omni_dev/claude/

client.rs

//! Claude client for commit message improvement.

use anyhow::{Context, Result};
use tracing::{debug, info, warn};

use crate::claude::error::is_structured_output_rejection as ai_error_is_schema_rejection;
use crate::claude::error::is_transient_ai_error as ai_error_is_transient;
use crate::claude::token_budget::TokenBudget;
use crate::claude::{ai::bedrock::BedrockAiClient, ai::claude::ClaudeAiClient};
use crate::claude::{
    ai::{AiClient, RequestOptions, ResponseFormat},
    error::ClaudeError,
    prompts, response_schema,
};
use crate::data::{
    amendments::{Amendment, AmendmentFile},
    context::CommitContext,
    RepositoryView, RepositoryViewForAI,
};
use crate::utils::secret_env::{secret_var, secret_var_any};

/// Returned when the full diff does not fit the token budget.
///
/// Carries the data needed for split dispatch so the caller can size
/// diff chunks appropriately.
struct BudgetExceeded {
    /// Available input tokens for this model (context window minus output reserve).
    available_input_tokens: usize,
}

/// Maximum retries for amendment parse/request failures (matches check retry count).
const AMENDMENT_PARSE_MAX_RETRIES: u32 = 2;

/// Claude client for commit message improvement.
pub struct ClaudeClient {
    /// AI client implementation.
    ai_client: Box<dyn AiClient>,
    /// Whether structured JSON-schema output is off for this client.
    ///
    /// Starts from the [`STRUCTURED_OUTPUT_DISABLE_ENV`] off-switch, and is
    /// latched on by [`send_with_optional_schema`](Self::send_with_optional_schema)
    /// the first time an endpoint rejects `output_config` (issue #1561). One
    /// latch per client covers a whole run: `check` shares a single
    /// `ClaudeClient` across its concurrent batch futures, so the rejection is
    /// discovered at most once however many commits are in flight.
    ///
    /// [`STRUCTURED_OUTPUT_DISABLE_ENV`]: crate::claude::backend::STRUCTURED_OUTPUT_DISABLE_ENV
    schema_disabled: std::sync::atomic::AtomicBool,
}

impl ClaudeClient {
    /// Creates a new Claude client with the provided AI client implementation.
    pub fn new(ai_client: Box<dyn AiClient>) -> Self {
        Self {
            ai_client,
            schema_disabled: std::sync::atomic::AtomicBool::new(false),
        }
    }

    /// Returns `self` with structured JSON-schema output switched off when
    /// `disabled`, whatever the backend advertises.
    ///
    /// Applied by [`create_default_claude_client_with`] from the
    /// [`STRUCTURED_OUTPUT_DISABLE_ENV`](crate::claude::backend::STRUCTURED_OUTPUT_DISABLE_ENV)
    /// off-switch, so an endpoint known to reject `output_config` never gets
    /// sent it (issue #1561).
    #[must_use]
    pub fn with_structured_output_disabled(self, disabled: bool) -> Self {
        self.schema_disabled
            .store(disabled, std::sync::atomic::Ordering::Relaxed);
        self
    }

    /// Whether structured output is currently off for this client — either
    /// switched off up front or latched off by a rejection.
    fn schema_disabled(&self) -> bool {
        self.schema_disabled
            .load(std::sync::atomic::Ordering::Relaxed)
    }

    /// Returns metadata about the AI client.
    pub fn get_ai_client_metadata(&self) -> crate::claude::ai::AiClientMetadata {
        self.ai_client.get_metadata()
    }

    /// Consumes the wrapper and returns the inner [`AiClient`].
    ///
    /// `ClaudeClient` is the commit-message-improvement entry point —
    /// callers that want to drive the AI directly for unrelated workflows
    /// extract the underlying client via this method rather than
    /// reimplementing the backend-dispatch ladder in
    /// [`create_default_claude_client`].
    #[must_use]
    pub fn into_ai_client(self) -> Box<dyn AiClient> {
        self.ai_client
    }

    /// Adjusts a structured-call system prompt for the active backend's
    /// response format.
    ///
    /// Backends advertising
    /// [`AiClientCapabilities::supports_response_schema`](crate::claude::ai::AiClientCapabilities::supports_response_schema)
    /// receive the [`prompts::JSON_SCHEMA_RESPONSE_OVERRIDE`] suffix, which
    /// instructs the model to emit a bare JSON object matching the schema
    /// supplied via [`RequestOptions`]. Other backends receive the prompt
    /// unchanged. Should be called once at the top of each entry point so
    /// the suffix is included in subsequent token-budget calculations.
    fn adjusted_system_prompt(&self, system_prompt: String) -> String {
        let format = ResponseFormat::from_capabilities(&self.ai_client.capabilities());
        prompts::apply_response_format_to_system_prompt(system_prompt, format)
    }

    /// Returns the cached schema when the active backend can enforce
    /// response schemas, or `None` when it cannot.
    ///
    /// Used to gate per-call schema attachment so call sites stay
    /// readable: build the schema unconditionally, gate attachment on
    /// capabilities.
    fn schema_if_supported<'a>(
        &self,
        schema: &'a serde_json::Value,
    ) -> Option<&'a serde_json::Value> {
        if self.ai_client.capabilities().supports_response_schema && !self.schema_disabled() {
            Some(schema)
        } else {
            None
        }
    }

    /// Strips the JSON-schema system-prompt suffix, recovering the YAML
    /// prompt byte-for-byte.
    ///
    /// [`adjusted_system_prompt`](Self::adjusted_system_prompt) appends
    /// [`prompts::JSON_SCHEMA_RESPONSE_OVERRIDE`] at the very end for
    /// schema-capable backends, and that suffix tells the model to disregard
    /// the YAML instructions and match "the JSON Schema provided to the
    /// runtime" — a schema a request without `output_config` no longer
    /// carries. Any dispatch that ends up not attaching the schema must
    /// therefore strip it, or the model is asked to satisfy a schema it was
    /// never shown.
    ///
    /// A no-op when the suffix is absent, so it is safe to apply
    /// unconditionally on the schema-less path.
    fn yaml_system_prompt(system_prompt: &str) -> &str {
        system_prompt
            .strip_suffix(prompts::JSON_SCHEMA_RESPONSE_OVERRIDE)
            .unwrap_or(system_prompt)
    }

    /// Dispatches a structured AI call with optional schema enforcement.
    ///
    /// When `schema` is `Some`, sends via
    /// [`AiClient::send_request_with_options`] so the backend can enforce
    /// the schema (e.g. `claude -p --json-schema <file>`); otherwise
    /// falls back to plain [`AiClient::send_request`]. Backends without
    /// schema support are expected to report
    /// [`AiClientCapabilities::supports_response_schema`](crate::claude::ai::AiClientCapabilities::supports_response_schema)
    /// `= false`, in which case [`schema_if_supported`](Self::schema_if_supported)
    /// at the call site returns `None` and we take the second branch.
    ///
    /// An endpoint that rejects the `output_config` field itself — a gateway
    /// named by `ANTHROPIC_BEDROCK_BASE_URL` that does not pass it through
    /// (issue #1561) — is retried once on the schema-less path and latches
    /// structured output off for the rest of the run. This is a deliberate,
    /// narrow exception to the rule that a permanent 4xx must fail loudly
    /// (#1333): the retry is a *different* request, not the same one re-issued,
    /// the trigger is a positively-identified rejection of the field by name
    /// (`ai_error_is_schema_rejection`), and it warns.
    async fn send_with_optional_schema(
        &self,
        system_prompt: &str,
        user_prompt: &str,
        schema: Option<&serde_json::Value>,
    ) -> Result<String> {
        let Some(s) = schema else {
            return self
                .ai_client
                .send_request(Self::yaml_system_prompt(system_prompt), user_prompt)
                .await;
        };

        let opts = RequestOptions::default().with_response_schema(s.clone());
        match self
            .ai_client
            .send_request_with_options(system_prompt, user_prompt, opts)
            .await
        {
            Ok(content) => Ok(content),
            Err(e) if ai_error_is_schema_rejection(&e) => {
                self.schema_disabled
                    .store(true, std::sync::atomic::Ordering::Relaxed);
                warn!(
                    error = %e,
                    "The AI endpoint rejected the structured-output field `output_config`; \
                     retrying on the YAML response path and disabling it for the rest of this \
                     run. A gateway in front of Bedrock/Anthropic that does not pass \
                     `output_config` through is the usual cause. Set \
                     OMNI_DEV_STRUCTURED_OUTPUT_DISABLE=true, or \
                     `supports_structured_output: false` for this model in \
                     ~/.omni-dev/models.yaml, to skip the rejected request entirely."
                );
                self.ai_client
                    .send_request(Self::yaml_system_prompt(system_prompt), user_prompt)
                    .await
                    .with_context(|| {
                        format!("YAML fallback after the endpoint rejected `output_config` ({e})")
                    })
            }
            Err(e) => Err(e),
        }
    }

    /// Validates that the prompt fits within the model's token budget.
    ///
    /// Estimates token counts and logs utilization before each AI request.
    /// Returns an error if the prompt exceeds available input tokens.
    fn validate_prompt_budget(&self, system_prompt: &str, user_prompt: &str) -> Result<()> {
        let metadata = self.ai_client.get_metadata();
        let budget = TokenBudget::from_metadata(&metadata);
        let estimate = budget.validate_prompt(system_prompt, user_prompt)?;

        debug!(
            model = %metadata.model,
            estimated_tokens = estimate.estimated_tokens,
            available_tokens = estimate.available_tokens,
            utilization_pct = format!("{:.1}%", estimate.utilization_pct),
            "Token budget check passed"
        );

        Ok(())
    }

    /// Builds a user prompt and validates it against the model's token budget.
    ///
    /// Serializes the repository view to YAML, constructs the user prompt, and
    /// checks that it fits within the available input tokens. Returns an error
    /// if the prompt exceeds the budget.
    fn build_prompt_fitting_budget(
        &self,
        ai_view: &RepositoryViewForAI,
        system_prompt: &str,
        build_user_prompt: &(impl Fn(&str) -> String + ?Sized),
    ) -> Result<String> {
        let metadata = self.ai_client.get_metadata();
        let budget = TokenBudget::from_metadata(&metadata);

        let yaml =
            crate::data::to_yaml(ai_view).context("Failed to serialize repository view to YAML")?;
        let user_prompt = build_user_prompt(&yaml);

        let estimate = budget.validate_prompt(system_prompt, &user_prompt)?;
        debug!(
            model = %metadata.model,
            estimated_tokens = estimate.estimated_tokens,
            available_tokens = estimate.available_tokens,
            utilization_pct = format!("{:.1}%", estimate.utilization_pct),
            "Token budget check passed"
        );

        Ok(user_prompt)
    }

    /// Tests whether the full diff fits the token budget.
    ///
    /// Returns `Ok(Ok(user_prompt))` when the full diff fits,
    /// `Ok(Err(BudgetExceeded))` when it does not, or a top-level error
    /// on serialization failure.
    ///
    /// Generic over the view type so the diff-driven path
    /// (`RepositoryViewForAI`) and the commit-driven path
    /// (`RepositoryViewForAiFromCommits`) can share the same budget check.
    fn try_full_diff_budget<V: serde::Serialize>(
        &self,
        ai_view: &V,
        system_prompt: &str,
        build_user_prompt: &(impl Fn(&str) -> String + ?Sized),
    ) -> Result<std::result::Result<String, BudgetExceeded>> {
        let metadata = self.ai_client.get_metadata();
        let budget = TokenBudget::from_metadata(&metadata);

        let yaml =
            crate::data::to_yaml(ai_view).context("Failed to serialize repository view to YAML")?;
        let user_prompt = build_user_prompt(&yaml);

        if let Ok(estimate) = budget.validate_prompt(system_prompt, &user_prompt) {
            debug!(
                model = %metadata.model,
                estimated_tokens = estimate.estimated_tokens,
                available_tokens = estimate.available_tokens,
                utilization_pct = format!("{:.1}%", estimate.utilization_pct),
                "Token budget check passed"
            );
            return Ok(Ok(user_prompt));
        }

        Ok(Err(BudgetExceeded {
            available_input_tokens: budget.available_input_tokens(),
        }))
    }

    /// Generates an amendment for a single commit whose diff exceeds the
    /// token budget by splitting it into file-level chunks.
    ///
    /// Uses [`pack_file_diffs`](crate::claude::diff_pack::pack_file_diffs) to
    /// create chunks, sends one AI request per chunk, then runs a merge pass
    /// to synthesize a single [`Amendment`].
    async fn generate_amendment_split(
        &self,
        commit: &crate::git::CommitInfo,
        repo_view_for_ai: &RepositoryViewForAI,
        system_prompt: &str,
        build_user_prompt: &(dyn Fn(&str) -> String + Sync),
        available_input_tokens: usize,
        fresh: bool,
    ) -> Result<Amendment> {
        use crate::claude::batch::{
            PER_COMMIT_METADATA_OVERHEAD_TOKENS, USER_PROMPT_TEMPLATE_OVERHEAD_TOKENS,
            VIEW_ENVELOPE_OVERHEAD_TOKENS,
        };
        use crate::claude::diff_pack::pack_file_diffs;
        use crate::claude::token_budget;
        use crate::git::commit::CommitInfoForAI;

        // Compute effective capacity for diff packing by subtracting overhead
        // that will be added when the full prompt is assembled. This mirrors
        // the calculation in `batch::plan_batches`.
        //
        // Each chunk includes the FULL original_message and diff_summary (not
        // just the partial diff), so we must subtract those from capacity.
        // We also subtract user prompt template overhead for instruction text.
        let system_prompt_tokens = token_budget::estimate_tokens(system_prompt);
        let commit_text_tokens = token_budget::estimate_tokens(&commit.original_message)
            + token_budget::estimate_tokens(&commit.analysis.diff_summary);
        let chunk_capacity = available_input_tokens
            .saturating_sub(system_prompt_tokens)
            .saturating_sub(VIEW_ENVELOPE_OVERHEAD_TOKENS)
            .saturating_sub(PER_COMMIT_METADATA_OVERHEAD_TOKENS)
            .saturating_sub(USER_PROMPT_TEMPLATE_OVERHEAD_TOKENS)
            .saturating_sub(commit_text_tokens);

        debug!(
            commit = %&commit.hash[..8],
            available_input_tokens,
            system_prompt_tokens,
            envelope_overhead = VIEW_ENVELOPE_OVERHEAD_TOKENS,
            metadata_overhead = PER_COMMIT_METADATA_OVERHEAD_TOKENS,
            template_overhead = USER_PROMPT_TEMPLATE_OVERHEAD_TOKENS,
            commit_text_tokens,
            chunk_capacity,
            "Split dispatch: computed chunk capacity"
        );

        let plan = pack_file_diffs(&commit.hash, &commit.analysis.file_diffs, chunk_capacity)
            .with_context(|| {
                format!(
                    "Failed to plan diff chunks for commit {}",
                    &commit.hash[..8]
                )
            })?;

        let total_chunks = plan.chunks.len();
        debug!(
            commit = %&commit.hash[..8],
            chunks = total_chunks,
            chunk_capacity,
            "Split dispatch: processing commit in chunks"
        );

        let mut chunk_amendments = Vec::with_capacity(total_chunks);
        for (i, chunk) in plan.chunks.iter().enumerate() {
            let mut partial = CommitInfoForAI::from_commit_info_partial_with_overrides(
                commit.clone(),
                &chunk.file_paths,
                &chunk.diff_overrides,
            )
            .with_context(|| {
                format!(
                    "Failed to build partial view for chunk {}/{} of commit {}",
                    i + 1,
                    total_chunks,
                    &commit.hash[..8]
                )
            })?;

            if fresh {
                partial.base.original_message =
                    "(Original message hidden - generate fresh message from diff)".to_string();
            }

            let partial_view = repo_view_for_ai.single_commit_view_for_ai(&partial);

            // Log the actual diff content size for this chunk
            let diff_content_len = partial.base.analysis.diff_content.len();
            let diff_content_tokens =
                token_budget::estimate_tokens_from_char_count(diff_content_len);
            debug!(
                commit = %&commit.hash[..8],
                chunk_index = i,
                diff_content_len,
                diff_content_tokens,
                "Split dispatch: chunk diff content size"
            );

            let user_prompt =
                self.build_prompt_fitting_budget(&partial_view, system_prompt, build_user_prompt)?;

            info!(
                commit = %&commit.hash[..8],
                chunk = i + 1,
                total_chunks,
                user_prompt_len = user_prompt.len(),
                "Split dispatch: sending chunk to AI"
            );

            let content = match self
                .send_with_optional_schema(
                    system_prompt,
                    &user_prompt,
                    self.schema_if_supported(response_schema::amendment_file_schema()),
                )
                .await
            {
                Ok(content) => content,
                Err(e) => {
                    // Log the underlying error before wrapping
                    tracing::error!(
                        commit = %&commit.hash[..8],
                        chunk = i + 1,
                        error = %e,
                        error_debug = ?e,
                        "Split dispatch: AI request failed"
                    );
                    return Err(e).with_context(|| {
                        format!(
                            "Chunk {}/{} failed for commit {}",
                            i + 1,
                            total_chunks,
                            &commit.hash[..8]
                        )
                    });
                }
            };

            info!(
                commit = %&commit.hash[..8],
                chunk = i + 1,
                response_len = content.len(),
                "Split dispatch: received chunk response"
            );

            let amendment_file = self.parse_amendment_response(&content).with_context(|| {
                format!(
                    "Failed to parse chunk {}/{} response for commit {}",
                    i + 1,
                    total_chunks,
                    &commit.hash[..8]
                )
            })?;

            if let Some(amendment) = amendment_file.amendments.into_iter().next() {
                chunk_amendments.push(amendment);
            }
        }

        self.merge_amendment_chunks(
            &commit.hash,
            &commit.original_message,
            &commit.analysis.diff_summary,
            &chunk_amendments,
        )
        .await
    }

    /// Runs an AI reduce pass to synthesize a single amendment from partial
    /// chunk amendments for the same commit.
    ///
    /// Follows the same pattern as
    /// [`refine_amendments_coherence`](Self::refine_amendments_coherence).
    async fn merge_amendment_chunks(
        &self,
        commit_hash: &str,
        original_message: &str,
        diff_summary: &str,
        chunk_amendments: &[Amendment],
    ) -> Result<Amendment> {
        let system_prompt =
            self.adjusted_system_prompt(prompts::AMENDMENT_CHUNK_MERGE_SYSTEM_PROMPT.to_string());
        let user_prompt = prompts::generate_chunk_merge_user_prompt(
            commit_hash,
            original_message,
            diff_summary,
            chunk_amendments,
        );

        self.validate_prompt_budget(&system_prompt, &user_prompt)?;

        let content = self
            .send_with_optional_schema(
                &system_prompt,
                &user_prompt,
                self.schema_if_supported(response_schema::amendment_file_schema()),
            )
            .await
            .context("Merge pass failed for chunk amendments")?;

        let amendment_file = self
            .parse_amendment_response(&content)
            .context("Failed to parse merge pass response")?;

        amendment_file
            .amendments
            .into_iter()
            .next()
            .context("Merge pass returned no amendments")
    }

    /// Generates an amendment for a single commit, using split dispatch
    /// if the full diff exceeds the token budget.
    ///
    /// Tries the full diff first. If it exceeds the budget and the commit
    /// has file-level diffs, falls back to
    /// [`generate_amendment_split`](Self::generate_amendment_split).
    async fn generate_amendment_for_commit(
        &self,
        commit: &crate::git::CommitInfo,
        repo_view_for_ai: &RepositoryViewForAI,
        system_prompt: &str,
        build_user_prompt: &(dyn Fn(&str) -> String + Sync),
        fresh: bool,
    ) -> Result<Amendment> {
        let mut ai_commit = crate::git::commit::CommitInfoForAI::from_commit_info(commit.clone())?;
        if fresh {
            ai_commit.base.original_message =
                "(Original message hidden - generate fresh message from diff)".to_string();
        }
        let single_view = repo_view_for_ai.single_commit_view_for_ai(&ai_commit);

        match self.try_full_diff_budget(&single_view, system_prompt, build_user_prompt)? {
            Ok(user_prompt) => {
                let amendment_file = self
                    .send_and_parse_amendment_with_retry(system_prompt, &user_prompt)
                    .await?;
                amendment_file
                    .amendments
                    .into_iter()
                    .next()
                    .context("AI returned no amendments for commit")
            }
            Err(exceeded) => {
                if commit.analysis.file_diffs.is_empty() {
                    anyhow::bail!(
                        "Token budget exceeded for commit {} but no file-level diffs available for split dispatch",
                        &commit.hash[..8]
                    );
                }
                self.generate_amendment_split(
                    commit,
                    repo_view_for_ai,
                    system_prompt,
                    build_user_prompt,
                    exceeded.available_input_tokens,
                    fresh,
                )
                .await
            }
        }
    }

    /// Checks a single commit whose diff exceeds the token budget by
    /// splitting it into file-level chunks.
    ///
    /// Uses [`pack_file_diffs`](crate::claude::diff_pack::pack_file_diffs) to
    /// create chunks, sends one check request per chunk, then merges results
    /// deterministically (issue union + dedup). Runs an AI reduce pass only
    /// when at least one chunk returns a suggestion.
    async fn check_commit_split(
        &self,
        commit: &crate::git::CommitInfo,
        repo_view: &RepositoryView,
        system_prompt: &str,
        valid_scopes: &[crate::data::context::ScopeDefinition],
        include_suggestions: bool,
        available_input_tokens: usize,
    ) -> Result<crate::data::check::CheckReport> {
        use crate::claude::batch::{
            PER_COMMIT_METADATA_OVERHEAD_TOKENS, USER_PROMPT_TEMPLATE_OVERHEAD_TOKENS,
            VIEW_ENVELOPE_OVERHEAD_TOKENS,
        };
        use crate::claude::diff_pack::pack_file_diffs;
        use crate::claude::token_budget;
        use crate::data::check::{CommitCheckResult, CommitIssue, IssueSeverity};
        use crate::git::commit::CommitInfoForAI;

        // Compute effective capacity for diff packing by subtracting overhead
        // that will be added when the full prompt is assembled. This mirrors
        // the calculation in `batch::plan_batches`.
        //
        // Each chunk includes the FULL original_message and diff_summary (not
        // just the partial diff), so we must subtract those from capacity.
        // We also subtract user prompt template overhead for instruction text.
        let system_prompt_tokens = token_budget::estimate_tokens(system_prompt);
        let commit_text_tokens = token_budget::estimate_tokens(&commit.original_message)
            + token_budget::estimate_tokens(&commit.analysis.diff_summary);
        let chunk_capacity = available_input_tokens
            .saturating_sub(system_prompt_tokens)
            .saturating_sub(VIEW_ENVELOPE_OVERHEAD_TOKENS)
            .saturating_sub(PER_COMMIT_METADATA_OVERHEAD_TOKENS)
            .saturating_sub(USER_PROMPT_TEMPLATE_OVERHEAD_TOKENS)
            .saturating_sub(commit_text_tokens);

        debug!(
            commit = %&commit.hash[..8],
            available_input_tokens,
            system_prompt_tokens,
            envelope_overhead = VIEW_ENVELOPE_OVERHEAD_TOKENS,
            metadata_overhead = PER_COMMIT_METADATA_OVERHEAD_TOKENS,
            template_overhead = USER_PROMPT_TEMPLATE_OVERHEAD_TOKENS,
            commit_text_tokens,
            chunk_capacity,
            "Check split dispatch: computed chunk capacity"
        );

        let plan = pack_file_diffs(&commit.hash, &commit.analysis.file_diffs, chunk_capacity)
            .with_context(|| {
                format!(
                    "Failed to plan diff chunks for commit {}",
                    &commit.hash[..8]
                )
            })?;

        let total_chunks = plan.chunks.len();
        debug!(
            commit = %&commit.hash[..8],
            chunks = total_chunks,
            chunk_capacity,
            "Check split dispatch: processing commit in chunks"
        );

        let build_user_prompt =
            |yaml: &str| prompts::generate_check_user_prompt(yaml, include_suggestions);

        let mut chunk_results = Vec::with_capacity(total_chunks);
        for (i, chunk) in plan.chunks.iter().enumerate() {
            let mut partial = CommitInfoForAI::from_commit_info_partial_with_overrides(
                commit.clone(),
                &chunk.file_paths,
                &chunk.diff_overrides,
            )
            .with_context(|| {
                format!(
                    "Failed to build partial view for chunk {}/{} of commit {}",
                    i + 1,
                    total_chunks,
                    &commit.hash[..8]
                )
            })?;

            partial.run_pre_validation_checks(valid_scopes);

            let partial_view = RepositoryViewForAI::from_repository_view(repo_view.clone())
                .context("Failed to enhance repository view with diff content")?
                .single_commit_view_for_ai(&partial);

            let user_prompt =
                self.build_prompt_fitting_budget(&partial_view, system_prompt, &build_user_prompt)?;

            let content = self
                .send_with_optional_schema(
                    system_prompt,
                    &user_prompt,
                    self.schema_if_supported(response_schema::check_response_schema()),
                )
                .await
                .with_context(|| {
                    format!(
                        "Check chunk {}/{} failed for commit {}",
                        i + 1,
                        total_chunks,
                        &commit.hash[..8]
                    )
                })?;

            let report = self
                .parse_check_response(&content, repo_view)
                .with_context(|| {
                    format!(
                        "Failed to parse check chunk {}/{} response for commit {}",
                        i + 1,
                        total_chunks,
                        &commit.hash[..8]
                    )
                })?;

            if let Some(result) = report.commits.into_iter().next() {
                chunk_results.push(result);
            }
        }

        // Deterministic merge: union issues, dedup by (rule, severity, section)
        let mut seen = std::collections::HashSet::new();
        let mut merged_issues: Vec<CommitIssue> = Vec::new();
        for result in &chunk_results {
            for issue in &result.issues {
                let key: (String, IssueSeverity, String) =
                    (issue.rule.clone(), issue.severity, issue.section.clone());
                if seen.insert(key) {
                    merged_issues.push(issue.clone());
                }
            }
        }

        let passes = chunk_results.iter().all(|r| r.passes);

        // AI reduce pass for suggestion/summary only when needed
        let has_suggestions = chunk_results.iter().any(|r| r.suggestion.is_some());

        let (merged_suggestion, merged_summary) = if has_suggestions {
            self.merge_check_chunks(
                &commit.hash,
                &commit.original_message,
                &commit.analysis.diff_summary,
                passes,
                &chunk_results,
                repo_view,
            )
            .await?
        } else {
            // Take first non-None summary
            let summary = chunk_results.iter().find_map(|r| r.summary.clone());
            (None, summary)
        };

        let original_message = commit
            .original_message
            .lines()
            .next()
            .unwrap_or("")
            .to_string();

        let merged_result = CommitCheckResult {
            hash: commit.hash.clone(),
            message: original_message,
            issues: merged_issues,
            suggestion: merged_suggestion,
            passes,
            summary: merged_summary,
        };

        Ok(crate::data::check::CheckReport::new(vec![merged_result]))
    }

    /// Runs an AI reduce pass to synthesize a single suggestion and summary
    /// from partial chunk check results for the same commit.
    ///
    /// Only called when at least one chunk returned a suggestion.
    async fn merge_check_chunks(
        &self,
        commit_hash: &str,
        original_message: &str,
        diff_summary: &str,
        passes: bool,
        chunk_results: &[crate::data::check::CommitCheckResult],
        repo_view: &RepositoryView,
    ) -> Result<(Option<crate::data::check::CommitSuggestion>, Option<String>)> {
        let suggestions: Vec<&crate::data::check::CommitSuggestion> = chunk_results
            .iter()
            .filter_map(|r| r.suggestion.as_ref())
            .collect();

        let summaries: Vec<Option<&str>> =
            chunk_results.iter().map(|r| r.summary.as_deref()).collect();

        let system_prompt =
            self.adjusted_system_prompt(prompts::CHECK_CHUNK_MERGE_SYSTEM_PROMPT.to_string());
        let user_prompt = prompts::generate_check_chunk_merge_user_prompt(
            commit_hash,
            original_message,
            diff_summary,
            passes,
            &suggestions,
            &summaries,
        );

        self.validate_prompt_budget(&system_prompt, &user_prompt)?;

        let content = self
            .send_with_optional_schema(
                &system_prompt,
                &user_prompt,
                self.schema_if_supported(response_schema::check_response_schema()),
            )
            .await
            .context("Merge pass failed for check chunk suggestions")?;

        let report = self
            .parse_check_response(&content, repo_view)
            .context("Failed to parse check merge pass response")?;

        let result = report.commits.into_iter().next();
        Ok(match result {
            Some(r) => (r.suggestion, r.summary),
            None => (None, None),
        })
    }

    /// Sends a raw prompt to the AI client and returns the text response.
    pub async fn send_message(&self, system_prompt: &str, user_prompt: &str) -> Result<String> {
        self.validate_prompt_budget(system_prompt, user_prompt)?;
        self.ai_client
            .send_request(system_prompt, user_prompt)
            .await
    }

    /// Creates a new Claude client with API key from environment variables.
    // omni-dev: coverage ignore reason="process-bound: hardcodes SystemEnv with no injectable seam (unlike create_default_claude_client_with's tested Default arm, which this mirrors) and has no in-crate caller; testing it would mean mutating real process env for shared credential vars"
    pub fn from_env(model: String) -> Result<Self> {
        // Try to get API key from environment variables
        let api_key = secret_var_any(
            &crate::utils::env::SystemEnv,
            &["CLAUDE_API_KEY", "ANTHROPIC_API_KEY"],
        )?
        .ok_or(ClaudeError::ApiKeyNotFound)?;

        let ai_client = ClaudeAiClient::new(model, api_key.expose_secret().to_string(), None)?;
        Ok(Self::new(Box::new(ai_client)))
    }
    // omni-dev: coverage end

    /// Generates commit message amendments from repository view.
    pub async fn generate_amendments(&self, repo_view: &RepositoryView) -> Result<AmendmentFile> {
        self.generate_amendments_with_options(repo_view, false)
            .await
    }

    /// Generates commit message amendments from repository view with options.
    ///
    /// If `fresh` is true, ignores existing commit messages and generates new ones
    /// based solely on the diff content.
    ///
    /// For single-commit views whose full diff exceeds the token budget,
    /// splits the diff into file-level chunks and dispatches multiple AI
    /// requests, then merges results. Multi-commit views fall back to
    /// progressive diff reduction (the caller retries individually on
    /// failure).
    pub async fn generate_amendments_with_options(
        &self,
        repo_view: &RepositoryView,
        fresh: bool,
    ) -> Result<AmendmentFile> {
        // Convert to AI-enhanced view with diff content
        let ai_repo_view =
            RepositoryViewForAI::from_repository_view_with_options(repo_view.clone(), fresh)
                .context("Failed to enhance repository view with diff content")?;

        let system_prompt = self.adjusted_system_prompt(prompts::SYSTEM_PROMPT.to_string());
        let build_user_prompt = |yaml: &str| prompts::generate_user_prompt(yaml);

        // Try full view first; fall back to per-commit split dispatch
        match self.try_full_diff_budget(&ai_repo_view, &system_prompt, &build_user_prompt)? {
            Ok(user_prompt) => {
                self.send_and_parse_amendment_with_retry(&system_prompt, &user_prompt)
                    .await
            }
            Err(_exceeded) => {
                let mut amendments = Vec::new();
                for commit in &repo_view.commits {
                    let amendment = self
                        .generate_amendment_for_commit(
                            commit,
                            &ai_repo_view,
                            &system_prompt,
                            &build_user_prompt,
                            fresh,
                        )
                        .await?;
                    amendments.push(amendment);
                }
                Ok(AmendmentFile { amendments })
            }
        }
    }

    /// Generates contextual commit message amendments with enhanced intelligence.
    pub async fn generate_contextual_amendments(
        &self,
        repo_view: &RepositoryView,
        context: &CommitContext,
    ) -> Result<AmendmentFile> {
        self.generate_contextual_amendments_with_options(repo_view, context, false)
            .await
    }

    /// Generates contextual commit message amendments with options.
    ///
    /// If `fresh` is true, ignores existing commit messages and generates new ones
    /// based solely on the diff content.
    ///
    /// For single-commit views whose full diff exceeds the token budget,
    /// splits the diff into file-level chunks and dispatches multiple AI
    /// requests, then merges results. Multi-commit views fall back to
    /// progressive diff reduction.
    pub async fn generate_contextual_amendments_with_options(
        &self,
        repo_view: &RepositoryView,
        context: &CommitContext,
        fresh: bool,
    ) -> Result<AmendmentFile> {
        // Convert to AI-enhanced view with diff content
        let ai_repo_view =
            RepositoryViewForAI::from_repository_view_with_options(repo_view.clone(), fresh)
                .context("Failed to enhance repository view with diff content")?;

        // Generate contextual prompts using intelligence
        let prompt_style = self.ai_client.get_metadata().prompt_style();
        let system_prompt = self.adjusted_system_prompt(
            prompts::generate_contextual_system_prompt_for_provider(context, prompt_style),
        );

        // Debug logging to troubleshoot custom commit type issue
        match &context.project.commit_guidelines {
            Some(guidelines) => {
                debug!(length = guidelines.len(), "Project commit guidelines found");
                debug!(guidelines = %guidelines, "Commit guidelines content");
            }
            None => {
                debug!("No project commit guidelines found");
            }
        }

        let build_user_prompt =
            |yaml: &str| prompts::generate_contextual_user_prompt(yaml, context);

        // Try full view first; fall back to per-commit split dispatch
        match self.try_full_diff_budget(&ai_repo_view, &system_prompt, &build_user_prompt)? {
            Ok(user_prompt) => {
                self.send_and_parse_amendment_with_retry(&system_prompt, &user_prompt)
                    .await
            }
            Err(_exceeded) => {
                let mut amendments = Vec::new();
                for commit in &repo_view.commits {
                    let amendment = self
                        .generate_amendment_for_commit(
                            commit,
                            &ai_repo_view,
                            &system_prompt,
                            &build_user_prompt,
                            fresh,
                        )
                        .await?;
                    amendments.push(amendment);
                }
                Ok(AmendmentFile { amendments })
            }
        }
    }

    /// Parses Claude's YAML response into an AmendmentFile.
    fn parse_amendment_response(&self, content: &str) -> Result<AmendmentFile> {
        // Extract YAML from potential markdown wrapper
        let yaml_content = self.extract_yaml_from_response(content);

        // Try to parse YAML using our hybrid YAML parser
        let amendment_file: AmendmentFile = crate::data::from_yaml(&yaml_content).map_err(|e| {
            debug!(
                error = %e,
                content_length = content.len(),
                yaml_length = yaml_content.len(),
                "YAML parsing failed"
            );
            debug!(content = %content, "Raw Claude response");
            debug!(yaml = %yaml_content, "Extracted YAML content");

            // Try to provide more helpful error messages for common issues
            if yaml_content.lines().any(|line| line.contains('\t')) {
                ClaudeError::AmendmentParsingFailed("YAML parsing error: Found tab characters. YAML requires spaces for indentation.".to_string())
            } else if yaml_content.lines().any(|line| line.trim().starts_with('-') && !line.trim().starts_with("- ")) {
                ClaudeError::AmendmentParsingFailed("YAML parsing error: List items must have a space after the dash (- item).".to_string())
            } else {
                ClaudeError::AmendmentParsingFailed(format!("YAML parsing error: {e}"))
            }
        })?;

        // Validate the parsed amendments
        amendment_file
            .validate()
            .map_err(|e| ClaudeError::AmendmentParsingFailed(format!("Validation error: {e}")))?;

        Ok(amendment_file)
    }

    /// Sends a prompt to the AI and parses the response as an [`AmendmentFile`],
    /// retrying on parse or request failures.
    ///
    /// Mirrors the retry pattern in [`check_commits_with_retry`](Self::check_commits_with_retry):
    /// up to [`AMENDMENT_PARSE_MAX_RETRIES`] additional attempts after the first
    /// failure. Logs a warning via `eprintln!` and a `debug!` trace on each retry.
    /// Returns the last error if all attempts are exhausted.
    async fn send_and_parse_amendment_with_retry(
        &self,
        system_prompt: &str,
        user_prompt: &str,
    ) -> Result<AmendmentFile> {
        let mut last_error = None;
        for attempt in 0..=AMENDMENT_PARSE_MAX_RETRIES {
            match self
                .send_with_optional_schema(
                    system_prompt,
                    user_prompt,
                    self.schema_if_supported(response_schema::amendment_file_schema()),
                )
                .await
            {
                Ok(content) => match self.parse_amendment_response(&content) {
                    Ok(amendment_file) => return Ok(amendment_file),
                    Err(e) => {
                        if attempt < AMENDMENT_PARSE_MAX_RETRIES {
                            eprintln!(
                                "warning: failed to parse amendment response (attempt {}), retrying...",
                                attempt + 1
                            );
                            debug!(error = %e, attempt = attempt + 1, "Amendment response parse failed, retrying");
                        }
                        last_error = Some(e);
                    }
                },
                // A permanent failure cannot succeed on a retry, so report it
                // immediately instead of repeating a doomed request (#1333).
                Err(e) if !ai_error_is_transient(&e) => return Err(e),
                Err(e) => {
                    if attempt < AMENDMENT_PARSE_MAX_RETRIES {
                        eprintln!(
                            "warning: AI request failed (attempt {}), retrying...",
                            attempt + 1
                        );
                        debug!(error = %e, attempt = attempt + 1, "AI request failed, retrying");
                    }
                    last_error = Some(e);
                }
            }
        }
        Err(last_error
            .unwrap_or_else(|| anyhow::anyhow!("Amendment generation failed after retries")))
    }

    /// Parses an AI response as PR content YAML.
    fn parse_pr_response(&self, content: &str) -> Result<crate::cli::git::PrContent> {
        let yaml_content = content.trim();
        crate::data::from_yaml(yaml_content)
            .context("Failed to parse AI response as YAML. AI may have returned malformed output.")
    }

    /// Generates PR content for a single commit whose diff exceeds the token
    /// budget by splitting it into file-level chunks.
    ///
    /// Analogous to [`generate_amendment_split`](Self::generate_amendment_split)
    /// but produces [`PrContent`](crate::cli::git::PrContent) instead of an
    /// amendment.
    async fn generate_pr_content_split(
        &self,
        commit: &crate::git::CommitInfo,
        repo_view_for_ai: &RepositoryViewForAI,
        system_prompt: &str,
        build_user_prompt: &(dyn Fn(&str) -> String + Sync),
        available_input_tokens: usize,
        pr_template: &str,
    ) -> Result<crate::cli::git::PrContent> {
        use crate::claude::batch::{
            PER_COMMIT_METADATA_OVERHEAD_TOKENS, USER_PROMPT_TEMPLATE_OVERHEAD_TOKENS,
            VIEW_ENVELOPE_OVERHEAD_TOKENS,
        };
        use crate::claude::diff_pack::pack_file_diffs;
        use crate::claude::token_budget;
        use crate::git::commit::CommitInfoForAI;

        // Compute effective capacity for diff packing by subtracting overhead
        // that will be added when the full prompt is assembled. This mirrors
        // the calculation in `batch::plan_batches`.
        //
        // Each chunk includes the FULL original_message and diff_summary (not
        // just the partial diff), so we must subtract those from capacity.
        // We also subtract user prompt template overhead for instruction text.
        let system_prompt_tokens = token_budget::estimate_tokens(system_prompt);
        let commit_text_tokens = token_budget::estimate_tokens(&commit.original_message)
            + token_budget::estimate_tokens(&commit.analysis.diff_summary);
        let chunk_capacity = available_input_tokens
            .saturating_sub(system_prompt_tokens)
            .saturating_sub(VIEW_ENVELOPE_OVERHEAD_TOKENS)
            .saturating_sub(PER_COMMIT_METADATA_OVERHEAD_TOKENS)
            .saturating_sub(USER_PROMPT_TEMPLATE_OVERHEAD_TOKENS)
            .saturating_sub(commit_text_tokens);

        debug!(
            commit = %&commit.hash[..8],
            available_input_tokens,
            system_prompt_tokens,
            envelope_overhead = VIEW_ENVELOPE_OVERHEAD_TOKENS,
            metadata_overhead = PER_COMMIT_METADATA_OVERHEAD_TOKENS,
            template_overhead = USER_PROMPT_TEMPLATE_OVERHEAD_TOKENS,
            commit_text_tokens,
            chunk_capacity,
            "PR split dispatch: computed chunk capacity"
        );

        let plan = pack_file_diffs(&commit.hash, &commit.analysis.file_diffs, chunk_capacity)
            .with_context(|| {
                format!(
                    "Failed to plan diff chunks for commit {}",
                    &commit.hash[..8]
                )
            })?;

        let total_chunks = plan.chunks.len();
        debug!(
            commit = %&commit.hash[..8],
            chunks = total_chunks,
            chunk_capacity,
            "PR split dispatch: processing commit in chunks"
        );

        let mut chunk_contents = Vec::with_capacity(total_chunks);
        for (i, chunk) in plan.chunks.iter().enumerate() {
            let partial = CommitInfoForAI::from_commit_info_partial_with_overrides(
                commit.clone(),
                &chunk.file_paths,
                &chunk.diff_overrides,
            )
            .with_context(|| {
                format!(
                    "Failed to build partial view for chunk {}/{} of commit {}",
                    i + 1,
                    total_chunks,
                    &commit.hash[..8]
                )
            })?;

            let partial_view = repo_view_for_ai.single_commit_view_for_ai(&partial);

            let user_prompt =
                self.build_prompt_fitting_budget(&partial_view, system_prompt, build_user_prompt)?;

            let content = self
                .send_with_optional_schema(
                    system_prompt,
                    &user_prompt,
                    self.schema_if_supported(response_schema::pr_content_schema()),
                )
                .await
                .with_context(|| {
                    format!(
                        "PR chunk {}/{} failed for commit {}",
                        i + 1,
                        total_chunks,
                        &commit.hash[..8]
                    )
                })?;

            let pr_content = self.parse_pr_response(&content).with_context(|| {
                format!(
                    "Failed to parse PR chunk {}/{} response for commit {}",
                    i + 1,
                    total_chunks,
                    &commit.hash[..8]
                )
            })?;

            chunk_contents.push(pr_content);
        }

        self.merge_pr_content_chunks(&chunk_contents, pr_template)
            .await
    }

    /// Runs an AI reduce pass to synthesize a single PR content from partial
    /// per-commit or per-chunk PR contents.
    async fn merge_pr_content_chunks(
        &self,
        partial_contents: &[crate::cli::git::PrContent],
        pr_template: &str,
    ) -> Result<crate::cli::git::PrContent> {
        let system_prompt =
            self.adjusted_system_prompt(prompts::PR_CONTENT_MERGE_SYSTEM_PROMPT.to_string());
        let user_prompt =
            prompts::generate_pr_content_merge_user_prompt(partial_contents, pr_template);

        self.validate_prompt_budget(&system_prompt, &user_prompt)?;

        let content = self
            .send_with_optional_schema(
                &system_prompt,
                &user_prompt,
                self.schema_if_supported(response_schema::pr_content_schema()),
            )
            .await
            .context("Merge pass failed for PR content chunks")?;

        self.parse_pr_response(&content)
            .context("Failed to parse PR content merge pass response")
    }

    /// Generates PR content for a single commit, using split dispatch if needed.
    async fn generate_pr_content_for_commit(
        &self,
        commit: &crate::git::CommitInfo,
        repo_view_for_ai: &RepositoryViewForAI,
        system_prompt: &str,
        build_user_prompt: &(dyn Fn(&str) -> String + Sync),
        pr_template: &str,
    ) -> Result<crate::cli::git::PrContent> {
        let ai_commit = crate::git::commit::CommitInfoForAI::from_commit_info(commit.clone())?;
        let single_view = repo_view_for_ai.single_commit_view_for_ai(&ai_commit);

        match self.try_full_diff_budget(&single_view, system_prompt, build_user_prompt)? {
            Ok(user_prompt) => {
                let content = self
                    .send_with_optional_schema(
                        system_prompt,
                        &user_prompt,
                        self.schema_if_supported(response_schema::pr_content_schema()),
                    )
                    .await?;
                self.parse_pr_response(&content)
            }
            Err(exceeded) => {
                if commit.analysis.file_diffs.is_empty() {
                    anyhow::bail!(
                        "Token budget exceeded for commit {} but no file-level diffs available for split dispatch",
                        &commit.hash[..8]
                    );
                }
                self.generate_pr_content_split(
                    commit,
                    repo_view_for_ai,
                    system_prompt,
                    build_user_prompt,
                    exceeded.available_input_tokens,
                    pr_template,
                )
                .await
            }
        }
    }

    /// Generates AI-powered PR content (title + description) from repository view and template.
    pub async fn generate_pr_content(
        &self,
        repo_view: &RepositoryView,
        pr_template: &str,
    ) -> Result<crate::cli::git::PrContent> {
        // Convert to AI-enhanced view with diff content
        let ai_repo_view = RepositoryViewForAI::from_repository_view(repo_view.clone())
            .context("Failed to enhance repository view with diff content")?;

        let system_prompt =
            self.adjusted_system_prompt(prompts::PR_GENERATION_SYSTEM_PROMPT.to_string());

        let build_user_prompt =
            |yaml: &str| prompts::generate_pr_description_prompt(yaml, pr_template);

        // Try full view first; fall back to per-commit split dispatch
        match self.try_full_diff_budget(&ai_repo_view, &system_prompt, &build_user_prompt)? {
            Ok(user_prompt) => {
                let content = self
                    .send_with_optional_schema(
                        &system_prompt,
                        &user_prompt,
                        self.schema_if_supported(response_schema::pr_content_schema()),
                    )
                    .await?;
                self.parse_pr_response(&content)
            }
            Err(_exceeded) => {
                let mut per_commit_contents = Vec::new();
                for commit in &repo_view.commits {
                    let pr = self
                        .generate_pr_content_for_commit(
                            commit,
                            &ai_repo_view,
                            &system_prompt,
                            &build_user_prompt,
                            pr_template,
                        )
                        .await?;
                    per_commit_contents.push(pr);
                }
                if per_commit_contents.len() == 1 {
                    return per_commit_contents
                        .into_iter()
                        .next()
                        .context("Per-commit PR contents unexpectedly empty");
                }
                self.merge_pr_content_chunks(&per_commit_contents, pr_template)
                    .await
            }
        }
    }

    /// Generates AI-powered PR content with project context (title + description).
    pub async fn generate_pr_content_with_context(
        &self,
        repo_view: &RepositoryView,
        pr_template: &str,
        context: &crate::data::context::CommitContext,
    ) -> Result<crate::cli::git::PrContent> {
        // Convert to AI-enhanced view with diff content
        let ai_repo_view = RepositoryViewForAI::from_repository_view(repo_view.clone())
            .context("Failed to enhance repository view with diff content")?;

        // Generate contextual prompts for PR description with provider-specific handling
        let prompt_style = self.ai_client.get_metadata().prompt_style();
        let system_prompt = self.adjusted_system_prompt(
            prompts::generate_pr_system_prompt_with_context_for_provider(context, prompt_style),
        );

        let build_user_prompt = |yaml: &str| {
            prompts::generate_pr_description_prompt_with_context(yaml, pr_template, context)
        };

        // Try full view first; fall back to per-commit split dispatch
        match self.try_full_diff_budget(&ai_repo_view, &system_prompt, &build_user_prompt)? {
            Ok(user_prompt) => {
                let content = self
                    .send_with_optional_schema(
                        &system_prompt,
                        &user_prompt,
                        self.schema_if_supported(response_schema::pr_content_schema()),
                    )
                    .await?;

                debug!(
                    content_length = content.len(),
                    "Received AI response for PR content"
                );

                let pr_content = self.parse_pr_response(&content)?;

                debug!(
                    parsed_title = %pr_content.title,
                    parsed_description_length = pr_content.description.len(),
                    parsed_description_preview = %pr_content.description.lines().take(3).collect::<Vec<_>>().join("\\n"),
                    "Successfully parsed PR content from YAML"
                );

                Ok(pr_content)
            }
            Err(_exceeded) => {
                let mut per_commit_contents = Vec::new();
                for commit in &repo_view.commits {
                    let pr = self
                        .generate_pr_content_for_commit(
                            commit,
                            &ai_repo_view,
                            &system_prompt,
                            &build_user_prompt,
                            pr_template,
                        )
                        .await?;
                    per_commit_contents.push(pr);
                }
                if per_commit_contents.len() == 1 {
                    return per_commit_contents
                        .into_iter()
                        .next()
                        .context("Per-commit PR contents unexpectedly empty");
                }
                self.merge_pr_content_chunks(&per_commit_contents, pr_template)
                    .await
            }
        }
    }

    /// Generates AI-powered PR content from commit messages only (no diff).
    ///
    /// Used by `omni-dev git branch create pr --from-commits`. Builds a
    /// payload that contains commit messages and metadata (hash, author,
    /// date, detected type/scope) but **no diff content** — the full diff
    /// files are never read from disk for this path. Falls back to a
    /// per-commit split dispatch if the commit-message payload exceeds the
    /// token budget (rare; commit messages are small).
    pub async fn generate_pr_content_with_context_from_commits(
        &self,
        repo_view: &RepositoryView,
        pr_template: &str,
        context: &crate::data::context::CommitContext,
    ) -> Result<crate::cli::git::PrContent> {
        use crate::data::RepositoryViewForAiFromCommits;

        let commits_view = RepositoryViewForAiFromCommits::from_repository_view(repo_view.clone());

        let prompt_style = self.ai_client.get_metadata().prompt_style();
        let system_prompt = self.adjusted_system_prompt(
            prompts::generate_pr_system_prompt_from_commits_with_context_for_provider(
                context,
                prompt_style,
            ),
        );

        let build_user_prompt = |yaml: &str| {
            prompts::generate_pr_description_prompt_from_commits_with_context(
                yaml,
                pr_template,
                context,
            )
        };

        match self.try_full_diff_budget(&commits_view, &system_prompt, &build_user_prompt)? {
            Ok(user_prompt) => {
                let content = self
                    .send_with_optional_schema(
                        &system_prompt,
                        &user_prompt,
                        self.schema_if_supported(response_schema::pr_content_schema()),
                    )
                    .await?;

                debug!(
                    content_length = content.len(),
                    "Received AI response for from-commits PR content"
                );

                self.parse_pr_response(&content)
            }
            Err(_exceeded) => {
                let mut per_commit_contents = Vec::new();
                for commit in &commits_view.commits {
                    let pr = self
                        .generate_pr_content_for_commit_from_commits(
                            commit,
                            &commits_view,
                            &system_prompt,
                            &build_user_prompt,
                        )
                        .await?;
                    per_commit_contents.push(pr);
                }
                if per_commit_contents.len() == 1 {
                    return per_commit_contents
                        .into_iter()
                        .next()
                        .context("Per-commit PR contents unexpectedly empty");
                }
                self.merge_pr_content_chunks(&per_commit_contents, pr_template)
                    .await
            }
        }
    }

    /// Per-commit dispatch for the `--from-commits` path.
    ///
    /// Builds a single-commit view of the commit-message payload and
    /// sends it. No file-level fallback exists for this path because the
    /// payload is already a single commit message — if it does not fit
    /// the budget, the commit message itself is pathologically large and
    /// we surface a clear error rather than silently truncating.
    async fn generate_pr_content_for_commit_from_commits(
        &self,
        commit: &crate::data::CommitInfoFromCommits,
        commits_view: &crate::data::RepositoryViewForAiFromCommits,
        system_prompt: &str,
        build_user_prompt: &(dyn Fn(&str) -> String + Sync),
    ) -> Result<crate::cli::git::PrContent> {
        let single_view = commits_view.single_commit_view_from_commits(commit);

        match self.try_full_diff_budget(&single_view, system_prompt, build_user_prompt)? {
            Ok(user_prompt) => {
                let content = self
                    .send_with_optional_schema(
                        system_prompt,
                        &user_prompt,
                        self.schema_if_supported(response_schema::pr_content_schema()),
                    )
                    .await?;
                self.parse_pr_response(&content)
            }
            Err(_exceeded) => {
                anyhow::bail!(
                    "Token budget exceeded for commit {} in --from-commits mode; commit message is too large to fit",
                    &commit.hash[..8.min(commit.hash.len())]
                )
            }
        }
    }

    /// Checks commit messages against guidelines and returns a report.
    ///
    /// Validates commit messages against project guidelines or defaults,
    /// returning a structured report with issues and suggestions.
    pub async fn check_commits(
        &self,
        repo_view: &RepositoryView,
        guidelines: Option<&str>,
        include_suggestions: bool,
    ) -> Result<crate::data::check::CheckReport> {
        self.check_commits_with_scopes(repo_view, guidelines, &[], include_suggestions)
            .await
    }

    /// Checks commit messages against guidelines with valid scopes and returns a report.
    ///
    /// Validates commit messages against project guidelines or defaults,
    /// using the provided valid scopes for scope validation.
    pub async fn check_commits_with_scopes(
        &self,
        repo_view: &RepositoryView,
        guidelines: Option<&str>,
        valid_scopes: &[crate::data::context::ScopeDefinition],
        include_suggestions: bool,
    ) -> Result<crate::data::check::CheckReport> {
        self.check_commits_with_retry(repo_view, guidelines, valid_scopes, include_suggestions, 2)
            .await
    }

    /// Checks commit messages with retry logic for parse failures.
    ///
    /// For single-commit views whose full diff exceeds the token budget,
    /// splits the diff into file-level chunks and dispatches multiple AI
    /// requests, then merges results. Multi-commit views fall back to
    /// progressive diff reduction (the caller retries individually on
    /// failure).
    async fn check_commits_with_retry(
        &self,
        repo_view: &RepositoryView,
        guidelines: Option<&str>,
        valid_scopes: &[crate::data::context::ScopeDefinition],
        include_suggestions: bool,
        max_retries: u32,
    ) -> Result<crate::data::check::CheckReport> {
        // Generate system prompt with scopes
        let system_prompt = self.adjusted_system_prompt(
            prompts::generate_check_system_prompt_with_scopes(guidelines, valid_scopes),
        );

        let build_user_prompt =
            |yaml: &str| prompts::generate_check_user_prompt(yaml, include_suggestions);

        let mut ai_repo_view = RepositoryViewForAI::from_repository_view(repo_view.clone())
            .context("Failed to enhance repository view with diff content")?;
        for commit in &mut ai_repo_view.commits {
            commit.run_pre_validation_checks(valid_scopes);
        }

        // Try full view first; fall back to per-commit split dispatch
        match self.try_full_diff_budget(&ai_repo_view, &system_prompt, &build_user_prompt)? {
            Ok(user_prompt) => {
                // Full view fits: send with retry loop
                let mut last_error = None;
                for attempt in 0..=max_retries {
                    match self
                        .send_with_optional_schema(
                            &system_prompt,
                            &user_prompt,
                            self.schema_if_supported(response_schema::check_response_schema()),
                        )
                        .await
                    {
                        Ok(content) => match self.parse_check_response(&content, repo_view) {
                            Ok(report) => return Ok(report),
                            Err(e) => {
                                if attempt < max_retries {
                                    eprintln!(
                                        "warning: failed to parse AI response (attempt {}), retrying...",
                                        attempt + 1
                                    );
                                    debug!(error = %e, attempt = attempt + 1, "Check response parse failed, retrying");
                                }
                                last_error = Some(e);
                            }
                        },
                        // A permanent failure cannot succeed on a retry, so
                        // report it immediately (#1333).
                        Err(e) if !ai_error_is_transient(&e) => return Err(e),
                        Err(e) => {
                            if attempt < max_retries {
                                eprintln!(
                                    "warning: AI request failed (attempt {}), retrying...",
                                    attempt + 1
                                );
                                debug!(error = %e, attempt = attempt + 1, "AI request failed, retrying");
                            }
                            last_error = Some(e);
                        }
                    }
                }
                Err(last_error.unwrap_or_else(|| anyhow::anyhow!("Check failed after retries")))
            }
            Err(_exceeded) => {
                // Per-commit split dispatch
                let mut all_results = Vec::new();
                for commit in &repo_view.commits {
                    let single_view = repo_view.single_commit_view(commit);
                    let mut single_ai_view =
                        RepositoryViewForAI::from_repository_view(single_view.clone())
                            .context("Failed to enhance single-commit view with diff content")?;
                    for c in &mut single_ai_view.commits {
                        c.run_pre_validation_checks(valid_scopes);
                    }

                    match self.try_full_diff_budget(
                        &single_ai_view,
                        &system_prompt,
                        &build_user_prompt,
                    )? {
                        Ok(user_prompt) => {
                            let content = self
                                .send_with_optional_schema(
                                    &system_prompt,
                                    &user_prompt,
                                    self.schema_if_supported(
                                        response_schema::check_response_schema(),
                                    ),
                                )
                                .await?;
                            let report = self.parse_check_response(&content, &single_view)?;
                            all_results.extend(report.commits);
                        }
                        Err(exceeded) => {
                            if commit.analysis.file_diffs.is_empty() {
                                anyhow::bail!(
                                    "Token budget exceeded for commit {} but no file-level diffs available for split dispatch",
                                    &commit.hash[..8]
                                );
                            }
                            let report = self
                                .check_commit_split(
                                    commit,
                                    &single_view,
                                    &system_prompt,
                                    valid_scopes,
                                    include_suggestions,
                                    exceeded.available_input_tokens,
                                )
                                .await?;
                            all_results.extend(report.commits);
                        }
                    }
                }
                Ok(crate::data::check::CheckReport::new(all_results))
            }
        }
    }

    /// Parses the check response from AI.
    fn parse_check_response(
        &self,
        content: &str,
        repo_view: &RepositoryView,
    ) -> Result<crate::data::check::CheckReport> {
        use crate::data::check::{
            AiCheckResponse, CheckReport, CommitCheckResult as CheckResultType,
        };

        // Extract YAML from potential markdown wrapper
        let yaml_content = self.extract_yaml_from_check_response(content);

        // Parse YAML response
        let ai_response: AiCheckResponse = crate::data::from_yaml(&yaml_content).map_err(|e| {
            debug!(
                error = %e,
                content_length = content.len(),
                yaml_length = yaml_content.len(),
                "Check YAML parsing failed"
            );
            debug!(content = %content, "Raw AI response");
            debug!(yaml = %yaml_content, "Extracted YAML content");
            ClaudeError::AmendmentParsingFailed(format!("Check response parsing error: {e}"))
        })?;

        // Create a map of commit hashes to original messages for lookup
        let commit_messages: std::collections::HashMap<&str, &str> = repo_view
            .commits
            .iter()
            .map(|c| (c.hash.as_str(), c.original_message.as_str()))
            .collect();

        // Convert AI response to CheckReport
        let results: Vec<CheckResultType> = ai_response
            .checks
            .into_iter()
            .map(|check| {
                let mut result: CheckResultType = check.into();
                // Fill in the original message from repo_view
                if let Some(msg) = commit_messages.get(result.hash.as_str()) {
                    result.message = msg.lines().next().unwrap_or("").to_string();
                } else {
                    // Try to find by prefix
                    for (hash, msg) in &commit_messages {
                        if hash.starts_with(&result.hash) || result.hash.starts_with(*hash) {
                            result.message = msg.lines().next().unwrap_or("").to_string();
                            break;
                        }
                    }
                }
                result
            })
            .collect();

        Ok(CheckReport::new(results))
    }

    /// Extracts YAML content from check response, handling markdown wrappers.
    fn extract_yaml_from_check_response(&self, content: &str) -> String {
        let content = content.trim();

        // If content already starts with "checks:", it's pure YAML - return as-is
        if content.starts_with("checks:") {
            return content.to_string();
        }

        // Try to extract from ```yaml blocks first
        if let Some(yaml_start) = content.find("```yaml") {
            if let Some(yaml_content) = content[yaml_start + 7..].split("```").next() {
                return yaml_content.trim().to_string();
            }
        }

        // Try to extract from generic ``` blocks
        if let Some(code_start) = content.find("```") {
            if let Some(code_content) = content[code_start + 3..].split("```").next() {
                let potential_yaml = code_content.trim();
                // Check if it looks like YAML (starts with expected structure)
                if potential_yaml.starts_with("checks:") {
                    return potential_yaml.to_string();
                }
            }
        }

        // If no markdown blocks found or extraction failed, return trimmed content
        content.to_string()
    }

    /// Refines individually-generated amendments for cross-commit coherence.
    ///
    /// Sends commit summaries and proposed messages to the AI for a second pass
    /// that normalizes scopes, detects rename chains, and removes redundancy.
    pub async fn refine_amendments_coherence(
        &self,
        items: &[(crate::data::amendments::Amendment, String)],
    ) -> Result<AmendmentFile> {
        let system_prompt =
            self.adjusted_system_prompt(prompts::AMENDMENT_COHERENCE_SYSTEM_PROMPT.to_string());
        let user_prompt = prompts::generate_amendment_coherence_user_prompt(items);

        self.validate_prompt_budget(&system_prompt, &user_prompt)?;

        let content = self
            .send_with_optional_schema(
                &system_prompt,
                &user_prompt,
                self.schema_if_supported(response_schema::amendment_file_schema()),
            )
            .await?;

        self.parse_amendment_response(&content)
    }

    /// Refines individually-generated check results for cross-commit coherence.
    ///
    /// Sends commit summaries and check outcomes to the AI for a second pass
    /// that ensures consistent severity, detects cross-commit issues, and
    /// normalizes scope validation.
    pub async fn refine_checks_coherence(
        &self,
        items: &[(crate::data::check::CommitCheckResult, String)],
        repo_view: &RepositoryView,
    ) -> Result<crate::data::check::CheckReport> {
        let system_prompt =
            self.adjusted_system_prompt(prompts::CHECK_COHERENCE_SYSTEM_PROMPT.to_string());
        let user_prompt = prompts::generate_check_coherence_user_prompt(items);

        self.validate_prompt_budget(&system_prompt, &user_prompt)?;

        let content = self
            .send_with_optional_schema(
                &system_prompt,
                &user_prompt,
                self.schema_if_supported(response_schema::check_response_schema()),
            )
            .await?;

        self.parse_check_response(&content, repo_view)
    }

    /// Extracts YAML content from Claude response, handling markdown wrappers.
    fn extract_yaml_from_response(&self, content: &str) -> String {
        let content = content.trim();

        // If content already starts with "amendments:", it's pure YAML - return as-is
        if content.starts_with("amendments:") {
            return content.to_string();
        }

        // Try to extract from ```yaml blocks first
        if let Some(yaml_start) = content.find("```yaml") {
            if let Some(yaml_content) = content[yaml_start + 7..].split("```").next() {
                return yaml_content.trim().to_string();
            }
        }

        // Try to extract from generic ``` blocks
        if let Some(code_start) = content.find("```") {
            if let Some(code_content) = content[code_start + 3..].split("```").next() {
                let potential_yaml = code_content.trim();
                // Check if it looks like YAML (starts with expected structure)
                if potential_yaml.starts_with("amendments:") {
                    return potential_yaml.to_string();
                }
            }
        }

        // If no markdown blocks found or extraction failed, return trimmed content
        content.to_string()
    }
}

/// Validates a beta header against the model registry.
fn validate_beta_header(model: &str, beta_header: &Option<(String, String)>) -> Result<()> {
    if let Some((ref key, ref value)) = beta_header {
        let registry = crate::claude::model_config::get_model_registry();
        let supported = registry.get_beta_headers(model);
        if !supported
            .iter()
            .any(|bh| bh.key == *key && bh.value == *value)
        {
            let available: Vec<String> = supported
                .iter()
                .map(|bh| format!("{}:{}", bh.key, bh.value))
                .collect();
            if available.is_empty() {
                anyhow::bail!("Model '{model}' does not support any beta headers");
            }
            anyhow::bail!(
                "Beta header '{key}:{value}' is not supported for model '{model}'. Supported: {}",
                available.join(", ")
            );
        }
    }
    Ok(())
}

/// Warns that `--beta-header` is ignored for a non-Anthropic backend.
///
/// Beta headers are an Anthropic-specific mechanism (`anthropic-beta`); the
/// OpenAI-compatible backends never transmit them. Rather than validate the
/// value against the Claude registry — which could hard-fail (`Model 'X' does
/// not support any beta headers`) for a header that would never be sent
/// (#1119) — these backends warn once and ignore it, matching the `claude-cli`
/// backend's warn-and-ignore behaviour so the flag has one consistent shape
/// across every non-Anthropic backend.
fn warn_beta_header_ignored(
    backend: crate::claude::backend::AiBackend,
    beta_header: Option<&(String, String)>,
) {
    if let Some((key, value)) = beta_header {
        warn!(
            "--beta-header '{key}:{value}' is ignored when OMNI_DEV_AI_BACKEND={} \
             (beta headers are Anthropic-specific and are not sent to this backend)",
            backend.env_value()
        );
    }
}

/// Creates a default Claude client using environment variables and settings.
///
/// Async because the Ollama branch probes the local server for its
/// loaded context length so token-budget checks reflect what the server
/// actually loaded the model with (registry values are an estimate that
/// can exceed the live limit). All other branches finish synchronously.
pub async fn create_default_claude_client(
    model: Option<String>,
    beta_header: Option<(String, String)>,
) -> Result<ClaudeClient> {
    create_default_claude_client_with(
        &crate::utils::settings::SettingsEnv::load(),
        model,
        beta_header,
    )
    .await
}

/// [`create_default_claude_client`] over an injected
/// [`EnvSource`](crate::utils::env::EnvSource).
///
/// This is the backend-dispatch boundary. Backend, model, and beta-header
/// selection delegate to the shared resolvers in [`crate::claude::backend`]
/// (also used by preflight); this function reads only the per-backend
/// credential/endpoint vars and builds the client. The production wrapper
/// passes `&SettingsEnv::load()`; tests pass a pure `MapEnv`, so dispatch is
/// tested without mutating the process environment or taking a lock
/// (issue #1030). The `Sync` bound keeps the returned future `Send` (the
/// reference is held across the Ollama context-length probe `.await`).
pub(crate) async fn create_default_claude_client_with(
    env: &(impl crate::utils::env::EnvSource + Sync),
    model: Option<String>,
    beta_header: Option<(String, String)>,
) -> Result<ClaudeClient> {
    use crate::claude::ai::claude_cli::ClaudeCliAiClient;
    use crate::claude::ai::openai::OpenAiAiClient;
    use crate::claude::backend::{self, AiBackend};

    let ai_backend = backend::resolve_backend(env)?;
    let beta_header = backend::resolve_beta_header(beta_header, env)?;
    let registry = crate::claude::model_config::get_model_registry();
    let model = backend::resolve_model(ai_backend, model.as_deref(), env, registry);
    // Resolved once, applied once below: the off-switch is a property of the
    // endpoint, not of any one backend, so every arm honours it (#1561).
    let schema_disabled = backend::resolve_structured_output_disabled(env);
    debug!(backend = ?ai_backend, model = %model, schema_disabled, "Resolved AI backend");

    let ai_client: Box<dyn AiClient> = match ai_backend {
        AiBackend::ClaudeCli => {
            // The `claude -p` subprocess negotiates betas itself, so the
            // beta header is deliberately not forwarded (and, uniquely, not
            // validated — this backend accepts short model aliases like
            // sonnet/opus/haiku that the registry does not know).
            if beta_header.is_some() {
                warn!(
                    "--beta-header is ignored when OMNI_DEV_AI_BACKEND=claude-cli \
                     (the CLI's --betas flag has different semantics and is not forwarded)"
                );
            }
            debug!(model = %model, "Creating claude -p subprocess client");
            Box::new(ClaudeCliAiClient::new(model))
        }
        AiBackend::Ollama => {
            warn_beta_header_ignored(AiBackend::Ollama, beta_header.as_ref());
            let base_url = env.var("OLLAMA_BASE_URL");
            let mut ai_client = OpenAiAiClient::new_ollama(model, base_url, None)?;
            match ai_client.probe_loaded_context_length().await {
                Some(source) => {
                    info!(
                        loaded_context_length = ai_client.loaded_context_length(),
                        source = source.as_str(),
                        model = %ai_client.get_metadata().model,
                        "Probed loaded context length from local server"
                    );
                }
                None => {
                    debug!(
                        "Loaded context length probe did not return a value; \
                         falling back to registry/default for token budget"
                    );
                }
            }
            Box::new(ai_client)
        }
        AiBackend::OpenAi => {
            debug!("Creating OpenAI client");
            warn_beta_header_ignored(AiBackend::OpenAi, beta_header.as_ref());

            let api_key = secret_var_any(env, &["OPENAI_API_KEY", "OPENAI_AUTH_TOKEN"])?
                .ok_or_else(|| {
                    debug!("Failed to get OpenAI API key");
                    ClaudeError::ApiKeyNotFound
                })?;
            debug!("OpenAI API key found");

            let ai_client =
                OpenAiAiClient::new_openai(model, api_key.expose_secret().to_string(), None)?;
            debug!("OpenAI client created successfully");
            Box::new(ai_client)
        }
        AiBackend::Bedrock => {
            validate_beta_header(&model, &beta_header)?;
            let auth_token = secret_var(env, "ANTHROPIC_AUTH_TOKEN")?
                .ok_or(ClaudeError::ApiKeyNotFound)?
                .expose_secret()
                .to_string();

            let base_url = env
                .var("ANTHROPIC_BEDROCK_BASE_URL")
                .ok_or(ClaudeError::ApiKeyNotFound)?;

            Box::new(BedrockAiClient::new(
                model,
                auth_token,
                base_url,
                beta_header,
            )?)
        }
        AiBackend::Default => {
            debug!("Creating direct Claude API client");
            validate_beta_header(&model, &beta_header)?;
            let api_key = secret_var_any(
                env,
                &[
                    "CLAUDE_API_KEY",
                    "ANTHROPIC_API_KEY",
                    "ANTHROPIC_AUTH_TOKEN",
                ],
            )?
            .ok_or(ClaudeError::ApiKeyNotFound)?;

            let ai_client =
                ClaudeAiClient::new(model, api_key.expose_secret().to_string(), beta_header)?;
            debug!("Claude client created successfully");
            Box::new(ai_client)
        }
    };

    Ok(ClaudeClient::new(ai_client).with_structured_output_disabled(schema_disabled))
}

#[cfg(test)]
#[allow(
    clippy::unwrap_used,
    clippy::expect_used,
    clippy::format_in_format_args
)]
mod tests {
    use super::*;
    use crate::claude::ai::{AiClient, AiClientCapabilities, AiClientMetadata};
    use std::future::Future;
    use std::pin::Pin;
    use std::sync::{Arc, Mutex};

    /// Mock AI client for testing — never makes real HTTP requests.
    struct MockAiClient;

    impl AiClient for MockAiClient {
        fn send_request<'a>(
            &'a self,
            _system_prompt: &'a str,
            _user_prompt: &'a str,
        ) -> Pin<Box<dyn Future<Output = Result<String>> + Send + 'a>> {
            Box::pin(async { Ok(String::new()) })
        }

        fn get_metadata(&self) -> AiClientMetadata {
            AiClientMetadata {
                provider: "Mock".to_string(),
                model: "mock-model".to_string(),
                max_context_length: 200_000,
                max_response_length: 8_192,
                active_beta: None,
            }
        }
    }

    fn make_client() -> ClaudeClient {
        ClaudeClient::new(Box::new(MockAiClient))
    }

    /// Mock AI client that records both prompts and per-call options
    /// (the schema attached, if any). Used to verify
    /// [`ClaudeClient::send_with_optional_schema`] dispatches via the
    /// options-aware method when a schema is provided and via the plain
    /// method otherwise.
    ///
    /// Returns the configured `response` string from both `send_request`
    /// and `send_request_with_options` so tests that need a parseable
    /// response (e.g. the refine_* coherence paths) can plug in canned
    /// YAML/JSON.
    struct SchemaRecordingMockAiClient {
        capabilities: AiClientCapabilities,
        response: String,
        recorded_options: Arc<Mutex<Vec<RequestOptions>>>,
        recorded_plain: Arc<Mutex<Vec<(String, String)>>>,
        /// When set, every `send_request_with_options` call fails with this
        /// error — how the #1561 tests stand in for an endpoint that rejects
        /// `output_config` while still answering plain requests.
        options_error: Option<ClaudeError>,
        /// When set, every `send_request` call fails with this error, for the
        /// case where the YAML fallback fails too.
        plain_error: Option<ClaudeError>,
    }
    impl SchemaRecordingMockAiClient {
        fn new(supports_response_schema: bool) -> Self {
            Self::with_response(supports_response_schema, String::new())
        }

        fn with_response(supports_response_schema: bool, response: String) -> Self {
            Self {
                capabilities: AiClientCapabilities {
                    supports_response_schema,
                },
                response,
                recorded_options: Arc::new(Mutex::new(Vec::new())),
                recorded_plain: Arc::new(Mutex::new(Vec::new())),
                options_error: None,
                plain_error: None,
            }
        }

        fn failing_options(mut self, error: ClaudeError) -> Self {
            self.options_error = Some(error);
            self
        }

        fn failing_plain(mut self, error: ClaudeError) -> Self {
            self.plain_error = Some(error);
            self
        }
    }

    impl AiClient for SchemaRecordingMockAiClient {
        fn send_request<'a>(
            &'a self,
            system_prompt: &'a str,
            user_prompt: &'a str,
        ) -> Pin<Box<dyn Future<Output = Result<String>> + Send + 'a>> {
            let plain = self.recorded_plain.clone();
            let sys = system_prompt.to_string();
            let usr = user_prompt.to_string();
            let response = self.response.clone();
            let error = self
                .plain_error
                .as_ref()
                .map(|e| ClaudeError::ApiRequestFailed(e.to_string()));
            Box::pin(async move {
                plain.lock().unwrap().push((sys, usr));
                match error {
                    Some(e) => Err(e.into()),
                    None => Ok(response),
                }
            })
        }

        fn capabilities(&self) -> AiClientCapabilities {
            self.capabilities
        }

        fn send_request_with_options<'a>(
            &'a self,
            _system_prompt: &'a str,
            _user_prompt: &'a str,
            options: RequestOptions,
        ) -> Pin<Box<dyn Future<Output = Result<String>> + Send + 'a>> {
            let recorded = self.recorded_options.clone();
            let response = self.response.clone();
            let error = self.options_error.as_ref().map(|e| match e {
                ClaudeError::ApiHttpError { status, body } => ClaudeError::ApiHttpError {
                    status: *status,
                    body: body.clone(),
                },
                other => ClaudeError::ApiRequestFailed(other.to_string()),
            });
            Box::pin(async move {
                recorded.lock().unwrap().push(options);
                match error {
                    Some(e) => Err(e.into()),
                    None => Ok(response),
                }
            })
        }

        fn get_metadata(&self) -> AiClientMetadata {
            AiClientMetadata {
                provider: "SchemaMock".to_string(),
                model: "schema-mock".to_string(),
                max_context_length: 200_000,
                max_response_length: 8_192,
                active_beta: None,
            }
        }
    }

    // ── ClaudeClient schema-routing helpers ───────────────────────────

    /// Backends that don't advertise schema support take the
    /// `send_request` branch in `send_with_optional_schema` regardless
    /// of whether a schema was supplied at the call site.
    #[tokio::test]
    async fn send_with_optional_schema_without_caps_uses_plain_send() {
        let inner = SchemaRecordingMockAiClient::new(false);
        let plain_log = inner.recorded_plain.clone();
        let opts_log = inner.recorded_options.clone();
        let client = ClaudeClient::new(Box::new(inner));

        let schema = serde_json::json!({"type": "object"});
        client
            .send_with_optional_schema(
                "sys",
                "usr",
                client.schema_if_supported(&schema), // → None
            )
            .await
            .unwrap();

        assert_eq!(plain_log.lock().unwrap().len(), 1);
        assert!(opts_log.lock().unwrap().is_empty());
    }

    /// Backends that advertise schema support take the
    /// `send_request_with_options` branch and receive the schema in the
    /// options struct.
    #[tokio::test]
    async fn send_with_optional_schema_with_caps_uses_options_send() {
        let inner = SchemaRecordingMockAiClient::new(true);
        let plain_log = inner.recorded_plain.clone();
        let opts_log = inner.recorded_options.clone();
        let client = ClaudeClient::new(Box::new(inner));

        let schema = serde_json::json!({"type": "object", "additionalProperties": false});
        client
            .send_with_optional_schema(
                "sys",
                "usr",
                client.schema_if_supported(&schema), // → Some
            )
            .await
            .unwrap();

        let recorded = opts_log.lock().unwrap();
        assert_eq!(recorded.len(), 1);
        assert_eq!(recorded[0].response_schema.as_ref(), Some(&schema));
        assert!(plain_log.lock().unwrap().is_empty());
    }

    // ── #1561: endpoint rejects `output_config` ──────────────────────

    /// The rejection an `output_config`-stripping gateway returns, verbatim
    /// from issue #1561.
    fn output_config_rejection() -> ClaudeError {
        ClaudeError::ApiHttpError {
            status: 400,
            body: String::from(
                r#"{"message":"output_config.format: Extra inputs are not permitted"}"#,
            ),
        }
    }

    fn schema_fixture() -> serde_json::Value {
        serde_json::json!({"type": "object", "additionalProperties": false})
    }

    /// A rejected `output_config` retries once on the plain path, and the
    /// retry's system prompt has the JSON-schema override stripped — the model
    /// must not be told to satisfy a schema the request no longer carries.
    #[tokio::test]
    async fn output_config_rejection_falls_back_to_yaml_path() {
        let inner = SchemaRecordingMockAiClient::with_response(true, "answer: ok".to_string())
            .failing_options(output_config_rejection());
        let plain_log = inner.recorded_plain.clone();
        let opts_log = inner.recorded_options.clone();
        let client = ClaudeClient::new(Box::new(inner));

        let schema = schema_fixture();
        let system_prompt = client.adjusted_system_prompt("base system prompt".to_string());
        assert!(system_prompt.ends_with(prompts::JSON_SCHEMA_RESPONSE_OVERRIDE));

        let out = client
            .send_with_optional_schema(&system_prompt, "usr", client.schema_if_supported(&schema))
            .await
            .expect("the YAML fallback should carry the call");

        assert_eq!(out, "answer: ok");
        assert_eq!(opts_log.lock().unwrap().len(), 1);
        let plain = plain_log.lock().unwrap();
        assert_eq!(plain.len(), 1);
        assert_eq!(plain[0].0, "base system prompt");
        assert!(
            !plain[0].0.contains("STRUCTURED OUTPUT OVERRIDE"),
            "fallback prompt still carries the JSON override: {}",
            plain[0].0
        );
    }

    /// The rejection latches structured output off, so a second call skips the
    /// schema path entirely rather than paying the rejected round-trip again.
    #[tokio::test]
    async fn output_config_rejection_latches_off_for_later_calls() {
        let inner = SchemaRecordingMockAiClient::with_response(true, "answer: ok".to_string())
            .failing_options(output_config_rejection());
        let plain_log = inner.recorded_plain.clone();
        let opts_log = inner.recorded_options.clone();
        let client = ClaudeClient::new(Box::new(inner));

        let schema = schema_fixture();
        for _ in 0..3 {
            client
                .send_with_optional_schema("sys", "usr", client.schema_if_supported(&schema))
                .await
                .expect("every call should succeed on the YAML path");
        }

        assert_eq!(
            opts_log.lock().unwrap().len(),
            1,
            "the rejected shape should be attempted at most once per client"
        );
        assert_eq!(plain_log.lock().unwrap().len(), 3);
        assert!(client.schema_if_supported(&schema).is_none());
    }

    /// Any other permanent failure still propagates — the fallback must not
    /// widen into a blanket retry of 4xx responses (#1333).
    #[tokio::test]
    async fn unrelated_client_error_does_not_fall_back() {
        let inner =
            SchemaRecordingMockAiClient::new(true).failing_options(ClaudeError::ApiHttpError {
                status: 400,
                body: String::from("max_tokens: must be positive"),
            });
        let plain_log = inner.recorded_plain.clone();
        let client = ClaudeClient::new(Box::new(inner));

        let schema = schema_fixture();
        let err = client
            .send_with_optional_schema("sys", "usr", client.schema_if_supported(&schema))
            .await
            .expect_err("an unrelated 400 must not be papered over");

        assert!(err.to_string().contains("max_tokens"), "{err}");
        assert!(plain_log.lock().unwrap().is_empty());
        assert!(client.schema_if_supported(&schema).is_some());
    }

    /// When the fallback fails too, both failures survive in the error chain.
    #[tokio::test]
    async fn failing_fallback_reports_both_errors() {
        let inner = SchemaRecordingMockAiClient::new(true)
            .failing_options(output_config_rejection())
            .failing_plain(ClaudeError::NetworkError(String::from("connection reset")));
        let client = ClaudeClient::new(Box::new(inner));

        let schema = schema_fixture();
        let err = client
            .send_with_optional_schema("sys", "usr", client.schema_if_supported(&schema))
            .await
            .expect_err("a failing fallback must surface");

        let chain = format!("{err:#}");
        assert!(chain.contains("output_config"), "{chain}");
        assert!(chain.contains("connection reset"), "{chain}");
    }

    /// The strip is exact and idempotent: it removes the override when present
    /// and leaves an unadorned YAML prompt untouched.
    #[test]
    fn yaml_system_prompt_strips_only_the_json_override() {
        let base = "base system prompt";
        let adjusted = prompts::apply_response_format_to_system_prompt(
            base.to_string(),
            ResponseFormat::JsonSchema,
        );
        assert_eq!(ClaudeClient::yaml_system_prompt(&adjusted), base);
        assert_eq!(ClaudeClient::yaml_system_prompt(base), base);
        assert_eq!(
            ClaudeClient::yaml_system_prompt(ClaudeClient::yaml_system_prompt(&adjusted)),
            base
        );
    }

    /// End-to-end reproduction of issue #1561 against a real
    /// [`BedrockAiClient`]: a gateway that answers any body carrying
    /// `output_config` with the reported `400` and passes everything else
    /// through. The call must succeed, having re-issued exactly one
    /// `output_config`-free request.
    #[tokio::test]
    async fn bedrock_gateway_rejecting_output_config_still_completes() {
        use wiremock::matchers::method;
        use wiremock::{Mock, MockServer, Request, ResponseTemplate};

        fn body_has_output_config(request: &Request) -> bool {
            serde_json::from_slice::<serde_json::Value>(&request.body)
                .is_ok_and(|body| body.get("output_config").is_some())
        }

        let server = MockServer::start().await;
        Mock::given(method("POST"))
            .and(body_has_output_config)
            .respond_with(ResponseTemplate::new(400).set_body_string(
                r#"{"message":"output_config.format: Extra inputs are not permitted"}"#,
            ))
            .expect(1)
            .mount(&server)
            .await;
        Mock::given(method("POST"))
            .and(|request: &Request| !body_has_output_config(request))
            .respond_with(ResponseTemplate::new(200).set_body_string(
                r#"{"id":"m","type":"message","role":"assistant","model":"m",
                    "content":[{"type":"text","text":"answer: ok"}],"stop_reason":"end_turn"}"#,
            ))
            .expect(1)
            .mount(&server)
            .await;

        let ai_client = BedrockAiClient::new(
            "claude-sonnet-4-6".to_string(),
            "token".to_string(),
            server.uri(),
            None,
        )
        .expect("bedrock client should build");
        let client = ClaudeClient::new(Box::new(ai_client));

        let schema = schema_fixture();
        let system_prompt = client.adjusted_system_prompt("base system prompt".to_string());
        let out = client
            .send_with_optional_schema(&system_prompt, "usr", client.schema_if_supported(&schema))
            .await
            .expect("the gateway rejection should be survivable");

        assert_eq!(out, "answer: ok");

        let received = server.received_requests().await.unwrap();
        assert_eq!(received.len(), 2, "expected one rejected then one retried");
        let retried: serde_json::Value = serde_json::from_slice(&received[1].body).unwrap();
        assert!(retried.get("output_config").is_none());
        assert_eq!(retried["system"], "base system prompt");
    }

    /// The env off-switch reaches the client, so an endpoint known to reject
    /// the field never gets sent it — not even once.
    #[tokio::test]
    async fn structured_output_disable_env_switches_the_schema_path_off() {
        let base = MapEnv::new()
            .with("CLAUDE_CODE_USE_BEDROCK", "true")
            .with("ANTHROPIC_AUTH_TOKEN", "token")
            .with("ANTHROPIC_BEDROCK_BASE_URL", "https://example.com")
            .with("OMNI_DEV_MODEL", "claude-sonnet-4-6");

        let schema = schema_fixture();

        let enabled = create_default_claude_client_with(&base, None, None)
            .await
            .expect("bedrock client should build");
        assert!(
            enabled.schema_if_supported(&schema).is_some(),
            "a flagged model should use the schema path by default"
        );

        let disabled = create_default_claude_client_with(
            &base.clone().with(
                crate::claude::backend::STRUCTURED_OUTPUT_DISABLE_ENV,
                "true",
            ),
            None,
            None,
        )
        .await
        .expect("bedrock client should build");
        assert!(disabled.schema_if_supported(&schema).is_none());
    }

    /// `adjusted_system_prompt` only appends the JSON-schema override
    /// suffix when the active backend supports schema enforcement.
    #[test]
    fn adjusted_system_prompt_adds_suffix_when_supported() {
        let client = ClaudeClient::new(Box::new(SchemaRecordingMockAiClient::new(true)));
        let result = client.adjusted_system_prompt("body".to_string());
        assert!(result.starts_with("body"));
        assert!(result.contains("STRUCTURED OUTPUT OVERRIDE"));
    }

    #[test]
    fn adjusted_system_prompt_passes_through_when_not_supported() {
        let client = ClaudeClient::new(Box::new(SchemaRecordingMockAiClient::new(false)));
        let result = client.adjusted_system_prompt("body".to_string());
        assert_eq!(result, "body");
    }

    #[test]
    fn schema_if_supported_returns_some_when_supported() {
        let client = ClaudeClient::new(Box::new(SchemaRecordingMockAiClient::new(true)));
        let schema = serde_json::json!({"type": "object"});
        let returned = client.schema_if_supported(&schema);
        assert!(returned.is_some());
        assert!(std::ptr::eq(
            std::ptr::from_ref(returned.unwrap()),
            std::ptr::addr_of!(schema)
        ));
    }

    #[test]
    fn schema_if_supported_returns_none_when_not_supported() {
        let client = ClaudeClient::new(Box::new(SchemaRecordingMockAiClient::new(false)));
        let schema = serde_json::json!({"type": "object"});
        assert!(client.schema_if_supported(&schema).is_none());
    }

    // ── refine_amendments_coherence / refine_checks_coherence ────────

    /// Exercises the full body of `refine_amendments_coherence`:
    /// adjusted_system_prompt → validate_prompt_budget → schema-aware
    /// dispatch → parse_amendment_response. Uses a schema-supporting
    /// mock so the schema attachment branch is taken too.
    #[tokio::test]
    async fn refine_amendments_coherence_round_trip() {
        let mock = SchemaRecordingMockAiClient::with_response(
            true, // supports_response_schema
            "amendments: []".to_string(),
        );
        let recorded_opts = mock.recorded_options.clone();
        let client = ClaudeClient::new(Box::new(mock));

        let amendment = crate::data::amendments::Amendment {
            commit: "abc123".to_string(),
            message: "feat: do thing".to_string(),
            summary: "did the thing".to_string(),
        };
        let items = vec![(amendment, "summary text".to_string())];

        let result = client
            .refine_amendments_coherence(&items)
            .await
            .expect("coherence refinement should succeed");
        assert!(result.amendments.is_empty());

        // Verify the schema-aware dispatch path was taken and that the
        // attached schema is the AmendmentFile schema.
        let recorded = recorded_opts.lock().unwrap();
        assert_eq!(recorded.len(), 1);
        let attached = recorded[0]
            .response_schema
            .as_ref()
            .expect("schema must be attached when capability is true");
        assert_eq!(
            attached,
            response_schema::amendment_file_schema(),
            "refine_amendments_coherence should attach the AmendmentFile schema"
        );
    }

    /// Same coverage shape as the amendment variant, but for the check
    /// coherence path. Uses `parse_check_response` which needs a
    /// repository view to map commit hashes back to messages — we
    /// supply an empty view.
    #[tokio::test]
    async fn refine_checks_coherence_round_trip() {
        let mock = SchemaRecordingMockAiClient::with_response(
            true, // supports_response_schema
            "checks: []".to_string(),
        );
        let recorded_opts = mock.recorded_options.clone();
        let client = ClaudeClient::new(Box::new(mock));

        let check = crate::data::check::CommitCheckResult {
            hash: "abc123".to_string(),
            message: "feat: do thing".to_string(),
            issues: Vec::new(),
            suggestion: None,
            passes: true,
            summary: Some("summary".to_string()),
        };
        let items = vec![(check, "summary text".to_string())];
        let dir = tempfile::TempDir::new().unwrap();
        let repo_view = make_test_repo_view(&dir);

        let result = client
            .refine_checks_coherence(&items, &repo_view)
            .await
            .expect("coherence refinement should succeed");
        assert_eq!(result.summary.total_commits, 0);

        let recorded = recorded_opts.lock().unwrap();
        assert_eq!(recorded.len(), 1);
        let attached = recorded[0]
            .response_schema
            .as_ref()
            .expect("schema must be attached when capability is true");
        assert_eq!(
            attached,
            response_schema::check_response_schema(),
            "refine_checks_coherence should attach the AiCheckResponse schema"
        );
    }

    /// Verifies the no-schema branch of refine_amendments_coherence —
    /// when the backend doesn't advertise schema support, dispatch
    /// falls through to plain `send_request` and no schema is attached.
    #[tokio::test]
    async fn refine_amendments_coherence_without_schema_capability() {
        let mock = SchemaRecordingMockAiClient::with_response(
            false, // supports_response_schema
            "amendments: []".to_string(),
        );
        let recorded_plain = mock.recorded_plain.clone();
        let recorded_opts = mock.recorded_options.clone();
        let client = ClaudeClient::new(Box::new(mock));

        let amendment = crate::data::amendments::Amendment {
            commit: "abc123".to_string(),
            message: "feat: do thing".to_string(),
            summary: String::new(),
        };
        let items = vec![(amendment, "summary".to_string())];

        client
            .refine_amendments_coherence(&items)
            .await
            .expect("coherence refinement should succeed without schema support");

        assert_eq!(recorded_plain.lock().unwrap().len(), 1);
        assert!(
            recorded_opts.lock().unwrap().is_empty(),
            "no-schema backend must not be reached via the options path"
        );
    }

    // ── extract_yaml_from_response ─────────────────────────────────

    #[test]
    fn extract_yaml_pure_amendments() {
        let client = make_client();
        let content = "amendments:\n  - commit: abc123\n    message: test";
        let result = client.extract_yaml_from_response(content);
        assert!(result.starts_with("amendments:"));
    }

    #[test]
    fn extract_yaml_with_markdown_yaml_block() {
        let client = make_client();
        let content = "Here is the result:\n```yaml\namendments:\n  - commit: abc\n```\n";
        let result = client.extract_yaml_from_response(content);
        assert!(result.starts_with("amendments:"));
    }

    #[test]
    fn extract_yaml_with_generic_code_block() {
        let client = make_client();
        let content = "```\namendments:\n  - commit: abc\n```";
        let result = client.extract_yaml_from_response(content);
        assert!(result.starts_with("amendments:"));
    }

    #[test]
    fn extract_yaml_with_whitespace() {
        let client = make_client();
        let content = "  \n  amendments:\n  - commit: abc\n  ";
        let result = client.extract_yaml_from_response(content);
        assert!(result.starts_with("amendments:"));
    }

    #[test]
    fn extract_yaml_fallback_returns_trimmed() {
        let client = make_client();
        let content = "  some random text  ";
        let result = client.extract_yaml_from_response(content);
        assert_eq!(result, "some random text");
    }

    // ── extract_yaml_from_check_response ───────────────────────────

    #[test]
    fn extract_check_yaml_pure() {
        let client = make_client();
        let content = "checks:\n  - commit: abc123";
        let result = client.extract_yaml_from_check_response(content);
        assert!(result.starts_with("checks:"));
    }

    #[test]
    fn extract_check_yaml_markdown_block() {
        let client = make_client();
        let content = "```yaml\nchecks:\n  - commit: abc\n```";
        let result = client.extract_yaml_from_check_response(content);
        assert!(result.starts_with("checks:"));
    }

    #[test]
    fn extract_check_yaml_generic_block() {
        let client = make_client();
        let content = "```\nchecks:\n  - commit: abc\n```";
        let result = client.extract_yaml_from_check_response(content);
        assert!(result.starts_with("checks:"));
    }

    #[test]
    fn extract_check_yaml_fallback() {
        let client = make_client();
        let content = "  unexpected content  ";
        let result = client.extract_yaml_from_check_response(content);
        assert_eq!(result, "unexpected content");
    }

    // ── parse_amendment_response ────────────────────────────────────

    #[test]
    fn parse_amendment_response_valid() {
        let client = make_client();
        let yaml = format!(
            "amendments:\n  - commit: \"{}\"\n    message: \"test message\"",
            "a".repeat(40)
        );
        let result = client.parse_amendment_response(&yaml);
        assert!(result.is_ok());
        assert_eq!(result.unwrap().amendments.len(), 1);
    }

    #[test]
    fn parse_amendment_response_invalid_yaml() {
        let client = make_client();
        let result = client.parse_amendment_response("not: valid: yaml: [{{");
        assert!(result.is_err());
    }

    #[test]
    fn parse_amendment_response_invalid_hash() {
        let client = make_client();
        let yaml = "amendments:\n  - commit: \"short\"\n    message: \"test\"";
        let result = client.parse_amendment_response(yaml);
        assert!(result.is_err());
    }

    // ── validate_beta_header ───────────────────────────────────────

    #[test]
    fn validate_beta_header_none_passes() {
        let result = validate_beta_header("claude-opus-4-1-20250805", &None);
        assert!(result.is_ok());
    }

    #[test]
    fn validate_beta_header_unsupported_fails() {
        let header = Some(("fake-key".to_string(), "fake-value".to_string()));
        let result = validate_beta_header("claude-opus-4-1-20250805", &header);
        assert!(result.is_err());
    }

    /// The 4.6 entries expose 1M context natively, which makes their
    /// `context-1m-2025-08-07` entry a no-op for limits. It is retained anyway
    /// because this validator hard-errors on any header absent from the
    /// registry — dropping it would newly reject a flag these callers can pass
    /// today (#1334).
    #[test]
    fn validate_beta_header_context_1m_still_accepted_on_4_6() {
        let header = Some((
            "anthropic-beta".to_string(),
            "context-1m-2025-08-07".to_string(),
        ));

        assert!(
            validate_beta_header("claude-opus-4-6", &header).is_ok(),
            "context-1m must stay accepted on Opus 4.6 even though 1M is now native"
        );
        assert!(
            validate_beta_header("claude-sonnet-4-6", &header).is_ok(),
            "context-1m must stay accepted on Sonnet 4.6 even though 1M is now native"
        );
    }

    /// The current generation never used the context beta header, so passing it
    /// is a genuine mistake and must surface rather than be silently ignored.
    #[test]
    fn validate_beta_header_context_1m_rejected_on_current_generation() {
        let header = Some((
            "anthropic-beta".to_string(),
            "context-1m-2025-08-07".to_string(),
        ));

        for model in ["claude-sonnet-5", "claude-opus-4-8", "claude-fable-5"] {
            let err = validate_beta_header(model, &header)
                .expect_err("current-generation models declare no beta headers");
            assert!(
                err.to_string()
                    .contains("does not support any beta headers"),
                "unexpected error for {model}: {err}"
            );
        }
    }

    // ── ClaudeClient::new / get_ai_client_metadata ─────────────────

    #[test]
    fn client_metadata() {
        let client = make_client();
        let metadata = client.get_ai_client_metadata();
        assert_eq!(metadata.provider, "Mock");
        assert_eq!(metadata.model, "mock-model");
    }

    // ── property tests ────────────────────────────────────────────

    mod prop {
        use super::*;
        use proptest::prelude::*;

        proptest! {
            #[test]
            fn yaml_response_output_trimmed(s in ".*") {
                let client = make_client();
                let result = client.extract_yaml_from_response(&s);
                prop_assert_eq!(&result, result.trim());
            }

            #[test]
            fn yaml_response_amendments_prefix_preserved(tail in ".*") {
                let client = make_client();
                let input = format!("amendments:{tail}");
                let result = client.extract_yaml_from_response(&input);
                prop_assert!(result.starts_with("amendments:"));
            }

            #[test]
            fn check_response_checks_prefix_preserved(tail in ".*") {
                let client = make_client();
                let input = format!("checks:{tail}");
                let result = client.extract_yaml_from_check_response(&input);
                prop_assert!(result.starts_with("checks:"));
            }

            #[test]
            fn yaml_fenced_block_strips_fences(
                content in "[a-zA-Z0-9: _\\-\n]{1,100}",
            ) {
                let client = make_client();
                let input = format!("```yaml\n{content}\n```");
                let result = client.extract_yaml_from_response(&input);
                prop_assert!(!result.contains("```"));
            }
        }
    }

    // ── ConfigurableMockAiClient tests ──────────────────────────────

    fn make_configurable_client(responses: Vec<Result<String>>) -> ClaudeClient {
        ClaudeClient::new(Box::new(
            crate::claude::test_utils::ConfigurableMockAiClient::new(responses),
        ))
    }

    fn make_test_repo_view(dir: &tempfile::TempDir) -> crate::data::RepositoryView {
        use crate::data::{AiInfo, FieldExplanation, WorkingDirectoryInfo};
        use crate::git::commit::FileChanges;
        use crate::git::{CommitAnalysis, CommitInfo};

        let diff_path = dir.path().join("0.diff");
        std::fs::write(&diff_path, "+added line\n").unwrap();

        crate::data::RepositoryView {
            versions: None,
            explanation: FieldExplanation::default(),
            working_directory: WorkingDirectoryInfo {
                clean: true,
                untracked_changes: Vec::new(),
            },
            remotes: Vec::new(),
            ai: AiInfo {
                scratch: String::new(),
            },
            branch_info: None,
            pr_template: None,
            pr_template_location: None,
            branch_prs: None,
            commits: vec![CommitInfo {
                hash: format!("{:0>40}", 0),
                author: "Test <test@test.com>".to_string(),
                date: chrono::Utc::now().fixed_offset(),
                original_message: "feat(test): add something".to_string(),
                in_main_branches: Vec::new(),
                analysis: CommitAnalysis {
                    detected_type: "feat".to_string(),
                    detected_scope: "test".to_string(),
                    proposed_message: "feat(test): add something".to_string(),
                    file_changes: FileChanges {
                        total_files: 1,
                        files_added: 1,
                        files_deleted: 0,
                        file_list: Vec::new(),
                    },
                    diff_summary: "file.rs | 1 +".to_string(),
                    diff_file: diff_path.to_string_lossy().to_string(),
                    file_diffs: Vec::new(),
                },
            }],
        }
    }

    fn valid_check_yaml() -> String {
        format!(
            "checks:\n  - commit: \"{hash}\"\n    passes: true\n    issues: []\n",
            hash = format!("{:0>40}", 0)
        )
    }

    #[tokio::test]
    async fn send_message_propagates_ai_error() {
        let client = make_configurable_client(vec![Err(anyhow::anyhow!("mock error"))]);
        let result = client.send_message("sys", "usr").await;
        assert!(result.is_err());
        assert!(result.unwrap_err().to_string().contains("mock error"));
    }

    #[tokio::test]
    async fn check_commits_succeeds_after_request_error() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_test_repo_view(&dir);
        // First attempt: request error; retries return valid response.
        let client = make_configurable_client(vec![
            Err(anyhow::anyhow!("rate limit")),
            Ok(valid_check_yaml()),
            Ok(valid_check_yaml()),
        ]);
        let result = client
            .check_commits_with_scopes(&repo_view, None, &[], false)
            .await;
        assert!(result.is_ok());
    }

    #[tokio::test]
    async fn check_commits_succeeds_after_parse_error() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_test_repo_view(&dir);
        // First attempt: AI returns malformed YAML; retry succeeds.
        let client = make_configurable_client(vec![
            Ok("not: valid: yaml: [[".to_string()),
            Ok(valid_check_yaml()),
            Ok(valid_check_yaml()),
        ]);
        let result = client
            .check_commits_with_scopes(&repo_view, None, &[], false)
            .await;
        assert!(result.is_ok());
    }

    #[tokio::test]
    async fn check_commits_fails_after_all_retries_exhausted() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_test_repo_view(&dir);
        let client = make_configurable_client(vec![
            Err(anyhow::anyhow!("first failure")),
            Err(anyhow::anyhow!("second failure")),
            Err(anyhow::anyhow!("final failure")),
        ]);
        let result = client
            .check_commits_with_scopes(&repo_view, None, &[], false)
            .await;
        assert!(result.is_err());
    }

    #[tokio::test]
    async fn check_commits_fails_when_all_parses_fail() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_test_repo_view(&dir);
        let client = make_configurable_client(vec![
            Ok("bad yaml [[".to_string()),
            Ok("bad yaml [[".to_string()),
            Ok("bad yaml [[".to_string()),
        ]);
        let result = client
            .check_commits_with_scopes(&repo_view, None, &[], false)
            .await;
        assert!(result.is_err());
    }

    // ── split dispatch tests ─────────────────────────────────────

    /// Creates a mock client with a constrained context window.
    ///
    /// The window is large enough that a single-file chunk fits, but too
    /// small for both files together (including system prompt overhead).
    fn make_small_context_client(responses: Vec<Result<String>>) -> ClaudeClient {
        // Context of 50k with more conservative token estimation (2.5 chars/token
        // vs 3.5) ensures per-file diffs fit in chunks without placeholders while
        // still being large enough to trigger split dispatch for multiple files.
        let mock = crate::claude::test_utils::ConfigurableMockAiClient::new(responses)
            .with_context_length(50_000);
        ClaudeClient::new(Box::new(mock))
    }

    /// Like [`make_small_context_client`] but also returns a handle to inspect
    /// how many mock responses remain unconsumed after the test runs.
    fn make_small_context_client_tracked(
        responses: Vec<Result<String>>,
    ) -> (ClaudeClient, crate::claude::test_utils::ResponseQueueHandle) {
        let mock = crate::claude::test_utils::ConfigurableMockAiClient::new(responses)
            .with_context_length(50_000);
        let handle = mock.response_handle();
        (ClaudeClient::new(Box::new(mock)), handle)
    }

    /// Creates a repo view with per-file diffs large enough to exceed the
    /// constrained context window, ensuring the split dispatch path triggers.
    fn make_large_diff_repo_view(dir: &tempfile::TempDir) -> crate::data::RepositoryView {
        use crate::data::{AiInfo, FieldExplanation, WorkingDirectoryInfo};
        use crate::git::commit::{FileChange, FileChanges, FileDiffRef};
        use crate::git::{CommitAnalysis, CommitInfo};

        let hash = "a".repeat(40);

        // Write a full (flat) diff file large enough to bust the budget.
        // With 50k context / 2.5 chars-per-token / 1.2 margin, available ≈ 41k tokens.
        // 120k chars → ~57,600 tokens → well over budget.
        let full_diff = "x".repeat(120_000);
        let flat_diff_path = dir.path().join("full.diff");
        std::fs::write(&flat_diff_path, &full_diff).unwrap();

        // Write two large per-file diff files (~30K chars each ≈ 14,400 tokens with
        // conservative 2.5 chars/token * 1.2 margin estimation)
        let diff_a = format!("diff --git a/src/a.rs b/src/a.rs\n{}\n", "a".repeat(30_000));
        let diff_b = format!("diff --git a/src/b.rs b/src/b.rs\n{}\n", "b".repeat(30_000));

        let path_a = dir.path().join("0000.diff");
        let path_b = dir.path().join("0001.diff");
        std::fs::write(&path_a, &diff_a).unwrap();
        std::fs::write(&path_b, &diff_b).unwrap();

        crate::data::RepositoryView {
            versions: None,
            explanation: FieldExplanation::default(),
            working_directory: WorkingDirectoryInfo {
                clean: true,
                untracked_changes: Vec::new(),
            },
            remotes: Vec::new(),
            ai: AiInfo {
                scratch: String::new(),
            },
            branch_info: None,
            pr_template: None,
            pr_template_location: None,
            branch_prs: None,
            commits: vec![CommitInfo {
                hash,
                author: "Test <test@test.com>".to_string(),
                date: chrono::Utc::now().fixed_offset(),
                original_message: "feat(test): large commit".to_string(),
                in_main_branches: Vec::new(),
                analysis: CommitAnalysis {
                    detected_type: "feat".to_string(),
                    detected_scope: "test".to_string(),
                    proposed_message: "feat(test): large commit".to_string(),
                    file_changes: FileChanges {
                        total_files: 2,
                        files_added: 2,
                        files_deleted: 0,
                        file_list: vec![
                            FileChange {
                                status: "A".to_string(),
                                file: "src/a.rs".to_string(),
                            },
                            FileChange {
                                status: "A".to_string(),
                                file: "src/b.rs".to_string(),
                            },
                        ],
                    },
                    diff_summary: " src/a.rs | 100 ++++\n src/b.rs | 100 ++++\n".to_string(),
                    diff_file: flat_diff_path.to_string_lossy().to_string(),
                    file_diffs: vec![
                        FileDiffRef {
                            path: "src/a.rs".to_string(),
                            diff_file: path_a.to_string_lossy().to_string(),
                            byte_len: diff_a.len(),
                        },
                        FileDiffRef {
                            path: "src/b.rs".to_string(),
                            diff_file: path_b.to_string_lossy().to_string(),
                            byte_len: diff_b.len(),
                        },
                    ],
                },
            }],
        }
    }

    fn valid_amendment_yaml(hash: &str, message: &str) -> String {
        format!("amendments:\n  - commit: \"{hash}\"\n    message: \"{message}\"")
    }

    #[tokio::test]
    async fn generate_amendments_split_dispatch() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_large_diff_repo_view(&dir);
        let hash = "a".repeat(40);

        // Responses: chunk 1 + chunk 2 + merge pass
        let client = make_small_context_client(vec![
            Ok(valid_amendment_yaml(&hash, "feat(a): add a.rs")),
            Ok(valid_amendment_yaml(&hash, "feat(b): add b.rs")),
            Ok(valid_amendment_yaml(&hash, "feat(test): add a.rs and b.rs")),
        ]);

        let result = client
            .generate_amendments_with_options(&repo_view, false)
            .await;

        assert!(result.is_ok(), "split dispatch failed: {:?}", result.err());
        let amendments = result.unwrap();
        assert_eq!(amendments.amendments.len(), 1);
        assert_eq!(amendments.amendments[0].commit, hash);
        assert!(amendments.amendments[0]
            .message
            .contains("add a.rs and b.rs"));
    }

    #[tokio::test]
    async fn generate_amendments_split_chunk_failure() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_large_diff_repo_view(&dir);
        let hash = "a".repeat(40);

        // First chunk succeeds, second chunk fails
        let client = make_small_context_client(vec![
            Ok(valid_amendment_yaml(&hash, "feat(a): add a.rs")),
            Err(anyhow::anyhow!("rate limit exceeded")),
        ]);

        let result = client
            .generate_amendments_with_options(&repo_view, false)
            .await;

        assert!(result.is_err());
    }

    #[tokio::test]
    async fn generate_amendments_no_split_when_fits() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_test_repo_view(&dir); // Small diff, no file_diffs
        let hash = format!("{:0>40}", 0);

        // Only one response needed — no split dispatch
        let client = make_configurable_client(vec![Ok(valid_amendment_yaml(
            &hash,
            "feat(test): improved message",
        ))]);

        let result = client
            .generate_amendments_with_options(&repo_view, false)
            .await;

        assert!(result.is_ok());
        assert_eq!(result.unwrap().amendments.len(), 1);
    }

    // ── check split dispatch tests ──────────────────────────────

    fn valid_check_yaml_for(hash: &str, passes: bool) -> String {
        format!(
            "checks:\n  - commit: \"{hash}\"\n    passes: {passes}\n    issues: []\n    summary: \"test summary\"\n"
        )
    }

    fn valid_check_yaml_with_issues(hash: &str) -> String {
        format!(
            concat!(
                "checks:\n",
                "  - commit: \"{hash}\"\n",
                "    passes: false\n",
                "    issues:\n",
                "      - severity: error\n",
                "        section: \"Subject Line\"\n",
                "        rule: \"imperative-mood\"\n",
                "        explanation: \"Subject uses past tense\"\n",
                "    suggestion:\n",
                "      message: \"feat(test): shorter subject\"\n",
                "      explanation: \"Shortened subject line\"\n",
                "    summary: \"Large commit with issues\"\n",
            ),
            hash = hash,
        )
    }

    fn valid_check_yaml_chunk_no_suggestion(hash: &str) -> String {
        format!(
            concat!(
                "checks:\n",
                "  - commit: \"{hash}\"\n",
                "    passes: true\n",
                "    issues: []\n",
                "    summary: \"chunk summary\"\n",
            ),
            hash = hash,
        )
    }

    #[tokio::test]
    async fn check_commits_split_dispatch() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_large_diff_repo_view(&dir);
        let hash = "a".repeat(40);

        // Responses: chunk 1 (issues + suggestion) + chunk 2 (issues + suggestion) + merge pass
        let client = make_small_context_client(vec![
            Ok(valid_check_yaml_with_issues(&hash)),
            Ok(valid_check_yaml_with_issues(&hash)),
            Ok(valid_check_yaml_with_issues(&hash)), // merge pass response
        ]);

        let result = client
            .check_commits_with_scopes(&repo_view, None, &[], true)
            .await;

        assert!(result.is_ok(), "split dispatch failed: {:?}", result.err());
        let report = result.unwrap();
        assert_eq!(report.commits.len(), 1);
        assert!(!report.commits[0].passes);
        // Dedup: both chunks report the same (rule, severity, section), so only 1 unique issue
        assert_eq!(report.commits[0].issues.len(), 1);
        assert_eq!(report.commits[0].issues[0].rule, "imperative-mood");
    }

    #[tokio::test]
    async fn check_commits_split_dispatch_no_merge_when_no_suggestions() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_large_diff_repo_view(&dir);
        let hash = "a".repeat(40);

        // Responses: chunk 1 + chunk 2, both passing with no suggestions
        // No merge pass needed — only 2 responses
        let client = make_small_context_client(vec![
            Ok(valid_check_yaml_chunk_no_suggestion(&hash)),
            Ok(valid_check_yaml_chunk_no_suggestion(&hash)),
        ]);

        let result = client
            .check_commits_with_scopes(&repo_view, None, &[], false)
            .await;

        assert!(result.is_ok(), "split dispatch failed: {:?}", result.err());
        let report = result.unwrap();
        assert_eq!(report.commits.len(), 1);
        assert!(report.commits[0].passes);
        assert!(report.commits[0].issues.is_empty());
        assert!(report.commits[0].suggestion.is_none());
        // First non-None summary from chunks
        assert_eq!(report.commits[0].summary.as_deref(), Some("chunk summary"));
    }

    #[tokio::test]
    async fn check_commits_split_chunk_failure() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_large_diff_repo_view(&dir);
        let hash = "a".repeat(40);

        // First chunk succeeds, second chunk fails
        let client = make_small_context_client(vec![
            Ok(valid_check_yaml_for(&hash, true)),
            Err(anyhow::anyhow!("rate limit exceeded")),
        ]);

        let result = client
            .check_commits_with_scopes(&repo_view, None, &[], false)
            .await;

        assert!(result.is_err());
    }

    #[tokio::test]
    async fn check_commits_no_split_when_fits() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_test_repo_view(&dir); // Small diff, no file_diffs
        let hash = format!("{:0>40}", 0);

        // Only one response needed — no split dispatch
        let client = make_configurable_client(vec![Ok(valid_check_yaml_for(&hash, true))]);

        let result = client
            .check_commits_with_scopes(&repo_view, None, &[], false)
            .await;

        assert!(result.is_ok());
        assert_eq!(result.unwrap().commits.len(), 1);
    }

    #[tokio::test]
    async fn check_commits_split_dedup_across_chunks() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_large_diff_repo_view(&dir);
        let hash = "a".repeat(40);

        // Chunk 1: two issues (error + warning)
        let chunk1 = format!(
            concat!(
                "checks:\n",
                "  - commit: \"{hash}\"\n",
                "    passes: false\n",
                "    issues:\n",
                "      - severity: error\n",
                "        section: \"Subject Line\"\n",
                "        rule: \"imperative-mood\"\n",
                "        explanation: \"Subject uses past tense\"\n",
                "      - severity: warning\n",
                "        section: \"Content\"\n",
                "        rule: \"body-required\"\n",
                "        explanation: \"Large change needs body\"\n",
            ),
            hash = hash,
        );

        // Chunk 2: same error (different wording) + new info issue
        let chunk2 = format!(
            concat!(
                "checks:\n",
                "  - commit: \"{hash}\"\n",
                "    passes: false\n",
                "    issues:\n",
                "      - severity: error\n",
                "        section: \"Subject Line\"\n",
                "        rule: \"imperative-mood\"\n",
                "        explanation: \"Subject line is too long\"\n",
                "      - severity: info\n",
                "        section: \"Style\"\n",
                "        rule: \"scope-suggestion\"\n",
                "        explanation: \"Consider more specific scope\"\n",
            ),
            hash = hash,
        );

        // No suggestions → no merge pass needed
        let client = make_small_context_client(vec![Ok(chunk1), Ok(chunk2)]);

        let result = client
            .check_commits_with_scopes(&repo_view, None, &[], false)
            .await;

        assert!(result.is_ok(), "split dispatch failed: {:?}", result.err());
        let report = result.unwrap();
        assert_eq!(report.commits.len(), 1);
        assert!(!report.commits[0].passes);
        // 3 unique issues: imperative-mood, body-required, scope-suggestion
        // (imperative-mood appears in both chunks but deduped)
        assert_eq!(report.commits[0].issues.len(), 3);
    }

    #[tokio::test]
    async fn check_commits_split_passes_only_when_all_chunks_pass() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_large_diff_repo_view(&dir);
        let hash = "a".repeat(40);

        // Chunk 1 passes, chunk 2 fails
        let client = make_small_context_client(vec![
            Ok(valid_check_yaml_for(&hash, true)),
            Ok(valid_check_yaml_for(&hash, false)),
        ]);

        let result = client
            .check_commits_with_scopes(&repo_view, None, &[], false)
            .await;

        assert!(result.is_ok(), "split dispatch failed: {:?}", result.err());
        let report = result.unwrap();
        assert!(
            !report.commits[0].passes,
            "should fail when any chunk fails"
        );
    }

    // ── multi-commit and PR generation paths ──────────────────────

    /// Creates a repo view with two small commits (fits budget without split dispatch).
    fn make_multi_commit_repo_view(dir: &tempfile::TempDir) -> crate::data::RepositoryView {
        use crate::data::{AiInfo, FieldExplanation, WorkingDirectoryInfo};
        use crate::git::commit::FileChanges;
        use crate::git::{CommitAnalysis, CommitInfo};

        let diff_a = dir.path().join("0.diff");
        let diff_b = dir.path().join("1.diff");
        std::fs::write(&diff_a, "+line a\n").unwrap();
        std::fs::write(&diff_b, "+line b\n").unwrap();

        let hash_a = "a".repeat(40);
        let hash_b = "b".repeat(40);

        crate::data::RepositoryView {
            versions: None,
            explanation: FieldExplanation::default(),
            working_directory: WorkingDirectoryInfo {
                clean: true,
                untracked_changes: Vec::new(),
            },
            remotes: Vec::new(),
            ai: AiInfo {
                scratch: String::new(),
            },
            branch_info: None,
            pr_template: None,
            pr_template_location: None,
            branch_prs: None,
            commits: vec![
                CommitInfo {
                    hash: hash_a,
                    author: "Test <test@test.com>".to_string(),
                    date: chrono::Utc::now().fixed_offset(),
                    original_message: "feat(a): add a".to_string(),
                    in_main_branches: Vec::new(),
                    analysis: CommitAnalysis {
                        detected_type: "feat".to_string(),
                        detected_scope: "a".to_string(),
                        proposed_message: "feat(a): add a".to_string(),
                        file_changes: FileChanges {
                            total_files: 1,
                            files_added: 1,
                            files_deleted: 0,
                            file_list: Vec::new(),
                        },
                        diff_summary: "a.rs | 1 +".to_string(),
                        diff_file: diff_a.to_string_lossy().to_string(),
                        file_diffs: Vec::new(),
                    },
                },
                CommitInfo {
                    hash: hash_b,
                    author: "Test <test@test.com>".to_string(),
                    date: chrono::Utc::now().fixed_offset(),
                    original_message: "feat(b): add b".to_string(),
                    in_main_branches: Vec::new(),
                    analysis: CommitAnalysis {
                        detected_type: "feat".to_string(),
                        detected_scope: "b".to_string(),
                        proposed_message: "feat(b): add b".to_string(),
                        file_changes: FileChanges {
                            total_files: 1,
                            files_added: 1,
                            files_deleted: 0,
                            file_list: Vec::new(),
                        },
                        diff_summary: "b.rs | 1 +".to_string(),
                        diff_file: diff_b.to_string_lossy().to_string(),
                        file_diffs: Vec::new(),
                    },
                },
            ],
        }
    }

    #[tokio::test]
    async fn generate_amendments_multi_commit() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_multi_commit_repo_view(&dir);
        let hash_a = "a".repeat(40);
        let hash_b = "b".repeat(40);

        let response = format!(
            concat!(
                "amendments:\n",
                "  - commit: \"{hash_a}\"\n",
                "    message: \"feat(a): improved a\"\n",
                "  - commit: \"{hash_b}\"\n",
                "    message: \"feat(b): improved b\"\n",
            ),
            hash_a = hash_a,
            hash_b = hash_b,
        );
        let client = make_configurable_client(vec![Ok(response)]);

        let result = client
            .generate_amendments_with_options(&repo_view, false)
            .await;

        assert!(
            result.is_ok(),
            "multi-commit amendment failed: {:?}",
            result.err()
        );
        let amendments = result.unwrap();
        assert_eq!(amendments.amendments.len(), 2);
    }

    #[tokio::test]
    async fn generate_contextual_amendments_multi_commit() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_multi_commit_repo_view(&dir);
        let hash_a = "a".repeat(40);
        let hash_b = "b".repeat(40);

        let response = format!(
            concat!(
                "amendments:\n",
                "  - commit: \"{hash_a}\"\n",
                "    message: \"feat(a): improved a\"\n",
                "  - commit: \"{hash_b}\"\n",
                "    message: \"feat(b): improved b\"\n",
            ),
            hash_a = hash_a,
            hash_b = hash_b,
        );
        let client = make_configurable_client(vec![Ok(response)]);
        let context = crate::data::context::CommitContext::default();

        let result = client
            .generate_contextual_amendments_with_options(&repo_view, &context, false)
            .await;

        assert!(
            result.is_ok(),
            "multi-commit contextual amendment failed: {:?}",
            result.err()
        );
        let amendments = result.unwrap();
        assert_eq!(amendments.amendments.len(), 2);
    }

    #[tokio::test]
    async fn generate_pr_content_succeeds() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_test_repo_view(&dir);

        let response = "title: \"feat: add something\"\ndescription: \"Adds a new feature.\"\n";
        let client = make_configurable_client(vec![Ok(response.to_string())]);

        let result = client.generate_pr_content(&repo_view, "").await;

        assert!(result.is_ok(), "PR generation failed: {:?}", result.err());
        let pr = result.unwrap();
        assert_eq!(pr.title, "feat: add something");
        assert_eq!(pr.description, "Adds a new feature.");
    }

    #[tokio::test]
    async fn generate_pr_content_with_context_succeeds() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_test_repo_view(&dir);
        let context = crate::data::context::CommitContext::default();

        let response = "title: \"feat: add something\"\ndescription: \"Adds a new feature.\"\n";
        let client = make_configurable_client(vec![Ok(response.to_string())]);

        let result = client
            .generate_pr_content_with_context(&repo_view, "", &context)
            .await;

        assert!(
            result.is_ok(),
            "PR generation with context failed: {:?}",
            result.err()
        );
        let pr = result.unwrap();
        assert_eq!(pr.title, "feat: add something");
    }

    #[tokio::test]
    async fn check_commits_multi_commit() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_multi_commit_repo_view(&dir);
        let hash_a = "a".repeat(40);
        let hash_b = "b".repeat(40);

        let response = format!(
            concat!(
                "checks:\n",
                "  - commit: \"{hash_a}\"\n",
                "    passes: true\n",
                "    issues: []\n",
                "  - commit: \"{hash_b}\"\n",
                "    passes: true\n",
                "    issues: []\n",
            ),
            hash_a = hash_a,
            hash_b = hash_b,
        );
        let client = make_configurable_client(vec![Ok(response)]);

        let result = client
            .check_commits_with_scopes(&repo_view, None, &[], false)
            .await;

        assert!(
            result.is_ok(),
            "multi-commit check failed: {:?}",
            result.err()
        );
        let report = result.unwrap();
        assert_eq!(report.commits.len(), 2);
        assert!(report.commits[0].passes);
        assert!(report.commits[1].passes);
    }

    // ── Multi-commit split dispatch helpers ──────────────────────────

    /// Creates a repo view with two large-diff commits whose combined view
    /// exceeds the constrained 25KB context window.
    fn make_large_multi_commit_repo_view(dir: &tempfile::TempDir) -> crate::data::RepositoryView {
        use crate::data::{AiInfo, FieldExplanation, WorkingDirectoryInfo};
        use crate::git::commit::{FileChange, FileChanges, FileDiffRef};
        use crate::git::{CommitAnalysis, CommitInfo};

        let hash_a = "a".repeat(40);
        let hash_b = "b".repeat(40);

        // Write flat diff files large enough to bust the 50K-token budget when combined.
        // Each 60k chars ≈ 28,800 tokens; combined ≈ 57,600 > 41,808 available.
        let diff_content_a = "x".repeat(60_000);
        let diff_content_b = "y".repeat(60_000);
        let flat_a = dir.path().join("flat_a.diff");
        let flat_b = dir.path().join("flat_b.diff");
        std::fs::write(&flat_a, &diff_content_a).unwrap();
        std::fs::write(&flat_b, &diff_content_b).unwrap();

        // Write per-file diff files for split dispatch
        let file_diff_a = format!("diff --git a/src/a.rs b/src/a.rs\n{}\n", "a".repeat(30_000));
        let file_diff_b = format!("diff --git a/src/b.rs b/src/b.rs\n{}\n", "b".repeat(30_000));
        let per_file_a = dir.path().join("pf_a.diff");
        let per_file_b = dir.path().join("pf_b.diff");
        std::fs::write(&per_file_a, &file_diff_a).unwrap();
        std::fs::write(&per_file_b, &file_diff_b).unwrap();

        crate::data::RepositoryView {
            versions: None,
            explanation: FieldExplanation::default(),
            working_directory: WorkingDirectoryInfo {
                clean: true,
                untracked_changes: Vec::new(),
            },
            remotes: Vec::new(),
            ai: AiInfo {
                scratch: String::new(),
            },
            branch_info: None,
            pr_template: None,
            pr_template_location: None,
            branch_prs: None,
            commits: vec![
                CommitInfo {
                    hash: hash_a,
                    author: "Test <test@test.com>".to_string(),
                    date: chrono::Utc::now().fixed_offset(),
                    original_message: "feat(a): add module a".to_string(),
                    in_main_branches: Vec::new(),
                    analysis: CommitAnalysis {
                        detected_type: "feat".to_string(),
                        detected_scope: "a".to_string(),
                        proposed_message: "feat(a): add module a".to_string(),
                        file_changes: FileChanges {
                            total_files: 1,
                            files_added: 1,
                            files_deleted: 0,
                            file_list: vec![FileChange {
                                status: "A".to_string(),
                                file: "src/a.rs".to_string(),
                            }],
                        },
                        diff_summary: " src/a.rs | 100 ++++\n".to_string(),
                        diff_file: flat_a.to_string_lossy().to_string(),
                        file_diffs: vec![FileDiffRef {
                            path: "src/a.rs".to_string(),
                            diff_file: per_file_a.to_string_lossy().to_string(),
                            byte_len: file_diff_a.len(),
                        }],
                    },
                },
                CommitInfo {
                    hash: hash_b,
                    author: "Test <test@test.com>".to_string(),
                    date: chrono::Utc::now().fixed_offset(),
                    original_message: "feat(b): add module b".to_string(),
                    in_main_branches: Vec::new(),
                    analysis: CommitAnalysis {
                        detected_type: "feat".to_string(),
                        detected_scope: "b".to_string(),
                        proposed_message: "feat(b): add module b".to_string(),
                        file_changes: FileChanges {
                            total_files: 1,
                            files_added: 1,
                            files_deleted: 0,
                            file_list: vec![FileChange {
                                status: "A".to_string(),
                                file: "src/b.rs".to_string(),
                            }],
                        },
                        diff_summary: " src/b.rs | 100 ++++\n".to_string(),
                        diff_file: flat_b.to_string_lossy().to_string(),
                        file_diffs: vec![FileDiffRef {
                            path: "src/b.rs".to_string(),
                            diff_file: per_file_b.to_string_lossy().to_string(),
                            byte_len: file_diff_b.len(),
                        }],
                    },
                },
            ],
        }
    }

    fn valid_pr_yaml(title: &str, description: &str) -> String {
        format!("title: \"{title}\"\ndescription: \"{description}\"\n")
    }

    // ── Multi-commit amendment split dispatch tests ──────────────────

    #[tokio::test]
    async fn generate_amendments_multi_commit_split_dispatch() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_large_multi_commit_repo_view(&dir);
        let hash_a = "a".repeat(40);
        let hash_b = "b".repeat(40);

        // Full view exceeds budget → per-commit fallback
        // Each commit fits individually (1 file each) → 1 response per commit
        let (client, handle) = make_small_context_client_tracked(vec![
            Ok(valid_amendment_yaml(&hash_a, "feat(a): improved a")),
            Ok(valid_amendment_yaml(&hash_b, "feat(b): improved b")),
        ]);

        let result = client
            .generate_amendments_with_options(&repo_view, false)
            .await;

        assert!(
            result.is_ok(),
            "multi-commit split dispatch failed: {:?}",
            result.err()
        );
        let amendments = result.unwrap();
        assert_eq!(amendments.amendments.len(), 2);
        assert_eq!(amendments.amendments[0].commit, hash_a);
        assert_eq!(amendments.amendments[1].commit, hash_b);
        assert!(amendments.amendments[0].message.contains("improved a"));
        assert!(amendments.amendments[1].message.contains("improved b"));
        assert_eq!(handle.remaining(), 0, "expected all responses consumed");
    }

    #[tokio::test]
    async fn generate_contextual_amendments_multi_commit_split_dispatch() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_large_multi_commit_repo_view(&dir);
        let hash_a = "a".repeat(40);
        let hash_b = "b".repeat(40);
        let context = crate::data::context::CommitContext::default();

        let (client, handle) = make_small_context_client_tracked(vec![
            Ok(valid_amendment_yaml(&hash_a, "feat(a): improved a")),
            Ok(valid_amendment_yaml(&hash_b, "feat(b): improved b")),
        ]);

        let result = client
            .generate_contextual_amendments_with_options(&repo_view, &context, false)
            .await;

        assert!(
            result.is_ok(),
            "multi-commit contextual split dispatch failed: {:?}",
            result.err()
        );
        let amendments = result.unwrap();
        assert_eq!(amendments.amendments.len(), 2);
        assert_eq!(amendments.amendments[0].commit, hash_a);
        assert_eq!(amendments.amendments[1].commit, hash_b);
        assert_eq!(handle.remaining(), 0, "expected all responses consumed");
    }

    // ── Multi-commit check split dispatch tests ──────────────────────

    #[tokio::test]
    async fn check_commits_multi_commit_split_dispatch() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_large_multi_commit_repo_view(&dir);
        let hash_a = "a".repeat(40);
        let hash_b = "b".repeat(40);

        // Full view exceeds budget → per-commit fallback
        let (client, handle) = make_small_context_client_tracked(vec![
            Ok(valid_check_yaml_for(&hash_a, true)),
            Ok(valid_check_yaml_for(&hash_b, true)),
        ]);

        let result = client
            .check_commits_with_scopes(&repo_view, None, &[], false)
            .await;

        assert!(
            result.is_ok(),
            "multi-commit check split dispatch failed: {:?}",
            result.err()
        );
        let report = result.unwrap();
        assert_eq!(report.commits.len(), 2);
        assert!(report.commits[0].passes);
        assert!(report.commits[1].passes);
        assert_eq!(handle.remaining(), 0, "expected all responses consumed");
    }

    // ── PR split dispatch tests ──────────────────────────────────────

    #[tokio::test]
    async fn generate_pr_content_split_dispatch() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_large_diff_repo_view(&dir);

        // Single large commit: full view exceeds budget → per-commit fallback
        // 1 commit with 2 file chunks → chunk 1 + chunk 2 + chunk merge pass
        // Single per-commit result → returned directly (no extra merge)
        let (client, handle) = make_small_context_client_tracked(vec![
            Ok(valid_pr_yaml("feat(a): add a.rs", "Adds a.rs module")),
            Ok(valid_pr_yaml("feat(b): add b.rs", "Adds b.rs module")),
            Ok(valid_pr_yaml(
                "feat(test): add modules",
                "Adds a.rs and b.rs",
            )),
        ]);

        let result = client.generate_pr_content(&repo_view, "").await;

        assert!(
            result.is_ok(),
            "PR split dispatch failed: {:?}",
            result.err()
        );
        let pr = result.unwrap();
        assert!(pr.title.contains("add modules"));
        assert_eq!(handle.remaining(), 0, "expected all responses consumed");
    }

    #[tokio::test]
    async fn generate_pr_content_multi_commit_split_dispatch() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_large_multi_commit_repo_view(&dir);

        // Full view exceeds budget → per-commit fallback
        // Each commit fits individually → 1 response per commit, then merge pass
        let (client, handle) = make_small_context_client_tracked(vec![
            Ok(valid_pr_yaml("feat(a): add module a", "Adds module a")),
            Ok(valid_pr_yaml("feat(b): add module b", "Adds module b")),
            Ok(valid_pr_yaml(
                "feat: add modules a and b",
                "Adds both modules",
            )),
        ]);

        let result = client.generate_pr_content(&repo_view, "").await;

        assert!(
            result.is_ok(),
            "PR multi-commit split dispatch failed: {:?}",
            result.err()
        );
        let pr = result.unwrap();
        assert!(pr.title.contains("modules"));
        assert_eq!(handle.remaining(), 0, "expected all responses consumed");
    }

    #[tokio::test]
    async fn generate_pr_content_with_context_from_commits_succeeds() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_test_repo_view(&dir);
        let context = crate::data::context::CommitContext::default();

        let response = "title: \"feat: from commits\"\ndescription: \"derived from commit\"\n";
        let client = make_configurable_client(vec![Ok(response.to_string())]);

        let result = client
            .generate_pr_content_with_context_from_commits(&repo_view, "", &context)
            .await;

        assert!(
            result.is_ok(),
            "PR from-commits generation failed: {:?}",
            result.err()
        );
        let pr = result.unwrap();
        assert_eq!(pr.title, "feat: from commits");
    }

    #[tokio::test]
    async fn generate_pr_content_with_context_from_commits_omits_diff_in_prompt() {
        let dir = tempfile::tempdir().unwrap();
        // Write a recognisable diff payload to the diff_file so we can prove it
        // is NOT being read into the prompt.
        let diff_path = dir.path().join("recognisable.diff");
        std::fs::write(
            &diff_path,
            "diff --git a/x b/x\n@@ -1 +1 @@\n-old\n+UNIQUE_DIFF_MARKER\n",
        )
        .unwrap();

        // Build a repo view whose single commit's diff_file points at the
        // marker file but whose commit message contains a distinct subject.
        let repo_view = {
            use crate::data::{AiInfo, FieldExplanation, WorkingDirectoryInfo};
            use crate::git::commit::FileChanges;
            use crate::git::{CommitAnalysis, CommitInfo};
            crate::data::RepositoryView {
                versions: None,
                explanation: FieldExplanation::default(),
                working_directory: WorkingDirectoryInfo {
                    clean: true,
                    untracked_changes: Vec::new(),
                },
                remotes: Vec::new(),
                ai: AiInfo {
                    scratch: String::new(),
                },
                branch_info: None,
                pr_template: None,
                pr_template_location: None,
                branch_prs: None,
                commits: vec![CommitInfo {
                    hash: format!("{:0>40}", 0),
                    author: "Test <test@test.com>".to_string(),
                    date: chrono::Utc::now().fixed_offset(),
                    original_message: "feat(test): UNIQUE_COMMIT_SUBJECT_MARKER".to_string(),
                    in_main_branches: Vec::new(),
                    analysis: CommitAnalysis {
                        detected_type: "feat".to_string(),
                        detected_scope: "test".to_string(),
                        proposed_message: "feat(test): unique".to_string(),
                        file_changes: FileChanges {
                            total_files: 1,
                            files_added: 1,
                            files_deleted: 0,
                            file_list: Vec::new(),
                        },
                        diff_summary: "UNIQUE_STAT_MARKER | 1 +".to_string(),
                        diff_file: diff_path.to_string_lossy().to_string(),
                        file_diffs: Vec::new(),
                    },
                }],
            }
        };
        let context = crate::data::context::CommitContext::default();

        let (client, _resp_handle, prompt_handle) =
            make_configurable_client_with_prompts(vec![Ok(
                "title: \"feat: x\"\ndescription: \"y\"\n".to_string(),
            )]);

        client
            .generate_pr_content_with_context_from_commits(&repo_view, "", &context)
            .await
            .unwrap();

        let prompts = prompt_handle.prompts();
        assert_eq!(prompts.len(), 1, "expected exactly one AI call");
        let (system_prompt, user_prompt) = &prompts[0];

        // Commit narrative IS present
        assert!(
            user_prompt.contains("UNIQUE_COMMIT_SUBJECT_MARKER"),
            "user prompt must include the commit subject"
        );
        // No diff content
        assert!(
            !user_prompt.contains("UNIQUE_DIFF_MARKER"),
            "user prompt must NOT include diff content: {user_prompt}"
        );
        assert!(
            !user_prompt.contains("diff --git"),
            "user prompt must NOT include diff hunks"
        );
        assert!(
            !user_prompt.contains("diff_content"),
            "user prompt must NOT include diff_content YAML field"
        );
        assert!(
            !user_prompt.contains("UNIQUE_STAT_MARKER"),
            "user prompt must NOT include diff_summary"
        );
        // System prompt references commits, not diff files
        assert!(
            !system_prompt.contains("diff files"),
            "system prompt must not mention diff files"
        );
    }

    #[tokio::test]
    async fn generate_pr_content_with_context_default_mode_includes_diff_in_prompt() {
        // Regression test locking in the byte-identical-default guarantee:
        // when --from-commits is OFF, the prompt MUST still contain diff content.
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_test_repo_view(&dir);
        let context = crate::data::context::CommitContext::default();

        let (client, _resp_handle, prompt_handle) =
            make_configurable_client_with_prompts(vec![Ok(
                "title: \"feat: x\"\ndescription: \"y\"\n".to_string(),
            )]);

        client
            .generate_pr_content_with_context(&repo_view, "", &context)
            .await
            .unwrap();

        let prompts = prompt_handle.prompts();
        assert_eq!(prompts.len(), 1);
        let (_system_prompt, user_prompt) = &prompts[0];
        assert!(
            user_prompt.contains("diff_content"),
            "default mode must still serialise diff_content into the prompt"
        );
    }

    #[tokio::test]
    async fn generate_pr_content_with_context_from_commits_multi_commit_per_commit_dispatch() {
        // Force per-commit fallback by giving each commit a large message
        // so the combined view busts the small (50K) context window.
        use crate::data::{AiInfo, FieldExplanation, WorkingDirectoryInfo};
        use crate::git::commit::FileChanges;
        use crate::git::{CommitAnalysis, CommitInfo};

        let dir = tempfile::tempdir().unwrap();
        let make_commit = |hash: String, marker: &str, msg_size: usize| {
            let diff_path = dir.path().join(format!("{}.diff", &hash[..4]));
            std::fs::write(&diff_path, "+x\n").unwrap();
            CommitInfo {
                hash,
                author: "Test <test@test.com>".to_string(),
                date: chrono::Utc::now().fixed_offset(),
                original_message: format!("feat({marker}): {}", "x".repeat(msg_size)),
                in_main_branches: Vec::new(),
                analysis: CommitAnalysis {
                    detected_type: "feat".to_string(),
                    detected_scope: marker.to_string(),
                    proposed_message: format!("feat({marker}): t"),
                    file_changes: FileChanges {
                        total_files: 1,
                        files_added: 1,
                        files_deleted: 0,
                        file_list: Vec::new(),
                    },
                    diff_summary: String::new(),
                    diff_file: diff_path.to_string_lossy().to_string(),
                    file_diffs: Vec::new(),
                },
            }
        };

        // Two commits, each with a ~80KB message → combined ~160KB chars,
        // exceeds the 50K-context-length mock's ~20K-token input budget.
        let repo_view = crate::data::RepositoryView {
            versions: None,
            explanation: FieldExplanation::default(),
            working_directory: WorkingDirectoryInfo {
                clean: true,
                untracked_changes: Vec::new(),
            },
            remotes: Vec::new(),
            ai: AiInfo {
                scratch: String::new(),
            },
            branch_info: None,
            pr_template: None,
            pr_template_location: None,
            branch_prs: None,
            commits: vec![
                make_commit("a".repeat(40), "a", 80_000),
                make_commit("b".repeat(40), "b", 80_000),
            ],
        };
        let context = crate::data::context::CommitContext::default();

        // Full view exceeds → per-commit fallback (2 calls) → merge (1 call).
        let (client, handle) = make_small_context_client_tracked(vec![
            Ok(valid_pr_yaml("feat(a): a", "did a")),
            Ok(valid_pr_yaml("feat(b): b", "did b")),
            Ok(valid_pr_yaml("feat: a and b", "did both")),
        ]);

        let result = client
            .generate_pr_content_with_context_from_commits(&repo_view, "", &context)
            .await;

        assert!(
            result.is_ok(),
            "from-commits per-commit dispatch failed: {:?}",
            result.err()
        );
        let pr = result.unwrap();
        assert!(
            pr.title.contains("and"),
            "unexpected merged title: {}",
            pr.title
        );
        assert_eq!(handle.remaining(), 0, "expected all responses consumed");
    }

    #[tokio::test]
    async fn generate_pr_content_with_context_from_commits_single_commit_per_commit_return() {
        // Force per-commit fallback with a single commit whose message busts
        // the budget on the full view but fits in the slimmer single-commit
        // view (no envelope around it). Exercises the
        // `per_commit_contents.len() == 1` return branch — no merge pass.
        use crate::data::{AiInfo, FieldExplanation, WorkingDirectoryInfo};
        use crate::git::commit::FileChanges;
        use crate::git::{CommitAnalysis, CommitInfo};

        let dir = tempfile::tempdir().unwrap();
        let diff_path = dir.path().join("0.diff");
        std::fs::write(&diff_path, "+x\n").unwrap();
        let commit = CommitInfo {
            hash: "a".repeat(40),
            author: "Test <test@test.com>".to_string(),
            date: chrono::Utc::now().fixed_offset(),
            // ~80KB message → busts the 50K-context client's input budget
            // when wrapped in the full envelope, but the slimmer single-commit
            // view (envelope stripped) still fits with the small context.
            original_message: format!("feat(only): {}", "x".repeat(80_000)),
            in_main_branches: Vec::new(),
            analysis: CommitAnalysis {
                detected_type: "feat".to_string(),
                detected_scope: "only".to_string(),
                proposed_message: "feat(only): m".to_string(),
                file_changes: FileChanges {
                    total_files: 1,
                    files_added: 1,
                    files_deleted: 0,
                    file_list: Vec::new(),
                },
                diff_summary: String::new(),
                diff_file: diff_path.to_string_lossy().to_string(),
                file_diffs: Vec::new(),
            },
        };
        let repo_view = crate::data::RepositoryView {
            versions: None,
            explanation: FieldExplanation::default(),
            working_directory: WorkingDirectoryInfo {
                clean: true,
                untracked_changes: Vec::new(),
            },
            remotes: Vec::new(),
            ai: AiInfo {
                scratch: String::new(),
            },
            branch_info: None,
            pr_template: None,
            pr_template_location: None,
            branch_prs: None,
            commits: vec![commit],
        };
        let context = crate::data::context::CommitContext::default();

        // Full envelope view exceeds → per-commit dispatch (1 call) → single
        // result returned directly, no merge.
        let (client, handle) = make_small_context_client_tracked(vec![Ok(valid_pr_yaml(
            "feat(only): direct return",
            "single commit body",
        ))]);

        let result = client
            .generate_pr_content_with_context_from_commits(&repo_view, "", &context)
            .await;

        assert!(
            result.is_ok(),
            "from-commits single-commit per-commit dispatch failed: {:?}",
            result.err()
        );
        let pr = result.unwrap();
        assert_eq!(pr.title, "feat(only): direct return");
        assert_eq!(
            handle.remaining(),
            0,
            "exactly one response should be consumed (no merge call)"
        );
    }

    #[tokio::test]
    async fn generate_pr_content_with_context_from_commits_bails_on_oversized_single_commit() {
        // A single commit whose own slim view still exceeds the budget triggers
        // the bail in `generate_pr_content_for_commit_from_commits`.
        use crate::data::{AiInfo, FieldExplanation, WorkingDirectoryInfo};
        use crate::git::commit::FileChanges;
        use crate::git::{CommitAnalysis, CommitInfo};

        let dir = tempfile::tempdir().unwrap();
        let diff_path = dir.path().join("0.diff");
        std::fs::write(&diff_path, "+x\n").unwrap();
        let commit = CommitInfo {
            hash: "c".repeat(40),
            author: "Test <test@test.com>".to_string(),
            date: chrono::Utc::now().fixed_offset(),
            // Message large enough that even the single-commit slim view
            // overflows the 50K context window's input budget.
            original_message: format!("feat: {}", "z".repeat(200_000)),
            in_main_branches: Vec::new(),
            analysis: CommitAnalysis {
                detected_type: "feat".to_string(),
                detected_scope: String::new(),
                proposed_message: "feat: oversized".to_string(),
                file_changes: FileChanges {
                    total_files: 0,
                    files_added: 0,
                    files_deleted: 0,
                    file_list: Vec::new(),
                },
                diff_summary: String::new(),
                diff_file: diff_path.to_string_lossy().to_string(),
                file_diffs: Vec::new(),
            },
        };
        let repo_view = crate::data::RepositoryView {
            versions: None,
            explanation: FieldExplanation::default(),
            working_directory: WorkingDirectoryInfo {
                clean: true,
                untracked_changes: Vec::new(),
            },
            remotes: Vec::new(),
            ai: AiInfo {
                scratch: String::new(),
            },
            branch_info: None,
            pr_template: None,
            pr_template_location: None,
            branch_prs: None,
            commits: vec![commit],
        };
        let context = crate::data::context::CommitContext::default();

        // No mock responses are needed; the call should bail before making
        // any AI request.
        let client = make_small_context_client(Vec::new());

        let result = client
            .generate_pr_content_with_context_from_commits(&repo_view, "", &context)
            .await;

        assert!(result.is_err(), "expected bail on oversized single commit");
        let msg = format!("{:#}", result.unwrap_err());
        assert!(
            msg.contains("Token budget exceeded"),
            "expected token-budget error message, got: {msg}"
        );
        assert!(
            msg.contains("--from-commits"),
            "error should reference the from-commits mode"
        );
    }

    #[tokio::test]
    async fn generate_pr_content_with_context_split_dispatch() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_large_multi_commit_repo_view(&dir);
        let context = crate::data::context::CommitContext::default();

        // Full view exceeds budget → per-commit fallback → merge pass
        let (client, handle) = make_small_context_client_tracked(vec![
            Ok(valid_pr_yaml("feat(a): add module a", "Adds module a")),
            Ok(valid_pr_yaml("feat(b): add module b", "Adds module b")),
            Ok(valid_pr_yaml(
                "feat: add modules a and b",
                "Adds both modules",
            )),
        ]);

        let result = client
            .generate_pr_content_with_context(&repo_view, "", &context)
            .await;

        assert!(
            result.is_ok(),
            "PR with context split dispatch failed: {:?}",
            result.err()
        );
        let pr = result.unwrap();
        assert!(pr.title.contains("modules"));
        assert_eq!(handle.remaining(), 0, "expected all responses consumed");
    }

    // ── prompt-recording split dispatch tests ────────────────────

    /// Like [`make_small_context_client_tracked`] but also returns a
    /// [`PromptRecordHandle`] for inspecting which prompts were sent.
    fn make_small_context_client_with_prompts(
        responses: Vec<Result<String>>,
    ) -> (
        ClaudeClient,
        crate::claude::test_utils::ResponseQueueHandle,
        crate::claude::test_utils::PromptRecordHandle,
    ) {
        let mock = crate::claude::test_utils::ConfigurableMockAiClient::new(responses)
            .with_context_length(50_000);
        let response_handle = mock.response_handle();
        let prompt_handle = mock.prompt_handle();
        (
            ClaudeClient::new(Box::new(mock)),
            response_handle,
            prompt_handle,
        )
    }

    /// Creates a default-context mock client that also records prompts.
    fn make_configurable_client_with_prompts(
        responses: Vec<Result<String>>,
    ) -> (
        ClaudeClient,
        crate::claude::test_utils::ResponseQueueHandle,
        crate::claude::test_utils::PromptRecordHandle,
    ) {
        let mock = crate::claude::test_utils::ConfigurableMockAiClient::new(responses);
        let response_handle = mock.response_handle();
        let prompt_handle = mock.prompt_handle();
        (
            ClaudeClient::new(Box::new(mock)),
            response_handle,
            prompt_handle,
        )
    }

    /// Creates a repo view with one commit containing a single large file
    /// whose diff exceeds the token budget. Because the per-file diff is
    /// loaded as a whole (hunk-level granularity from the packer is lost
    /// at the dispatch layer), the split dispatch path will fail with a
    /// budget error. This helper exists to test that the error propagates
    /// cleanly rather than silently degrading.
    fn make_single_oversized_file_repo_view(
        dir: &tempfile::TempDir,
    ) -> crate::data::RepositoryView {
        use crate::data::{AiInfo, FieldExplanation, WorkingDirectoryInfo};
        use crate::git::commit::{FileChange, FileChanges, FileDiffRef};
        use crate::git::{CommitAnalysis, CommitInfo};

        let hash = "c".repeat(40);

        // A single file diff large enough (~80K bytes ≈ 25K tokens) to
        // exceed the 25K context window budget even for a single chunk.
        let diff_content = format!(
            "diff --git a/src/big.rs b/src/big.rs\n{}\n",
            "x".repeat(80_000)
        );

        let flat_diff_path = dir.path().join("full.diff");
        std::fs::write(&flat_diff_path, &diff_content).unwrap();

        let per_file_path = dir.path().join("0000.diff");
        std::fs::write(&per_file_path, &diff_content).unwrap();

        crate::data::RepositoryView {
            versions: None,
            explanation: FieldExplanation::default(),
            working_directory: WorkingDirectoryInfo {
                clean: true,
                untracked_changes: Vec::new(),
            },
            remotes: Vec::new(),
            ai: AiInfo {
                scratch: String::new(),
            },
            branch_info: None,
            pr_template: None,
            pr_template_location: None,
            branch_prs: None,
            commits: vec![CommitInfo {
                hash,
                author: "Test <test@test.com>".to_string(),
                date: chrono::Utc::now().fixed_offset(),
                original_message: "feat(big): add large module".to_string(),
                in_main_branches: Vec::new(),
                analysis: CommitAnalysis {
                    detected_type: "feat".to_string(),
                    detected_scope: "big".to_string(),
                    proposed_message: "feat(big): add large module".to_string(),
                    file_changes: FileChanges {
                        total_files: 1,
                        files_added: 1,
                        files_deleted: 0,
                        file_list: vec![FileChange {
                            status: "A".to_string(),
                            file: "src/big.rs".to_string(),
                        }],
                    },
                    diff_summary: " src/big.rs | 80 ++++\n".to_string(),
                    diff_file: flat_diff_path.to_string_lossy().to_string(),
                    file_diffs: vec![FileDiffRef {
                        path: "src/big.rs".to_string(),
                        diff_file: per_file_path.to_string_lossy().to_string(),
                        byte_len: diff_content.len(),
                    }],
                },
            }],
        }
    }

    /// A small single-file commit whose diff fits within the token budget.
    ///
    /// Exercises the non-split path: `generate_amendments_with_options` →
    /// `try_full_diff_budget` succeeds → single AI request → amendment
    /// returned directly. Verifies exactly one request is made and the
    /// user prompt contains the actual diff content.
    #[tokio::test]
    async fn amendment_single_file_under_budget_no_split() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_test_repo_view(&dir);
        let hash = format!("{:0>40}", 0);

        let (client, response_handle, prompt_handle) =
            make_configurable_client_with_prompts(vec![Ok(valid_amendment_yaml(
                &hash,
                "feat(test): improved message",
            ))]);

        let result = client
            .generate_amendments_with_options(&repo_view, false)
            .await;

        assert!(result.is_ok());
        assert_eq!(result.unwrap().amendments.len(), 1);
        assert_eq!(response_handle.remaining(), 0);

        let prompts = prompt_handle.prompts();
        assert_eq!(
            prompts.len(),
            1,
            "expected exactly one AI request, no split"
        );

        let (_, user_prompt) = &prompts[0];
        assert!(
            user_prompt.contains("added line"),
            "user prompt should contain the diff content"
        );
    }

    /// A two-file commit that exceeds the token budget when combined.
    ///
    /// Exercises the file-level split path: `generate_amendments_with_options`
    /// → `try_full_diff_budget` fails → `generate_amendment_for_commit` →
    /// `try_full_diff_budget` fails again → `generate_amendment_split` →
    /// `pack_file_diffs` creates 2 chunks (one file each) → 2 AI requests
    /// → `merge_amendment_chunks` reduce pass → 1 merged amendment.
    ///
    /// Verifies that each chunk's user prompt contains only its file's diff
    /// content, and the merge prompt contains both partial amendment messages.
    #[tokio::test]
    async fn amendment_two_chunks_prompt_content() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_large_diff_repo_view(&dir);
        let hash = "a".repeat(40);

        let (client, response_handle, prompt_handle) =
            make_small_context_client_with_prompts(vec![
                Ok(valid_amendment_yaml(&hash, "feat(a): add a.rs")),
                Ok(valid_amendment_yaml(&hash, "feat(b): add b.rs")),
                Ok(valid_amendment_yaml(&hash, "feat(test): add a.rs and b.rs")),
            ]);

        let result = client
            .generate_amendments_with_options(&repo_view, false)
            .await;

        assert!(result.is_ok(), "split dispatch failed: {:?}", result.err());
        let amendments = result.unwrap();
        assert_eq!(amendments.amendments.len(), 1);
        assert!(amendments.amendments[0]
            .message
            .contains("add a.rs and b.rs"));
        assert_eq!(response_handle.remaining(), 0);

        let prompts = prompt_handle.prompts();
        assert_eq!(prompts.len(), 3, "expected 2 chunks + 1 merge = 3 requests");

        // Chunk 1 should contain file-a diff content (repeated 'a' chars)
        let (_, chunk1_user) = &prompts[0];
        assert!(
            chunk1_user.contains("aaa"),
            "chunk 1 prompt should contain file-a diff content"
        );

        // Chunk 2 should contain file-b diff content (repeated 'b' chars)
        let (_, chunk2_user) = &prompts[1];
        assert!(
            chunk2_user.contains("bbb"),
            "chunk 2 prompt should contain file-b diff content"
        );

        // Merge pass: system prompt is the synthesis prompt
        let (merge_sys, merge_user) = &prompts[2];
        assert!(
            merge_sys.contains("synthesiz"),
            "merge system prompt should contain synthesis instructions"
        );
        // Merge user prompt should contain both partial messages
        assert!(
            merge_user.contains("feat(a): add a.rs") && merge_user.contains("feat(b): add b.rs"),
            "merge user prompt should contain both partial amendment messages"
        );
    }

    /// A single file whose diff exceeds the budget even after split dispatch.
    ///
    /// Exercises the budget-error path: `generate_amendment_for_commit` →
    /// budget exceeded → `generate_amendment_split` → `pack_file_diffs`
    /// plans hunk-level chunks → but `from_commit_info_partial` loads the
    /// full per-file diff (deduplicates the repeated path) →
    /// Oversized files that can't be split get placeholders and proceed.
    ///
    /// Verifies that files too large for the budget are replaced with
    /// placeholder text indicating the file was omitted, rather than
    /// failing with a "prompt too large" error.
    #[tokio::test]
    async fn amendment_single_oversized_file_gets_placeholder() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_single_oversized_file_repo_view(&dir);
        let hash = "c".repeat(40);

        // The file is too large for the full budget but gets a placeholder.
        // With 50k context, the placeholder is small enough to fit in a
        // single request. Provide a second response in case the system prompt
        // is large enough to trigger split dispatch.
        let (client, _, prompt_handle) = make_small_context_client_with_prompts(vec![
            Ok(valid_amendment_yaml(&hash, "feat(big): add large module")),
            Ok(valid_amendment_yaml(&hash, "feat(big): add large module")),
        ]);

        let result = client
            .generate_amendments_with_options(&repo_view, false)
            .await;

        // Should succeed (either single request or split with placeholder)
        assert!(
            result.is_ok(),
            "expected success with placeholder, got: {result:?}"
        );

        // One request (placeholder makes it fit in single request)
        assert!(
            prompt_handle.request_count() >= 1,
            "expected at least 1 request, got {}",
            prompt_handle.request_count()
        );
    }

    /// A two-chunk split where the second chunk's AI request fails.
    ///
    /// Exercises the error-propagation path within `generate_amendment_split`:
    /// chunk 1 succeeds → chunk 2 returns `Err` → the `?` operator in the
    /// loop body propagates the error immediately, skipping the merge pass.
    ///
    /// Verifies that exactly 2 requests are recorded (no further processing)
    /// and the overall result is `Err` (no silent degradation).
    #[tokio::test]
    async fn amendment_chunk_failure_stops_dispatch() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_large_diff_repo_view(&dir);
        let hash = "a".repeat(40);

        // First chunk succeeds, second chunk fails
        let (client, _, prompt_handle) = make_small_context_client_with_prompts(vec![
            Ok(valid_amendment_yaml(&hash, "feat(a): add a.rs")),
            Err(anyhow::anyhow!("rate limit exceeded")),
        ]);

        let result = client
            .generate_amendments_with_options(&repo_view, false)
            .await;

        assert!(result.is_err());

        // Exactly 2 requests: chunk 1 (success) + chunk 2 (failure)
        let prompts = prompt_handle.prompts();
        assert_eq!(
            prompts.len(),
            2,
            "should stop after the failing chunk, got {} requests",
            prompts.len()
        );

        // The first request should reference one of the files
        let (_, first_user) = &prompts[0];
        assert!(
            first_user.contains("src/a.rs") || first_user.contains("src/b.rs"),
            "first chunk prompt should reference a file"
        );
    }

    /// Two-chunk amendment split dispatch, focused on the reduce pass inputs.
    ///
    /// Exercises `merge_amendment_chunks` which calls
    /// `generate_chunk_merge_user_prompt` to assemble the merge prompt from:
    /// the commit hash, original message, diff_summary, and the partial
    /// amendment messages returned by each chunk.
    ///
    /// Verifies that the merge (3rd) request's user prompt contains all of:
    /// both partial messages, the original commit message, the diff_summary
    /// file paths, and the commit hash.
    #[tokio::test]
    async fn amendment_reduce_pass_prompt_content() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_large_diff_repo_view(&dir);
        let hash = "a".repeat(40);

        let (client, _, prompt_handle) = make_small_context_client_with_prompts(vec![
            Ok(valid_amendment_yaml(
                &hash,
                "feat(a): add module a implementation",
            )),
            Ok(valid_amendment_yaml(
                &hash,
                "feat(b): add module b implementation",
            )),
            Ok(valid_amendment_yaml(
                &hash,
                "feat(test): add modules a and b",
            )),
        ]);

        let result = client
            .generate_amendments_with_options(&repo_view, false)
            .await;

        assert!(result.is_ok());

        let prompts = prompt_handle.prompts();
        assert_eq!(prompts.len(), 3);

        // The merge pass is the last (3rd) request
        let (merge_system, merge_user) = &prompts[2];

        // System prompt should be the amendment chunk merge prompt
        assert!(
            merge_system.contains("synthesiz"),
            "merge system prompt should contain synthesis instructions"
        );

        // User prompt should contain the partial messages from chunks
        assert!(
            merge_user.contains("feat(a): add module a implementation"),
            "merge user prompt should contain chunk 1's partial message"
        );
        assert!(
            merge_user.contains("feat(b): add module b implementation"),
            "merge user prompt should contain chunk 2's partial message"
        );

        // User prompt should contain the original commit message
        assert!(
            merge_user.contains("feat(test): large commit"),
            "merge user prompt should contain the original commit message"
        );

        // User prompt should contain the diff_summary referencing both files
        assert!(
            merge_user.contains("src/a.rs") && merge_user.contains("src/b.rs"),
            "merge user prompt should contain the diff_summary"
        );

        // User prompt should reference the commit hash
        assert!(
            merge_user.contains(&hash),
            "merge user prompt should reference the commit hash"
        );
    }

    /// Two-chunk check split dispatch with issue deduplication and merge.
    ///
    /// Exercises `check_commit_split` which:
    /// 1. Dispatches 2 chunk requests (one per file)
    /// 2. Collects issues from both chunks into a `HashSet` keyed by
    ///    `(rule, severity, section)` — duplicates are dropped
    /// 3. Detects that both chunks have suggestions → calls
    ///    `merge_check_chunks` for the AI reduce pass
    ///
    /// Chunk 1 reports: `error:imperative-mood:Subject Line` +
    ///                   `warning:body-required:Content`
    /// Chunk 2 reports: `error:imperative-mood:Subject Line` (duplicate) +
    ///                   `info:scope-suggestion:Style` (new)
    ///
    /// Verifies: 3 unique issues after dedup, suggestion from merge pass,
    /// and the merge prompt contains both partial suggestions + diff_summary.
    #[tokio::test]
    async fn check_split_dedup_and_merge_prompt() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_large_diff_repo_view(&dir);
        let hash = "a".repeat(40);

        // Chunk 1: error (imperative-mood) + warning (body-required) + suggestion
        let chunk1_yaml = format!(
            concat!(
                "checks:\n",
                "  - commit: \"{hash}\"\n",
                "    passes: false\n",
                "    issues:\n",
                "      - severity: error\n",
                "        section: \"Subject Line\"\n",
                "        rule: \"imperative-mood\"\n",
                "        explanation: \"Subject uses past tense\"\n",
                "      - severity: warning\n",
                "        section: \"Content\"\n",
                "        rule: \"body-required\"\n",
                "        explanation: \"Large change needs body\"\n",
                "    suggestion:\n",
                "      message: \"feat(a): shorter subject for a\"\n",
                "      explanation: \"Shortened subject for file a\"\n",
                "    summary: \"Adds module a\"\n",
            ),
            hash = hash,
        );

        // Chunk 2: same error (different explanation) + new info issue + suggestion
        let chunk2_yaml = format!(
            concat!(
                "checks:\n",
                "  - commit: \"{hash}\"\n",
                "    passes: false\n",
                "    issues:\n",
                "      - severity: error\n",
                "        section: \"Subject Line\"\n",
                "        rule: \"imperative-mood\"\n",
                "        explanation: \"Subject line is way too long\"\n",
                "      - severity: info\n",
                "        section: \"Style\"\n",
                "        rule: \"scope-suggestion\"\n",
                "        explanation: \"Consider more specific scope\"\n",
                "    suggestion:\n",
                "      message: \"feat(b): shorter subject for b\"\n",
                "      explanation: \"Shortened subject for file b\"\n",
                "    summary: \"Adds module b\"\n",
            ),
            hash = hash,
        );

        // Merge pass (called because suggestions exist)
        let merge_yaml = format!(
            concat!(
                "checks:\n",
                "  - commit: \"{hash}\"\n",
                "    passes: false\n",
                "    issues: []\n",
                "    suggestion:\n",
                "      message: \"feat(test): add modules a and b\"\n",
                "      explanation: \"Combined suggestion\"\n",
                "    summary: \"Adds modules a and b\"\n",
            ),
            hash = hash,
        );

        let (client, response_handle, prompt_handle) =
            make_small_context_client_with_prompts(vec![
                Ok(chunk1_yaml),
                Ok(chunk2_yaml),
                Ok(merge_yaml),
            ]);

        let result = client
            .check_commits_with_scopes(&repo_view, None, &[], true)
            .await;

        assert!(result.is_ok(), "split dispatch failed: {:?}", result.err());
        let report = result.unwrap();
        assert_eq!(report.commits.len(), 1);
        assert!(!report.commits[0].passes);
        assert_eq!(response_handle.remaining(), 0);

        // Dedup: 3 unique (rule, severity, section) tuples
        //  - imperative-mood / error / Subject Line   (appears in both → deduped)
        //  - body-required    / warning / Content
        //  - scope-suggestion / info / Style
        assert_eq!(
            report.commits[0].issues.len(),
            3,
            "expected 3 unique issues after dedup, got {:?}",
            report.commits[0]
                .issues
                .iter()
                .map(|i| &i.rule)
                .collect::<Vec<_>>()
        );

        // Suggestion should come from the merge pass
        assert!(report.commits[0].suggestion.is_some());
        assert!(
            report.commits[0]
                .suggestion
                .as_ref()
                .unwrap()
                .message
                .contains("add modules a and b"),
            "suggestion should come from the merge pass"
        );

        // Prompt content assertions
        let prompts = prompt_handle.prompts();
        assert_eq!(prompts.len(), 3, "expected 2 chunks + 1 merge");

        // Chunk prompts should collectively cover both files
        let (_, chunk1_user) = &prompts[0];
        let (_, chunk2_user) = &prompts[1];
        let combined_chunk_prompts = format!("{chunk1_user}{chunk2_user}");
        assert!(
            combined_chunk_prompts.contains("src/a.rs")
                && combined_chunk_prompts.contains("src/b.rs"),
            "chunk prompts should collectively cover both files"
        );

        // Merge pass prompt should contain partial suggestions
        let (merge_sys, merge_user) = &prompts[2];
        assert!(
            merge_sys.contains("synthesiz") || merge_sys.contains("reviewer"),
            "merge system prompt should be the check chunk merge prompt"
        );
        assert!(
            merge_user.contains("feat(a): shorter subject for a")
                && merge_user.contains("feat(b): shorter subject for b"),
            "merge user prompt should contain both partial suggestions"
        );
        // Merge prompt should contain the diff_summary
        assert!(
            merge_user.contains("src/a.rs") && merge_user.contains("src/b.rs"),
            "merge user prompt should contain the diff_summary"
        );
    }

    // ── Amendment retry tests ──────────────────────────────────────────

    #[tokio::test]
    async fn amendment_retry_parse_failure_then_success() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_test_repo_view(&dir);
        let hash = format!("{:0>40}", 0);

        let (client, response_handle, prompt_handle) = make_configurable_client_with_prompts(vec![
            Ok("not valid yaml {{[".to_string()),
            Ok(valid_amendment_yaml(&hash, "feat(test): improved")),
        ]);

        let result = client
            .generate_amendments_with_options(&repo_view, false)
            .await;

        assert!(
            result.is_ok(),
            "should succeed after retry: {:?}",
            result.err()
        );
        assert_eq!(result.unwrap().amendments.len(), 1);
        assert_eq!(response_handle.remaining(), 0, "both responses consumed");
        assert_eq!(prompt_handle.request_count(), 2, "exactly 2 AI requests");
    }

    #[tokio::test]
    async fn amendment_retry_request_failure_then_success() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_test_repo_view(&dir);
        let hash = format!("{:0>40}", 0);

        let (client, response_handle, prompt_handle) = make_configurable_client_with_prompts(vec![
            Err(anyhow::anyhow!("rate limit")),
            Ok(valid_amendment_yaml(&hash, "feat(test): improved")),
        ]);

        let result = client
            .generate_amendments_with_options(&repo_view, false)
            .await;

        assert!(
            result.is_ok(),
            "should succeed after retry: {:?}",
            result.err()
        );
        assert_eq!(result.unwrap().amendments.len(), 1);
        assert_eq!(response_handle.remaining(), 0);
        assert_eq!(prompt_handle.request_count(), 2);
    }

    #[tokio::test]
    async fn amendment_retry_all_attempts_exhausted() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_test_repo_view(&dir);

        let (client, response_handle, prompt_handle) = make_configurable_client_with_prompts(vec![
            Ok("bad yaml 1".to_string()),
            Ok("bad yaml 2".to_string()),
            Ok("bad yaml 3".to_string()),
        ]);

        let result = client
            .generate_amendments_with_options(&repo_view, false)
            .await;

        assert!(result.is_err(), "should fail after all retries exhausted");
        assert_eq!(response_handle.remaining(), 0, "all 3 responses consumed");
        assert_eq!(
            prompt_handle.request_count(),
            3,
            "exactly 3 AI requests (1 + 2 retries)"
        );
    }

    #[tokio::test]
    async fn amendment_retry_success_first_attempt() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_test_repo_view(&dir);
        let hash = format!("{:0>40}", 0);

        let (client, response_handle, prompt_handle) =
            make_configurable_client_with_prompts(vec![Ok(valid_amendment_yaml(
                &hash,
                "feat(test): works first time",
            ))]);

        let result = client
            .generate_amendments_with_options(&repo_view, false)
            .await;

        assert!(result.is_ok());
        assert_eq!(response_handle.remaining(), 0);
        assert_eq!(prompt_handle.request_count(), 1, "only 1 request, no retry");
    }

    #[tokio::test]
    async fn amendment_retry_mixed_request_and_parse_failures() {
        let dir = tempfile::tempdir().unwrap();
        let repo_view = make_test_repo_view(&dir);
        let hash = format!("{:0>40}", 0);

        let (client, response_handle, prompt_handle) = make_configurable_client_with_prompts(vec![
            Err(anyhow::anyhow!("network error")),
            Ok("invalid yaml {{".to_string()),
            Ok(valid_amendment_yaml(&hash, "feat(test): third time")),
        ]);

        let result = client
            .generate_amendments_with_options(&repo_view, false)
            .await;

        assert!(
            result.is_ok(),
            "should succeed on third attempt: {:?}",
            result.err()
        );
        assert_eq!(result.unwrap().amendments.len(), 1);
        assert_eq!(response_handle.remaining(), 0);
        assert_eq!(prompt_handle.request_count(), 3, "all 3 attempts used");
    }

    // ── create_default_claude_client factory ───────────────────────
    //
    // Backend dispatch is the env-parsing boundary; tests inject a pure
    // `MapEnv` into `create_default_claude_client_with` rather than mutating
    // the process environment, so they need no lock and run in parallel.

    use crate::test_support::env::MapEnv;

    #[tokio::test]
    async fn factory_claude_cli_backend_dispatches_to_claude_cli_client() {
        let env = MapEnv::new().with("OMNI_DEV_AI_BACKEND", "claude-cli");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("factory should succeed");
        let metadata = client.get_ai_client_metadata();
        assert_eq!(metadata.provider, "Claude CLI");
        // Default model falls through to the registry's claude default.
        assert_eq!(metadata.model, "claude-sonnet-5");
    }

    #[tokio::test]
    async fn factory_claude_cli_backend_honours_model_precedence() {
        // CLAUDE_MODEL has higher precedence than CLAUDE_CODE_MODEL.
        let env = MapEnv::new()
            .with("OMNI_DEV_AI_BACKEND", "claude-cli")
            .with("CLAUDE_CODE_MODEL", "opus")
            .with("CLAUDE_MODEL", "haiku");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("factory should succeed");
        let metadata = client.get_ai_client_metadata();
        assert_eq!(metadata.provider, "Claude CLI");
        assert_eq!(metadata.model, "haiku");
    }

    #[tokio::test]
    async fn factory_claude_cli_backend_explicit_model_wins_over_env() {
        let env = MapEnv::new()
            .with("OMNI_DEV_AI_BACKEND", "claude-cli")
            .with("CLAUDE_MODEL", "haiku");

        let client = create_default_claude_client_with(&env, Some("opus".to_string()), None)
            .await
            .expect("factory should succeed");
        let metadata = client.get_ai_client_metadata();
        assert_eq!(metadata.model, "opus");
    }

    #[tokio::test]
    async fn factory_claude_cli_backend_accepts_underscore_alias() {
        let env = MapEnv::new().with("OMNI_DEV_AI_BACKEND", "claude_cli");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("factory should succeed");
        let metadata = client.get_ai_client_metadata();
        assert_eq!(metadata.provider, "Claude CLI");
    }

    #[tokio::test]
    async fn factory_claude_cli_backend_ignores_beta_header_without_validation() {
        // The claude-cli arm warns and drops the beta header instead of
        // validating it — the model may be a short alias ("sonnet") the
        // registry does not know, and even an unsupported header/model pair
        // must not error on this backend.
        let env = MapEnv::new()
            .with("OMNI_DEV_AI_BACKEND", "claude-cli")
            .with("CLAUDE_MODEL", "sonnet")
            .with("OMNI_DEV_BETA_HEADER", "anthropic-beta:not-a-real-beta");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("claude-cli must ignore the beta header, not validate it");
        let metadata = client.get_ai_client_metadata();
        assert_eq!(metadata.provider, "Claude CLI");
        assert_eq!(metadata.model, "sonnet");
        assert_eq!(
            metadata.active_beta, None,
            "beta header must not be forwarded"
        );
    }

    #[tokio::test]
    async fn factory_ollama_branch_probes_loaded_context_length() {
        use wiremock::matchers::{method, path};
        use wiremock::{Mock, MockServer, ResponseTemplate};

        // Activate an INFO subscriber for this (current-thread) test so the
        // probe's success `info!` event fires in-process and its fields are
        // evaluated. Discard the formatted output — we assert on the probed
        // value below, not the log line.
        let _log_guard = tracing::subscriber::set_default(
            tracing_subscriber::fmt()
                .with_max_level(tracing::Level::INFO)
                .with_writer(std::io::sink)
                .finish(),
        );

        let server = MockServer::start().await;
        Mock::given(method("GET"))
            .and(path("/api/v0/models"))
            .respond_with(ResponseTemplate::new(200).set_body_json(serde_json::json!({
                "data": [
                    { "id": "lm-loaded", "state": "loaded", "loaded_context_length": 6144_u64 }
                ]
            })))
            .mount(&server)
            .await;

        let env = MapEnv::new()
            .with("USE_OLLAMA", "true")
            .with("OLLAMA_BASE_URL", &server.uri())
            .with("OLLAMA_MODEL", "lm-loaded");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("factory should succeed");
        let metadata = client.get_ai_client_metadata();
        assert_eq!(metadata.provider, "Ollama");
        assert_eq!(metadata.model, "lm-loaded");
        // The probed value (6144) overrides the registry/default.
        assert_eq!(metadata.max_context_length, 6144);
    }

    #[tokio::test]
    async fn factory_ollama_branch_falls_back_when_probe_fails() {
        use wiremock::matchers::{method, path};
        use wiremock::{Mock, MockServer, ResponseTemplate};

        let server = MockServer::start().await;
        Mock::given(method("GET"))
            .and(path("/api/v0/models"))
            .respond_with(ResponseTemplate::new(500))
            .mount(&server)
            .await;
        Mock::given(method("POST"))
            .and(path("/api/show"))
            .respond_with(ResponseTemplate::new(500))
            .mount(&server)
            .await;

        let env = MapEnv::new()
            .with("USE_OLLAMA", "true")
            .with("OLLAMA_BASE_URL", &server.uri())
            .with("OLLAMA_MODEL", "no-such-model");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("factory should succeed");
        let metadata = client.get_ai_client_metadata();
        // Probe failure → fall back to the registry estimate (which
        // resolves to FALLBACK_INPUT_CONTEXT for unknown models).
        let registry_value =
            crate::claude::model_config::get_model_registry().get_input_context("no-such-model");
        assert_eq!(metadata.max_context_length, registry_value);
    }

    /// LM Studio path is tested above. This complements it by exercising
    /// the Ollama-native fallthrough through the factory, so the
    /// info-log arm fires for both `ProbeSource` variants.
    #[tokio::test]
    async fn factory_ollama_branch_probes_via_ollama_native() {
        use wiremock::matchers::{method, path};
        use wiremock::{Mock, MockServer, ResponseTemplate};

        let server = MockServer::start().await;
        Mock::given(method("GET"))
            .and(path("/api/v0/models"))
            .respond_with(ResponseTemplate::new(404))
            .mount(&server)
            .await;
        Mock::given(method("POST"))
            .and(path("/api/show"))
            .respond_with(ResponseTemplate::new(200).set_body_json(serde_json::json!({
                "model_info": { "llama.context_length": 12288_u64 }
            })))
            .mount(&server)
            .await;

        let env = MapEnv::new()
            .with("USE_OLLAMA", "true")
            .with("OLLAMA_BASE_URL", &server.uri())
            .with("OLLAMA_MODEL", "ollama-native-model");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("factory should succeed");
        let metadata = client.get_ai_client_metadata();
        assert_eq!(metadata.max_context_length, 12288);
    }

    // The OpenAI / Bedrock / default-Claude branches construct their clients
    // synchronously (no network probe, unlike Ollama), so a pure MapEnv drives
    // them with no env mutation. These cover the non-Ollama dispatch arms.

    #[tokio::test]
    async fn factory_openai_branch_builds_client() {
        let env = MapEnv::new()
            .with("USE_OPENAI", "true")
            .with("OPENAI_MODEL", "gpt-4.1")
            .with("OPENAI_API_KEY", "sk-test");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("factory should succeed");
        assert_eq!(client.get_ai_client_metadata().model, "gpt-4.1");
    }

    #[tokio::test]
    async fn factory_openai_branch_errors_without_api_key() {
        let env = MapEnv::new().with("USE_OPENAI", "true");
        let result = create_default_claude_client_with(&env, None, None).await;
        assert!(result.is_err());
    }

    #[tokio::test]
    async fn factory_bedrock_branch_builds_client() {
        let env = MapEnv::new()
            .with("CLAUDE_CODE_USE_BEDROCK", "true")
            .with("ANTHROPIC_MODEL", "claude-sonnet-4-6")
            .with("ANTHROPIC_AUTH_TOKEN", "tok")
            .with("ANTHROPIC_BEDROCK_BASE_URL", "https://bedrock.example.com");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("factory should succeed");
        assert_eq!(client.get_ai_client_metadata().model, "claude-sonnet-4-6");
    }

    #[tokio::test]
    async fn factory_bedrock_branch_errors_without_auth_token() {
        let env = MapEnv::new().with("CLAUDE_CODE_USE_BEDROCK", "true");
        let result = create_default_claude_client_with(&env, None, None).await;
        assert!(result.is_err());
    }

    #[tokio::test]
    async fn factory_bedrock_branch_errors_without_base_url() {
        let env = MapEnv::new()
            .with("CLAUDE_CODE_USE_BEDROCK", "true")
            .with("ANTHROPIC_AUTH_TOKEN", "tok");
        let result = create_default_claude_client_with(&env, None, None).await;
        assert!(result.is_err());
    }

    #[tokio::test]
    async fn factory_default_claude_branch_builds_client() {
        let env = MapEnv::new()
            .with("ANTHROPIC_MODEL", "claude-opus-4-6")
            .with("CLAUDE_API_KEY", "sk-test");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("factory should succeed");
        assert_eq!(client.get_ai_client_metadata().model, "claude-opus-4-6");
    }

    #[tokio::test]
    async fn factory_default_claude_branch_errors_without_api_key() {
        let result = create_default_claude_client_with(&MapEnv::new(), None, None).await;
        assert!(result.is_err());
    }

    #[tokio::test]
    async fn factory_default_claude_branch_errors_on_conflicting_key_and_file() {
        // CLAUDE_API_KEY and its _FILE companion both set in the same layer is
        // a secret_env::SecretEnvError::Conflict (issue #2006), a distinct
        // failure from the plain-unset ApiKeyNotFound case above — exercising
        // the `?` that propagates it out of secret_var_any in this arm.
        let key_file = crate::utils::secret_env::file_var_name("CLAUDE_API_KEY");
        let env = MapEnv::new()
            .with("CLAUDE_API_KEY", "sk-test")
            .with(&key_file, "/tmp/should-not-be-read");
        let result = create_default_claude_client_with(&env, None, None).await;
        assert!(result.is_err());
    }

    // ── shared-resolver behaviors introduced by #1118 ───────────────

    #[tokio::test]
    async fn factory_default_claude_branch_honours_claude_model_chain() {
        // The headline #1118 bug: CLAUDE_MODEL / CLAUDE_CODE_MODEL were
        // silently ignored by the direct-API branch.
        let env = MapEnv::new()
            .with("CLAUDE_MODEL", "claude-opus-4-6")
            .with("ANTHROPIC_MODEL", "claude-sonnet-4-6")
            .with("CLAUDE_API_KEY", "sk-test");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("factory should succeed");
        assert_eq!(client.get_ai_client_metadata().model, "claude-opus-4-6");
    }

    #[tokio::test]
    async fn factory_bedrock_branch_honours_claude_model_chain() {
        let env = MapEnv::new()
            .with("CLAUDE_CODE_USE_BEDROCK", "true")
            .with("CLAUDE_CODE_MODEL", "claude-opus-4-6")
            .with("ANTHROPIC_AUTH_TOKEN", "tok")
            .with("ANTHROPIC_BEDROCK_BASE_URL", "https://bedrock.example.com");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("factory should succeed");
        assert_eq!(client.get_ai_client_metadata().model, "claude-opus-4-6");
    }

    #[tokio::test]
    async fn factory_omni_dev_model_beats_provider_var() {
        let env = MapEnv::new()
            .with("USE_OPENAI", "true")
            .with("OMNI_DEV_MODEL", "gpt-4.1")
            .with("OPENAI_MODEL", "gpt-5-mini")
            .with("OPENAI_API_KEY", "sk-test");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("factory should succeed");
        assert_eq!(client.get_ai_client_metadata().model, "gpt-4.1");
    }

    #[tokio::test]
    async fn factory_backend_env_var_beats_legacy_use_flags() {
        // OMNI_DEV_AI_BACKEND=openai wins even though USE_OLLAMA is set.
        let env = MapEnv::new()
            .with("OMNI_DEV_AI_BACKEND", "openai")
            .with("USE_OLLAMA", "true")
            .with("OPENAI_API_KEY", "sk-test");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("factory should succeed");
        assert_eq!(client.get_ai_client_metadata().model, "gpt-5-mini");
    }

    #[tokio::test]
    async fn factory_backend_default_value_forces_direct_api() {
        // `--ai-backend default` propagates as OMNI_DEV_AI_BACKEND=default and
        // must override the USE_* soup, dispatching the direct Claude client.
        let env = MapEnv::new()
            .with("OMNI_DEV_AI_BACKEND", "default")
            .with("USE_OLLAMA", "true")
            .with("CLAUDE_API_KEY", "sk-test");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("factory should succeed");
        assert_eq!(client.get_ai_client_metadata().model, "claude-sonnet-5");
    }

    #[tokio::test]
    async fn factory_unknown_backend_value_is_hard_error() {
        let env = MapEnv::new()
            .with("OMNI_DEV_AI_BACKEND", "junk")
            .with("CLAUDE_API_KEY", "sk-test");

        let err = create_default_claude_client_with(&env, None, None)
            .await
            .map(|_| ())
            .expect_err("unknown backend must error");
        assert!(format!("{err:#}").contains("junk"));
    }

    #[tokio::test]
    async fn factory_beta_header_env_var_is_applied() {
        let env = MapEnv::new()
            .with("ANTHROPIC_MODEL", "claude-3-7-sonnet-20250219")
            .with(
                "OMNI_DEV_BETA_HEADER",
                "anthropic-beta:output-128k-2025-02-19",
            )
            .with("CLAUDE_API_KEY", "sk-test");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("factory should accept a registry-supported beta header");
        // The beta header raises the model's max output tokens.
        assert_eq!(client.get_ai_client_metadata().max_response_length, 128_000);
    }

    #[tokio::test]
    async fn factory_beta_header_env_var_malformed_is_hard_error() {
        let env = MapEnv::new()
            .with("OMNI_DEV_BETA_HEADER", "no-colon-here")
            .with("CLAUDE_API_KEY", "sk-test");

        let err = create_default_claude_client_with(&env, None, None)
            .await
            .map(|_| ())
            .expect_err("malformed beta header must error");
        assert!(format!("{err:#}").contains("no-colon-here"));
    }

    /// `--beta-header` is Anthropic-specific; the OpenAI backend must warn and
    /// ignore it rather than validate it against the Claude registry (which
    /// would hard-fail for a header that is never sent — #1119).
    #[tokio::test]
    async fn factory_openai_backend_ignores_beta_header_without_validation() {
        let env = MapEnv::new()
            .with("OMNI_DEV_AI_BACKEND", "openai")
            .with("OPENAI_API_KEY", "sk-openai")
            .with("OPENAI_MODEL", "gpt-4o")
            .with("OMNI_DEV_BETA_HEADER", "anthropic-beta:not-a-real-beta");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("openai backend must ignore --beta-header, not validate it");
        assert_eq!(
            client.get_ai_client_metadata().active_beta,
            None,
            "beta header must not reach the OpenAI client"
        );
    }

    /// Ollama mirrors the OpenAI arm: warn and ignore, never validate. Points
    /// at a closed port so the context-length probe fails fast and non-fatally.
    #[tokio::test]
    async fn factory_ollama_backend_ignores_beta_header_without_validation() {
        let env = MapEnv::new()
            .with("OMNI_DEV_AI_BACKEND", "ollama")
            .with("OLLAMA_BASE_URL", "http://127.0.0.1:1")
            .with("OLLAMA_MODEL", "llama2")
            .with("OMNI_DEV_BETA_HEADER", "anthropic-beta:not-a-real-beta");

        let client = create_default_claude_client_with(&env, None, None)
            .await
            .expect("ollama backend must ignore --beta-header, not validate it");
        assert_eq!(
            client.get_ai_client_metadata().active_beta,
            None,
            "beta header must not reach the Ollama client"
        );
    }
}