agent-harness-design
Designing agent harnesses and tool systems — risk taxonomy for tools, permission decisions, draft/commit pattern, structured tool…
Use when work moves between Claude Code and Codex, a session resumes from a handoff, or an agent may rewrite an existing implementation. Loads the shared CONTINUITY.json contract, preserves accepted decisions, checks Git baseline and scope, and requires explicit replan mode for
$ npx -y skills add AnastasiyaW/codex-claude-code-config --skill cross-harness-continuation --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/cross-harness-continuationContext preview
The summary Claude sees to decide when to auto-load this skill.
Use when work moves between Claude Code and Codex, a session resumes from a handoff, or an agent may rewrite an existing implementation. Loads the shared CONTINUITY.json contract, preserves accepted decisions, checks Git baseline and scope, and requires explicit replan mode for
name: cross-harness-continuation description: "Use when work moves between Claude Code and Codex, a session resumes from a handoff, or an agent may rewrite an existing implementation. Loads the shared CONTINUITY.json contract, preserves accepted decisions, checks Git baseline and scope, and requires explicit replan mode for intentional redesign. Do NOT use for fresh work with no prior agent's decisions to preserve, for a single-harness session that no one else is continuing, or as a general handoff-writing helper (see the session-handoff rule); this skill governs the contract between agents, not the notes one agent leaves itself."
Use this skill when a task continues after another agent, especially when the handoff is Claude -> Codex, Codex -> Claude, or a plugin has already received several partial fixes.
The repository is the source of truth for code. The continuation contract is the source of truth for intent that Git cannot express:
`<repo>/.claude/continuity/CONTINUITY.json`
The canonical shape is in `references/CONTINUITY.example.json`. It must contain:
1. Read `CONTINUITY.json`, the newest handoff, `AGENTS.md`/`CLAUDE.md`, and the current Git status. Treat the live checkout as authoritative over stale prose. 2. Verify the baseline commit, branch, dirty paths, and the claimed files. Preserve existing changes unless a test or the contract proves they are wrong. 3. Make one focused continuation change. Use `Edit` for existing files. Do not replace a whole file merely because another implementation is easier to write. 4. Run the smallest relevant test, then the independent verifier for a non-trivial or cross-module change. Record commands and outcomes in the contract or handoff. 5. Update the contract with what was preserved, what changed, what failed, and the single next step. Commit a clean, descriptive checkpoint when the project workflow allows it.
Replanning is valid when measured evidence shows the current design is wrong, the requirements changed, or an explicit user decision authorizes a redesign. Declare it in the process that launches the agent:
AGENT_CONTINUITY_MODE=replan AGENT_CONTINUITY_REASON=<specific evidence and intended redesign>
The reason is required. Without it, the guard blocks whole-file overwrites, out-of-scope edits, and near-whole-file replacements. A replan must also be recorded in the decision log/handoff; an environment flag alone is not proof.
Do not revert a plugin because its code “looks different”. First classify the diff as intentional, unverified, or conflicting; reconstruct the baseline from Git; run the plugin's focused tests and a fresh read-only verifier; then repair only confirmed regressions. Keep the old implementation available through Git or an isolated worktree until the replacement passes its acceptance checks.
contract. Do not infer scope from prose alone.
and inspect the dependency boundary before declaring continuation complete.
skill protects serial handoff and cannot resolve a concurrent merge for you.
in the private archive or project docs and link them from the contract.
| Symptom | Cause | Fix | |---|---|---| | `No CONTINUITY.json found` | The prior session left only prose handoff | Create the contract from the current Git baseline before editing existing code | | `scope violation` | The task expanded beyond the claimed slice | Finish the slice or update scope with a recorded reason | | `blocks Write over existing tracked file` | Whole-file overwrite would erase prior intent | Use a focused `Edit`, or explicitly enter replan mode | | `near-whole-file replacement` | An edit is acting like a rewrite | Split the edit and test each step, or replan with evidence | | Claude and Codex disagree | No shared decision/evidence was recorded | Re-read `preserve`, `do_not_redo`, tests, and Git; do not “pick a side” by intuition |
A practical configuration kit for Claude Code, Codex, and other coding agents. It contains architectural principles, enforcement hooks, skills, drop-in rules, starter templates, and dynamic-workflow commands.
Designing agent harnesses and tool systems — risk taxonomy for tools, permission decisions, draft/commit pattern, structured tool…
Use when designing, auditing, refactoring, or explaining an agentic harness for any domain, especially when work must continue…
Практическая инженерия диффузионных моделей: архитектуры, обучение, инференс, оптимизация памяти. Использовать при любых задачах…
Expert prompt engineering for FLUX.2 [klein] image generation and editing model. Use this skill whenever the user wants to create…
Plan or review LoRA and edit-training work specifically for FLUX.2 Klein or Qwen-Image-Edit, including paired datasets,…
Forensic image-to-prompt compiler for image generation models. Use this skill whenever the user wants to: convert/describe an…