cogamer-improve

name	cogamer.improve
description	Implement one targeted CvC code change from analysis, verify improvement, submit if better

Improve

Implement one targeted code change based on an analysis, verify it improves scores, and submit.

Announce at start: "I'm using the improve skill to implement and verify a fix from the analysis."

Prerequisites

Run /cogamer.analyze first to produce cogamer/analysis.md with a diagnosis.

Steps

1. Read Analysis

Read cogamer/analysis.md. Understand the weakness, root cause, and proposed fix.

2. Eval Baseline

softmax cogames eval -m machina_1 -p class=cvc_policy.cogamer_policy.CvCPolicy -e 10 --format json --seed 42

Record the average reward as the baseline.

3. Implement Fix

Make one focused code change as described in the analysis. Touch only the files identified. Keep the diff minimal.

Possible change surfaces:

• Engine logic: src/cvc_policy/agent/main.py, roles.py, targeting.py, pressure.py, scoring.py, navigation.py
• Programs: src/cvc_policy/programs.py (program table, LLM prompt, LLM response parsing)
• Parameters: src/cvc_policy/agent/types.py
• Game state: src/cvc_policy/game_state.py

4. Eval After Change

Run the same eval across 5 seeds:

for seed in 42 43 44 45 46; do
  softmax cogames eval -m machina_1 -p class=cvc_policy.cogamer_policy.CvCPolicy -e 5 --seed $seed --format json
done

5. Decide

• Improved: average reward increased -> keep the change, submit
• Regressed: average reward decreased -> git checkout the changed files, report what didn't work
• Neutral: no significant change -> keep if the change is clearly more correct, otherwise revert

6. Submit if Improved

Get the player name from softmax cogames player list. Submit automatically:

cogames upload -p class=cvc_policy.cogamer_policy.CvCPolicy -n <cogamer-name>

Do NOT ask for confirmation. Log the submission.

7. Update Analysis

Append the result to cogamer/analysis.md:

## Result
- baseline: X
- after: Y
- verdict: improved/regressed/neutral
- submitted: yes/no

Principles

• One change per cycle. Isolate what works.
• Revert on regression. No exceptions.
• Fix root causes, not symptoms. If the analysis says "mining is slow", don't add a band-aid — fix the actual pathfinding or scoring logic.