Verify each finding against the current code and only fix it if needed.

Outside diff comments:
In `@app/src/services/chatService.ts`:
- Around line 143-186: Add namespaced debug logs inside each progress-event
callback (the callbacks registered for listeners.onInferenceStart,
onIterationStart, onToolCall, onToolResult, onSubagentSpawned, and the
onSubagentDone handlers) so every received payload is logged before forwarding;
include grep-friendly prefix (e.g. "[realtime][chat]"), and correlation fields
thread_id, request_id, round, event (use EVENTS.* symbol name) and success
(boolean) in the log entry. Ensure logs are emitted only in dev/debug builds if
needed (wrap with the app's dev-check) and keep the logging placement in the
existing cb/onCompleted/onFailed functions so sequencing/cancellation can be
traced end-to-end.

In `@src/openhuman/channels/providers/web.rs`:
- Around line 489-500: Remove the dead helper by deleting the parse_tool_args
function (fn parse_tool_args(arguments: &str) -> Value) from
src/openhuman/channels/providers/web.rs and also remove the associated unit
test(s) that only reference this helper; search for references to
parse_tool_args in the repo and ensure any test file (or #[cfg(test)] block)
that solely exists to exercise this function is removed or refactored to no
longer call it, then run cargo test to confirm no remaining references.

---

Nitpick comments:
In `@src/openhuman/channels/providers/web.rs`:
- Around line 396-406: The output field currently builds JSON via string
formatting (output: Some(format!(...))) which is fragile; replace that with
serde_json::json! to construct a proper JSON value (e.g., json!({"output_chars":
output_chars, "elapsed_ms": elapsed_ms})) and then serialize it to a String
(to_string() or serde_json::to_string()) before assigning to the output field;
update the import to use serde_json::json if not already present and ensure this
change is applied where the struct is constructed (the block setting skill_id,
args, output, success, round, etc.).

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@src/openhuman/agent/harness/session/turn.rs`:
- Around line 754-758: The emit_progress method currently discards errors from
tx.try_send(event) which can hide Full or Closed failures; update emit_progress
(and the on_progress send path) to capture the Result from tx.try_send and log
failures including the error and the event context (use the existing logger or
processLogger), distinguishing Closed vs Full so Closed logs at warning/error
and Full logs at debug/info; optionally consider backing off and retrying
critical events (e.g., tool completion) or switching to a bounded buffer with
drop policy, but minimally ensure every try_send error is logged with error
details and the AgentProgress identifier in emit_progress.

In `@src/openhuman/channels/providers/web.rs`:
- Around line 264-274: The progress bridge currently drains buffered progress_rx
after the inference task is cancelled, causing stale events to be published for
a cancelled request; modify the code that creates and runs the bridge so it can
be canceled or validated: have spawn_progress_bridge return a JoinHandle and
store it alongside the main inference task (or return a cancellable future) so
cancel_chat can abort the bridge handle, or add a request validation check
inside the bridge’s publishing loop that compares the captured request_id
against the current in-flight request before emitting; update
agent.set_on_progress usage to ensure progress_tx is closed/removed when the
bridge is aborted to avoid draining leftover events.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@tests/json_rpc_e2e.rs`:
- Around line 491-534: The SSE parsing currently decodes each incoming chunk
with from_utf8 which can drop bytes when a multibyte UTF-8 sequence is split
across chunk boundaries and only looks for "\n\n" frames; change
read_sse_event_by_type (and read_first_sse_event) to buffer at the byte level
(Vec<u8>), append each chunk as bytes, search the buffer for either b"\r\n\r\n"
or b"\n\n" to extract complete frame bytes, then decode the extracted frame once
with std::str::from_utf8 (handling the Result rather than silently using ""),
parse lines for "data:" prefixes, and only then json-decode the joined payload;
this preserves UTF-8 integrity across chunk boundaries and accepts both CRLF and
LF SSE frame terminators.

---

Nitpick comments:
In `@tests/json_rpc_e2e.rs`:
- Around line 444-486: The helper read_first_sse_event is unused; either delete
the entire function (and its #[allow(dead_code)] annotation) to remove dead
code, or refactor the parsing logic into a shared helper (e.g.,
parse_sse_payload or parse_sse_events_from_stream) that accepts a bytes stream
or buffered text and returns serde_json::Value, then replace duplicate parsing
in other test paths with calls to that new helper and remove
read_first_sse_event. Ensure references are updated for the function name
read_first_sse_event and that serde_json::from_str error handling is preserved
in the consolidated helper.