AgentSkillsCN

handoff-haiku

仅限 Haiku 模型编排器使用的内部技能。不适用于 Sonnet 或 Opus——请改用 /handoff。在不依赖学习判断的前提下,机械式地保留会话上下文。

SKILL.md
--- frontmatter
name: handoff-haiku
description: Internal skill for Haiku model orchestrators only. Not for Sonnet or Opus — use /handoff instead. Mechanical session context preservation without learnings judgment.
user-invocable: false

Skill: handoff-haiku

Mechanical session handoff for Haiku models. Preserves context without judgment calls. For full handoff protocol with learnings processing, use /handoff skill (Sonnet/Opus).

Target Model: Haiku (mechanical preservation)

Protocol

When invoked, update session.md mechanically:

1. Review Conversation

  • Identify completed tasks (with commit hashes/file refs)
  • Identify pending/remaining tasks
  • Note any blockers or gotchas

2. Update session.md

Merge rules — read current session.md first, then apply:

  • REPLACE these sections with fresh content: "Completed This Session", "Next Steps"
  • MERGE these sections (carry forward unresolved items + add new): "Pending Tasks", "Blockers / Gotchas"
  • ADD "Session Notes" section if new observations to record
  • PRESERVE UNCHANGED all other existing sections (especially "Recent Learnings", "Reference Files", any "Prior Session" content)

How to apply:

  1. Read current session.md
  2. Update header (date, status line)
  3. Replace sections listed under REPLACE with fresh content
  4. For MERGE sections: preserve unresolved items from previous session, add new items, mark completed items with [x]
  5. Add Session Notes section if new observations exist
  6. Keep everything else exactly as-is

MERGE semantics for Pending Tasks:

  • Review all pending tasks from previous session.md
  • Keep all unresolved tasks (mark with [ ] or carry forward as-is)
  • Add new pending tasks from current session
  • Mark completed tasks with [x] if they were completed in current session

Task metadata format:

Use this convention when writing tasks:

code
- [ ] **Task Name** — `command` | model | restart?

Examples:

code
- [ ] **Implement ambient awareness** — `/plan-adhoc plans/ambient-awareness/design.md` | sonnet
- [ ] **Design runbook identifiers** — `/design plans/runbook-identifiers/problem.md` | opus | restart

Field rules:

  • Command: Backtick-wrapped command to start the task
  • Model: haiku, sonnet, or opus (default: sonnet if omitted)
  • Restart: Optional flag - only include if restart needed (omit = no restart)

Mechanical merge: Preserve metadata format verbatim when carrying forward unresolved items. No judgment needed - copy unchanged.

MERGE semantics for Blockers / Gotchas:

  • Review all blockers/gotchas from previous session.md
  • Keep items that are still relevant (unresolved blockers, active gotchas)
  • Add new blockers/gotchas discovered in current session
  • Drop items that were resolved in current session (document resolution in Completed or Session Notes if significant)

Template for sections being replaced/added:

Completed This Session

[Bullet list of completed work with commit hashes/file refs]

Pending Tasks

[Bullet list with checkboxes]

Session Notes

[Raw observations, discoveries, issues encountered - NO FILTERING] [Preserve verbatim what happened, let standard model judge later]

Blockers / Gotchas

[Any blockers or warnings for next agent]

Next Steps

[Immediate next action]

CRITICAL: Do NOT delete existing sections not shown above. session.md may contain sections from prior handoffs (Recent Learnings, Reference Files, etc.) that MUST be preserved.

3. Report Completion

Report line count: "Session handoff complete. [X] lines."

Key Differences from Full Handoff

"Session Notes" not "Recent Learnings":

  • No judgment about what qualifies as a learning
  • Dump raw observations for standard model to process later
  • Preserve verbatim what happened

No learnings staging:

  • Don't call add-learning.py
  • Let standard model handle learnings judgment

No size advice:

  • Just report completion with line count
  • Let user decide on session continuity

Principles

Preserve everything mechanically:

  • Don't filter or judge importance
  • Include specifics: commit hashes, file paths, line numbers, metrics
  • Document what happened and why

Trust the next agent:

  • Standard model will review Session Notes for learnings
  • Standard model will decide on consolidation needs
  • Preserve context completely without filtering