Context Handoff

Session continuity for coding agents. Write compressed state at session end, read it at session start. The next session skips re-discovery and starts working immediately.

Why this matters: LLMs are stateless. Every new session starts from zero. Without a handoff, agents waste 30-50% of their context window just rediscovering what the last session already knew — files, decisions, blockers. This skill solves that by compressing session state into a lightweight file that the next session reads in seconds.

Session Start Behavior

If

.context/handoff.md

exists (injected via the plugin's SessionStart hook or read manually), use it as your primary orientation:

1. Read the handoff file.
2. Note the branch, what was done, what's in progress, and what's next.
3. Check alignment: does the handoff's

   session_name

   match what the user is asking now? If unrelated, ask: "I have a handoff from a previous session on [session_name]. Is this related, or should I start fresh?"
4. Verify files exist before trusting the handoff. Spot-check 1-2 files from "Key Files" to confirm they're on disk. If missing, tell the user.
5. If aligned, start from the "Next" checklist — do NOT re-explore files already in "Key Files" unless the handoff indicates uncertainty.
6. If stale (branch mismatch, files changed since

   updated

   timestamp), mention this and do targeted re-discovery only on changed areas.

Session End Behavior

When the session is ending (user says goodbye, you're about to /clear, or context is getting heavy):

Step 1: Read Config

Check

.context/config.yaml

for mode setting. Default:

mode: ask

.

Step 2: Introspect

Gather state from:

• git diff --stat

  and

  git branch --show-current

• Files you read, created, or edited during this session
• Decisions made (especially where you chose between alternatives)
• Blockers hit or gotchas discovered
• The user's original task and any refinements

Step 3: Draft Handoff

Write a handoff document with these sections:

---
updated: [current ISO 8601 timestamp]
branch: [current branch from git branch --show-current]
session_name: '[brief label for this work]'
context_pressure: [low | medium | high]
---

## Done

[Completed work — each item references specific files:line numbers]

## In Progress

[Partially completed work — what state it's in]

## Blocked

[External dependencies, failing tests, missing info]

## Next

- [ ] [Ordered actionable checklist]

## Decisions

[Non-obvious choices and why they were made]

## Key Files

[All files relevant to the task, with line refs where useful]

Step 4: Write or Ask

• yolo mode: Write

  .context/handoff.md

  directly. Tell the user.
• ask mode: Show draft, incorporate feedback, then write.

Step 5: Verify Quality

After writing, check against targets:

• 30-80 lines, under 1k tokens
• Every "Done" item has a

  file:line

  reference
• "Next" items are actionable without reading conversation history
• A new agent reading only the handoff starts productive work in 2-3 tool calls

If the handoff fails any target, revise before saving.

Compaction Philosophy

The handoff is a compression of truth, not a transcript. Think of it like a teammate's sticky note: just enough to get you going, nothing more.

Include (high signal):

• Key files with line numbers — lets the next session jump to the right code
• Decisions and why — prevents re-debating the same choices
• Blockers and gotchas — the next session will hit the same walls without these
• Next steps as ordered checklist — clear starting point

Exclude (noise):

• Full file contents — files are on disk, just reference them
• Code snippets — reference by

  file:line

  instead. Exception: a one-line type signature the next session needs without reading the file
• Tool call transcripts — the handoff replaces these
• Exploratory dead ends — unless they're gotchas worth warning about
• Verbose error output — summarize the error and what fixed it

Subagent Strategy

For large repos (>10 key files or heavy git history), delegate the introspection step to a read-only subagent to keep the main context clean:

1. Spawn a background subagent with Haiku model and read-only tools (Read, Grep, Glob, Bash for

   git

   commands)
2. The subagent gathers: git diff, changed files, branch info, file summaries
3. It returns a compressed state summary (not raw output)
4. Use the summary to draft the handoff in the main conversation

This prevents file-reading noise from consuming the main context window.

Manual Checkpoint

Invoke

/session-handoff

mid-session to checkpoint progress without ending the session. This writes the handoff and the session continues.

Config Reference

.context/config.yaml

(optional, all fields have defaults):

handoff:
  mode: ask # yolo | ask
  auto_trigger: true # true: activate on session end. false: only on /session-handoff
  include_git_diff: true # include git diff --stat
  max_key_files: 20 # cap on listed files

When

auto_trigger

is false, the skill only writes a handoff when explicitly invoked. It will not activate automatically at session end.