Greptile Summary

This PR adds text.verbosity forwarding to the OpenAI Responses API path, following the same pattern already used for service_tier and reasoning.effort. Both alias styles (textVerbosity / text_verbosity) are supported, null-override suppression is implemented, and existing text payload fields are preserved during injection. Test coverage is thorough across the payload mutation, alias canonicalization, invalid-value warning, and null-suppression cases.

Key observations:

• The private normalizeOpenAITextVerbosity function is duplicated between openai-ws-stream.ts and openai-stream-wrappers.ts with identical logic. More importantly, the WebSocket send path uses the local copy, which silently discards invalid verbosity values without logging a warning. The HTTP/payload wrapper path correctly uses resolveOpenAITextVerbosity and emits a log.warn for the same scenario. Consolidating to a single shared utility (or importing resolveOpenAITextVerbosity in openai-ws-stream.ts) would close the diagnostic gap and remove the duplication.
• The createOpenAITextVerbosityWrapper guard (model.api !== "openai-responses") is intentionally broader than createOpenAIServiceTierWrapper (which also checks model.provider === "openai" and the base URL). This is consistent with how createOpenAIResponsesContextManagementWrapper handles the same scope, so it appears deliberate.
• The resolveAliasedParamValue + null-suppression flow in applyExtraParamsToAgent correctly handles all four cases: no value, valid value, invalid value, and null override.

Confidence Score: 4/5

• Safe to merge with one minor inconsistency to address: invalid verbosity values are silently dropped on the WebSocket path where the HTTP path would log a warning.
• The implementation is well-structured, closely mirrors existing patterns (service_tier, reasoning.effort), and has good test coverage across edge cases. The only notable gap is the missing log.warn for invalid textVerbosity values when using the WebSocket transport, which won't cause a runtime error but can make misconfigured settings harder to diagnose. No data-loss or correctness bugs were found.
• src/agents/openai-ws-stream.ts — the duplicated normalizeOpenAITextVerbosity and the missing warning for invalid values on the WS path.

This is a comment left during a code review.
Path: src/agents/openai-ws-stream.ts
Line: 604-613

Comment:
**No warning logged for invalid verbosity in WS path**

The WebSocket path silently discards invalid `textVerbosity` values (e.g. `"loud"`) with no feedback to the user. The HTTP path in `openai-stream-wrappers.ts` uses `resolveOpenAITextVerbosity`, which emits a `log.warn` when normalization fails. This inconsistency means a misconfigured verbosity will surface a warning on SSE transport but be silently ignored on WebSocket transport, making it harder to debug.

Consider calling `resolveOpenAITextVerbosity` here (after importing it from `openai-stream-wrappers.ts`) instead of the local `normalizeOpenAITextVerbosity`, which would give users the same warning regardless of transport:

```ts
// in imports at top of file
import { resolveOpenAITextVerbosity } from "./pi-embedded-runner/openai-stream-wrappers.js";

// then replace lines 604–613 with:
      const textVerbosity = resolveOpenAITextVerbosity(
        streamOpts as Record<string, unknown> | undefined,
      );
      if (textVerbosity !== undefined) {
        const existingText =
          extraParams.text && typeof extraParams.text === "object"
            ? (extraParams.text as Record<string, unknown>)
            : {};
        extraParams.text = { ...existingText, verbosity: textVerbosity };
      }
```

As a bonus, this would eliminate the duplicate `normalizeOpenAITextVerbosity` definition in this file.

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

_{Last reviewed commit: 7d4aa5c}

Follow-up pushed to finish the user-visible/E2E gap.

New commit on this PR branch: b04cbcc215

Delta in this commit:

1. createOpenAITextVerbosityWrapper() now also applies to openai-codex-responses.
2. Codex path now force-applies configured verbosity (so explicit model param overrides Codex default medium).
3. Added tests:
   • Codex responses receive configured verbosity.
   • OpenAI responses preserve caller-specified payload.text.verbosity.
4. /status now surfaces configured text verbosity as Text: <level> for the active model.
5. Added status test for text verbosity visibility.

Re-ran targeted suite:

• src/agents/pi-embedded-runner-extraparams.test.ts
• src/agents/openai-ws-stream.test.ts
• src/auto-reply/status.test.ts

Result: 145/145 passing.

Addressed the remaining Greptile review note in 1754a4c306.

What changed:

• removed the duplicate WS-local normalizeOpenAITextVerbosity() helper
• reused shared resolveOpenAITextVerbosity() in openai-ws-stream.ts
• invalid WS textVerbosity values now emit the same log.warn(...) as the HTTP path
• added WS test coverage for the invalid-value warning case

Re-ran targeted suite:

• src/agents/pi-embedded-runner-extraparams.test.ts
• src/agents/openai-ws-stream.test.ts
• src/auto-reply/status.test.ts

Result: 146/146 passing.

Rebased/refresh pass is pushed on top of current upstream/main (force-pushed head: e4c94c910e).

What this preserves from the original PR intent:

• forwards textVerbosity / text_verbosity for Responses payloads
• applies configured text verbosity for both OpenAI Responses and Codex Responses
  • Codex path intentionally overrides existing payload.text.verbosity
  • OpenAI Responses path preserves caller-provided payload.text.verbosity
• WS transport now uses shared resolveOpenAITextVerbosity()
  • invalid WS values emit the same warning as HTTP path
• status output shows configured text verbosity as Text: <level>

Targeted suite re-run on refreshed head:

• src/agents/pi-embedded-runner-extraparams.test.ts
• src/agents/openai-ws-stream.test.ts
• src/auto-reply/status.test.ts

Result: 167/167 passing.