Recover a crashed session
Rebuild the most-recently-active prior Claude Code session for the current folder from its raw on-disk transcript, inject a budgeted working-state digest into THIS session, and offer native claude --resume for a full thread continuation.
When to use
- A session crashed or was killed mid-task, you reopened Claude here, no
/checkpoint was saved, and you don't know the session ID.
- You want to keep working in the session you already have open (digest injection), or get the original thread back natively.
Not this skill: transient API/rate-limit blip mid-turn -> retry. Deliberately saved checkpoint -> /checkpoint resume. Topic search across all past sessions/machines -> /jacked.
Requirements
Needs a current jacked CLI on PATH. Handle these two failure cases explicitly; do not work around them by hand-parsing transcripts:
jacked not found -> tell the user to install or repair jacked, then stop.jacked recover reports No such command 'recover' / exits 2 -> the on-PATH jacked is outdated (it predates this command). Tell the user to upgrade with uv tool install claude-jacked --force && jacked install. In a uv-managed repo checkout, uv run jacked recover ... runs the current source instead.
On-disk source of truth (for self-verification only). Transcripts always survive at ~/.claude/projects/<encoded-cwd>/<session-id>.jsonl, where <encoded-cwd> is the cwd with /, \, and . all replaced by -. If jacked recover finds nothing or a near-empty list, the cause is almost always the wrong folder or a non-default CLAUDE_CONFIG_DIR/CLAUDE_PROJECTS_DIR (an index/path mismatch), not data loss — point the user at that path to confirm the files exist before concluding the session is gone. Do not invent state.
Steps
Find candidates. Run:
jacked recover --json --exclude "$CLAUDE_CODE_SESSION_ID"
Passing $CLAUDE_CODE_SESSION_ID at the shell excludes the session you are running in (it would otherwise rank newest). If this errors with No such command 'recover', your jacked is outdated — see Requirements.
No candidates (count is 0 or chosen is null) -> tell the user no recoverable session was found for this folder (fresh repo, wrong folder, or nothing crashed here) and stop. Do not invent state.
Sanity-check the auto-pick. If chosen.last_prompt is itself a /recover invocation, that is the live session leaking through — drop it and use the next candidate (or re-run with that id added to --exclude).
Present and confirm — before injecting. Show the chosen session and the alternates: ai_title, age, git_branch, and last_prompt. The recommended chosen is the newest session with real substance — if the very newest session is near-empty (almost no messages) it is skipped as the recommendation but still appears in the candidate list, so the user can pick it as an alternate. Ask: "Recover this one, or pick an alternate ()?" Wait for the user. Do not inject until they confirm.
Inject the digest. On confirmation:
jacked recover --session <id> --digest
The output IS the recovered working state — read it. Besides the todos, files, plan, and sub-agent findings, it surfaces two crash-critical facts when present: an In-flight intent block (the crashed agent's final reasoning — why the next step was next) and a Failed actions list (e.g. Bash: pytest -q → FAILED: 3 errors), since the last command/edit that errored is usually the single most load-bearing recovery fact. It ends with a claude --resume <id> line and, if it was trimmed to fit, a budget note.
Scale the digest with --depth {brief|standard|full} (default standard): use --depth full for a heavy multi-file session (more requests, tool actions, files, and a larger char budget), --depth brief for a quick one-liner recovery. --budget N overrides the char budget independently.
Offer native resume. Tell the user: "For a true continuation that preserves Claude's internal state, run claude --resume <id> in a fresh terminal. The digest above lets us continue right here instead."
Re-anchor and continue. Summarize in 1-2 lines: "You were working on X; last step was Y; next was Z." MEMORY.md already carries standing project conventions. Then continue the work.
Manual restart of /goal or /loop
If the recovered digest contains a "Manual restart required" block, the crashed session was driving a /goal or /loop that cannot be auto-resumed (these only run when pasted into a live Claude Code session). Surface that exact command to the user and tell them to paste it into Claude Code themselves to restart it — do not try to run it yourself.
Wrong pick
If the user says it is the wrong session, re-run step 5 with the alternate's id from the candidate list.
Incomplete last turn
When the digest flags "the last turn may be incomplete," treat that work as in-progress, not finished — verify it before building on it.
1---2name: recover3description: Use when a Claude Code session in this folder crashed or was lost mid-work (computer died, terminal closed, Claude broke) and the user reopened Claude without a checkpoint and does not know the session ID — rebuild the last session for THIS folder from its on-disk transcript and continue. Any request to recover, restore, or resume a crashed or dead session qualifies. NOT for transient API/rate-limit errors mid-turn (use retry), NOT for deliberately saved state (use /checkpoint resume), NOT for finding an old session by topic across machines (use /jacked).4---56# Recover a crashed session78Rebuild the most-recently-active prior Claude Code session **for the current folder** from its raw on-disk transcript, inject a budgeted working-state digest into THIS session, and offer native `claude --resume` for a full thread continuation.910## When to use11- A session crashed or was killed mid-task, you reopened Claude here, no `/checkpoint` was saved, and you don't know the session ID.12- You want to keep working in the session you already have open (digest injection), or get the original thread back natively.1314**Not this skill:** transient API/rate-limit blip mid-turn -> `retry`. Deliberately saved checkpoint -> `/checkpoint resume`. Topic search across all past sessions/machines -> `/jacked`.1516## Requirements17Needs a current `jacked` CLI on PATH. Handle these two failure cases explicitly; do not work around them by hand-parsing transcripts:18- **`jacked` not found** -> tell the user to install or repair jacked, then stop.19- **`jacked recover` reports `No such command 'recover'` / exits 2** -> the on-PATH `jacked` is outdated (it predates this command). Tell the user to upgrade with `uv tool install claude-jacked --force && jacked install`. In a uv-managed repo checkout, `uv run jacked recover ...` runs the current source instead.2021**On-disk source of truth (for self-verification only).** Transcripts always survive at `~/.claude/projects/<encoded-cwd>/<session-id>.jsonl`, where `<encoded-cwd>` is the cwd with `/`, `\`, and `.` all replaced by `-`. If `jacked recover` finds nothing or a near-empty list, the cause is almost always the wrong folder or a non-default `CLAUDE_CONFIG_DIR`/`CLAUDE_PROJECTS_DIR` (an index/path mismatch), **not** data loss — point the user at that path to confirm the files exist before concluding the session is gone. Do not invent state.2223## Steps24251. **Find candidates.** Run:26 ```bash27 jacked recover --json --exclude "$CLAUDE_CODE_SESSION_ID"28 ```29 Passing `$CLAUDE_CODE_SESSION_ID` at the shell excludes the session you are running in (it would otherwise rank newest). If this errors with `No such command 'recover'`, your `jacked` is outdated — see Requirements.30312. **No candidates** (`count` is 0 or `chosen` is null) -> tell the user no recoverable session was found for this folder (fresh repo, wrong folder, or nothing crashed here) and stop. Do not invent state.32333. **Sanity-check the auto-pick.** If `chosen.last_prompt` is itself a `/recover` invocation, that is the live session leaking through — drop it and use the next candidate (or re-run with that id added to `--exclude`).34354. **Present and confirm — before injecting.** Show the chosen session and the alternates: `ai_title`, `age`, `git_branch`, and `last_prompt`. The recommended `chosen` is the newest session with real substance — if the very newest session is near-empty (almost no messages) it is skipped as the recommendation but still appears in the candidate list, so the user can pick it as an alternate. Ask: "Recover this one, or pick an alternate (<ids>)?" Wait for the user. Do not inject until they confirm.36375. **Inject the digest.** On confirmation:38 ```bash39 jacked recover --session <id> --digest40 ```41 The output IS the recovered working state — read it. Besides the todos, files, plan, and sub-agent findings, it surfaces two crash-critical facts when present: an **In-flight intent** block (the crashed agent's final reasoning — why the next step was next) and a **Failed actions** list (e.g. `Bash: pytest -q → FAILED: 3 errors`), since the last command/edit that *errored* is usually the single most load-bearing recovery fact. It ends with a `claude --resume <id>` line and, if it was trimmed to fit, a budget note.4243 Scale the digest with `--depth {brief|standard|full}` (default `standard`): use `--depth full` for a heavy multi-file session (more requests, tool actions, files, and a larger char budget), `--depth brief` for a quick one-liner recovery. `--budget N` overrides the char budget independently.44456. **Offer native resume.** Tell the user: "For a true continuation that preserves Claude's internal state, run `claude --resume <id>` in a fresh terminal. The digest above lets us continue right here instead."46477. **Re-anchor and continue.** Summarize in 1-2 lines: "You were working on X; last step was Y; next was Z." `MEMORY.md` already carries standing project conventions. Then continue the work.4849## Manual restart of /goal or /loop50If the recovered digest contains a "Manual restart required" block, the crashed session was driving a `/goal` or `/loop` that **cannot be auto-resumed** (these only run when pasted into a live Claude Code session). Surface that exact command to the user and tell them to paste it into Claude Code themselves to restart it — do not try to run it yourself.5152## Wrong pick53If the user says it is the wrong session, re-run step 5 with the alternate's id from the candidate list.5455## Incomplete last turn56When the digest flags "the last turn may be incomplete," treat that work as in-progress, not finished — verify it before building on it.
jackneil