researchforge-plan · git:20260720.c0c9164 · 2026-07-20 · sha256 5cf1c87af2e26fda
researchforge-plan git:20260720.c0c9164A
Immutable. This exact content never changes and is served at /api/v1/blob/5cf1c87af2e26fda.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_ycv6kqaf2ckmltxl/5cf1c87af2e26fdaad82facaed7d69f777e9c60140ced2dc609067cdde8077d2.json
Then /plugin install researchforge-plan@markdownregistry-ycv6kqaf2ckmltxl in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
--- name: researchforge-plan description: Design experiment variants for a hypothesis — write patches, import the plan for validation, and get it approved. Use after a baseline exists, when the user wants to plan or implement experiments. --- # Plan experiments You author the plan and patches; the engine validates every layer before anything can run. One plan targets one hypothesis. ## 1. Export planning context ```bash researchforge experiment plan <hypothesis-id> --json ``` This writes `.researchforge/experiments/context.json`: the hypothesis, the approved contract (metric, constraints, **editable and protected paths**, limits), baseline metrics, repository scan, and the embedded schema for the plan file. Read it fully before writing anything. ## 2. Write the plan and patches - `.researchforge/experiments/plan.yaml` conforming to the embedded schema: `hypothesis_id`, `approach_summary`, and one entry per experiment (`key`, `title`, `change_summary`, `patch_file`). - One unified diff per variant under `.researchforge/experiments/patches/`, applying cleanly to the frozen baseline commit. Hard constraints the engine enforces (do not fight them): - patches may only touch `permissions.editable_paths`; a patch touching a protected path is recorded as **rejected** at import and will never run — never edit benchmarks, evaluation code, or `.researchforge/` to make a metric look better; - changed files are extracted by git from the patch itself, never from your description; - keep variants small and single-idea: the funnel measures one change at a time; - to BUILD ON a prior result, set `parent:` on an entry (another key in this plan or an `exp-NNN` from the context's `prior_experiments`) — the parent's patch chain is applied first and your diff must be written against that combined state. A branched child's improvement is still measured against the baseline; report its increment over the parent separately and honestly. ## 3. Import (engine validates, six layers) ```bash researchforge experiment import .researchforge/experiments/plan.yaml --json ``` On failure, the `--json` payload lists exactly what to fix (schema fields, unknown hypothesis, patch that does not apply, protected-path violations). Fix the plan or patch files and re-import. ## 4. Approval — the user's decision Show the user the imported plan (`researchforge experiment show <plan-id> --json`), including worst-case wall time, then ask them to run: ```bash researchforge experiment approve <plan-id> ``` Only use `--yes` if the user explicitly approved this plan in this conversation. Next: the researchforge-run skill executes the funnel. ## Rules - The Python engine is the boundary: never work around a validation error, a protected path, or an approval gate — fix the artifact or ask the user. - Approvals belong to the user: never pass `--yes` or type a confirmation unless the user explicitly approved that step in this conversation. - Ground every summary in stored data: quote only numbers returned by `--json` output or files under `.researchforge/` — never invent metrics.