# Adapt Agent Backup > Implement a new local Code Agent conversation backup adapter in context-harness. Use when asked to add support for backing up another AI coding assistant, adapt a new Agent source, create collectors/hooks/config/tests for a new context-harness source, or prepare a PR that lets context-harness sync a currently unsupported Agent. - **Type:** Skill - **Install:** `agentstack add skill-yinjialu-context-harness-adapt-agent-backup` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [yinjialu](https://agentstack.voostack.com/s/yinjialu) - **Installs:** 0 - **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [yinjialu](https://github.com/yinjialu) - **Source:** https://github.com/yinjialu/context-harness/tree/main/skills/adapt-agent-backup ## Install ```sh agentstack add skill-yinjialu-context-harness-adapt-agent-backup ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # Adapt Agent Backup Use this skill when adding support for a new Code Agent source to `context-harness`. ## First Read Before editing, read: - `docs/agent-backup-adapter-architecture.md` - Existing collectors in `context_harness/collectors/` - Existing tests for Codex and Claude Code in `tests/` ## Workflow 1. Identify the source id. - Use lowercase kebab-case, for example `my-agent`. - Infer the target Agent from the current runtime, app name, CLI name, repository metadata, process environment, or user prompt. - Derive the id from that Agent name unless the user explicitly provides one. - Ask the user for the target Agent only if it cannot be inferred with reasonable confidence. - Use the same id in config, CLI, output directories, state keys, tests, and docs. 2. Check whether the source is already supported. - Search for the source id and common aliases in `context_harness/`, `tests/`, `skills/`, `README.md`, and `README.zh-CN.md`. - Inspect `context_harness/cli.py` sync choices, `context_harness/config.py`, `context_harness/init.py`, `context_harness/collectors/`, and `context_harness/hooks/`. - If support is complete, do not add a duplicate adapter. Instead, verify it with existing tests or a smoke command, then report that the source is already supported. - If support is partial or broken, repair only the missing or failing pieces and preserve the existing source id and archive/state compatibility. 3. Inspect the target Agent's local transcript format. - Find default transcript paths. - Identify session id, title, timestamps, roles, user messages, assistant messages, tool/noise events, and internal meta events. - Check whether the Agent has a reliable stop/session hook. If not, implement manual sync only. - Use synthetic or anonymized fixtures only. 4. Add config and initialization. - Update `context_harness/config.py` with a new source in `AppConfig` and `load_config`. - Update `context_harness/init.py` with default config and `conversations/` directory creation. 5. Add a collector. - Create `context_harness/collectors/.py`. - Parse raw transcript files into `Conversation` and `Message`. - Ignore bad JSON lines, empty messages, tools, and internal metadata. - Return `None` for sessions with no user-visible conversation. - Reuse `render_conversation`, `read_state`, `write_state`, and the existing `SyncResult` pattern. - Support `latest`, `all_sessions`, and `session_path`. 6. Register CLI commands. - Add the source to `sync` choices in `context_harness/cli.py`. - Dispatch to `sync_`. - Respect `enabled = false` with `_disabled_result`. 7. Add hooks only when reliable. - Create `context_harness/hooks/.py` only if the Agent has a stable hook/config mechanism. - Hook commands should call `sync --hook-stdin`. - Preserve unrelated user settings and make hook installation idempotent. - Register hook install support in `context_harness/cli.py`. 8. Update docs and skills. - Update `README.md` and `README.zh-CN.md`. - Add the source to the supported local AI coding assistants list in both README files. - Prefer the target Agent's official product/docs URL for the link. - Prefer an official favicon/logo URL from the Agent's website or docs for the icon; if no reliable official icon URL exists, use a plain text link instead of adding an unofficial asset. - Update `skills/init-context/SKILL.md` and `skills/sync-conversations/SKILL.md`. - Keep skills as workflow instructions; keep behavior in the CLI. 9. Add focused tests. - Add fixtures under `tests/fixtures/`. - Cover config defaults/custom paths, init directories/default config, collector archive output, skipped unchanged archives, missing source directory, latest mtime behavior, hook `session_path` override, noise filtering, bad JSON tolerance, filename stability, CLI sync, disabled source behavior, and hook idempotence if hooks are implemented. 10. Verify. - Run `uv sync` if dependencies are not installed. - Run `uv run pytest`. - Manually smoke test `uv run context-harness --context-home sync --latest 1` when a fixture or local sample is available. - Keep changes small and complete; do not refactor unrelated modules. 11. Prepare the PR for automated review. - Add label `agent-backup-adapter`. This label marks the PR as an Agent backup adapter and allows repository automation to run adapter-specific checks. - Add label `ai-review` only when an explicit Claude AI review is desired. This is a manual shortcut label and does not replace `agent-backup-adapter`. - Include `Source id: ` in the PR body. - Include transcript path/format assumptions, hook support status, verification commands, and remaining risks. - Optional AI review enhancement: if the repository has Codex GitHub integration enabled, comment `@codex review`; if Claude Code GitHub Actions is enabled, either add label `ai-review` or comment `@claude review this adapter PR against the adapt-agent-backup checklist`. Any one of these is acceptable as an extra AI review signal. ## PR Checklist - The adapter produces Markdown archives in `context_home/conversations//`. - The adapter writes incremental state to `context_home/state/-sync-state.json`. - README supported-Agent lists, examples, default config, skills, CLI choices, and tests all use the same source id. - Real user transcripts, memory, logs, state, and private machine config are not committed. - PR has label `agent-backup-adapter`, and its body includes `Source id: `. - PR has label `ai-review` only when a manual Claude AI review should be requested. - PR description states transcript format assumptions, hook support status, commands run, and remaining risks. - The diff avoids unrelated rewrites and leaves existing behavior intact outside the new adapter path. ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [yinjialu](https://github.com/yinjialu) - **Source:** [yinjialu/context-harness](https://github.com/yinjialu/context-harness) - **License:** MIT Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-yinjialu-context-harness-adapt-agent-backup - Seller: https://agentstack.voostack.com/s/yinjialu - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.