Harness Handoff — Deep Context Transfer

Harness Handoff — Deep Context Transfer

---

description: Trigger when user says "harness handoff", "handoff context", "new agent takeover", "full project context", "load all context", "crash recovery", "transfer context", "project overview full". Provides deep context recovery (~8k tokens) for new agent handoff or crash recovery. Do NOT trigger for: "continue writing X function" (just code), "help me understand this file" (single file read), "what is this project" (too vague, use harness guide).

Behavior

Provide deep context recovery for new agent handoff or crash recovery scenarios. Core principle: Index-First, Fragment-on-Demand. For lightweight same-session /​compact recovery, use harness resume (~3k tokens).

---

Difference from Claude Code Default Behavior

Claude Code auto-loads only CLAUDE.md. This skill adds:

• a) task_plan Phase anchor (not just project overview)
• b) progress.md tail (recent activity context)
• c) docs/ INDEX tree (document navigation map)
• d) MEMORY.md project-specific memory
• e) .harness state (project scaffold status)
• f) git log (recent commit history)

---

Read-tool Steps (1-5, deterministic file reads)

Step 1: CLAUDE.md

• Read project CLAUDE.md
• If >300 lines → read tail 300 (recent conventions/​rules are more valuable)

Step 2: task_plan.md Phase Anchor

• Use regex ^#{1,3}\s.*([Pp]hase) to locate current Phase
• Extract ±50 lines around match
• WIP signals: unchecked - [ ], 🚧, WIP, in progress, (doing)
• If no task_plan.md → skip

Step 3: progress.md

• <500 lines → read all
• ≥500 lines → read tail 200

Step 4: docs/ INDEX Tree

• Read docs/​INDEX.md if exists
• Glob docs/​**/​INDEX.md, read up to 10 files
• If none → skip

Step 5: MEMORY.md Best-Match

target = cwd.replace('/', '-').lstrip('-')
candidates = glob('~/​.claude/​projects/​*/​memory/​MEMORY.md')
matches = [c for c in candidates if c.parents[1].name.lstrip('-') == target]

• len == 1 → read it
• len == 0 → skip
• len > 1 → list candidates, ask user to choose

---

Bash-tool Steps (6-7, may fail gracefully)

Step 6: .harness/ Directory Check

test -d .harness && ls .harness/ || echo "No .harness directory"

Step 7: git log

git log --oneline -20 2>/​dev/​null || echo "No git history or not a git repo"

---

Output: 3-Segment Structured Summary

═══ Project Overview ═══
Main goal (from CLAUDE.md Project Overview or task_plan title)
Git status (recent 20 commits / "no git")
.harness status (contents / "none")

═══ Current Work ═══
Current Phase / latest commit / last stopped at / unfinished items

═══ Available Resources ═══
Document structure (docs/ INDEX tree)
Memory categories (MEMORY.md sections)
Harness state (.harness/ if present)

Next: Tell me what you want to do, I'll read specific doc fragments on demand

---

Token Budget

Target 8-12k tokens:

• CLAUDE.md ≤300 lines ≈ 2-3k
• task_plan Phase ±50 lines ≈ 2k
• progress tail 200 lines ≈ 2k
• 10 INDEX files ≈ 2-3k
• MEMORY.md ≈ 1-1.5k
• git log -20 ≈ 0.3k

If exceeded: truncate progress to 100 lines first, then INDEX to 5 files.

---

Empty Rendering

• If a row field is empty → omit the entire row
• If all 3 rows in a segment are empty → omit segment with its title
• If everything is empty → print fallback:

  No project context found. Say harness to initialize a new project.

Graceful Degradation

Any file missing → output "Step skipped (file not found)" and continue. Never error out.