dcdpr / jp / 25394478219

65%

Build Type

push

github

Committed by web-flow

Commit Message

feat(cli, render): Role-aware turn headers and unified rendering pipeline (#599)

Replace the plain `---` HR separator between turns with labeled
role-boundary headers. Every turn boundary now shows who is speaking
next: `── alice ──…` for the user (using the `author` field stamped on
the `ChatRequest`, falling back to `user`) and `── jp (anthropic/…) ──…`
for the assistant (using `assistant.name` + the resolved model id).

The new `TurnView` type (`render/turn_view.rs`) owns the chat and
structured sub-renderers and tracks per-turn assistant-header state.
Both `TurnRenderer` (replay, `jp conversation print`) and
`TurnCoordinator` (live streaming) now route their rendering through a
`TurnView`, so role attribution, content-kind transitions, and
structured-output dispatch are consistent across both flows.

On the tool-execution side, the parallel `pending_tool_calls` cache and
three-tuple `streaming_perm_results` stash are replaced by a
`PendingTools` scratchpad and a `build_execution_plan` function that
derives an ordered `ExecutionPlan` by walking the conversation stream.
`Action::ExecuteTools` no longer carries the call list — the shell
derives it from the stream, making "the stream is the source of truth"
hold at the API level rather than by convention.

Three bugs are fixed in the same pass:

- **Resume signal**: picking `Continue` while the LLM stream is still
alive mapped to `InterruptAction::Resume`, but the handler was returning
`LoopAction::Break`, dropping the live `SelectAll` and forcing a
redundant HTTP request. Now returns `LoopAction::Continue`.

- **Sparse plan indices**: when an unavailable tool appeared before an
approved tool in the same turn, the approved executor received plan
index 1 while the internal `results` Vec had length 1, causing an
index-out-of-bounds panic. Fixed by re-basing plan indices to contiguous
local positions inside `execute_with_prompting` and pairing each
response back with its original plan index on output.

... (continued)

Coverage Stats

272 of 324 new or added lines in 10 files covered. (83.95%)

4 existing lines in 4 files now uncovered.

24862 of 38441 relevant lines covered (64.68%)

214.16 hits per line