improve-context

Improve context

Use the evaluation system as the referee. Do not promote a context change because the new answer looks more plausible.

1. Diagnose

Read the failing evaluation record and its analysis trace. Classify the context failure:

• missing information;
• wrong information;
• stale information;
• conflicting information;
• correct information was not supplied;
• supplied information was ignored or misapplied;
• the problem belongs in a skill, helper, workflow, or evaluator rather than context.

Run /context-trace on the exact question. Preserve its JSON fingerprint.

2. Propose

For edits to guides, reviewed queries or semantic definitions, use /maintain-context to inspect the appropriate template, validate dependencies and preserve human review.

Create working/context-change-proposal.yaml with:

• observed failure and evidence;
• smallest proposed change;
• representation and delivery;
• source and owner;
• status proposed;
• cases expected to improve;
• possible regressions;
• required reviewers;
• acceptance criteria.

Do not change trusted context until the user approves the proposal.

3. Preserve the baseline

Run the working and regression cases before the edit. Keep the manifest paths and run IDs. Freeze the suite, data snapshot, model, evaluator, and trial count. Record the proposal's target case IDs and use the installed suite agreed for this experiment. Do not silently replace it with a legacy policy fixture. Lock outputs before grading; keep evaluation references outside candidate context. Do not launch model runs if the user has paused them.

4. Apply one change

Make only the approved context change. Show the diff. Do not change the evaluator or expected results in the same comparison.

5. Rerun and inspect

Run the same working and regression cases. Inspect:

• aggregate score;
• target cases;
• regressions;
• cases still failing;
• context manifest;
• context citations and SQL-use evidence.

6. Write the receipt