Handoff AI work without carrying the session

Key takeaways

• Part 4 aligns session lifetime with task lifetime and makes work resumable by another subscription without a transcript
• Use a fresh session for a new task, compaction for a long continuation, and an evidence handoff for a provider switch
• Reusing a session can repeatedly carry old conversations and files into an unrelated request
• A useful handoff contains changed files, verification results, confirmed facts, rejected hypotheses, and one next action
• This part produces handoff.md and a transition checklist for Part 5's tool-output budget

Treat session transitions as a state machine

• Move each session through fresh → active → compact or handoff → closed

  • fresh reads Part 3's guidance and the current task contract
  • active accumulates only evidence needed for the task
  • compact applies only while the same goal continues
  • handoff closes a session or changes its owner

• A new goal creates a stronger boundary than the convenience of the current chat

  • An unrelated review request during a bug fix is a new task
  • Analyzing a test failure caused by that fix remains a substep of the same task

• Anthropic says each turn sends prior conversation, project context and read files, plus the new prompt¹

  • It recommends /clear for a task change and /compact for a long continuation
  • This is documented Claude Code behavior; do not assume every product meters identically

Decide between fresh and compact

• Continue only when all three answers are yes

  • The completion condition remains the same
  • The allowed paths substantially overlap
  • Existing session evidence directly supports the next decision

• Compaction reduces a conversation; it does not create a task boundary

  • Compacting between diagnosis and implementation of one bug can be appropriate
  • Start fresh when moving from the finished bug to a new feature

• Gemini CLI exposes maximum retained turns and a compression threshold²

  • model.maxSessionTurns bounds user, model, and tool turns kept in a session
  • model.chatCompression.contextPercentageThreshold controls automatic and manual compression timing
  • An aggressive value can lose evidence, so validate it against the baseline

Transfer evidence instead of conversation

• Use six provider-neutral Markdown fields
  • Goal and completion condition
  • Current commit and changed files
  • Commands run and results
  • Confirmed facts
  • Rejected hypotheses with reasons
  • One next action or unresolved question

# WEB-41 handoff

- Goal: stabilize tied search-result ordering
- State: commit abc123, modified apps/web/lib/tech/search.ts
- Verified: focused test 14/14 passed, typecheck not run
- Facts: score calculation is stable; final sort lacks an ID tie-breaker
- Rejected: locale collation is not involved because IDs differ before localization
- Next: add the ID tie-breaker test, then run the web typecheck

• Avoid both raw transcripts and narrative summaries

  • A transcript forces the next session to re-read failed attempts and outputs
  • Long prose blurs confirmed fact, inference, and next action

• Label inference separately from fact

  • Put executed and observed results under Facts
  • Put possibilities under Open questions so the next agent revalidates them

Test whether the handoff actually recovers work

• A fresh session without the original chat should complete the next verification from the handoff alone

  • Three repeated questions suggest missing essential evidence
  • Broad re-exploration suggests missing paths or confirmed facts

• Measure recovery cost, not handoff brevity by itself

  • Count turns to the first valid change
  • Count duplicate commands and human hints
  • Record changes reverted because of a handoff error

• Frequent handoffs do not always save usage

  • A tiny single-file edit can cost less than documenting its state
  • A reasoning-heavy design task may first need a stronger decision record

Hand output boundaries to the next part

• Part 5 caps the evidence inserted into a handoff

  • Keep only command, count, and duration for passing tests
  • Keep the first error and nearby context for failures
  • Pass a file list and relevant hunks instead of an unbounded diff

• Part 4 is complete when a new provider session resumes without a transcript

  • Test one Codex-to-Claude or Claude-to-Codex transition
  • Keep Part 1's restriction to official subscription sign-in without API keys

Action plan

• When a new goal appears, update handoff.md and open a new session
• Within one task, compact when context grows or the response becomes anchored to old hypotheses
• Standardize the protocol only after five transitions reduce duplicate commands and turns to a valid change