Load Session

Load Session

Load a previous Claude Code session as a compact summary (default) or full transcript.

Design: the transcript never flows through main context. The parser writes it to a temp file, the session-summarizer sub-agent reads the file directly, and the file is deleted afterward. Cost is roughly constant regardless of session size.

Phase 1 — Prereqs & browse

Check Python: python3 --version 2>/dev/null || python --version 2>/dev/null. Store the working command as {PY}. If neither works, stop.

If $ARGUMENTS is non-empty, treat it as a session name and skip to Phase 2. Otherwise list sessions:

{PY} "${CLAUDE_SKILL_DIR}/scripts/list-sessions.py" --page 1 --compact

Returns {page, total_pages, total, sessions: [{i, uuid, jsonl_path, name, project, branch, msgs, size, last}]}.

Render as a table via AskUserQuestion:

Select a session (page {page}/{total_pages}, {total} total)

#	Session	Project	Branch	Msgs	Size	Last

Enter a number, N/P for next/prev page, or C to cancel.

On selection, extract that row's uuid and jsonl_path directly — no second resolve call needed. N/P → re-run with updated --page. C → stop.

Phase 2 — Estimate & choose mode

If we arrived from $ARGUMENTS (no jsonl_path in hand), resolve:

{PY} "${CLAUDE_SKILL_DIR}/scripts/resolve-session.py" "<name-or-uuid>"

Stop on error. Refuse to load the current session into itself — if resolved uuid matches ${CLAUDE_SESSION_ID}, stop with "Cannot load the current session into itself."

Estimate cost:

{PY} "${CLAUDE_SKILL_DIR}/scripts/estimate-context.py" "<jsonl_path>"

Returns {file_size_human, user_messages, assistant_messages, estimated_tokens_text, context_pct_full, duration_minutes, budget_warning}.

Ask via AskUserQuestion:

Session: {name} Messages: {user_messages} user + {assistant_messages} assistant | Duration: {duration_minutes} min | Size: {file_size_human}

1. Compact summary (recommended) — ~500–1000 tokens
2. Full transcript — ~{estimated_tokens_text} tokens{warning}
3. Cancel

If budget_warning is true, append ⚠ uses {context_pct_full}% of context window to option 2.

1 → Phase 3a · 2 → Phase 3b · 3 → stop.

Phase 3a — Compact mode

Detect template:

{PY} "${CLAUDE_SKILL_DIR}/scripts/detect-template.py" "<jsonl_path>"

Returns {recommended_template, confidence, reason}.

Pick template via AskUserQuestion (mark recommended_template with "← recommended ({reason})"):

1. Quick — fast single-pass gist
2. Default — Chain-of-Density entity-dense summary
3. Decisions — architectural choices and trade-offs
4. Actions — files changed, what was done, why
5. Discussion — questions, findings, open threads
6. Timeline — chronological phases
7. Handoff — goals/actions/assessment/next steps
8. Custom — provide your own prompt

For 1–7, read ${CLAUDE_SKILL_DIR}/templates/summary-{template_name}.md as {template_text}. For 8, ask for a path or inline text; if it's an existing file path, read it, otherwise use verbatim.

Write transcript to temp file:

{PY} "${CLAUDE_SKILL_DIR}/scripts/parse-session.py" "<jsonl_path>" --out-file auto

Returns {transcript_path, char_count, estimated_tokens, entries}. Do not print the transcript. Capture transcript_path.

Delegate to sub-agent via Agent tool with subagent_type: "bix-session:load-session:summarizer", description: "Summarize session: {name}", and prompt:

transcript_path: {transcript_path}

Read the transcript file and summarize it using the template below. After summarizing, delete the transcript file with `rm -f`.

---
TEMPLATE:

{template_text}

The fully-qualified bix-session:load-session: prefix is required — without it the call silently falls back to general-purpose on the parent model.

On error or empty result, run rm -f "<transcript_path>" yourself and suggest a different template or full mode.

Inject:

---

Loaded session: {name} (compact — {template_name})

{summary_text}

---

Proceed to Phase 4.

Phase 3b — Full mode

Full transcript must reach main context (that's the point), but still write through a temp file so cleanup is uniform:

{PY} "${CLAUDE_SKILL_DIR}/scripts/parse-session.py" "<jsonl_path>" --out-file auto

Read the file with the Read tool, output its contents to the user, then rm -f "<transcript_path>".

Phase 4 — Summary

Session loaded: {name} Mode: compact ({template_name}) | full Estimated tokens added: ~{N}

Rules

• Never load the current session into itself — compare resolved UUID against ${CLAUDE_SESSION_ID}.
• Never print the transcript via Bash — always use --out-file, consumed by the sub-agent (compact) or Read tool (full).
• Always delegate compact summarization to bix-session:load-session:summarizer. Never inline. Never via claude --resume.
• Always clean up the temp file — sub-agent does it in compact mode; skill does it in full mode. SessionStart hook is a safety net for crashes.
• Always show token estimates before full mode and require confirmation.
• Scripts return JSON with an error field — check and surface it.