Greptile Summary

This PR fixes a real HTTP 400 failure for Volcano Engine DeepSeek (and any other non-OpenAI provider using the openai-responses API) by stripping prompt_cache_key and prompt_cache_retention — fields unconditionally injected by pi-ai — from request payloads for providers not in a new explicit allowlist (openai, openai-codex, opencode, azure-openai-responses, github-copilot).

The implementation reuses the existing streamWithPayloadPatch / onPayload interception pattern correctly: delete on a missing key is a no-op, so Anthropic and other non-OpenAI API types are unaffected in practice, and the workspaceDir threading through createAnthropicToolPayloadCompatibilityWrapper (addressed in prior review threads) is preserved.

Key observations:

• The new createStripPromptCacheFieldsWrapper does not include the model.api !== "openai-responses" short-circuit that sibling wrappers like createParallelToolCallsWrapper use. Since prompt_cache_key is only injected for openai-responses calls, this is harmless but inconsistent with the established pattern.
• No unit tests were added for the new stripping behaviour. Given the existing per-feature test files in extra-params.*.test.ts, a focused test asserting stripping for non-listed providers and preservation for listed providers would improve confidence and catch future regressions.

Confidence Score: 4/5

• Safe to merge — the fix is narrowly scoped, uses the established onPayload interception pattern, and is backward compatible for all existing providers.
• The logic is correct and the regression surface is small: delete on a missing key is a no-op, so only providers that actually use openai-responses with an unlisted provider ID are affected. The two gaps (missing API-type guard and no tests) are style/coverage concerns, not functional bugs.
• src/agents/pi-embedded-runner/extra-params.ts — verify the PROMPT_CACHE_KEY_PROVIDERS allowlist is complete if any new openai-responses-backed provider is added in the future.

This is a comment left during a code review.
Path: src/agents/pi-embedded-runner/extra-params.ts
Line: 44-51

Comment:
**Missing API-type guard, unlike sibling wrappers**

`createParallelToolCallsWrapper` (line 172) short-circuits immediately when `model.api` is neither `"openai-completions"` nor `"openai-responses"`:

```ts
if (model.api !== "openai-completions" && model.api !== "openai-responses") {
  return underlying(model, context, options);
}
```

`createStripPromptCacheFieldsWrapper` has no such guard. `prompt_cache_key` / `prompt_cache_retention` are only injected by pi-ai for `openai-responses` payloads, so the `delete` calls are no-ops for every other API type — but the wrapper still installs an extra `onPayload` interceptor on every Anthropic, Google, Moonshot, etc. call made through providers not in the allowlist.

Adding the guard would be consistent with the existing pattern and make the intent explicit:

```suggestion
function createStripPromptCacheFieldsWrapper(baseStreamFn: StreamFn | undefined): StreamFn {
  const underlying = baseStreamFn ?? streamSimple;
  return (model, context, options) => {
    if (model.api !== "openai-responses") {
      return underlying(model, context, options);
    }
    return streamWithPayloadPatch(underlying, model, context, options, (payload) => {
      delete payload.prompt_cache_key;
      delete payload.prompt_cache_retention;
    });
  };
}
```

How can I resolve this? If you propose a fix, please make it concise.

---

This is a comment left during a code review.
Path: src/agents/pi-embedded-runner/extra-params.ts
Line: 248-250

Comment:
**No test coverage for the new stripping logic**

The `extra-params.*.test.ts` suite is fairly comprehensive — there are dedicated files for OpenAI, Google, xAI, OpenRouter cache-control, and cache-retention defaults. No test file was added (or updated) to assert that `prompt_cache_key` and `prompt_cache_retention` are stripped for unlisted providers (e.g. a "volcano" or "custom-openai-responses" provider), and that they are _kept_ for providers in `PROMPT_CACHE_KEY_PROVIDERS`.

A failing assertion here would surface if the allowlist accidentally drops a supported provider or if `streamWithPayloadPatch` changes its interception order.

How can I resolve this? If you propose a fix, please make it concise.

_{Last reviewed commit: "fix: strip prompt_ca..."}

Review Conversations Status

All bot review comments have been addressed:

greptile-apps[bot] — workspaceDir silently dropped (P1)

✅ Resolved — The initial push was based on a stale copy of extra-params.ts. After rebasing on latest main in 672148d5, workspaceDir is preserved in the function signature (line 205) and correctly forwarded to createAnthropicToolPayloadCompatibilityWrapper (line 269).

greptile-apps[bot] — workspaceDir no longer forwarded to Anthropic tool-payload wrapper (P1)

✅ Resolved — Same rebase fix. The call now correctly passes { config: cfg, workspaceDir }.

chatgpt-codex-connector[bot] — Restore resolver options for Anthropic payload compatibility (P1)

✅ Resolved — Same root cause and fix as above.

Verification

• Current diff is 1 commit ahead of main, 0 behind, 1 file changed
• workspaceDir confirmed present at line 205 (signature) and line 269 (forwarded)
• The only change is the addition of createStripPromptCacheFieldsWrapper + PROMPT_CACHE_KEY_PROVIDERS set

Closing in favor of a cleaner approach. The allowlist-based fix duplicates provider knowledge already encoded in openai-stream-wrappers.ts. The replacement PR will fold the stripping logic into createOpenAIResponsesContextManagementWrapper using the existing isDirectOpenAIBaseUrl() check — no new provider lists, no new files.