Proposals

minerva propose (and minerva_propose) turns signals into a ranked list. The last run is stored at ~/.agents/.minerva/proposals.json.

Kind	Meaning	Apply?
new_skill	Draft a SKILL.md from a repeated pattern	yes
update_skill	Merge an observed pattern into an existing skill's ## Observed patterns section	yes, idempotent
load_gap	Catalog skill matched the prompt but was not loaded	no — load it in the harness

Each proposal has a stable id (kind + hash of the skill name).

Proposals come from shell_family, repeat_sequence, retry_loop, and load_gap signals. Names are derived from the anchor: git-workflow, go-strreplace-loop, bash-retry. user_correction and long_manual show up in analyze only.

Applying an update

## Observed patterns

- repeated shell family git (11 sessions; last cursor 6d2fd597)
- retry after error: Bash (2 sessions; last codex 01a0466f)

Minerva owns that section and nothing else. A legacy ## Observed later paragraph from earlier releases is folded into it on the next apply.

minerva propose --since 30d --json
minerva propose apply new_skill-ab12cd34

Drafts include frontmatter (name, description starting with “Use when”), the observed pattern, and up to five session evidence rows.

Minerva does not call an LLM to write the body. The host agent can rewrite the draft after apply if you want richer prose.