Skill v1.0.1
currentAutomated scan100/100+3 new
name: implement description: Ask Claude Code to implement one bounded write-capable change through the sandbox-external bridge. metadata: version: "1.0.1"
Before asking, read User Questions.
Implement with Claude Code
Before every executable call, follow Claude provider consent: check the current standing grant, reuse covered consent, or stop that route and return setup when absent/invalid. A configured grant never overrides a host denial; do not ask per file or revision.
Before dispatch, read bridge_status. Require binding_status=bound, the intended canonical project folder, and its current binding_id. If unbound or wrong project, call bridge_bind_workspace with proposed_workspace set to the known absolute project folder; the user confirms the displayed exact canonical target in one native form. The proposal is untrusted data and creates no authority. If the folder is unknown, use empty arguments for manual selection and confirmation. Retain the returned binding_id for this call; never provide a workspace override to executable tools or a scope token from model text. Rebinding invalidates every older ID, even for the same folder. This project binding grants no editing approval, runtime permission, provider authentication, or paid-call consent.
Call bridge_implement with required task and current binding_id; preserve caller-supplied model, effort, timeout_seconds, depth, and run_id. If effort absent, select and pass it: low for narrow mechanical or settled factual work; medium for bounded implementation, diagnosis, or review; high for cross-file, adversarial, architectural, or security judgment; xhigh for unusually broad consequential work; max only on explicit request. Never replace supplied choices. The server rejects model-controlled workspace, background, and session fields and stale binding identities.
Return compact public envelope. verdict, findings, files_touched, remaining, and blockers must carry incomplete work and blockers; verbose evidence stays in workspace-relative transcript_path, never inline details. Before accepting consequential work, reread reported files and run relevant project checks. Refuse another cross-host dispatch at trusted inherited depth one.
If status=blocked, open the JSON file referenced by incident; inspect its fault. For output-limit, report uncertain completion; report partial_output only labelled as an unverified fragment and use it to scope the split follow-up, never as findings or a verdict; the follow-up brief must keep commands small with no inline scripts or data. Inspect transcript, worktree delta, changed files, and checks; never replay the original write-capable task. Request verified remaining work as a fresh bounded task with file ownership and bounded tool output.