From an external review of v0.5.0.
Measured
skills/commitlore-commits/SKILL.md mentions harvest 10 times and capture once. Its description — the field that decides when Claude Code loads it — triggers only on explicit requests:
"commit this with commitlore", "write a commitlore record for this change", "harvest a commit message", "commitlore 기록 남겨서 커밋해줘"
Every one requires the user to already know CommitLore exists and to name it.
Why that is a gap now
commitlore capture runs prepare → verify → stage as one command, and the MCP write-side exists. The skill that ships to Claude Code still walks through the older path: harvest → write a JSON draft to a file → harvest-verify → paste trailers into the message.
So the product moved and the agent instruction did not. A user who says "fix this bug and commit it" gets no record, however much decision context the session produced — not because the tool cannot, but because nothing tells the agent to consider it.
What the workflow should be
ordinary commit request
→ does this change carry decision context? (usually no)
→ MCP prepare → draft → MCP verify
→ show one candidate
→ on approval, stage
→ ordinary commit
It has to work without the user saying "commitlore". Most commits should still produce nothing — that is the existing contract and it should not change.
The manual harvest path stays, as an advanced section rather than the lead.
Depends on
Sequencing with the approval contract matters — "show one candidate" has no enforceable meaning until something can hold a candidate before it is staged.
Not in scope
The record vocabulary, the verification rules, and "most commits carry nothing". This is about which path the agent is taught to take.
From an external review of v0.5.0.
Measured
skills/commitlore-commits/SKILL.mdmentionsharvest10 times andcaptureonce. Itsdescription— the field that decides when Claude Code loads it — triggers only on explicit requests:Every one requires the user to already know CommitLore exists and to name it.
Why that is a gap now
commitlore captureruns prepare → verify → stage as one command, and the MCP write-side exists. The skill that ships to Claude Code still walks through the older path: harvest → write a JSON draft to a file → harvest-verify → paste trailers into the message.So the product moved and the agent instruction did not. A user who says "fix this bug and commit it" gets no record, however much decision context the session produced — not because the tool cannot, but because nothing tells the agent to consider it.
What the workflow should be
It has to work without the user saying "commitlore". Most commits should still produce nothing — that is the existing contract and it should not change.
The manual harvest path stays, as an advanced section rather than the lead.
Depends on
Sequencing with the approval contract matters — "show one candidate" has no enforceable meaning until something can hold a candidate before it is staged.
Not in scope
The record vocabulary, the verification rules, and "most commits carry nothing". This is about which path the agent is taught to take.