feat(prompt): add core tool taxonomy block

[LING71671]

Summary

• prepend a compact core tool taxonomy block to the composed system prompt
• generate taxonomy tool names from the eager native-tool list instead of duplicating a separate prompt-only list
• add prompt tests covering the taxonomy text, eager-tool sync, and locale-bookend guard behavior

Refs #2084

Verification

• cargo test -p codewhale-tui --bin codewhale-tui tool_taxonomy --all-features
• cargo test -p codewhale-tui --bin codewhale-tui compose_prompt --all-features
• cargo test -p codewhale-tui --bin codewhale-tui prompts::tests --all-features
• cargo fmt --all --check
• git diff --check
• cargo clippy -p codewhale-tui --all-targets --all-features

Greptile Summary

This PR prepends a compact ## Core Tool Taxonomy block to the assembled system prompt, grouping core tools into discovery, git-inspection, and verification categories. Tool names are derived at runtime from DEFAULT_ACTIVE_NATIVE_TOOLS (via the new default_active_native_tool_names accessor in engine.rs) so the taxonomy cannot silently diverge from the eager-load list.

• render_core_tool_taxonomy_block correctly returns Option<String> per group, eliminating the malformed-sentence case; run_tests is filtered from Plan-mode prompts via core_taxonomy_tools_for_mode.
• Three new tests cover taxonomy prefix order, Plan/Agent mode differences, and the sync invariant; the CJK guard is now scoped to BASE_PROMPT and each per-mode taxonomy string.

Confidence Score: 5/5

Safe to merge — the taxonomy block is generated entirely from the existing eagerly-loaded tool list, making the prompt change deterministic and self-consistent.

The logic is correct: render_core_tool_group returns Option so empty groups produce no output, the Plan-mode filter is explicit and tested, and the sync test guarantees taxonomy constants stay in step with the engine tool list.

No files require special attention. crates/tui/src/prompts.rs is the main change and its new tests cover the key invariants.

Important Files Changed

Filename	Overview
crates/tui/src/core/engine.rs	Adds a thin pub(crate) accessor that re-exports tool_catalog::DEFAULT_ACTIVE_NATIVE_TOOLS so prompts.rs can reference it without accessing a private sub-module directly. Change is minimal and correct.
crates/tui/src/prompts.rs	Prepends a mode-aware Core Tool Taxonomy block to the composed system prompt; correctly guards empty groups via Option, filters run_tests from Plan mode, and adds three new tests covering taxonomy content, eager-tool sync, and the scoped CJK guard.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[compose_prompt_with_approval_and_model] --> B[render_core_tool_taxonomy_block mode]
    B --> C[core_taxonomy_tools_for_mode mode]
    C --> D[default_active_native_tool_names]
    C -->|Plan mode| E[filter out run_tests]
    B --> F[render_core_tool_group DISCOVERY]
    B --> G[render_core_tool_group GIT]
    B --> H[render_core_tool_group VERIFICATION]
    F -->|Some| I[Use ... for discovery.]
    G -->|Some| J[Use ... for git inspection.]
    H -->|Some if not Plan| K[Use ... for verification.]
    I & J & K --> L[Core Tool Taxonomy block]
    L --> M[parts array assembled]
    M --> N[Final system prompt]

Loading

_{Reviews (2): Last reviewed commit: "fix(prompt): make tool taxonomy mode-awa..." | Re-trigger Greptile}

[gemini-code-assist]

Code Review

This pull request introduces a dynamically generated "Core Tool Taxonomy" block prepended to the system prompt, helping guide the model on which tools to use for discovery, git inspection, and verification. It also includes corresponding unit tests and relaxes a test assertion regarding CJK characters. The review feedback suggests making the taxonomy block dynamic based on the active AppMode (for instance, omitting run_tests in read-only Plan mode) and formatting the output cleanly when certain tool groups are empty.

[gemini-code-assist]

The taxonomy block is currently static and does not account for the active mode. In Plan mode, code-execution tools like run_tests are not registered in the tool catalog to maintain read-only safety. However, the taxonomy prompt will still instruct the model to use run_tests for verification, which can lead to model confusion or hallucinations.

Additionally, if any tool group is empty, the current formatting will produce awkward strings with double spaces (e.g., Use for verification.).

We should update render_core_tool_taxonomy_block to accept the AppMode and dynamically filter out unavailable tools, while cleanly formatting only the non-empty tool groups.

fn render_core_tool_taxonomy_block(mode: AppMode) -> String {
    let core_tools = crate::core::engine::default_active_native_tool_names();
    let discovery = render_core_tool_group(TOOL_TAXONOMY_DISCOVERY, core_tools);
    let git = render_core_tool_group(TOOL_TAXONOMY_GIT, core_tools);
    let verification = if mode == AppMode::Plan {
        String::new()
    } else {
        render_core_tool_group(TOOL_TAXONOMY_VERIFICATION, core_tools)
    };

    let mut parts = Vec::new();
    if !discovery.is_empty() {
        parts.push(format!("Use {discovery} for discovery."));
    }
    if !git.is_empty() {
        parts.push(format!("Use {git} for git inspection."));
    }
    if !verification.is_empty() {
        parts.push(format!("Use {verification} for verification."));
    }

    format!("## Core Tool Taxonomy\n\n{}", parts.join(" "))
}

[gemini-code-assist]

Pass the mode parameter to render_core_tool_taxonomy_block to dynamically filter out tools that are not registered in the current mode (such as run_tests in Plan mode).

Suggested change

	let tool_taxonomy = render_core_tool_taxonomy_block();
	let tool_taxonomy = render_core_tool_taxonomy_block(mode);