Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro

Run ID: f8b26e02-871e-48e5-99a9-ba866b179e96

Reviewing files that changed from the base of the PR and between 5d39a59 and 60106b6.

• docs/development/CONFIG_FIELDS.md
• docs/reference/public-api.md

• GitHub Check: Greptile Review

docs/development/CONFIG_FIELDS.md (2)

34-34: lgtm - promptCacheRetention documented correctly.

the entry matches the guideline spec and aligns with lib/request/request-transformer.ts:206-216 fallback resolution logic. the common values 5m|1h|24h|7d are appropriate server-side cache duration strings.

---

78-78: lgtm - responseContinuation documented correctly.

the default false matches lib/config.ts:925-937 where normalizeBoolean falls back to false, and the schema in lib/schemas.ts:50-54 confirms it's z.boolean().optional().

docs/reference/public-api.md (3)

73-78: lgtm - hosted tool support documented correctly.

the enumeration of supported hosted tools matches the implementation in lib/request/request-transformer.ts:307-368:

• tool_search filtered when !capabilities.toolSearch
• computer / computer_use_preview filtered when !capabilities.computerUse
• namespace bundles with recursive nested tool filtering

the filtering behavior is accurately described.

---

71-71: verify the text.format preservation claim.

the documentation states text.format is preserved when verbosity defaults are applied, but the provided context snippets don't show this preservation logic. we need evidence that the request transformer handles text.format correctly.

run the following script to find text.format handling:

#!/bin/bash
# description: search for text.format preservation logic in request transformation

# search for text.format handling
rg -nP --type=ts -C5 'text\.format|text\[.format.\]' lib/request/

# search for verbosity default application
rg -nP --type=ts -C5 'textVerbosity|text.*verbosity' lib/request/

---

70-70: documentation claims are accurate – all verified in code.

the auto-fill behavior at index.ts:1354-1359 correctly implements the documented statement: previous_response_id gets auto-filled from stored session state when both getResponseContinuation(pluginConfig) returns true and the field isn't already set in the request body. test case at test/index.test.ts:1323-1380 confirms the follow-up request receives the stored value.

text.format preservation at request-transformer.ts:989 happens through the spread operator (...body.text) which leaves any existing format field untouched. prompt_cache_retention fallback correctly chains body.prompt_cache_retention → body.providerOptions?.openai?.promptCacheRetention → modelConfig.promptCacheRetention at request-transformer.ts:210-214. all four hosted tool types (tool_search, mcp, computer/computer_use_preview, namespace) have type definitions in lib/types.ts and filtering logic in request-transformer.ts:305-352. all five sse synthesis fields (output_text, reasoning_summary_text, commentary_text, final_answer_text, phase_text) are assigned in lib/request/response-handler.ts:199-322.

the section is ready as written.