claude-archive-search

--- name: claude-archive-search description: Use when searching your own past conversations, memory, subagent runs, or tool output stored under ~/.claude/, or when recovering files lost from a prior Write tool call. --- # Claude Archive Search Where everything under `~/.claude/` lives, and how to grep it fast. ## Layout ``` ~/.claude/ CLAUDE.md user-global instructions settings.json settings projects// per-project archive (slug = cwd with / → -) .jsonl main conversation transcript (user+assistant msgs) /subagents/agent-.jsonl subagent conversation transcript /subagents/agent-.meta.json {agentType, description} /tool-results/.txt externalized large tool output (>~2MB) memory/MEMORY.md auto-memory index (always in context) memory/*.md auto-memory entries transcripts/ses_*.jsonl cross-project / desktop session transcripts sessions/.json pid → sessionId + cwd mapping session-env// per-session env locks plans/*.md saved plans agents/*.md agent definitions commands/ slash-command definitions hooks/ user hook scripts skills//SKILL.md user skills plugins/cache/// installed plugin skills + agents ``` Slug for this repo: `-Users-emmahyde-projects-sector` (prefix dash + `/`→`-`). ## Transcript JSONL format Each line is one event. Useful fields on message events: - `sessionId`, `cwd`, `gitBranch`, `timestamp` - `isSidechain: true` ⇒ subagent message - `message.content[]` — array of `{type: text|tool_use|tool_result, ...}` - `tool_use`: `{id, name, input}` — Write keeps full `input.content` inline - `tool_result`: `{tool_use_id, content}` — large outputs replaced by `…saved to: ` pointing into `tool-results/` ## Find things fast | Goal | Command | |---|---| | Project slug from cwd | `echo "-${PWD//\//-}"` | | All transcripts for this project | `ls ~/.claude/projects/"$(echo "-${PWD//\//-}")"/*.jsonl` | | Recent transcripts (10 newest) | `ls -t ~/.claude/projects//*.jsonl \| head -10` | | Full-text search conversation | `rg -l 'PATTERN' ~/.claude/projects//` | | Search memory | `rg 'PATTERN' ~/.claude/projects//memory/` | | Search across ALL projects | `rg 'PATTERN' ~/.claude/projects/` | | Subagent prompts only | `rg '"isSidechain":true.*"type":"user"' ~/.claude/projects//*/subagents/*.jsonl` | | Subagent metadata | `cat ~/.claude/projects///subagents/*.meta.json` | | Map session → cwd | `cat ~/.claude/sessions/*.json \| jq '{sessionId, cwd}'` | | User prompts in a session | `jq -r 'select(.type=="user" and .message.role=="user") \| .message.content' < FILE.jsonl` | | Tool calls of type X | `jq -r 'select(.message.content[]?.type=="tool_use" and .message.content[]?.name=="Write") \| .message.content[]?.input.file_path' < FILE.jsonl` | ## Structured extraction from JSONL ```bash # All Write calls (path + size) from one transcript jq -r '.message.content[]? | select(.type=="tool_use" and .name=="Write") | "\(.input.file_path)\t\(.input.content|length)"' FILE.jsonl # All text assistant responses jq -r 'select(.type=="assistant") | .message.content[]? | select(.type=="text") | .text' FILE.jsonl # Subagent type + task for a session for m in ~/.claude/projects///subagents/*.meta.json; do jq -r '[.agentType, .description] | @tsv' "$m" done ``` ## Recover lost Write output Use `recover-writes.py` in this skill dir. It extracts every `Write` tool call from transcripts and reconstructs the files (or a markdown report). ```bash # Report every Write across all transcripts in current project (stdout markdown) ~/.claude/skills/claude-archive-search/recover-writes.py # Filter to a path substring, write a markdown report ~/.claude/skills/claude-archive-search/recover-writes.py --path Sector.Engine/Crew -o writes.md # Restore latest version of each written file to disk under ./recovered/ ~/.claude/skills/claude-archive-search/recover-writes.py --restore --out recovered/ # Scope to one transcript or UUID session ~/.claude/skills/claude-archive-search/recover-writes.py --session 090e7f30-f644-4254-a301-b8b0d685c9ed ~/.claude/skills/claude-archive-search/recover-writes.py --transcript ~/.claude/projects//.jsonl # Include subagent writes too ~/.claude/skills/claude-archive-search/recover-writes.py --include-subagents ``` Run `--help` for all flags. Defaults: current project slug inferred from `$PWD`, main transcripts only, markdown to stdout, latest write per path wins. ## Notes / gotchas - `projects//*.jsonl` and `projects///subagents/*.jsonl` are **separate files**; a full session ≈ main + all matching subagents. - `transcripts/ses_*.jsonl` is a distinct tree (desktop / MCP sessions) with a different record shape — not the same as project transcripts. - Write `input.content` is always inline in the jsonl; Edit tool stores `old_string`/`new_string` only. To reconstruct post-Edit state you'd need the pre-Edit Read plus ordered Edit replays — `recover-writes.py` does not do this. - Large tool *results* (not Writes) get externalized to `tool-results/*.txt` with a `` stub in the jsonl referencing the path — grep the text files directly for content. - `memory/MEMORY.md` is injected into every conversation; individual `memory/*.md` files are not.