session-recovery

name

session-recovery

description

Recover lost agent session content and file changes from on-disk conversation logs. Streaming and OOM-safe on 700MB+ daily JSONL. Two commands: search.py for keyword search across recent sessions (with hit snippets, file-op listing, JSON for agent consumption); extract.py for pulling full write/edit content from a single session by ID prefix, with optional replay-rebuild for pure-edit sequences and safe restore-to-disk (refuses silent overwrites without --yes). Multi-agent aware via --agent main|all|a,b. Configurable data root via --root flag or SESSION_RECOVERY_ROOT env var (default ~/.openclaw/agents/). Trigger when user wants to find lost session content, recover files written by an agent, locate which session modified a file, search session history by keyword, or rebuild a file from an edit replay. Also triggers on 找回会话, 会话被覆盖, 历史会话搜索, 文件被删了, session 丢了, 找回某个文件, 重放编辑.

session-recovery

• Version: 1.0.1
• License: MIT
• Author: Evan Song · github.com/Songhonglei
• Repository: https://github.com/Songhonglei/better-agent-skills

Recover lost agent session content and file changes from on-disk conversation logs.

从本地会话日志中找回丢失的 agent 对话和文件改动。

Built for OpenClaw's on-disk session format (~/.openclaw/agents/<agent>/sessions/*.jsonl). Streaming and OOM-safe for production-scale environments (5000+ sessions, 700MB+ daily JSONL).

Compatibility

Platform	Status
OpenClaw	✅ Full (primary target)
Other agent frameworks (CC / Codex / Hermes)	🛣️ Roadmap
Linux	✅ Full
macOS	✅ Full
Windows (WSL)	✅ Full
Python	≥ 3.8 (uses pathlib, datetime.fromisoformat)

Quick start

# Search recent sessions for a keyword
python3 scripts/search.py "index.html" --days 2

# Search by specific date, also list write/edit ops
python3 scripts/search.py "skill logo" --date 2026-06-13 --extract-files

# Search all agents
python3 scripts/search.py "anything" --agent all

# Filter by real msg timestamp (slower, more accurate when mtime lies)
python3 scripts/search.py "stuff" --days 7 --by-content-time

# JSON output for agent consumption
python3 scripts/search.py "stuff" --json

# Once you know the session ID, extract file ops
python3 scripts/extract.py 21f68359 --file-filter index.html

# Show full content (warning: can be large)
python3 scripts/extract.py 21f68359 --file-filter index.html --show-content

# Restore to original path (non-interactive needs --yes)
python3 scripts/extract.py 21f68359 --file-filter index.html --restore --yes

# Restore to a specific path
python3 scripts/extract.py 21f68359 --restore-to ~/out/index.html --yes

# Pure-edit session (no write baseline): replay edits to rebuild
python3 scripts/extract.py 21f68359 --file-filter app.py --rebuild \
    --restore-to ~/out/app.py --yes

Custom data root

If your OpenClaw install puts agents elsewhere, override:

# Per-call
python3 scripts/search.py "foo" --root /custom/openclaw/agents

# Or environment variable
export SESSION_RECOVERY_ROOT=/custom/openclaw/agents
python3 scripts/search.py "foo"

Priority: --root > SESSION_RECOVERY_ROOT env > ~/.openclaw/agents/.

Data sources

Primary: JSONL conversations at <root>/<agent>/sessions/*.jsonl (always present). Contains full tool calls, including write content and edit deltas.

Secondary (OpenClaw-specific): QMD archive at <root>/<agent>/qmd/sessions/*.md (may not exist on all installs). Plain-text dialogue; survives session reset but does NOT contain full code — use for finding which session, then extract from JSONL.

Recovery priority: non-reset JSONL > reset+QMD > QMD only > reset only.

Safety

• Restore refuses silent overwrites in non-TTY environments. You must pass --yes to confirm in agent / CI / script contexts.
• Target-already-exists is warned even with --yes (writes are destructive).
• --show-content prints full file content which may be huge — prefer --restore-to for large files, or preview first.
• --max-files N is an OOM guard (default 300). When hit, output explicitly flags "may have missed results"; narrow the time window or raise the limit.

Notes

• Session IDs support prefix match (first 8 hex chars usually enough).
• Time filter defaults to file mtime, which can lie if a session was last modified hours after its content was written. Use --by-content-time for accurate (slower) filtering by real message timestamps.
• --limit is per agent, per source. With --agent all total results multiply by agent count and source count.
• Field-name compatibility built in: write/edit ops accept any of file_path|path, content|new_string|newText, oldText|old_string, arguments|input|parameters.

License

MIT (see LICENSE)