automation access were ruled out by authenticated control trials in which Claude successfully changed the same automation through runtime API discovery.

  4. added
    ready-for-devScoped for contribution. Managed by issue-readiness-check.yml.
    on Aug 30, 2026
  5. ameyypawar commented on Sep 2, 2026

    @ameyypawar

    The skills=[] in the boundary evidence looks like a consequence of the profile-launch path rather than skills projection being absent. Canvas does build agent_context with the enabled/disabled skill lists — buildAgentContext in agent-server-adapter.ts, which sets skills, disabled_skills and load_public_skills: false. But agent_profile_id and agent_settings are mutually exclusive on the start request, so a launch from an agent profile, which is what the ACP preset uses, carries no agent_context at all.

    Worth noting the npm client isn't the constraint: agent_launch_additions is already typed on the start request in the pinned @openhands/typescript-client 1.39.0, and this repo already casts to send a field the generated client doesn't type (from_event_id). What's missing is an agent-server that accepts skills_append.

    One process note: OpenHands/software-agent-sdk#4717 carries a closing keyword for this issue, and cross-repo closing keywords do close on merge, so this will close when that lands.

  6. openhands-ai commented on Sep 2, 2026

    @openhands-ai

    @ameyypawar it looks like you haven't created an OpenHands account yet. Please sign up at OpenHands Cloud and try again.

  7. gokay-ai commented on Sep 14, 2026

    @gokay-ai
    Contributor

    Looking into Claude Code ACP not receiving Canvas-enabled built-in skills — will bridge enabled skills into the ACP session/profile path.

  8. added 2 commits that reference this issue on Oct 1, 2026
    f38b247
    a623feb
  9. added
    feature-map:mappedLinked from a feature-map note. May be reproduced or blocked; see the note for evidence.
    on Oct 12, 2026
  10. jpelletier1 commented on Oct 16, 2026

    @jpelletier1
    Contributor

    @DevinVinson worth prioritizing given this seems to be a longstanding issue with ACP

  11. removed this from the milestone on Oct 19, 2026
  12. simonrosenberg commented on Oct 23, 2026

    @simonrosenberg
    Member

    Root cause and fix, as a stack of three PRs:

    1. fix(launch): keep client-supplied skills on ACP agents under native sourcing software-agent-sdk#6153 (SDK Fix field deprecation in runloop runtime client #6152): under the default native ACP skill sourcing, finalize() cleared an ACP agent's skills after merging agent_launch_additions.skills, so skills a client sends at launch were dropped. It now strips only server-managed skills.
    2. fix(acp): send the enabled catalog skills when launching an ACP profile #18834: Agent Canvas sends the enabled catalog skills in agent_launch_additions.skills for ACP profiles (supersedes the draft fix: project enabled Canvas skills into Claude Code ACP sessions #17421, whose system_message_suffix_append channel is capped at 32 KiB while openhands-automation alone is about 48 KB).
    3. feat(acp): expose skills to every ACP agent as files plus a path catalog software-agent-sdk#6171 (SDK chore(deps): bump the version-all group in /frontend with 5 updates #6170): ACP CLIs have no invoke_skill tool, so the skills are written to <persistence_dir>/acp/skills/<name>/SKILL.md and listed in the prompt with their path, for every ACP agent.

    Checked with real Claude Code and Codex sessions: a skill with no trigger word, and a request without one, now load in both. Canvas needs an agent-server release carrying #6153 and #6171 before it can pin it.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    agent-canvasbugSomething isn't workingfeature-map:mappedLinked from a feature-map note. May be reproduced or blocked; see the note for evidence.priority:mediumMaterial bug with limited impact or a reasonable workaround.ready-for-devScoped for contribution. Managed by issue-readiness-check.yml.skillsRelated to skills

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions