Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 2 additions & 1 deletion .claude/skills/import-benchmark/SKILL.md
Original file line number Diff line number Diff line change
Expand Up @@ -217,5 +217,6 @@ detailed prerequisites and single/multiple datasets should remain readable.
The catalogue presents one Original procedure. Put a published Harbor agent in
`benchmark.json.published.agent` so the builder promotes it to the reference
group. For a standard benchmark, follow import-standard's `reference_agent`
and paper-model declarations. Keep participant choice on Agent/Memory and
and paper-model declarations, or its own agents (`agents.json`, `start.driver`)
when the upstream ships several agent scripts. Keep participant choice on Agent/Memory and
model filtering on Model; do not add separate Original/your-model setup rows.
7 changes: 6 additions & 1 deletion .claude/skills/import-standard/SKILL.md
Original file line number Diff line number Diff line change
Expand Up @@ -447,7 +447,12 @@ upstream rule as a Nullius defect. Show one **Original** procedure. Model select
**Like the paper** filters documented models; **Any of your choice** uses the
provider catalogue. Do not invent paper models. Declare `reference_agent` on
`setups.original` when upstream runs a named agent; otherwise the builder shows
**No agent · Original**, or **Original memory** for a memory benchmark. The
**No agent · Original**, or **Original memory** for a memory benchmark. An
upstream that ships several agent scripts brings them as its own agents:
`agents.json` beside `standard.json`, each entry an agent with `start.driver`
and its published `models` (docs/standard.md, "A benchmark's own agents"),
`options.agent` written with its label, summary and setups only; recipes then
name one as `<benchmark>:<key>`. The
`original` / `model` driver methods preserve upstream model aliases; they are
not separate user-facing setups. Recipes select `others: ["original"]`, one
`model`, and `model_choice: "paper" | "any"`.
3 changes: 3 additions & 0 deletions .claude/skills/remove-integration/SKILL.md
Original file line number Diff line number Diff line change
Expand Up @@ -21,6 +21,9 @@ the user's requested scope, including logs and stored data when requested.
- Catalogue entries live under `app/nullius/catalogue/`. Capabilities can also
have leftover folders under the legacy `features/` directory. Remove the
whole entry folder: an empty folder can still be selected by the loader.
- A benchmark's own agents are entries of its folder's `agents.json`: they go
with the benchmark. Removing one agent alone is removing its entry, and its
saved recipes (`<benchmark>:<key>`) then name an agent the compile refuses.
- For requested runtime cleanup, locate the application's actual profile,
catalogue, runs, workspaces and standard installs through its settings and
path APIs. With `uv tool`, use the installed tool's interpreter; the shell's
Expand Down
Loading
Loading