| name | aicoder-harvest-and-promote |
| description | Harvest learnings from target sessions and promote proven overlay patterns to the shared control-plane base |
| user-invocable | false |
aicoder-harvest-and-promote
Improvement and harvest workflow for the aicoder-opencode control plane.
When to use
- After a successful target session.
- Periodically (weekly) to keep shared skills current with what is working.
- When a pattern in a target overlay has proven reliable across multiple sessions.
Harvest procedure
After a target session:
- Review what the session produced: did any skill or agent pattern work
unexpectedly well or fail?
- Check
targets/<target>/overlay/.agents/skills/ for any local skills that
were added or modified during the session.
- For each candidate: determine whether it is target-specific or cross-repo
general.
Promotion criteria
Promote a local overlay skill to the shared base when:
- It has worked correctly in at least 2 independent sessions.
- It makes no assumptions about the target repo's product code.
- It would serve at least one other target unchanged or with minimal
parameterization.
Promotion procedure
- Move the skill source to
.agents/skills/<skill-name>/SKILL.md.
- Add it to
.agents/skills/.manifest.json if one exists.
- Update the
instructions array in opencode.jsonc.
- Remove or thin the local overlay copy to a shim or pointer.
- Run
bun run check to confirm nothing broke.
Model registry improvement
After sessions that used specific models:
- Note which provider routes were actually called vs. skipped.
- If a model's top provider route consistently failed: update
config/models.jsonc
to reorder or disable it.
- Run
make manage-models to interactively adjust, then make validate-dr-repo
to confirm the change does not break the target path.
Plugin improvement signals
- A plugin tool was called in every session: probably good, keep it.
- A plugin tool was never called: consider whether it should stay in the hot set.
- A plugin returned wrong data: fix the source in
src/plugins/, propagate shim
updates to any affected target overlays.
Avoid
- Promoting skills that encode product-specific assumptions.
- Demoting stable skills based on one bad session.
- Changing the model registry without running a test probe first.
