dcc-cua
Project-owned DCC-CUA routing skill for bounded application UI automation. Use this skill whenever the user says dcc-cua, DCC CUA, our dcc-cua, or 我们的 dcc-cua, including browser and other non-DCC UI tasks. These phrases are a hard route to the dcc-mcp project UI Control stack and
Install
npx skills add https://github.com/dcc-mcp/dcc-mcp-core/tree/main/skills/dcc-cua
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install dcc-mcp-dcc-mcp-core@llmmart
git clone https://github.com/dcc-mcp/dcc-mcp-core.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole dcc-mcp/dcc-mcp-core collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
DCC-CUA — Project UI Control Router
Use this Skill as the canonical route whenever the user explicitly names
dcc-cua, DCC CUA, our dcc-cua, or 我们的 dcc-cua.
Non-substitution contract
An explicit DCC-CUA request is a hard routing boundary, regardless of whether the target is Maya, Chrome, a browser, or another desktop application.
- Use the project-owned
dcc-cuaruntime and DCC-MCPui-controlsurface. - Never load or call generic Codex/OpenAI Computer Use, the
computer-useSkill,@oai/sky, or Browser/Chrome automation plugins for that request. - Never treat a DCC-CUA runtime, binding, readiness, or permission failure as permission to change providers.
- Repair the project route when safely possible. Otherwise report the exact blocker and stop.
- Use a generic provider only after the user explicitly retracts the DCC-CUA requirement or explicitly requests that provider by name.
This boundary is provider selection, not an authorization bypass. DCC-CUA task grants, target binding, interruption, and confirmation policy still apply.
Embedding developers may use Core's constructor-only managed_launcher
dependency only for an exclusively transferred local MCP child. Read
owned worker launch before
integration. The embedding must supply the actual authorized launcher; this
preparatory interface does not establish independent supervisor readiness,
attach to a shared worker, export a lease, or permit changing Job policy.
Automatic route from search
The contract above also covers requests that never name dcc-cua. When a
skill or tool search returns no usable interface, the search surface attaches
fallback advice naming this Skill, with a reason code:
| reason | meaning |
|---|---|
no_candidate |
nothing in the catalog matched the request |
low_confidence |
the closest matches are lexical near-misses, not a real interface |
no_executable_interface |
the closest match is documentation-only and declares no callable tool |
The advice is additive: the hits are still returned, and the routing criteria never reorder or drop them.
The advice is blocked, with reason code cua_runtime_unavailable, when this
runtime is not usable. A blocked response is a blocker to report or repair —
it is not permission to switch provider. The non-substitution contract above
applies unchanged to the automatic route.
Ranking is deliberately not involved. This Skill declares
layer: infrastructure, which the search rank policy demotes to 0.35 so it
stays out of neutral discovery. That must not be "fixed" by raising the
coefficient: the route is explicit, not a ranking effect.
Runtime preflight
Use the official component contract; do not download an arbitrary executable:
dcc-mcp-cli components status dcc-cua
dcc-mcp-cli components ensure dcc-cua --yes
dcc-cua manifest
dcc-cua ping
Run components ensure only when installation or repair is authorized. It
consumes the official versionless manifest, verifies the declared SHA-256, and
reconciles the independently released companion executable.
The two binaries come from different releases: dcc-mcp-cli ships with
dcc-mcp-core (https://github.com/dcc-mcp/dcc-mcp-core/releases), while the
dcc-cua runtime is an independent companion released from
https://github.com/dcc-mcp/dcc-cua/releases. components ensure dcc-cua
fetches the runtime from that dcc-cua release, never from a Core release.
For an exact, already qualified offline package, use
dcc-mcp-cli components install-verified dcc-cua --help. Supply the canonical
--target-core-cli and its --expected-core-cli-sha256, --source-root, exact
--version, new and --previous-manifest paths and their pinned SHA-256 values,
plus --expected-current-sha256, --expected-new-sha256, and --yes.
Obtain manifest digests from independently verified official publications.
The command disables automatic update maintenance and gateway startup, copies
only declared files into a private verified staging tree, and checks the Core
anchor and old foreground under the installation lock. It preserves historical
backup files and rejects a pending package journal with recovery required;
recovery is a separate operation. An error after commit can require cleanup;
read back installed hashes and the journal before reporting an unchanged
installation or retrying. Installing files does not reconnect loaded callers.
Retryable rename/delete failures retain prepared ownership only when a live
OS file identity still matches. Windows requires the OS and filesystem to
support FileIdInfo / 128-bit file IDs;
availability of the older basic handle-information API alone is insufficient.
An unsupported or unreadable initial identity check preserves the ownership
ticket without attempting consumption. If identity cannot be rechecked after
a failed operation, the ticket remains relinquished and inspection is required.
Do not assume compatibility with every older Windows version or filesystem,
or delete an ambiguous sidecar to force recovery.
These checks complement the installation lock, not a filesystem sandbox
against unrelated writers that ignore that lock.
For semantic application profiles:
dcc-cua profiles
dcc-cua profile --id <profile-id>
Do not invent a profile ID. Runtime-advertised capabilities are authoritative.
Execution order
- Prefer a typed DCC-MCP host tool when it directly expresses the operation.
- Use DCC-CUA only for the UI behavior that typed host tools cannot expose.
- Bind one exact target with process ID and native window handle whenever the surface supports them.
- Open one scoped session with the minimum task grant required for the work.
- Take a fresh observation before each action that depends on UI state.
- Act using stable semantic control or DOM references when available.
- Wait for a typed state transition and verify the real final state.
- Stop the session on success, failure, interruption, or abandonment.
An input sent acknowledgement is not completion evidence.
For native application menu bars, prefer the negotiated native_menu_path
route through ui_control__act(action="invoke_menu", menu_path=[...]) when a
semantic menu click or Alt mnemonic cannot prove that a popup opened. A menu
invocation invalidates the current observation; honor verification_required
and verify the popup or resulting application state with a fresh snapshot.
DCC-host route
For a registered DCC instance, use the DCC-MCP UI Control tools or their CLI projection:
dcc-mcp-cli load-skill ui-control --instance-id <instance-id> --output toon
dcc-mcp-cli ui-control snapshot --instance-id <instance-id> --json '{"session_id":"ui","process_id":1234,"window_handle":5678}'
dcc-mcp-cli ui-control act --instance-id <instance-id> --json '{"session_id":"ui","control_id":"ok","action":"click","snapshot_id":"<snapshot-id>"}'
dcc-mcp-cli ui-control act --instance-id <instance-id> --json '{"session_id":"ui","action":"invoke_menu","menu_path":["Window","Arrange","Left"]}'
dcc-mcp-cli ui-control stop --instance-id <instance-id> --json '{"session_id":"ui"}'
Use the same exact instance and session throughout the action chain. Do not switch to another DCC process because it looks similar.
Browser and non-DCC route
Browser work remains inside DCC-CUA. Use the Host's typed browser surface and
browser_dom capabilities, not an in-app Browser or Chrome plugin.
- Bind the exact browser PID and window handle first.
- Bind the exact tab/target returned by DCC-CUA; do not infer it from a title alone when an exact target identifier exists.
- Keep connection-scoped sessions and capabilities on one Host connection.
- Use DOM/semantic references from the latest observation rather than stale coordinates.
browser_prepareand existing-profile attachment require both the Host grant and the session task grant advertised by the runtime contract.- Authentication challenges, CAPTCHAs, purchases, account/security changes, and unexpected permission prompts remain trusted human boundaries.
Target and evidence invariants
- Preserve PID and native window handle in observations and audit records.
- Treat a changed PID, window handle, tab target, or session owner as a fresh binding that requires a fresh observation.
- Keep
fullreadiness strict. If only an exact-window or typed-browser route is independently ready, report that route-specific degraded readiness. - Honor Escape/user interruption immediately and do not resume without fresh authorization and observation.
- Do not expose local usernames, internal package paths, browser profile data, credentials, tokens, or unrelated window titles in public evidence.
- Verify success by reading the destination state after the mutation.
Failure behavior
When DCC-CUA cannot complete the request, report:
- the exact component/runtime version,
- the exact target identity that was bound,
- the failing readiness, capability, permission, or action stage,
- the last safe observation or typed error, and
- the safe next repair step.
Do not mention a generic Computer Use fallback unless the user asks for one.
Files (dcc-mcp-core)
-
agents
-
openai.yaml 404 B
Loading…
-
-
SKILL.md 10.5 KB
--- name: dcc-cua description: >- Project-owned DCC-CUA routing skill for bounded application UI automation. Use this skill whenever the user says dcc-cua, DCC CUA, our dcc-cua, or 我们的 dcc-cua, including browser and other non-DCC UI tasks. These phrases are a hard route to the dcc-mcp project UI Control stack and take precedence over generic Codex/OpenAI Computer Use, the computer-use Skill, @oai/sky, and Browser/Chrome plugins. If DCC-CUA is unavailable, repair or report the project route; never silently fall back to a generic computer-use provider. license: MIT-0 allowed-tools: Bash Read metadata: dcc-mcp: dcc: python recall-context: app_type: python domain: ui-automation workflow_stage: automation task_category: mutate layer: infrastructure compatibility: Cross-platform routing contract. The current DCC-CUA host is installed and verified through the official dcc-mcp-cli component manifest; exact platform capabilities remain runtime-discovered. version: "0.1.0" search-hint: "dcc-cua DCC CUA dcc cua our dcc-cua our dcc cua 我们的 dcc-cua 我们的 dcc cua project-owned UI control browser DOM exact PID HWND computer use automation" tags: "dcc-cua, dcc-ui-control, ui-control, browser-dom, exact-window, computer-use, infrastructure" openclaw: emoji: "🖱️" homepage: https://github.com/dcc-mcp/dcc-mcp-core/tree/main/skills/dcc-cua --- # DCC-CUA — Project UI Control Router Use this Skill as the canonical route whenever the user explicitly names `dcc-cua`, `DCC CUA`, `our dcc-cua`, or `我们的 dcc-cua`. ## Non-substitution contract An explicit DCC-CUA request is a hard routing boundary, regardless of whether the target is Maya, Chrome, a browser, or another desktop application. - Use the project-owned `dcc-cua` runtime and DCC-MCP `ui-control` surface. - Never load or call generic Codex/OpenAI Computer Use, the `computer-use` Skill, `@oai/sky`, or Browser/Chrome automation plugins for that request. - Never treat a DCC-CUA runtime, binding, readiness, or permission failure as permission to change providers. - Repair the project route when safely possible. Otherwise report the exact blocker and stop. - Use a generic provider only after the user explicitly retracts the DCC-CUA requirement or explicitly requests that provider by name. This boundary is provider selection, not an authorization bypass. DCC-CUA task grants, target binding, interruption, and confirmation policy still apply. Embedding developers may use Core's constructor-only `managed_launcher` dependency only for an exclusively transferred local MCP child. Read [owned worker launch](../../docs/guide/ui-control-managed-worker.md) before integration. The embedding must supply the actual authorized launcher; this preparatory interface does not establish independent supervisor readiness, attach to a shared worker, export a lease, or permit changing Job policy. ## Automatic route from search The contract above also covers requests that never name `dcc-cua`. When a skill or tool search returns no usable interface, the search surface attaches `fallback` advice naming this Skill, with a reason code: | reason | meaning | |---|---| | `no_candidate` | nothing in the catalog matched the request | | `low_confidence` | the closest matches are lexical near-misses, not a real interface | | `no_executable_interface` | the closest match is documentation-only and declares no callable tool | The advice is **additive**: the hits are still returned, and the routing criteria never reorder or drop them. The advice is blocked, with reason code `cua_runtime_unavailable`, when this runtime is not usable. A blocked response is a blocker to report or repair — it is not permission to switch provider. The non-substitution contract above applies unchanged to the automatic route. Ranking is deliberately not involved. This Skill declares `layer: infrastructure`, which the search rank policy demotes to `0.35` so it stays out of neutral discovery. That must not be "fixed" by raising the coefficient: the route is explicit, not a ranking effect. ## Runtime preflight Use the official component contract; do not download an arbitrary executable: ```bash dcc-mcp-cli components status dcc-cua dcc-mcp-cli components ensure dcc-cua --yes dcc-cua manifest dcc-cua ping ``` Run `components ensure` only when installation or repair is authorized. It consumes the official versionless manifest, verifies the declared SHA-256, and reconciles the independently released companion executable. The two binaries come from different releases: `dcc-mcp-cli` ships with `dcc-mcp-core` (`https://github.com/dcc-mcp/dcc-mcp-core/releases`), while the `dcc-cua` runtime is an independent companion released from `https://github.com/dcc-mcp/dcc-cua/releases`. `components ensure dcc-cua` fetches the runtime from that dcc-cua release, never from a Core release. For an exact, already qualified offline package, use `dcc-mcp-cli components install-verified dcc-cua --help`. Supply the canonical `--target-core-cli` and its `--expected-core-cli-sha256`, `--source-root`, exact `--version`, new and `--previous-manifest` paths and their pinned SHA-256 values, plus `--expected-current-sha256`, `--expected-new-sha256`, and `--yes`. Obtain manifest digests from independently verified official publications. The command disables automatic update maintenance and gateway startup, copies only declared files into a private verified staging tree, and checks the Core anchor and old foreground under the installation lock. It preserves historical backup files and rejects a pending package journal with `recovery required`; recovery is a separate operation. An error after commit can require cleanup; read back installed hashes and the journal before reporting an unchanged installation or retrying. Installing files does not reconnect loaded callers. Retryable rename/delete failures retain prepared ownership only when a live OS file identity still matches. Windows requires the OS and filesystem to support [`FileIdInfo` / 128-bit file IDs](https://learn.microsoft.com/en-us/windows/win32/api/winbase/ns-winbase-file_id_info); availability of the older basic handle-information API alone is insufficient. An unsupported or unreadable initial identity check preserves the ownership ticket without attempting consumption. If identity cannot be rechecked after a failed operation, the ticket remains relinquished and inspection is required. Do not assume compatibility with every older Windows version or filesystem, or delete an ambiguous sidecar to force recovery. These checks complement the installation lock, not a filesystem sandbox against unrelated writers that ignore that lock. For semantic application profiles: ```bash dcc-cua profiles dcc-cua profile --id <profile-id> ``` Do not invent a profile ID. Runtime-advertised capabilities are authoritative. ## Execution order 1. Prefer a typed DCC-MCP host tool when it directly expresses the operation. 2. Use DCC-CUA only for the UI behavior that typed host tools cannot expose. 3. Bind one exact target with process ID and native window handle whenever the surface supports them. 4. Open one scoped session with the minimum task grant required for the work. 5. Take a fresh observation before each action that depends on UI state. 6. Act using stable semantic control or DOM references when available. 7. Wait for a typed state transition and verify the real final state. 8. Stop the session on success, failure, interruption, or abandonment. An `input sent` acknowledgement is not completion evidence. For native application menu bars, prefer the negotiated `native_menu_path` route through `ui_control__act(action="invoke_menu", menu_path=[...])` when a semantic menu click or Alt mnemonic cannot prove that a popup opened. A menu invocation invalidates the current observation; honor `verification_required` and verify the popup or resulting application state with a fresh snapshot. ## DCC-host route For a registered DCC instance, use the DCC-MCP UI Control tools or their CLI projection: ```bash dcc-mcp-cli load-skill ui-control --instance-id <instance-id> --output toon dcc-mcp-cli ui-control snapshot --instance-id <instance-id> --json '{"session_id":"ui","process_id":1234,"window_handle":5678}' dcc-mcp-cli ui-control act --instance-id <instance-id> --json '{"session_id":"ui","control_id":"ok","action":"click","snapshot_id":"<snapshot-id>"}' dcc-mcp-cli ui-control act --instance-id <instance-id> --json '{"session_id":"ui","action":"invoke_menu","menu_path":["Window","Arrange","Left"]}' dcc-mcp-cli ui-control stop --instance-id <instance-id> --json '{"session_id":"ui"}' ``` Use the same exact instance and session throughout the action chain. Do not switch to another DCC process because it looks similar. ## Browser and non-DCC route Browser work remains inside DCC-CUA. Use the Host's typed browser surface and `browser_dom` capabilities, not an in-app Browser or Chrome plugin. - Bind the exact browser PID and window handle first. - Bind the exact tab/target returned by DCC-CUA; do not infer it from a title alone when an exact target identifier exists. - Keep connection-scoped sessions and capabilities on one Host connection. - Use DOM/semantic references from the latest observation rather than stale coordinates. - `browser_prepare` and existing-profile attachment require both the Host grant and the session task grant advertised by the runtime contract. - Authentication challenges, CAPTCHAs, purchases, account/security changes, and unexpected permission prompts remain trusted human boundaries. ## Target and evidence invariants - Preserve PID and native window handle in observations and audit records. - Treat a changed PID, window handle, tab target, or session owner as a fresh binding that requires a fresh observation. - Keep `full` readiness strict. If only an exact-window or typed-browser route is independently ready, report that route-specific degraded readiness. - Honor Escape/user interruption immediately and do not resume without fresh authorization and observation. - Do not expose local usernames, internal package paths, browser profile data, credentials, tokens, or unrelated window titles in public evidence. - Verify success by reading the destination state after the mutation. ## Failure behavior When DCC-CUA cannot complete the request, report: 1. the exact component/runtime version, 2. the exact target identity that was bound, 3. the failing readiness, capability, permission, or action stage, 4. the last safe observation or typed error, and 5. the safe next repair step. Do not mention a generic Computer Use fallback unless the user asks for one.
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.