Extract WTA prompt building and turn metrics

[vanzue]

Summary

• extract ACP prompt templates, assembly, pane/context resolution, and prompt audit logging into focused prompt modules
• extract concurrent per-turn timing, ACP timing logs, and response-latency telemetry into turn_metrics.rs
• establish one-way dependencies from client to prompt_builder to prompt_context, while keeping callbacks and session orchestration in client.rs
• reduce client.rs by roughly 1,670 lines and update the architecture refactor plan

Validation

• cargo test --manifest-path tools/wta/Cargo.toml --quiet (1213 passed)
• cargo build --manifest-path tools/wta/Cargo.toml --quiet
• manual Debug Intelligent Terminal smoke testing of normal prompts, pane context, autofix, multi-tab routing, and permission/tool-call sequencing
• cargo clippy --manifest-path tools/wta/Cargo.toml --all-targets --quiet remains blocked by the pre-existing clippy::never_loop error in src/ui/recommendations.rs; this PR does not modify that file

[Copilot]

Pull request overview

This PR refactors the WTA ACP client implementation by extracting prompt construction/context resolution and per-turn timing/telemetry into dedicated modules, reducing client.rs size and clarifying module dependencies within tools/wta/src/protocol/acp/.

Changes:

• Added prompt_builder.rs to centralize prompt template selection, per-session template memoization, prompt assembly, and per-turn audit logging.
• Expanded prompt_context.rs to own pane/context resolution (active vs. source pane for autofix), shell identity detection, and shared terminal-buffer capture helpers/providers.
• Added turn_metrics.rs to own prompt timing state, timing logs, and response-latency telemetry; updated call sites and tests accordingly.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
tools/wta/src/protocol/acp/turn_metrics.rs	New module for per-turn timing state/logs and response-latency telemetry.
tools/wta/src/protocol/acp/prompt_context.rs	Moves terminal context assembly and pane/shell resolution helpers into a focused module with tests.
tools/wta/src/protocol/acp/prompt_builder.rs	New module for prompt template memoization, prompt assembly, and turn audit logging.
tools/wta/src/protocol/acp/mod.rs	Wires new prompt/timing modules into the ACP module tree.
tools/wta/src/protocol/acp/mock_agent_tests.rs	Updates imports to use the new prompt/timing modules.
tools/wta/src/protocol/acp/client.rs	Switches prompt assembly/timing responsibilities to the extracted modules.
tools/wta/src/app.rs	Updates import of prompt_timing_log to the new module.
doc/specs/wta-architecture-refactor.md	Updates the refactor plan/status to reflect the new module boundaries.

[Copilot]

Pull request overview

Copilot reviewed 8 out of 8 changed files in this pull request and generated no new comments.

Comments suppressed due to low confidence (2)

tools/wta/src/protocol/acp/prompt_builder.rs:243

• log_turn_trace calls prompt_text.chars().count() to decide whether to include a tail snippet. Since prompts can be large (template + captured context), this forces an O(n) full scan of the prompt on every turn just for logging. You can keep the same behavior with a bounded scan by checking whether there is any character after HEAD_LEN + TAIL_LEN.

    let tail = if prompt_text.chars().count() > HEAD_LEN + TAIL_LEN {
        snippet(prompt_text, TAIL_LEN, false)
    } else {
        String::new()
    };

tools/wta/src/protocol/acp/prompt_context.rs:38

• truncate_for_prompt always allocates truncated and then (when the input is already within budget) allocates again via text.to_string(). Returning truncated in the non-truncation case avoids an extra allocation/copy on every call.

    if chars.next().is_none() {
        text.to_string()
    } else {
        format!("{truncated}\n...<truncated>")
    }

[Copilot]

Pull request overview

Copilot reviewed 8 out of 8 changed files in this pull request and generated no new comments.

Comments suppressed due to low confidence (2)

tools/wta/src/protocol/acp/prompt_context.rs:38

• truncate_for_prompt always allocates truncated, and when the input is already within budget it allocates again via text.to_string(). This doubles the work on the common non-truncation path; you can return truncated directly (it already contains the full text when no extra char remains).

fn truncate_for_prompt(text: &str, max_chars: usize) -> String {
    let mut chars = text.chars();
    let truncated: String = chars.by_ref().take(max_chars).collect();
    if chars.next().is_none() {
        text.to_string()
    } else {
        format!("{truncated}\n...<truncated>")
    }

tools/wta/src/protocol/acp/prompt_builder.rs:243

• log_turn_trace uses prompt_text.chars().count() to decide whether to include a tail snippet, which scans the entire (potentially large) prompt every turn. You can make this check bounded by looking for the (HEAD_LEN+TAIL_LEN+1)th character via skip(...).next() instead.

    let head = snippet(prompt_text, HEAD_LEN, true);
    let tail = if prompt_text.chars().count() > HEAD_LEN + TAIL_LEN {
        snippet(prompt_text, TAIL_LEN, false)
    } else {
        String::new()
    };