apply-mining-patches

name	apply-mining-patches
description	Apply the diff/patch outputs from a /mine-transcripts run to lessons.md and rules/*.md after human review. Use after running /mine-transcripts and reviewing the proposed diffs. Triggers on /apply-mining-patches, "apply the mining patches", "apply mining run", "approve and apply the trajectory diffs".
when_to_use	After /mine-transcripts produces diffs and you've reviewed them. Skips approved-then-undone work.
allowed-tools	Bash, Read, Write, Edit
arguments	["run-timestamp"]

/apply-mining-patches — Apply approved trajectory-mining diffs

/apply-mining-patches <run-ts> applies the lessons + rules diffs from ~/.claude/telemetry/mining-runs/<run-ts>/ to live files, after the user has had a chance to review.

What you do

1. Resolve the run

• If <run-ts> is "latest", glob ~/.claude/telemetry/mining-runs/2* and pick the most recent
• Otherwise use the exact timestamp directory

2. Surface what's there

Read and summarize:

• manifest.json — sessions mined, signal counts
• lessons.diff — proposed additions to ~/.claude/lessons.md
• agent-tuning.md — agents with ≥2 occurrences worth tuning
• Any rules-<name>.diff files

3. Confirm with user

Present a summary like:

Mining run 20260510T120000Z:
  - 12 sessions, 7 corrections, 3 patterns, 4 failure modes
  - lessons.diff: +45 lines (3 new lesson entries)
  - agent-tuning.md: 1 agent recommended (flake-detector, 4 occurrences)
  - rules-self-improvement.diff: +8 lines

Apply? (yes/no/preview)

If preview, show the actual diff content (truncated). If no, exit.

4. Apply diffs

For lessons.diff: append to ~/.claude/lessons.md.

cat <(echo) <(cat <run-dir>/lessons.diff) >> ~/.claude/lessons.md

For rule diffs: append to the named rule file.

For agent tuning candidates: do NOT apply automatically. Surface the /tune-agent <name> command and let the user run it explicitly. Reflexion patches require their own approval.

5. Mark applied + record predictions

After successful application, write to <run-dir>/applied.marker with the apply timestamp. This prevents accidental double-application.

Then record a falsifiable prediction per applied rule diff (AHE decision observability — /rl-status scores these when due):

python3 ~/.claude/rl/prediction_ledger.py add \
  --change "rule: <rule-name> (mining run <run-ts>)" \
  --metric correction_rate --agent session --direction down --horizon-days 45

A rule whose prediction scores WRONG twice is a sunset candidate per rules/writing-rules.md — cite the ledger entries in the rent audit.

6. Commit (if in a git repo)

If the user wants, commit the lessons/rules updates with a message like:

docs(memory): apply mining run 20260510T120000Z

3 new lessons, 1 rule update from 12 sessions of trajectory mining.
Refs: ~/.claude/telemetry/mining-runs/20260510T120000Z/

Anti-patterns

• Applying without showing the user the actual content first — patches can be wrong
• Auto-applying agent-tuning candidates — those need /tune-agent's own approval flow
• Applying a run that's already marked applied — fail loudly
• Discarding a run's directory after apply — keep it for reproducibility
• Editing the original miner output file — the diffs are immutable; if a diff is wrong, that's a bug in the miner, fix the miner