handoff ยท diff
git:20261003.db01775 to git:20261015.c0bb9cf
6 added, 127 removed. Audit A to A.
---
name: handoff
description: Write a self-contained context summary that a new session can continue from.
when_to_use: |
Use when:
- Context window is getting long or session quality is degrading
- Approaching context window capacity
- User says "handoff", "context is long", or "start fresh"
- Switching to a new Claude Code session mid-task
user-invocable: true
disable-model-invocation: true
shell: bash
argument-hint: optional notes about what to include
allowed-tools:
- Read
- Glob
- Bash(git status *)
- Bash(git diff *)
- Bash(git log *)
- Bash(git branch --show-current)
- mcp__plugin_oh-my-claudeagent_omca__boulder_progress
- mcp__plugin_oh-my-claudeagent_omca__notepad_read
- mcp__plugin_oh-my-claudeagent_omca__notepad_list
---
- # Handoff: session context summarization
-
- ## Tool restrictions
-
- Read-only. No Write/Edit/Agent. MCP tools: `boulder_progress`, `notepad_read`, `notepad_list`. `boulder_write` is deliberately absent: it is the only writer of the plan registry, and a summarizer has nothing to register.
-
- Self-contained handoff summary for new-session continuation.
-
- ## PHASE 0: VALIDATE
-
- Confirm meaningful work exists. For a nearly empty session, tell the user there is nothing to hand off.
-
- ## PHASE 1: GATHER CONTEXT
-
- Run in parallel:
-
- ```bash
- git diff --stat HEAD~10..HEAD 2>/dev/null || git log --oneline -10
- git status --porcelain
- git branch --show-current
- git log --oneline -5
- ```
-
- Also: `boulder_progress()`, `notepad_read` for active plan sections (load it if needed: `ToolSearch({query: "select:mcp__plugin_oh-my-claudeagent_omca__notepad_read", max_results: 1})`), the plan files in `<plans-dir>` (the `plansDirectory` setting when set, relative to the project root, otherwise `~/.claude/plans`). Treat the plan file's numbered checkboxes and `boulder_progress()` as the primary source, and `TaskList()`, where that tool exists, as a supplement.
-
- ## PHASE 2: EXTRACT
-
- First person ("I did...", "I told you..."):
- - What was done, what remains
- - KEY FILES (max 10, workspace-relative)
- - USER REQUESTS verbatim, no paraphrasing
- - EXPLICIT CONSTRAINTS verbatim, no inventing
-
- ## PHASE 3: FORMAT OUTPUT
-
- ```
- HANDOFF CONTEXT
- ===============
-
- USER REQUESTS (AS-IS)
- ---------------------
- - [Exact verbatim user requests]
-
- GOAL
- ----
- [One sentence: what should be done next]
-
- PROGRESS
- --------
- Tasks completed: X / Y total
- - [x] [Completed task name]
- - [ ] [Pending task name]
- (Use the plan file's checkboxes and `boulder_progress()`, plus `TaskList()` counts where that tool exists. Never invent a count: if no source is available, write "unknown".)
-
- NOTEPAD SUMMARY
- ---------------
- [Only include if an active plan with notepad data exists]
- Plan: [plan-name]
- - learnings: N entries - [one-line summary of key insight]
- - issues: N entries - [one-line summary of open issue if any]
- - decisions: N entries - [list verbatim from notepad decisions section, one per line]
- - problems: N entries - [summary if non-zero]
-
- WORK COMPLETED
- --------------
- - [First person bullets of what was done]
- - [Include specific file paths]
-
- CURRENT STATE
- -------------
- - [Current state of codebase or task]
- - [Build/test status if applicable]
-
- REMAINING WORK
- --------------
- - [ ] [Unchecked plan task - copy exact task text]
- - [ ] [Next logical step if not in plan]
- (Pull from plan file checkboxes, and from `TaskList()` state where that tool exists)
-
- KEY FILES
- ---------
- - [path/to/file] - [brief role]
- (Maximum 10 files, prioritized by importance)
-
- IMPORTANT DECISIONS
- -------------------
- - [Technical decisions and why - pull from notepad decisions section if available]
- - [Trade-offs considered]
-
- EXPLICIT CONSTRAINTS
- --------------------
- - [Verbatim constraints only - from user or existing AGENTS.md]
- - If none: None
-
- CONTEXT FOR CONTINUATION
- ------------------------
- - [What the next session needs to know]
- - [Warnings or gotchas]
- ```
-
- Rules:
- - Plain text, no markdown headers in output
- - No bold/italic/code fences in content fields
- - Only what matters for continuation
- - USER REQUESTS and CONSTRAINTS: verbatim only
- - PROGRESS: actual counts from the plan file and `boulder_progress()`, supplemented by `TaskList()` where available. Write "unknown" rather than a guess
- - NOTEPAD SUMMARY: omit if no active plan/notepad data
- - REMAINING WORK: verbatim from plan file
-
- ## PHASE 4: INSTRUCT
-
- ```
- TO CONTINUE IN A NEW SESSION:
+ # Handoff
- 1. Start a new Claude Code session: claude
- 2. Paste the HANDOFF CONTEXT above as your first message
- 3. Add: "Continue from the handoff context above. [Your next task]"
- ```
+ Run only when the user explicitly asks for a handoff. A long conversation alone is not authorization. Read-only: do not change files, register or bind a plan, delegate, or create a session.
- ## Invocation
+ Write compact, self-contained continuation context: the goal and exact constraints, verified progress, relevant changed files, decisions, blockers, and the next action. Quote requests and explicit constraints accurately; do not invent requirements. Inspect only the status, diff, plan, or notepad details needed to make those statements accurate. Use `boulder_progress` only for this session's explicitly bound plan; do not select another session's plan. Read relevant notepad sections with `notepad_read` when needed.
- Handoff is a user-driven workflow, so this skill sets `disable-model-invocation: true`: it
- runs only when you type `/oh-my-claudeagent:handoff`. Claude cannot load it on its own, and
- it is not preloaded into subagents. With keyword triggers on, a prompt that says "handoff"
- gets a one-line hint suggesting the command; the hint does not start the workflow.
+ Separate observed results from assumptions and pending checks. Do not invent task counts, dump metadata or transcripts, include credentials, or imply historical evidence verifies the current checkout. Copy remaining plan tasks accurately. If there is no meaningful work to transfer, say so. Let the user choose whether to start a new session and paste the context there.
- ## Constraints
+ ## Claude Code
- - No programmatic session creation
- - No sensitive information (API keys, credentials)
- - Max 10 KEY FILES
- - User requests quoted verbatim
+ This skill is user-invoked through `/oh-my-claudeagent:handoff`; do not load it automatically. Find a requested plan in the configured `plansDirectory`, or `~/.claude/plans` by default. The user can start a new session with `claude` and paste the handoff as the first message. Do not create the session programmatically.