feat(cache): project mode prompts per request

[Hmbown]

Harvests PR #2687 by @LeoAlex0 for the v0.9.0 stewardship branch.

This keeps the stable system prompt mode-agnostic and projects mode, approval policy, and tool taxonomy as request-time user-role runtime metadata. That preserves stored history/prefix-cache bytes and avoids appending extra system messages for strict chat-template providers.

Stewardship polish from the source PR:

• Preserves the existing turn-metadata cache tests from the stewardship branch.
• Keeps the replan replay guard at the real <= 2 behavior.
• Tightens cache-inspect coverage to require both deduplicated=false and deduplicated=true tool-result metadata.
• Adds changelog credit for @LeoAlex0.

Issue #2722 remains open as the broader v0.9 harvest tracker.

Verification:

• cargo fmt --all -- --check
• git diff --check origin/codex/v0.9.0-stewardship..HEAD
• python3 scripts/check-coauthor-trailers.py --range origin/codex/v0.9.0-stewardship..HEAD --check-authors
• cargo test -p codewhale-tui --locked --bin codewhale-tui runtime_prompt -- --nocapture
• cargo test -p codewhale-tui --locked --bin codewhale-tui turn_metadata -- --nocapture
• cargo test -p codewhale-tui --locked --bin codewhale-tui cache_inspect -- --nocapture
• cargo test -p codewhale-tui --locked --bin codewhale-tui prompt -- --nocapture
• cargo test -p codewhale-tui --locked --bin codewhale-tui capacity -- --nocapture
• cargo test -p codewhale-tui --locked --bin codewhale-tui post_tool_replay_invoked_when_high_non_severe_risk -- --nocapture
• cargo test -p codewhale-tui --locked --bin codewhale-tui error_escalation_triggers_replan_when_severe_or_repeated_failures -- --nocapture
• cargo clippy -p codewhale-tui --locked --all-targets --all-features -- -D warnings

[greptile-apps]

Hmbown has reached the 50-review limit for trial accounts. To continue receiving code reviews, upgrade your plan.

[gemini-code-assist]

Code Review

This pull request refactors prompt handling by projecting mode, approval, and tool-taxonomy metadata as transient, request-time runtime prompts instead of mutating the stored system prompt, thereby keeping provider prefix caches byte-stable. Feedback on these changes highlights a potential issue where appending the runtime prompt as a separate user message can create consecutive user messages, which violates the alternating role requirements of strict API providers like Anthropic. To resolve this, it is recommended to merge the runtime prompt content into the last message if it is also a user message.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Appending the runtime prompt as a separate user message can result in consecutive user messages in the message list (e.g., when the last message in the session history is also a user message).

Strict API providers (such as Anthropic Claude) and strict Jinja chat templates (used by many open-source models via Ollama, vLLM, Hugging Face, etc.) strictly enforce alternating roles (user/assistant). Sending consecutive user messages will cause validation errors (e.g., 400 Bad Request or TemplateError) and crash the application.

To prevent this, we should check if the last message in the history is a user message (and not a tool result), and if so, merge the runtime prompt's content into that last message's content array. This preserves the prefix cache of all previous turns (since the current turn's message is at the end and is new anyway) while ensuring perfect compatibility with strict providers and templates.

        let mut messages = self.session.messages.clone();
        let runtime_msg = self.runtime_prompt_message();
        let merged = if let Some(last_msg) = messages.last_mut() {
            let is_tool_result = last_msg.content.iter().any(|block| {
                matches!(block, ContentBlock::ToolResult { .. })
            });
            if last_msg.role == "user" && !is_tool_result {
                last_msg.content.extend(runtime_msg.content);
                true
            } else {
                false
            }
        } else {
            false
        };
        if !merged {
            messages.push(runtime_msg);
        }
        messages