Session Handoff

Guards:

• Not plan mode. This skill writes files. If in Claude Code plan mode, exit first.
• Not shadowing. NEVER generate handoff-like documents freeform. Freeform summaries look right but lack chain tracking, self-validation, and evidence mining. Only this skill produces handoffs.
• Not when discussing. Only run when the user explicitly asks to CREATE a handoff right now. If ambiguous ("what does handoff do", "edit the handoff file"), ask first.

Typical use: ~75% context. You have a lot of conversation to mine — extract maximum value before closing. On 1M context that's ~750K tokens of history.

The user should not need to provide anything — /handoff alone is sufficient.

Arguments: $ARGUMENTS

Compact mode

If $ARGUMENTS contains the word compact-mode, the compact-handoff mod started this run. It replaces the conversation with this handoff when your turn ends, so the turn must finish with nothing pending:

• Run Steps 1-6 exactly as normal. Anything else in $ARGUMENTS is the framing hint.
• Step 7: one line only (file path, line count, chain seq). The compaction replaces this turn seconds later, so a full report is wasted output.
• Ask nothing in this turn. Where a step says to ask (e.g. Tier B "Any doubt"), take its stated default and note the choice in the handoff.
• Step 8: don't ask. Follow the Yes flow in references/close-session.md (append ## Session Closed, commit, print the paste prompt), skipping its step 4.
• End your turn with this exact last line, path relative to the repo root: COMPACT-HANDOFF-READY: {handoff path}
• If no handoff file could be written, end with COMPACT-HANDOFF-FAILED: {reason} instead. The mod then leaves the context untouched.

---

Agent Strategy

Parallelize independent research. Launch in one message.

What	Mode	Why
Step 1A (git/beads/ls)	Parallel Bash, never agents	Cheap commands; agent bootup wastes 15K+ each
Step 1B context agents (OV, stale-refs, bible)	Parallel Bash inline; agents only if parent handoff >500 lines	Independent research
Step 1C (conversation mining)	Main agent only	Only you have the history
Steps 5+6 (beads/memory writes)	Parallel Bash	Independent writes

---

Step 1: Deep Context Gathering

1A: External State (parallel Bash — never agents)

Run in one message as inline Bash calls:

Commands	Returns
git log --oneline -20, git diff --stat, git status -s | head -30, git branch --show-current	Branch, recent commits, uncommitted changes
bd list --status=in_progress, bd list --status=open --priority=0,1, bd stats (skip if bd unavailable)	Active/open beads
ls plans/handoffs/, ls .claude/handoffs/, ls plans/*.md	Existing handoff files

1B: Chain Detection

Resolve the chain tag (use first that applies):

1. Epic exists → use epic name/ID
2. 1-4 beads → use all bead IDs (e.g., myproject-xxxx, myproject-yyyy)
3. 5+ beads → pick 2-3 most relevant to the primary work stream
4. No beads/epic → generate fallback: python -c "import secrets; print(secrets.token_hex(4))" → standalone-{hex}

Find prior handoff in this chain (two tiers, stop at first match):

• Tier A — Paste Prompt (deterministic). Did the user start this session by pasting something like Read HANDOFF_foo_date.md (seq 2, chain-x) and continue...? If yes, that file is the parent. Read its header. Continuation — seq = parent's + 1.

• Tier B — Bead/Epic Scan (heuristic, skips auto-handoffs).

  grep -l "Chain:.*{chain_tag}" plans/handoffs/HANDOFF_*.md 2>/dev/null \
    | xargs grep -L "^\*\*Auto:\*\* true" 2>/dev/null
  

  A shared bead is a CANDIDATE, not proof of continuation. Before claiming the match as parent:

  1. Read the candidate's ## Where We're Going section.
  2. Is current session work a direct follow-on of those steps? (Same feature/fix, continuing the named next-actions?)
  3. Clear continuation → inherit chain, increment seq, set parent.
  4. Unclear or unrelated → treat this as seq 1 (new chain). Add a ## Related Handoffs section listing the sibling file as reference only, NOT parent. A bead can host many independent work streams (brainstorm → impl → testing) — don't conflate them.
  5. Any doubt → ask the user: "Found prior handoff {file} on same bead. Is this session a continuation? (default: new chain)"

Neither tier matches: seq 1, parent: none.

1B-3/4: Context Agents (parallel, inline Bash unless parent is huge)

Once chain tag resolved, launch in parallel:

Task	Returns
OV Recall (if available): /memory-recall with 2-3 keyword searches	Prior decisions, failed approaches
Parent Context (if parent exists): READ FULL PARENT — extract Goal, Where We Are, Key Decisions, What We Tried, Where We're Going, Open Questions, code identifiers	Parent summary for "Since Last Handoff" + identifier list
Reference Docs: ls plans/*BIBLE* plans/*bible* *BIBLE* CLAUDE.md .claude/CLAUDE.md and read if found	Project context
Stale Refs (if parent): Grep each parent identifier against current codebase	List of identifiers NOT found

Parent reading is MANDATORY when a parent exists. "Since Last Handoff" requires comparing what was planned vs what happened. Skip agents that don't apply.

1C: Conversation Mining

If arguments were provided ($ARGUMENTS), use as a soft hint for framing. Conversation is ground truth.

Choose mining pass and announce it (this is not optional):

Pass	When	Strategy
Quick	<100K context tokens	Single pass with extraction checklist below
Deep	100K-500K context tokens, OR 1M context + 20+ tool calls	Two passes — read references/mining-deep-chunked.md
Chunked	500K+ context tokens, OR 1M context + 50+ tool calls or >1hr work	Map-reduce — read references/mining-deep-chunked.md

Write: "Mining with {Quick/Deep/Chunked} pass ({reason})." before starting. Don't use numeric names — they collide with chain seq numbers.

For Deep or Chunked, read references/mining-deep-chunked.md NOW for the multi-pass protocol. Don't do map-reduce from memory.

Extraction checklist (apply per pass):

• Goals & objectives (user's target, overarching epic)
• Work completed (every file modified, function changed, with specifics)
• Approaches tried (chronological, successful and failed)
• Failed approaches + why (MOST expensive to re-discover)
• Test results & measurements (raw numbers)
• Data files created (paths to JSON/CSV/logs)
• Decisions made + rejected alternatives
• Discoveries & gotchas
• Code analysis (signatures, thresholds, constants)
• User preferences expressed
• Remaining questions
• Dependencies on other work

If you're skimming, STOP. Re-read. Details are the value.

---

Step 2: Choose Output Location

First directory that exists (or create plans/handoffs/):

1. plans/handoffs/
2. .claude/handoffs/

Step 3: Generate File Name

• With beads: HANDOFF_{chain_tag}_{slug}_{YYYY-MM-DD}.md (e.g., HANDOFF_myproject-xxxx_auth-rewrite_2026-03-19.md). For multi-bead, use primary bead only.
• No beads: HANDOFF_{slug}_{YYYY-MM-DD}.md
• Slug: 2-4 word kebab-case.
• Collision: append _2, _3, etc.

Step 4: Write the Handoff File

Read references/output-template.md for the full file structure. It defines every section (Goal, Where We Are, What We Tried, Evidence & Data, etc.) with composition rules.

Line Budget

Check your system prompt for context window size.

	Standard (200K)	Extended (1M)
Target (aim for ceiling)	300-400 lines	500-800 lines
Hard minimum	150 lines	250 lines
Light session min	80 lines	120 lines
Split threshold	400 lines	800 lines

Target the CEILING. An 800-line handoff on 1M context is ~0.7% of the window — negligible cost, huge savings. Too-long is cheap; too-short costs hours of re-investigation.

Two-Phase Write

1. Phase 1 — Initial Write. Compose and write everything in ONE Write call. All sections. Phase 1 MUST hit the pass minimum on its own — Quick: 150+ standard / 250+ extended, Deep: 300+, Chunked: 500+. Phase 1 is not a rough draft; it's the baseline. If you find yourself planning to "flesh it out in Phase 2," stop and expand the current sections before writing.
2. Phase 2 — Gap Research. After writing, count lines. Read your file back. Scan conversation for data you didn't capture (tables skipped, mid-session feedback missed, measurements without numbers, approaches mentioned but not detailed). Use Edit to append toward the ceiling. Phase 2 is for gaps, not for baseline. If Phase 2 needs to add 50+ lines to reach the minimum, Phase 1 was under-mined.

Phase 2 is MANDATORY for Deep and Chunked passes. Optional for Quick — but run it if your first pass is below the ceiling by >20%.

Splitting: only if final file exceeds threshold. Under threshold = one file.

---

Step 4-CHECK: Self-Validation

Read references/validation.md and run every check. If any fails, expand thin sections before proceeding.

---

Steps 5 + 6: Update Beads & Persist Memory (parallel Bash)

# Beads (if in_progress work exists)
bd update {id} --notes "Handoff written. See {file_path}"

# Memory (if bd remember available)
bd remember "Handoff: {path}. Chain: {chain_tag} seq {N}. Status: {status}. Next: {next action}"

Step 7: Report

Tell the user concisely:

• File path(s) and line count(s)
• Chain info (tag, seq, new vs continuation)
• Self-check outcome (and what was expanded if first pass failed)
• The Next Action

Step 8: Ask to Close Session

Ask:

Handoff complete. Ready to close this session?

• Yes — I'll commit, mark "session closed", give you a paste prompt for the next session.
• No — We keep working. Say "close session" when done.

(Defaults to commit — say "close without commit" to skip.)

Based on the answer, read references/close-session.md and follow the flow for Yes / No / "close without commit".

---

Cleanup: Archiving Completed Chains

When bead/epic is closed and all work is done:

# Find all handoffs for a bead or epic
grep -l 'Chain:.*{bead_id_or_epic}' plans/handoffs/HANDOFF_*.md plans/handoffs/PLAN_*.md 2>/dev/null

# Archive (don't delete — old decisions are useful)
mkdir -p plans/handoffs/archive/
mv {files} plans/handoffs/archive/