session-hub-recall

name	session-hub-recall
description	Use when you need to find, inspect, or hand off local Codex, Claude Code, or Gemini session history in this repository's Session Hub tool. Trigger this skill for requests about past agent runs, prior fixes, earlier conversations, continuation context, Claude-to-Codex handoff, or when you need to search local session history before redoing work.

Session Hub Recall

Use Session Hub as the local memory layer for prior agent work.

Prefer the CLI over reading raw session files directly. The CLI already normalizes Codex, Claude, and Gemini history and keeps a persistent local index.

Commands

Build or refresh the local index first when recall quality matters or when the history may have changed:

session-hub index

Search compactly across providers:

session-hub search --query "<terms>" --json
session-hub search --query "<terms>" --provider claude --state active --json
session-hub search --query "<terms>" --refresh-index --json

Inspect one result in more detail:

session-hub show <item-id> --json
session-hub show <item-id> --query "<terms>" --json

Export Claude history into a Codex continuation pack:

session-hub export <claude-item-id> --ownership-confirmed --json
session-hub export <claude-item-id> --ownership-confirmed --handoff-to-codex --json

Workflow

1. If the user asks about previous work, old fixes, prior debugging, or "did we already solve this", run session-hub index unless you know the index is fresh enough.
2. Run session-hub search --json with focused keywords.
3. Read itemId, provider, title, snippet, and updatedAt from search results. Do not paste full raw session paths to the user unless needed.
4. For the most relevant result, run session-hub show <item-id> --json.
5. Use returned snippets as evidence for your answer. Only go to raw files if the snippets are insufficient.
6. If the user wants to continue earlier Claude work in Codex, use session-hub export ... --ownership-confirmed.

Query Guidance

• Start narrow: feature name, error string, file name, tool name, branch name, project name.
• If results are sparse, remove adjectives and extra prose.
• If results are noisy, restrict --provider or --state.
• If a user refers to "the last time" or "most recent", sort your judgment by updatedAt from search results.

Guardrails

• Do not scan raw ~/.codex, ~/.claude, or ~/.gemini trees first if Session Hub can answer it.
• Do not use export for non-Claude items. export is Claude-only.
• Do not run Claude export without --ownership-confirmed.
• Prefer --json for agent use so downstream steps stay structured.
• show is for inspection; search is for discovery. Do not treat one as a substitute for the other.

Output Expectations

When answering the user after recall:

• Say which prior session or provider you found.
• Quote only short, high-signal snippets.
• Distinguish between what the prior session actually did and your current inference.
• If recall was inconclusive, say that and continue with fresh investigation.