| name | find-session |
| description | Search past Claude Code AND Codex conversations by keyword and return session IDs for resuming. Use when the user asks to find, locate, or look for a past/previous conversation, or needs a session ID to resume a conversation with claude --resume or codex resume. NOT for searching within a conversation (use /reflect), NOT as a replacement for the resume command itself. |
Find Session
Search session JSONL files for keyword matches and present session IDs. Covers
two sources: Claude Code (~/.claude/projects/<slug>/<uuid>.jsonl) and
Codex (~/.codex/sessions/<Y>/<M>/<D>/rollout-<ts>-<uuid>.jsonl). Each
result is tagged with its source and the correct resume command.
Step 1: Run the search
Run (relative to the skill base directory shown above):
python3 scripts/find_session.py [--all] [--json] [--source claude|codex|both] [--include-subagents] KEYWORDS
- Default: searches the current project's sessions only (fast, no noise), both sources
--all (or -all): searches across all projects and shows [source · project] labels--source claude|codex|both: limit to one source (default both). Shorthand: --claude, --codex, --both--include-subagents: include Codex subagent sessions (guardian auto-review judges, spawned children) — excluded by default as non-resumable noise--json: outputs structured JSON (useful for programmatic follow-up); error paths also return JSON
Replace KEYWORDS with the user's search terms (space-separated). Omit keywords to list the most recent sessions. Always shell-quote arguments that contain special characters.
The current-project scope resolves the project root via git (or cwd). Claude sessions are matched by directory slug; Codex sessions are matched by the cwd recorded in each transcript's session_meta.
Step 2: Check for Haiku summarization
Read config.json and check the # UNTITLED_COUNT=N TOTAL_SHOWN=M hint line on stderr.
- Is
use_haiku_summary true in config.json? If no → skip to Step 3.
- Is
UNTITLED_COUNT >= 2? If no → skip to Step 3.
- Is
TOTAL_SHOWN <= haiku_threshold? If no → skip to Step 3.
- All conditions met → use
AskUserQuestion to ask: "N untitled sessions found. Summarize them with Haiku before showing results?"
- If user agrees:
a. Re-run the script with
--json to get session file paths (and the source of each).
b. Spawn a single subagent (model: read haiku_model from config.json) to summarize each untitled session. For claude sessions, read the first 15 messages of the JSONL. For codex sessions, the first lines are mostly injected boilerplate (session_meta, # AGENTS.md, <environment_context>) — the --json preview field already holds the first real user message, so prefer it (or read past the injected role: user records). Return a one-line summary per session.
c. Present all results as a table: date | source | project | session ID | title or summary | resume command. Use the per-result resume_cmd from --json (claude --resume <id> vs codex resume <id>). Mark Haiku-generated summaries in italics.
- If user declines → go to Step 3.
Step 3: Present results
Show the script output as-is. Each result already prints its own resume line — claude --resume <id> for Claude sessions, codex resume <id> for Codex sessions. The user copies the matching line to resume.
Error handling
If the script prints "No matching sessions found" or "No session directory found":
- Confirm the working directory is inside a Claude Code project.
- Suggest the user try
--all to search across all projects.
- For older sessions that may have been compacted, suggest the user try semantic search with claude-mem if available.
Gotchas
Empty results despite valid keywords — The script resolves the project root via git rev-parse --show-toplevel (falls back to cwd). If run outside a git repo, it uses cwd directly. Confirm the resolved path maps to the right ~/.claude/projects/ slug.
Compacted sessions missing content — Old sessions may have been compacted (summarized). The JSONL may no longer contain the original user messages verbatim, so keyword search will miss them.
Stale titles after /rename — The script uses the latest custom-title in the JSONL, not the first. If a session was renamed multiple times, the most recent name is shown.
Punctuation in search terms — Terms like "claude-mem" are matched as literal substrings. Plain alphanumeric terms use word-boundary matching. Hyphenated terms work as a single phrase, not split into parts.
--all noise — claude-mem observer directories are auto-skipped. Use exclude_slugs in config.json to skip any other noisy project dirs by their exact ~/.claude/projects/ slug name. (Codex subagent sessions are excluded separately via the --include-subagents toggle.)
Codex sessions are always "untitled" — Codex has no /rename, so every Codex result is shown by preview (first real user message), never a title. Expect the Haiku-summarization branch (Step 2) to trigger more often once Codex results are in the mix.
Codex preview skips injected boilerplate — Codex records # AGENTS.md, <environment_context>, <skill>, <turn_aborted>, and <<ccr: blocks as role: user. These are filtered from both the preview and keyword matching, so matches reflect genuine user input — not injected context. amq-squad agent bootstraps ("You are a fresh amq-squad agent…") are kept, as they identify the session's role/workstream.
Codex resume command differs — resume Codex sessions with codex resume <uuid> (UUID = the trailing id in the rollout-… filename, which equals payload.id), not claude --resume. Each result prints the correct command.
--all keyword search reads every transcript — across all projects that is ~12k Claude files plus ~1k Codex files, so expect a few seconds. A ripgrep/grep pre-filter narrows the Codex scan when a query is present (falls back to a full scan if neither tool is installed). The default current-project scope stays fast (~1s) because Codex files are skipped by cwd after reading only their first line.
