# Handoff > Agent-to-agent cross-session transfer — pointers not bodies. Channel prompt (user /handoff) or artifact (model-invoked subagent). Pack and Goal are open vocabulary. Orchestrator for continuing work in a fresh chat. Not same-session continuation without a new chat. - Skill: `csark0812/handoff` (Agent Skill, multi-file: 5 files) - Install (CLI): `npx skillmds@latest add csark0812/handoff` - Raw SKILL.md: https://api.skillmd.com/api/skills/csark0812/handoff/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: csark0812 (https://skillmd.com/u/csark0812) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/csark0812/handoff --- # Handoff **Orchestrator** — cross-session A2A (`channel:prompt` vs `channel:artifact`). Shared vocabulary → [process-skill-composition.md](https://raw.githubusercontent.com/csark0812/toolbox/main/references/process-skill-composition.md). References: [pack.md](references/pack.md) · [output.md](references/output.md) · [handoff-subagent-dispatch.md](references/handoff-subagent-dispatch.md) When you calibrate claims, read [research-basis.md](references/research-basis.md). ## Entry gate - **New chat** intended — same-session continuation without transfer is out of scope. - **Channel** chosen — `prompt` (user) or `artifact` (model-invoked). ## Non-negotiables 1. **Pointers not bodies** — paths, URLs, SHAs. Never paste plans, PRDs, or diffs into handoff. 2. **Honest state** — done vs in-progress vs broken. When you claim progress, cite commits or tests. 3. **Redact** secrets, tokens, and PII before write or paste. 4. **Channel prompt** — user `/handoff` → **no disk** (`_agent/handoffs/` forbidden). 5. **Channel artifact** — model-invoked → subagent write mandatory. Coordinator-primary write is a **violation**. ## Workflow 1. **Pick the channel, pack, and goal** — use the closest rows in [pack.md](references/pack.md). These values control production only. 2. **Gather** — minimal bullets from thread. Omit empty categories. 3. **Deliver** — - `channel:prompt` → fenced receiving prompt per [output.md](references/output.md) in pragmatic STE - `channel:artifact` → [handoff-subagent-dispatch.md](references/handoff-subagent-dispatch.md) → paste stub only ## Execution interface map Use this map when the request says "handoff via X client." 1. Pick channel from producer intent: - `prompt` for user-directed handoff (`/handoff`) - `artifact` for model-invoked context-pressure continuation 2. Build one canonical payload from `pack`, `goal`, and thread claims. 3. Route to the client interface: - `prompt`: emit receiving prompt block and hand it to next chat startup. - `artifact`: write `/_agent/handoffs/.md`, then emit only the read stub. 4. Validate output with the client-specific rule. ### Cross-client default behavior | Interface | Transport rule | Producer action | Required receive artifact | | ------------------------------- | ------------------------- | ------------------------------------------------------------------------------------------------------ | ----------------------------------------------- | | Codex | Chat thread controls | `channel:prompt` for user handoff, `channel:artifact` for model-invoked | `channel:prompt` block or handoff artifact path | | Cursor | Conversation continuation | `channel:prompt` into fresh chat unless artifact hook is explicitly available | Prompt block | | Claude Code | Chat continuation | `channel:prompt` into fresh chat; if client-specific task API exists, pass block as first user message | Prompt block | | Claude API / GPT-like interface | Chat continuation | `channel:prompt` is required; map fields to a user-visible handoff message | Prompt block | | ChatGPT UI / web chat | Chat continuation | `channel:prompt` with `Open workspace` target first line | Prompt block | | GitHub Copilot Chat | IDE continuation | `channel:prompt` only; no shared artifact path unless local host exposes handoff write | Prompt block | | Generic MCP / custom client | API-specific | map to canonical payload fields; if no artifact support, downgrade to prompt | Prompt block (or artifact path if supported) | When an interface does not expose an artifact API, always fall back to `channel:prompt`. ### Codex API branch Use this when the ask explicitly targets Codex API execution. 1. Build `channel:prompt` payload from plan claims. 2. Prefer one of: - start fresh chat with the prompt payload if orchestration is in user flow. - use `create_thread` for a new Codex task, then `send_message_to_thread` with the prompt block for handoff bootstrap. 3. For `channel:artifact`, call `handoff` with `channel:artifact` via your orchestration flow, spawn the subagent writer path, then emit only the stub pointing to `/_agent/handoffs/.md`. Avoid fabricating any unsupported Codex tool calls. If a required API is unavailable, document fallback at `channel:prompt`. ### High-volume interface notes - For clients with both UI and API modes, document the mode explicitly in your own orchestration. - If a client has a structured conversation API, preserve the same fields and map only into its initial message payload. - For IDE-native chat clients, prefer prompt mode with explicit workspace. ### Interface execution boundaries - Don’t hard-code any one chat stack’s control API as the only supported path. - Do not include `channel` and `pack` inside receiving content. - Don’t skip the redaction rule for any interface. - Don’t write handoff files from the coordinator for `channel:artifact`; spawn/coordinate the subagent path when required. If the next session needs only review state, use consumer overlay blocks instead of `Pack: full`. ## Consumer bindings Project-specific injected context is appended on skill read.