Skip to content

tobihagemann/turbo / claude/skills/recall-session

recall-session

Verified

Recall what a past Claude Code session did and why by locating its transcript. Use when the user asks to "recall the session", "why did we do X", or "why was this changed".

Recall Session

Locate the Claude Code transcript behind a change, a passage, or another project's session, and recover what that session did and why. Useful for answering reviewer questions, writing post-hoc explanations, or recovering forgotten context.

Inputs

Accept any of:

  • A commit SHA
  • A file path, optionally with a line number (<path>:<line>)
  • A reviewer question plus surrounding context (file and line)
  • A passage with no commit behind it, because the project is not a git repository or the change is uncommitted
  • Another project's session to check, named by its directory

If only a file is given, git blame resolves the commit that last touched the line.

Step 1: Locate the Transcript

Call ${CLAUDE_SKILL_DIR}/scripts/find_transcript.py with either --commit <sha> or --file <path>[:<line>]. Pass --cwd /path/to/repo when searching a different repo than the current working directory.

python3 ${CLAUDE_SKILL_DIR}/scripts/find_transcript.py --file <path>:<line>
python3 ${CLAUDE_SKILL_DIR}/scripts/find_transcript.py --commit <sha>

When there is no commit to resolve, because the project is not a git repository or --file returns no-commit for an uncommitted line, run --phrase with a distinctive stretch of the passage's current text, taken from a single line, instead. It skips git, ranks the project's transcripts by how often they handle the phrase, and returns the same JSON with commit: null.

python3 ${CLAUDE_SKILL_DIR}/scripts/find_transcript.py --phrase "<distinctive text from the passage>"

For another project's session, run --phrase with --cwd <dir>, taking a phrase distinctive to that session, such as a command it ran or a file path it touched.

With --commit or --file, the script:

  1. Resolves the commit via git rev-parse or git blame
  2. Enumerates directories under the effective Claude configuration home's projects/ tree and filters their transcripts by authoritative cwd records, covering path encoding, truncation, collisions, and subdirectory launches
  3. Ranks candidate transcripts whose mtime is within --window-days of the commit (default 14)
  4. Scores candidates by mentions of touched files and tool-use edits on them
  5. Extracts cleaned user prompts and substantive assistant text from the top candidates

The JSON output has status, commit, project_dir (the top candidate's directory when matched), project_dirs (directories with matching cwd records, limited to the time window when there is a commit), and candidates with session_id, score, match_reasons, and excerpts.

Status values:

  • ok — candidates returned
  • no-commit — couldn't resolve a commit
  • no-transcripts — the effective Claude configuration home has no project transcript directories
  • no-match — transcripts exist but none match the touched files in the window, or none mention the phrase

Step 2: Read and Synthesize

If status is ok:

  1. Start with the top candidate (highest score). Read its excerpts first.
  2. If the excerpts already answer the question, stop. If they are thin or ambiguous, read the full transcript at jsonl_path directly for more context.
  3. Ignore candidates with low scores or scores far below the top — they are false positives.
  4. When the answer spans multiple excerpts, quote the most specific one.
  5. In phrase mode, the last match at reason marks the last time the session handled the passage. When the question is about the passage's current state, read the transcript around that point.

Synthesize a concise summary tied to the question being answered:

  • For a change or a passage, lead with the why. The diff already shows the what.
  • For another project's session, report what it did against what the request asks about.
  • Quote the implementer's own words when they already say it well.
  • Keep it to one or two paragraphs. Don't narrate the whole session.

If status is no-match or no-transcripts and the session's date (the commit's, or the one the user gives) is older than the transcript cleanup period (cleanupPeriodDays, 30 days by default), say its transcript was likely pruned. Use AskUserQuestion to ask whether a backup holds an older copy of the Claude configuration home, the directory that holds projects/ (~/.claude unless CLAUDE_CONFIG_DIR is set), taking its path through a freeform answer. When the user names one, re-run Step 1 with CLAUDE_CONFIG_DIR set to that directory and continue from its status.

If status is still anything other than ok, report that no transcript was found. For a change or a passage, fall back to reading the commit diff, when there is one, and the surrounding code, and say so explicitly so it's clear whether the answer still holds up. For another project's session, report the status alone.

Step 3: Output

Return the answer in this shape:

**Commit:** <short-sha> — <subject>
**Transcript:** <session-id> (score <N>)

<one or two paragraphs answering the question, quoting the implementer where useful>

If no transcript was found for a change or a passage:

**Commit:** <short-sha> — <subject>
**Transcript:** none found (<status>)

<fallback explanation derived from reading the commit and current code>

In either shape, without a commit, write the first line as **Commit:** none (not a git repository) or **Commit:** none (uncommitted). For another project's session, write it as **Project:** <the directory passed as --cwd>.

Then use the TaskList tool and proceed to any remaining task.

Rules

  • Treat excerpts as evidence, not ground truth. The implementer's intent at the time may have changed. If the current code contradicts an excerpt, note the discrepancy.
  • Only read the full .jsonl if the excerpts are insufficient. These files are large.
  • Never quote noise prefixes like <command-message> or skill-loading stubs. The script filters these out, but stay alert if reading a transcript directly.
  • If multiple candidates have similar high scores, name both and prefer the one whose time window contains the commit, or, without a commit, the one whose tool calls carry the phrase.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 471kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall