liuyixin-louis/agentroom/searching-agent-sessions

Searching Agent Sessions — Execution Rules

Goal & Scope

Help the user quickly find past coding agent sessions by topic, project, agent, or time. Uses CASS CLI as the search backend. Returns concise results with ready-to-copy resume commands.

CASS Binary

The skill expects cass to be on PATH. If not found, check these locations:

# Installed via install-cass.sh (added to shell profile)
which cass
 
# Or use the binary directly from the build
CASS="$(git rev-parse --show-toplevel 2>/dev/null)/search-backend/cass/target/release/cass"

If CASS is not installed, direct the user to run:

./scripts/install-cass.sh && source ~/.zshrc
cass index --full

Core Commands

1. Search by keyword/topic

cass search "<query>" --json --limit 20 --highlight

Optional filters:

• --agent <agent> — filter by agent: claude-code, codex, gemini
• --days <N> — limit to last N days
• --workspace <path> — filter by workspace/project path
• --mode semantic — use semantic search (better for conceptual queries)
• --mode hybrid — combine lexical + semantic

2. Recent sessions / timeline

cass timeline --json --group-by none --since 7d
cass timeline --json --group-by none --agent gemini --since 30d

3. Export a session transcript

cass export "<source_path>" --format markdown
cass export "<source_path>" --format json

4. Find related sessions

cass context "<source_path>" --json --limit 5

5. Session analytics

cass analytics tokens --days 7 --group-by day --json
cass analytics tools --days 7 --json

Procedure

1. Parse the user's intent — extract: search terms, agent filter, time range, project/workspace hint.

1. Choose search strategy:

• Specific keywords → cass search "<query>" --json --limit 20
• Conceptual/vague query → cass search "<query>" --mode hybrid --json --limit 20
• "Recent sessions" / "what did I do today" → cass timeline --json --group-by none --since <range>
• "Sessions about project X" → cass search "<project>" --json or --workspace filter if path known

1. Run the command and parse JSON output.

1. Deduplicate to parent sessions — multiple hits often come from subagent files under the same parent session. Extract the parent UUID from each source_path (for subagent paths like <uuid>/subagents/agent-*.jsonl, the parent is <uuid>.jsonl). Group by parent UUID and show each unique parent session once.

1. Resolve workspace from parent session — for each unique parent session, read the cwd from the parent JSONL file (see Workspace Resolution). Do NOT use the workspace field from CASS search hits — it may come from subagents and point to a subdirectory.

1. Present results as a compact list per session (see Output Format below).

1. Always include resume command — for every session result, provide a ready-to-copy command block with cd <cwd> from step 5.

Workspace Resolution

CRITICAL: Every resume command MUST include cd <workspace> && before the resume command. The user opens a fresh terminal and pastes the command — without cd the agent resumes in the wrong directory and loses project context.

WARNING: Do NOT trust the workspace field from CASS search hits. CASS hits often come from subagent files whose workspace points to a subdirectory (e.g., project/src-tauri), not the directory where claude was originally invoked. Using the wrong directory means --resume will fail with "No conversation found" because Claude Code maps sessions to project slugs derived from the launch cwd.

After getting search results, always resolve workspace by reading the parent session file directly.

Claude Code sessions

Path pattern: ~/.claude/projects/<project-slug>/<uuid>.jsonl or .../<uuid>/subagents/agent-*.jsonl

1. Deduplicate to parent session first. For subagent files (<uuid>/subagents/agent-xxx.jsonl), extract the parent UUID from the path and use <project-slug>/<uuid>.jsonl as the main session file.
2. Read the `cwd` from the parent session JSONL (not subagent files):

``bash head -5 "<parent_session_file>.jsonl" | python3 -c " import sys, json for line in sys.stdin: try: d = json.loads(line.strip()) cwd = d.get('cwd') or d.get('workspace') if cwd: print(cwd) break except: pass " ``

1. Use the `cwd` value as the `cd` target in the resume command. This is the directory where claude was launched, and --resume only works from a directory that maps to the same project slug.
2. Cross-check with project slug: The <project-slug> in the path is the cwd with / replaced by -. If cwd extraction fails, reverse the slug: strip leading -, replace remaining - with /, prepend / — that's the workspace.

Codex sessions

Path pattern: ~/.codex/sessions/YYYY/MM/DD/rollout-...-<uuid>.jsonl

1. Read the first line (type=session_meta), workspace is in payload.cwd:

``bash head -1 "<session_file>" | python3 -c " import sys, json d = json.loads(sys.stdin.readline()) print(d.get('payload', {}).get('cwd', '')) " ``

Gemini sessions

Path pattern: ~/.gemini/tmp/<project-dir>/chats/session-*.json

The <project-dir> can be either:

• A human-readable project name (e.g., my-project)
• A SHA256 hash of the real project directory (legacy format)

To resolve workspace:

1. Named folders — if the folder name is human-readable (not a 64-char hex string), it's the project name. Check common project locations:

• ~/Projects/<name> or ~/<name>
• Read the session JSON's projectHash field, then check ~/.gemini/trustedFolders.json — keys are real folder paths; SHA256-hash each key and compare against projectHash.

1. SHA256 hash folders — read ~/.gemini/trustedFolders.json — keys are real folder paths. SHA256-hash each key and compare against the hash in the session path.

Resume Command Construction

Extract session ID (UUID) from source_path using regex: [a-f0-9]{8}-[a-f0-9]{4}-[a-f0-9]{4}-[a-f0-9]{4}-[a-f0-9]{12}

Gemini UUID — IMPORTANT

Gemini session filenames only contain a short 8-char ID (e.g., session-2026-03-03T03-16-90638149.json), but gemini --resume requires the full UUID. You MUST read the session JSON and extract the sessionId field:

python3 -c "
import json
with open('<session_file>') as f:
    data = json.load(f)
print(data.get('sessionId', ''))
"

Do NOT use the short ID from the filename — it will fail with "Invalid session identifier".

Resume commands by agent:

• Claude Code: claude --resume <uuid> --dangerously-skip-permissions
• Codex: codex resume <uuid> (NOTE: user's shell may alias codex to already include --dangerously-bypass-approvals-and-sandbox — do NOT add the flag by default to avoid doubling)
• Gemini: gemini --resume <full-uuid> --yolo (MUST be full UUID from sessionId field, NOT the short ID from filename)

Output Format — CRITICAL

For each unique session found, output in this exact format:

### Session N — <Agent> — <relative time>
 
**Topic:** <title or first meaningful snippet, ~1-2 lines>
 
**Resume:**

cd <workspace> && <resume_command>

Rules:

• Always include the resume command block with cd <workspace> && prepended when workspace is known. If workspace cannot be resolved, just show the resume command alone.
• The user will copy-paste these commands into their own terminal. Do NOT offer to run them. Do NOT use osascript.
• Group hits by unique session — do NOT show the same session file multiple times.
• Show top 5-8 unique sessions max.
• Keep topic/title concise — one or two lines, not a wall of text.
• Show the source path abbreviated (last 2-3 components) on a separate line if useful.
• If session ID cannot be extracted (rare), note "resume unavailable" instead of the command block.

Example output

### Session 1 — Gemini — Feb 18
 
**Topic:** Headless agent infra — cron scheduling, auth profiles, CLI runner architecture
 
**Resume:**

cd ~/my-project && gemini --resume 6737dd0a-1b2c-4d5e-8f9a-0b1c2d3e4f5a --yolo

 
---
 
### Session 2 — Claude Code — Feb 17
 
**Topic:** Desktop app Tauri setup and CASS integration
 
**Resume:**

cd ~/Projects/my-app && claude --resume 513d51d0-033e-4583-afe4-7eb448e8a3b2 --dangerously-skip-permissions

Quality Checklist

• [ ] Used --json flag for machine-parseable output
• [ ] Deduplicated by source_path — each session appears once
• [ ] Every session has a resume command block (or explicit "unavailable" note)
• [ ] Workspace resolved when possible (especially for Gemini hash paths)
• [ ] Did not dump raw JSON to the user
• [ ] Did not offer to "run" or "execute" — user copies commands themselves
• [ ] Respected agent/time/workspace filters from user query