| name | rgit-recall |
| description | Use when the user wants to recall, resurrect, bring back, or re-apply a previously captured feature/idea onto today's codebase (e.g. "bring back the re-ranking retrieval step"). |
rgit-recall
Drives the recall → compose → regenerate half of the research-git loop. The stored capsule is a spec; the regenerator rebuilds it onto today's code.
Prerequisites: the repo is rgit init-ed and the research-git MCP server is connected (it exposes recall and compose).
Locating the agent definitions. On Claude Code the plugin runtime resolves agent paths for you. On other CLIs (Codex, Gemini, opencode) this skill is symlinked into ~/.agents/skills/rgit-recall, so resolve the plugin root once and reference the agent from there:
SKILL_REAL=$(realpath ~/.agents/skills/rgit-recall 2>/dev/null || readlink -f ~/.agents/skills/rgit-recall)
PLUGIN_ROOT=$(dirname "$(dirname "$SKILL_REAL")")
The agents/capsule-regenerator.md reference below lives at $PLUGIN_ROOT/agents/capsule-regenerator.md.
Process
1. Recall the capsule(s)
Take the user's natural-language ask and call the MCP tool recall(query). It returns matches, each with its depends_on subgraph. Show the user a short list (name + intent) and confirm which feature(s) to bring back. Default to the top match if unambiguous. If nothing matches, tell the user and stop (suggest list_features to browse).
2. Resolve the full feature set
Include each chosen capsule plus its depends_on dependencies (a feature often needs its prerequisites). Collect the final list of feature_ids.
3. Compose the regeneration brief (against current code)
Call the MCP tool compose(feature_ids). It returns, per feature: intent, knobs, data_assumptions, resurrection_guide, the reference code_slices, the live current_source of each touched symbol, and any conflicts (symbols touched by more than one chosen feature).
4. Dispatch the capsule-regenerator subagent (on subscription)
Dispatch a subagent using the capsule-regenerator agent definition (agents/capsule-regenerator.md). Pass the full brief verbatim plus repo_root. The subagent edits the working tree to re-implement the feature(s) onto today's code, resolves conflicts, sanity-checks syntax, and returns an applied report with provenance (clean vs adapted) per feature.
5. Review + close the loop
Show the user the resulting working-tree diff (git diff) and the subagent's provenance/adaptation notes. Do not commit, run, or freeze for them — the human runs the experiment, and that run is what freezes the reproducible artifact.
Hand them a complete, paste-ready command with the real capsule id filled in — never a template with <placeholders>. If you don't already know their test command from the conversation, ask for it first. Example of the shape (with a real id):
rgit run --from feat_ab12 -- python eval.py --retrieval rerank
That records a new run node, freezes a byte-exact artifact, links a produced edge from the source capsule, and (on approving the resulting proposal) establishes variant_of back to the original. If the subagent returned an updated_resurrection_guide, write it to a file and pass --refresh-guide-file <path> on that same run — placed before the --, since every rgit flag goes before the separator and everything after -- is the experiment command:
rgit run --from feat_ab12 --refresh-guide-file guide.md -- python eval.py --retrieval rerank
Notes
- Reproducibility stays intact. The subagent only authors; the human's
rgit run is the only thing that freezes the artifact — the agent is never in the replay path.
- Sibling flow: capture/segmentation is the
rgit-capture skill (capsule-segmenter agent).
