tokenix 0.65.1

use anyhow::Result;
use std::path::{Path, PathBuf};
use std::time::{SystemTime, UNIX_EPOCH};

use crate::chunker::count_tokens;
use crate::store::{log_hook_event, HookEvent};

mod shell;

const POST_HOOK_TOOLS: &[&str] = &["Bash", "ListDirectory"];
const BASH_MAX_LINES: usize = 100;
const BASH_HEAD_LINES: usize = 40;
const BASH_TAIL_LINES: usize = 15;

/// Bash-aware compression: checks user TOML filters first, then built-in heuristics.
const UNFILTERED_LOG_CAP: u64 = 262_144; // 256 KB before rotation

fn log_unfiltered_cmd(cmd: &str) {
    let cmd = cmd.trim();
    // Skip empties and multi-line commands (for-loops, heredocs): they log one
    // entry per line and pollute the "what to filter next" signal with fragments
    // like `for`, `done`, and stray regex pieces rather than real command names.
    if cmd.is_empty() || cmd.contains('\n') {
        return;
    }
    // Write to the global ~/.tokenix/ dir, not the project dir, to avoid
    // accidentally committing internal tokenix logs.
    let log_path = match crate::store::global_dir() {
        Some(h) => h.join("unfiltered_cmds.log"),
        None => return,
    };
    if let Some(parent) = log_path.parent() {
        let _ = std::fs::create_dir_all(parent);
    }
    // Rotate at the cap so the log cannot grow unbounded (one generation kept).
    if std::fs::metadata(&log_path).is_ok_and(|m| m.len() >= UNFILTERED_LOG_CAP) {
        let _ = std::fs::rename(&log_path, log_path.with_extension("log.1"));
    }
    // Shell input is where credentials actually live (`curl -H "Authorization:
    // Bearer …"`, `psql postgres://user:pw@host`). The failure tee already
    // masks them before writing; this log did not, and it is long-lived.
    let entry = format!("{}\n", crate::conversation_audit::redact_credentials(cmd));
    use std::io::Write;
    if let Ok(mut f) = std::fs::OpenOptions::new()
        .create(true)
        .append(true)
        .open(&log_path)
    {
        let _ = f.write_all(entry.as_bytes());
        crate::store::restrict_to_owner(&log_path);
    }
}

pub fn compress_bash_output(cmd: &str, s: &str) -> String {
    compress_bash_output_for_stream(cmd, s, false, None)
}

fn compress_bash_output_for_stream(
    cmd: &str,
    s: &str,
    is_stderr: bool,
    exit_ok: Option<bool>,
) -> String {
    // User-defined TOML filters take priority over built-in heuristics.
    let user_filters = crate::filters::load_filter_groups_for_command(cmd);
    if let Some(f) = crate::filters::find_filter_ranked(cmd, &user_filters) {
        if is_stderr && !f.filter_stderr {
            return compress_output(s);
        }
        // The global ceiling applies to filtered output too: a filter's own caps
        // are per-filter, and a passthrough fallback (`passthrough_when_emptied`,
        // failure-signal passthrough, `never_worse`) can hand back the full raw
        // output, which is exactly the shape this cap exists to bound.
        return enforce_token_budget(&crate::filters::apply_filter_with_exit(s, f, exit_ok));
    }

    // No filter matched — record for later analysis (tokenix filter list).
    log_unfiltered_cmd(cmd);

    let raw_lines: Vec<&str> = s.lines().collect();
    if is_status_poll_output(&raw_lines) {
        let out = compress_status_poll(&raw_lines);
        if out.len() < s.len() {
            // Early return still owes the global ceiling: this path skips
            // `compress_output`, where `enforce_token_budget` normally runs.
            return enforce_token_budget(&out);
        }
    }

    let base = compress_output(s);
    let lines: Vec<&str> = base.lines().collect();

    // `cargo metadata`: a single huge JSON blob compact_json cannot shrink — summarize
    // it to package count + workspace members instead of letting 500k tokens through.
    if is_cargo_metadata_command(cmd) {
        let out = compress_cargo_metadata(&base);
        if out.len() < base.len() {
            return out;
        }
    }

    // `cargo tree`: highly repetitive (subtrees repeat, marked `(*)`); collapse to the
    // unique crate set.
    if is_cargo_tree_command(cmd) {
        let out = compress_cargo_tree(&lines);
        if out.len() < base.len() {
            return out;
        }
    }

    // Plain grep: strip the indentation in matched content and group by file.
    if is_grep_command(cmd) {
        let out = compress_grep(&lines);
        if out.len() < base.len() {
            return out;
        }
    }

    // ps: keep the header + busiest processes by %CPU.
    if is_ps_command(cmd) {
        let out = compress_ps(&lines);
        if out.len() < base.len() {
            return out;
        }
    }

    // Cargo: always try (it filters signal from noise regardless of total length)
    if is_cargo_output(&lines) {
        let cargo_out = compress_cargo(&lines);
        if cargo_out.len() < base.len() {
            return cargo_out;
        }
    }

    if is_path_listing_command(cmd) {
        let listing_out = compress_path_listing(&lines);
        if listing_out.len() < base.len() {
            return listing_out;
        }
    }

    if is_git_status_command(cmd) {
        let status_out = compress_git_status(&lines);
        if status_out.len() < base.len() {
            return status_out;
        }
    }

    if is_git_log_command(cmd) || is_git_log(&lines) {
        let log_out = compress_git_log(&lines);
        if log_out.len() < base.len() {
            return log_out;
        }
    }

    if is_git_diff_command(cmd) {
        let diff_out = compress_git_diff(&lines);
        if diff_out.len() < base.len() {
            return diff_out;
        }
    }

    // Kubernetes: compress kubectl output
    if is_kubectl_command(cmd) {
        let kube_out = compress_kubectl(&lines);
        if kube_out.len() < base.len() {
            return kube_out;
        }
    }

    // npm/yarn/pnpm/bun: compress package manager output
    if is_pkg_manager_command(cmd) {
        let pkg_out = compress_pkg_manager(&lines);
        if pkg_out.len() < base.len() {
            return pkg_out;
        }
    }

    // Terraform: compress terraform output
    if is_terraform_command(cmd) {
        let tf_out = compress_terraform(&lines);
        if tf_out.len() < base.len() {
            return tf_out;
        }
    }

    // Docker compose: compress compose output
    if is_docker_compose_command(cmd) {
        let dc_out = compress_docker_compose(&lines);
        if dc_out.len() < base.len() {
            return dc_out;
        }
    }

    // Make/ninja/cmake: compress build system output
    if is_build_command(cmd) {
        let build_out = compress_build(&lines);
        if build_out.len() < base.len() {
            return build_out;
        }
    }

    // Generic: aggressive truncation for very long output
    if lines.len() > BASH_MAX_LINES * 2 {
        return aggressive_truncate(&lines);
    }

    if lines.len() <= BASH_MAX_LINES {
        return base;
    }

    truncate_head_tail(&lines, BASH_HEAD_LINES, BASH_TAIL_LINES)
}

fn is_path_listing_command(cmd: &str) -> bool {
    let trimmed = cmd.trim();
    // Plain ls / ls with flags / recursive find
    trimmed == "ls"
        || trimmed == "ls -R"
        || trimmed.starts_with("ls ")
        // POSIX find
        || trimmed.starts_with("find ")
        // Windows cmd / PowerShell dir
        || trimmed == "dir"
        || trimmed.starts_with("dir ")
        // PowerShell Get-ChildItem and its aliases
        || trimmed.starts_with("Get-ChildItem")
        || trimmed.starts_with("get-childitem")
        || trimmed == "gci"
        || trimmed.starts_with("gci ")
        // tree command (Unix and Windows)
        || trimmed == "tree"
        || trimmed.starts_with("tree ")
}

fn is_cargo_output(lines: &[&str]) -> bool {
    lines.iter().take(50).any(|l| {
        let t = l.trim();
        t.starts_with("Compiling ")
            || t.starts_with("Finished ")
            || t.starts_with("error[E")
            || t.contains("test result:")
    })
}

fn compress_cargo(lines: &[&str]) -> String {
    let mut out: Vec<&str> = Vec::new();
    let mut in_diagnostic = false;
    // Inside a test-failure stdout block (`---- <test> stdout ----` up to the
    // `test result:` summary). The reason a test failed — a custom panic message,
    // a pretty-assertion colour diff, a backtrace — is free-form and matches no
    // fixed prefix, so the block is captured verbatim instead of line-by-line.
    let mut in_failure_block = false;
    let mut warning_count: u32 = 0;
    const MAX_WARNINGS: u32 = 5;

    for line in lines {
        let t = line.trim();

        if !in_failure_block && t.starts_with("---- ") && t.ends_with("----") {
            in_failure_block = true;
        }
        if in_failure_block {
            out.push(line);
            // `test result:` ends the failure section; a `running ` line marks the
            // start of a fresh test binary, so it also closes a truncated block.
            if t.starts_with("test result:") || t.starts_with("running ") {
                in_failure_block = false;
            }
            continue;
        }

        let is_error = (t.starts_with("error[") || t == "error" || t.starts_with("error: "))
            && !t.starts_with("error_");
        let is_warning = t.starts_with("warning[") || t.starts_with("warning: ");
        let is_context = t.starts_with("-->")
            || (t.starts_with('|') && t.len() > 1)
            || t.starts_with("= note:")
            || t.starts_with("= help:")
            || t.starts_with("help:");
        let is_summary = t.starts_with("Finished ")
            || t.starts_with("error: aborting")
            || t.contains("test result:")
            || t.starts_with("running ")
            || t.starts_with("FAILED")
            || (t.starts_with("test ") && (t.ends_with("ok") || t.ends_with("FAILED")));
        // A panic outside a cargo-test stdout block (e.g. a plain binary run)
        // still carries the failure reason and must survive compression.
        let is_panic = t.contains("panicked at");

        if is_error || is_panic {
            out.push(line);
            in_diagnostic = true;
        } else if is_warning && warning_count < MAX_WARNINGS {
            out.push(line);
            in_diagnostic = true;
            warning_count += 1;
        } else if is_context && in_diagnostic {
            out.push(line);
        } else if is_summary {
            out.push(line);
            in_diagnostic = false;
        } else {
            in_diagnostic = false;
        }
    }

    if warning_count >= MAX_WARNINGS {
        out.push("  ... (additional warnings omitted)");
    }

    out.join("\n")
}

fn is_git_log(lines: &[&str]) -> bool {
    lines.iter().take(5).any(|l| l.starts_with("commit "))
}

fn is_git_log_command(cmd: &str) -> bool {
    let cmd = cmd.trim();
    cmd == "git log" || cmd.starts_with("git log ")
}

fn is_git_status_command(cmd: &str) -> bool {
    let cmd = cmd.trim();
    cmd == "git status" || cmd.starts_with("git status ")
}

fn is_git_diff_command(cmd: &str) -> bool {
    let cmd = cmd.trim();
    cmd == "git diff" || cmd.starts_with("git diff ")
}

fn compress_git_log(lines: &[&str]) -> String {
    let oneline: Vec<&str> = lines
        .iter()
        .map(|line| line.trim())
        .filter(|line| {
            line.len() > 8
                && line.chars().take_while(|c| c.is_ascii_hexdigit()).count() >= 7
                && line.chars().nth(7).is_some_and(|c| c.is_whitespace())
        })
        .collect();
    if oneline.len() >= 3 {
        let first = oneline.first().copied().unwrap_or_default();
        let last = oneline.last().copied().unwrap_or_default();
        return format!(
            "git log: {} commits\nfirst: {first}\nlast: {last}",
            oneline.len()
        );
    }

    const MAX_COMMITS: usize = 20;
    let mut commit_count: usize = 0;
    let mut keep_until: usize = 0;

    for (i, line) in lines.iter().enumerate() {
        if line.starts_with("commit ") {
            commit_count += 1;
            if commit_count > MAX_COMMITS {
                break;
            }
        }
        keep_until = i + 1;
    }

    if keep_until >= lines.len() {
        return lines.join("\n");
    }
    let omitted = lines.len() - keep_until;
    format!(
        "{}\n[... {} more lines omitted (>{} commits)]",
        lines[..keep_until].join("\n"),
        omitted,
        MAX_COMMITS
    )
}

/// Compact `git status` (verbose or `--short`/porcelain) to one terse
/// `CODE path` line per change (`M`/`A`/`D`/`R`/`T`/`C`/`??`). Verbose
/// `modified:   file` lines are far longer than the porcelain form, so this
/// rewrites them; porcelain lines pass through; the prose/headers are dropped.
fn compress_git_status(lines: &[&str]) -> String {
    const VERBOSE: &[(&str, &str)] = &[
        ("modified:", "M"),
        ("new file:", "A"),
        ("deleted:", "D"),
        ("renamed:", "R"),
        ("typechange:", "T"),
        ("copied:", "C"),
        ("both modified:", "U"),
    ];
    let mut out: Vec<String> = Vec::new();
    let mut in_untracked = false;
    for line in lines {
        let t = line.trim();
        if t.is_empty() {
            continue;
        }
        // Section markers from verbose output.
        if t.starts_with("Untracked files") {
            in_untracked = true;
            continue;
        }
        // Hint lines like `(use "git add"...)` sit inside a section — skip them
        // without changing the section state (untracked filenames follow).
        if t.starts_with('(') {
            continue;
        }
        if t.starts_with("Changes ")
            || t.starts_with("On branch")
            || t.starts_with("Your branch")
            || t.starts_with("no changes")
            || t.starts_with("nothing to commit")
            || t.contains("working tree clean")
        {
            in_untracked = false;
            continue;
        }
        // Verbose "modified:   file" → "M file".
        if let Some((_, code)) = VERBOSE.iter().find(|(kw, _)| t.starts_with(kw)) {
            let file = t.split_once(':').map(|x| x.1).unwrap_or("").trim();
            out.push(format!("{code} {file}"));
            continue;
        }
        // Porcelain ("?? file", " M file", "MM file"): already terse — keep.
        let code2: String = t.chars().take(2).collect();
        let is_porcelain = code2 == "??"
            || code2
                .chars()
                .all(|c| matches!(c, 'M' | 'A' | 'D' | 'R' | 'C' | 'U' | 'T' | ' '))
                && code2 != "  ";
        if is_porcelain && t.len() >= 3 {
            out.push(t.to_string());
            continue;
        }
        // Bare filename under "Untracked files:".
        if in_untracked {
            out.push(format!("?? {t}"));
        }
    }
    if out.is_empty() {
        return "git status: clean".to_string();
    }
    out.join("\n")
}

fn compress_git_diff(lines: &[&str]) -> String {
    let files = lines
        .iter()
        .filter_map(|line| line.strip_prefix("diff --git "))
        .count();
    let hunks = lines.iter().filter(|line| line.starts_with("@@")).count();
    let additions = lines
        .iter()
        .filter(|line| line.starts_with('+') && !line.starts_with("+++"))
        .count();
    let deletions = lines
        .iter()
        .filter(|line| line.starts_with('-') && !line.starts_with("---"))
        .count();
    let mut keep = Vec::new();
    for line in lines {
        if line.starts_with("diff --git ")
            || line.starts_with("@@")
            || line.starts_with("+++")
            || line.starts_with("---")
        {
            keep.push(*line);
        }
        if keep.len() >= 40 {
            break;
        }
    }
    if files == 0 && hunks == 0 {
        return lines.join("\n");
    }
    // A header-only summary throws away every changed line, which is the whole
    // point of running `git diff`. Only worth it when the diff is big enough
    // that reading it in full is the actual problem.
    const DIFF_SUMMARY_MIN_LINES: usize = 200;
    if lines.len() < DIFF_SUMMARY_MIN_LINES {
        return lines.join("\n");
    }
    format!(
        "git diff: files={files} hunks={hunks} +{additions} -{deletions}\n{}",
        keep.join("\n")
    )
}

fn is_status_poll_output(lines: &[&str]) -> bool {
    lines
        .iter()
        .take(80)
        .filter(|line| {
            let t = line.trim();
            t.starts_with("phase=") || t.starts_with("status=") || t.starts_with("state=")
        })
        .count()
        >= 3
}

fn compress_status_poll(lines: &[&str]) -> String {
    let mut out = Vec::new();
    let mut last_poll: Option<String> = None;
    let mut last_count = 0usize;

    let flush = |out: &mut Vec<String>, last: &mut Option<String>, count: &mut usize| {
        if let Some(value) = last.take() {
            if *count > 1 {
                out.push(format!("{value} x{count}"));
            } else {
                out.push(value);
            }
        }
        *count = 0;
    };

    for line in lines {
        let t = line.trim();
        let is_poll =
            t.starts_with("phase=") || t.starts_with("status=") || t.starts_with("state=");

        if is_poll {
            match &last_poll {
                Some(last) if last == t => {
                    last_count += 1;
                }
                _ => {
                    flush(&mut out, &mut last_poll, &mut last_count);
                    last_poll = Some(t.to_string());
                    last_count = 1;
                }
            }
            continue;
        }

        flush(&mut out, &mut last_poll, &mut last_count);
        out.push(line.to_string());
    }

    flush(&mut out, &mut last_poll, &mut last_count);
    out.join("\n")
}

/// Aggressive truncation for very long output - keeps only errors, head, tail
fn aggressive_truncate(lines: &[&str]) -> String {
    const HEAD_KEEP: usize = 20;
    const TAIL_KEEP: usize = 10;
    const MAX_TOTAL: usize = HEAD_KEEP + TAIL_KEEP + 1;

    if lines.len() <= MAX_TOTAL {
        return lines.join("\n");
    }

    // Priority: error/warning lines
    let mut priority = Vec::new();
    let mut other_head = Vec::new();
    let mut other_tail = Vec::new();

    for (i, line) in lines.iter().enumerate() {
        let t = line.trim();
        let is_priority = t.starts_with("error")
            || t.starts_with("warning")
            || t.starts_with("FAIL")
            || t.starts_with("panic")
            || t.contains("error[")
            || t.contains("warning[")
            || t.contains("ERROR")
            || t.contains("FAILED");

        if is_priority && priority.len() < 50 {
            priority.push(line.to_string());
        } else if i < HEAD_KEEP {
            other_head.push(line.to_string());
        } else if i >= lines.len() - TAIL_KEEP {
            other_tail.push(line.to_string());
        }
    }

    let mut result = Vec::new();
    result.extend(other_head);
    if !priority.is_empty() {
        result.push("[PRIORITY LINES]".to_string());
        result.extend(priority);
    }
    result.push(format!(
        "[... {} lines omitted ...]",
        lines.len() - result.len() - other_tail.len()
    ));
    result.extend(other_tail);
    result.join("\n")
}

fn is_kubectl_command(cmd: &str) -> bool {
    let t = cmd.trim();
    t == "kubectl" || t.starts_with("kubectl ") || t == "k" || t.starts_with("k ")
}

fn compress_kubectl(lines: &[&str]) -> String {
    // kubectl get: tabular - keep header + first/last few rows
    // kubectl describe: verbose - summarize
    // kubectl logs: similar to plain logs
    let first_nonempty = lines.iter().find(|l| !l.trim().is_empty());
    if let Some(first) = first_nonempty {
        let t = first.trim();
        if t.starts_with("NAME") && t.contains("READY") {
            // kubectl get pods/nodes/etc - tabular
            let header = first.to_string();
            let data: Vec<&str> = lines
                .iter()
                .skip(1)
                .filter(|l| !l.trim().is_empty())
                .copied()
                .collect();
            if data.len() <= 20 {
                return lines.join("\n");
            }
            let mut out = vec![header];
            out.extend(data.iter().take(10).map(|s| s.to_string()));
            out.push(format!("... {} more rows ...", data.len() - 10));
            out.extend(data.iter().rev().take(5).rev().map(|s| s.to_string()));
            return out.join("\n");
        }
        if t.starts_with("Name:") || t.starts_with("Namespace:") {
            // kubectl describe: verbose, but the reason anyone runs it is the
            // Events/conditions tail. Replacing it with a fixed placeholder
            // fabricated a response and destroyed exactly that; keep a bounded
            // head/tail view instead.
            return truncate_head_tail(lines, 20, 25);
        }
    }
    lines.join("\n")
}

fn is_pkg_manager_command(cmd: &str) -> bool {
    let t = cmd.trim();
    t.starts_with("npm ")
        || t.starts_with("yarn ")
        || t.starts_with("pnpm ")
        || t.starts_with("bun ")
        || t.starts_with("cargo ")
        || t.starts_with("pip ")
        || t.starts_with("uv ")
        || t.starts_with("composer ")
        || t.starts_with("gem ")
        || t.starts_with("go ")
        || t.starts_with("gradle ")
        || t.starts_with("maven ")
        || t.starts_with("mvn ")
        || t.starts_with("dotnet ")
}

fn compress_pkg_manager(lines: &[&str]) -> String {
    // Generic package manager output: progress bars, downloading, extracting
    // Keep errors, summary, final status
    let mut result = Vec::new();
    let mut in_progress = false;

    for line in lines {
        let t = line.trim();
        if t.is_empty() {
            continue;
        }
        let is_progress = t.starts_with("Progress:")
            || t.starts_with("Downloading")
            || t.starts_with("Extracting")
            || t.starts_with("Installing")
            || t.starts_with("Building")
            || t.starts_with("Compiling")
            || t.starts_with("Resolving")
            || t.starts_with("Fetching")
            || t.contains("█")
            || t.contains("░")
            || t.contains("▓")
            || t.chars().filter(|c| *c == '=' || *c == '>').count() > 10;

        if is_progress {
            if !in_progress {
                result.push("[package operations...]".to_string());
                in_progress = true;
            }
        } else {
            in_progress = false;
            // Case-insensitive: package managers print `ERROR:`, `Error`, and
            // `npm ERR!` — a case-sensitive `starts_with("error")` dropped every
            // one of them once the first 30 lines were used up.
            let lower = t.to_ascii_lowercase();
            let is_important = [
                "error", "err!", "warning", "warn", "fail", "success", "done", "added", "removed",
                "updated", "npm err",
            ]
            .iter()
            .any(|p| lower.starts_with(p))
                || lower.contains("err!")
                || lower.contains("vulnerab")
                || lower.contains("audit");
            if is_important || result.len() < 30 {
                result.push(line.to_string());
            }
        }
    }
    result.join("\n")
}

fn is_terraform_command(cmd: &str) -> bool {
    let t = cmd.trim();
    t.starts_with("terraform ")
}

fn compress_terraform(lines: &[&str]) -> String {
    // Terraform: very verbose, keep plan summary, errors, apply status
    let mut result = Vec::new();
    let mut in_plan = false;

    for line in lines {
        let t = line.trim();
        if t.is_empty() {
            continue;
        }
        if t.starts_with("Plan:")
            || t.starts_with("Apply complete")
            || t.starts_with("Destroy complete")
        {
            result.push(line.to_string());
            continue;
        }
        if t.starts_with("Error:") || t.starts_with("Warning:") {
            result.push(line.to_string());
            continue;
        }
        if t.starts_with("+ ") || t.starts_with("~ ") || t.starts_with("- ") || t.starts_with("/ ")
        {
            if !in_plan {
                result.push("[resource changes...]".to_string());
                in_plan = true;
            }
        } else {
            in_plan = false;
        }
        if result.len() < 40 {
            result.push(line.to_string());
        }
    }
    result.join("\n")
}

fn is_docker_compose_command(cmd: &str) -> bool {
    let t = cmd.trim();
    t.starts_with("docker-compose ") || t.starts_with("docker compose ")
}

fn compress_docker_compose(lines: &[&str]) -> String {
    // Docker compose: container status, logs
    let mut result = Vec::new();
    for line in lines {
        let t = line.trim();
        if t.is_empty() {
            continue;
        }
        if t.starts_with("Creating")
            || t.starts_with("Starting")
            || t.starts_with("Stopping")
            || t.starts_with("Removing")
            || t.starts_with("Building")
            || t.starts_with("Pulling")
        {
            result.push("[container ops...]".to_string());
        } else if t.starts_with("error")
            || t.starts_with("Error")
            || t.starts_with("FAIL")
            || result.len() < 30
        {
            result.push(line.to_string());
        }
    }
    result.join("\n")
}

fn is_build_command(cmd: &str) -> bool {
    let t = cmd.trim();
    t.starts_with("make ")
        || t.starts_with("ninja ")
        || t.starts_with("cmake ")
        || t.starts_with("bazel ")
        || t.starts_with("sbt ")
        || t.starts_with("mvn ")
        || t.starts_with("gradle ")
        || t.starts_with("cargo build")
        || t.starts_with("cargo test")
        || t.starts_with("go build")
        || t.starts_with("go test")
        || t.starts_with("dotnet build")
        || t.starts_with("dotnet test")
}

fn compress_build(lines: &[&str]) -> String {
    // Build systems: lots of "Compiling", "Building", "Linking" lines
    // Keep errors, warnings, final status
    let mut result = Vec::new();
    let mut in_compile = false;

    for line in lines {
        let t = line.trim();
        if t.is_empty() {
            continue;
        }
        let is_compile = t.starts_with("Compiling")
            || t.starts_with("Building")
            || t.starts_with("Linking")
            || t.starts_with("Generating")
            || t.starts_with("Running")
            || t.starts_with("CC ")
            || t.starts_with("CXX ")
            || t.starts_with("LD ")
            || t.starts_with("AR ")
            || t.starts_with("[")
                && (t.contains("%")
                    || t.contains("/") && t.chars().filter(|c| c.is_ascii_digit()).count() > 3);

        if is_compile {
            if !in_compile {
                result.push("[build steps...]".to_string());
                in_compile = true;
            }
        } else {
            in_compile = false;
            let is_important = t.starts_with("error")
                || t.starts_with("warning")
                || t.starts_with("FAIL")
                || t.starts_with("Error")
                || t.starts_with("Warning")
                || t.starts_with("Finished")
                || t.starts_with("Built")
                || t.starts_with("Build complete")
                || t.starts_with("Build failed")
                || t.contains("test result:");
            if is_important || result.len() < 50 {
                result.push(line.to_string());
            }
        }
    }
    result.join("\n")
}

fn compress_path_listing(lines: &[&str]) -> String {
    let paths: Vec<&str> = lines
        .iter()
        .map(|line| line.trim())
        .filter(|line| {
            !line.is_empty()
                && !line.ends_with(':')
                && !line.contains(" -> ")
                && (line.contains('/') || line.contains('\\'))
        })
        .collect();
    if paths.len() < 4 {
        return lines.join("\n");
    }

    let mut counts = std::collections::BTreeMap::<String, usize>::new();
    for path in paths {
        let normalized = path.replace('\\', "/");
        let dir = normalized
            .rsplit_once('/')
            .map(|(dir, _)| dir)
            .unwrap_or(".")
            .to_string();
        *counts.entry(dir).or_insert(0) += 1;
    }

    let total_files: usize = counts.values().sum();
    // Collapse to top-level directories (first path component) and show only the
    // busiest few — a full per-leaf-dir listing balloons on deep trees (e.g. a
    // `find .` that descends into target/). This keeps the high-signal shape.
    let mut top: std::collections::BTreeMap<String, usize> = std::collections::BTreeMap::new();
    for (dir, count) in &counts {
        let head = dir
            .trim_start_matches("./")
            .split('/')
            .next()
            .filter(|s| !s.is_empty())
            .unwrap_or(".")
            .to_string();
        *top.entry(head).or_insert(0) += count;
    }
    let mut ranked: Vec<(String, usize)> = top.into_iter().collect();
    ranked.sort_by(|a, b| b.1.cmp(&a.1).then(a.0.cmp(&b.0)));
    const MAX_DIRS: usize = 8;
    let mut out = vec![format!(
        "{} files across {} top-level dir(s):",
        total_files,
        ranked.len()
    )];
    for (dir, count) in ranked.iter().take(MAX_DIRS) {
        out.push(format!("{}/ ({})", dir, count));
    }
    if ranked.len() > MAX_DIRS {
        out.push(format!("... +{} more dir(s)", ranked.len() - MAX_DIRS));
    }
    out.join("\n")
}

fn is_cargo_metadata_command(cmd: &str) -> bool {
    cmd.contains("cargo metadata")
}

/// Summarize `cargo metadata` (a single multi-hundred-KB JSON blob) to the package
/// count + workspace members. The full transitive metadata is almost never the
/// signal an agent needs, and it otherwise passes through uncompressed.
fn compress_cargo_metadata(s: &str) -> String {
    let Ok(v) = serde_json::from_str::<serde_json::Value>(s.trim()) else {
        return s.to_string();
    };
    let n_pkgs = v
        .get("packages")
        .and_then(|p| p.as_array())
        .map(|a| a.len())
        .unwrap_or(0);
    let members: Vec<String> = v
        .get("workspace_members")
        .and_then(|m| m.as_array())
        .map(|a| {
            a.iter()
                .filter_map(|m| m.as_str())
                .map(|id| id.split([' ', '@']).next().unwrap_or(id).to_string())
                .collect()
        })
        .unwrap_or_default();
    let root = v
        .get("resolve")
        .and_then(|r| r.get("root"))
        .and_then(|r| r.as_str())
        .map(|id| id.split([' ', '@']).next().unwrap_or(id).to_string());
    let mut out = format!("cargo metadata: {n_pkgs} packages in the dependency graph");
    if !members.is_empty() {
        out.push_str(&format!("\nworkspace members: {}", members.join(", ")));
    }
    if let Some(root) = root {
        out.push_str(&format!("\nroot: {root}"));
    }
    out
}

fn is_cargo_tree_command(cmd: &str) -> bool {
    let t = cmd.trim();
    t == "cargo tree" || t.starts_with("cargo tree ")
}

/// Collapse `cargo tree` to its unique crate set. The tree repeats whole subtrees
/// (marked `(*)`) and draws box characters; the useful signal is which crates are
/// in the graph.
fn compress_cargo_tree(lines: &[&str]) -> String {
    let mut crates: std::collections::BTreeSet<String> = std::collections::BTreeSet::new();
    for line in lines {
        // Strip leading tree-drawing characters and whitespace.
        let stripped = line.trim_start_matches([
            ' ', '|', '`', '+', '-', '\u{2502}', '\u{251c}', '\u{2514}', '\u{2500}',
        ]);
        let stripped = stripped.trim();
        if stripped.is_empty() {
            continue;
        }
        // A crate line looks like "name v1.2.3" or "name v1.2.3 (proc-macro)".
        let mut it = stripped.split_whitespace();
        if let (Some(name), Some(ver)) = (it.next(), it.next()) {
            if ver.starts_with('v') && name.chars().next().is_some_and(|c| c.is_ascii_alphabetic())
            {
                crates.insert(format!("{name} {ver}"));
            }
        }
    }
    if crates.is_empty() {
        return lines.join("\n");
    }
    const MAX_SHOWN: usize = 15;
    let total = crates.len();
    let shown: Vec<String> = crates.into_iter().take(MAX_SHOWN).collect();
    let suffix = if total > MAX_SHOWN {
        format!(" (+{} more)", total - MAX_SHOWN)
    } else {
        String::new()
    };
    format!(
        "cargo tree: {} unique crates\n{}{}",
        total,
        shown.join(", "),
        suffix
    )
}

fn is_ps_command(cmd: &str) -> bool {
    let t = cmd.trim();
    t == "ps" || t.starts_with("ps ")
}

/// Compact `ps` output to the header + the busiest processes by %CPU. A raw
/// `ps aux` is hundreds of lines dominated by idle kernel threads; the signal is
/// what is actually consuming the machine.
fn compress_ps(lines: &[&str]) -> String {
    const TOP: usize = 4;
    const WIDTH: usize = 85;
    let trunc = |l: &str| -> String {
        if l.chars().count() > WIDTH {
            format!("{}…", l.chars().take(WIDTH).collect::<String>())
        } else {
            l.to_string()
        }
    };
    let nonempty: Vec<&str> = lines
        .iter()
        .copied()
        .filter(|l| !l.trim().is_empty())
        .collect();
    if nonempty.len() <= TOP + 1 {
        return nonempty
            .iter()
            .map(|l| trunc(l))
            .collect::<Vec<_>>()
            .join("\n");
    }
    let header = nonempty[0];
    // Locate the CPU column from the header instead of assuming index 2: that
    // holds for `ps aux` (USER PID %CPU …) but not for `ps -ef` (UID PID PPID C
    // …) or a custom `ps -eo`, where index 2 is PPID and the "sorted by %CPU"
    // claim was simply false.
    let cpu_col = header
        .split_whitespace()
        .position(|h| h.eq_ignore_ascii_case("%cpu"))
        .or_else(|| {
            header
                .split_whitespace()
                .position(|h| h.eq_ignore_ascii_case("c"))
        });
    let mut rows: Vec<&str> = nonempty[1..].to_vec();
    let sorted = match cpu_col {
        Some(col) => {
            let cpu_of = |l: &str| -> f32 {
                l.split_whitespace()
                    .nth(col)
                    .and_then(|c| c.parse::<f32>().ok())
                    .unwrap_or(0.0)
            };
            rows.sort_by(|a, b| {
                cpu_of(b)
                    .partial_cmp(&cpu_of(a))
                    .unwrap_or(std::cmp::Ordering::Equal)
            });
            true
        }
        None => false,
    };
    let mut out = vec![trunc(header)];
    for r in rows.iter().take(TOP) {
        out.push(trunc(r));
    }
    out.push(format!(
        "... {} more process(es) ({}, top {} shown)",
        rows.len() - TOP,
        if sorted {
            "sorted by CPU"
        } else {
            "original order"
        },
        TOP
    ));
    out.join("\n")
}

fn is_grep_command(cmd: &str) -> bool {
    let t = cmd.trim();
    // Plain grep only — not `git grep` (handled elsewhere) or ripgrep (rg.toml).
    (t == "grep" || t.starts_with("grep ")) && !t.starts_with("grep -V")
}

/// Compact plain grep output: drop blank lines, strip the indentation inside each
/// `path:line:CONTENT` match, and cap the number of lines.
fn compress_grep(lines: &[&str]) -> String {
    const MAX_MATCHES: usize = 50;
    let mut out: Vec<String> = Vec::new();
    let mut shown = 0usize;
    for line in lines {
        if line.trim().is_empty() {
            continue;
        }
        let colons: Vec<usize> = line.match_indices(':').map(|(i, _)| i).collect();
        let mut split_idx = None;
        for &idx in &colons {
            let prev_idx = colons.iter().rev().copied().find(|&p| p < idx);
            let start = prev_idx.map(|p| p + 1).unwrap_or(0);
            let part = &line[start..idx];
            if !part.is_empty() && part.chars().all(|c| c.is_ascii_digit()) {
                split_idx = Some(idx);
                break;
            }
        }
        let split_idx = split_idx.or_else(|| colons.last().copied());

        let compact = match split_idx {
            Some(idx) => {
                let (head, content) = line.split_at(idx + 1);
                format!("{head}{}", content.trim())
            }
            None => line.trim().to_string(),
        };
        if shown >= MAX_MATCHES {
            out.push(format!("... +{} more match line(s)", lines.len() - shown));
            break;
        }
        out.push(compact);
        shown += 1;
    }
    if out.is_empty() {
        return "grep: no matches".to_string();
    }
    out.join("\n")
}

fn truncate_head_tail(lines: &[&str], head: usize, tail: usize) -> String {
    let total = lines.len();
    if total <= head + tail {
        return lines.join("\n");
    }
    let omitted = total - head - tail;
    format!(
        "{}\n[... {} lines omitted ...]\n{}",
        lines[..head].join("\n"),
        omitted,
        lines[total - tail..].join("\n")
    )
}

/// Minimum length of an unbroken base64 run before it is treated as an embedded
/// blob and redacted. Prose, code, hashes, and even minified JS break the base64
/// alphabet with whitespace or punctuation long before this length, so a run this
/// long is effectively always a real payload (embedded image, tar/binary dump).
const BASE64_MIN: usize = 512;

/// Label a base64 run by its decoded magic bytes so the redaction says what was
/// dropped (`png image base64`) instead of an opaque `base64 blob`. Matching is on
/// the base64-encoded prefix — the first few base64 chars deterministically encode
/// the leading file-magic bytes.
fn base64_blob_kind(run: &[u8]) -> &'static str {
    let head = std::str::from_utf8(&run[..run.len().min(16)]).unwrap_or("");
    if head.starts_with("iVBORw0KGgo") {
        "png image base64"
    } else if head.starts_with("/9j/") {
        "jpeg image base64"
    } else if head.starts_with("R0lGOD") {
        "gif image base64"
    } else if head.starts_with("UklGR") {
        "webp/riff base64"
    } else if head.starts_with("JVBERi0") {
        "pdf base64"
    } else if head.starts_with("H4sI") {
        "gzip base64"
    } else {
        "base64 blob"
    }
}

/// A long run of base64-alphabet bytes is only a real blob if it looks like
/// encoded binary — mixed-case letters, or a base64 symbol (`+ / = - _`). Pure-hex
/// or all-digit runs (checksum manifests, id lists, hex dumps) are spared: real
/// base64 of binary is mixed-case/symbol-rich with overwhelming probability, so
/// this avoids silently eating a list of sha256 hashes or a numeric column.
fn looks_like_base64(run: &[u8]) -> bool {
    // `-` and `_` are base64url characters, but they are also how humans join
    // words: a long kebab/snake identifier chain (`build-step-1-step-2-...`,
    // a joined list of slugs) hit this and was silently redacted. Real base64
    // payloads do not decompose into many short separator-delimited groups, so
    // reject that shape before anything else.
    let separator_groups = run
        .split(|b| matches!(b, b'-' | b'_'))
        .filter(|g| !g.is_empty())
        .count();
    if separator_groups >= 4 {
        let avg_group = run.len() / separator_groups;
        if avg_group < 12 {
            return false;
        }
    }

    let mut has_lower = false;
    let mut has_upper = false;
    for &b in run {
        match b {
            // `+ / =` cannot appear in an identifier, so they are proof on their
            // own. `-` and `_` are NOT: treating them as proof made the very
            // first hyphen of a lowercase kebab chain end the scan with "base64"
            // before the mixed-case test could spare it.
            b'+' | b'/' | b'=' => return true,
            b'a'..=b'z' => has_lower = true,
            b'A'..=b'Z' => has_upper = true,
            _ => {}
        }
        if has_lower && has_upper {
            return true;
        }
    }
    false
}

/// Redact long base64 / data-URI blobs (embedded PNGs, tar/binary dumps) that
/// otherwise replay tens of thousands of tokens of pure noise into context. A
/// `data:<mime>;base64,` prefix is preserved (its `:`/`;`/`,` break the run), so
/// only the payload is dropped. Operates on bytes but only ever removes ASCII
/// base64 runs, so the result stays valid UTF-8.
fn strip_base64_blobs(s: &str) -> String {
    if s.len() < BASE64_MIN {
        return s.to_string();
    }
    let is_b64 = |b: u8| b.is_ascii_alphanumeric() || matches!(b, b'+' | b'/' | b'=' | b'-' | b'_');
    let bytes = s.as_bytes();
    let mut out: Vec<u8> = Vec::with_capacity(bytes.len());
    let mut i = 0;
    let mut changed = false;
    while i < bytes.len() {
        if is_b64(bytes[i]) {
            let start = i;
            while i < bytes.len() && is_b64(bytes[i]) {
                i += 1;
            }
            let run = i - start;
            let slice = &bytes[start..i];
            if run >= BASE64_MIN && looks_like_base64(slice) {
                let kind = base64_blob_kind(slice);
                out.extend_from_slice(format!("[{kind} omitted: {run} chars]").as_bytes());
                changed = true;
            } else {
                out.extend_from_slice(slice);
            }
        } else {
            out.push(bytes[i]);
            i += 1;
        }
    }
    if !changed {
        return s.to_string();
    }
    String::from_utf8(out).unwrap_or_else(|_| s.to_string())
}

/// Minimum encoded width of an interior line and minimum consecutive-line count for
/// a *wrapped* base64 block (MIME/PEM/`base64`/`openssl` wrap at 64–76 cols, which
/// the contiguous scanner misses because each line is under `BASE64_MIN`).
const WRAP_MIN_LINE: usize = 60;
const WRAP_MIN_LINES: usize = 5;

/// Collapse line-wrapped base64 blocks (PEM certs/keys, `base64 file`, binary git
/// patches) that `strip_base64_blobs` cannot see because the newline every ~76
/// chars keeps each run under `BASE64_MIN`. A block is ≥`WRAP_MIN_LINES` consecutive
/// pure-base64 lines of width ≥`WRAP_MIN_LINE` (plus one optional shorter padded
/// tail line). PEM `-----BEGIN/END-----` markers contain spaces/dashes-with-spaces
/// so they are not base64 lines and survive, keeping the block's context.
fn strip_wrapped_base64(s: &str) -> String {
    if s.len() < BASE64_MIN {
        return s.to_string();
    }
    let is_b64_line = |t: &str| {
        !t.is_empty()
            && t.bytes()
                .all(|b| b.is_ascii_alphanumeric() || matches!(b, b'+' | b'/' | b'=' | b'-' | b'_'))
    };
    let trailing = s.ends_with('\n');
    let lines: Vec<&str> = s.trim_end_matches('\n').split('\n').collect();
    let mut out: Vec<String> = Vec::with_capacity(lines.len());
    let mut i = 0;
    let mut changed = false;
    while i < lines.len() {
        let t = lines[i].trim();
        if is_b64_line(t) && t.len() >= WRAP_MIN_LINE {
            let start = i;
            let mut chars = 0usize;
            while i < lines.len()
                && is_b64_line(lines[i].trim())
                && lines[i].trim().len() >= WRAP_MIN_LINE
            {
                chars += lines[i].trim().len();
                i += 1;
            }
            // The final padded chunk of a real block is a shorter line, but a normal
            // all-alnum word (`done`, `trailer`) also qualifies — absorbing it would
            // eat legit text. Leave it: a single ≤76-char line is negligible tokens.
            let count = i - start;
            if count >= WRAP_MIN_LINES
                && chars >= BASE64_MIN
                && looks_like_base64(lines[start].trim().as_bytes())
            {
                let kind = base64_blob_kind(lines[start].trim().as_bytes());
                out.push(format!(
                    "[{kind} omitted: {chars} chars across {count} lines]"
                ));
                changed = true;
            } else {
                for line in &lines[start..i] {
                    out.push((*line).to_string());
                }
            }
        } else {
            out.push(lines[i].to_string());
            i += 1;
        }
    }
    if !changed {
        return s.to_string();
    }
    let mut joined = out.join("\n");
    if trailing {
        joined.push('\n');
    }
    joined
}

/// Redact both single-line/data-URI and line-wrapped base64 blobs.
fn redact_base64_blobs(s: &str) -> String {
    strip_wrapped_base64(&strip_base64_blobs(s))
}

/// The line terminator that dominates `s`.
///
/// Every `\r\n` also contains a `\n`, so CRLF dominates when it accounts for
/// more than half of all newlines.
pub(crate) fn dominant_eol(s: &str) -> &'static str {
    let crlf = s.matches("\r\n").count();
    let lf = s.matches('\n').count();
    if crlf > 0 && crlf * 2 > lf {
        "\r\n"
    } else {
        "\n"
    }
}

/// Restore `eol` over output that the line pipeline has normalized to LF.
///
/// Every transform in this crate splits with `.lines()` — which consumes `\r\n`
/// and `\n` alike — and rejoins with `"\n"`. On a CRLF input that silently hands
/// the agent a copy where *every* line ending differs from the bytes on disk. If
/// the agent then quotes those lines into an exact-match edit (SEARCH/REPLACE,
/// `str_replace`), the anchor cannot match: measured elsewhere as successful
/// patch application dropping from 27/40 to 15/40 when compression rewrote the
/// bytes an edit tool has to reproduce. Windows is this project's primary
/// platform, so this is the common case, not the corner case.
pub(crate) fn restore_eol(out: String, eol: &str) -> String {
    if eol != "\r\n" || out.is_empty() {
        return out;
    }
    // Normalize first so a partially-CRLF result cannot become `\r\r\n`.
    out.replace("\r\n", "\n").replace('\n', "\r\n")
}

pub fn compress_output(s: &str) -> String {
    let eol = dominant_eol(s);
    restore_eol(compress_output_inner(s), eol)
}

fn compress_output_inner(s: &str) -> String {
    // Redact embedded base64 blobs first: a data URI inside a JSON string would
    // otherwise survive JSON compaction at full weight.
    let stripped = redact_base64_blobs(s);
    let s: &str = &stripped;

    // JSON compaction first: if output is pure JSON or NDJSON, compact and return early.
    // The other transforms (ANSI, emoji, blank lines) don't apply to JSON.
    let compacted = compact_json(s);
    if compacted != s {
        return enforce_token_budget(&compacted);
    }
    let s = strip_ansi(s);
    let s = remove_emojis(&s);
    let s = collapse_blank_lines(&s);
    let s = group_repeated_blocks(&s);
    let s = group_repeated_lines(&s);

    // Additional generic aggressive compression
    enforce_token_budget(&generic_aggressive_compress(&s))
}

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

    #[test]
    fn crlf_survives_the_compression_pipeline() {
        // Regression: `.lines()` + `join("\n")` laundered CRLF into LF, so every
        // line the agent got back differed from disk by one byte.
        let src = "fn main() {\r\n    println!(\"hi\");\r\n}\r\n";
        let out = compress_output(src);
        assert!(
            out.contains("\r\n"),
            "CRLF input must come back CRLF, got {out:?}"
        );
        assert!(
            !out.replace("\r\n", "").contains('\n'),
            "no bare LF may survive in a CRLF payload: {out:?}"
        );
    }

    #[test]
    fn lf_input_stays_lf() {
        let src = "fn main() {\n    println!(\"hi\");\n}\n";
        assert!(!compress_output(src).contains('\r'));
    }

    #[test]
    fn dominant_eol_picks_the_majority() {
        assert_eq!(dominant_eol("a\r\nb\r\nc\r\n"), "\r\n");
        assert_eq!(dominant_eol("a\nb\nc\n"), "\n");
        assert_eq!(dominant_eol(""), "\n");
        // Mixed endings: a lone CRLF in a mostly-LF payload must not flip it.
        assert_eq!(dominant_eol("a\nb\nc\r\nd\n"), "\n");
    }

    #[test]
    fn restore_eol_is_idempotent() {
        let once = restore_eol("a\nb\n".to_string(), "\r\n");
        assert_eq!(once, "a\r\nb\r\n");
        assert_eq!(restore_eol(once.clone(), "\r\n"), once);
    }
}

/// Absolute ceiling on how many tokens any single compressed output may cost.
/// The line-based caps (`generic_aggressive_compress`) cannot bound output that
/// is few-but-enormous lines (a one-line JSON payload, a minified bundle, a grep
/// hit with megabyte-long lines), and the `compact_json` early return bypasses
/// them entirely — a 3 MB tool result compacted by 10% still costs ~700k tokens.
/// Override with `TOKENIX_MAX_OUTPUT_TOKENS`; `0` disables the cap.
const DEFAULT_MAX_OUTPUT_TOKENS: usize = 8000;

fn max_output_tokens() -> usize {
    std::env::var("TOKENIX_MAX_OUTPUT_TOKENS")
        .ok()
        .and_then(|v| v.trim().parse::<usize>().ok())
        .unwrap_or(DEFAULT_MAX_OUTPUT_TOKENS)
}

/// Clip `s` to the global token budget, keeping a head and a tail window so both
/// the command's opening context and its (usually load-bearing) final lines
/// survive. Char boundaries are respected, so UTF-8 output stays valid.
fn enforce_token_budget(s: &str) -> String {
    let budget = max_output_tokens();
    if budget == 0 || s.is_empty() {
        return s.to_string();
    }
    let tokens = count_tokens(s);
    if tokens <= budget {
        return s.to_string();
    }

    // Convert the token budget into a char budget. `count_tokens` is chars/4
    // rounded up, so this ratio is ~4 by construction — it is written as a
    // division rather than a constant so the conversion stays correct if
    // `count_tokens` ever becomes a real BPE count (roadmap).
    let char_count = s.chars().count();
    let chars_per_token = (char_count as f64 / tokens as f64).max(1.0);
    let char_budget = (budget as f64 * chars_per_token) as usize;
    if char_budget >= char_count {
        return s.to_string();
    }
    let head_chars = char_budget * 3 / 4;
    let tail_chars = char_budget.saturating_sub(head_chars);

    let head_end = byte_index_at_char(s, head_chars);
    let tail_start = byte_index_at_char(s, char_count.saturating_sub(tail_chars));
    if tail_start <= head_end {
        return s.to_string();
    }

    let omitted = tokens.saturating_sub(budget);
    let facts = match preserved_identifiers(&s[head_end..tail_start]) {
        Some(ids) => format!("\n[ids kept from the omitted range: {ids}]"),
        None => String::new(),
    };
    format!(
        "{}\n[... {} tokens omitted by tokenix output cap ({} max; set TOKENIX_MAX_OUTPUT_TOKENS to change) ...]{}\n{}",
        &s[..head_end],
        omitted,
        budget,
        facts,
        &s[tail_start..]
    )
}

/// Identifiers a caller cannot reconstruct or guess: commit SHAs, UUIDs, URLs,
/// and compiler/error codes. Truncating a range that contained the one SHA the
/// agent needed forces a re-run, so they are carried over into the marker.
/// Everything else (prose, repeated log lines) is genuinely re-derivable and is
/// left dropped.
fn preserved_identifiers(omitted: &str) -> Option<String> {
    const MAX_FACTS: usize = 12;
    const MAX_CHARS: usize = 240;

    let mut facts: Vec<&str> = Vec::new();
    for token in omitted.split(|c: char| {
        c.is_whitespace() || matches!(c, '"' | '\'' | ',' | '(' | ')' | '[' | ']' | '{' | '}')
    }) {
        let token = token.trim_matches(|c: char| matches!(c, '.' | ':' | ';' | '`'));
        if token.len() < 5 || facts.len() >= MAX_FACTS || facts.contains(&token) {
            continue;
        }
        if is_identifier_like(token) {
            facts.push(token);
        }
    }
    if facts.is_empty() {
        return None;
    }
    let mut out = String::new();
    for fact in facts {
        if out.len() + fact.len() + 2 > MAX_CHARS {
            break;
        }
        if !out.is_empty() {
            out.push_str(", ");
        }
        out.push_str(fact);
    }
    (!out.is_empty()).then_some(out)
}

fn is_identifier_like(token: &str) -> bool {
    if token.starts_with("http://") || token.starts_with("https://") {
        return true;
    }
    // UUID: 8-4-4-4-12 hex groups.
    let groups: Vec<&str> = token.split('-').collect();
    if groups.len() == 5
        && [8, 4, 4, 4, 12]
            == [
                groups[0].len(),
                groups[1].len(),
                groups[2].len(),
                groups[3].len(),
                groups[4].len(),
            ]
        && groups
            .iter()
            .all(|g| g.chars().all(|c| c.is_ascii_hexdigit()))
    {
        return true;
    }
    // Git-style SHA: 7-40 hex chars, mixed letters and digits so plain numbers
    // and English words ("added", "decade") are not mistaken for one.
    let is_hex = token.len() >= 7
        && token.len() <= 40
        && token.chars().all(|c| c.is_ascii_hexdigit())
        && token.chars().any(|c| c.is_ascii_digit())
        && token.chars().any(|c| c.is_ascii_alphabetic());
    if is_hex {
        return true;
    }
    // Compiler/runtime error codes: E0308, ERR_MODULE_NOT_FOUND, TS2345.
    let code_like = token.chars().next().is_some_and(|c| c.is_ascii_uppercase())
        && token.chars().any(|c| c.is_ascii_digit())
        && token
            .chars()
            .all(|c| c.is_ascii_uppercase() || c.is_ascii_digit() || c == '_');
    code_like
}

fn byte_index_at_char(s: &str, char_idx: usize) -> usize {
    s.char_indices()
        .nth(char_idx)
        .map(|(i, _)| i)
        .unwrap_or(s.len())
}

/// Longest block a repeat scan will consider. Loop bodies that spam a context
/// window (a PowerShell exception block, a retry banner, a k8s poll report) are
/// typically a handful of lines; scanning wider costs time for no real gain.
const MAX_REPEAT_BLOCK_LINES: usize = 12;

/// Collapse consecutive repetitions of a multi-line *block*.
///
/// `group_repeated_lines` only sees runs of identical adjacent lines, so a
/// watch/poll loop that re-emits the same 8-line stanza hundreds of times slips
/// through untouched (measured: a single monitoring session cost ~617k tokens
/// this way). A block repeated 3+ times is kept once, annotated with the count.
fn group_repeated_blocks(s: &str) -> String {
    let trailing_newline = s.ends_with('\n');
    let source = if trailing_newline {
        &s[..s.len() - 1]
    } else {
        s
    };
    let lines: Vec<&str> = source.split('\n').collect();
    // Below 6 lines there is nothing a 3x block repeat can be built from; above
    // the ceiling the quadratic-ish scan is not worth it (the token cap and the
    // line caps already bound outputs that large).
    if lines.len() < 6 || lines.len() > 100_000 {
        return s.to_string();
    }

    let mut result = String::with_capacity(s.len());
    let mut i = 0;
    while i < lines.len() {
        let mut collapsed = false;
        // Prefer the smallest repeating unit: a 2-line stanza repeated 6x should
        // report 6, not 3 repeats of a 4-line block.
        for width in 2..=MAX_REPEAT_BLOCK_LINES.min((lines.len() - i) / 2) {
            let block = &lines[i..i + width];
            // A block of identical lines is a run, not a stanza: leave it to
            // `group_repeated_lines`, which reports it as one line + a count
            // instead of an arbitrary N-line window.
            if block.iter().all(|l| *l == block[0]) {
                continue;
            }
            let mut reps = 1;
            while i + (reps + 1) * width <= lines.len()
                && &lines[i + reps * width..i + (reps + 1) * width] == block
            {
                reps += 1;
            }
            if reps >= 3 {
                for line in block {
                    result.push_str(line);
                    result.push('\n');
                }
                result.push_str(&format!(
                    "[block of {} lines repeated {}x]\n",
                    width,
                    reps - 1
                ));
                i += reps * width;
                collapsed = true;
                break;
            }
        }
        if !collapsed {
            result.push_str(lines[i]);
            result.push('\n');
            i += 1;
        }
    }
    if !trailing_newline && result.ends_with('\n') {
        result.pop();
    }
    result
}

fn generic_aggressive_compress(s: &str) -> String {
    let lines: Vec<&str> = s.lines().collect();
    if lines.len() <= 50 {
        return s.to_string();
    }

    // Detect and compress common patterns
    let mut result: Vec<String> = Vec::new();
    let mut i = 0;

    while i < lines.len() {
        let line = lines[i];

        // Skip progress bars and spinners
        if line.contains("█")
            || line.contains("░")
            || line.contains("▓")
            || line
                .chars()
                .filter(|c| *c == '=' || *c == '>' || *c == '#')
                .count()
                > 15
        {
            if result.is_empty() || !result.last().unwrap().contains("progress") {
                result.push("[progress bar omitted]".to_string());
            }
            i += 1;
            continue;
        }

        // Collapse repeated similar lines (download, extracting, etc.)
        let prefixes = [
            "Downloading ",
            "Extracting ",
            "Installing ",
            "Fetching ",
            "Resolving ",
            "Building ",
            "Compiling ",
            "Generating ",
            "Running ",
            "Testing ",
            "Checking ",
            "Verifying ",
        ];
        let mut matched_prefix = None;
        for prefix in &prefixes {
            if line.starts_with(prefix) {
                matched_prefix = Some(*prefix);
                break;
            }
        }
        if let Some(prefix) = matched_prefix {
            let mut count = 1;
            let mut j = i + 1;
            while j < lines.len() && lines[j].starts_with(prefix) && count < 100 {
                count += 1;
                j += 1;
            }
            if count >= 3 {
                result.push(format!("{} ({} similar lines)", line, count - 1));
                i = j;
                continue;
            }
        }

        result.push(line.to_string());
        i += 1;
    }

    // If still too long, apply head/tail truncation
    if result.len() > 100 {
        const HEAD: usize = 40;
        const TAIL: usize = 20;
        if result.len() > HEAD + TAIL {
            let omitted = result.len() - HEAD - TAIL;
            let mut truncated = result[..HEAD].to_vec();
            truncated.push(format!("[... {} lines omitted ...]", omitted));
            truncated.extend(result[result.len() - TAIL..].to_vec());
            return truncated.join("\n");
        }
    }

    result.join("\n")
}

/// Compact pretty-printed JSON (pure JSON or NDJSON) into single-line form.
/// Returns the original string unchanged if not JSON or if already compact.
fn compact_json(s: &str) -> String {
    let trimmed = s.trim();

    // Case 0: too short to be meaningful JSON
    if trimmed.len() < 2 {
        return s.to_string();
    }

    // Case 1: entire output is a JSON object or array
    if (trimmed.starts_with('{') || trimmed.starts_with('['))
        && serde_json::from_str::<serde::de::IgnoredAny>(trimmed).is_ok()
    {
        let compact = strip_json_whitespace(trimmed);
        if compact.len() < trimmed.len() {
            return if s.ends_with('\n') {
                compact + "\n"
            } else {
                compact
            };
        }
    }

    // Case 2: NDJSON — every non-empty line is a JSON object or array
    let lines: Vec<&str> = trimmed.lines().collect();
    if lines.len() > 1
        && lines.iter().all(|l| {
            let t = l.trim();
            t.is_empty()
                || (t.starts_with('{') && serde_json::from_str::<serde_json::Value>(t).is_ok())
                || (t.starts_with('[') && serde_json::from_str::<serde_json::Value>(t).is_ok())
        })
    {
        let compacted: String = lines
            .iter()
            .filter_map(|l| {
                let t = l.trim();
                if t.is_empty() {
                    return None;
                }
                Some(strip_json_whitespace(t))
            })
            .collect::<Vec<_>>()
            .join("\n");
        let result = if s.ends_with('\n') {
            compacted + "\n"
        } else {
            compacted
        };
        if result.len() < s.len() {
            return result;
        }
    }

    s.to_string()
}

/// Drop insignificant whitespace from *already-validated* JSON text, preserving
/// the bytes that matter. Re-serializing through `serde_json::Value` would be
/// lossy: integers past `u64` degrade to floats in scientific notation,
/// duplicate keys collapse, and key order is rewritten.
fn strip_json_whitespace(json: &str) -> String {
    let mut out = String::with_capacity(json.len());
    let mut in_string = false;
    let mut escaped = false;
    for c in json.chars() {
        if in_string {
            out.push(c);
            if escaped {
                escaped = false;
            } else if c == '\\' {
                escaped = true;
            } else if c == '"' {
                in_string = false;
            }
            continue;
        }
        match c {
            '"' => {
                in_string = true;
                out.push(c);
            }
            c if c.is_ascii_whitespace() => {}
            _ => out.push(c),
        }
    }
    out
}

fn compress_command_streams(
    command_str: &str,
    stdout_raw: &str,
    stderr_raw: &str,
    success: bool,
) -> (String, String) {
    let stdout_compressed = if stdout_raw.trim().is_empty() {
        if stderr_raw.trim().is_empty() && success {
            compress_bash_output_for_stream(command_str, stdout_raw, false, Some(success))
        } else {
            String::new()
        }
    } else {
        compress_bash_output_for_stream(command_str, stdout_raw, false, Some(success))
    };

    let stderr_compressed = if stderr_raw.trim().is_empty() {
        String::new()
    } else {
        compress_bash_output_for_stream(command_str, stderr_raw, true, Some(success))
    };

    (
        keep_final_newline(stdout_raw, stdout_compressed),
        keep_final_newline(stderr_raw, stderr_compressed),
    )
}

/// Filters rejoin lines without a trailing terminator. When the agent's
/// harness merges both streams, a missing one glues this stream's last line
/// onto the other's first (`[... omitted ...]test result: FAILED`).
fn keep_final_newline(raw: &str, mut out: String) -> String {
    if !out.is_empty() && raw.ends_with('\n') && !out.ends_with('\n') {
        out.push_str(if raw.ends_with("\r\n") { "\r\n" } else { "\n" });
    }
    out
}

/// Remove ANSI/VT100 escape sequences (CSI, OSC, and single-char sequences).
pub(crate) fn strip_ansi(s: &str) -> String {
    let bytes = s.as_bytes();
    let mut result: Vec<u8> = Vec::with_capacity(bytes.len());
    let mut i = 0;
    while i < bytes.len() {
        if bytes[i] != 0x1b {
            result.push(bytes[i]);
            i += 1;
            continue;
        }
        i += 1;
        if i >= bytes.len() {
            break;
        }
        match bytes[i] {
            b'[' => {
                i += 1;
                // CSI: skip until final byte (0x40–0x7E)
                while i < bytes.len() {
                    let b = bytes[i];
                    i += 1;
                    if (0x40..=0x7E).contains(&b) {
                        break;
                    }
                }
            }
            b']' => {
                i += 1;
                // OSC: skip until BEL or ST (ESC \)
                while i < bytes.len() {
                    if bytes[i] == 0x07 {
                        i += 1;
                        break;
                    }
                    if bytes[i] == 0x1b && i + 1 < bytes.len() && bytes[i + 1] == b'\\' {
                        i += 2;
                        break;
                    }
                    i += 1;
                }
            }
            b if b.is_ascii() => {
                i += 1; // single-char sequence: ESC + one ASCII byte
            }
            _ => {
                // ESC followed by a non-ASCII byte is not an escape sequence.
                // Consuming it would split a multi-byte character and make the
                // whole result invalid UTF-8 — which used to blank the entire
                // output via `unwrap_or_default()`. Drop the stray ESC only.
            }
        }
    }
    // ANSI sequences are pure ASCII; remaining bytes are still valid UTF-8.
    // Fall back to lossy rather than an empty string: losing one character is
    // recoverable, losing the whole command output is not.
    match String::from_utf8(result) {
        Ok(s) => s,
        Err(e) => String::from_utf8_lossy(e.as_bytes()).into_owned(),
    }
}

/// Remove emoji characters by unicode code-point range.
fn remove_emojis(s: &str) -> String {
    s.chars().filter(|&c| !is_emoji_char(c)).collect()
}

fn is_emoji_char(c: char) -> bool {
    matches!(c,
        '\u{1F000}'..='\u{1FFFF}' // Emoticons, misc symbols and pictographs, transport, etc.
        | '\u{2600}'..='\u{26FF}' // Misc symbols (☀☁⚡ etc.)
        | '\u{2700}'..='\u{27BF}' // Dingbats (✈✉✔ etc.)
        | '\u{FE00}'..='\u{FE0F}' // Variation selectors (emoji presentation)
        | '\u{200D}'              // Zero-width joiner (emoji combiner)
        | '\u{20E3}'              // Combining enclosing keycap
    )
}

/// Collapse 3+ consecutive newlines down to 2 (one blank line between paragraphs).
fn collapse_blank_lines(s: &str) -> String {
    let mut result = String::with_capacity(s.len());
    let mut newline_run = 0usize;
    for c in s.chars() {
        if c == '\n' {
            newline_run += 1;
            if newline_run <= 2 {
                result.push('\n');
            }
        } else {
            newline_run = 0;
            result.push(c);
        }
    }
    result
}

/// Group consecutive identical lines that appear 3+ times into one line + annotation.
/// Lines appearing 1–2 times in a row are left unchanged.
/// Also performs fuzzy grouping for common patterns (e.g., progress bars, file listings).
fn group_repeated_lines(s: &str) -> String {
    let trailing_newline = s.ends_with('\n');
    let source = if trailing_newline {
        &s[..s.len() - 1]
    } else {
        s
    };
    let lines: Vec<&str> = source.split('\n').collect();
    let mut result = String::with_capacity(s.len());
    let mut i = 0;
    while i < lines.len() {
        let line = lines[i];

        // 1. Exact match grouping
        let mut end = i + 1;
        while end < lines.len() && lines[end] == line {
            end += 1;
        }
        let count = end - i;
        if count >= 3 {
            result.push_str(line);
            result.push('\n');
            result.push_str(&format!("[repeated {}x]\n", count - 1));
            i = end;
            continue;
        }

        // 2. Fuzzy grouping (similarity)
        if let Some(fuzzy_count) = try_fuzzy_group(&lines, i) {
            if fuzzy_count >= 3 {
                result.push_str(line);
                result.push_str(" ... (and ");
                result.push_str(&(fuzzy_count - 1).to_string());
                result.push_str(" similar lines)\n");
                i += fuzzy_count;
                continue;
            }
        }

        result.push_str(line);
        result.push('\n');
        i += 1;
    }
    if !trailing_newline && result.ends_with('\n') {
        result.pop();
    }
    result
}

fn try_fuzzy_group(lines: &[&str], start: usize) -> Option<usize> {
    let line = lines[start];
    if line.len() < 5 {
        return None;
    }

    // Patterns for fuzzy grouping:
    let prefixes = [
        "Removing ",
        "Compiling ",
        "Installing ",
        "Download ",
        "Extracting ",
        "Checked ",
        "test ",
    ];

    for prefix in prefixes {
        if line.starts_with(prefix) {
            if prefix == "test " && !line.contains(" ... ok") {
                continue;
            }
            let mut count = 1;
            for next_line in lines.iter().skip(start + 1) {
                if next_line.starts_with(prefix) {
                    count += 1;
                } else {
                    break;
                }
            }
            if count >= 3 {
                return Some(count);
            }
        }
    }
    None
}

fn find_repo_root() -> PathBuf {
    let cwd = std::env::current_dir().unwrap_or_else(|_| PathBuf::from("."));
    crate::store::find_project_root(&cwd)
}

pub fn now_ts() -> f64 {
    SystemTime::now()
        .duration_since(UNIX_EPOCH)
        .unwrap()
        .as_secs_f64()
}

/// Extract plain text from a PostToolUse tool_response value.
/// Handles: plain string, {"output": "..."}, and content-array format.
fn extract_response_text(response: &serde_json::Value) -> Option<String> {
    if let Some(s) = response.as_str() {
        return Some(s.to_string());
    }
    if let Some(s) = response["output"].as_str() {
        return Some(s.to_string());
    }
    // Claude Code Bash tool_response: { stdout, stderr, interrupted, ... }.
    let stdout = response["stdout"].as_str().unwrap_or("");
    let stderr = response["stderr"].as_str().unwrap_or("");
    if !stdout.is_empty() || !stderr.is_empty() {
        let mut combined = stdout.to_string();
        if !stderr.is_empty() {
            if !combined.is_empty() {
                combined.push('\n');
            }
            combined.push_str(stderr);
        }
        return Some(combined);
    }
    if let Some(arr) = response["content"].as_array() {
        let text: String = arr
            .iter()
            .filter_map(|item| {
                if item["type"].as_str() == Some("text") {
                    item["text"].as_str().map(str::to_string)
                } else {
                    None
                }
            })
            .collect::<Vec<_>>()
            .join("\n");
        if !text.is_empty() {
            return Some(text);
        }
    }
    None
}

/// Output dialect for a PostToolUse hook, selected by which agent invoked it.
#[derive(Debug, PartialEq)]
enum PostDialect {
    /// Claude Code / Codex: PostToolUse cannot replace or shorten a tool result,
    /// so compression here is a no-op — exit 0 silently without logging savings.
    ClaudeNoop,
    /// GitHub Copilot CLI: print `{"modifiedResult":{...}}` JSON on stdout, exit 0.
    CopilotJson,
}

/// A PostToolUse payload normalized across Claude Code and Copilot CLI formats.
struct PostHookInput {
    tool_name: String,
    command: String,
    text: String,
    dialect: PostDialect,
}

/// Decode Copilot's `toolArgs`, which may arrive as a JSON-encoded string or an object.
fn decode_tool_args(v: &serde_json::Value) -> serde_json::Value {
    match v.as_str() {
        Some(raw) => serde_json::from_str(raw).unwrap_or(serde_json::Value::Null),
        None => v.clone(),
    }
}

/// Map an agent's shell/list tool name onto tokenix's internal POST_HOOK_TOOLS name.
/// Copilot uses `bash`/`powershell`; Claude Code already uses `Bash`/`ListDirectory`.
fn normalize_post_tool(name: &str) -> String {
    match name.to_ascii_lowercase().as_str() {
        "bash"
        | "powershell"
        | "shell"
        | "run_shell_command"
        | "default_api:run_shell_command"
        | "run_command"
        | "default_api:run_command"
        | "get_terminal_output"
        | "default_api:get_terminal_output" => "Bash".to_string(),
        "listdirectory" | "default_api:list_directory" => "ListDirectory".to_string(),
        _ => name.to_string(),
    }
}

/// Extract the LLM-facing text from a Copilot `toolResult` (object or string,
/// camelCase `textResultForLlm` or VS Code snake_case `text_result_for_llm`).
fn extract_copilot_result(tr: &serde_json::Value) -> Option<String> {
    if let Some(s) = tr.as_str() {
        return Some(s.to_string());
    }
    tr["textResultForLlm"]
        .as_str()
        .or_else(|| tr["text_result_for_llm"].as_str())
        .map(str::to_string)
}

/// Normalize a PostToolUse payload. Copilot CLI sends camelCase `toolName`/`toolResult`;
/// Claude Code sends snake_case `tool_name`/`tool_response`.
fn parse_post_input(v: &serde_json::Value) -> Option<PostHookInput> {
    // GitHub Copilot CLI: camelCase toolName + toolResult + (maybe string-encoded) toolArgs.
    if let Some(raw_name) = v["toolName"].as_str() {
        let args = decode_tool_args(&v["toolArgs"]);
        let command = args["command"]
            .as_str()
            .or_else(|| args["CommandLine"].as_str())
            .or_else(|| args["commandLine"].as_str())
            .or_else(|| args["command_line"].as_str())
            .unwrap_or("")
            .to_string();
        return Some(PostHookInput {
            tool_name: normalize_post_tool(raw_name),
            command,
            text: extract_copilot_result(&v["toolResult"])?,
            dialect: PostDialect::CopilotJson,
        });
    }

    // Claude Code / Codex: snake_case tool_name + tool_response.
    let raw_name = v["tool_name"].as_str()?;
    let command = v["tool_input"]["command"]
        .as_str()
        .or_else(|| v["tool_input"]["CommandLine"].as_str())
        .or_else(|| v["tool_input"]["commandLine"].as_str())
        .or_else(|| v["tool_input"]["command_line"].as_str())
        .unwrap_or("")
        .to_string();
    Some(PostHookInput {
        tool_name: normalize_post_tool(raw_name),
        command,
        text: extract_response_text(&v["tool_response"])?,
        dialect: PostDialect::ClaudeNoop,
    })
}

pub fn run_hook_post() -> Result<()> {
    let raw_stdin = std::io::read_to_string(std::io::stdin()).unwrap_or_default();
    let clean = raw_stdin.trim_start_matches('\u{feff}').trim();

    let v: serde_json::Value = match serde_json::from_str(clean) {
        Ok(v) => v,
        Err(_) => std::process::exit(0),
    };

    let input = match parse_post_input(&v) {
        Some(i) if !i.text.is_empty() => i,
        _ => std::process::exit(0),
    };

    // Redaction runs first and unconditionally — on every tool this hook
    // fires on, independent of dialect. This is the path that reaches tool
    // results the PreToolUse Bash rewrite never sees (Read results, MCP
    // results, ListDirectory, ...). Historically Claude/Codex non-shell
    // results exited here immediately on the (stale) belief that Claude Code
    // PostToolUse can never replace a result — as of Claude Code v2.1.121+,
    // `hookSpecificOutput.updatedToolOutput` is honored for all tools, not
    // only MCP, so that early exit was silently skipping real redaction.
    let (redacted, secret_hit) = crate::secrets_scan::redact_known_secrets(&input.text);

    let is_shell = POST_HOOK_TOOLS.contains(&input.tool_name.as_str());
    let compressed = if input.tool_name == "Bash" {
        compress_bash_output(&input.command, &redacted)
    } else if is_shell {
        compress_output(&redacted)
    } else if input.dialect == PostDialect::ClaudeNoop {
        // Non-shell, Claude/Codex dialect: no command-oriented compression
        // heuristics apply — redaction above is the whole job.
        redacted
    } else {
        // Non-shell, Copilot dialect: unchanged behavior, redact embedded
        // base64 blobs — the dominant token waste in agent histories — on
        // top of the secret redaction already applied.
        redact_base64_blobs(&redacted)
    };

    let changed = compressed != input.text;
    if !secret_hit && !changed {
        std::process::exit(0);
    }

    let repo_root = find_repo_root();
    let original_tokens = count_tokens(&input.text) as i64;
    let actual_tokens = count_tokens(&compressed) as i64;

    if secret_hit {
        let _ = log_hook_event(
            &repo_root,
            &HookEvent {
                ts: now_ts(),
                tool: input.tool_name.clone(),
                action: "intercepted".to_string(),
                phase: "post".to_string(),
                reason: "secret-redacted".to_string(),
                saved_tokens: 0,
                actual_tokens,
                original_estimate: original_tokens,
                input_preview: clean.chars().take(200).collect(),
                command: input.command.clone(),
            },
        );
    }
    if changed {
        let _ = log_hook_event(
            &repo_root,
            &HookEvent {
                ts: now_ts(),
                tool: input.tool_name.clone(),
                action: "intercepted".to_string(),
                phase: "post".to_string(),
                reason: "size-compressed".to_string(),
                saved_tokens: (original_tokens - actual_tokens).max(0),
                actual_tokens,
                original_estimate: original_tokens,
                input_preview: clean.chars().take(200).collect(),
                command: input.command.clone(),
            },
        );
    }

    match input.dialect {
        PostDialect::ClaudeNoop => {
            // As of Claude Code v2.1.121+, updatedToolOutput is honored for
            // all tools. Fail open on older installs: the field is simply
            // ignored and the original tool result stands — never a
            // blocking error, never malformed stdout.
            let out = serde_json::json!({
                "hookSpecificOutput": {
                    "hookEventName": "PostToolUse",
                    "updatedToolOutput": compressed,
                }
            });
            println!("{}", serde_json::to_string(&out).unwrap_or_default());
        }
        PostDialect::CopilotJson => {
            let out = serde_json::json!({
                "modifiedResult": {
                    "resultType": "success",
                    "textResultForLlm": compressed,
                }
            });
            println!("{}", serde_json::to_string(&out).unwrap_or_default());
        }
    }
    std::process::exit(0);
}

/// Pick the PowerShell binary: prefer PowerShell 7+ (`pwsh`, UTF-8 native) and
/// fall back to Windows PowerShell 5.1 (`powershell`) when pwsh is absent.
fn powershell_program() -> &'static str {
    static PROGRAM: std::sync::OnceLock<&'static str> = std::sync::OnceLock::new();
    PROGRAM.get_or_init(|| {
        let probe = std::process::Command::new("pwsh")
            .args(["-NoProfile", "-NonInteractive", "-Command", "$null"])
            .stdout(std::process::Stdio::null())
            .stderr(std::process::Stdio::null())
            .status();
        if probe.is_ok() {
            "pwsh"
        } else {
            "powershell"
        }
    })
}

/// Below this, the recovery hint costs more than the content it points at.
const RECOVERY_HINT_MIN_DROPPED_BYTES: usize = 500;

const TEE_MAX_FILES: usize = 20;
const TEE_MAX_BYTES: usize = 1_000_000;
const TEE_MIN_RAW_BYTES: usize = 500;

/// Failure tee: when a command fails and compression dropped content, persist
/// the full raw output under `~/.tokenix/tee/` so the agent can Read the rest
/// instead of re-running the command (the most expensive recovery path — the
/// re-run pays the raw cost twice and still goes through the hook). Returns
/// the file path to reference in a recovery hint. `TOKENIX_TEE=0` disables.
fn tee_raw_output(command_str: &str, stdout_raw: &str, stderr_raw: &str) -> Option<PathBuf> {
    if std::env::var("TOKENIX_TEE").is_ok_and(|v| v == "0") {
        return None;
    }
    let dir = crate::store::global_dir()?.join("tee");
    std::fs::create_dir_all(&dir).ok()?;
    crate::store::restrict_to_owner(&dir);

    let epoch = std::time::SystemTime::now()
        .duration_since(std::time::UNIX_EPOCH)
        .ok()?
        .as_secs();
    let slug: String = command_str
        .chars()
        .map(|c| if c.is_ascii_alphanumeric() { c } else { '-' })
        .take(40)
        .collect::<String>()
        .trim_matches('-')
        .to_string();
    let path = dir.join(format!("{epoch}_{slug}.log"));

    // The command line itself is the most reliable place to find a credential in
    // a failing run (`curl -H "Authorization: …"`, `psql postgres://u:pw@h`), and
    // this file outlives the session. Mask it like every other persisted view.
    let mut body = format!(
        "$ {}\n",
        crate::conversation_audit::redact_credentials(command_str)
    );
    let mut push_stream = |label: &str, raw: &str| {
        if !raw.trim().is_empty() {
            body.push_str(&format!("--- {label} ---\n"));
            let mut cut = raw.len().min(TEE_MAX_BYTES);
            while cut > 0 && !raw.is_char_boundary(cut) {
                cut -= 1;
            }
            // A failing command is exactly the one that dumps its environment or
            // echoes an auth header. The masking is shape-targeted (auth headers,
            // URL userinfo, `token=`/`password=`, PEM headers), so stack traces
            // and compiler errors — the reason this file exists — survive intact.
            body.push_str(&crate::conversation_audit::redact_credentials(&raw[..cut]));
            if cut < raw.len() {
                body.push_str("\n[tee capped at 1MB]");
            }
            if !body.ends_with('\n') {
                body.push('\n');
            }
        }
    };
    push_stream("stdout", stdout_raw);
    push_stream("stderr", stderr_raw);
    std::fs::write(&path, &body).ok()?;
    crate::store::restrict_to_owner(&path);

    // Rotation: keep the newest TEE_MAX_FILES logs (epoch prefix sorts).
    if let Ok(entries) = std::fs::read_dir(&dir) {
        let mut logs: Vec<PathBuf> = entries
            .flatten()
            .map(|e| e.path())
            .filter(|p| p.extension().and_then(|e| e.to_str()) == Some("log"))
            .collect();
        if logs.len() > TEE_MAX_FILES {
            logs.sort();
            for old in &logs[..logs.len() - TEE_MAX_FILES] {
                let _ = std::fs::remove_file(old);
            }
        }
    }
    Some(path)
}

/// Build the child `Command` for `tokenix run`, shared by the compressed and
/// raw execution paths so the two can never drift on how the command is
/// actually invoked.
fn build_shell_command(
    command_str: &str,
    shell: &str,
    cwd: Option<&Path>,
) -> std::process::Command {
    let is_powershell = matches!(shell, "pwsh" | "powershell");
    let mut cmd = if is_powershell {
        // Force UTF-8 so captured bytes decode cleanly (Windows PowerShell 5.1
        // otherwise emits UTF-16/codepage output that from_utf8_lossy mangles).
        let wrapped =
            format!("[Console]::OutputEncoding=[System.Text.Encoding]::UTF8; {command_str}");
        let mut c = std::process::Command::new(powershell_program());
        c.args(["-NoProfile", "-NonInteractive", "-Command", &wrapped]);
        c
    } else if cfg!(windows) {
        match shell::git_bash() {
            Some(bash) => {
                let mut c = std::process::Command::new(bash);
                c.args(["-c", command_str]);
                c
            }
            None => {
                let mut c = std::process::Command::new("cmd");
                c.args(["/C", command_str]);
                c
            }
        }
    } else {
        let mut c = std::process::Command::new("sh");
        c.args(["-c", command_str]);
        c
    };

    // `tokenix run --path <dir>` must actually run there.
    if let Some(dir) = cwd {
        cmd.current_dir(dir);
    }
    cmd
}

/// Run the command and print its stdout/stderr byte-for-byte, with no
/// compression, dedup, or recall stash — `tokenix run --raw` / `TOKENIX_RAW=1`.
///
/// Exists because auto-detecting "this output feeds a script, not an LLM" is
/// not a safe deterministic signal here: the agent harness that is the normal
/// caller of `tokenix run` also captures stdout through a pipe, so a piped
/// script and a harness invocation are indistinguishable by `isatty` alone.
/// An explicit opt-out is the only form of this guard that cannot misfire.
pub fn run_command_raw(command_str: &str, shell: &str, cwd: Option<&Path>) -> Result<i32> {
    let mut cmd = build_shell_command(command_str, shell, cwd);
    // Inherit rather than capture: streams the child's bytes straight through
    // untouched (encoding included), and avoids buffering output this path
    // never needs to look at.
    cmd.stdout(std::process::Stdio::inherit());
    cmd.stderr(std::process::Stdio::inherit());
    let status = cmd.status()?;
    Ok(status.code().unwrap_or(0))
}

pub fn run_command_and_compress(command_str: &str, shell: &str, cwd: Option<&Path>) -> Result<i32> {
    let is_powershell = matches!(shell, "pwsh" | "powershell");
    let output = build_shell_command(command_str, shell, cwd).output()?;

    let stdout_raw = String::from_utf8_lossy(&output.stdout);
    let stderr_raw = String::from_utf8_lossy(&output.stderr);

    let (stdout_compressed, mut stderr_compressed) = compress_command_streams(
        command_str,
        &stdout_raw,
        &stderr_raw,
        output.status.success(),
    );

    // Failure tee: the compressed view of a failure may not be enough to fix
    // it. Persist the raw output and point at it, so recovery is a targeted
    // Read instead of a full re-run. Only when compression actually dropped
    // content and the raw is big enough to be worth a file.
    let raw_len = stdout_raw.len() + stderr_raw.len();
    let compressed_len = stdout_compressed.len() + stderr_compressed.len();
    if !output.status.success() && raw_len >= TEE_MIN_RAW_BYTES && compressed_len + 80 < raw_len {
        if let Some(path) = tee_raw_output(command_str, &stdout_raw, &stderr_raw) {
            // Say "credentials masked" rather than implying a byte-for-byte copy:
            // the agent must not conclude a redacted token was the literal value.
            stderr_compressed.push_str(&format!(
                "\n[full output (credentials masked): {}]\n",
                path.display()
            ));
        }
    }

    let repo_root = find_repo_root();
    // Dedup and recall are scoped per project: the index is machine-wide, and a
    // pointer to "the identical output you already have" is only true within the
    // checkout that produced it.
    let project = repo_root.to_string_lossy().to_string();

    // Cross-call dedup: agents re-run `git status`/`cargo check`/`kubectl get`
    // and pay full price for byte-identical output every time. Only on success
    // — a repeated failure still needs its error text in front of the model.
    let mut stdout_compressed = stdout_compressed;
    let mut deduped_against: Option<String> = None;
    if output.status.success() {
        let stdout_tokens = count_tokens(&stdout_compressed);
        if let Some(hit) = crate::recall::find_identical(
            &project,
            command_str,
            &stdout_compressed,
            stdout_tokens,
            now_ts(),
        ) {
            let marker = crate::recall::dedup_marker(&hit, now_ts());
            // Never trade a short output for a longer marker.
            if marker.len() < stdout_compressed.len() {
                deduped_against = Some(hit.key.clone());
                stdout_compressed = format!("{marker}\n");
            }
        }
        if deduped_against.is_none() && stdout_tokens >= 1 {
            let dropped = stdout_raw.len().saturating_sub(stdout_compressed.len());
            let key = crate::recall::remember(
                &project,
                command_str,
                &stdout_compressed,
                &stdout_raw,
                stdout_tokens,
                now_ts(),
            );
            // Advertise the recovery key on the *first* clipped run, not only on a
            // later dedup hit. Compression is reversible by design, but the key
            // was being stashed and never shown, so the one output most worth
            // recovering — a big first result the caps just trimmed — had no way
            // back short of re-running the command.
            if let Some(key) = key {
                if dropped >= RECOVERY_HINT_MIN_DROPPED_BYTES {
                    if !stdout_compressed.ends_with('\n') {
                        stdout_compressed.push('\n');
                    }
                    stdout_compressed.push_str(&format!(
                        "[tokenix: {dropped} bytes not shown — `tokenix retrieve {key}` returns the full output]\n"
                    ));
                }
            }
        }
    }

    // Print to standard streams
    print!("{}", stdout_compressed);
    eprint!("{}", stderr_compressed);

    // Write log event of the actual execution savings
    // Capture raw output if a `tokenix filter record` session is active.
    crate::recordings::capture(&repo_root, command_str, &stdout_raw, &stderr_raw);
    let original_tokens = (count_tokens(&stdout_raw) + count_tokens(&stderr_raw)) as i64;
    let actual_tokens =
        (count_tokens(&stdout_compressed) + count_tokens(&stderr_compressed)) as i64;
    let saved = (original_tokens - actual_tokens).max(0);

    // Log every measured run, including the ones that saved nothing.
    //
    // Skipping `saved == 0` made `gain`'s denominator the set of commands where
    // compression already worked, so `pct_saved` answered "how much do we save
    // when we save" — a number that cannot go down and is not the savings rate
    // anyone reads it as. A run that went through `tokenix run` and came back
    // the same size is a real measurement and belongs in the sample.
    let (action, reason) = if saved > 0 {
        ("intercepted", "compressed command output")
    } else {
        ("pass", "no reduction available")
    };
    let _ = log_hook_event(
        &repo_root,
        &HookEvent {
            ts: now_ts(),
            tool: if is_powershell { "PowerShell" } else { "Bash" }.to_string(),
            action: action.to_string(),
            phase: "ToolOutputCompressed".to_string(),
            reason: reason.to_string(),
            saved_tokens: saved,
            actual_tokens,
            original_estimate: original_tokens,
            input_preview: command_str.chars().take(200).collect(),
            command: command_str.to_string(),
        },
    );

    Ok(output.status.code().unwrap_or(0))
}

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

    #[test]
    fn run_command_raw_preserves_exit_code() {
        // stdout/stderr content is inherited straight to the process' own
        // streams (asserting the bytes would mean capturing the whole test
        // binary's output); the exit code is the one signal this path can
        // check without that. Manually verified byte-exactness against
        // `tokenix run --raw` in homologation.
        let code = run_command_raw("exit 7", "auto", None).unwrap();
        assert_eq!(code, 7);
    }

    #[test]
    fn compact_json_is_lossless() {
        // Regression: round-tripping through serde_json::Value turned integers
        // past u64 into scientific notation, collapsed duplicate keys, and
        // alphabetized the object.
        let big = r#"{ "id" : 123456789012345678901234567890 , "b" : 1 , "a" : 2 }"#;
        let out = compact_json(big);
        assert!(out.contains("123456789012345678901234567890"), "{out}");
        assert!(out.starts_with(r#"{"id":"#), "key order changed: {out}");
        // Whitespace inside strings is preserved.
        assert_eq!(compact_json(r#"{ "k" : "a  b" }"#), r#"{"k":"a  b"}"#);
    }

    #[test]
    fn strip_ansi_does_not_blank_output_on_esc_before_multibyte() {
        // Regression: ESC + a UTF-8 lead byte consumed half a character, so
        // `from_utf8` failed and `unwrap_or_default()` wiped the whole output.
        let raw = String::from_utf8(vec![b'o', b'k', 0x1b, 0xC3, 0xA9, b'!']).unwrap();
        let out = strip_ansi(&raw);
        assert!(out.starts_with("ok"), "output was blanked: {out:?}");
        assert!(out.ends_with('!'), "tail lost: {out:?}");
    }

    #[test]
    fn pkg_manager_keeps_uppercase_error_lines() {
        // Regression: `starts_with("error")` was case-sensitive, so `ERROR:` and
        // `npm ERR!` were dropped once the first 30 lines were used up.
        let mut lines: Vec<&str> = (0..40).map(|_| "added 1 package").collect();
        lines.push("ERROR: could not resolve dependency");
        lines.push("npm ERR! code ERESOLVE");
        let out = compress_pkg_manager(&lines);
        assert!(out.contains("ERROR: could not resolve"), "{out}");
        assert!(out.contains("npm ERR!"), "{out}");
    }

    #[test]
    fn kebab_identifier_chain_is_not_redacted_as_base64() {
        // Regression: `-` counted as proof of base64, so the scan returned true
        // on the first hyphen of a lowercase slug chain.
        let run = "build-step-one-build-step-two-build-step-three".repeat(20);
        assert!(!looks_like_base64(run.as_bytes()));
        // A real base64url payload is mixed-case and still detected.
        let b64 = "aGVsbG9Xb3JsZFRoaXNJc0Jhc2U2NA".repeat(20);
        assert!(looks_like_base64(b64.as_bytes()));
    }

    #[test]
    fn small_git_diff_keeps_its_changed_lines() {
        let diff = "diff --git a/x b/x\n--- a/x\n+++ b/x\n@@ -1 +1 @@\n-old\n+new";
        let lines: Vec<&str> = diff.lines().collect();
        let out = compress_git_diff(&lines);
        assert!(out.contains("+new") && out.contains("-old"), "{out}");
    }

    #[test]
    fn kubectl_describe_is_not_replaced_by_a_placeholder() {
        let mut lines = vec!["Name:         my-pod", "Namespace:    default"];
        for i in 0..80 {
            lines.push(if i == 79 {
                "  Warning  BackOff  pod crashlooping"
            } else {
                "  filler: value"
            });
        }
        let out = compress_kubectl(&lines);
        assert!(!out.contains("<resource details>"), "fabricated: {out}");
        assert!(out.contains("BackOff"), "events tail lost: {out}");
    }

    #[test]
    fn strips_data_uri_image_blob() {
        let blob = "Zm9vQmFyBaz1".repeat(200); // 2400 mixed-case base64 chars
        let n = blob.len();
        let raw = format!("here is the image data:image/png;base64,{blob} end");
        let out = compress_output(&raw);
        assert!(out.contains("data:image/png;base64,"), "prefix kept: {out}");
        assert!(
            out.contains(&format!("[base64 blob omitted: {n} chars]")),
            "redacted: {out}"
        );
        assert!(!out.contains(&blob), "raw blob must be gone");
        assert!(out.starts_with("here is the image"));
        assert!(out.trim_end().ends_with("end"));
    }

    #[test]
    fn strips_bare_base64_blob() {
        // A tar/binary dump with no data: prefix (the `unknown-large` audit shape).
        let blob = "iVBORw0KGgoAAAANSUhEUg".repeat(60); // > 512 base64 chars, PNG magic
        let out = strip_base64_blobs(&blob);
        assert_eq!(
            out,
            format!("[png image base64 omitted: {} chars]", blob.len())
        );
    }

    #[test]
    fn strips_line_wrapped_base64_block() {
        // 10 wrapped lines of 64 base64 chars — the MIME/`base64 file` shape the
        // contiguous scanner misses (each line < BASE64_MIN).
        let line = "Zm9v".repeat(16); // 64 mixed-case base64 chars per line
        let block = std::iter::repeat_n(line.as_str(), 10)
            .collect::<Vec<_>>()
            .join("\n");
        let raw = format!("preamble\n{block}\ntrailer\n");
        let out = strip_wrapped_base64(&raw);
        assert!(out.contains("omitted:"), "collapsed: {out}");
        assert!(out.contains("across 10 lines"), "line count: {out}");
        assert!(out.starts_with("preamble\n"));
        // A trailing all-alnum word must NOT be absorbed into the blob.
        assert!(out.ends_with("trailer\n"), "trailer kept: {out}");
        assert!(!out.contains(&line), "raw blob gone");
    }

    #[test]
    fn strips_pem_body_keeps_markers() {
        // 64-char lines (PEM wrap width) so each clears WRAP_MIN_LINE.
        let body = std::iter::repeat_n(
            "MIIDdummyBase64ContentLinePadded0123456789ABCDEFabcdefXYZ01234567",
            8,
        )
        .collect::<Vec<_>>()
        .join("\n");
        let pem = format!("-----BEGIN CERTIFICATE-----\n{body}\n-----END CERTIFICATE-----\n");
        let out = strip_wrapped_base64(&pem);
        assert!(
            out.contains("-----BEGIN CERTIFICATE-----"),
            "begin kept: {out}"
        );
        assert!(out.contains("-----END CERTIFICATE-----"), "end kept: {out}");
        assert!(out.contains("omitted:"), "body collapsed: {out}");
    }

    #[test]
    fn wrapped_stripper_spares_short_or_prose_blocks() {
        // 4 long mixed-case base64 lines (>512 chars total, past the length gate)
        // but below WRAP_MIN_LINES → untouched (exercises the line-count gate).
        let short = format!("{a}\n{a}\n{a}\n{a}", a = "Zm9v".repeat(35));
        assert_eq!(strip_wrapped_base64(&short), short);
        // Consecutive prose lines with spaces are not base64 lines.
        let prose = "the quick brown fox jumped\n".repeat(30);
        assert_eq!(strip_wrapped_base64(&prose), prose);
    }

    #[test]
    fn base64_blob_kind_labels_by_magic() {
        let jpeg = format!("/9j/{}", "A".repeat(600));
        assert!(strip_base64_blobs(&jpeg).contains("[jpeg image base64 omitted:"));
        let gif = format!("R0lGOD{}", "A".repeat(600));
        assert!(strip_base64_blobs(&gif).contains("[gif image base64 omitted:"));
        let opaque = "Zm9v".repeat(150); // 600 mixed-case chars, no magic prefix
        assert!(strip_base64_blobs(&opaque).contains("[base64 blob omitted:"));
    }

    #[test]
    fn base64_stripper_spares_hex_and_digit_runs() {
        // A long lowercase-hex run (concatenated checksums) is NOT base64 — sparing
        // it avoids silently eating a sha256 manifest or hex dump.
        let hex = "deadbeef0123456789abcdef".repeat(30); // 720 hex chars, single case
        assert_eq!(strip_base64_blobs(&hex), hex);
        // A long numeric id column is spared too.
        let digits = "1234567890".repeat(70); // 700 digits
        assert_eq!(strip_base64_blobs(&digits), digits);
    }

    #[test]
    fn wrapped_stripper_spares_hash_manifest() {
        // 8 lines of 64-char lowercase hex (a `.sha256` manifest) must not collapse.
        let line = "0123456789abcdef".repeat(4); // 64 hex chars
        let manifest = std::iter::repeat_n(line.as_str(), 8)
            .collect::<Vec<_>>()
            .join("\n");
        assert_eq!(strip_wrapped_base64(&manifest), manifest);
    }

    #[test]
    fn base64_stripper_spares_normal_text() {
        // Long minified-looking code stays intact: punctuation breaks the run.
        let code = "const x=(a,b)=>{return a+b;};".repeat(80);
        assert_eq!(strip_base64_blobs(&code), code);
        // A short token below threshold is untouched.
        let short = "AAAABBBBCCCCDDDD";
        assert_eq!(strip_base64_blobs(short), short);
        // UTF-8 around a blob survives.
        let mixed = format!("café ☕ {} über", "Zm9vYmFy".repeat(80));
        let out = strip_base64_blobs(&mixed);
        assert!(out.starts_with("café ☕ "));
        assert!(out.trim_end().ends_with("über"));
    }

    #[test]
    fn strips_ansi_colors() {
        assert_eq!(strip_ansi("\x1b[32mOK\x1b[0m"), "OK");
        assert_eq!(strip_ansi("\x1b[1;31mError\x1b[0m: bad"), "Error: bad");
    }

    #[test]
    fn git_status_verbose_to_porcelain() {
        let raw = "On branch main\nYour branch is up to date with 'origin/main'.\n\nChanges not staged for commit:\n  (use \"git add <file>...\")\n\tmodified:   src/main.rs\n\tdeleted:    old.rs\n\nUntracked files:\n  (use \"git add <file>...\")\n\tnew.rs\n";
        let lines: Vec<&str> = raw.lines().collect();
        let out = compress_git_status(&lines);
        assert_eq!(out, "M src/main.rs\nD old.rs\n?? new.rs");
        // Clean repo → terse marker.
        let clean = ["On branch main", "nothing to commit, working tree clean"];
        assert_eq!(compress_git_status(&clean), "git status: clean");
    }

    #[test]
    fn grep_strips_indentation() {
        let lines = [
            "src/a.rs:10:    let x = 5;",
            "src/b.rs:2:        fn main() {}",
            "",
        ];
        let out = compress_grep(&lines);
        assert_eq!(out, "src/a.rs:10:let x = 5;\nsrc/b.rs:2:fn main() {}");
    }

    // Golden/safety-net for the dominant savings path: `cargo build` is ~96% of
    // all tokens tokenix saves globally, and it flows through `compress_cargo`.
    // This pins the exact kept-vs-dropped behavior so any accidental change to the
    // cargo heuristic shows up as a failing snapshot instead of a silent regression.
    #[test]
    fn cargo_build_golden_keeps_signal_drops_noise() {
        let raw = "\
   Compiling libc v0.2.1
   Compiling serde v1.0.2
   Compiling tokenix v0.57.1
warning: function is never used: `foo`
error[E0425]: cannot find value `y`
    Finished `dev` profile [optimized] target(s) in 4.2s";
        let lines: Vec<&str> = raw.lines().collect();
        let out = compress_cargo(&lines);

        // Exact snapshot: warning + error + Finished survive; every `Compiling` drops.
        assert_eq!(
            out,
            "warning: function is never used: `foo`\n\
             error[E0425]: cannot find value `y`\n\
             \x20   Finished `dev` profile [optimized] target(s) in 4.2s"
        );
        // Invariants (robust to formatting tweaks in the snapshot above).
        assert!(!out.contains("Compiling "), "build noise dropped");
        assert!(out.contains("error[E0425]"), "errors preserved");
        assert!(out.contains("Finished "), "summary preserved");
        assert!(out.len() < raw.len(), "net reduction");
        // Deterministic: same input → same output.
        assert_eq!(compress_cargo(&lines), out);
    }

    #[test]
    fn cargo_metadata_summarized() {
        let json = r#"{"packages":[{"name":"a","version":"1.0.0"},{"name":"b","version":"2.0.0"}],"workspace_members":["tokenix 0.1.0 (path+file:///x)"],"resolve":{"root":"tokenix 0.1.0 (path+file:///x)"}}"#;
        let out = compress_cargo_metadata(json);
        assert!(out.contains("2 packages"));
        assert!(out.contains("workspace members: tokenix"));
        assert!(out.len() < json.len());
    }

    #[test]
    fn cargo_tree_dedupes() {
        let lines = [
            "tokenix v0.1.0",
            "├── anyhow v1.0.0",
            "│   └── anyhow v1.0.0 (*)",
            "└── serde v1.0.0",
        ];
        let out = compress_cargo_tree(&lines);
        assert!(out.starts_with("cargo tree: "));
        assert!(out.contains("anyhow v1.0.0"));
        assert!(out.contains("serde v1.0.0"));
        // anyhow appears twice in input but once in the unique set.
        assert_eq!(out.matches("anyhow v1.0.0").count(), 1);
    }

    #[test]
    fn command_streams_do_not_turn_stderr_errors_into_success() {
        let (stdout, stderr) = compress_command_streams(
            "cargo build",
            "",
            "error: could not find `Cargo.toml` in this directory\n",
            false,
        );

        assert_eq!(stdout, "");
        assert!(
            stderr.contains("could not find"),
            "stderr error must be preserved, got: {stderr:?}"
        );
        assert!(
            !stderr.contains("build succeeded"),
            "stderr must not emit success sentinel: {stderr:?}"
        );
    }

    #[test]
    fn cargo_warnings_on_stderr_reach_the_filter() {
        let warning = "warning: unused variable: `x`\n --> src/lib.rs:2:9\n  |\n2 |     let x = 5;\n  |         ^ help: if this is intentional, prefix it with an underscore: `_x`\n  |\n  = note: `#[warn(unused_variables)]` on by default\n\n";
        let stderr_raw = format!("   Compiling demo v0.1.0\n{warning}    Finished `dev` profile\n");
        let (_, stderr) = compress_command_streams("cargo clippy", "", &stderr_raw, true);
        assert!(
            stderr.contains("warning: unused variable: `x`\n --> src/lib.rs:2:9"),
            "{stderr}"
        );
        assert!(
            !stderr.contains("let x = 5"),
            "rustc writes to stderr; the cargo filter must see it: {stderr}"
        );
        assert!(
            stderr.ends_with('\n'),
            "a merged stdout would glue onto the last line: {stderr:?}"
        );
    }

    #[test]
    fn silent_success_stays_empty() {
        // on_empty sentinels were removed: a silent successful command must not
        // fabricate a "succeeded" string — genuinely empty output stays empty.
        let (stdout, stderr) = compress_command_streams("cargo build", "", "", true);
        assert_eq!(stdout, "");
        assert_eq!(stderr, "");
    }

    #[test]
    fn ps_keeps_top_by_cpu() {
        let lines = [
            "USER PID %CPU %MEM CMD",
            "u 1 0.0 0.1 idle",
            "u 2 99.0 5.0 hot",
            "u 3 0.1 0.2 warm",
            "u 4 0.0 0.0 idle2",
            "u 5 0.0 0.0 idle3",
            "u 6 0.0 0.0 idle4",
        ];
        let out = compress_ps(&lines);
        let hot_pos = out.find("hot").expect("busiest process kept");
        let idle_pos = out.find("idle3");
        // The 99% process is kept and ranks above the idle ones (which may be dropped).
        assert!(out.starts_with("USER PID"));
        assert!(idle_pos.is_none() || hot_pos < idle_pos.unwrap());
    }

    #[test]
    fn path_listing_collapses_to_top_level() {
        let lines = [
            "./src/a.rs",
            "./src/b.rs",
            "./target/debug/x.rs",
            "./target/debug/y.rs",
            "./benchmark/c.rs",
        ];
        let out = compress_path_listing(&lines);
        assert!(out.contains("5 files across"));
        assert!(out.contains("target/ (2)") || out.contains("src/ (2)"));
        assert!(out.len() < lines.join("\n").len());
    }

    #[test]
    fn cargo_test_failure_detail_is_preserved() {
        // A failing `cargo test` must keep the WHOLE failure block — the reason a
        // test failed is free-form (custom panic message, pretty-assertion colour
        // diff, backtrace) and matches no fixed prefix. Compression should still
        // drop noise (Compiling lines, passing tests) but never the failure signal.
        let raw = "\
Compiling foo v0.1.0
running 2 tests
test tests::ok_one ... ok
test tests::adds ... FAILED

failures:

---- tests::adds stdout ----
thread 'tests::adds' panicked at src/lib.rs:10:9:
custom failure: widget count drifted by 1
Diff < left / right > :
<4
>5
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

failures:
    tests::adds

test result: FAILED. 1 passed; 1 failed; 0 ignored; 0 measured; 0 filtered out";
        let lines: Vec<&str> = raw.lines().collect();
        let out = compress_cargo(&lines);
        assert!(out.contains("panicked at"), "panic line must be preserved");
        // Free-form lines that no prefix matches — the stateful block capture keeps them.
        assert!(
            out.contains("custom failure: widget count drifted by 1"),
            "custom panic message must be preserved"
        );
        assert!(
            out.contains("Diff < left / right > :") && out.contains("<4") && out.contains(">5"),
            "pretty-assertion diff must be preserved"
        );
        assert!(
            out.contains("---- tests::adds stdout ----"),
            "failing test name marker must be preserved"
        );
        assert!(
            out.contains("test result: FAILED"),
            "summary must be preserved"
        );
        // Still compresses: the noisy Compiling line is dropped.
        assert!(!out.contains("Compiling foo"), "noise should be dropped");
    }

    #[test]
    fn strips_osc_sequences() {
        assert_eq!(strip_ansi("\x1b]0;title\x07text"), "text");
    }

    #[test]
    fn removes_emojis() {
        assert_eq!(remove_emojis("🚀 Build done"), " Build done");
        assert_eq!(remove_emojis("no emojis here"), "no emojis here");
    }

    #[test]
    fn collapses_blank_lines() {
        let input = "a\n\n\n\n\nb";
        let output = collapse_blank_lines(input);
        assert_eq!(output, "a\n\nb");
    }

    #[test]
    fn groups_repeated_lines() {
        let input = "line1\nline1\nline1\nline1\nline2\n";
        let output = group_repeated_lines(input);
        assert_eq!(output, "line1\n[repeated 3x]\nline2\n");
    }

    #[test]
    fn does_not_group_two_identical_lines() {
        let input = "a\na\nb\n";
        assert_eq!(group_repeated_lines(input), "a\na\nb\n");
    }

    #[test]
    fn groups_repeated_multiline_blocks() {
        // The monitoring-loop shape: an 3-line stanza re-emitted every poll.
        let stanza = "SetValueInvocationException:\n  Line | 3 | $RawUI.CursorPosition\n  Exception setting CursorPosition\n";
        let input = format!("{}header\n", stanza.repeat(5));
        let output = group_repeated_blocks(&input);
        assert!(
            output.contains("[block of 3 lines repeated 4x]"),
            "got: {output}"
        );
        assert_eq!(output.matches("SetValueInvocationException").count(), 1);
        assert!(output.ends_with("header\n"));
    }

    #[test]
    fn block_grouping_prefers_smallest_unit() {
        let input = "a\nb\na\nb\na\nb\na\nb\nz\n";
        let output = group_repeated_blocks(input);
        assert_eq!(output, "a\nb\n[block of 2 lines repeated 3x]\nz\n");
    }

    #[test]
    fn block_grouping_defers_identical_line_runs_to_line_grouping() {
        // "a" x6 must stay a line run, not become a 2-line "block".
        let input = "a\na\na\na\na\na\nz\n";
        assert_eq!(group_repeated_blocks(input), input);
        assert_eq!(group_repeated_lines(input), "a\n[repeated 5x]\nz\n");
    }

    #[test]
    fn block_grouping_leaves_two_repeats_alone() {
        let input = "a\nb\nc\na\nb\nc\nz\n";
        assert_eq!(group_repeated_blocks(input), input);
    }

    #[test]
    fn block_grouping_preserves_non_repeating_output() {
        let input = "one\ntwo\nthree\nfour\nfive\nsix\nseven\n";
        assert_eq!(group_repeated_blocks(input), input);
    }

    #[test]
    fn token_budget_caps_single_giant_line() {
        // The gap this closes: one enormous line has no line count to truncate on.
        let giant = format!(
            "prefix {} suffix",
            "lorem ipsum dolor sit amet ".repeat(20_000)
        );
        let out = enforce_token_budget(&giant);
        assert!(out.contains("tokens omitted by tokenix output cap"));
        assert!(count_tokens(&out) < count_tokens(&giant) / 2);
        assert!(out.starts_with("prefix "));
        assert!(out.ends_with(" suffix"));
    }

    #[test]
    fn token_budget_keeps_identifiers_from_the_dropped_range() {
        // A dropped SHA/UUID/URL cannot be re-derived — losing it forces the
        // re-run the cap exists to avoid.
        let filler = "routine log line with nothing worth keeping\n".repeat(6000);
        let s = format!(
            "start\n{filler}commit 9fceb02d1a3e5b7c failed at https://ci.example.com/run/42 with E0308 and id 550e8400-e29b-41d4-a716-446655440000\n{filler}end\n"
        );
        let out = enforce_token_budget(&s);
        assert!(
            out.contains("ids kept from the omitted range"),
            "got: {}",
            &out[..400.min(out.len())]
        );
        assert!(out.contains("9fceb02d1a3e5b7c"));
        assert!(out.contains("https://ci.example.com/run/42"));
        assert!(out.contains("E0308"));
        assert!(out.contains("550e8400-e29b-41d4-a716-446655440000"));
    }

    #[test]
    fn base64_stripper_spares_long_kebab_identifier_chains() {
        // Found by smoke-testing a real command: `-` is a base64url character,
        // so a long hyphen-joined slug list was redacted as a blob (data loss).
        let chain = (1..=80)
            .map(|i| format!("chunk{i}"))
            .collect::<Vec<_>>()
            .join("-");
        assert!(chain.len() > BASE64_MIN, "fixture must reach the run floor");
        let out = redact_base64_blobs(&format!("SMOKE-{chain}"));
        assert!(out.contains("chunk80"), "identifier chain was eaten: {out}");
        assert!(!out.contains("omitted"));

        // Snake_case chains are the same shape.
        let snake = (1..=80)
            .map(|i| format!("Step{i}"))
            .collect::<Vec<_>>()
            .join("_");
        assert!(redact_base64_blobs(&snake).contains("Step80"));
    }

    #[test]
    fn base64_stripper_still_redacts_base64url_payloads() {
        // A real base64url blob has few, long separator-free stretches — the
        // guard above must not give it a pass.
        let blob = format!(
            "{}-{}",
            "Zm9vQmFyBaz1".repeat(30),
            "QmFyRm9vYmF6".repeat(30)
        );
        let out = redact_base64_blobs(&blob);
        assert!(
            out.contains("omitted"),
            "real base64url must still be cut: {out}"
        );
    }

    #[test]
    fn identifier_scan_ignores_ordinary_prose() {
        // "decade"/"facade" are hex-only words; plain numbers are not identifiers.
        assert_eq!(
            preserved_identifiers("the decade long facade of 123456 lines added"),
            None
        );
    }

    #[test]
    fn token_budget_leaves_small_output_untouched() {
        let s = "short output\nsecond line\n";
        assert_eq!(enforce_token_budget(s), s);
    }

    #[test]
    fn token_budget_is_utf8_safe() {
        let s = "café ☕ ".repeat(30_000);
        let out = enforce_token_budget(&s);
        assert!(out.len() < s.len());
        // Round-tripping through String proves no char boundary was split.
        assert!(std::str::from_utf8(out.as_bytes()).is_ok());
    }

    #[test]
    fn compress_output_caps_giant_compacted_json() {
        // compact_json returned early with no cap: a huge payload shrank ~10% and
        // still cost hundreds of thousands of tokens.
        let items: Vec<String> = (0..40_000)
            .map(|i| format!("{{\n  \"id\": {i},\n  \"name\": \"item number {i}\"\n}}"))
            .collect();
        let raw = format!("[\n{}\n]", items.join(",\n"));
        let out = compress_output(&raw);
        assert!(
            count_tokens(&out) <= DEFAULT_MAX_OUTPUT_TOKENS + 64,
            "cap not applied: {} tokens",
            count_tokens(&out)
        );
        assert!(out.contains("tokens omitted by tokenix output cap"));
    }

    #[test]
    fn compacts_pretty_json_object() {
        let input = "{\n  \"status\": \"ok\",\n  \"count\": 42\n}\n";
        let output = compact_json(input);
        // key order is not guaranteed; verify it compacted (shorter) and parses to same value
        assert!(output.len() < input.len(), "should be shorter");
        assert!(output.ends_with('\n'));
        let v_in: serde_json::Value = serde_json::from_str(input.trim()).unwrap();
        let v_out: serde_json::Value = serde_json::from_str(output.trim()).unwrap();
        assert_eq!(v_in, v_out);
    }

    #[test]
    fn compacts_pretty_json_array() {
        let input = "[\n  1,\n  2,\n  3\n]";
        let output = compact_json(input);
        assert_eq!(output, "[1,2,3]");
    }

    #[test]
    fn passes_through_already_compact_json() {
        let input = "{\"a\":1}\n";
        assert_eq!(compact_json(input), input);
    }

    #[test]
    fn compacts_ndjson() {
        let input = "{ \"level\": \"info\", \"msg\": \"started\" }\n{ \"level\": \"error\", \"msg\": \"failed\" }\n";
        let output = compact_json(input);
        assert_eq!(
            output,
            "{\"level\":\"info\",\"msg\":\"started\"}\n{\"level\":\"error\",\"msg\":\"failed\"}\n"
        );
    }

    #[test]
    fn passes_through_plain_text() {
        let input = "On branch main\nnothing to commit\n";
        assert_eq!(compact_json(input), input);
    }

    #[test]
    fn compress_is_idempotent_on_clean_input() {
        let clean = "hello\nworld\n";
        assert_eq!(compress_output(clean), clean);
    }

    #[test]
    fn full_compression_pipeline() {
        let input = "\x1b[32m🚀 Starting\x1b[0m\n\n\n\nline\nline\nline\nline\ndone\n";
        let output = compress_output(input);
        assert!(output.contains("Starting"));
        assert!(!output.contains("\x1b["));
        assert!(!output.contains('🚀'));
        assert!(output.contains("[repeated"));
        assert!(!output.contains("\n\n\n"));
    }

    #[test]
    fn bash_short_output_passes_through() {
        let input = "hello\nworld\n";
        assert_eq!(compress_bash_output("", input), input);
    }

    #[test]
    fn bash_generic_truncation_over_100_lines() {
        let lines: String = (1..=150).map(|i| format!("line {}\n", i)).collect();
        let out = compress_bash_output("", &lines);
        assert!(out.contains("lines omitted"), "should truncate: {}", out);
        assert!(out.contains("line 1\n"));
        assert!(out.contains("line 150"));
    }

    #[test]
    fn bash_path_listing_groups_by_directory() {
        let input = [
            "src/main.rs",
            "src/query.rs",
            "src/hook.rs",
            "benchmark/samples/database_client.ts",
        ]
        .join("\n");

        let out = compress_bash_output("ls -R", &input);
        assert!(
            out.contains("4 files across 2 top-level dir(s)"),
            "output: {}",
            out
        );
        assert!(out.contains("src/ (3)"), "output: {}", out);
        assert!(out.contains("benchmark/ (1)"), "output: {}", out);
    }

    #[test]
    fn bash_cargo_extracts_errors() {
        let mut input = String::new();
        for i in 0..60 {
            input.push_str(&format!("Compiling crate{} v0.1.0\n", i));
        }
        input.push_str("error[E0425]: cannot find value `foo`\n");
        input.push_str("  --> src/main.rs:3:5\n");
        input.push_str("   |\n");
        input.push_str("3  |     foo();\n");
        input.push_str("error: aborting due to 1 previous error\n");
        input.push_str("Finished dev in 1.23s\n");

        let out = compress_bash_output("", &input);
        assert!(out.contains("error[E0425]"), "should keep error: {}", out);
        assert!(out.contains("Finished"), "should keep summary: {}", out);
        assert!(
            !out.contains("Compiling crate0"),
            "should strip Compiling lines"
        );
    }

    #[test]
    fn bash_git_log_truncated_after_20_commits() {
        let mut input = String::new();
        for i in 0..30 {
            input.push_str(&format!("commit {:040}\n", i));
            input.push_str("Author: Test\nDate: Today\n\n    message\n\n");
        }
        let out = compress_bash_output("", &input);
        assert!(out.contains("lines omitted"), "should truncate: {}", out);
    }

    #[test]
    fn bash_status_poll_collapses_repeated_states() {
        let input = "\
phase=Pending
phase=Pending
phase=Pending
phase=Running
phase=Running
=== LOGS ===
daily job failed validation
";
        let out = compress_bash_output(
            "for i in $(seq 1 40); do phase=$(kubectl get pod); echo phase=$phase; done",
            input,
        );
        assert!(out.contains("phase=Pending x3"), "output: {out}");
        assert!(out.contains("phase=Running x2"), "output: {out}");
        assert!(out.contains("daily job failed validation"), "output: {out}");
        assert!(
            !out.contains("phase=Pending\nphase=Pending"),
            "output: {out}"
        );
    }

    #[test]
    fn parses_claude_post_input() {
        let v = serde_json::json!({
            "tool_name": "Bash",
            "tool_input": {"command": "git status"},
            "tool_response": "On branch main\n"
        });
        let input = parse_post_input(&v).unwrap();
        assert_eq!(input.tool_name, "Bash");
        assert_eq!(input.command, "git status");
        assert_eq!(input.text, "On branch main\n");
        assert_eq!(input.dialect, PostDialect::ClaudeNoop);
    }

    #[test]
    fn parses_claude_bash_stdout_stderr_shape() {
        // Real Claude Code Bash PostToolUse payload: tool_response is an object
        // with stdout/stderr, not `output`/`content`.
        let v = serde_json::json!({
            "tool_name": "Bash",
            "tool_input": {"command": "npm install"},
            "tool_response": {
                "stdout": "added 120 packages in 3s\n",
                "stderr": "npm warn deprecated foo\n",
                "interrupted": false,
                "isImage": false
            }
        });
        let input = parse_post_input(&v).unwrap();
        assert_eq!(input.tool_name, "Bash");
        assert_eq!(input.command, "npm install");
        assert!(input.text.contains("added 120 packages"));
        assert!(input.text.contains("npm warn deprecated foo"));
        assert_eq!(input.dialect, PostDialect::ClaudeNoop);
    }

    #[test]
    fn parses_copilot_post_input_camelcase() {
        let v = serde_json::json!({
            "toolName": "bash",
            "toolArgs": {"command": "git diff"},
            "toolResult": {"resultType": "success", "textResultForLlm": "diff output"}
        });
        let input = parse_post_input(&v).unwrap();
        assert_eq!(input.tool_name, "Bash"); // bash → Bash
        assert_eq!(input.command, "git diff");
        assert_eq!(input.text, "diff output");
        assert_eq!(input.dialect, PostDialect::CopilotJson);
    }

    #[test]
    fn parses_copilot_post_input_string_encoded_args() {
        // Copilot may send toolArgs as a JSON-encoded string.
        let v = serde_json::json!({
            "toolName": "powershell",
            "toolArgs": "{\"command\":\"git status\"}",
            "toolResult": {"textResultForLlm": "status output"}
        });
        let input = parse_post_input(&v).unwrap();
        assert_eq!(input.tool_name, "Bash"); // powershell → Bash
        assert_eq!(input.command, "git status");
        assert_eq!(input.text, "status output");
    }

    #[test]
    fn extract_copilot_result_handles_both_casings() {
        let camel = serde_json::json!({"textResultForLlm": "a"});
        let snake = serde_json::json!({"text_result_for_llm": "b"});
        let plain = serde_json::Value::String("c".to_string());
        assert_eq!(extract_copilot_result(&camel).as_deref(), Some("a"));
        assert_eq!(extract_copilot_result(&snake).as_deref(), Some("b"));
        assert_eq!(extract_copilot_result(&plain).as_deref(), Some("c"));
    }

    #[test]
    fn normalize_post_tool_maps_shells_to_bash() {
        assert_eq!(normalize_post_tool("bash"), "Bash");
        assert_eq!(normalize_post_tool("powershell"), "Bash");
        assert_eq!(normalize_post_tool("Bash"), "Bash"); // Claude casing preserved
        assert_eq!(normalize_post_tool("run_command"), "Bash");
        assert_eq!(normalize_post_tool("default_api:run_command"), "Bash");
        assert_eq!(normalize_post_tool("ListDirectory"), "ListDirectory");
        assert_eq!(normalize_post_tool("view"), "view"); // unmapped → unchanged
    }

    #[test]
    fn parses_claude_post_input_run_command() {
        let v = serde_json::json!({
            "tool_name": "default_api:run_command",
            "tool_input": {"CommandLine": "git diff"},
            "tool_response": "diff output"
        });
        let input = parse_post_input(&v).unwrap();
        assert_eq!(input.tool_name, "Bash");
        assert_eq!(input.command, "git diff");
        assert_eq!(input.text, "diff output");
        assert_eq!(input.dialect, PostDialect::ClaudeNoop);
    }
}