aptu_core/ai/

provider.rs

// SPDX-License-Identifier: Apache-2.0

//! AI provider trait and shared implementations.
//!
//! Defines the `AiProvider` trait that all AI providers must implement,
//! along with default implementations for shared logic like prompt building,
//! request sending, and response parsing.

use anyhow::{Context, Result};
use async_trait::async_trait;
use regex::Regex;
use reqwest::Client;
use secrecy::SecretString;
use std::sync::LazyLock;
use tracing::{debug, instrument};

use super::AiResponse;
use super::types::{
    ChatCompletionRequest, ChatCompletionResponse, ChatMessage, IssueDetails, ResponseFormat,
    TriageResponse,
};
use crate::history::AiStats;

use super::prompts::{
    build_create_system_prompt, build_pr_label_system_prompt, build_pr_review_system_prompt,
    build_release_notes_system_prompt, build_triage_system_prompt,
};

/// Parses JSON response from AI provider, detecting truncated responses.
///
/// If the JSON parsing fails with an EOF error (indicating the response was cut off),
/// returns a `TruncatedResponse` error that can be retried. Other JSON errors are
/// wrapped as `InvalidAIResponse`.
///
/// # Arguments
///
/// * `text` - The JSON text to parse
/// * `provider` - The name of the AI provider (for error context)
///
/// # Returns
///
/// Parsed value of type T, or an error if parsing fails
fn parse_ai_json<T: serde::de::DeserializeOwned>(text: &str, provider: &str) -> Result<T> {
    match serde_json::from_str::<T>(text) {
        Ok(value) => Ok(value),
        Err(e) => {
            // Check if this is an EOF error (truncated response)
            if e.is_eof() {
                Err(anyhow::anyhow!(
                    crate::error::AptuError::TruncatedResponse {
                        provider: provider.to_string(),
                    }
                ))
            } else {
                Err(anyhow::anyhow!(crate::error::AptuError::InvalidAIResponse(
                    e
                )))
            }
        }
    }
}

/// Maximum length for issue body to stay within token limits.
pub const MAX_BODY_LENGTH: usize = 4000;

/// Maximum number of comments to include in the prompt.
pub const MAX_COMMENTS: usize = 5;

/// Maximum number of files to include in PR review prompt.
pub const MAX_FILES: usize = 20;

/// Maximum total diff size (in characters) for PR review prompt.
pub const MAX_TOTAL_DIFF_SIZE: usize = 50_000;

/// Maximum number of labels to include in the prompt.
pub const MAX_LABELS: usize = 30;

/// Maximum number of milestones to include in the prompt.
pub const MAX_MILESTONES: usize = 10;

/// Maximum characters per file's full content included in the PR review prompt.
/// Content pre-truncated by `fetch_file_contents` may already be within this limit,
/// but the prompt builder applies it as a second safety cap.
pub const MAX_FULL_CONTENT_CHARS: usize = 4_000;

/// Estimated overhead for XML tags, section headers, and schema preamble added by
/// `build_pr_review_user_prompt`. Used to ensure the prompt budget accounts for
/// non-content characters when estimating total prompt size.
const PROMPT_OVERHEAD_CHARS: usize = 1_000;

/// Preamble appended to every user-turn prompt to request a JSON response matching the schema.
const SCHEMA_PREAMBLE: &str = "\n\nRespond with valid JSON matching this schema:\n";

/// Matches `<pull_request>` and `</pull_request>` tags (case-insensitive) used as prompt
/// delimiters. These must be stripped from user-controlled fields to prevent prompt injection.
///
/// The pattern is a fixed literal with no quantifiers or alternation, so `ReDoS` is not a
/// concern: regex engine complexity is O(n) in the input length regardless of content.
static XML_DELIMITERS: LazyLock<Regex> =
    LazyLock::new(|| Regex::new(r"(?i)</?pull_request>").expect("valid regex"));

/// Removes `<pull_request>` / `</pull_request>` XML delimiter tags from a user-supplied
/// string, preventing prompt injection via XML tag smuggling.
///
/// Tags are removed entirely (replaced with empty string) rather than substituted with a
/// placeholder. A visible placeholder such as `[sanitized]` could cause the LLM to reason
/// about the substitution marker itself, which is unnecessary and potentially confusing.
///
/// Nested or malformed XML is not a concern: the only delimiter this code inserts into the
/// prompt is the exact string `<pull_request>` / `</pull_request>` (no attributes, no
/// nesting). Stripping those two fixed forms is sufficient to prevent a user-supplied value
/// from breaking out of the delimiter boundary.
///
/// Applied to all user-controlled fields that appear inside the `<pull_request>` block:
/// `pr.title`, `pr.body`, `file.filename`, `file.status`, and each file's patch content.
fn sanitize_prompt_field(s: &str) -> String {
    XML_DELIMITERS.replace_all(s, "").into_owned()
}

/// AI provider trait for issue triage and creation.
///
/// Defines the interface that all AI providers must implement.
/// Default implementations are provided for shared logic.
#[async_trait]
pub trait AiProvider: Send + Sync {
    /// Returns the name of the provider (e.g., "gemini", "openrouter").
    fn name(&self) -> &str;

    /// Returns the API URL for this provider.
    fn api_url(&self) -> &str;

    /// Returns the environment variable name for the API key.
    fn api_key_env(&self) -> &str;

    /// Returns the HTTP client for making requests.
    fn http_client(&self) -> &Client;

    /// Returns the API key for authentication.
    fn api_key(&self) -> &SecretString;

    /// Returns the model name.
    fn model(&self) -> &str;

    /// Returns the maximum tokens for API responses.
    fn max_tokens(&self) -> u32;

    /// Returns the temperature for API requests.
    fn temperature(&self) -> f32;

    /// Returns the maximum retry attempts for rate-limited requests.
    ///
    /// Default implementation returns 3. Providers can override
    /// to use a different retry limit.
    fn max_attempts(&self) -> u32 {
        3
    }

    /// Returns the circuit breaker for this provider (optional).
    ///
    /// Default implementation returns None. Providers can override
    /// to provide circuit breaker functionality.
    fn circuit_breaker(&self) -> Option<&super::CircuitBreaker> {
        None
    }

    /// Builds HTTP headers for API requests.
    ///
    /// Default implementation includes Authorization and Content-Type headers.
    /// Providers can override to add custom headers.
    fn build_headers(&self) -> reqwest::header::HeaderMap {
        let mut headers = reqwest::header::HeaderMap::new();
        if let Ok(val) = "application/json".parse() {
            headers.insert("Content-Type", val);
        }
        headers
    }

    /// Validates the model configuration.
    ///
    /// Default implementation does nothing. Providers can override
    /// to enforce constraints (e.g., free tier validation).
    fn validate_model(&self) -> Result<()> {
        Ok(())
    }

    /// Returns the custom guidance string for system prompt injection, if set.
    ///
    /// Default implementation returns `None`. Providers that store custom guidance
    /// (e.g., from `AiConfig`) override this to supply it.
    fn custom_guidance(&self) -> Option<&str> {
        None
    }

    /// Sends a chat completion request to the provider's API (HTTP-only, no retry).
    ///
    /// Default implementation handles HTTP headers, error responses (401, 429).
    /// Does not include retry logic - use `send_and_parse()` for retry behavior.
    #[instrument(skip(self, request), fields(provider = self.name(), model = self.model()))]
    async fn send_request_inner(
        &self,
        request: &ChatCompletionRequest,
    ) -> Result<ChatCompletionResponse> {
        use secrecy::ExposeSecret;
        use tracing::warn;

        use crate::error::AptuError;

        let mut req = self.http_client().post(self.api_url());

        // Add Authorization header
        req = req.header(
            "Authorization",
            format!("Bearer {}", self.api_key().expose_secret()),
        );

        // Add custom headers from provider
        for (key, value) in &self.build_headers() {
            req = req.header(key.clone(), value.clone());
        }

        let response = req
            .json(request)
            .send()
            .await
            .context(format!("Failed to send request to {} API", self.name()))?;

        // Check for HTTP errors
        let status = response.status();
        if !status.is_success() {
            if status.as_u16() == 401 {
                anyhow::bail!(
                    "Invalid {} API key. Check your {} environment variable.",
                    self.name(),
                    self.api_key_env()
                );
            } else if status.as_u16() == 429 {
                warn!("Rate limited by {} API", self.name());
                // Parse Retry-After header (seconds), default to 0 if not present
                let retry_after = response
                    .headers()
                    .get("Retry-After")
                    .and_then(|h| h.to_str().ok())
                    .and_then(|s| s.parse::<u64>().ok())
                    .unwrap_or(0);
                debug!(retry_after, "Parsed Retry-After header");
                return Err(AptuError::RateLimited {
                    provider: self.name().to_string(),
                    retry_after,
                }
                .into());
            }
            let error_body = response.text().await.unwrap_or_default();
            anyhow::bail!(
                "{} API error (HTTP {}): {}",
                self.name(),
                status.as_u16(),
                error_body
            );
        }

        // Parse response
        let completion: ChatCompletionResponse = response
            .json()
            .await
            .context(format!("Failed to parse {} API response", self.name()))?;

        Ok(completion)
    }

    /// Sends a chat completion request and parses the response with retry logic.
    ///
    /// This method wraps both HTTP request and JSON parsing in a single retry loop,
    /// allowing truncated responses to be retried. Includes circuit breaker handling.
    ///
    /// # Arguments
    ///
    /// * `request` - The chat completion request to send
    ///
    /// # Returns
    ///
    /// A tuple of (parsed response, stats) extracted from the API response
    ///
    /// # Errors
    ///
    /// Returns an error if:
    /// - API request fails (network, timeout, rate limit)
    /// - Response cannot be parsed as valid JSON (including truncated responses)
    #[instrument(skip(self, request), fields(provider = self.name(), model = self.model()))]
    async fn send_and_parse<T: serde::de::DeserializeOwned + Send>(
        &self,
        request: &ChatCompletionRequest,
    ) -> Result<(T, AiStats)> {
        use tracing::{info, warn};

        use crate::error::AptuError;
        use crate::retry::{extract_retry_after, is_retryable_anyhow};

        // Check circuit breaker before attempting request
        if let Some(cb) = self.circuit_breaker()
            && cb.is_open()
        {
            return Err(AptuError::CircuitOpen.into());
        }

        // Start timing (outside retry loop to measure total time including retries)
        let start = std::time::Instant::now();

        // Custom retry loop that respects retry_after from RateLimited errors
        let mut attempt: u32 = 0;
        let max_attempts: u32 = self.max_attempts();

        // Helper function to avoid closure-in-expression clippy warning
        #[allow(clippy::items_after_statements)]
        async fn try_request<T: serde::de::DeserializeOwned>(
            provider: &(impl AiProvider + ?Sized),
            request: &ChatCompletionRequest,
        ) -> Result<(T, ChatCompletionResponse)> {
            // Send HTTP request
            let completion = provider.send_request_inner(request).await?;

            // Extract message content
            let content = completion
                .choices
                .first()
                .and_then(|c| {
                    c.message
                        .content
                        .clone()
                        .or_else(|| c.message.reasoning.clone())
                })
                .context("No response from AI model")?;

            debug!(response_length = content.len(), "Received AI response");

            // Parse JSON response (inside retry loop, so truncated responses are retried)
            let parsed: T = parse_ai_json(&content, provider.name())?;

            Ok((parsed, completion))
        }

        let (parsed, completion): (T, ChatCompletionResponse) = loop {
            attempt += 1;

            let result = try_request(self, request).await;

            match result {
                Ok(success) => break success,
                Err(err) => {
                    // Check if error is retryable
                    if !is_retryable_anyhow(&err) || attempt >= max_attempts {
                        return Err(err);
                    }

                    // Extract retry_after if present, otherwise use exponential backoff
                    let delay = if let Some(retry_after_duration) = extract_retry_after(&err) {
                        debug!(
                            retry_after_secs = retry_after_duration.as_secs(),
                            "Using Retry-After value from rate limit error"
                        );
                        retry_after_duration
                    } else {
                        // Use exponential backoff with jitter: 1s, 2s, 4s + 0-500ms
                        let backoff_secs = 2_u64.pow(attempt.saturating_sub(1));
                        let jitter_ms = fastrand::u64(0..500);
                        std::time::Duration::from_millis(backoff_secs * 1000 + jitter_ms)
                    };

                    let error_msg = err.to_string();
                    warn!(
                        error = %error_msg,
                        delay_secs = delay.as_secs(),
                        attempt,
                        max_attempts,
                        "Retrying after error"
                    );

                    // Drop err before await to avoid holding non-Send value across await
                    drop(err);
                    tokio::time::sleep(delay).await;
                }
            }
        };

        // Record success in circuit breaker
        if let Some(cb) = self.circuit_breaker() {
            cb.record_success();
        }

        // Calculate duration (total time including any retries)
        #[allow(clippy::cast_possible_truncation)]
        let duration_ms = start.elapsed().as_millis() as u64;

        // Build AI stats from usage info (trust API's cost field)
        let (input_tokens, output_tokens, cost_usd) = if let Some(usage) = completion.usage {
            (usage.prompt_tokens, usage.completion_tokens, usage.cost)
        } else {
            // If no usage info, default to 0
            debug!("No usage information in API response");
            (0, 0, None)
        };

        let ai_stats = AiStats {
            provider: self.name().to_string(),
            model: self.model().to_string(),
            input_tokens,
            output_tokens,
            duration_ms,
            cost_usd,
            fallback_provider: None,
            prompt_chars: 0,
        };

        // Emit structured metrics
        info!(
            duration_ms,
            input_tokens,
            output_tokens,
            cost_usd = ?cost_usd,
            model = %self.model(),
            "AI request completed"
        );

        Ok((parsed, ai_stats))
    }

    /// Analyzes a GitHub issue using the provider's API.
    ///
    /// Returns a structured triage response with summary, labels, questions, duplicates, and usage stats.
    ///
    /// # Arguments
    ///
    /// * `issue` - Issue details to analyze
    ///
    /// # Errors
    ///
    /// Returns an error if:
    /// - API request fails (network, timeout, rate limit)
    /// - Response cannot be parsed as valid JSON
    #[instrument(skip(self, issue), fields(issue_number = issue.number, repo = %format!("{}/{}", issue.owner, issue.repo)))]
    async fn analyze_issue(&self, issue: &IssueDetails) -> Result<AiResponse> {
        debug!(model = %self.model(), "Calling {} API", self.name());

        // Build request
        let system_content = if let Some(override_prompt) =
            super::context::load_system_prompt_override("triage_system").await
        {
            override_prompt
        } else {
            Self::build_system_prompt(self.custom_guidance())
        };

        let request = ChatCompletionRequest {
            model: self.model().to_string(),
            messages: vec![
                ChatMessage {
                    role: "system".to_string(),
                    content: Some(system_content),
                    reasoning: None,
                },
                ChatMessage {
                    role: "user".to_string(),
                    content: Some(Self::build_user_prompt(issue)),
                    reasoning: None,
                },
            ],
            response_format: Some(ResponseFormat {
                format_type: "json_object".to_string(),
                json_schema: None,
            }),
            max_tokens: Some(self.max_tokens()),
            temperature: Some(self.temperature()),
        };

        // Send request and parse JSON with retry logic
        let (triage, ai_stats) = self.send_and_parse::<TriageResponse>(&request).await?;

        debug!(
            input_tokens = ai_stats.input_tokens,
            output_tokens = ai_stats.output_tokens,
            duration_ms = ai_stats.duration_ms,
            cost_usd = ?ai_stats.cost_usd,
            "AI analysis complete"
        );

        Ok(AiResponse {
            triage,
            stats: ai_stats,
        })
    }

    /// Creates a formatted GitHub issue using the provider's API.
    ///
    /// Takes raw issue title and body, formats them using AI (conventional commit style,
    /// structured body), and returns the formatted content with suggested labels.
    ///
    /// # Arguments
    ///
    /// * `title` - Raw issue title from user
    /// * `body` - Raw issue body/description from user
    /// * `repo` - Repository name for context (owner/repo format)
    ///
    /// # Errors
    ///
    /// Returns an error if:
    /// - API request fails (network, timeout, rate limit)
    /// - Response cannot be parsed as valid JSON
    #[instrument(skip(self), fields(repo = %repo))]
    async fn create_issue(
        &self,
        title: &str,
        body: &str,
        repo: &str,
    ) -> Result<(super::types::CreateIssueResponse, AiStats)> {
        debug!(model = %self.model(), "Calling {} API for issue creation", self.name());

        // Build request
        let system_content = if let Some(override_prompt) =
            super::context::load_system_prompt_override("create_system").await
        {
            override_prompt
        } else {
            Self::build_create_system_prompt(self.custom_guidance())
        };

        let request = ChatCompletionRequest {
            model: self.model().to_string(),
            messages: vec![
                ChatMessage {
                    role: "system".to_string(),
                    content: Some(system_content),
                    reasoning: None,
                },
                ChatMessage {
                    role: "user".to_string(),
                    content: Some(Self::build_create_user_prompt(title, body, repo)),
                    reasoning: None,
                },
            ],
            response_format: Some(ResponseFormat {
                format_type: "json_object".to_string(),
                json_schema: None,
            }),
            max_tokens: Some(self.max_tokens()),
            temperature: Some(self.temperature()),
        };

        // Send request and parse JSON with retry logic
        let (create_response, ai_stats) = self
            .send_and_parse::<super::types::CreateIssueResponse>(&request)
            .await?;

        debug!(
            title_len = create_response.formatted_title.len(),
            body_len = create_response.formatted_body.len(),
            labels = create_response.suggested_labels.len(),
            input_tokens = ai_stats.input_tokens,
            output_tokens = ai_stats.output_tokens,
            duration_ms = ai_stats.duration_ms,
            "Issue formatting complete with stats"
        );

        Ok((create_response, ai_stats))
    }

    /// Builds the system prompt for issue triage.
    #[must_use]
    fn build_system_prompt(custom_guidance: Option<&str>) -> String {
        let context = super::context::load_custom_guidance(custom_guidance);
        build_triage_system_prompt(&context)
    }

    /// Builds the user prompt containing the issue details.
    #[must_use]
    fn build_user_prompt(issue: &IssueDetails) -> String {
        use std::fmt::Write;

        let mut prompt = String::new();

        prompt.push_str("<issue_content>\n");
        let _ = writeln!(prompt, "Title: {}\n", issue.title);

        // Truncate body if too long
        let body = if issue.body.len() > MAX_BODY_LENGTH {
            format!(
                "{}...\n[Body truncated - original length: {} chars]",
                &issue.body[..MAX_BODY_LENGTH],
                issue.body.len()
            )
        } else if issue.body.is_empty() {
            "[No description provided]".to_string()
        } else {
            issue.body.clone()
        };
        let _ = writeln!(prompt, "Body:\n{body}\n");

        // Include existing labels
        if !issue.labels.is_empty() {
            let _ = writeln!(prompt, "Existing Labels: {}\n", issue.labels.join(", "));
        }

        // Include recent comments (limited)
        if !issue.comments.is_empty() {
            prompt.push_str("Recent Comments:\n");
            for comment in issue.comments.iter().take(MAX_COMMENTS) {
                let comment_body = if comment.body.len() > 500 {
                    format!("{}...", &comment.body[..500])
                } else {
                    comment.body.clone()
                };
                let _ = writeln!(prompt, "- @{}: {}", comment.author, comment_body);
            }
            prompt.push('\n');
        }

        // Include related issues from search (for context)
        if !issue.repo_context.is_empty() {
            prompt.push_str("Related Issues in Repository (for context):\n");
            for related in issue.repo_context.iter().take(10) {
                let _ = writeln!(
                    prompt,
                    "- #{} [{}] {}",
                    related.number, related.state, related.title
                );
            }
            prompt.push('\n');
        }

        // Include repository structure (source files)
        if !issue.repo_tree.is_empty() {
            prompt.push_str("Repository Structure (source files):\n");
            for path in issue.repo_tree.iter().take(20) {
                let _ = writeln!(prompt, "- {path}");
            }
            prompt.push('\n');
        }

        // Include available labels
        if !issue.available_labels.is_empty() {
            prompt.push_str("Available Labels:\n");
            for label in issue.available_labels.iter().take(MAX_LABELS) {
                let description = if label.description.is_empty() {
                    String::new()
                } else {
                    format!(" - {}", label.description)
                };
                let _ = writeln!(
                    prompt,
                    "- {} (color: #{}){}",
                    label.name, label.color, description
                );
            }
            prompt.push('\n');
        }

        // Include available milestones
        if !issue.available_milestones.is_empty() {
            prompt.push_str("Available Milestones:\n");
            for milestone in issue.available_milestones.iter().take(MAX_MILESTONES) {
                let description = if milestone.description.is_empty() {
                    String::new()
                } else {
                    format!(" - {}", milestone.description)
                };
                let _ = writeln!(prompt, "- {}{}", milestone.title, description);
            }
            prompt.push('\n');
        }

        prompt.push_str("</issue_content>");
        prompt.push_str(SCHEMA_PREAMBLE);
        prompt.push_str(crate::ai::prompts::TRIAGE_SCHEMA);

        prompt
    }

    /// Builds the system prompt for issue creation/formatting.
    #[must_use]
    fn build_create_system_prompt(custom_guidance: Option<&str>) -> String {
        let context = super::context::load_custom_guidance(custom_guidance);
        build_create_system_prompt(&context)
    }

    /// Builds the user prompt for issue creation/formatting.
    #[must_use]
    fn build_create_user_prompt(title: &str, body: &str, _repo: &str) -> String {
        format!(
            "Please format this GitHub issue:\n\nTitle: {title}\n\nBody:\n{body}{}{}",
            SCHEMA_PREAMBLE,
            crate::ai::prompts::CREATE_SCHEMA
        )
    }

    /// Reviews a pull request using the provider's API.
    ///
    /// Analyzes PR metadata and file diffs to provide structured review feedback.
    ///
    /// # Arguments
    ///
    /// * `pr` - Pull request details including files and diffs
    ///
    /// # Errors
    ///
    /// Returns an error if:
    /// - API request fails (network, timeout, rate limit)
    /// - Response cannot be parsed as valid JSON
    #[instrument(skip(self, pr, ast_context, call_graph), fields(pr_number = pr.number, repo = %format!("{}/{}", pr.owner, pr.repo)))]
    async fn review_pr(
        &self,
        pr: &super::types::PrDetails,
        mut ast_context: String,
        mut call_graph: String,
        review_config: &crate::config::ReviewConfig,
    ) -> Result<(super::types::PrReviewResponse, AiStats)> {
        debug!(model = %self.model(), "Calling {} API for PR review", self.name());

        // Estimate preliminary size; enforce drop order for budget control
        let mut estimated_size = pr.title.len()
            + pr.body.len()
            + pr.files
                .iter()
                .map(|f| f.patch.as_ref().map_or(0, String::len))
                .sum::<usize>()
            + pr.files
                .iter()
                .map(|f| f.full_content.as_ref().map_or(0, String::len))
                .sum::<usize>()
            + ast_context.len()
            + call_graph.len()
            + PROMPT_OVERHEAD_CHARS;

        let max_prompt_chars = review_config.max_prompt_chars;

        // Drop call_graph if over budget
        if estimated_size > max_prompt_chars {
            tracing::warn!(
                section = "call_graph",
                chars = call_graph.len(),
                "Dropping section: prompt budget exceeded"
            );
            let dropped_chars = call_graph.len();
            call_graph.clear();
            estimated_size -= dropped_chars;
        }

        // Drop ast_context if still over budget
        if estimated_size > max_prompt_chars {
            tracing::warn!(
                section = "ast_context",
                chars = ast_context.len(),
                "Dropping section: prompt budget exceeded"
            );
            let dropped_chars = ast_context.len();
            ast_context.clear();
            estimated_size -= dropped_chars;
        }

        // Step 3: Drop largest file patches first if still over budget
        let mut pr_mut = pr.clone();
        if estimated_size > max_prompt_chars {
            // Collect files with their patch sizes
            let mut file_sizes: Vec<(usize, usize)> = pr_mut
                .files
                .iter()
                .enumerate()
                .map(|(idx, f)| (idx, f.patch.as_ref().map_or(0, String::len)))
                .collect();
            // Sort by patch size descending
            file_sizes.sort_by(|a, b| b.1.cmp(&a.1));

            for (file_idx, patch_size) in file_sizes {
                if estimated_size <= max_prompt_chars {
                    break;
                }
                if patch_size > 0 {
                    tracing::warn!(
                        file = %pr_mut.files[file_idx].filename,
                        patch_chars = patch_size,
                        "Dropping file patch: prompt budget exceeded"
                    );
                    pr_mut.files[file_idx].patch = None;
                    estimated_size -= patch_size;
                }
            }
        }

        // Step 4: drop full_content on all files
        if estimated_size > max_prompt_chars {
            for file in &mut pr_mut.files {
                if let Some(fc) = file.full_content.take() {
                    estimated_size = estimated_size.saturating_sub(fc.len());
                    tracing::warn!(
                        bytes = fc.len(),
                        filename = %file.filename,
                        "prompt budget: dropping full_content"
                    );
                }
            }
        }

        tracing::info!(
            prompt_chars = estimated_size,
            max_chars = max_prompt_chars,
            "PR review prompt assembled"
        );

        // Build request
        let system_content = if let Some(override_prompt) =
            super::context::load_system_prompt_override("pr_review_system").await
        {
            override_prompt
        } else {
            Self::build_pr_review_system_prompt(self.custom_guidance())
        };

        // Assemble full prompt to measure actual size
        let assembled_prompt =
            Self::build_pr_review_user_prompt(&pr_mut, &ast_context, &call_graph);
        let actual_prompt_chars = assembled_prompt.len();

        tracing::info!(
            actual_prompt_chars,
            estimated_prompt_chars = estimated_size,
            max_chars = max_prompt_chars,
            "Actual assembled prompt size vs. estimate"
        );

        let request = ChatCompletionRequest {
            model: self.model().to_string(),
            messages: vec![
                ChatMessage {
                    role: "system".to_string(),
                    content: Some(system_content),
                    reasoning: None,
                },
                ChatMessage {
                    role: "user".to_string(),
                    content: Some(assembled_prompt),
                    reasoning: None,
                },
            ],
            response_format: Some(ResponseFormat {
                format_type: "json_object".to_string(),
                json_schema: None,
            }),
            max_tokens: Some(self.max_tokens()),
            temperature: Some(self.temperature()),
        };

        // Send request and parse JSON with retry logic
        let (review, mut ai_stats) = self
            .send_and_parse::<super::types::PrReviewResponse>(&request)
            .await?;

        ai_stats.prompt_chars = actual_prompt_chars;

        debug!(
            verdict = %review.verdict,
            input_tokens = ai_stats.input_tokens,
            output_tokens = ai_stats.output_tokens,
            duration_ms = ai_stats.duration_ms,
            prompt_chars = ai_stats.prompt_chars,
            "PR review complete with stats"
        );

        Ok((review, ai_stats))
    }

    /// Suggests labels for a pull request using the provider's API.
    ///
    /// Analyzes PR title, body, and file paths to suggest relevant labels.
    ///
    /// # Arguments
    ///
    /// * `title` - Pull request title
    /// * `body` - Pull request description
    /// * `file_paths` - List of file paths changed in the PR
    ///
    /// # Errors
    ///
    /// Returns an error if:
    /// - API request fails (network, timeout, rate limit)
    /// - Response cannot be parsed as valid JSON
    #[instrument(skip(self), fields(title = %title))]
    async fn suggest_pr_labels(
        &self,
        title: &str,
        body: &str,
        file_paths: &[String],
    ) -> Result<(Vec<String>, AiStats)> {
        debug!(model = %self.model(), "Calling {} API for PR label suggestion", self.name());

        // Build request
        let system_content = if let Some(override_prompt) =
            super::context::load_system_prompt_override("pr_label_system").await
        {
            override_prompt
        } else {
            Self::build_pr_label_system_prompt(self.custom_guidance())
        };

        let request = ChatCompletionRequest {
            model: self.model().to_string(),
            messages: vec![
                ChatMessage {
                    role: "system".to_string(),
                    content: Some(system_content),
                    reasoning: None,
                },
                ChatMessage {
                    role: "user".to_string(),
                    content: Some(Self::build_pr_label_user_prompt(title, body, file_paths)),
                    reasoning: None,
                },
            ],
            response_format: Some(ResponseFormat {
                format_type: "json_object".to_string(),
                json_schema: None,
            }),
            max_tokens: Some(self.max_tokens()),
            temperature: Some(self.temperature()),
        };

        // Send request and parse JSON with retry logic
        let (response, ai_stats) = self
            .send_and_parse::<super::types::PrLabelResponse>(&request)
            .await?;

        debug!(
            label_count = response.suggested_labels.len(),
            input_tokens = ai_stats.input_tokens,
            output_tokens = ai_stats.output_tokens,
            duration_ms = ai_stats.duration_ms,
            "PR label suggestion complete with stats"
        );

        Ok((response.suggested_labels, ai_stats))
    }

    /// Builds the system prompt for PR review.
    #[must_use]
    fn build_pr_review_system_prompt(custom_guidance: Option<&str>) -> String {
        let context = super::context::load_custom_guidance(custom_guidance);
        build_pr_review_system_prompt(&context)
    }

    /// Builds the user prompt for PR review.
    ///
    /// All user-controlled fields (title, body, filename, status, patch) are sanitized via
    /// [`sanitize_prompt_field`] before being written into the prompt to prevent prompt
    /// injection via XML tag smuggling.
    #[must_use]
    fn build_pr_review_user_prompt(
        pr: &super::types::PrDetails,
        ast_context: &str,
        call_graph: &str,
    ) -> String {
        use std::fmt::Write;

        let mut prompt = String::new();

        prompt.push_str("<pull_request>\n");
        let _ = writeln!(prompt, "Title: {}\n", sanitize_prompt_field(&pr.title));
        let _ = writeln!(prompt, "Branch: {} -> {}\n", pr.head_branch, pr.base_branch);

        // PR description - sanitize before truncation
        let sanitized_body = sanitize_prompt_field(&pr.body);
        let body = if sanitized_body.is_empty() {
            "[No description provided]".to_string()
        } else if sanitized_body.len() > MAX_BODY_LENGTH {
            format!(
                "{}...\n[Description truncated - original length: {} chars]",
                &sanitized_body[..MAX_BODY_LENGTH],
                sanitized_body.len()
            )
        } else {
            sanitized_body
        };
        let _ = writeln!(prompt, "Description:\n{body}\n");

        // File changes with limits
        prompt.push_str("Files Changed:\n");
        let mut total_diff_size = 0;
        let mut files_included = 0;
        let mut files_skipped = 0;

        for file in &pr.files {
            // Check file count limit
            if files_included >= MAX_FILES {
                files_skipped += 1;
                continue;
            }

            let _ = writeln!(
                prompt,
                "- {} ({}) +{} -{}\n",
                sanitize_prompt_field(&file.filename),
                sanitize_prompt_field(&file.status),
                file.additions,
                file.deletions
            );

            // Include patch if available (sanitize then truncate large patches)
            if let Some(patch) = &file.patch {
                const MAX_PATCH_LENGTH: usize = 2000;
                let sanitized_patch = sanitize_prompt_field(patch);
                let patch_content = if sanitized_patch.len() > MAX_PATCH_LENGTH {
                    format!(
                        "{}...\n[Patch truncated - original length: {} chars]",
                        &sanitized_patch[..MAX_PATCH_LENGTH],
                        sanitized_patch.len()
                    )
                } else {
                    sanitized_patch
                };

                // Check if adding this patch would exceed total diff size limit
                let patch_size = patch_content.len();
                if total_diff_size + patch_size > MAX_TOTAL_DIFF_SIZE {
                    let _ = writeln!(
                        prompt,
                        "```diff\n[Patch omitted - total diff size limit reached]\n```\n"
                    );
                    files_skipped += 1;
                    continue;
                }

                let _ = writeln!(prompt, "```diff\n{patch_content}\n```\n");
                total_diff_size += patch_size;
            }

            // Include full file content if available (cap at MAX_FULL_CONTENT_CHARS)
            if let Some(content) = &file.full_content {
                let sanitized = sanitize_prompt_field(content);
                let displayed = if sanitized.len() > MAX_FULL_CONTENT_CHARS {
                    sanitized[..MAX_FULL_CONTENT_CHARS].to_string()
                } else {
                    sanitized
                };
                let _ = writeln!(
                    prompt,
                    "<file_content path=\"{}\">\n{}\n</file_content>\n",
                    sanitize_prompt_field(&file.filename),
                    displayed
                );
            }

            files_included += 1;
        }

        // Add truncation message if files were skipped
        if files_skipped > 0 {
            let _ = writeln!(
                prompt,
                "\n[{files_skipped} files omitted due to size limits (MAX_FILES={MAX_FILES}, MAX_TOTAL_DIFF_SIZE={MAX_TOTAL_DIFF_SIZE})]"
            );
        }

        prompt.push_str("</pull_request>");
        if !ast_context.is_empty() {
            prompt.push_str(ast_context);
        }
        if !call_graph.is_empty() {
            prompt.push_str(call_graph);
        }
        prompt.push_str(SCHEMA_PREAMBLE);
        prompt.push_str(crate::ai::prompts::PR_REVIEW_SCHEMA);

        prompt
    }

    /// Builds the system prompt for PR label suggestion.
    #[must_use]
    fn build_pr_label_system_prompt(custom_guidance: Option<&str>) -> String {
        let context = super::context::load_custom_guidance(custom_guidance);
        build_pr_label_system_prompt(&context)
    }

    /// Builds the user prompt for PR label suggestion.
    #[must_use]
    fn build_pr_label_user_prompt(title: &str, body: &str, file_paths: &[String]) -> String {
        use std::fmt::Write;

        let mut prompt = String::new();

        prompt.push_str("<pull_request>\n");
        let _ = writeln!(prompt, "Title: {title}\n");

        // PR description
        let body_content = if body.is_empty() {
            "[No description provided]".to_string()
        } else if body.len() > MAX_BODY_LENGTH {
            format!(
                "{}...\n[Description truncated - original length: {} chars]",
                &body[..MAX_BODY_LENGTH],
                body.len()
            )
        } else {
            body.to_string()
        };
        let _ = writeln!(prompt, "Description:\n{body_content}\n");

        // File paths
        if !file_paths.is_empty() {
            prompt.push_str("Files Changed:\n");
            for path in file_paths.iter().take(20) {
                let _ = writeln!(prompt, "- {path}");
            }
            if file_paths.len() > 20 {
                let _ = writeln!(prompt, "- ... and {} more files", file_paths.len() - 20);
            }
            prompt.push('\n');
        }

        prompt.push_str("</pull_request>");
        prompt.push_str(SCHEMA_PREAMBLE);
        prompt.push_str(crate::ai::prompts::PR_LABEL_SCHEMA);

        prompt
    }

    /// Generate release notes from PR summaries.
    ///
    /// # Arguments
    ///
    /// * `prs` - List of PR summaries to synthesize
    /// * `version` - Version being released
    ///
    /// # Returns
    ///
    /// Structured release notes with theme, highlights, and categorized changes.
    #[instrument(skip(self, prs))]
    async fn generate_release_notes(
        &self,
        prs: Vec<super::types::PrSummary>,
        version: &str,
    ) -> Result<(super::types::ReleaseNotesResponse, AiStats)> {
        let system_content = if let Some(override_prompt) =
            super::context::load_system_prompt_override("release_notes_system").await
        {
            override_prompt
        } else {
            let context = super::context::load_custom_guidance(self.custom_guidance());
            build_release_notes_system_prompt(&context)
        };
        let prompt = Self::build_release_notes_prompt(&prs, version);
        let request = ChatCompletionRequest {
            model: self.model().to_string(),
            messages: vec![
                ChatMessage {
                    role: "system".to_string(),
                    content: Some(system_content),
                    reasoning: None,
                },
                ChatMessage {
                    role: "user".to_string(),
                    content: Some(prompt),
                    reasoning: None,
                },
            ],
            response_format: Some(ResponseFormat {
                format_type: "json_object".to_string(),
                json_schema: None,
            }),
            temperature: Some(0.7),
            max_tokens: Some(self.max_tokens()),
        };

        let (parsed, ai_stats) = self
            .send_and_parse::<super::types::ReleaseNotesResponse>(&request)
            .await?;

        debug!(
            input_tokens = ai_stats.input_tokens,
            output_tokens = ai_stats.output_tokens,
            duration_ms = ai_stats.duration_ms,
            "Release notes generation complete with stats"
        );

        Ok((parsed, ai_stats))
    }

    /// Build the user prompt for release notes generation.
    #[must_use]
    fn build_release_notes_prompt(prs: &[super::types::PrSummary], version: &str) -> String {
        let pr_list = prs
            .iter()
            .map(|pr| {
                format!(
                    "- #{}: {} (by @{})\n  {}",
                    pr.number,
                    pr.title,
                    pr.author,
                    pr.body.lines().next().unwrap_or("")
                )
            })
            .collect::<Vec<_>>()
            .join("\n");

        format!(
            "Generate release notes for version {version} based on these merged PRs:\n\n{pr_list}{}{}",
            SCHEMA_PREAMBLE,
            crate::ai::prompts::RELEASE_NOTES_SCHEMA
        )
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    /// Shared struct for parse_ai_json error-path tests.
    /// The field is only used via serde deserialization; `_message` silences dead_code.
    #[derive(Debug, serde::Deserialize)]
    struct ErrorTestResponse {
        _message: String,
    }

    struct TestProvider;

    impl AiProvider for TestProvider {
        fn name(&self) -> &'static str {
            "test"
        }

        fn api_url(&self) -> &'static str {
            "https://test.example.com"
        }

        fn api_key_env(&self) -> &'static str {
            "TEST_API_KEY"
        }

        fn http_client(&self) -> &Client {
            unimplemented!()
        }

        fn api_key(&self) -> &SecretString {
            unimplemented!()
        }

        fn model(&self) -> &'static str {
            "test-model"
        }

        fn max_tokens(&self) -> u32 {
            2048
        }

        fn temperature(&self) -> f32 {
            0.3
        }
    }

    #[test]
    fn test_build_system_prompt_contains_json_schema() {
        let system_prompt = TestProvider::build_system_prompt(None);
        // Schema description strings are unique to the schema file and must NOT appear in the
        // system prompt after moving schema injection to the user turn.
        assert!(
            !system_prompt
                .contains("A 2-3 sentence summary of what the issue is about and its impact")
        );

        // Schema MUST appear in the user prompt
        let issue = IssueDetails::builder()
            .owner("test".to_string())
            .repo("repo".to_string())
            .number(1)
            .title("Test".to_string())
            .body("Body".to_string())
            .labels(vec![])
            .comments(vec![])
            .url("https://github.com/test/repo/issues/1".to_string())
            .build();
        let user_prompt = TestProvider::build_user_prompt(&issue);
        assert!(
            user_prompt
                .contains("A 2-3 sentence summary of what the issue is about and its impact")
        );
        assert!(user_prompt.contains("suggested_labels"));
    }

    #[test]
    fn test_build_user_prompt_with_delimiters() {
        let issue = IssueDetails::builder()
            .owner("test".to_string())
            .repo("repo".to_string())
            .number(1)
            .title("Test issue".to_string())
            .body("This is the body".to_string())
            .labels(vec!["bug".to_string()])
            .comments(vec![])
            .url("https://github.com/test/repo/issues/1".to_string())
            .build();

        let prompt = TestProvider::build_user_prompt(&issue);
        assert!(prompt.starts_with("<issue_content>"));
        assert!(prompt.contains("</issue_content>"));
        assert!(prompt.contains("Respond with valid JSON matching this schema"));
        assert!(prompt.contains("Title: Test issue"));
        assert!(prompt.contains("This is the body"));
        assert!(prompt.contains("Existing Labels: bug"));
    }

    #[test]
    fn test_build_user_prompt_truncates_long_body() {
        let long_body = "x".repeat(5000);
        let issue = IssueDetails::builder()
            .owner("test".to_string())
            .repo("repo".to_string())
            .number(1)
            .title("Test".to_string())
            .body(long_body)
            .labels(vec![])
            .comments(vec![])
            .url("https://github.com/test/repo/issues/1".to_string())
            .build();

        let prompt = TestProvider::build_user_prompt(&issue);
        assert!(prompt.contains("[Body truncated"));
        assert!(prompt.contains("5000 chars"));
    }

    #[test]
    fn test_build_user_prompt_empty_body() {
        let issue = IssueDetails::builder()
            .owner("test".to_string())
            .repo("repo".to_string())
            .number(1)
            .title("Test".to_string())
            .body(String::new())
            .labels(vec![])
            .comments(vec![])
            .url("https://github.com/test/repo/issues/1".to_string())
            .build();

        let prompt = TestProvider::build_user_prompt(&issue);
        assert!(prompt.contains("[No description provided]"));
    }

    #[test]
    fn test_build_create_system_prompt_contains_json_schema() {
        let system_prompt = TestProvider::build_create_system_prompt(None);
        // Schema description strings are unique to the schema file and must NOT appear in system prompt.
        assert!(
            !system_prompt
                .contains("Well-formatted issue title following conventional commit style")
        );

        // Schema MUST appear in the user prompt
        let user_prompt =
            TestProvider::build_create_user_prompt("My title", "My body", "test/repo");
        assert!(
            user_prompt.contains("Well-formatted issue title following conventional commit style")
        );
        assert!(user_prompt.contains("formatted_body"));
    }

    #[test]
    fn test_build_pr_review_user_prompt_respects_file_limit() {
        use super::super::types::{PrDetails, PrFile};

        let mut files = Vec::new();
        for i in 0..25 {
            files.push(PrFile {
                filename: format!("file{i}.rs"),
                status: "modified".to_string(),
                additions: 10,
                deletions: 5,
                patch: Some(format!("patch content {i}")),
                full_content: None,
            });
        }

        let pr = PrDetails {
            owner: "test".to_string(),
            repo: "repo".to_string(),
            number: 1,
            title: "Test PR".to_string(),
            body: "Description".to_string(),
            head_branch: "feature".to_string(),
            base_branch: "main".to_string(),
            url: "https://github.com/test/repo/pull/1".to_string(),
            files,
            labels: vec![],
            head_sha: String::new(),
        };

        let prompt = TestProvider::build_pr_review_user_prompt(&pr, "", "");
        assert!(prompt.contains("files omitted due to size limits"));
        assert!(prompt.contains("MAX_FILES=20"));
    }

    #[test]
    fn test_build_pr_review_user_prompt_respects_diff_size_limit() {
        use super::super::types::{PrDetails, PrFile};

        // Create patches that will exceed the limit when combined
        // Each patch is ~30KB, so two will exceed 50KB limit
        let patch1 = "x".repeat(30_000);
        let patch2 = "y".repeat(30_000);

        let files = vec![
            PrFile {
                filename: "file1.rs".to_string(),
                status: "modified".to_string(),
                additions: 100,
                deletions: 50,
                patch: Some(patch1),
                full_content: None,
            },
            PrFile {
                filename: "file2.rs".to_string(),
                status: "modified".to_string(),
                additions: 100,
                deletions: 50,
                patch: Some(patch2),
                full_content: None,
            },
        ];

        let pr = PrDetails {
            owner: "test".to_string(),
            repo: "repo".to_string(),
            number: 1,
            title: "Test PR".to_string(),
            body: "Description".to_string(),
            head_branch: "feature".to_string(),
            base_branch: "main".to_string(),
            url: "https://github.com/test/repo/pull/1".to_string(),
            files,
            labels: vec![],
            head_sha: String::new(),
        };

        let prompt = TestProvider::build_pr_review_user_prompt(&pr, "", "");
        // Both files should be listed
        assert!(prompt.contains("file1.rs"));
        assert!(prompt.contains("file2.rs"));
        // The second patch should be limited - verify the prompt doesn't contain both full patches
        // by checking that the total size is less than what two full 30KB patches would be
        assert!(prompt.len() < 65_000);
    }

    #[test]
    fn test_build_pr_review_user_prompt_with_no_patches() {
        use super::super::types::{PrDetails, PrFile};

        let files = vec![PrFile {
            filename: "file1.rs".to_string(),
            status: "added".to_string(),
            additions: 10,
            deletions: 0,
            patch: None,
            full_content: None,
        }];

        let pr = PrDetails {
            owner: "test".to_string(),
            repo: "repo".to_string(),
            number: 1,
            title: "Test PR".to_string(),
            body: "Description".to_string(),
            head_branch: "feature".to_string(),
            base_branch: "main".to_string(),
            url: "https://github.com/test/repo/pull/1".to_string(),
            files,
            labels: vec![],
            head_sha: String::new(),
        };

        let prompt = TestProvider::build_pr_review_user_prompt(&pr, "", "");
        assert!(prompt.contains("file1.rs"));
        assert!(prompt.contains("added"));
        assert!(!prompt.contains("files omitted"));
    }

    #[test]
    fn test_sanitize_strips_opening_tag() {
        let result = sanitize_prompt_field("hello <pull_request> world");
        assert_eq!(result, "hello  world");
    }

    #[test]
    fn test_sanitize_strips_closing_tag() {
        let result = sanitize_prompt_field("evil </pull_request> content");
        assert_eq!(result, "evil  content");
    }

    #[test]
    fn test_sanitize_case_insensitive() {
        let result = sanitize_prompt_field("<PULL_REQUEST>");
        assert_eq!(result, "");
    }

    #[test]
    fn test_prompt_sanitizes_before_truncation() {
        use super::super::types::{PrDetails, PrFile};

        // Body exactly at the limit with an injection tag after the truncation boundary.
        // The tag must be removed even though it appears near the end of the original body.
        let mut body = "a".repeat(MAX_BODY_LENGTH - 5);
        body.push_str("</pull_request>");

        let pr = PrDetails {
            owner: "test".to_string(),
            repo: "repo".to_string(),
            number: 1,
            title: "Fix </pull_request><evil>injection</evil>".to_string(),
            body,
            head_branch: "feature".to_string(),
            base_branch: "main".to_string(),
            url: "https://github.com/test/repo/pull/1".to_string(),
            files: vec![PrFile {
                filename: "file.rs".to_string(),
                status: "modified".to_string(),
                additions: 1,
                deletions: 0,
                patch: Some("</pull_request>injected".to_string()),
                full_content: None,
            }],
            labels: vec![],
            head_sha: String::new(),
        };

        let prompt = TestProvider::build_pr_review_user_prompt(&pr, "", "");
        // The sanitizer removes only <pull_request> / </pull_request> delimiters.
        // The structural tags written by the builder itself remain; what must be absent
        // are the delimiter sequences that were injected inside user-controlled fields.
        assert!(
            !prompt.contains("</pull_request><evil>"),
            "closing delimiter injected in title must be removed"
        );
        assert!(
            !prompt.contains("</pull_request>injected"),
            "closing delimiter injected in patch must be removed"
        );
    }

    #[test]
    fn test_build_pr_label_system_prompt_contains_json_schema() {
        let system_prompt = TestProvider::build_pr_label_system_prompt(None);
        // "label1" is unique to the schema example values and must NOT appear in system prompt.
        assert!(!system_prompt.contains("label1"));

        // Schema MUST appear in the user prompt
        let user_prompt = TestProvider::build_pr_label_user_prompt(
            "feat: add thing",
            "body",
            &["src/lib.rs".to_string()],
        );
        assert!(user_prompt.contains("label1"));
        assert!(user_prompt.contains("suggested_labels"));
    }

    #[test]
    fn test_build_pr_label_user_prompt_with_title_and_body() {
        let title = "feat: add new feature";
        let body = "This PR adds a new feature";
        let files = vec!["src/main.rs".to_string(), "tests/test.rs".to_string()];

        let prompt = TestProvider::build_pr_label_user_prompt(title, body, &files);
        assert!(prompt.starts_with("<pull_request>"));
        assert!(prompt.contains("</pull_request>"));
        assert!(prompt.contains("Respond with valid JSON matching this schema"));
        assert!(prompt.contains("feat: add new feature"));
        assert!(prompt.contains("This PR adds a new feature"));
        assert!(prompt.contains("src/main.rs"));
        assert!(prompt.contains("tests/test.rs"));
    }

    #[test]
    fn test_build_pr_label_user_prompt_empty_body() {
        let title = "fix: bug fix";
        let body = "";
        let files = vec!["src/lib.rs".to_string()];

        let prompt = TestProvider::build_pr_label_user_prompt(title, body, &files);
        assert!(prompt.contains("[No description provided]"));
        assert!(prompt.contains("src/lib.rs"));
    }

    #[test]
    fn test_build_pr_label_user_prompt_truncates_long_body() {
        let title = "test";
        let long_body = "x".repeat(5000);
        let files = vec![];

        let prompt = TestProvider::build_pr_label_user_prompt(title, &long_body, &files);
        assert!(prompt.contains("[Description truncated"));
        assert!(prompt.contains("5000 chars"));
    }

    #[test]
    fn test_build_pr_label_user_prompt_respects_file_limit() {
        let title = "test";
        let body = "test";
        let mut files = Vec::new();
        for i in 0..25 {
            files.push(format!("file{i}.rs"));
        }

        let prompt = TestProvider::build_pr_label_user_prompt(title, body, &files);
        assert!(prompt.contains("file0.rs"));
        assert!(prompt.contains("file19.rs"));
        assert!(!prompt.contains("file20.rs"));
        assert!(prompt.contains("... and 5 more files"));
    }

    #[test]
    fn test_build_pr_label_user_prompt_empty_files() {
        let title = "test";
        let body = "test";
        let files: Vec<String> = vec![];

        let prompt = TestProvider::build_pr_label_user_prompt(title, body, &files);
        assert!(prompt.contains("Title: test"));
        assert!(prompt.contains("Description:\ntest"));
        assert!(!prompt.contains("Files Changed:"));
    }

    #[test]
    fn test_parse_ai_json_with_valid_json() {
        #[derive(serde::Deserialize)]
        struct TestResponse {
            message: String,
        }

        let json = r#"{"message": "hello"}"#;
        let result: Result<TestResponse> = parse_ai_json(json, "test-provider");
        assert!(result.is_ok());
        let response = result.unwrap();
        assert_eq!(response.message, "hello");
    }

    #[test]
    fn test_parse_ai_json_with_truncated_json() {
        let json = r#"{"message": "hello"#;
        let result: Result<ErrorTestResponse> = parse_ai_json(json, "test-provider");
        assert!(result.is_err());
        let err = result.unwrap_err();
        assert!(
            err.to_string()
                .contains("Truncated response from test-provider")
        );
    }

    #[test]
    fn test_parse_ai_json_with_malformed_json() {
        let json = r#"{"message": invalid}"#;
        let result: Result<ErrorTestResponse> = parse_ai_json(json, "test-provider");
        assert!(result.is_err());
        let err = result.unwrap_err();
        assert!(err.to_string().contains("Invalid JSON response from AI"));
    }

    #[tokio::test]
    async fn test_load_system_prompt_override_returns_none_when_absent() {
        let result =
            super::super::context::load_system_prompt_override("__nonexistent_test_override__")
                .await;
        assert!(result.is_none());
    }

    #[tokio::test]
    async fn test_load_system_prompt_override_returns_content_when_present() {
        use std::io::Write;
        let dir = tempfile::tempdir().expect("create tempdir");
        let file_path = dir.path().join("test_override.md");
        let mut f = std::fs::File::create(&file_path).expect("create file");
        writeln!(f, "Custom override content").expect("write file");
        drop(f);

        let content = tokio::fs::read_to_string(&file_path).await.ok();
        assert_eq!(content.as_deref(), Some("Custom override content\n"));
    }

    #[test]
    fn test_build_pr_review_prompt_omits_call_graph_when_oversized() {
        use super::super::types::{PrDetails, PrFile};

        // Arrange: simulate review_pr dropping call_graph due to budget.
        // When call_graph is oversized, review_pr clears it before calling build_pr_review_user_prompt.
        let pr = PrDetails {
            owner: "test".to_string(),
            repo: "repo".to_string(),
            number: 1,
            title: "Budget drop test".to_string(),
            body: "body".to_string(),
            head_branch: "feat".to_string(),
            base_branch: "main".to_string(),
            url: "https://github.com/test/repo/pull/1".to_string(),
            files: vec![PrFile {
                filename: "lib.rs".to_string(),
                status: "modified".to_string(),
                additions: 1,
                deletions: 0,
                patch: Some("+line".to_string()),
                full_content: None,
            }],
            labels: vec![],
            head_sha: String::new(),
        };

        // Act: call build_pr_review_user_prompt with empty call_graph (dropped by review_pr)
        // and non-empty ast_context (retained because it fits after call_graph drop)
        let ast_context = "Y".repeat(500);
        let call_graph = "";
        let prompt = TestProvider::build_pr_review_user_prompt(&pr, &ast_context, call_graph);

        // Assert: call_graph absent, ast_context present
        assert!(
            !prompt.contains(&"X".repeat(10)),
            "call_graph content must not appear in prompt after budget drop"
        );
        assert!(
            prompt.contains(&"Y".repeat(10)),
            "ast_context content must appear in prompt (fits within budget)"
        );
    }

    #[test]
    fn test_build_pr_review_prompt_omits_ast_after_call_graph() {
        use super::super::types::{PrDetails, PrFile};

        // Arrange: simulate review_pr dropping both call_graph and ast_context due to budget.
        let pr = PrDetails {
            owner: "test".to_string(),
            repo: "repo".to_string(),
            number: 1,
            title: "Budget drop test".to_string(),
            body: "body".to_string(),
            head_branch: "feat".to_string(),
            base_branch: "main".to_string(),
            url: "https://github.com/test/repo/pull/1".to_string(),
            files: vec![PrFile {
                filename: "lib.rs".to_string(),
                status: "modified".to_string(),
                additions: 1,
                deletions: 0,
                patch: Some("+line".to_string()),
                full_content: None,
            }],
            labels: vec![],
            head_sha: String::new(),
        };

        // Act: call build_pr_review_user_prompt with both empty (dropped by review_pr)
        let ast_context = "";
        let call_graph = "";
        let prompt = TestProvider::build_pr_review_user_prompt(&pr, ast_context, call_graph);

        // Assert: both absent, PR title retained
        assert!(
            !prompt.contains(&"C".repeat(10)),
            "call_graph content must not appear after budget drop"
        );
        assert!(
            !prompt.contains(&"A".repeat(10)),
            "ast_context content must not appear after budget drop"
        );
        assert!(
            prompt.contains("Budget drop test"),
            "PR title must be retained in prompt"
        );
    }

    #[test]
    fn test_build_pr_review_prompt_drops_patches_when_over_budget() {
        use super::super::types::{PrDetails, PrFile};

        // Arrange: simulate review_pr dropping patches due to budget.
        // Create 3 files with patches of different sizes.
        let pr = PrDetails {
            owner: "test".to_string(),
            repo: "repo".to_string(),
            number: 1,
            title: "Patch drop test".to_string(),
            body: "body".to_string(),
            head_branch: "feat".to_string(),
            base_branch: "main".to_string(),
            url: "https://github.com/test/repo/pull/1".to_string(),
            files: vec![
                PrFile {
                    filename: "large.rs".to_string(),
                    status: "modified".to_string(),
                    additions: 100,
                    deletions: 50,
                    patch: Some("L".repeat(5000)),
                    full_content: None,
                },
                PrFile {
                    filename: "medium.rs".to_string(),
                    status: "modified".to_string(),
                    additions: 50,
                    deletions: 25,
                    patch: Some("M".repeat(3000)),
                    full_content: None,
                },
                PrFile {
                    filename: "small.rs".to_string(),
                    status: "modified".to_string(),
                    additions: 10,
                    deletions: 5,
                    patch: Some("S".repeat(1000)),
                    full_content: None,
                },
            ],
            labels: vec![],
            head_sha: String::new(),
        };

        // Act: simulate review_pr dropping largest patches first
        let mut pr_mut = pr.clone();
        pr_mut.files[0].patch = None; // Drop largest patch
        pr_mut.files[1].patch = None; // Drop medium patch
        // Keep smallest patch

        let ast_context = "";
        let call_graph = "";
        let prompt = TestProvider::build_pr_review_user_prompt(&pr_mut, ast_context, call_graph);

        // Assert: largest patches absent, smallest present
        assert!(
            !prompt.contains(&"L".repeat(10)),
            "largest patch must be absent after drop"
        );
        assert!(
            !prompt.contains(&"M".repeat(10)),
            "medium patch must be absent after drop"
        );
        assert!(
            prompt.contains(&"S".repeat(10)),
            "smallest patch must be present"
        );
    }

    #[test]
    fn test_build_pr_review_prompt_drops_full_content_as_last_resort() {
        use super::super::types::{PrDetails, PrFile};

        // Arrange: simulate review_pr dropping full_content as last resort.
        let pr = PrDetails {
            owner: "test".to_string(),
            repo: "repo".to_string(),
            number: 1,
            title: "Full content drop test".to_string(),
            body: "body".to_string(),
            head_branch: "feat".to_string(),
            base_branch: "main".to_string(),
            url: "https://github.com/test/repo/pull/1".to_string(),
            files: vec![
                PrFile {
                    filename: "file1.rs".to_string(),
                    status: "modified".to_string(),
                    additions: 10,
                    deletions: 5,
                    patch: None,
                    full_content: Some("F".repeat(5000)),
                },
                PrFile {
                    filename: "file2.rs".to_string(),
                    status: "modified".to_string(),
                    additions: 10,
                    deletions: 5,
                    patch: None,
                    full_content: Some("C".repeat(3000)),
                },
            ],
            labels: vec![],
            head_sha: String::new(),
        };

        // Act: simulate review_pr dropping all full_content
        let mut pr_mut = pr.clone();
        for file in &mut pr_mut.files {
            file.full_content = None;
        }

        let ast_context = "";
        let call_graph = "";
        let prompt = TestProvider::build_pr_review_user_prompt(&pr_mut, ast_context, call_graph);

        // Assert: no file_content XML blocks appear
        assert!(
            !prompt.contains("<file_content"),
            "file_content blocks must not appear when full_content is cleared"
        );
        assert!(
            !prompt.contains(&"F".repeat(10)),
            "full_content from file1 must not appear"
        );
        assert!(
            !prompt.contains(&"C".repeat(10)),
            "full_content from file2 must not appear"
        );
    }
}