agent-harness-design
Designing agent harnesses and tool systems — risk taxonomy for tools, permission decisions, draft/commit pattern, structured tool…
Validate a proposed DeepSeek API integration before any key or project context is sent: check thinking-mode tool-call history, strict-schema assumptions, bounded output, and provider data boundaries. Use when integrating DeepSeek, adding DeepSeek tool calls or streaming,
$ npx -y skills add AnastasiyaW/codex-claude-code-config --skill deepseek-provider-contract --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/deepseek-provider-contractContext preview
The summary Claude sees to decide when to auto-load this skill.
Validate a proposed DeepSeek API integration before any key or project context is sent: check thinking-mode tool-call history, strict-schema assumptions, bounded output, and provider data boundaries. Use when integrating DeepSeek, adding DeepSeek tool calls or streaming,
name: deepseek-provider-contract description: "Validate a proposed DeepSeek API integration before any key or project context is sent: check thinking-mode tool-call history, strict-schema assumptions, bounded output, and provider data boundaries. Use when integrating DeepSeek, adding DeepSeek tool calls or streaming, debugging DeepSeek 400 after a tool call, or evaluating a DeepSeek harness/MCP. Do NOT use for a generic model comparison, ordinary local coding, or to send a repository transcript to a provider by default."
This is a local, offline contract check. It neither reads `DEEPSEEK_API_KEY` nor sends code, transcripts, or secrets. A third-party harness is an implementation candidate, not a trusted route.
1. Classify the proposed prompt/data. Private archive, credentials, customer data, and whole transcripts stay local unless the user explicitly authorizes that provider and scope. 2. Freeze the provider model, endpoint, request budget, and fallback in the integration plan. Do not inherit defaults from a community wrapper. 3. Save a redacted request-history fixture that explicitly declares whether the **outbound** request carries tools, then run:
The fixture shape is `{"outbound_request":{"tools":[...]},"messages":[...]}`; use `"tools": []` only for an outbound tools-free request.
python skills/operational/deepseek-provider-contract/scripts/validate_deepseek_history.py fixture.json
4. Run the provider only in a bounded, opt-in experiment. Record model, endpoint, max output, cache-usage fields, request digest, and observed result. 5. Compare against the incumbent on the same frozen task before adding any route to a workflow. A passing history fixture proves message lifecycle only; it does not prove quality, price, availability, or data handling.
message, including `reasoning_content` and `tool_calls`. Do not rebuild a reduced assistant message from `content` alone.
from prior assistant turns in its history. Do not reduce this to only the most recent tool-call turn; check the current provider documentation because model and endpoint semantics change.
supported JSON Schema subset. Never call a schema "strict" merely because it looks valid locally.
correctness promise.
Adopt a DeepSeek route only when all are true:
1. The offline history validator and its regression tests pass. 2. A redacted live canary succeeds with the exact frozen integration. 3. The frozen benchmark beats or materially supplements the current route. 4. Provider data scope and failure fallback are written down.
Otherwise keep this skill and the fixture only; do not install a global MCP.
prior assistant `reasoning_content`; a tools-free request has different context semantics. Treat the offline validator as a narrow fixture check, not a substitute for the live API contract.
those separately.
tools-bearing request request fail with HTTP 400 even if the first request succeeded.
the official documentation at the integration date.
| Symptom | Likely cause | Fix | |---|---|---| | HTTP 400 after a tool result | Assistant reasoning was discarded | Preserve the complete assistant tool-call message and rerun the validator | | Tool JSON is rejected in strict mode | Wrong endpoint or unsupported schema | Use the beta endpoint, validate the schema subset, or turn strict mode off | | Wrapper works once then loses context | History was normalized too aggressively | Compare its persisted fixture with the provider contract | | Temptation to add global MCP | Scope and data boundary were skipped | Keep it opt-in until a frozen canary and benchmark pass |
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…