Summary

Implements #2381 — show intent summary before file approval prompt.

When the model invokes write/modify/delete tools, its preceding text content is extracted as an "intent summary" and displayed in the approval view. Users can now see why a change is being made before reviewing what will change, addressing the "blind approval" problem described in the issue.

Before (current flow)

Agent analyzes → triggers write_file → approval prompt appears → user approves blind

After (new flow)

Agent analyzes → explains intent in text → approval prompt shows intent + diff → user decides with context

Motivation

The current approval prompt shows the diff (what will change) but not the reasoning behind it. Users are asked to approve changes they haven't been given context for. This PR adds a lightweight, non-blocking intent summary layer:

1. Do I agree with this approach? (intent summary)
2. Does this diff match what I agreed to? (existing approval)

Changes

How it works

Engine layer (turn_loop.rs)

After building all ToolExecutionPlans for a turn, the engine checks:

• Does this turn contain write tools? (!read_only && approval_required)
• Did the model produce text content before the tool calls? (current_text_visible)

If both are true, the text is passed as intent_summary in the ApprovalRequired event. If the model jumped straight to tool calls without explanation, intent_summary is None and the approval proceeds normally.

TUI layer (widgets/mod.rs)

The approval widget renders the intent summary above the tool parameters:

  [DESTRUCTIVE]  write_file
  Type: File modification
  About: Write content to src/main.rs
  Impact: Overwrites existing file

  Intent: I'm adding a new error handler for the timeout case.
  The current code silently drops the error, which makes
  debugging difficult in production.

  Params: {"path": "src/main.rs", "content": "..."}

  [1] Approve once
  [2] Approve for session
  [3] Deny

i18n

• English: Intent:
• Chinese (zh-Hans): 意图：
• Remaining lines indicator: … (+N lines) / … (还有 N 行)

Design decisions

• Non-blocking: If the model has no explanation, the approval still proceeds. No extra round-trip, no token cost, no latency.
• Zero overhead: No additional LLM calls. Reuses the text the model already generated.
• Backward compatible: YOLO mode (auto-approve) bypasses the approval view entirely, so intent summary is never shown. Approval cache works as-is.
• Graceful degradation: In narrow terminals where max_width == 0, the intent section is skipped entirely.
• No new dependencies: Uses existing truncate_with_ellipsis utility.

Testing

• All existing approval tests pass
• Build passes with RUSTFLAGS="-Dwarnings" (no dead code warnings)
• cargo fmt --check passes
• Pre-existing test failures (non_recoverable_engine_error_enters_offline_mode, session_id_captured_from_post_response_and_replayed) are unrelated and fail on main

Closes #2381

sequenceDiagram
    participant M as Model
    participant E as Engine (turn_loop.rs)
    participant EV as Event Bus (events.rs)
    participant RT as RuntimeThreads
    participant UI as TUI UI (ui.rs)
    participant W as ApprovalWidget (widgets/mod.rs)

    M->>E: Text content (explains intent)
    Note over E: current_text_visible += text
    M->>E: Tool call (write_file / exec_command)
    Note over E: has_write_tools = true
    E->>E: approval_intent_summary(current_text_visible) cap at 2000 chars
    E->>EV: "ApprovalRequired { intent_summary: Some("...") }"
    EV->>RT: "approval.required JSON { intent_summary: "..." }"
    EV->>UI: EngineEvent::ApprovalRequired
    UI->>UI: push_approval_request_view(..., intent_summary)
    UI->>W: ApprovalRequest::new_with_intent(..., intent_summary)
    Note over W: Render intent up to 3 lines + N more indicator
    W-->>M: User approves / denies

_{Reviews (3): Last reviewed commit: "fix(tui): harden approval intent summari..." | Re-trigger Greptile}