Context Handover

Context Handover

Saves session state so the next session can resume without information loss.

State files

.harness/state.json — session lifecycle (read/write):

{
  "version": "1.0",
  "session": { "status": "idle", "started_at": "ISO 8601", "last_session": "ISO 8601" },
  "position": { "phase": "01-discuss", "active_task": "...", "resume_file": "...", "stopped_at": "..." }
}

.harness/phases/XX-name/.continue-here.json — resume context (write):

{
  "version": "1.0",
  "phase": "01-discuss",
  "task": 1,
  "total_tasks": 3,
  "status": "in_progress",
  "last_updated": "ISO 8601",
  "current_state": "...",
  "completed_work": ["..."],
  "remaining_work": ["..."],
  "decisions_made": ["..."],
  "blockers": [],
  "context": "...",
  "next_action": "..."
}

When to trigger

• Proactive: After each tool call, check token usage. If ≥70% used and remaining tokens are insufficient to complete Review + Compound for the current phase — trigger immediately.
• Safety net: ≥80% total usage — if proactive trigger was missed.

Do NOT invoke /compact programmatically. Instruct the user to type it.

Phase detection (priority order)

1. .harness/state.json → position.phase — use if present
2. Active GitHub issue labels → phase:discuss / plan / execute / verify
3. Issue title/body keywords — "design"/"spec"/"ADR" → discuss; "PRD"/"story" → plan; "implement"/"build"/"fix" → execute; "test"/"QA" → verify
4. Default to execute — note "phase inferred by default" in handoff doc

Execution sequence

• 1. Invoke memory system — check .claude/settings.json hooks first:

  • If the active memory system has a Stop hook configured — it fires automatically on session end, no action needed.
  • Otherwise — call the memory system's update mechanism (e.g. mem0, or append to MEMORY.md) manually. It is a black box — trigger the update and move on.
  • If no memory system is detected — write key decisions into the decisions_made array of .continue-here.json (fallback).
  • Budget: <5% of remaining context.

• 2. Write .harness/state.json — atomic read → mutate → write:

  • Set session.status: "idle"
  • Set session.last_session: <current ISO timestamp>
  • Set position.stopped_at: <one-line description of last action>
  • Set position.resume_file: <path to .continue-here.json>
  • Budget: <1% of remaining context.

• 3. Write .harness/phases/XX-name/.continue-here.json — derive XX-name from position.phase. Rules:

  • All string values: factual, no padding. completed_work and remaining_work are arrays of strings.
  • Reference artifacts by path only — never inline content.
  • next_action: one concrete sentence starting with a verb.
  • Budget: <5% of remaining context.

• 4. Post GitHub handover comment — only if docs/agents/issue-tracker.md exists AND active task has a GitHub issue number. Format:

  ## Handover — YYYY-MM-DD HH:mm
  
  **Phase:** [current phase]
  **Session summary:** [1–3 sentences]
  **Next step:** [specific pick-up point]
  
  _[N] of ~[effort_estimate] context window(s) used so far (token budget: 1 = ~150K–200K)._
  
  ---
  _🤖 Posted by `/context-handover` (AI-generated)_
  

• 5. Output to user:

  • "Handover complete. Resume file: .harness/phases/XX-name/.continue-here.json."
  • "Start your next session with /session-start."
  • "To compact this session now, type /compact."

Graceful degradation

See phase-budgets.md for per-phase handoff content and session budget tables.