interactive shell sends prompt_cache_key on chat-completions endpoints but zero exec doesn't

[cjgeremillo4638pam-oss]

Version / branch / commit

zero 0.3.0 (installed via npm install -g @gitlawb/zero)

OS and environment

Windows 10/11, Command Prompt (cmd.exe) Provider: custom openai-compatible endpoint (NVIDIA NIM, https://integrate.api.nvidia.com/v1) apiFormat: chat-completions Model: minimaxai/minimax-m2.7

Steps to reproduce

1. Run zero setup, select "Custom OpenAI-compatible" (or the "NVIDIA NIM" preset),
   set baseURL to https://integrate.api.nvidia.com/v1, paste a valid NVIDIA NIM API key,
   set model to minimaxai/minimax-m2.7
2. Confirm setup is valid: zero doctor reports provider.config and provider.connectivity as pass
3. Launch the interactive TUI: zero
4. Send any prompt (e.g. "hello")
5. Observe error: "provider request error: Validation: Unsupported parameter(s): prompt_cache_key"

For comparison:
6. Run the same prompt via headless mode instead: zero exec -p "hello"
7. This succeeds and returns a normal response

Expected behavior

The interactive TUI should return a response just like zero exec does, since both use the same
provider config and the endpoint doesn't support prompt_cache_key. Ideally the client should
either omit prompt_cache_key for openai-compatible providers that don't advertise support for it,
or expose a config flag to disable it per-provider.

Actual behavior

The interactive TUI fails every request with:
provider request error: Validation: Unsupported parameter(s): prompt_cache_key

zero exec with an identical provider/model config works correctly and returns valid completions,
indicating the interactive shell's request path adds a parameter that the one-shot exec path does not.

Relevant logs, screenshots, or error messages

provider request error: Validation: Unsupported parameter(s): prompt_cache_key

(zero doctor output, redacted — attach if helpful, e.g.:)
[pass] provider.config - Provider config loaded for nvidia-nim.
baseURL: https://integrate.api.nvidia.com/v1 | credentialConfigured: set | model: minimaxai/minimax-m2.7 | name: nvidia-nim | provider: openai-compatible
[pass] provider.connectivity - Provider endpoint reachable (200).
endpoint: https://integrate.api.nvidia.com/v1/models | statusCode: 200

[PierrunoYT]

Working on it!

[CengSin]

Set the environment variable export ZERO_DISABLE_PROMPT_CACHE_KEY=1, then relaunch the interactive shell.

[gnanam1990]

Fix PR opened

A fix is up in: #636

Root cause

• Interactive TUI always creates a session and sets PromptCacheKey from the session ID.
• The OpenAI-compatible chat-completions transport then serializes prompt_cache_key.
• Strict gateways (NVIDIA NIM, etc.) reject unknown fields; plain zero exec usually has no session ID so it never sent the field — which is why exec worked and the TUI failed.

Fix

• Omit prompt_cache_key for provider_kind: openai-compatible.
• Keep it for official OpenAI so multi-turn prefix-cache routing still works.
• Existing workaround still works: ZERO_DISABLE_PROMPT_CACHE_KEY=1.

Local workaround until this ships

export ZERO_DISABLE_PROMPT_CACHE_KEY=1

Then relaunch the interactive shell.

Thanks for the clear repro and exec vs TUI comparison — that made the session-ID path difference obvious.