gather-evidence

Gather evidence

Treat human feedback as a claim or requested outcome to investigate. Do not edit code, reply to the person, resolve their thread, or change their review state.

Define the evidence cases

Read the complete feedback and its nearby context. Split independent claims into separate evidence cases. Record each case using the evidence fields:

• the exact feedback source
• the requested behavior or result
• the expected and reported behavior
• the exact check and required environment
• the current target, pull request, commit, build, or deployment

Do not translate a concrete observation into a proxy. Ask for missing context only when no direct check can resolve it.

Inspect the current target

Resolve the current target before gathering evidence. For pull request feedback, record the current head SHA and confirm that the local checkout or deployed build matches it. A later push invalidates the evidence.

Gather direct evidence from the interface named by the feedback. Inspect the code and diff for source claims. Run the exact interaction for behavioral claims. Use the applicable driver for a UI, CLI, or runtime. Before changing a mobile simulator or running mobile automation, invoke running-mobile-simulators.

Follow the direct-evidence rules. Record the command or interaction, target identity, environment, observation, and artifact location. Do not use a passing check, internal call, element presence, or agent report as proof of a different property.

Return a verdict

Return one verdict for each evidence case:

• confirmed. Direct evidence reproduces the reported failure or shows that the requested outcome is absent from the current target.
• resolved. Direct evidence shows the requested outcome on the current target and does not reproduce the reported failure.
• blocked. The required target, environment, capability, or observation is unavailable.
• decision-needed. The feedback requires a product or preference decision that evidence cannot answer.

For confirmed, pass the evidence case to patina-mode's matching playbook. Use Bug fix for a defect and Feature or Refactoring for a requested change. After the target changes, run gather-evidence again before preparing a response.

Report

Return the feedback source, verdict, target identity, direct observation, and artifact or command for each case. State exactly what blocks a blocked case. Leave every human-authored conversation open for the operator.

When a case has local video evidence, add a compact Linear handoff directly after that case:

• Write a short Markdown note the operator can paste without rewriting. Name the verdict, tested target, observation, and video filename. Include build, deployment, and device identity when available. Write one note per evidence case and name every video that proves it.
• Before cleanup, copy any video from a simulator, lease, or temporary directory that cleanup will remove to a path the current run will preserve. Leave the copy on disk at the reported path.
• Put one clickable file link per video immediately after its note. Render it as [filename.mp4](/absolute/path/filename.mp4). Wrap a target containing spaces in angle brackets. Keep the local link outside the paste-ready note because a Linear reader cannot open the operator's local path.
• If the host cannot render a clickable local file, return the exact absolute path and state that the operator must locate it manually. Report only actions that occurred; a link or path is not a copy, attachment, or upload.