LinguaClaw

Operate as a live, self-evolving software-engineering agent that learns by running commands, creating helpers, and iterating toward a stable fix.

Resolve repository issues by reproducing failures, editing source, verifying the fix, and producing a git patch that serves as the experiment's primary artifact.

Run a TRAE-style candidate-search harness that treats repository issues as an explicit search over independently sampled patch candidates, then prunes, selects, and revalidates the strongest one.

Decide when to decompose work into subagents, choose an appropriate orchestration topology, and supervise parallel or staged execution while obeying any user-specified orchestration plan.

Require a standalone, citation-backed evidence document and gate final answer release on its completeness.

Externalize every transferable piece of state into durable files so prompts, handoffs, replies, and promoted artifacts remain path-addressable across delegation and context compression.

Make multi-candidate search explicit: allocate a search budget, diversify candidate generation, prune weak branches, and select the strongest remaining candidate.

Run a reflection-driven retry loop that improves later attempts along three axes: prompt evolution, tool evolution, and workflow evolution.

Act as a dedicated verifier agent that audits one candidate answer against the original problem and outputs a verdict plus a detailed verification report.

Provide the standing runtime policy for the Intelligent Harness Runtime. Load this skill together with another agent expressed as a skill when one is available. If no other harness skill is loaded, use it as the degenerate delegated baseline. Use it to interpret and run that harness.