sl-sessions

name	sl-sessions
description	Search and ask questions about Claude Code session history. Use when asking what was worked on, what was tried before, how a problem was investigated across sessions, what happened recently, or any question about past agent sessions. Also use when the user references prior sessions, previous attempts, or past investigations — even without saying 'sessions' explicitly.

/sl-sessions

Search Claude Code session history and synthesize findings about what was worked on, tried, decided, or learned in prior sessions.

Usage

/sl-sessions [question or topic]
/sl-sessions

Pre-resolved context

Git branch (pre-resolved): !git rev-parse --abbrev-ref HEAD 2>/dev/null || true

If the line above resolved to a plain branch name (like feat/my-branch), use it for branch filtering and pass it to the synthesis subagent. If it still contains a backtick command string or is empty, derive the branch at runtime instead.

Repo root (pre-resolved): !git rev-parse --show-toplevel 2>/dev/null || true

If the line above resolved to a path, take its last path component as the repo folder name and use that for session discovery. If it is empty or still contains a backtick command string, derive the repo name at runtime instead.

Note: 2026

The current year is 2026. Use this when interpreting session timestamps.

Guardrails

These rules apply at all times during orchestration and synthesis.

• Never read entire session files into context. Session files can be 1-7MB. Always use the extraction scripts to filter first, then reason over the filtered output.
• Never extract or reproduce tool call inputs/outputs verbatim. Summarize what was attempted and what happened.
• Never include thinking or reasoning block content. Claude Code thinking blocks are internal reasoning and not actionable.
• Never analyze the current session. Its conversation history is already available to the caller.
• Surface technical content, not personal content. Sessions contain everything — credentials, frustration, half-formed opinions. Use judgment about what belongs in a technical summary and what doesn't.
• Fail fast on access errors. If session discovery fails on permissions, report the issue immediately. Do not retry the same operation with different tools or approaches — repeated retries waste tokens without changing the outcome.

Execution

If no question argument is provided, ask what the user wants to know about their session history. Use AskUserQuestion (call ToolSearch with select:AskUserQuestion first if its schema isn't loaded). Fall back to numbered options in chat only if it is unavailable. Never silently skip the question.

Step 1 — Determine scan window

Infer a time range from the user's question. Start narrow; widen only if a narrow scan finds nothing relevant.

Signal	Initial scan window
"today", "this morning"	1 day
"recently", "last few days", "this week", or no time signal	7 days
"last few weeks", "this month"	30 days
"last few months", broad feature history	90 days

Claude Code retains session history for ~30 days by default. Wider windows may find nothing on Claude Code unless the user has extended retention.

Step 2 — Discover sessions and extract metadata

Run the discovery + metadata pipeline (preserving the null-delimited xargs hardening that lets extract-metadata.py run in batch mode):

bash scripts/discover-sessions.sh <repo> <days> | tr '\n' '\0' | xargs -0 python3 scripts/extract-metadata.py --cwd-filter <repo>

Each output line is a JSON object describing a session (platform, file, size, ts, session, branch). The final _meta line carries files_processed and parse_errors.

If the inventory's _meta line shows files_processed: 0, return "no relevant prior sessions" and stop.

If parse_errors > 0, note that some sessions could not be parsed and proceed with what was returned.

Step 3 — Filter and rank

Apply these filters in order to pick the sessions worth deep-diving:

1. Branch filter. Keep sessions where branch == dispatch_branch exactly, or where the branch name contains a keyword from the question's topic (e.g., a question about "auth middleware" matches branches feat/auth-fix, chore/auth-refactor).

2. If the branch filter returned zero sessions:

   • Derive 2-4 keywords from the question's topic. For "a recent crash in the auth middleware where session-validation rejects valid tokens", derive auth,middleware,session,token (or similar).
   • Re-invoke the discovery pipeline with --keyword K1,K2,... appended to the extract-metadata.py invocation. The script returns sessions with non-zero match_count plus per-keyword counts.
   • If files_matched: 0, return "no relevant prior sessions" and stop. Do not extract anything.
   • If files_matched > 0, treat those sessions as candidates. Rank by match_count, break ties by per-keyword counts.

3. Drop sessions outside the scan window. Use last_ts when available, fall back to ts. Discard sessions where both fall before the window start.

4. Exclude the current session — its conversation history is already available to the caller.

5. Apply the deep-dive cap. Take at most 5 sessions total. Narrow by branch-match → match_count → file size > 30KB → recency.

6. Proceed only if at least one session remains after filtering. Otherwise return "no relevant prior sessions" and stop.

Note: gitBranch is captured at the first user message only. A session that began on main and did substantive work on a feature branch via mid-session git checkout records branch: "main". Branch-match returning nothing is not conclusive evidence — that's why the keyword-filter fallback in step 2 is required.

Step 4 — Set up scratch space

Create a per-run throwaway scratch directory:

SCRATCH=$(mktemp -d -t sl-sessions-XXXXXX)

Capture the absolute path; thread it into Step 5 and Step 6. The OS handles cleanup on session end; an explicit rm -rf "$SCRATCH" at the end of Step 7 is harmless and makes intent explicit.

Step 5 — Extract per-session content (file-mediated)

For each selected session, run the skeleton extractor with --output so content writes directly to the scratch file — extraction bytes never round-trip through the orchestrator's tool results:

python3 scripts/extract-skeleton.py --output "$SCRATCH/<session-id>.skeleton.txt" < <session-file>

Stdout receives only a one-line JSON status ({"_meta": true, "wrote": "...", "bytes": N, ...}). Capture bytes and parse_errors from each status line.

Conditional tail-extract — if a skeleton terminates mid-investigation (last visible turn is a tool call with no resolution, or the assistant is mid-debugging without a conclusion), re-extract with a tail shape:

python3 scripts/extract-skeleton.py --output "$SCRATCH/<session-id>.skeleton.tail.txt" < <session-file>

(The skeleton script does not accept a tail:N cap directly; if a tail-only view is needed, post-process the scratch file in shell with tail -n 50 after extraction. Use this only when the head output suggests the session was truncated mid-investigation.)

Conditional errors-mode — for sessions where investigation dead-ends are likely valuable:

python3 scripts/extract-errors.py --output "$SCRATCH/<session-id>.errors.txt" < <session-file>

Use selectively — only when understanding what went wrong adds value.

Step 6 — Dispatch synthesis subagent

Dispatch the sl-session-historian subagent via the Agent tool. Omit the mode parameter so the user's configured permission settings apply. Run on the mid-tier model (e.g., model: "sonnet") — the synthesizer doesn't need frontier reasoning.

The dispatch prompt is the agent's input contract. Pass these fields:

• problem_topic — one sentence naming the concrete question. Lift from the user's argument or, if missing, from the answer to the no-arg prompt.
• scratch_dir — absolute path to $SCRATCH.
• sessions — an array of objects, one per extracted session, each with:
  • path — absolute path to the skeleton file (and optionally errors_path for the errors file when extracted)
  • platform — always claude
  • branch — git branch when present
  • ts and last_ts — session timestamps
  • match_count and keyword_matches — when keyword filtering was used
• output_schema — the structure the agent's response should follow. Default schema:

  Structure your response with these sections (omit any with no findings):
  - What was tried before
  - What didn't work
  - Key decisions
  - Related context
  

  When the caller (e.g., sl-compound) supplies a schema in the skill argument, pass it through verbatim.

Example dispatch shape:

Synthesize findings from these prior sessions:

Problem topic: <one-line topic>

Sessions to read (paths in $SCRATCH):
1. /tmp/sl-sessions-XXXX/abc123.skeleton.txt
   platform=claude branch=feat/auth-fix ts=2026-05-01
2. /tmp/sl-sessions-XXXX/def456.skeleton.txt  errors=/tmp/sl-sessions-XXXX/def456.errors.txt
   platform=claude branch=chore/auth-refactor ts=2026-05-03
...

Output schema:
- What was tried before
- What didn't work
- Key decisions
- Related context

Filter rule: only surface findings directly relevant to this specific problem.
Ignore unrelated work from the same sessions or branches.

The agent reads each path via the platform's native file-read tool and returns prose findings. Bulk extraction content lives only in the agent's subagent context — the orchestrator's working state stays at file paths plus small inventory metadata.

Step 7 — Return findings

Return the synthesizer's output text to the caller verbatim. If discovery or keyword filtering returned zero sessions (Step 2 or Step 3), return the literal string no relevant prior sessions instead.

Optionally clean up scratch:

rm -rf "$SCRATCH"

The OS handles cleanup eventually regardless; the explicit cleanup is for readers who expect it.

Output

When the caller (typically a user typing /sl-sessions, or another skill invoking sl-sessions via the platform's skill-invocation primitive) does not specify an output format, include a brief header noting what was searched:

**Sessions searched**: [count] Claude Code | [date range]

Then the synthesizer's prose findings. When the caller supplies a schema, honor it verbatim and omit the default header.

Time budget

Stop as soon as a complete answer is available. A confident "no relevant prior sessions" within seconds is a complete answer; do not extend the search to fill time. The structural caps in Step 3 (max 5 sessions deep-dived) and Step 5 (conditional tail/errors extraction) bound runtime by construction.

Error handling

If the discovery pipeline fails (e.g., unreadable home directory, permission failure), surface the error to the caller. Do not substitute git log, file listings, or other sources — this skill's contract is session metadata and synthesis.

If extraction --output write fails (disk full, permission), surface a clear error and do not dispatch the synthesizer with partial paths.

If _meta reports parse_errors > 0 from any script, note partial extraction in the dispatch prompt and proceed; the synthesizer flags partial in findings.