feat(chat): add the raw llm request viewer #342

Merged

rwachters merged 2 commits from feat/raw-llm-request-viewer into master

2026-10-11 22:01:26 +00:00

rwachters commented

2026-10-11 22:00:11 +00:00

(Migrated from github.com)

Summary

Lets the owner of a chat session inspect the exact provider request the server prepared for the latest assistant step, so a surprising reply can be traced back to what was actually sent.

Changes

Added a provider-neutral RawLlmRequestPayload (method, url) and a raw_llm_request informational event on both the non-streaming and streaming chat surfaces. A step that issues several requests emits one event per request, and the latest one wins.
Captured the prepared request in the LLM client path through an onRequestPrepared observer, forwarded through the retrying decorator, the turn runners and the transport mappers; the streaming client is now suspending so the observer runs in the turn's own coroutine before dispatch.
Masked the credential before any event exists: the URL and every header value get the same min(8, length/2)-prefix masking, a short key becomes ****, and no Authorization header is synthesized for a keyless provider.
Added an app-side capture slot with overwrite semantics, cleared by a state reset and by an accepted send, plus a read-only dialog with an explicit empty state, a Copy action and a top-bar menu entry.
Copied document contains only the four wire fields; the body size shown in the dialog is computed client-side from the body.

Implementation notes

The new payload and the new event subtypes are an additive wire change that requires a lock-step server and client upgrade: an older client cannot decode the new event subtypes.
A capture never affects the transcript: the repository treats both new events as diagnostic no-ops that leave the cached session untouched.
The capture must never break the completion it describes, so a non-cancellation failure inside the observer is logged and swallowed while cancellation is rethrown.
The body is forwarded verbatim and is not masked; all current strategies send a pre-serialized body that contains no credential.

Validation

./gradlew :server:testClasses :server:test
./gradlew :app:desktopMainClasses :app:desktopTestClasses :app:desktopTest

Risks and follow-up

Android and WebAssembly targets were not compiled in the validation environment; no exhaustive when over the touched sealed types exists in androidMain or wasmJsMain.
Follow-up candidates deliberately left out of this change: the mapper KDoc overstates body masking, the accepted/refused-send and tool-loop paths have no dedicated tests, and the copied body relies on Any.toString() for non-string bodies.

## Summary Lets the owner of a chat session inspect the exact provider request the server prepared for the latest assistant step, so a surprising reply can be traced back to what was actually sent. ## Changes - Added a provider-neutral `RawLlmRequestPayload` (method, url) and a `raw_llm_request` informational event on both the non-streaming and streaming chat surfaces. A step that issues several requests emits one event per request, and the latest one wins. - Captured the prepared request in the LLM client path through an `onRequestPrepared` observer, forwarded through the retrying decorator, the turn runners and the transport mappers; the streaming client is now suspending so the observer runs in the turn's own coroutine before dispatch. - Masked the credential before any event exists: the URL and every header value get the same `min(8, length/2)`-prefix masking, a short key becomes `****`, and no `Authorization` header is synthesized for a keyless provider. - Added an app-side capture slot with overwrite semantics, cleared by a state reset and by an accepted send, plus a read-only dialog with an explicit empty state, a Copy action and a top-bar menu entry. - Copied document contains only the four wire fields; the body size shown in the dialog is computed client-side from the body. ## Implementation notes - The new payload and the new event subtypes are an additive wire change that requires a lock-step server and client upgrade: an older client cannot decode the new event subtypes. - A capture never affects the transcript: the repository treats both new events as diagnostic no-ops that leave the cached session untouched. - The capture must never break the completion it describes, so a non-cancellation failure inside the observer is logged and swallowed while cancellation is rethrown. - The body is forwarded verbatim and is not masked; all current strategies send a pre-serialized body that contains no credential. ## Validation - `./gradlew :server:testClasses :server:test` - `./gradlew :app:desktopMainClasses :app:desktopTestClasses :app:desktopTest` ## Risks and follow-up - Android and WebAssembly targets were not compiled in the validation environment; no exhaustive `when` over the touched sealed types exists in `androidMain` or `wasmJsMain`. - Follow-up candidates deliberately left out of this change: the mapper KDoc overstates body masking, the accepted/refused-send and tool-loop paths have no dedicated tests, and the copied body relies on `Any.toString()` for non-string bodies.