lean-ctx-server 3.12.1

// SPDX-License-Identifier: Apache-2.0
//! `ctx_compose` — task composer (Phase 2 of the efficiency epic).
//!
//! The biggest agent win is a single "rich per call" tool that returns ranked
//! files *with* inline bodies, replacing the typical search → read → outline →
//! read chain (3-5 calls) with one.
//!
//! lean-ctx already has the building blocks as separate tools; this composes
//! them into one response for a natural-language task:
//!   1. extracted keywords,
//!   2. locally ranked files (admitted BM25),
//!   3. exact match locations (index-backed `ctx_search`),
//!   4. the body of the most relevant symbol, inline.

use std::sync::mpsc;
use std::time::Duration;

use crate::tools::CrpMode;
use lean_ctx::core::tokens::count_tokens;

/// Wall-time budget for the semantic-ranking stage. The exact-match and symbol
/// stages are index-backed and cheap; only semantic ranking can hit a cold
/// `O(corpus)` BM25 build. We never let that block the agent loop: past the
/// budget (4s, tuned for cold-start coverage #902) we return the independently
/// admitted sections. Late ranking output is discarded; no legacy index is used
/// as a fallback. Override via `LEAN_CTX_COMPOSE_BUDGET_MS`.
const DEFAULT_SEMANTIC_BUDGET_MS: u64 = 4000;
/// A one-shot `lean-ctx call` has no agent loop to keep responsive, and its
/// caller asked for this result: discarding a ranking after 4s only wasted it
/// (#2049). Its own watchdog bounds the whole call.
const ONE_SHOT_SEMANTIC_BUDGET_MS: u64 = 120_000;
// A timed-out worker retains its permit until it actually finishes. Repeated
// requests cannot accumulate an unbounded set of fresh corpus scans.
static RANKING_WORKERS: tokio::sync::Semaphore = tokio::sync::Semaphore::const_new(4);

fn semantic_budget() -> Duration {
    let ms = std::env::var("LEAN_CTX_COMPOSE_BUDGET_MS")
        .ok()
        .and_then(|v| v.parse::<u64>().ok())
        .filter(|&v| v > 0)
        .unwrap_or(if lean_ctx::core::conversation::is_one_shot_process() {
            ONE_SHOT_SEMANTIC_BUDGET_MS
        } else {
            DEFAULT_SEMANTIC_BUDGET_MS
        });
    Duration::from_millis(ms)
}

/// Token budget for the inlined symbol bodies. Submodular selection fills it
/// with the most coverage-effective, non-redundant set of symbols.
/// Override via `LEAN_CTX_COMPOSE_SYMBOL_TOKENS`.
const DEFAULT_SYMBOL_BUDGET_TOKENS: usize = 600;

fn symbol_budget_tokens() -> usize {
    std::env::var("LEAN_CTX_COMPOSE_SYMBOL_TOKENS")
        .ok()
        .and_then(|v| v.parse::<usize>().ok())
        .filter(|&v| v > 0)
        .unwrap_or(DEFAULT_SYMBOL_BUDGET_TOKENS)
}

pub(crate) fn kernel_supplement_budget(project_root: &str) -> usize {
    let config = lean_ctx::core::context_kernel::activation::load_config(project_root);
    (symbol_budget_tokens() / 5)
        .min(lean_ctx::core::context_kernel::activation::supplement_budget(&config))
}

/// Words that carry no retrieval signal — dropped from keyword extraction.
const STOPWORDS: &[&str] = &[
    "the",
    "and",
    "for",
    "with",
    "that",
    "this",
    "from",
    "into",
    "how",
    "where",
    "what",
    "does",
    "are",
    "was",
    "use",
    "used",
    "uses",
    "add",
    "all",
    "any",
    "can",
    "get",
    "set",
    "via",
    "out",
    "its",
    "his",
    "her",
    "you",
    "your",
    "our",
    "find",
    "show",
    "list",
    "make",
    "when",
    "then",
    "has",
    "have",
    "had",
    "not",
    "but",
    "see",
    "function",
    "method",
    "class",
    "code",
    "file",
    "files",
    "implement",
    "implementation",
];

/// Extract up to `max` distinct identifier-ish keywords from a task, preserving
/// original case (symbol lookups are case-sensitive) and first-seen order.
fn extract_keywords(task: &str, max: usize) -> Vec<String> {
    let mut seen = std::collections::HashSet::new();
    let mut out = Vec::new();
    for raw in task.split(|c: char| !(c.is_alphanumeric() || c == '_')) {
        if raw.len() < 3 {
            continue;
        }
        if STOPWORDS.contains(&raw.to_ascii_lowercase().as_str()) {
            continue;
        }
        if seen.insert(raw.to_string()) {
            out.push(raw.to_string());
            if out.len() >= max {
                break;
            }
        }
    }
    out
}

/// Order `keywords` from most to least specific using the resident BM25 index's
/// per-token document frequency (how many chunks contain the token). Rarer =
/// more specific = better as the "exact matches" seed. A token absent from the
/// corpus (df 0) sinks to the end — grepping it yields nothing useful.
///
/// Non-blocking and best-effort: if the resident index isn't warm yet we return
/// the keywords in their original first-seen order (the previous behaviour), so
/// this can only improve the seed, never stall the call to build an index.
fn order_by_specificity(keywords: &[String], project_root: &str) -> Vec<String> {
    let Some(index) = resident_index(project_root) else {
        return keywords.to_vec();
    };
    rank_by_doc_freq(keywords, &index.doc_freqs)
}

/// Pure ranking core: choose the exact-match seed keyword.
///
/// The seed feeds a case-sensitive regex grep, so raw rarity is the wrong sort:
/// the rarest task token is often a lowercase prose word (`measurand`) that the
/// index counts case-insensitively but the grep then misses against `Measurand`
/// — 0 hits, worse than before. Instead prefer *code identifiers* (camelCase or
/// snake_case: `GetMaxCurrent`, `CurrentGetter`), which grep straight to code,
/// over acronyms (`OCPP`) and prose (`current`) that also match READMEs. Within
/// each class, rarer (lower document frequency) wins; absent tokens (df 0) sink.
/// Keys are lowercased in `doc_freqs`; a stable sort keeps first-seen order on
/// ties, so a task with no identifiers degrades to the previous rarity order.
fn rank_by_doc_freq(
    keywords: &[String],
    doc_freqs: &std::collections::HashMap<String, usize>,
) -> Vec<String> {
    let df = |kw: &String| match doc_freqs.get(&kw.to_ascii_lowercase()) {
        Some(&n) if n > 0 => n,
        _ => usize::MAX,
    };
    // Class 0 = code identifier (grep-friendly), class 1 = acronym/prose.
    let rank_key = |kw: &String| (u8::from(!is_code_identifier(kw)), df(kw));
    let mut ranked = keywords.to_vec();
    ranked.sort_by_key(rank_key);
    ranked
}

/// True for tokens that read as code identifiers — snake_case (`get_max_current`)
/// or camelCase/PascalCase with an internal capital (`GetMaxCurrent`). A leading
/// capital alone (`Current`) or an all-caps acronym (`OCPP`) does not qualify:
/// those match prose and file boilerplate as readily as code.
fn is_code_identifier(kw: &str) -> bool {
    if kw.contains('_') {
        return true;
    }
    let has_lower = kw.chars().any(|c| c.is_ascii_lowercase());
    let internal_upper = kw.chars().skip(1).any(|c| c.is_ascii_uppercase());
    has_lower && internal_upper
}

/// Fetch the already-resident BM25 index for `project_root` without triggering a
/// build. Returns `None` when nothing is cached yet (cold start).
fn resident_index(
    project_root: &str,
) -> Option<std::sync::Arc<lean_ctx::core::bm25_index::BM25Index>> {
    let cache = crate::tools::ctx_semantic_search::get_thread_cache()?;
    lean_ctx::core::bm25_cache::get_or_background(&cache, std::path::Path::new(project_root))
}

/// Run admitted local ranking under the caller's source authority and deadline.
/// The inherited view closes with the owning request, so a late worker cannot
/// switch to an unprotected policy or role while finishing.
fn ranked_files_budgeted(task: &str, project_root: &str, crp_mode: CrpMode) -> String {
    ranked_files_with_slots(task, project_root, crp_mode, &RANKING_WORKERS)
}

fn ranked_files_with_slots(
    task: &str,
    project_root: &str,
    crp_mode: CrpMode,
    slots: &'static tokio::sync::Semaphore,
) -> String {
    // Admit in the caller's scope: a role that denies ctx_search must not
    // obtain ranked source through compose, whatever the worker inherits.
    let role = lean_ctx::core::roles::active_role();
    if role
        .tools
        .denied
        .iter()
        .any(|denied| denied == "ctx_search")
    {
        return "ERR: source search is not authorized".into();
    }
    let Ok(slot) = slots.try_acquire() else {
        return "(local BM25 ranking busy: retry after active requests finish. Other sections are independently admitted.)".into();
    };
    let (tx, rx) = mpsc::channel::<String>();
    let task_owned = task.to_string();
    let root_owned = project_root.to_string();
    // The resident BM25 cache is thread-local to the request thread. Without
    // it the worker deserialized the whole persisted index on every call
    // (seconds on a mid-size repo), regularly overran the budget and made the
    // composed output depend on timing (focus analysis 2026-10-15).
    let resident_cache = crate::tools::ctx_semantic_search::get_thread_cache();

    lean_ctx::core::task_spine::TaskSpine::spawn_thread(move || {
        if let Some(cache) = resident_cache {
            crate::tools::ctx_semantic_search::set_thread_cache(cache);
        }
        let _slot = slot;
        let ranked = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
            crate::tools::ctx_semantic_search::handle_for_tool(
                "ctx_compose",
                &task_owned,
                &root_owned,
                8,
                crp_mode,
                None,
                None,
                Some("bm25"),
                Some(false),
                Some(false),
            )
        }))
        .unwrap_or_else(|_| {
            tracing::warn!("[ctx_compose: semantic ranking panicked; omitting section]");
            String::new()
        });
        // A timed-out receiver drops the result. This fresh view is never
        // persisted or substituted with a legacy shared index.
        let _ = tx.send(ranked);
    });

    match rx.recv_timeout(semantic_budget()) {
        Ok(ranked) => ranked.trim().to_string(),
        Err(_) => deferred_ranking_note().to_string(),
    }
}

/// Do not replay unadmitted index diagnostics or promise a warmed cache: each
/// local ranking view is freshly admitted, including after a timeout.
fn deferred_ranking_note() -> &'static str {
    "(local BM25 ranking deferred: source admission did not finish within this call's budget; \
     late ranking output is discarded. Other sections are independently admitted. \
     LEAN_CTX_COMPOSE_BUDGET_MS raises the budget.)"
}

/// Prose and documentation formats. An identifier's mentions there rank below
/// its definition and its uses in code.
const PROSE_GLOBS: &str = "*.{md,mdx,markdown,rst,txt,adoc}";

/// Exact-match evidence for `primary`, best first: definitions, then uses in
/// code, and prose mentions only when code has none. A single capped search in
/// walk order used to fill all ten slots with whatever came first — root-level
/// docs and agent rule files — so the code the task asked about never showed.
fn exact_matches_section(
    primary: &str,
    project_root: &str,
    crp_mode: CrpMode,
    allow_secret: bool,
) -> String {
    let search = |pattern: &str, max: usize, exclude: Option<&str>| {
        crate::tools::ctx_search::handle_filtered(
            pattern,
            project_root,
            None,
            max,
            crp_mode,
            true,
            allow_secret,
            false,
            exclude,
            None,
        )
        .text
    };
    let found = |text: &str| {
        let t = text.trim_start();
        !(t.is_empty() || t.starts_with("0 matches") || t.starts_with("ERR"))
    };
    let mut out = String::new();
    let definition = format!(
        r"\b(?:fn|struct|enum|trait|type|const|static|mod|class|interface|def|func|function)\s+{}\b",
        regex::escape(primary)
    );
    let defs = search(&definition, 5, None);
    if found(&defs) {
        out.push_str("Definitions:\n");
        out.push_str(defs.trim());
        out.push('\n');
    }
    let code = search(primary, 10, Some(PROSE_GLOBS));
    if found(&code) {
        out.push_str("Uses in code:\n");
        out.push_str(code.trim());
        out.push('\n');
    } else {
        out.push_str(search(primary, 10, None).trim());
        out.push('\n');
    }
    out
}

/// Small-file inlining (owner decision 2026-10-31). Benchmark transcripts
/// (Claude Code + Haiku, ledgerlite + Dispatch edit tasks): after an
/// orientation call the agent spent its next model step reading the files the
/// overview had named — every full read there was a file of at most 61 lines /
/// ~520 tokens, at most four per step, ~1.6k tokens together; larger files were
/// read by symbol or line range instead. Those small files are delivered here
/// in full, verbatim, so the overview call already carries them.
const SMALL_FILE_MAX_LINES: usize = 150;
const SMALL_FILE_MAX_TOKENS: usize = 800;
const SMALL_FILES_MAX: usize = 4;
/// Token budget of the whole section. `LEAN_CTX_COMPOSE_INLINE_TOKENS`
/// overrides it; `0` turns inlining off.
const DEFAULT_SMALL_FILES_BUDGET_TOKENS: usize = 2_000;

fn small_files_budget_tokens() -> usize {
    std::env::var("LEAN_CTX_COMPOSE_INLINE_TOKENS")
        .ok()
        .and_then(|v| v.parse::<usize>().ok())
        .unwrap_or(DEFAULT_SMALL_FILES_BUDGET_TOKENS)
}

/// Files the sections above name, in the order they name them: ranked
/// results (`path :: symbol`), exact matches (`path:line …`) and symbol
/// bodies (`File: path (…)`).
fn named_files(composed: &str) -> Vec<String> {
    let mut seen = std::collections::HashSet::new();
    let mut files = Vec::new();
    for line in composed.lines() {
        let candidate = if let Some((path, _)) = line.split_once(" :: ") {
            path.trim()
        } else if let Some(rest) = line.strip_prefix("File: ") {
            rest.split(" (").next().unwrap_or("").trim()
        } else {
            let Some((path, rest)) = line.split_once(':') else {
                continue;
            };
            if !rest.starts_with(|c: char| c.is_ascii_digit()) {
                continue;
            }
            path.trim()
        };
        if candidate.is_empty()
            || candidate.contains(char::is_whitespace)
            || !candidate.contains('.')
            || candidate.starts_with('[')
        {
            continue;
        }
        if seen.insert(candidate.to_string()) {
            files.push(candidate.to_string());
        }
    }
    files
}

/// The named files that are small, admitted exactly as `ctx_read` admits a
/// read (path jail, binary and policy checks, secret-like paths only for a
/// role that may read them), in full and byte-exact inside a fence. Empty when
/// none qualifies. Deterministic: order and selection depend only on the
/// sections above and the file contents.
fn small_files_section(composed: &str, project_root: &str, allow_secret: bool) -> String {
    let budget = small_files_budget_tokens();
    if budget == 0 {
        return String::new();
    }
    let mut used = 0usize;
    let mut inlined = 0usize;
    let mut section = String::new();
    let mut remaining_bytes = lean_ctx::core::limits::max_read_bytes();
    for rel in named_files(composed) {
        if inlined >= SMALL_FILES_MAX {
            break;
        }
        let path = std::path::Path::new(project_root).join(&rel);
        if !allow_secret && lean_ctx::core::io_boundary::is_secret_like(&path).is_some() {
            continue;
        }
        let Ok(meta) = std::fs::metadata(&path) else {
            continue;
        };
        // Cheap pre-filter before reading: 150 lines of code stay far below this.
        if !meta.is_file() || meta.len() > 16 * 1024 {
            continue;
        }
        let Ok(content) = crate::tools::ctx_read::read_file_for_tool_rooted_budgeted(
            &path.to_string_lossy(),
            project_root,
            "ctx_read",
            &mut remaining_bytes,
        ) else {
            continue;
        };
        let lines = content.lines().count();
        let tokens = count_tokens(&content);
        if lines == 0
            || lines > SMALL_FILE_MAX_LINES
            || tokens > SMALL_FILE_MAX_TOKENS
            || used + tokens > budget
        {
            continue;
        }
        let longest_run = content.split(|c| c != '`').map(str::len).max().unwrap_or(0);
        let fence = "`".repeat(longest_run.max(2) + 1);
        if section.is_empty() {
            section.push_str("\n## Small files (full, verbatim)\n");
        }
        section.push_str(&format!("File: {rel} ({lines} lines)\n{fence}\n{content}"));
        if !content.ends_with('\n') {
            section.push('\n');
        }
        section.push_str(&fence);
        section.push('\n');
        used += tokens;
        inlined += 1;
    }
    section
}

/// Compose a single rich response for `task`.
pub fn handle(task: &str, project_root: &str, crp_mode: CrpMode) -> (String, usize) {
    lean_ctx::core::policy::runtime::with_project_source_view(project_root, || {
        handle_in_view(task, project_root, crp_mode)
    })
    .unwrap_or_else(|reason| {
        (
            format!(
                "ERROR: context withheld: source authority changed or could not be verified \
                 ({reason})"
            ),
            0,
        )
    })
}

fn handle_in_view(task: &str, project_root: &str, crp_mode: CrpMode) -> (String, usize) {
    let task = task.trim();
    if task.is_empty() {
        return ("ERROR: task is required".to_string(), 0);
    }

    let protected = lean_ctx::core::policy::runtime::is_active();
    let keywords = extract_keywords(task, 6);
    let allow_secret = lean_ctx::core::roles::active_role().io.allow_secret_paths;

    let mut out = String::new();
    out.push_str(&format!("TASK: {task}\n"));
    if keywords.is_empty() {
        out.push_str("KEYWORDS: (none extracted — using full task for ranking)\n");
    } else {
        out.push_str(&format!("KEYWORDS: {}\n", keywords.join(", ")));
    }

    // 1. Fresh source admission precedes ranking under the same pinned view as
    //    the rest of this response, including in the bounded worker.
    out.push_str("\n## Ranked files (local BM25)\n");
    {
        let _phase = lean_ctx::core::phase_timing::enter("ctx_compose: ranking");
        out.push_str(&ranked_files_budgeted(task, project_root, crp_mode));
    }
    if protected {
        // These legacy stores do not retain enough original-source provenance
        // to reauthorize derived records after a policy or rights change.
        out.push_str("\nStored graph and memory enrichment withheld: source authorization is unavailable. Fresh exact matches and symbols follow.");
    }
    out.push('\n');

    // 2. Exact match locations for the most specific identifier-shaped keyword.
    // Broad prose words and acronyms create repository-wide README/Dockerfile
    // noise. Within identifiers, the resident index ranks the rarest one first.
    let exact_phase = lean_ctx::core::phase_timing::enter("ctx_compose: exact matches");
    let ranked_keywords = if protected {
        keywords.clone()
    } else {
        order_by_specificity(&keywords, project_root)
    };
    if let Some(primary) = ranked_keywords
        .iter()
        .find(|keyword| is_code_identifier(keyword))
    {
        out.push_str(&format!("\n## Exact matches: '{primary}'\n"));
        out.push_str(&exact_matches_section(
            primary,
            project_root,
            crp_mode,
            allow_secret,
        ));
    }

    // 3. Inline the symbol bodies that best cover the task keywords. Rather
    //    than just the first match, select the non-redundant *set* of symbols
    //    with maximal keyword coverage under a token budget via submodular
    //    greedy (1−1/e optimal). Two keywords resolving to the same symbol, or
    //    a symbol whose body adds no new keyword, are naturally pruned.
    drop(exact_phase);
    let symbols_phase = lean_ctx::core::phase_timing::enter("ctx_compose: symbol bodies");
    use lean_ctx::core::context_packing::CoverageItem;
    let mut snippets: Vec<String> = Vec::new();
    let mut items: Vec<CoverageItem> = Vec::new();
    for kw in &keywords {
        if let Some((rendered, toks)) =
            crate::tools::ctx_symbol::best_symbol_snippet_for_task(kw, task, project_root)
        {
            // The snippet always covers its triggering keyword, plus any other
            // task keyword its body textually surfaces (a more central symbol).
            let mut terms: std::collections::HashSet<String> =
                std::collections::HashSet::from([kw.clone()]);
            for other in &keywords {
                if other != kw && rendered.contains(other.as_str()) {
                    terms.insert(other.clone());
                }
            }
            if let Some(index) = snippets.iter().position(|snippet| snippet == &rendered) {
                items[index].terms.extend(terms);
                continue;
            }
            items.push(CoverageItem {
                terms,
                cost: toks.max(1),
            });
            snippets.push(rendered);
        }
    }
    if !items.is_empty() {
        let selected = lean_ctx::core::context_packing::greedy_max_coverage(
            &items,
            symbol_budget_tokens(),
            |_| 1.0,
        );
        let mut seen = std::collections::HashSet::new();
        let mut header_written = false;
        for idx in selected {
            let rendered = snippets[idx].trim();
            if rendered.is_empty() || !seen.insert(rendered.to_string()) {
                continue;
            }
            if !header_written {
                out.push_str("\n## Top symbols (bodies)\n");
                header_written = true;
            }
            out.push_str(rendered);
            out.push('\n');
        }
    }

    // 4. (No associative "Related" section: measured 2026-10-16 over 882
    //    replies it added no relevant file the ranked list had missed, at
    //    ~10 % precision and ~110 tokens when shown.)
    drop(symbols_phase);

    // 4b. Small files in full: the bodies the agent would read next anyway.
    {
        let _phase = lean_ctx::core::phase_timing::enter("ctx_compose: small files");
        out.push_str(&small_files_section(&out, project_root, allow_secret));
    }

    // 5. Context Kernel enrichment — cross-store context from knowledge,
    //    session state and the indexes that the lexical pipeline misses.
    //    Budget: 20% of symbol budget. Graceful no-op if kernel returns None.
    if !protected {
        use lean_ctx::core::context_kernel::context_dedup::dedup_kernel_blocks;
        let _phase = lean_ctx::core::phase_timing::enter("ctx_compose: kernel enrichment");

        let budget = kernel_supplement_budget(project_root);
        if let Some(enrichment) =
            lean_ctx::core::context_kernel::bridge::kernel_enrich(task, project_root, budget)
                .filter(|enrichment| !enrichment.blocks.is_empty())
        {
            let blocks =
                dedup_kernel_blocks(&enrichment.blocks, &mut std::collections::HashSet::new());
            if !blocks.is_empty() {
                out.push_str("\n## Context Kernel\n");
                out.push_str(&blocks);
                out.push('\n');
            }
        }
    }

    let sent = count_tokens(&out);
    (out, sent)
}

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

    #[test]
    fn ranking_worker_keeps_original_source_policy_and_full_role() {
        let _env = lean_ctx::core::data_dir::test_env_lock();
        let policy = lean_ctx::core::policy::load(
            "name = \"compose-worker\"\nversion = \"1.0.0\"\ndescription = \"test\"\n\
             [filters]\nclassification = \"block\"\nblocked_labels = [\"CONFIDENTIAL\"]\n\
             [redaction]\ncustomer = \"ACC-1234\"\n",
        )
        .unwrap();
        let _policy = lean_ctx::core::policy::runtime::TestPolicyOverride::set(Some(policy));
        let dir = tempfile::tempdir().unwrap();
        std::fs::write(
            dir.path().join("public.rs"),
            "fn authenticate() { let customer = \"ACC-1234\"; }\n",
        )
        .unwrap();
        std::fs::write(
            dir.path().join("private.rs"),
            "// CONFIDENTIAL\nfn authenticate_private() { let hidden = \"CLASSIFIED_CANARY\"; }\n",
        )
        .unwrap();
        let root = dir.path().to_str().unwrap();
        let rank = || {
            lean_ctx::core::policy::runtime::with_project_source_view(root, || {
                ranked_files_budgeted("authenticate", root, CrpMode::Off)
            })
            .unwrap()
        };
        let role = lean_ctx::core::roles::load_role("coder").unwrap();
        lean_ctx::core::roles::with_test_active_role(role.clone(), || {
            let ranked = rank();
            assert!(ranked.contains("public.rs"), "{ranked}");
            assert!(ranked.contains("REDACTED"), "{ranked}");
            assert!(!ranked.contains("ACC-1234"));
            assert!(!ranked.contains("private.rs"));
            assert!(!ranked.contains("CLASSIFIED_CANARY"));
            static ONE_SLOT: tokio::sync::Semaphore = tokio::sync::Semaphore::const_new(1);
            let occupied = ONE_SLOT.try_acquire().unwrap();
            let rank_with_limit = || {
                lean_ctx::core::policy::runtime::with_project_source_view(root, || {
                    ranked_files_with_slots("authenticate", root, CrpMode::Off, &ONE_SLOT)
                })
                .unwrap()
            };
            let busy = rank_with_limit();
            assert!(busy.contains("ranking busy"));
            assert!(!busy.contains("public.rs"));
            drop(occupied);
            let recovered = rank_with_limit();
            assert!(recovered.contains("public.rs"));
            assert!(recovered.contains("REDACTED"));
            assert!(!recovered.contains("CLASSIFIED_CANARY"));
        });
        let mut denied = role;
        denied.tools.denied.push("ctx_search".into());
        lean_ctx::core::roles::with_test_active_role(denied, || {
            let ranked = rank();
            assert!(ranked.starts_with("ERR:"), "{ranked}");
            assert!(!ranked.contains("public.rs"));
        });
    }

    #[test]
    fn rank_by_doc_freq_puts_rare_identifier_first() {
        // The evcc/#993 shape: an "OCPP … GetMaxCurrent" task. `Current` and
        // `OCPP` are common tokens; `GetMaxCurrent` is rare. The rare one must
        // seed the exact-match grep so it lands on code, not README/Dockerfile.
        let keywords = vec![
            "OCPP".to_string(),
            "GetMaxCurrent".to_string(),
            "Current".to_string(),
        ];
        let doc_freqs = std::collections::HashMap::from([
            ("ocpp".to_string(), 120),
            ("current".to_string(), 400),
            ("getmaxcurrent".to_string(), 3),
        ]);
        let ranked = rank_by_doc_freq(&keywords, &doc_freqs);
        assert_eq!(ranked.first().unwrap(), "GetMaxCurrent");
        assert_eq!(ranked.last().unwrap(), "Current");
    }

    /// Benchmark replay (feature-export-csv): after `ctx_compose` the agent's
    /// next step read cli.py and test_core.py in full. Named small files now
    /// come with the overview, verbatim; a large one and a secret-like one
    /// named by the same sections do not.
    #[test]
    fn named_small_files_are_inlined_verbatim_within_budget() {
        let _env = lean_ctx::core::data_dir::test_env_lock();
        let _policy = lean_ctx::core::policy::runtime::TestPolicyOverride::set(None);
        let dir = tempfile::tempdir().unwrap();
        std::fs::write(
            dir.path().join("pyproject.toml"),
            "[project]\nname = \"x\"\n",
        )
        .unwrap();
        std::fs::create_dir(dir.path().join("pkg")).unwrap();
        let cli = "import argparse\n\n\ndef main():\n    return \"```\"\n";
        std::fs::write(dir.path().join("pkg/cli.py"), cli).unwrap();
        let big: String = (0..400).fold(String::new(), |mut acc, i| {
            use std::fmt::Write as _;
            let _ = writeln!(acc, "value_{i} = {i}");
            acc
        });
        std::fs::write(dir.path().join("pkg/big.py"), &big).unwrap();
        std::fs::write(dir.path().join(".env"), "TOKEN=abc\n").unwrap();
        let composed = "## Ranked files (local BM25)\n\
                        pkg/big.py :: value_1 [Variable] (L2-2)\n\
                        pkg/cli.py :: main [Function] (L4-5)\n\
                        ## Exact matches: 'main'\n\
                        .env:1 TOKEN=abc\n\
                        pkg/cli.py:4 def main():\n";
        let root = dir.path().to_str().unwrap();
        let section = lean_ctx::core::policy::runtime::with_project_source_view(root, || {
            small_files_section(composed, root, false)
        })
        .unwrap();
        assert!(
            section.contains(&format!("File: pkg/cli.py (5 lines)\n````\n{cli}````\n")),
            "{section}"
        );
        assert!(!section.contains("value_399"), "{section}");
        assert!(!section.contains("TOKEN=abc"), "{section}");
        assert_eq!(section.matches("File: ").count(), 1, "{section}");
    }

    #[test]
    fn rank_by_doc_freq_sinks_absent_tokens_and_is_stable() {
        // A token absent from the corpus (df 0) is useless as a grep seed and
        // must sort last; equal-df tokens keep their original order.
        let keywords = vec![
            "absent".to_string(),
            "alpha".to_string(),
            "beta".to_string(),
        ];
        let doc_freqs =
            std::collections::HashMap::from([("alpha".to_string(), 5), ("beta".to_string(), 5)]);
        let ranked = rank_by_doc_freq(&keywords, &doc_freqs);
        assert_eq!(ranked, vec!["alpha", "beta", "absent"]);
    }

    #[test]
    fn rank_prefers_code_identifier_over_rarer_prose_word() {
        // The regression the case-sensitive grep exposed: a rarer lowercase prose
        // token (`measurand`, df 4) must NOT beat a camelCase identifier
        // (`GetMaxCurrent`, df 30) as the seed — the identifier greps to code,
        // the prose word whiffs against `Measurand`.
        let keywords = vec!["measurand".to_string(), "GetMaxCurrent".to_string()];
        let doc_freqs = std::collections::HashMap::from([
            ("measurand".to_string(), 4),
            ("getmaxcurrent".to_string(), 30),
        ]);
        let ranked = rank_by_doc_freq(&keywords, &doc_freqs);
        assert_eq!(ranked.first().unwrap(), "GetMaxCurrent");
    }

    #[test]
    fn is_code_identifier_classifies_camel_snake_vs_prose_and_acronym() {
        assert!(is_code_identifier("GetMaxCurrent"));
        assert!(is_code_identifier("CurrentGetter"));
        assert!(is_code_identifier("get_max_current"));
        // Leading-cap word and all-caps acronym are not code identifiers.
        assert!(!is_code_identifier("Current"));
        assert!(!is_code_identifier("OCPP"));
        assert!(!is_code_identifier("charger"));
    }

    #[test]
    fn extract_keywords_drops_stopwords_and_short_tokens() {
        let kw = extract_keywords("How does the BM25Index cache work for ctx_search?", 6);
        assert!(kw.contains(&"BM25Index".to_string()));
        assert!(kw.contains(&"cache".to_string()));
        assert!(kw.contains(&"ctx_search".to_string()));
        assert!(!kw.iter().any(|k| k == "the" || k == "How" || k == "for"));
    }

    #[test]
    fn extract_keywords_dedups_and_caps() {
        let kw = extract_keywords("alpha alpha beta gamma delta epsilon zeta eta", 3);
        assert_eq!(kw.len(), 3);
        assert_eq!(kw[0], "alpha");
    }

    #[test]
    fn exact_matches_choose_specific_identifier_not_first_broad_keyword() {
        let keywords = extract_keywords(
            "OCPP charger GetMaxCurrent Current.Offered measurand CurrentGetter",
            6,
        );
        assert!(keywords.iter().any(|keyword| keyword == "GetMaxCurrent"));
        assert!(keywords.iter().any(|keyword| is_code_identifier(keyword)));
        assert!(!is_code_identifier("OCPP"));

        let prose = extract_keywords("Fix semantic ranking exact matches", 6);
        assert!(prose.iter().all(|keyword| !is_code_identifier(keyword)));
    }

    /// Regression: root-level docs that mention the identifier came first in
    /// walk order and filled every slot; the definition in code never showed.
    #[test]
    fn exact_matches_rank_definition_and_code_above_prose() {
        // The source view re-verifies policy and role after the search; a
        // parallel test switching either (Windows CI) would void it.
        let _env = lean_ctx::core::data_dir::test_env_lock();
        let dir = tempfile::tempdir().unwrap();
        let root = dir.path();
        std::fs::write(root.join("AGENTS.md"), "use widget_factory\n".repeat(30)).unwrap();
        std::fs::create_dir_all(root.join("src")).unwrap();
        std::fs::write(
            root.join("src/lib.rs"),
            "pub fn widget_factory() -> u8 { 1 }\n",
        )
        .unwrap();
        std::fs::write(
            root.join("src/main.rs"),
            "fn main() { lib::widget_factory(); }\n",
        )
        .unwrap();
        let root_str = root.to_str().unwrap();
        let out = lean_ctx::core::policy::runtime::with_project_source_view(root_str, || {
            exact_matches_section("widget_factory", root_str, CrpMode::Off, false)
        })
        .expect("source view");

        let defs = out.find("Definitions:").expect("definition section");
        let uses = out.find("Uses in code:").expect("code section");
        assert!(defs < uses, "{out}");
        assert!(out.contains("lib.rs:1"), "definition found: {out}");
        assert!(out.contains("main.rs:1"), "use found: {out}");
        assert!(
            !out.contains("AGENTS.md"),
            "prose only when code has none: {out}"
        );

        let prose_only =
            lean_ctx::core::policy::runtime::with_project_source_view(root_str, || {
                exact_matches_section("use", root_str, CrpMode::Off, false)
            })
            .expect("source view");
        assert!(prose_only.contains("AGENTS.md"), "{prose_only}");
    }

    #[test]
    fn empty_task_is_rejected() {
        let (out, tok) = handle("   ", "/tmp", CrpMode::Off);
        assert!(out.starts_with("ERROR"));
        assert_eq!(tok, 0);
    }

    #[test]
    fn handle_withholds_context_when_project_authority_cannot_be_verified() {
        let root = tempfile::tempdir().unwrap();
        let missing = root.path().join("missing-project");
        let (output, tokens) = handle(
            "find authentication bugs",
            missing.to_str().unwrap(),
            CrpMode::Tdd,
        );
        assert_eq!(tokens, 0);
        assert!(output.starts_with("ERROR"));
        assert!(output.contains("source authority"));
        assert!(!output.contains("TASK:"));
    }

    #[test]
    fn deferred_ranking_note_is_deterministic_and_has_no_timing() {
        // Issue #498 / #1366: elapsed_ms must never appear in the note — it
        // varies between calls and defeats provider prompt caching.
        let a = deferred_ranking_note();
        let b = deferred_ranking_note();
        assert_eq!(a, b, "deferred note must be byte-stable across calls");
        assert!(
            !a.contains("elapsed"),
            "deferred note must not embed timing data: {a}"
        );
    }

    #[test]
    fn deferred_note_does_not_promise_cached_or_late_delivery() {
        let note = deferred_ranking_note();
        assert!(note.contains("late ranking output is discarded"));
        assert!(note.contains("independently admitted"));
        assert!(!note.contains("warming"));
        assert!(!note.contains("next call"));
    }
}