From openltm
Mines LTM memories from past git commits. Use when onboarding a repo into LTM, backfilling history after enabling gitLearn, or harvesting patterns after a sprint.
How this skill is triggered — by the user, by Claude, or both
Slash command
/openltm:GitLearnThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Extract durable LTM memories from past git commits by delegating the read-and-extract
Extract durable LTM memories from past git commits by delegating the read-and-extract
work to the dedicated git-learner agent (shipped with this plugin). The agent
reads the diffs and stores memories via the learn MCP tool, so this path needs
no API key and keeps the main thread's context clean (raw diffs stay in the agent).
| Invocation arg | Commits processed |
|---|---|
| (none) | last 10 |
--commits N | last N |
--since YYYY-MM-DD | all commits since that date |
gitLearnEnabled for the first time.The background post-commit hook (GitCommit.ts) calls an LLM API directly and needs
a configured key. This skill runs interactively instead, so it spawns the
git-learner agent via the Agent tool. Spawning an agent is valid here because
skills execute in the live Claude Code session, which has the Agent tool available.
The agent carries its own extraction rubric (signal-to-noise rules, category mapping,
storage fields), so this skill only has to supply the scope.
Determine the commit range from the invocation arg (default: last 10). Capture the repo root so the subagent runs git in the right directory:
git rev-parse --show-toplevel
git log --pretty=format:'%H %s' -<N> # or --since="<date>"
git-learner agent onceCall the Agent tool once with subagent_type: "git-learner". A single agent processes
the whole batch — do not spawn one per commit. The agent's system prompt already holds
the extraction rubric, so the spawn prompt only supplies a <scope> block:
<scope>
REPO_ROOT: <repo root from git rev-parse --show-toplevel>
PROJECT_NAME: <repo directory basename>
COMMITS:
<one commit hash per line, from Step 1>
</scope>
Pass nothing else — no rubric, no instructions. The agent knows what to keep, what to skip, how to map categories, and which fields to store.
Relay the subagent's table and total. Memories are stored with
source: "git-commit:<hash>" and file-path tags, queryable via
mcp__plugin_openltm_memory__recall.
mcp__plugin_openltm_memory__recall query="git commit patterns" — check what's
already stored, so the subagent reinforces rather than duplicates.npx claudepluginhub rohirik/openltm --plugin openltmExtracts learnings from conversations, audit history, and git logs; manages memories and cleans stale entries. Use after tracks, retrospectives, or maintenance.
Extracts recurring patterns from git commits/PRs/handoffs/CLAUDE.md via bash/grep; recommends artifacts (skills/rules/hooks/agents) using frequency thresholds.
Summarizes recent codebase activity over a time period (default 2 weeks): features landed, fixes, refactors, focus/neglected areas, and contributor patterns. Uses git log or timewarp tools for quick catch-up.