feat: release terminal conversation surface and stabilize tool/sub-run UX

[whatevertogo]

一、PR 概述

发布 terminal conversation surface，并统一 server / cli / web 的会话读模型，同时补齐工具流、子会话入口和 prompt cache metrics 的整条链路。

---

二、背景 / 动机

• 问题现象：现有 master 缺少完整的 terminal chat surface，conversation surface 在 server / web / cli 之间语义分散；工具调用、子会话入口、prompt metrics 也存在投影和展示不一致。
• 影响范围：terminal client、conversation HTTP/SSE 契约、session-runtime 查询与事件投影、web chat 工具块/子会话视图、prompt metrics 展示。
• 复现方式：在 master 上无法获得完整的 terminal 工作台体验，工具块和子会话相关行为也不稳定。

---

三、改动内容

• 新增 terminal client、read model、renderer、bottom pane、overlay 和 capability 适配，形成完整 CLI 工作台。
• 引入统一的 conversation / terminal HTTP 契约与 server projector，收口 server / cli / web 的对话 surface。
• 修复工具调用链路：实时 tool stream、参数/metadata 透传、结果面板、子会话入口恢复。
• 接入并展示 prompt cache / provider cache 指标，修正 prompt metrics 占位值问题。
• 同步更新 OpenSpec、协议 fixture、测试与文档。

改动类型：

• 新功能
• 缺陷修复
• 重构/整理
• 性能优化
• 文档更新
• 测试补充
• 依赖/配置变更

---

四、实现说明

• 核心思路：以 authoritative conversation surface 为中心，统一 durable/live 投影与多端读模型；terminal 侧采用独立 app/state/render/ui 分层；web 侧保留 block-preserving projection 和 sub-run tree。
• 关键设计/权衡：优先修正协议和投影根因，而不是在前端做临时聚合；对于工具与子会话，保留 metadata 直通并在 reducer/tree 层做幂等与拓扑修复。
• 主要涉及文件/模块：
  • crates/cli/**
  • crates/client/**
  • crates/protocol/src/http/conversation/**
  • crates/protocol/src/http/terminal/**
  • crates/server/src/http/routes/conversation.rs
  • crates/server/src/http/terminal_projection.rs
  • crates/session-runtime/src/query/**
  • crates/session-runtime/src/turn/**
  • frontend/src/lib/api/conversation.ts
  • frontend/src/lib/subRunView.ts
  • frontend/src/components/Chat/**

---

五、行为变化（对外影响）

• 之前是：master 没有完整 terminal chat surface；工具输出、子会话入口和 prompt metrics 在不同端表现不一致，部分能力会丢失或延后出现。
• 现在是：CLI / web / server 共享统一 conversation surface；terminal 可直接使用完整工作台；工具块支持实时结果与子会话入口；prompt metrics 能显示真实缓存相关指标。
• 兼容性影响：存在 conversation / terminal wire contract 与前端投影调整；仓库本身不承诺向后兼容。

---

六、测试与验证

自测结果：

• 已本地验证通过
• 已跑单测
• 已跑集成测试/端到端测试

测试命令：

cargo check --workspace
cargo test -p astrcode-server terminal_projection
cargo test -p astrcode-session-runtime
cd frontend && npm run typecheck
cd frontend && npm run test -- src/lib/api/conversation.test.ts src/lib/subRunView.test.ts src/components/Chat/ToolCallBlock.test.tsx

---

七、Checklist

• 代码已通过 cargo fmt --all -- --check
• 代码已通过 cargo clippy --all-targets --all-features -- -D warnings
• 测试已通过 cargo test --workspace --exclude astrcode
• 前端类型检查已通过 npm run typecheck
• 前端 lint 已通过 npm run lint
• 前端格式检查已通过 npm run format:check
• 已添加必要的注释
• 已更新相关文档
• 无 Breaking Changes（或已明确说明）
• 已关联相关 Issue

---

八、关联 Issue

暂无直接关联 Issue。

Summary by Sourcery

Release a unified terminal conversation surface and align tool, sub-run, and prompt metrics behavior across server, CLI, and web surfaces.

New Features:

• Add richer CLI terminal projection for multi-step turns, including multiple assistant/thinking blocks and tool boundaries.
• Expose tool call inputs, structured outputs, and metadata (including child sessions) throughout terminal, server, and web projections.
• Track and surface prompt cache reuse metrics and provider cache support in stored events and UI displays.

Bug Fixes:

• Ensure assistant/thinking blocks are ordered correctly around tool calls in durable projections and finalized reliably.
• Broadcast live tool start/result events alongside deltas so terminal and web stay in sync with runtime state.
• Prevent empty assistant-final events from being persisted when a step only runs tools without visible output.
• Make conversation snapshot append_block envelopes idempotent by block id and support metadata replacement patches.

Enhancements:

• Group tool stream messages under their parent tool call in the chat UI and render results with JSON/code viewers and nested scroll containment.
• Improve tool call blocks with localized labels, collapsible summaries, structured arg/result rendering, timing/truncation pills, and child-session navigation.
• Propagate prompt provider cache metrics and LLM usage into prompt metric events, and distinguish provider KV cache hit rate from prompt composer reuse in the UI.
• Stabilize sub-run tree reconstruction by tracking spawned agent refs via tool metadata and treating topology changes as full rebuilds.
• Extend terminal tool call blocks with input/metadata fields and add metadata-aware patching across server/CLI/web projections.

Tests:

• Expand Rust and frontend tests to cover tool live event broadcasting, prompt cache metric propagation, conversation append/patch behavior, tool block rendering, and sub-run recovery from spawn metadata.

[sourcery-ai]

Reviewer's Guide

Unifies the conversation/terminal projection across server, CLI, and web while completing the terminal chat surface, fixing tool/sub-run streaming semantics, and wiring real prompt cache metrics end‑to‑end.

Sequence diagram for live tool call streaming and terminal projection

sequenceDiagram
    actor User
    participant CLI as CliClient
    participant Server
    participant SessionRuntime
    participant SessionState
    participant TerminalProj as TerminalDeltaProjector
    participant Frontend

    User->>CLI: invoke tool via chat
    CLI->>Server: POST /conversation ...
    Server->>SessionRuntime: run turn

    Note over SessionRuntime: invoke_single_tool
    SessionRuntime->>SessionState: append durable ToolCallStart event
    SessionRuntime->>SessionState: broadcast_tool_start(turn_id, tool_call)
    SessionState-->>Server: AgentEvent.ToolCallStart (Live)
    Server->>TerminalProj: project ToolCallStart
    TerminalProj->>TerminalProj: start_tool_call(turn_id, tool_call_id, tool_name, input, source=Live)
    TerminalProj->>TerminalProj: split_after_live_tool_boundary() on TurnBlockRefs
    TerminalProj-->>Frontend: TerminalDeltaDto.AppendBlock(ToolCall with input, metadata)

    loop Tool stdout/stderr stream
        SessionRuntime->>SessionState: broadcast_tool_output_delta
        SessionState-->>Server: AgentEvent.ToolCallDelta (Live)
        Server->>TerminalProj: project ToolCallDelta
        TerminalProj-->>Frontend: TerminalDeltaDto.PatchBlock(append tool output)
    end

    SessionRuntime->>SessionState: append durable ToolCallResult event
    SessionRuntime->>SessionState: broadcast_tool_result(turn_id, result{duration_ms, metadata})
    SessionState-->>Server: AgentEvent.ToolCallResult (Live)
    Server->>TerminalProj: project ToolCallResult
    TerminalProj->>TerminalProj: replace_tool_summary
    TerminalProj->>TerminalProj: replace_tool_metadata
    TerminalProj-->>Frontend: TerminalDeltaDto.PatchBlock(ReplaceSummary)
    TerminalProj-->>Frontend: TerminalDeltaDto.PatchBlock(ReplaceMetadata)

    Frontend->>Frontend: applyConversationEnvelope(append_block/patch_block)
    Frontend->>Frontend: projectConversationState -> ToolCallMessage + ToolStreamMessage
    Frontend->>Frontend: MessageList groups ToolStream by toolCallId
    Frontend->>Frontend: ToolCallBlock(streams=[...]) renders grouped tool result, args, child session entry

Loading

Class diagram for updated terminal projection block lifecycle

classDiagram
    class TerminalDeltaProjector {
        -blocks: Vec~TerminalBlockDto~
        -block_index: HashMap~String, usize~
        -turn_blocks: HashMap~String, TurnBlockRefs~
        -tool_blocks: HashMap~String, ToolBlockRefs~
        +append_markdown_streaming_block(turn_id: &str, delta: &str, kind: BlockKind) Vec~TerminalDeltaDto~
        +start_tool_call(turn_id: &str, tool_call_id: &str, tool_name: &str, input: Option~Value~, source: ProjectionSource) Vec~TerminalDeltaDto~
        +append_assistant_message(turn_id: &str, content: &str, reasoning_content: Option~&str~) Vec~TerminalDeltaDto~
        +complete_turn_blocks(turn_id: &str) Vec~TerminalDeltaDto~
        -replace_tool_metadata(index: usize, metadata: &Value) bool
    }

    class TurnBlockRefs {
        +current_thinking: Option~String~
        +current_assistant: Option~String~
        +historical_thinking: Vec~String~
        +historical_assistant: Vec~String~
        +pending_thinking: Vec~String~
        +pending_assistant: Vec~String~
        +thinking_count: usize
        +assistant_count: usize
        +current_or_next_block_id(turn_id: &str, kind: BlockKind) String
        +block_id_for_finalize(turn_id: &str, kind: BlockKind) String
        +split_after_live_tool_boundary()
        +split_after_durable_tool_boundary()
        +all_block_ids() Vec~String~
    }

    class ToolBlockRefs {
        +call_block_id: Option~String~
        +stdout_block_id: Option~String~
        +stderr_block_id: Option~String~
    }

    class ProjectionSource {
        <<enum>>
        Durable
        Live
        +is_durable(self) bool
        +is_live(self) bool
    }

    class TerminalBlockDto {
        <<enum>>
        Thinking
        Assistant
        ToolCall
        ToolStream
        User
    }

    class TerminalToolCallBlockDto {
        +id: String
        +turn_id: Option~String~
        +tool_call_id: Option~String~
        +tool_name: String
        +status: TerminalBlockStatusDto
        +input: Option~Value~
        +summary: Option~String~
        +metadata: Option~Value~
    }

    class TerminalDeltaDto {
        <<enum>>
        AppendBlock
        PatchBlock
        ReplaceAll
    }

    class TerminalBlockPatchDto {
        <<enum>>
        ReplaceSummary(summary: String)
        ReplaceMetadata(metadata: Value)
        SetStatus(status: TerminalBlockStatusDto)
        AppendMarkdown(markdown: String)
    }

    class BlockKind {
        <<enum>>
        Thinking
        Assistant
    }

    TerminalDeltaProjector --> TurnBlockRefs : manages per-turn blocks
    TerminalDeltaProjector --> ToolBlockRefs : manages per-tool blocks
    TerminalDeltaProjector --> ProjectionSource : uses
    TerminalDeltaProjector --> TerminalBlockDto : produces
    TerminalDeltaProjector --> TerminalDeltaDto : emits
    TerminalDeltaDto --> TerminalBlockPatchDto : PatchBlock
    TerminalBlockDto --> TerminalToolCallBlockDto : ToolCall variant
    TurnBlockRefs --> BlockKind

    class Helpers {
        +turn_scoped_block_id(turn_id: &str, role: &str, ordinal: usize) String
    }

    Helpers ..> TurnBlockRefs

Loading

Class diagram for prompt cache metrics propagation

classDiagram
    class PromptBuildCacheMetrics {
        <<struct>>
        +reuse_hits: u32
        +reuse_misses: u32
    }

    class PromptBuildOutput {
        +system_prompt: String
        +system_prompt_blocks: Vec~SystemPromptBlock~
        +cache_metrics: PromptBuildCacheMetrics
        +metadata: Value
    }

    class PromptProvider {
        <<trait>>
        +build_prompt(request: PromptBuildRequest) PromptBuildOutput
    }

    class ComposerPromptProvider {
        +build_prompt(request: PromptBuildRequest) PromptBuildOutput
        -summarize_prompt_cache_metrics(output: PromptBuildOutput) PromptBuildCacheMetrics
    }

    class KernelGateway {
        +supports_cache_metrics() bool
    }

    class AssemblePromptRequest {
        +gateway: KernelGateway
        +step_index: usize
        +prompt_facts_provider: PromptFactsProvider
        +session_state: SessionState
    }

    class prompt_metrics_event {
        +prompt_metrics_event(turn_id: &str, agent: &AgentEventContext, step_index: usize, snapshot: PromptTokenSnapshot, truncated_tool_results: usize, cache_metrics: PromptBuildCacheMetrics, provider_cache_metrics_supported: bool) StorageEvent
    }

    class apply_prompt_metrics_usage {
        +apply_prompt_metrics_usage(events: &mut [StorageEvent], step_index: usize, usage: Option~LlmUsage~)
    }

    class StorageEvent {
        +turn_id: Option~String~
        +payload: StorageEventPayload
    }

    class StorageEventPayload {
        <<enum>>
        +PromptMetrics_variant: PromptMetrics
        +other_variants: ...
    }

    class PromptMetrics {
        +step_index: u32
        +provider_input_tokens: Option~u32~
        +provider_output_tokens: Option~u32~
        +cache_creation_input_tokens: Option~u32~
        +cache_read_input_tokens: Option~u32~
        +provider_cache_metrics_supported: bool
        +prompt_cache_reuse_hits: u32
        +prompt_cache_reuse_misses: u32
        +truncated_tool_results: u32
    }

    class LlmUsage {
        +input_tokens: u32
        +output_tokens: u32
        +cache_creation_input_tokens: u32
        +cache_read_input_tokens: u32
    }

    class RunStep {
        +run_single_step_with(...)
    }

    PromptProvider <|.. ComposerPromptProvider
    ComposerPromptProvider --> PromptBuildOutput : returns
    PromptBuildOutput --> PromptBuildCacheMetrics

    AssemblePromptRequest ..> KernelGateway
    AssemblePromptRequest ..> PromptProvider
    AssemblePromptRequest ..> prompt_metrics_event

    prompt_metrics_event --> StorageEvent
    StorageEvent --> StorageEventPayload
    StorageEventPayload --> PromptMetrics

    RunStep ..> apply_prompt_metrics_usage : after LLM output
    apply_prompt_metrics_usage --> StorageEvent : mutates PromptMetrics
    apply_prompt_metrics_usage ..> LlmUsage

    class FrontendPromptMetricsMessage {
        +providerInputTokens?: number
        +providerOutputTokens?: number
        +cacheCreationInputTokens?: number
        +cacheReadInputTokens?: number
        +providerCacheMetricsSupported: boolean
        +promptCacheReuseHits: number
        +promptCacheReuseMisses: number
    }

    class calculateCacheHitRatePercent {
        +calculateCacheHitRatePercent(metrics: PromptMetricsMessage) number | null
    }

    class calculatePromptReuseRatePercent {
        +calculatePromptReuseRatePercent(metrics: PromptMetricsMessage) number | null
    }

    StorageEventPayload --> FrontendPromptMetricsMessage : via HTTP projection
    calculateCacheHitRatePercent ..> FrontendPromptMetricsMessage
    calculatePromptReuseRatePercent ..> FrontendPromptMetricsMessage

Loading

File-Level Changes

Change	Details	Files
Stabilize terminal delta projection so assistant/thinking blocks and tool calls are correctly ordered across live vs durable streams and can carry tool input/metadata.	Refactors TurnBlockRefs to track current, pending, and historical assistant/thinking blocks with per-turn ordinal IDs. Derives block IDs lazily via turn_scoped_block_id and splits blocks at live/durable tool boundaries to preserve ordering around tool calls. Adds support for ToolCallStart input and ToolCallResult metadata in terminal projections, including ReplaceMetadata patches and tests for multi-step turns.	crates/server/src/http/terminal_projection.rs crates/protocol/src/http/terminal/v1.rs crates/protocol/tests/terminal_conformance.rs crates/protocol/tests/fixtures/terminal/v1/snapshot.json
Broadcast full tool lifecycle events (start, streaming, result) over the live channel and project them consistently into conversation state and UI.	Emits ToolCallStart and ToolCallResult live events from the tool cycle with duration attached, updating tests for both immediate and buffered modes. Treats append_block envelopes as idempotent upserts keyed by block id and supports replace_metadata patches on tool call blocks in the conversation read model. Ensures CLI and web projections accept tool input/metadata patches and update existing tool call blocks instead of duplicating them.	crates/session-runtime/src/turn/tool_cycle.rs crates/session-runtime/src/turn/events.rs crates/session-runtime/src/turn/request.rs crates/cli/src/state/conversation.rs frontend/src/lib/api/conversation.ts frontend/src/lib/api/conversation.test.ts
Improve tool call and tool stream UX in the web chat, including grouping streams under calls, rich JSON/text surfaces, and child session navigation.	Rewrites ToolCallBlock to be a collapsible panel that shows a formatted summary, grouped ToolStreamMessage results, structured JSON views, and argument/metadata pills. Groups consecutive toolStream messages by toolCallId in MessageList and passes them into ToolCallBlock instead of rendering streams separately. Introduces ToolCodePanel, nested scroll containment hook, and extends ToolStreamBlock and ToolJsonView to share styling and support structured JSON rendering. Adds tests for ToolCallBlock rendering of args, grouped output, and child session navigation button.	frontend/src/components/Chat/ToolCallBlock.tsx frontend/src/components/Chat/ToolCallBlock.test.tsx frontend/src/components/Chat/MessageList.tsx frontend/src/components/Chat/ToolStreamBlock.tsx frontend/src/components/Chat/ToolJsonView.tsx frontend/src/components/Chat/ToolCodePanel.tsx frontend/src/components/Chat/useNestedScrollContainment.ts
Make the sub-run tree resilient to missing lifecycle events and late-arriving spawn metadata so nested sub-runs are still discoverable.	Extends message fingerprints and remapping logic to treat changes that affect sub-run topology (including spawn metadata) as non-patchable, forcing rebuild when needed. Builds the sub-run index by walking all messages (not only roots) and populates parentSubRunId/childSessionId from spawn metadata when lifecycle events are missing. Adds tests for recovering nested spawned sub-runs from tool metadata and for updating the sub-run tree when spawn metadata arrives later on the same tool call.	frontend/src/lib/subRunView.ts frontend/src/lib/subRunView.test.ts
Wire prompt build cache metrics through the stack and expose them in prompt metrics events and UI.	Introduces PromptBuildCacheMetrics and adds it to PromptBuildOutput; adapters and test providers now populate/cache default metrics. Emits prompt_metrics_event with reuse hits/misses and provider_cache_metrics_supported, and backfills provider cache usage via apply_prompt_metrics_usage when LLM usage arrives. Kernel gateway exposes supports_cache_metrics and assemble_prompt_request passes cache metrics and support flags into prompt metrics events, with tests covering reuse counts. Extends cache hit calculation utilities and prompt metrics UI to distinguish provider KV cache hit rate from prompt composer reuse rate and display both in assistant and metrics messages.	crates/core/src/ports.rs crates/core/src/lib.rs crates/adapter-prompt/src/core_port.rs crates/session-runtime/src/turn/events.rs crates/session-runtime/src/turn/request.rs crates/session-runtime/src/turn/runner/step/mod.rs crates/session-runtime/src/turn/runner/step/tests.rs crates/session-runtime/src/turn/test_support.rs crates/kernel/src/gateway/mod.rs frontend/src/components/Chat/AssistantMessage.tsx frontend/src/components/Chat/PromptMetricsMessage.tsx frontend/src/lib/utils.ts
Tighten assistant output persistence semantics so empty assistant messages from tool-only steps are not stored.	Derives reasoning content and signature once, checks for non-empty assistant or reasoning content, and only emits AssistantFinal events when there is persistable content. Updates tests to assert that tool-only interrupted steps do not record AssistantFinal events.	crates/session-runtime/src/turn/runner/step/mod.rs crates/session-runtime/src/turn/runner/step/tests.rs

---

Tips and commands

Interacting with Sourcery

• Trigger a new review: Comment @sourcery-ai review on the pull request.
• Continue discussions: Reply directly to Sourcery's review comments.
• Generate a GitHub issue from a review comment: Ask Sourcery to create an
  issue from a review comment by replying to it. You can also reply to a
  review comment with @sourcery-ai issue to create an issue from it.
• Generate a pull request title: Write @sourcery-ai anywhere in the pull
  request title to generate a title at any time. You can also comment
  @sourcery-ai title on the pull request to (re-)generate the title at any time.
• Generate a pull request summary: Write @sourcery-ai summary anywhere in
  the pull request body to generate a PR summary at any time exactly where you
  want it. You can also comment @sourcery-ai summary on the pull request to
  (re-)generate the summary at any time.
• Generate reviewer's guide: Comment @sourcery-ai guide on the pull
  request to (re-)generate the reviewer's guide at any time.
• Resolve all Sourcery comments: Comment @sourcery-ai resolve on the
  pull request to resolve all Sourcery comments. Useful if you've already
  addressed all the comments and don't want to see them anymore.
• Dismiss all Sourcery reviews: Comment @sourcery-ai dismiss on the pull
  request to dismiss all existing Sourcery reviews. Especially useful if you
  want to start fresh with a new review - don't forget to comment
  @sourcery-ai review to trigger a new review!

Customizing Your Experience

Access your dashboard to:

• Enable or disable review features such as the Sourcery-generated pull request
  summary, the reviewer's guide, and others.
• Change the review language.
• Add, remove or edit custom review instructions.
• Adjust other review settings.

Getting Help

• Contact our support team for questions or feedback.
• Visit our documentation for detailed guides and information.
• Keep in touch with the Sourcery team by following us on X/Twitter, LinkedIn or GitHub.

[sourcery-ai]

Hey - I've found 1 issue, and left some high level feedback:

• In TurnBlockRefs::block_id_for_finalize you use pending_*: Vec<String> with remove(0), which is O(n); if a turn can accumulate many pending blocks, consider switching these to VecDeque and using pop_front to avoid repeated shifting.
• The new useNestedScrollContainment hook is attached to multiple scrollable components; to avoid subtle UX issues, you might want to guard against horizontal wheel events (e.g. trackpad two‑finger scroll) or modifier-key scrolling so they can still bubble to the outer container when appropriate.

Prompt for AI Agents

Please address the comments from this code review:

## Overall Comments
- In `TurnBlockRefs::block_id_for_finalize` you use `pending_*: Vec<String>` with `remove(0)`, which is O(n); if a turn can accumulate many pending blocks, consider switching these to `VecDeque` and using `pop_front` to avoid repeated shifting.
- The new `useNestedScrollContainment` hook is attached to multiple scrollable components; to avoid subtle UX issues, you might want to guard against horizontal wheel events (e.g. trackpad two‑finger scroll) or modifier-key scrolling so they can still bubble to the outer container when appropriate.

## Individual Comments

### Comment 1
<location path="frontend/src/components/Chat/useNestedScrollContainment.ts" line_range="11-20" />
<code_context>
+      return;
+    }
+
+    const onWheel = (event: WheelEvent) => {
+      const canScroll = container.scrollHeight > container.clientHeight + 1;
+      if (!canScroll) {
+        return;
+      }
+
+      const atTop = container.scrollTop <= 0 && event.deltaY < 0;
+      const atBottom =
+        container.scrollTop + container.clientHeight >= container.scrollHeight - 1 &&
+        event.deltaY > 0;
+
+      if (!atTop && !atBottom) {
+        event.stopPropagation();
+        return;
</code_context>
<issue_to_address>
**issue (bug_risk):** Scroll containment prevents parent scrolling at container boundaries due to `preventDefault`.

Because `event.preventDefault()` is still called when `atTop` or `atBottom`, neither the inner nor outer container scrolls at the edges, making the scroll feel stuck. Keep the current `stopPropagation` behavior in the middle of the content, but at the top/bottom let the event bubble (no `preventDefault`) so the parent can scroll. Adjust the conditions so `preventDefault` only applies while the inner container can continue scrolling, not when it has reached its bounds.
</issue_to_address>

---

Sourcery is free for open source - if you like our reviews please consider sharing them ✨

• X
• Mastodon
• LinkedIn
• Facebook

_{Help me be more useful! Please click 👍 or 👎 on each comment and I'll use the feedback to improve your reviews.}

[Copilot]

Pull request overview

This PR releases a complete “terminal conversation surface” and aligns the read-model/projection semantics across server/CLI/web, with a focus on stabilizing tool streaming, sub-run navigation, and prompt/provider cache metrics end-to-end.

Changes:

• Frontend: improve tool-call rendering (group tool streams, richer JSON/code surfaces, child-session entry) and fix prompt metrics display (provider KV cache + prompt reuse).
• Session/runtime & adapters: broadcast tool start/result live events and plumb prompt cache reuse metrics into PromptMetrics events.
• Protocol/server: extend terminal/conversation contracts to carry tool input/metadata and update terminal projection + fixtures/tests accordingly.

Reviewed changes

Copilot reviewed 35 out of 35 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
frontend/src/lib/utils.ts	Refines provider cache hit rate calculation and adds prompt reuse rate helper.
frontend/src/lib/toolDisplay.ts	Extracts child-session/sub-run navigation targets from tool metadata.
frontend/src/lib/subRunView.ts	Improves sub-run topology/fingerprint handling to recover nested spawned sub-runs.
frontend/src/lib/subRunView.test.ts	Adds coverage for recovering nested spawned sub-runs and fingerprint updates.
frontend/src/lib/api/conversation.ts	Projects tool args/metadata and makes append_block idempotent + adds replace_metadata patch handling.
frontend/src/lib/api/conversation.test.ts	Adds tests for append_block upsert behavior and replace_metadata patching.
frontend/src/components/Chat/useNestedScrollContainment.ts	Adds wheel-event containment hook for nested scroll regions.
frontend/src/components/Chat/ToolStreamBlock.tsx	Renders stdout tool stream as JSON when possible; otherwise uses a shared code panel.
frontend/src/components/Chat/ToolJsonView.tsx	Adds scroll modes/default open behavior and adjusts JSON rendering.
frontend/src/components/Chat/ToolCodePanel.tsx	Introduces a reusable code panel surface for tool outputs.
frontend/src/components/Chat/ToolCallBlock.tsx	Adds collapsible tool block UI, stream grouping, args/metadata surfaces, and child-session navigation.
frontend/src/components/Chat/ToolCallBlock.test.tsx	Adds SSR-based tests for tool block summary/body rendering and child-session CTA.
frontend/src/components/Chat/PromptMetricsMessage.tsx	Displays provider cache and prompt reuse metrics in prompt metrics card.
frontend/src/components/Chat/MessageList.tsx	Groups tool streams under tool calls during rendering.
frontend/src/components/Chat/AssistantMessage.tsx	Updates cache indicator logic to prefer provider KV cache and fallback to prompt reuse.
crates/session-runtime/src/turn/tool_cycle.rs	Broadcasts ToolCallStart/ToolCallResult over live channel; updates tests accordingly.
crates/session-runtime/src/turn/test_support.rs	Updates noop prompt provider to include cache_metrics; removes an assertion helper.
crates/session-runtime/src/turn/submit.rs	Updates test prompt provider output shape with cache_metrics.
crates/session-runtime/src/turn/runner/step/tests.rs	Adjusts step fixtures + asserts interrupted tool-only steps don’t persist empty AssistantFinal.
crates/session-runtime/src/turn/runner/step/mod.rs	Backfills prompt metrics with LLM usage and avoids persisting empty assistant finals.
crates/session-runtime/src/turn/request.rs	Emits prompt cache reuse counts + provider cache support flag; adds test coverage.
crates/session-runtime/src/turn/manual_compact.rs	Updates test prompt provider output shape with cache_metrics.
crates/session-runtime/src/turn/events.rs	Adds prompt cache reuse counts and backfills provider cache token usage into prompt metrics events.
crates/server/src/http/terminal_projection.rs	Supports multi-block assistant/thinking across tool boundaries; carries tool input/metadata into terminal blocks.
crates/server/src/bootstrap/capabilities.rs	Updates test prompt provider output shape with cache_metrics.
crates/protocol/tests/terminal_conformance.rs	Updates terminal snapshot conformance to include tool input field.
crates/protocol/tests/fixtures/terminal/v1/snapshot.json	Fixture update: tool_call blocks include input.
crates/protocol/src/http/terminal/v1.rs	Extends terminal tool_call DTO + adds replace_metadata patch.
crates/kernel/src/gateway/mod.rs	Exposes supports_cache_metrics on KernelGateway for prompt metrics flagging.
crates/core/src/ports.rs	Introduces PromptBuildCacheMetrics and adds cache_metrics to PromptBuildOutput.
crates/core/src/lib.rs	Re-exports PromptBuildCacheMetrics.
crates/cli/src/state/conversation.rs	Handles ReplaceMetadata patch for tool call blocks.
crates/application/src/agent/test_support.rs	Updates test prompt provider output shape with cache_metrics.
crates/adapter-prompt/src/core_port.rs	Summarizes prompt cache reuse hits/misses into PromptBuildOutput.cache_metrics.
crates/adapter-llm/src/anthropic.rs	Adds helper clarifying which system layers are cacheable.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

useNestedScrollContainment calls event.preventDefault() when the nested scroller is already at the top/bottom. This can prevent the parent scroll container from scrolling, effectively trapping the user when they wheel past the boundary. Consider only stopping propagation while the element can still scroll, and at boundaries allow the event to bubble (or rely on CSS overscroll-behavior: contain for bounce control) instead of preventing default.

Suggested change

	return;
	}
	event.preventDefault();
	}

[Copilot]

renderPrimitiveValue now renders full string values with no truncation/preview. Large tool outputs (e.g., long file contents encoded in JSON) can create huge DOM/text nodes and noticeably degrade render performance. Consider restoring a length-limited preview with an expand action, or collapsing long strings behind a <details>.

[whatevertogo]

@codex review it

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5c0ad3d74c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".