| name | conversation-history-sparse |
| description | Read ~/.claude/projects and ~/.codex/sessions sparingly to infer what the current directory/worktree/branch was doing, with strict sampling limits. |
| type | analysis |
| scope | project |
Conversation History Sparse
Use this skill when the user asks to recover recent context from conversation history without consuming much context window.
Canonical Workflow
Use the repo-local Claude skill as the primary procedure and command cookbook:
.claude/skills/conversation-history-sparse.md
This Codex skill is intentionally thin and delegates operational details there to avoid duplicate instructions.
Codex-Specific Rules
- Keep extraction sparse (metadata + tiny prompt samples only).
- Prefer
multi_tool_use.parallel for independent metadata reads.
- Never dump full JSONL lines if they contain large payloads.
- Summarize evidence in 4 sections:
- Branch/PR
- Claude sparse history
- Codex sparse history
- Inference
Trigger Cues
- "what was this worktree doing"
- "read convo history sparingly"
- "inspect ~/.codex/sessions and ~/.claude/projects"
- "recover session context without using much tokens"