| name | session-handoff |
| description | Create a session-level continuation pack when the current coding window should offload context before continuing in a new window. Use when users ask to整理当前 session, prepare a new window to continue the same unfinished task, or create a handoff/context-offload summary from current progress, blockers, plans, memory, and lessons. Do not use for reusable experience capture, standalone plan updates, or full transcript archival. |
Session Handoff
Purpose
Use this skill to offload the current session into a continuation pack before opening a new window for the same unfinished task.
The primary output is one Markdown handoff file under docs/session-handoff/ that helps the next window quickly recover task state, blockers, next action, and relevant lessons without manually rebuilding context.
Activation Cues
Activate when any condition is true:
- User explicitly asks to整理当前 session, 做 handoff, 做 context offload, or prepare a new window to continue the same task.
- The current task is unfinished and the user wants a session-level summary for next-window continuation.
- User asks to summarize what this window has done, what remains, and what the next window should read first.
Do not activate when any condition is true:
- The request is only to create reusable lessons or an experience card.
- The request is only to update plan status or lifecycle governance.
- The request is only to preserve long-term project memory without a window handoff need.
- The request is full transcript archival or raw chat export.
Scope Limit
This skill manages session-level continuation packs.
This skill does:
- summarize current-session progress for next-window continuation
- aggregate high-signal refs from plans, memory, experience, and process docs
- preserve blockers, next actions, and avoid-repeat guidance
- point the next window to the right execution entrypoint before coding when direct coding would be unsafe
This skill does not:
- own plan lifecycle status
- own reusable experience-card persistence
- own project memory truth
- archive full chat transcripts
- auto-write back to plan, memory, or experience docs in MVP
- author or refresh implementation plans itself
Ownership Boundary
sdd-plan-maintainer owns plan lifecycle and plan status truth under docs/plans/.
layered-project-memory owns continuity facts under docs/memory/.
experience-capture owns reusable cross-task experience under docs/experience/.
session-handoff owns session-level aggregation for next-window continuation.
Boundary rule:
- Read canonical artifact surfaces under
docs/ rather than depending on a specific skill name or output.
- Treat plan, memory, and experience docs as source of truth; the handoff pack is only a pointer-first aggregation layer.
- If a next-window plan review or plan creation should happen before coding, point to the right owner and current entrypoint, but do not write the plan content here.
Script Decision
Use the skill-bundled scripts/smoke.sh for structure and contract checks only.
No business runtime script is required in MVP.
Workflow Contract
- Confirm the request is a current-window offload for same-task continuation.
- Collect related plan, memory, experience, and process refs from
docs/ and other referenced evidence files.
- Write one handoff pack to
docs/session-handoff/SHO-YYYYMMDD-NNNN.md.
- Keep the pack pointer-first and scoped to the current session, even when the session spans multiple subtasks.
- Ensure the pack includes fixed sections:
Session ID
Generated At
Offload Reason
Session Goal
Overview
Related Refs
Subtask Snapshots
Avoid Repeat
Next Window Boot
Source of Truth
- Limit
Avoid Repeat to 1-3 high-signal items.
- New-window consumption order must be stated clearly:
- read
Offload Reason and Overview
- read
Subtask Snapshots
- read related plan refs
- read related memory refs
- read related experience refs and
Avoid Repeat
- then start new planning
- When direct coding would be unsafe, add one concise
Before coding hint inside Next Window Boot that points to the current execution entrypoint or the next planning owner.
Keep it pointer-first; do not turn the handoff into a plan or execution contract.
Hard Rules
- MVP is explicit-trigger only; no
suggest-once.
- Output exactly one Markdown continuation pack per trigger.
- Do not create
.json sidecars or SESSION_INDEX.json in MVP.
- Do not auto-refresh or mutate plan, memory, experience, or process docs in MVP.
- Keep references repo-relative whenever possible.
- Do not copy large logs or long document excerpts; summarize and point.
- If the current session invalidates an older plan, mark that plan as historical or reference-only in the pack and point to the new diagnostic entry, but do not mutate plan lifecycle state in this skill.
- Do not author plan steps, acceptance criteria, or execution contracts inside the handoff pack; only point to the next planning owner when needed.
Resource Map
- Read
references/positioning-boundary.md for scope and layer split.
- Read
references/trigger-rules.md for trigger and non-trigger rules.
- Read
references/canonical-surfaces.md for artifact-surface reading rules.
- Read
references/process-protocol.md for execution flow and handoff order.
- Read
references/output-contract.md for Markdown section contract.
- Read
references/resume-entrypoint-guidance.md when the next window may need a plan review or a plan refresh before coding.
- Read
references/regression-cases.md for trigger-boundary regression checks.
- Read
references/examples/README.md for examples and non-examples.
- Use the skill-bundled
scripts/smoke.sh for deterministic smoke checks.