Summary

Replaces low-level direct calls to chat modules in acpSession.ts with GeminiClient.sendMessageStream(). This establishes parity with core terminal features—including chat compression, token masking safeguards, context validation limits, and execution tracing.

Details

• Prompt Pipeline Refactoring: Migrates Session.prompt() stream logic from calling GeminiChat.sendMessageStream() to GeminiClient.sendMessageStream().
• Structured Events Mapping: Converts underlying core state notifications (GeminiEventType) to line-delimited protocol interactions format.
• Enhanced Telemetry Tracking: Introduces variable client identifiers distinguishing incoming requests sources (tui, acp, cli-command, or a2a-server).
• Automated Test Suites Extensions: Appends core integration case parameters tests into packages/cli/src/acp/acpSession.test.ts.

Before the change

Prior to this change, ACP interactive turn system bypassed the high-level client lifecycle manager (GeminiClient) and orchestrated low-level calls directly via internal packages/core components.

sequenceDiagram
    participant ACP Session
    participant Routing System
    participant GeminiChat

    ACP Session->>Routing System: router.route(RoutingContext)
    Routing System-->>ACP Session: Returns model target
    ACP Session->>GeminiChat: sendMessageStream(model, request parts)
    loop Stream Processing
        GeminiChat-->>ACP Session: yield Raw StreamEvent.CHUNK
        Note over ACP Session: - Manually parses usage data<br/>- Manually formats token info<br/>- Processes raw candidates content
    end

Pipeline Characteristics

• Direct Interaction: Low-level consumption of raw response streams directly from GeminiChat.sendMessageStream().
• Duplicated Responsibilities: Low-level token limits calculations, candidate conversions, and parsing were written directly inside the Session stream.
• Bypassed Engine Features: ACP turn streams operated outside established core features:
  • Automated session history compression
  • Context optimization / Output text masking
  • Context constraints enforcement (overflow warnings)
  • Process steering (BeforeAgent & AfterAgent orchestration hooks)

After/With this change

With this commit, the interactive turn orchestrator in acpSession.ts was updated to use the high-level API provided by GeminiClient.sendMessageStream().

sequenceDiagram
    participant ACP Session
    participant GeminiClient
    participant Core Orchestrator

    ACP Session->>GeminiClient: sendMessageStream(currentParts)
    opt Client Interceptors
        GeminiClient->>Core Orchestrator: Triggers Hooks & History Management
    end
    loop Engine Events
        GeminiClient-->>ACP Session: yield GeminiEventType (Content, Thought, ToolCall, Finished)
        Note over ACP Session: Maps clean protocol payloads<br/>to frontends (e.g., VS Code IDE)
    end

Pipeline Characteristics

• Abstracted Routing: Bounded turn orchestration is performed implicitly by the GeminiClient pipeline.
• Decoupled Event Model: Stream consumers use high-level events defined in GeminiEventType rather than low-level JSON-derived candidate chunks.
• Feature Parity: ACP flows leverage the identical capabilities driving primary terminal workflows:
  • Loop detection tracking and resolution hooks
  • Safety masking constraints and runtime compression optimizations
  • Trace tracking tagged with explicit contextual identification parameters (clientName set as acp, tui, cli-command, or acp-* IDE extensions)

Summary Comparison

Related Issues

maintaince-gemini-cli#1604

How to Validate

Connect to ACP via Jetbrains and ensure all regular prompt workflows behave as expected.

Pre-Merge Checklist

• [NA] Updated relevant documentation and README (if needed)
• [✅] Added/updated tests (if needed)
• [NA] Noted breaking changes (if any)
• [✅] Validated on required platforms/methods:
  • [✅] MacOS
    • npm run
    • npx
    • Docker
    • Podman
    • Seatbelt
  • Windows
    • npm run
    • npx
    • Docker
  • Linux
    • npm run
    • npx
    • Docker