drep/analysis/

code_quality.rs

//! The code-quality analyzer: render a payload, ask the LLM, turn the
//! response into findings — and never report an unanalyzed file as clean.
//!
//! The analyzer enforces these contracts:
//!
//! 1. No language, no analysis. drep has no opinion on a file type it does
//!    not claim, and silently returning an empty result is the correct
//!    behavior — *not* a failure. The CLI surfaces "no language" by simply
//!    not including the file in the work set.
//! 2. Empty hunks → empty result, no LLM call.
//! 3. Build the payload with `payload::render`. `None` → empty result.
//! 4. Cache first. A hit is parsed exactly as a `Complete` response would
//!    be, and the duplicate is silent — the caller's view is identical.
//! 5. Concurrency. A limiter slot is acquired before the LLM call and held
//!    for the duration. A cache hit must not acquire a slot: the slot
//!    represents in-flight HTTP work, and a cache read is not in-flight.
//! 6. `Ok(Extracted::Complete)` → parse, store in the cache.
//! 7. `Ok(Extracted::Truncated)` → parse the partial result AND mark the
//!    file failed. Never cache a truncated response — caching it makes one
//!    truncation permanent for the whole TTL, and this layer does not know
//!    about `--fail-on` (a caller deciding otherwise would make
//!    `failed_files` depend on a CLI flag, which is the wrong layering).
//! 8. `Err(LlmError::*)` → no findings, file in `failed_files` with the
//!    specific LLM layer that failed.
//!
//! The five rules around out-of-range lines, missing fields, unknown
//! severities, and `issues` itself being absent are at the boundary between
//! "model misreported" and "we could not understand the response". The first
//! is a *finding* we drop; the others are *file-level failures*, because a
//! file we did not fully understand must never be reported clean.

use std::path::Path;

use futures::future::join_all;
use serde_json::Value;

use crate::analysis::findings::{Finding, LlmSeverity};
use crate::analysis::payload;
use crate::analysis::prompt::build_analysis_prompt;
use crate::analysis::response_contract::{
    CATEGORY, COMPILE_FAILURE, ISSUES, LINE, MESSAGE, SEVERITY, SUGGESTION,
};
use crate::analysis::result::{AnalysisResult, FailureReason, ProviderFailure};
use crate::diff::hunks::Hunk;
use crate::diff::hunks::group_by_file;
use crate::languages;
use crate::llm::cache::Cache;
use crate::llm::chain::{ChainError, ProviderChain};
use crate::llm::error::LlmError;
use crate::llm::json_parsing::Extracted;

/// The code-quality analyzer.
///
/// Built once per process. The `chain` and the `cache` are both passed in
/// rather than constructed here: they are process-wide resources. The chain in
/// particular carries the concurrency limiter for each provider *and* the
/// record of which providers have been demoted, so a second chain would both
/// double the in-flight requests against one endpoint and re-discover a dead
/// provider the first one already knew about.
///
/// The model and temperature are **not** duplicated onto this struct. They
/// belong to the provider that ends up answering, which is not known until the
/// chain has run, and a second copy is exactly what lets a request go to one
/// model while the cache key names another.
pub struct CodeQualityAnalyzer {
    pub(crate) chain: ProviderChain,
    pub(crate) cache: Cache,
    cache_only: bool,
}

impl CodeQualityAnalyzer {
    /// Build from a provider chain and a shared cache.
    ///
    /// Infallible: `ProviderChain::new` has already rejected an empty or
    /// misconfigured chain, so there is nothing left for this to validate.
    /// `cache` is a parameter rather than constructed here so the key stays
    /// independent of the `Cache` root (see [`Cache::key`]) and a test can
    /// point it at a `TempDir`.
    pub fn new(chain: ProviderChain, cache: Cache) -> Self {
        Self {
            chain,
            cache,
            cache_only: false,
        }
    }

    /// Select a cache-only pass. The analyzer keeps parsing cached responses
    /// through the normal schema path; only the backend request is forbidden.
    pub fn with_cache_only(mut self, cache_only: bool) -> Self {
        self.cache_only = cache_only;
        self
    }

    /// The provider chain, for a caller reporting which providers served.
    pub fn chain(&self) -> &ProviderChain {
        &self.chain
    }

    /// Analyze one file's hunks.
    ///
    /// Returns an [`AnalysisResult`] populated with whatever the file
    /// produced: findings, a failure marker, or both. The result is never a
    /// bare `Vec<Finding>`; the failure axis is part of the return type so
    /// the caller cannot forget it.
    pub async fn analyze_file(&self, hunks: &[Hunk]) -> AnalysisResult {
        self.analyze_files_in_mode(std::iter::once(hunks), self.cache_only)
            .await
    }

    async fn analyze_file_in_mode(&self, hunks: &[Hunk], cache_only: bool) -> AnalysisResult {
        // Rule 1: no language, no analysis. `languages::detect` on the
        // first hunk's file path is enough because every hunk in `by_file`
        // shares a path (the diff module groups by file).
        let Some(first) = hunks.first() else {
            return AnalysisResult::default();
        };
        let Some(language) = languages::detect(&first.file_path) else {
            return AnalysisResult::default();
        };

        // Rule 3: payload. `render` returns `None` only for an empty slice,
        // which the `hunks.first()` guard above has already excluded - rule 2
        // and rule 3 are the same check, so a second `hunks.is_empty()` here
        // would be unreachable rather than defensive.
        let Some(payload) = payload::render(language, hunks) else {
            return AnalysisResult::default();
        };

        // The size ceiling is enforced on the *rendered payload*, so it holds
        // for every input mode. Checking the file size during paths-mode input
        // resolution - as this used to, and still does as a pre-filter - left
        // `--staged` and `--diff` unguarded, which are the two modes a commit
        // gate actually runs in: a newly-added 5 MB file reached the model
        // whole. Too large is a *failure*, not a skip; a file drep declined to
        // analyze is not clean.
        let rendered = u64::try_from(payload.text.len()).unwrap_or(u64::MAX);
        if rendered > payload::PAYLOAD_MAX_BYTES {
            return AnalysisResult::failed(
                first.file_path.clone(),
                FailureReason::PayloadTooLarge {
                    bytes: rendered,
                    limit: payload::PAYLOAD_MAX_BYTES,
                },
            );
        }

        let system_prompt = build_analysis_prompt(language);

        // Rules 4, 5 and the failover loop all live in the chain: the cache is
        // consulted per provider (a hit costs no concurrency slot, because the
        // slot represents in-flight HTTP work), and the key comes back naming
        // whoever answered. Computing a key here would be computing it for a
        // provider that may not be the one that serves the file.
        let served = if cache_only {
            match self
                .chain
                .cached_json(&system_prompt, &payload.text, &self.cache)
            {
                Some(served) => Ok(served),
                None => {
                    return AnalysisResult::failed(
                        first.file_path.clone(),
                        FailureReason::CacheMiss,
                    );
                }
            }
        } else {
            self.chain
                .complete_json(&system_prompt, &payload.text, &self.cache)
                .await
        };

        match served {
            // Rule 6: complete → parse, store in the cache.
            // Rules 6 and 7 in one arm. Both never-cache rules live in the
            // `if let` below rather than being split across two arms, where
            // the `Complete` arm's guard could only ever be true.
            Ok(served) => {
                let result = parse_response(&payload, &first.file_path, &served.extracted);
                // Cache only a `Complete` response we fully understood. A
                // truncated one is a prefix, and a body can be valid JSON and
                // still schema-invalid - a missing `issues` array, a record
                // with an unknown severity - which yields a file-level
                // failure. Caching either replays it for the whole TTL
                // instead of letting the next run ask again.
                //
                // `served.key`, never a key computed here: the entry must be
                // filed under the model that produced it, or a later run with
                // the head restored gets a hit that never came from the head.
                //
                // The write itself is best-effort: a cache failure is a
                // diagnostic, not a failure of the analysis.
                if let Extracted::Complete(value) = &served.extracted
                    && !served.from_cache
                    && result.failed_files.is_empty()
                {
                    let _ = self.cache.put(&served.key, value);
                }
                result
            }
            // Rule 8: no provider produced an answer → no findings, file in
            // `failed_files` with every provider's reason. The detail is kept
            // rather than discarded, so the CLI can render a line the user can
            // act on.
            Err(err) => AnalysisResult::failed(first.file_path.clone(), chain_failure_reason(err)),
        }
    }

    /// Analyze many files concurrently, bounded by the limiter.
    ///
    /// Each entry of `by_file` is one file's hunks; the limiter bounds the
    /// in-flight requests, so we spawn them all and let it queue. The
    /// per-file results are merged with [`AnalysisResult::merge`].
    pub async fn analyze_files(&self, by_file: &[Vec<Hunk>]) -> AnalysisResult {
        self.analyze_files_in_mode(by_file.iter().map(Vec::as_slice), self.cache_only)
            .await
    }

    /// Analyze selected cache misses without rebuilding the provider chain.
    ///
    /// Push-gate mode uses this after its cache-only pass, preserving sticky
    /// demotion and backend diagnostics while avoiding a second pass over
    /// every cache hit in the diff.
    pub async fn analyze_files_live(&self, by_file: &[&[Hunk]]) -> AnalysisResult {
        self.analyze_files_in_mode(by_file.iter().copied(), false)
            .await
    }

    async fn analyze_files_in_mode<'a>(
        &self,
        by_file: impl IntoIterator<Item = &'a [Hunk]>,
        cache_only: bool,
    ) -> AnalysisResult {
        let groups: Vec<HunkGroup<'a>> = by_file.into_iter().flat_map(partition_hunks).collect();
        let futures = groups
            .iter()
            .map(|group| self.analyze_file_in_mode(group.as_slice(), cache_only));
        let results = join_all(futures).await;
        let mut merged = AnalysisResult::default();
        for result in results {
            merged.merge(result);
        }
        merged
    }
}

/// A correctly grouped borrowed slice, or owned groups recovered from a mixed one.
enum HunkGroup<'a> {
    Borrowed(&'a [Hunk]),
    Owned(Vec<Hunk>),
}

impl HunkGroup<'_> {
    fn as_slice(&self) -> &[Hunk] {
        match self {
            Self::Borrowed(hunks) => hunks,
            Self::Owned(hunks) => hunks,
        }
    }
}

/// Preserve the zero-copy normal path and partition only malformed input.
fn partition_hunks(hunks: &[Hunk]) -> Vec<HunkGroup<'_>> {
    let mixed = hunks
        .first()
        .is_some_and(|first| hunks.iter().any(|hunk| hunk.file_path != first.file_path));
    if !mixed {
        return vec![HunkGroup::Borrowed(hunks)];
    }
    group_by_file(hunks.iter().cloned())
        .into_iter()
        .map(HunkGroup::Owned)
        .collect()
}

/// Map an `LlmError` to the failure reason the caller carries in
/// `AnalysisResult::failed_files`.
///
/// Distinct from the parsing-path reasons because the LLM layer's failure
/// modes are a different axis. HTTP failures preserve a numeric status and
/// process backends preserve a stable typed kind, so callers never have to
/// recover policy from human-readable messages.
pub(crate) fn into_failure_reason(err: LlmError) -> FailureReason {
    match err {
        LlmError::Transport { status, message } => FailureReason::Transport { status, message },
        LlmError::Unparseable(message) => FailureReason::Unparseable(message),
        LlmError::ModelStopped { finish, message } => {
            FailureReason::ModelStopped { finish, message }
        }
        // `NotConfigured` is a configuration failure at the LLM boundary —
        // not a connectivity failure, but indistinguishable from one to the
        // gate, which only cares whether the file was analyzed. Mapping to
        // `Transport { status: None }` keeps the exit code 2 path uniform
        // without inventing a new variant the JSON output would have to
        // distinguish.
        LlmError::NotConfigured(message) => FailureReason::Transport {
            status: None,
            message,
        },
        LlmError::Backend { kind, message } => FailureReason::Backend { kind, message },
    }
}

/// Map a whole-chain failure to the reason the caller carries.
///
/// **A one-provider chain collapses to that provider's own reason.** That keeps
/// a single-provider config - what `drep init` writes - reporting exactly what
/// it reported before failover existed, down to the JSON `kind`.
///
/// The trigger is the chain's *length*, not the number of providers that
/// failed. Those differ precisely where it matters: a two-provider chain
/// stopped at the head by a 401 produces one attempt, and collapsing it would
/// discard the provider index and the model name just as the user is asking
/// "I configured a fallback - why didn't it run?".
fn chain_failure_reason(err: ChainError) -> FailureReason {
    let mut attempts = err.attempts;
    if err.chain_len == 1
        && let Some(only) = attempts.pop()
    {
        // Move the attempt out so the error string is not cloned on the
        // overwhelmingly common one-provider path.
        return into_failure_reason(only.error);
    }
    FailureReason::ChainFailed(
        attempts
            .into_iter()
            .map(|attempt| ProviderFailure {
                provider: attempt.provider,
                model: attempt.model,
                reason: into_failure_reason(attempt.error),
                skipped: attempt.skipped,
            })
            .collect(),
    )
}

/// Turn an `Extracted` value into an [`AnalysisResult`].
///
/// A free function, not a method: it reads no analyzer state, and keeping it
/// free means the whole parsing core is testable without a `MockServer`, a
/// `Cache` and a `TempDir`.
///
/// The truncation flag is **read off the discriminant**, never passed in
/// beside it. An earlier shape took `extracted` *and* a `truncated: bool`,
/// which let `(Extracted::Truncated(v), false)` compile and report a
/// truncated file as clean - the single outcome this module exists to
/// prevent, resting on four call sites agreeing by convention.
fn parse_response(
    payload: &payload::Payload,
    file_path: &Path,
    extracted: &Extracted,
) -> AnalysisResult {
    let mut result = AnalysisResult::default();

    let (value, truncated) = match extracted {
        Extracted::Complete(value) => (value, false),
        // Rule 7: a truncated response is a prefix of what the model meant,
        // so the file is unanalyzed however good the partial findings look.
        Extracted::Truncated(value) => (value, true),
    };

    // The response shape is `{"issues": [...], "summary": "..."}`. Anything
    // else is malformed, and the whole file is unanalyzed.
    let Some(issues) = value.get(ISSUES).and_then(Value::as_array) else {
        // Truncation wins over "no `issues` array": a response cut off before
        // it reached `issues` has no array *because* it was truncated, and
        // reporting that as a malformed record hides the real cause.
        let reason = if truncated {
            FailureReason::Truncated
        } else {
            FailureReason::MalformedFinding("response has no `issues` array".to_owned())
        };
        result.failed_files.insert(file_path.to_path_buf(), reason);
        return result;
    };

    // Every finding carries the same path string. Allocate it only once the
    // response has an issues array; the missing-array failure above does not
    // need it.
    let path_string = file_path.to_string_lossy().into_owned();
    let mut failure = truncated.then_some(FailureReason::Truncated);

    // `issues: []` is a legitimate clean result: empty findings, no failure.
    for issue in issues {
        match parse_issue(issue, payload, &path_string) {
            IssueOutcome::Finding(finding) => result.findings.push(finding),
            IssueOutcome::Dropped => result.dropped_out_of_range += 1,
            // Do not return early: a malformed record in the middle of an
            // otherwise-valid array should still let the well-formed records
            // through. The failure class is "we do not fully understand the
            // response", not "every record is wrong".
            // First reason wins, matching `union_failures`: the reasons are not
            // meaningfully combinable, and the last-writer version reported
            // whichever malformed record happened to sit at the end of the
            // array rather than the one that first told us the response was
            // not understood.
            IssueOutcome::Malformed(detail) => {
                failure.get_or_insert(FailureReason::MalformedFinding(detail));
            }
        }
    }

    if let Some(reason) = failure {
        result.failed_files.insert(file_path.to_path_buf(), reason);
    }
    result
}

/// Parse one issue record into a [`Finding`], a drop, or a malformed
/// marker with a reason.
///
/// The three outcomes are deliberately distinct:
///
/// - `Finding`: a valid record attributed to a real line in the
///   payload. The caller adds it to `findings`.
/// - `Dropped`: a valid record whose line was not in
///   `payload.valid_lines`. The caller increments
///   `dropped_out_of_range`. The file is **not** marked failed: we
///   understood the record perfectly, it was simply about code the
///   model was never shown.
/// - `Malformed`: an unparseable record (unknown severity, missing
///   field, non-integer `line`). The caller adds the file to
///   `failed_files`. We cannot know what the record meant, so we
///   cannot trust the rest of the response either.
fn parse_issue(issue: &Value, payload: &payload::Payload, file_path: &str) -> IssueOutcome {
    // `line` must be a positive integer. A missing field, a string,
    // a float, or a non-positive number are all malformed.
    let Some(line) = issue.get(LINE).and_then(Value::as_u64) else {
        return IssueOutcome::Malformed("missing or non-integer `line`".to_owned());
    };
    // `Value::as_u64` already rejects non-integers; the only
    // remaining "not a positive integer" case is zero. A line of
    // zero is a model artifact, not a real line.
    if line == 0 {
        return IssueOutcome::Malformed("`line` is zero".to_owned());
    }
    // A line beyond `u32` is a model artifact of exactly the same class as
    // a line of zero, and this module's whole thesis is that we do not guess.
    // Clamping to `u32::MAX` happened to land in `Dropped` because that value
    // is never in `valid_lines` - correct by accident, via a silent clamp.
    let Ok(line) = u32::try_from(line) else {
        return IssueOutcome::Malformed("`line` is beyond u32".to_owned());
    };

    // `severity` must be one of the levels the prompt asked for. Anything
    // else is malformed: we cannot map it to a `Severity`, and we do not
    // silently coerce. The vocabulary lives on `LlmSeverity` so the prompt
    // and this parser cannot list different levels.
    let Some(severity_str) = issue.get(SEVERITY).and_then(Value::as_str) else {
        return IssueOutcome::Malformed("missing `severity`".to_owned());
    };
    let Ok(severity) = severity_str.parse::<LlmSeverity>() else {
        return IssueOutcome::Malformed(format!("unknown severity `{severity_str}`"));
    };
    let severity = severity.to_severity();

    // `message` is the only remaining required field. Missing or
    // non-string is malformed.
    let Some(message) = issue.get(MESSAGE).and_then(Value::as_str) else {
        return IssueOutcome::Malformed("missing or non-string `message`".to_owned());
    };

    // `category` is optional, but a *present* non-string one is malformed, not
    // a missing one. `.get().and_then(as_str).unwrap_or("unknown")` collapses
    // the two, so `"category": 7` was reported as the finding kind "unknown" -
    // a response we demonstrably did not understand, recorded as understood and
    // then cached for the whole TTL. Same rule `severity` and `message` follow.
    let kind = match issue.get(CATEGORY) {
        None => "unknown".to_owned(),
        Some(Value::String(text)) => text.clone(),
        Some(_) => return IssueOutcome::Malformed("non-string `category`".to_owned()),
    };
    let message = message.to_owned();
    let asserts_compile_failure = match issue.get(COMPILE_FAILURE) {
        None => false,
        Some(Value::Bool(value)) => *value,
        Some(_) => return IssueOutcome::Malformed("non-boolean `compile_failure`".to_owned()),
    };

    // `suggestion` is optional on the same terms. Absent or empty → `None`, not
    // `Some("")`: an empty suggestion is not a suggestion. A present non-string
    // one is malformed rather than silently absent.
    let suggestion = match issue.get(SUGGESTION) {
        None => None,
        Some(Value::String(text)) if text.is_empty() => None,
        Some(Value::String(text)) => Some(text.clone()),
        Some(_) => return IssueOutcome::Malformed("non-string `suggestion`".to_owned()),
    };

    // Membership is checked LAST, after the record's shape.
    //
    // Shape asks "did the model answer in our schema"; membership asks "did it
    // talk about code we sent". A record with an unknown severity is evidence
    // the response's vocabulary is wrong, and that contaminates the records we
    // did accept - so it must fail the file even when the record also cites a
    // line we never sent. Checking membership first would let a demonstrably
    // schema-violating response be reported as fully understood.
    //
    // A well-formed record about code we did not send is different: we
    // understood it perfectly, it is simply out of scope. It is dropped and
    // counted, never clamped onto the nearest valid line, which would attach a
    // real-looking finding to arbitrary code. Pinned by
    // `out_of_range_line_is_dropped_not_clamped`.
    if !payload.valid_lines.contains(&line) {
        return IssueOutcome::Dropped;
    }

    IssueOutcome::Finding(Finding {
        kind,
        severity,
        file_path: file_path.to_owned(),
        line,
        column: None,
        message,
        suggestion,
        asserts_compile_failure,
        fingerprint: None,
    })
}

/// What one issue record became after parsing.
enum IssueOutcome {
    /// A valid record attributed to a real line in the payload.
    Finding(Finding),
    /// A valid record whose line was not in `payload.valid_lines`.
    Dropped,
    /// An unparseable record, with a reason naming what was wrong.
    Malformed(String),
}

#[cfg(test)]
mod partition_tests {
    use super::{HunkGroup, partition_hunks};
    use crate::diff::hunks::Hunk;
    use std::path::PathBuf;

    #[test]
    fn already_grouped_hunks_remain_borrowed() {
        let hunks = [Hunk::whole_file(PathBuf::from("same.rs"), "one\ntwo\n")];
        let groups = partition_hunks(&hunks);
        let [HunkGroup::Borrowed(group)] = groups.as_slice() else {
            panic!("an already-grouped slice must stay on the zero-copy path");
        };
        assert!(std::ptr::eq(group.as_ptr(), hunks.as_ptr()));
    }
}