Mine agent transcripts

Past corrections are a materialized process. They are the spec for what the environment should have prevented.

Read the history

From any worktree of this repository:

python3 .agents/skills/mine-agent-transcripts/scan.py --days 30
python3 .agents/skills/mine-agent-transcripts/scan.py --days 90 --corrections

--corrections narrows to user turns that look like a correction. --full keeps long messages intact. The script is read-only, resolves the project root through git rev-parse --git-common-dir, and matches every session and worktree of the project.

Set OPENCODE_DB to read another OpenCode database. The script covers OpenCode sessions only, not Claude Code or Cursor history.

Cluster, do not fix

Group the turns by what the agent kept getting wrong. A cluster needs at least two turns and the same underlying shape:

• the same wrong file or layer touched
• the same missing verification step
• the same wording correction, repeated
• the same question the agent should have answered from the repository

Write one line per cluster with links to the turns. Ignore preferences that appear once and never again; they are noise, not process.

Hand off the clusters

Pass the cluster list to harden-environment. Each cluster becomes one constraint, at the cheapest layer that makes the mistake impossible. If no constraint is possible, the cluster is a judgment call: write the rule once, with the why, and stop repeating it.

Rules

• Never commit the transcript dump. It contains prompts, paths, and user content.
• Never quote a transcript as evidence in a PR without the user's approval.
• Never turn a single annoyed message into a repository rule. Two occurrences minimum, or the rule is just noise with extra maintenance.
• The output is evidence, not the fix. The fix lives in the environment.