lean-ctx-server 3.12.3

// SPDX-License-Identifier: Apache-2.0
use lean_ctx::core::archive;
use lean_ctx::core::context_admission::recovery;
use lean_ctx::core::context_handles::HandleRegistry;
use lean_ctx::core::context_ledger::ContextLedger;

/// The text of either outcome of [`handle_outcome`], for callers that only
/// print it. The registered tool takes the typed outcome.
pub fn handle(args: &serde_json::Value) -> String {
    match handle_outcome(args) {
        Ok(text) | Err(text) => text,
    }
}

/// `Ok` is the reply; `Err` is the text of a failed retrieval — the same text
/// [`handle`] answers. Retrieval declares every source
/// (`CtxExpandTool::publishes_every_source`), so its failure must reach the
/// registered tool as a failure and never as a reply that could be recorded
/// as a delivery. `list` and `search_all` declare nothing and answer as before.
pub fn handle_outcome(args: &serde_json::Value) -> Result<String, String> {
    let action = args
        .get("action")
        .and_then(|v| v.as_str())
        .unwrap_or("retrieve");

    match action {
        "list" => Ok(handle_list(args)),
        "search_all" => Ok(handle_search_all(args)),
        _ => handle_retrieve(args),
    }
}

fn normalized_handle_ref(id: &str) -> Option<&str> {
    let clean = id.strip_prefix('@').unwrap_or(id);
    if clean.len() < 2
        || !matches!(clean.chars().next()?, 'F' | 'S' | 'K' | 'M' | 'P')
        || !clean[1..].chars().all(|c| c.is_ascii_digit())
    {
        return None;
    }
    Some(clean)
}

/// Try to resolve a handle reference (@F1, @K1, etc.) to a file path.
/// Returns None if the ID is not a handle reference.
pub fn resolve_handle_ref(id: &str) -> Option<String> {
    let clean = normalized_handle_ref(id)?;
    let ledger = ContextLedger::load();
    let mut registry = HandleRegistry::new();
    for entry in &ledger.entries {
        if let (Some(item_id), Some(kind)) = (&entry.id, &entry.kind) {
            let phi = entry.phi.unwrap_or(0.5);
            let view_costs = entry.view_costs.clone().unwrap_or_else(|| {
                lean_ctx::core::context_field::ViewCosts::from_full_tokens(entry.original_tokens)
            });
            registry.register(
                item_id.clone(),
                *kind,
                &entry.path,
                &format!("{} {}L", entry.path, entry.original_tokens),
                &view_costs,
                phi,
                entry
                    .state
                    .as_ref()
                    .is_some_and(|s| *s == lean_ctx::core::context_field::ContextState::Pinned),
            );
        }
    }

    registry.resolve(clean).map(|h| h.source_path.clone())
}

fn handle_retrieve(args: &serde_json::Value) -> Result<String, String> {
    let Some(id) = args.get("id").and_then(|v| v.as_str()) else {
        return Err("ERROR: 'id' parameter is required. Use ctx_expand(action=\"list\") to see available archives, or pass a handle ref like @F1.".to_string());
    };

    // Handle reference resolution: @F1, @K1, @S1, etc.
    if let Some(path) = resolve_handle_ref(id) {
        // Expanding a file that was read compressed is that read's signal;
        // after a full read it is nobody's. An unknown read leaves the
        // current task's signals unknown.
        use lean_ctx::core::context_store::task_signals::{self, Signal};
        let last = lean_ctx::core::bounce_tracker::global()
            .lock()
            .ok()
            .and_then(|tracker| tracker.last_read(&path));
        match last {
            Some((true, Some(origin))) => task_signals::record(&origin, Signal::Expand),
            Some((false, _)) => {}
            Some((true, None)) | None => task_signals::record_unattributed(),
        }
        let mode = args.get("mode").and_then(|v| v.as_str()).unwrap_or("full");
        return Ok(format!(
            "[handle:{id} -> {path}]\nUse ctx_read(path=\"{path}\", mode=\"{mode}\") to load content."
        ));
    }

    // Unified tee-store handle (#482 / #936). One resolver, fixed precedence:
    // the proxy's prune / live-compression stubs (`proxy_<hash>`), the JSON
    // crusher's lossy originals (`json_<hash>`), AND every compressed shell
    // command's already-teed verbatim output (`<slug>_<8hex>.log`) all live in
    // the shared content-addressed store and resolve here, before the reference
    // (`ref_`) and archive (hex) stores below. The agent pulls back just the
    // slice it needs (head / tail / search / json_path / range) instead of
    // re-injecting the whole original — the surgical front-end the issue calls
    // "preferred when available". Admit the complete original before selecting.
    if let Some(path) = lean_ctx::core::ccr::resolve_tee(id) {
        // The tee store keeps no lineage: the reply's sources are unknown.
        archive::authority::mark_unbound();
        return expand_tee_file(&path, args);
    }

    // Resolve the entry's content once, then run the shared selector ladder.
    // Archive IDs are hex-only; reference IDs are `ref_`-prefixed — the prefix
    // picks the exact store, so the two stores differ only in *how* content is
    // resolved, never in *how* selectors are dispatched (#498). Resolving up
    // front also drops the archive path's per-selector disk re-reads.
    if id.starts_with("ref_") {
        let Some(content) = crate::server::reference_store::resolve(id) else {
            return Err(format!(
                "Reference '{id}' unavailable, expired (5-min TTL), or no longer authorized."
            ));
        };
        archive::authority::mark_unbound();
        // Recovery is a new delivery: re-admitted under the current policy
        // before any selector sees it (G5).
        let content = match recovery::admit_recovered(&content, "reference") {
            Ok(content) => content,
            Err(reason) => return Err(format!("ERROR: {reason}")),
        };
        return Ok(dispatch_selectors(id, &content, "Reference", args));
    }
    let archive_id = if id.starts_with("shell_") {
        match archive::resolve_alias(id) {
            Some(archive_id) => archive_id,
            None => match archive_finished_job(id)? {
                FinishedJob::Archived(archive_id) => archive_id,
                // A job that still runs, or that finished and printed
                // nothing, is a state of the job: a reply, not a failure.
                FinishedJob::Running(message) | FinishedJob::NoOutput(message) => {
                    return Ok(message);
                }
            },
        }
    } else {
        id.to_string()
    };
    // One resolver per id shape. A 64-hex id addresses a copy bound to a
    // lineage (project files, caller-supplied sources or both): it resolves
    // under the authority that holds NOW, and a refusal is a typed,
    // content-free tool error. Any other id is a public archive as before.
    let recovery_id = archive::RecoveryId::parse(&archive_id);
    let resolved = match &recovery_id {
        Some(recovery_id) => archive::retrieve_recovery_checked(recovery_id).map_err(Some),
        None => archive::retrieve(&archive_id).ok_or(None),
    };
    let content = match resolved {
        Ok(content) => content,
        Err(None | Some(archive::RecoveryRefusal::Missing)) => {
            return Err(format!(
                "Archive '{archive_id}' not found or expired. Use ctx_expand(action=\"list\") to see available archives. Archives stored before 3.11.5 belong to no project; the owner can adopt them with `lean-ctx migrate legacy-context`."
            ));
        }
        Err(Some(refusal)) => {
            return Err(format!(
                "ERROR: archive '{archive_id}' is not released ({refusal})"
            ));
        }
    };
    // The reply below is derived from this copy and from nothing else, so it
    // takes the copy's lineage: the checked resolver has just registered the
    // released text, and that lineage goes into the call's source capture
    // (this tool declares its sources; the release point records the reply
    // as delivered). A public archive has no lineage, and neither has a copy
    // the register does not know: unknown. The lineage is looked up for the
    // copy as resolved, and published only after the last step that can fail:
    // a call that fails publishes nothing.
    let lineage = recovery_id
        .and_then(|_| archive::delivered::lineage_of(&archive::delivered::digest(&content)));
    let content = match recovery::admit_recovered(&content, "archive") {
        Ok(content) => content,
        Err(reason) => return Err(format!("ERROR: {reason}")),
    };
    match lineage {
        Some(lineage) => lineage.publish_files(),
        None => archive::authority::mark_unbound(),
    }
    // K10: the agent fetched a truncated output back in full through its
    // pointer (no head/tail/search/json/range selector), counted by the
    // output class of the tool that produced it.
    {
        use lean_ctx_hive::metrics_v2::signals;
        if full_fetch(args)
            && signals::recording()
            && let Some(tool) = archive::tool_of(&archive_id)
            && lean_ctx::core::firewall::is_firewallable_tool(&tool)
        {
            signals::refetched(&signals::output_filter(&tool));
        }
    }
    Ok(dispatch_selectors(&archive_id, &content, "Archive", args))
}

/// Whether a retrieval returns the whole archive: no selector that cuts it
/// down ([`dispatch_selectors`]; a range needs both ends).
fn full_fetch(args: &serde_json::Value) -> bool {
    let has = |key: &str| args.get(key).is_some_and(|v| !v.is_null());
    let json_keys = args.get("json_keys").and_then(serde_json::Value::as_bool) == Some(true);
    let range = has("start_line") && has("end_line");
    !(has("head") || has("tail") || has("search") || has("json_path") || json_keys || range)
}

/// What [`archive_finished_job`] found for a known job.
enum FinishedJob {
    /// The archive id its output is stored under.
    Archived(String),
    /// The job still runs: the message that says so.
    Running(String),
    /// The job printed nothing: the message that says so.
    NoOutput(String),
}

/// GH #2005: a background job is archived when a status poll first reports it
/// finished, so `ctx_expand(id=<jobId>)` used to find nothing before that poll
/// and answered "expired" for a running job, a finished one and an unknown ID
/// alike. Ask the job table instead: archive a finished job's output now (the
/// same redaction and alias as the status path), and name the other two cases.
///
/// This is not a second way to the retained text: it takes the status path's
/// own checks. Output the job record holds back, or that the current storage
/// policy no longer admits, is refused with the status path's wording — it is
/// neither archived nor reported as a job that printed nothing.
///
/// `Err` is a failed retrieval (unknown job, withheld output, no archive). A
/// job that still runs and a finished job without output are replies
/// ([`FinishedJob::Running`], [`FinishedJob::NoOutput`]).
fn archive_finished_job(id: &str) -> Result<FinishedJob, String> {
    use crate::server::background_shell::{JobState, snapshot};
    let Some(snapshot) = snapshot(id) else {
        return Err(format!(
            "Background job '{id}' does not exist, or finished more than 5 minutes ago \
             without being polled or expanded."
        ));
    };
    let (output, label) = match &snapshot.state {
        JobState::Running { .. } => {
            return Ok(FinishedJob::Running(format!(
                "Background job '{id}' is still running. Expand it once it has finished, or \
                 read its progress with ctx_shell(background_action=\"status\", job_id=\"{id}\")."
            )));
        }
        JobState::Completed { output, exit_code } => (output.as_str(), format!("exit {exit_code}")),
        JobState::Cancelled { output } => (output.as_str(), "cancelled".to_string()),
    };
    let withheld = || {
        format!(
            "Background job '{id}' finished ({label}). \
             [lean-ctx gateway: retained output withheld by storage policy]"
        )
    };
    if snapshot.output_withheld {
        return Err(withheld());
    }
    if output.trim().is_empty() {
        return Ok(FinishedJob::NoOutput(format!(
            "Background job '{id}' finished ({label}) without output."
        )));
    }
    let Some(admission) =
        crate::tools::registered::ctx_shell_background::reauthorize_retained_output(
            id, &snapshot, output,
        )
    else {
        return Err(withheld());
    };
    let redacted = lean_ctx::core::redaction::redact_text_if_enabled(&admission.rendered());
    archive::store_background("ctx_shell", id, &redacted, None)
        .map(|stored| FinishedJob::Archived(stored.id))
        .ok_or_else(|| {
            format!(
                "Background job '{id}' finished ({label}), but its output could not be archived."
            )
        })
}

/// Apply the structured selector ladder (head / tail / json_keys / search /
/// range / full) to already-resolved `content`. Resolving once and formatting
/// in-memory lets the archive and reference stores share a single code path and
/// the same `archive::format_*` formatters, so output is byte-identical
/// regardless of which store backed the ID (#498). `noun` is the capitalised
/// store name used in messages — `"Archive"` or `"Reference"`.
fn dispatch_selectors(id: &str, content: &str, noun: &str, args: &serde_json::Value) -> String {
    let label = format!("{} {id}", noun.to_ascii_lowercase());

    if let Some(n) = args.get("head").and_then(serde_json::Value::as_u64) {
        let n = n as usize;
        return format!(
            "{noun} {id} head {n}:\n{}",
            archive::format_range(content, 1, n)
        );
    }
    if let Some(n) = args.get("tail").and_then(serde_json::Value::as_u64) {
        let n = n as usize;
        let total = content.lines().count();
        let start = if total > n { total - n + 1 } else { 1 };
        return format!(
            "{noun} {id} tail {n}:\n{}",
            archive::format_range(content, start, total)
        );
    }
    if args.get("json_keys").and_then(serde_json::Value::as_bool) == Some(true)
        || args.get("json_path").is_some()
    {
        let path = args.get("json_path").and_then(|v| v.as_str());
        return match archive::format_json_keys(content, path, &label) {
            Some(out) => out,
            None => format!(
                "{noun} '{id}' is not valid JSON. Use ctx_expand(id=\"{id}\") for raw content."
            ),
        };
    }
    if let Some(pattern) = args.get("search").and_then(|v| v.as_str()) {
        return archive::format_search(content, pattern, &label);
    }

    let start = args
        .get("start_line")
        .and_then(serde_json::Value::as_u64)
        .map(|v| v as usize);
    let end = args
        .get("end_line")
        .and_then(serde_json::Value::as_u64)
        .map(|v| v as usize);
    if let (Some(s), Some(e)) = (start, end) {
        return format!(
            "{noun} {id} lines {s}-{e}:\n{}",
            archive::format_range(content, s, e)
        );
    }

    let lines = content.lines().count();
    let chars = content.len();
    format!("{noun} {id} ({chars} chars, {lines} lines):\n{content}")
}

/// Surgical retrieval over a CCR proxy tee file (#482). Mirrors the archive
/// selectors (head / tail / search / json_path / range / full) but operates on
/// the verbatim tee content on disk, so the agent pulls back only the slice it
/// needs rather than undoing the proxy's compression with a full re-inject.
///
/// `Err` is the error text of a tee file that could not be read or admitted.
fn expand_tee_file(path: &std::path::Path, args: &serde_json::Value) -> Result<String, String> {
    match lean_ctx::core::ccr::read_tee_detailed(path) {
        Ok(content) => Ok(tee_selectors(path, &content, args)),
        Err(reason) => Err(format!("ERROR: {reason}")),
    }
}

/// The selector ladder of [`expand_tee_file`] over the admitted tee `content`.
fn tee_selectors(path: &std::path::Path, content: &str, args: &serde_json::Value) -> String {
    let label = path.file_name().and_then(|n| n.to_str()).unwrap_or("ccr");

    if let Some(n) = args.get("head").and_then(serde_json::Value::as_u64) {
        return format!(
            "[ccr {label}] head {n}:\n{}",
            head_lines(&content, n as usize)
        );
    }
    if let Some(n) = args.get("tail").and_then(serde_json::Value::as_u64) {
        return format!(
            "[ccr {label}] tail {n}:\n{}",
            tail_lines(&content, n as usize)
        );
    }
    if args.get("json_keys").and_then(serde_json::Value::as_bool) == Some(true)
        || args.get("json_path").is_some()
    {
        let jp = args.get("json_path").and_then(|v| v.as_str());
        return match json_view(&content, jp) {
            Some(out) => format!("[ccr {label}] json {}:\n{out}", jp.unwrap_or("(keys)")),
            None => format!(
                "[ccr {label}] not valid JSON or path not found. Use ctx_expand(id=\"{label}\") for raw content."
            ),
        };
    }
    if let Some(pattern) = args.get("search").and_then(|v| v.as_str()) {
        return format!(
            "[ccr {label}] search \"{pattern}\":\n{}",
            search_lines(&content, pattern)
        );
    }
    let start = args
        .get("start_line")
        .and_then(serde_json::Value::as_u64)
        .map(|v| v as usize);
    let end = args
        .get("end_line")
        .and_then(serde_json::Value::as_u64)
        .map(|v| v as usize);
    if let (Some(s), Some(e)) = (start, end) {
        return format!(
            "[ccr {label}] lines {s}-{e}:\n{}",
            range_lines(&content, s, e)
        );
    }

    let lines = content.lines().count();
    format!(
        "[ccr {label}] ({} chars, {lines} lines):\n{content}",
        content.len()
    )
}

fn head_lines(s: &str, n: usize) -> String {
    s.lines().take(n).collect::<Vec<_>>().join("\n")
}

fn tail_lines(s: &str, n: usize) -> String {
    let v: Vec<&str> = s.lines().collect();
    let start = v.len().saturating_sub(n);
    v[start..].join("\n")
}

/// 1-indexed inclusive line range, clamped to the available lines.
fn range_lines(s: &str, start: usize, end: usize) -> String {
    let v: Vec<&str> = s.lines().collect();
    let a = start.saturating_sub(1).min(v.len());
    let b = end.min(v.len());
    if a >= b {
        return String::new();
    }
    v[a..b].join("\n")
}

/// Case-insensitive, like the archive and reference stores
/// (`archive::format_search`) — #2172: the tee store alone matched case-exactly,
/// so the same `search` that finds a line in an archive answered "no lines
/// match" for a tee log that contains it.
fn search_lines(s: &str, pattern: &str) -> String {
    let pattern_lower = pattern.to_lowercase();
    let hits: Vec<String> = s
        .lines()
        .enumerate()
        .filter(|(_, l)| l.to_lowercase().contains(&pattern_lower))
        .map(|(i, l)| format!("{}: {}", i + 1, l))
        .collect();
    if hits.is_empty() {
        format!("(no lines match \"{pattern}\")")
    } else {
        hits.join("\n")
    }
}

/// `json_path` navigation over the tee content: object segments by key, array
/// segments by numeric index, dot-separated. Empty path lists the root keys.
/// Objects render as their key list; scalars/arrays pretty-print.
fn json_view(s: &str, path: Option<&str>) -> Option<String> {
    let root: serde_json::Value = serde_json::from_str(s).ok()?;
    let target = match path {
        Some(p) if !p.is_empty() => navigate_json(&root, p)?,
        _ => &root,
    };
    if let Some(obj) = target.as_object() {
        Some(obj.keys().cloned().collect::<Vec<_>>().join("\n"))
    } else {
        serde_json::to_string_pretty(target).ok()
    }
}

fn navigate_json<'a>(v: &'a serde_json::Value, path: &str) -> Option<&'a serde_json::Value> {
    let mut cur = v;
    for seg in path.split('.').filter(|s| !s.is_empty()) {
        cur = match seg.parse::<usize>() {
            Ok(idx) => cur.get(idx)?,
            Err(_) => cur.get(seg)?,
        };
    }
    Some(cur)
}

fn handle_search_all(args: &serde_json::Value) -> String {
    let query = match args.get("query").and_then(|v| v.as_str()) {
        Some(q) if !q.is_empty() => q,
        _ => return "ERROR: 'query' parameter required for search_all.".to_string(),
    };
    let limit = args
        .get("limit")
        .and_then(serde_json::Value::as_u64)
        .unwrap_or(10) as usize;

    let results = lean_ctx::core::archive_fts::search(query, limit);
    if results.is_empty() {
        return format!(
            "No archives match \"{query}\". Indexed: {} entries.",
            lean_ctx::core::archive_fts::entry_count()
        );
    }

    let mut out = format!("{} result(s) for \"{}\":\n", results.len(), query);
    for r in &results {
        // Index rows are recovered text too: each row is re-admitted as a
        // whole, and a row the gateway withholds is listed without content.
        let row = format!("{} | …{}…", r.command, r.snippet);
        match recovery::admit_recovered(&row, "archive-index") {
            Ok(row) => out.push_str(&format!("  {} | {} | {row}\n", r.archive_id, r.tool)),
            Err(_) => out.push_str(&format!(
                "  {} | {} | [withheld by the context gateway]\n",
                r.archive_id, r.tool
            )),
        }
    }
    out.push_str("\nRetrieve full: ctx_expand(id=\"<archive_id>\")");
    out
}

fn handle_list(args: &serde_json::Value) -> String {
    let session_id = args.get("session_id").and_then(|v| v.as_str());
    let entries = archive::list_entries(session_id);

    if entries.is_empty() {
        return "No archives found.".to_string();
    }

    let mut out = format!("{} archive(s):\n", entries.len());
    for e in &entries {
        out.push_str(&format!(
            "  {} | {} | {} | {} chars ({} tok) | {}\n",
            e.id,
            e.tool,
            e.command,
            e.size_chars,
            e.size_tokens,
            e.created_at.format("%H:%M:%S")
        ));
    }
    out.push_str("\nRetrieve: ctx_expand(id=\"<id>\")");
    out.push_str("\nSearch: ctx_expand(id=\"<id>\", search=\"ERROR\")");
    out.push_str("\nRange: ctx_expand(id=\"<id>\", start_line=10, end_line=50)");
    out.push_str("\nHead/Tail: ctx_expand(id=\"<id>\", head=120) | tail=40");
    out.push_str("\nJSON: ctx_expand(id=\"<id>\", json_keys=true) | json_path=\"data.items\"");
    out
}

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

    #[test]
    fn handle_missing_id_returns_error() {
        let result = handle(&json!({}));
        assert!(result.contains("ERROR"));
        assert!(result.contains("id"));
    }

    #[test]
    fn handle_nonexistent_returns_not_found() {
        let result = handle(&json!({"id": "nonexistent_xyz"}));
        assert!(result.contains("not found"));
    }

    /// `ctx_expand` resolves a lineage-bound recovery copy through the one
    /// checked resolver: a file-bound id as before; a copy bound to a
    /// caller-supplied source only while its presentation is current, with a
    /// typed, content-free error otherwise. Before, the handler refused every
    /// lineage-bound copy as "not found".
    #[test]
    fn lineage_bound_copies_resolve_under_current_authority_only() {
        use lean_ctx::core::archive::delivered::Lineage;
        use lean_ctx::core::archive::presented::{PresentedBinding, record_presentation};
        let data = lean_ctx::core::data_dir::isolated_data_dir();
        let _community = lean_ctx::core::policy::runtime::TestPolicyOverride::set(None);
        let write_config = |config: &str| {
            let dir = lean_ctx::core::paths::config_dir_read_only().expect("isolated config dir");
            std::fs::create_dir_all(&dir).expect("config dir");
            std::fs::write(dir.join("config.toml"), config).expect("config");
        };
        write_config("[context_gateway]\nmode = \"governed\"\n");
        let directory = tempfile::tempdir().expect("project");
        let root = directory.path().canonicalize().expect("canonical root");
        lean_ctx::core::policy::runtime::REQUEST_PROJECT.sync_scope(
            std::cell::RefCell::new(Some(root.clone())),
            || {
                // File-bound copy: resolves, and is refused typed once its
                // source is gone.
                let source = root.join("kept.rs");
                let original = "pub fn kept_quartz() {}\n";
                std::fs::write(&source, original).expect("source");
                let admitted = crate::tools::ctx_read::read_file_for_tool_rooted(
                    source.to_str().expect("utf-8 path"),
                    root.to_str().expect("utf-8 root"),
                    "ctx_read",
                )
                .expect("admitted read");
                let files = lean_ctx::core::archive::authority::ArchiveAuthority::file(
                    &root,
                    source.to_str().expect("utf-8 path"),
                    "ctx_read",
                    &admitted,
                )
                .expect("bound read");
                let file_id =
                    archive::store_recovery_with_lineage(original, &Lineage::of_files(files))
                        .expect("stored");
                let expanded = handle(&json!({"id": file_id.as_str()}));
                assert!(expanded.contains("kept_quartz"), "{expanded}");
                std::fs::remove_file(&source).expect("delete the source");
                let refused = handle(&json!({"id": file_id.as_str()}));
                assert!(
                    refused.contains("recovery.source_no_longer_admitted"),
                    "{refused}"
                );
                assert!(!refused.contains("kept_quartz"), "{refused}");

                // Copy bound to a caller-supplied source.
                let prepared = "## leanctx-source-v1\ninvoice ledger basalt\n";
                let binding =
                    PresentedBinding::admitted_now(&root, "ref-one", "source-one", "sha256:aa")
                        .expect("policy identity available");
                assert!(record_presentation(&binding));
                let id = archive::store_recovery_with_lineage(
                    prepared,
                    &Lineage::of_presented(vec![binding.clone()]),
                )
                .expect("stored");
                let expanded = handle(&json!({"id": id.as_str()}));
                assert!(expanded.contains("invoice ledger basalt"), "{expanded}");

                // Presentation lapsed.
                std::fs::remove_dir_all(
                    data.path()
                        .join("archives")
                        .join(lean_ctx::core::archive::presented::REGISTER_DIRECTORY),
                )
                .expect("presentation register");
                let lapsed = handle(&json!({"id": id.as_str()}));
                assert!(lapsed.contains("recovery.source_not_presented"), "{lapsed}");
                assert!(!lapsed.contains("basalt"), "{lapsed}");
                assert!(record_presentation(&binding));
                assert!(handle(&json!({"id": id.as_str()})).contains("basalt"));

                // Policy changed.
                write_config("[context_gateway]\nmode = \"sovereign\"\n");
                let changed = handle(&json!({"id": id.as_str()}));
                assert!(
                    changed.contains("recovery.source_no_longer_admitted"),
                    "{changed}"
                );
                assert!(!changed.contains("basalt"), "{changed}");

                // An unknown canonical id is still "not found".
                let absent = handle(&json!({"id": "0".repeat(64)}));
                assert!(absent.contains("not found"), "{absent}");
            },
        );
    }

    #[test]
    fn handle_list_empty() {
        let result = handle(&json!({"action": "list"}));
        assert!(
            result.contains("No archives") || result.contains("archive(s)"),
            "unexpected: {result}"
        );
    }

    #[test]
    fn only_an_unselected_retrieval_is_a_full_fetch_back() {
        assert!(full_fetch(&json!({"id": "abc"})));
        assert!(full_fetch(&json!({"id": "abc", "start_line": 5})));
        assert!(full_fetch(&json!({"id": "abc", "json_keys": false})));
        for selected in [
            json!({"id": "abc", "head": 10}),
            json!({"id": "abc", "tail": 10}),
            json!({"id": "abc", "search": "ERROR"}),
            json!({"id": "abc", "json_keys": true}),
            json!({"id": "abc", "json_path": "a.b"}),
            json!({"id": "abc", "start_line": 1, "end_line": 80}),
        ] {
            assert!(!full_fetch(&selected), "{selected}");
        }
    }

    #[test]
    fn text_selectors_slice_correctly() {
        let body = (1..=10)
            .map(|i| format!("line {i}"))
            .collect::<Vec<_>>()
            .join("\n");
        assert_eq!(head_lines(&body, 2), "line 1\nline 2");
        assert_eq!(tail_lines(&body, 2), "line 9\nline 10");
        assert_eq!(range_lines(&body, 3, 4), "line 3\nline 4");
        assert!(search_lines(&body, "line 7").contains("7: line 7"));
        assert!(search_lines(&body, "zzz").contains("no lines match"));
    }

    /// #2172: a tee-log search answered "no lines match" for a line the log
    /// holds when the case differed; archives always matched it.
    #[test]
    fn tee_search_matches_like_the_archive_search() {
        let log = "Compiling lean-ctx\nerror[E0308]: mismatched types\nRegistrationController ok";
        for pattern in ["ERROR[E0308]", "registrationcontroller"] {
            let hits = search_lines(log, pattern);
            assert!(!hits.contains("no lines match"), "{pattern}: {hits}");
            assert!(
                !archive::format_search(log, pattern, "archive x").starts_with("No matches"),
                "precondition: the archive search finds {pattern}"
            );
        }
        assert!(search_lines(log, "ERROR[E0308]").starts_with("2: error[E0308]"));
    }

    #[test]
    fn json_view_lists_keys_and_navigates() {
        let doc = r#"{"a":{"b":[10,20,30]},"c":1}"#;
        assert_eq!(json_view(doc, None).unwrap(), "a\nc");
        assert_eq!(json_view(doc, Some("a")).unwrap(), "b");
        assert_eq!(json_view(doc, Some("a.b.1")).unwrap(), "20");
        assert!(json_view(doc, Some("a.missing")).is_none());
        assert!(json_view("not json", None).is_none());
    }

    #[test]
    fn ctx_expand_retrieves_proxy_tee_handle_surgically() {
        let _lock = lean_ctx::core::data_dir::test_env_lock();
        let _community = lean_ctx::core::policy::runtime::TestPolicyOverride::set(None);
        // Mimic what the proxy does: persist a verbatim original to the tee store
        // and hand the agent its content-addressed handle.
        let original = (1..=60)
            .map(|i| format!("output row {i:03}"))
            .collect::<Vec<_>>()
            .join("\n");
        assert!(original.len() >= lean_ctx::core::ccr::MIN_TEE_BYTES);
        let tee_handle = lean_ctx::core::ccr::persist(&original).expect("tee handle");

        // Full content via the handle path (proxy-only fallback also reads this).
        let full = handle(&json!({"id": tee_handle}));
        assert!(full.contains("output row 001") && full.contains("output row 060"));

        // Surgical slices via the bare hash form the stub can also carry.
        let hash = lean_ctx::core::hasher::hash_short(&original);
        let head = handle(&json!({"id": hash, "head": 2}));
        assert!(head.contains("output row 001") && !head.contains("output row 010"));
        let search = handle(&json!({"id": hash, "search": "row 042"}));
        assert!(search.contains("output row 042") && !search.contains("output row 001"));
    }

    #[test]
    fn ctx_expand_retrieves_shell_tee_output_surgically() {
        let _lock = lean_ctx::core::data_dir::test_env_lock();
        let _community = lean_ctx::core::policy::runtime::TestPolicyOverride::set(None);
        // Every compressed shell command already tees its verbatim output to the
        // shared store (#936). With the unified resolver, ctx_expand can slice
        // that tee surgically — the agent no longer has to re-read the whole file.
        let original = (1..=60)
            .map(|i| format!("api response row {i:03}"))
            .collect::<Vec<_>>()
            .join("\n");
        let tee_path =
            lean_ctx::shell::save_tee("gh api /repos/foo/bar", &original).expect("shell tee saved");
        let name = std::path::Path::new(&tee_path)
            .file_name()
            .and_then(|n| n.to_str())
            .unwrap()
            .to_string();

        // Full content via the bare shell-tee basename the footer advertises.
        let full = handle(&json!({ "id": name.clone() }));
        assert!(full.contains("api response row 001") && full.contains("api response row 060"));

        // Surgical slices: head and search, same ladder as proxy/archive handles.
        let head = handle(&json!({ "id": name.clone(), "head": 2 }));
        assert!(head.contains("api response row 001") && !head.contains("api response row 010"));
        let search = handle(&json!({ "id": name, "search": "row 042" }));
        assert!(search.contains("api response row 042") && !search.contains("row 001"));
    }

    #[test]
    fn ctx_expand_retrieves_json_crush_original() {
        let _lock = lean_ctx::core::data_dir::test_env_lock();
        let _community = lean_ctx::core::policy::runtime::TestPolicyOverride::set(None);
        // The lossy JSON crusher persists its verbatim original under json_<hash>
        // (#936). The dropped columns must be recoverable through the same
        // ctx_expand path the stub/footer advertises.
        let original = (1..=50)
            .map(|i| format!(r#"{{"id":{i},"ts":"2026-06-22T10:00:{i:02}Z"}}"#))
            .collect::<Vec<_>>()
            .join("\n");
        assert!(original.len() >= lean_ctx::core::ccr::MIN_TEE_BYTES);
        let handle_path = lean_ctx::core::ccr::persist_json(&original).expect("json tee handle");
        assert!(handle_path.contains("json_"), "json_ prefix: {handle_path}");

        let out = handle(&json!({ "id": handle_path }));
        assert!(
            out.contains("2026-06-22T10:00:42Z"),
            "dropped column recoverable: {out}"
        );
    }

    /// #1911: the HTML→markdown crush footer tells the agent to call
    /// `ctx_expand(id=<html_ handle>, search=…)`; that exact call must slice the
    /// persisted page instead of reporting "not found or expired".
    #[test]
    fn ctx_expand_retrieves_html_crush_original() {
        let _lock = lean_ctx::core::data_dir::test_env_lock();
        let _community = lean_ctx::core::policy::runtime::TestPolicyOverride::set(None);
        let original = (1..=40)
            .map(|i| format!("<h2>Section {i}</h2><p>body of section {i}</p>"))
            .collect::<Vec<_>>()
            .join("\n");
        let handle_path = lean_ctx::core::ccr::persist_html(&original).expect("html tee handle");
        assert!(handle_path.contains("html_"), "html_ prefix: {handle_path}");

        let out = handle(&json!({ "id": handle_path, "search": "Section 7<" }));
        assert!(
            out.contains("body of section 7<") && !out.contains("body of section 8<"),
            "html section recoverable: {out}"
        );
        assert!(!out.contains("not found"), "{out}");
    }

    #[test]
    fn ctx_expand_resolves_reference_store_ids() {
        let _data = lean_ctx::core::data_dir::isolated_data_dir();
        let _policy = lean_ctx::core::policy::runtime::TestPolicyOverride::set(None);
        // #498: `ref_`-prefixed IDs route to the in-memory reference store, not
        // the on-disk archive. Exercises the resolve-then-dispatch ladder end to
        // end through the public `handle` entry point.
        let body = (1..=40)
            .map(|i| format!("ref row {i}"))
            .collect::<Vec<_>>()
            .join("\n");
        let id = crate::server::reference_store::store(&body).unwrap();
        assert!(id.starts_with("ref_"), "store must mint a ref_ id: {id}");

        let full = handle(&json!({"id": id}));
        assert!(
            full.contains("Reference") && full.contains("ref row 1") && full.contains("ref row 40"),
            "full: {full}"
        );

        let head = handle(&json!({"id": id, "head": 3}));
        assert!(head.contains("ref row 1") && head.contains("ref row 3"));
        assert!(!head.contains("ref row 4"), "head leaked row 4: {head}");

        let tail = handle(&json!({"id": id, "tail": 2}));
        assert!(tail.contains("ref row 39") && tail.contains("ref row 40"));
        assert!(!tail.contains("ref row 38"), "tail leaked row 38: {tail}");

        let search = handle(&json!({"id": id, "search": "ref row 7"}));
        assert!(search.contains("ref row 7") && !search.contains("ref row 1\n"));

        let range = handle(&json!({"id": id, "start_line": 5, "end_line": 6}));
        assert!(range.contains("ref row 5") && range.contains("ref row 6"));
        assert!(!range.contains("ref row 4") && !range.contains("ref row 7"));
    }

    #[test]
    fn ctx_expand_reference_json_keys_and_missing() {
        let _data = lean_ctx::core::data_dir::isolated_data_dir();
        let _policy = lean_ctx::core::policy::runtime::TestPolicyOverride::set(None);
        let id = crate::server::reference_store::store(r#"{"a":1,"b":[1,2,3]}"#).unwrap();
        let keys = handle(&json!({"id": id, "json_keys": true}));
        assert!(keys.contains("object (2 keys)"), "got: {keys}");
        assert!(keys.contains("array(3)"), "got: {keys}");

        // An expired/unknown ref must explain the 5-min TTL, not fall through to
        // the archive's "not found" message.
        let missing = handle(&json!({"id": "ref_deadbeefcafef00d"}));
        assert!(
            missing.contains("unavailable") && missing.contains("5-min TTL"),
            "got: {missing}"
        );
    }

    /// The formatted reply (selector applied, header added) of a file-bound
    /// recovery copy is registered under the copy's file lineage and fails
    /// together with its source. A copy bound to a caller-presented source
    /// resolves, but its reply stays unknown: the capture carries files only.
    #[test]
    fn a_file_bound_reply_is_registered_and_fails_with_its_source() {
        use lean_ctx::core::archive::delivered::{Lineage, LineageRefusal};
        use lean_ctx::core::archive::presented::{PresentedBinding, record_presentation};
        use lean_ctx::core::context_admission::stored::lineage_support::{
            deliver, governed, project, reusable, scoped,
        };
        let _data = governed();
        let _community = lean_ctx::core::policy::runtime::TestPolicyOverride::set(None);
        let (_directory, root) = project();
        let source = format!("{root}/kept.rs");
        let original = "pub fn kept_quartz() {}\nfn second_line() {}\n";
        std::fs::write(&source, original).expect("source");
        let (file_id, presented_id) = scoped(&root, || {
            let admitted =
                crate::tools::ctx_read::read_file_for_tool_rooted(&source, &root, "ctx_read")
                    .expect("admitted read");
            let files = lean_ctx::core::archive::authority::ArchiveAuthority::file(
                std::path::Path::new(&root),
                &source,
                "ctx_read",
                &admitted,
            )
            .expect("bound read");
            let file_id = archive::store_recovery_with_lineage(original, &Lineage::of_files(files))
                .expect("stored");
            let binding = PresentedBinding::admitted_now(
                std::path::Path::new(&root),
                "ref-one",
                "source-one",
                "sha256:aa",
            )
            .expect("policy identity available");
            assert!(record_presentation(&binding));
            let presented_id = archive::store_recovery_with_lineage(
                "## leanctx-source-v1\ninvoice ledger basalt\n",
                &Lineage::of_presented(vec![binding]),
            )
            .expect("stored");
            (file_id, presented_id)
        });

        let (reply, recorded) = deliver(&root, || {
            handle(&json!({"id": file_id.as_str(), "head": 1}))
        });
        assert!(reply.contains("kept_quartz"), "{reply}");
        assert!(
            !reply.contains("second_line"),
            "the selector applied: {reply}"
        );
        assert_eq!(recorded, 1, "the formatted reply is registered");
        assert_eq!(reusable(&root, &reply), Ok(()));

        let (presented, recorded) =
            deliver(&root, || handle(&json!({"id": presented_id.as_str()})));
        assert!(presented.contains("invoice ledger basalt"), "{presented}");
        assert_eq!(recorded, 0);
        assert_eq!(reusable(&root, &presented), Err(LineageRefusal::Unknown));

        std::fs::remove_file(&source).expect("delete the source");
        assert_eq!(
            reusable(&root, &reply),
            Err(LineageRefusal::SourceNoLongerAdmitted)
        );
        let (refused, recorded) = deliver(&root, || {
            handle(&json!({"id": file_id.as_str(), "head": 1}))
        });
        assert!(
            refused.contains("recovery.source_no_longer_admitted"),
            "{refused}"
        );
        assert_eq!(recorded, 0);
    }
}