zeph_core/agent/

agent_access_impl.rs

// SPDX-FileCopyrightText: 2026 Andrei G <bug-ops>
// SPDX-License-Identifier: MIT OR Apache-2.0

//! Implementation of [`zeph_commands::traits::agent::AgentAccess`] for [`Agent<C>`].
//!
//! Each method in `AgentAccess` returns a formatted `String` result (without sending to the
//! channel directly), so that `CommandContext::sink` does not conflict with this borrow.
//! The one exception is methods for subsystems that are already channel-free (memory, graph).
//!
//! [`Agent<C>`]: super::Agent

use std::fmt::Write as _;
use std::future::Future;
use std::pin::Pin;
use std::sync::Arc;
use std::time::Duration;

use tracing::Instrument as _;
use zeph_commands::CommandError;
use zeph_commands::traits::agent::AgentAccess;
use zeph_db;
use zeph_llm::provider::LlmProvider as _;
use zeph_memory::semantic::SemanticMemory;
use zeph_memory::{Edge, Entity, GraphExtractionConfig, GraphStore, MessageId, extract_and_store};

use super::{Agent, error::AgentError};
use crate::channel::Channel;

impl<C: Channel + Send + 'static> Agent<C> {
    fn resolve_graph_store(&self) -> Result<(Arc<SemanticMemory>, Arc<GraphStore>), String> {
        let Some(memory) = self.services.memory.persistence.memory.clone() else {
            return Err("Graph memory is not enabled.".to_owned());
        };
        let Some(store) = memory.graph_store.clone() else {
            if self.services.memory.extraction.graph_config.enabled {
                return Err(
                    "Graph memory enabled but vector store unavailable (Qdrant unreachable)."
                        .to_owned(),
                );
            }
            return Err("Graph memory is not enabled.".to_owned());
        };
        Ok((memory, store))
    }
}

/// Outcome of resolving an entity by display name against the graph store: either the
/// entity was found, or a user-facing message that the caller should return as-is (no
/// match, or the store timed out).
enum EntityLookup {
    Found(Entity),
    Message(String),
}

/// Outcome of a graph-store call bounded by [`with_graph_store_timeout`]'s 5s deadline:
/// either it completed, or it timed out (Qdrant unreachable).
enum StoreCallOutcome<T> {
    Completed(T),
    TimedOut,
}

/// Runs `fut` under a 5s timeout — the deadline shared by `resolve_entity_by_name` and the
/// edge lookups in `graph_facts`/`graph_history`. Maps a store error to [`CommandError`];
/// logs and reports a timeout via [`StoreCallOutcome::TimedOut`] so callers only need to
/// turn that into their own user-facing message.
async fn with_graph_store_timeout<T>(
    fut: impl Future<Output = Result<T, zeph_memory::MemoryError>>,
) -> Result<StoreCallOutcome<T>, CommandError> {
    match tokio::time::timeout(Duration::from_secs(5), fut).await {
        Ok(Ok(v)) => Ok(StoreCallOutcome::Completed(v)),
        Ok(Err(e)) => Err(CommandError::new(e.to_string())),
        Err(_) => {
            tracing::warn!("graph store call timed out after 5s (Qdrant unreachable)");
            Ok(StoreCallOutcome::TimedOut)
        }
    }
}

/// Resolves `name` to an [`Entity`] via [`GraphStore::find_entity_by_name`], bounded by a
/// 5s timeout — the lookup block shared by `graph_facts` and `graph_history`.
async fn resolve_entity_by_name(
    store: &GraphStore,
    name: &str,
) -> Result<EntityLookup, CommandError> {
    let matches = match with_graph_store_timeout(store.find_entity_by_name(name)).await? {
        StoreCallOutcome::Completed(v) => v,
        StoreCallOutcome::TimedOut => {
            return Ok(EntityLookup::Message(
                "Graph store unavailable (Qdrant unreachable).".to_owned(),
            ));
        }
    };
    let Some(entity) = matches.into_iter().next() else {
        return Ok(EntityLookup::Message(format!(
            "No entity found matching '{name}'."
        )));
    };
    Ok(EntityLookup::Found(entity))
}

/// Builds the `entity_id -> display_name` lookup map shared by `graph_facts` and
/// `graph_history`: seeds `entity`'s own name, inserts a placeholder for every edge
/// endpoint, then resolves each placeholder via [`GraphStore::find_entity_by_id`] (5s
/// timeout), falling back to `#{id}` when the lookup fails or times out.
async fn build_entity_name_map(
    store: &GraphStore,
    entity: &Entity,
    edges: &[Edge],
) -> std::collections::HashMap<i64, String> {
    let mut entity_names: std::collections::HashMap<i64, String> = std::collections::HashMap::new();
    entity_names.insert(entity.id.0, entity.name.clone());
    for edge in edges {
        entity_names.entry(edge.source_entity_id).or_default();
        entity_names.entry(edge.target_entity_id).or_default();
    }
    for (&id, name_val) in &mut entity_names {
        if name_val.is_empty() {
            let result =
                tokio::time::timeout(Duration::from_secs(5), store.find_entity_by_id(id)).await;
            if let Ok(Ok(Some(other))) = result {
                *name_val = other.name;
            } else {
                *name_val = format!("#{id}");
            }
        }
    }
    entity_names
}

/// Run Stage-2 LLM semantic scan for all skills in the plugin at `source`.
///
/// Skills are scanned concurrently with up to 4 in-flight at a time
/// (`buffer_unordered(4)`). An aggregate 5-minute timeout wraps the whole
/// batch; each individual scan is already bounded by `SCAN_TIMEOUT` (30 s) in
/// `SkillSemanticScanner`. Returns `Some(err_msg)` when any skill is blocked,
/// `None` when all skills pass.
///
/// Each future carries its own `skill_name` so that the rejection message names
/// the correct skill regardless of completion order (which differs from input
/// order when futures complete out-of-order with `buffer_unordered`).
async fn semantic_scan_plugin_add(
    scanner: &zeph_skills::semantic_scanner::SkillSemanticScanner,
    source: &str,
    managed_dir: Option<std::path::PathBuf>,
    mcp_allowed: Vec<String>,
    base_shell_allowed: Vec<String>,
) -> Result<Option<String>, CommandError> {
    use futures::stream::StreamExt as _;
    use zeph_skills::semantic_scanner::ScanVerdict;

    let plugins_dir = zeph_plugins::PluginManager::default_plugins_dir();
    let mgr_dir =
        managed_dir.unwrap_or_else(|| zeph_config::defaults::default_vault_dir().join("skills"));
    let mgr =
        zeph_plugins::PluginManager::new(plugins_dir, mgr_dir, mcp_allowed, base_shell_allowed);

    let source_owned = source.to_owned();
    let scan_inputs = tokio::task::spawn_blocking(move || mgr.scan_targets(&source_owned))
        .await
        .map_err(|e| CommandError(format!("plugin scan_targets panicked: {e}")))?
        .map_err(|e| CommandError(format!("plugin add failed: {e}")))?;

    tracing::info!(
        plugin.source = %source,
        skills_count = scan_inputs.len(),
        "plugins.add: running Stage-2 semantic scan"
    );

    // Scan all skills concurrently with up to 4 in-flight. Each individual scan is
    // already bounded by SCAN_TIMEOUT (30 s); the outer 5-min cap guards the batch.
    // Each future owns its skill_name so verdicts carry the correct name regardless
    // of buffer_unordered completion order (which is not the same as input order).
    let scan_futs: Vec<_> = scan_inputs
        .iter()
        .map(|input| {
            let name = input.skill_name.clone();
            let purpose = input.declared_purpose.clone();
            let md = input.skill_md.clone();
            async move {
                let verdict = scanner.scan(&name, &purpose, &md).await;
                (name, verdict)
            }
        })
        .collect();

    let verdicts: Vec<_> = tokio::time::timeout(
        std::time::Duration::from_mins(5),
        futures::stream::iter(scan_futs)
            .buffer_unordered(4)
            .collect::<Vec<_>>(),
    )
    .await
    .map_err(|_| CommandError("plugin scan timed out after 300s".to_owned()))?;

    for (skill_name, verdict_result) in verdicts {
        let verdict = verdict_result.map_err(|e| {
            CommandError(format!(
                "plugin add failed: semantic scan error for skill {skill_name:?}: {e}"
            ))
        })?;
        match verdict {
            ScanVerdict::Allow => {
                tracing::debug!(
                    skill = %skill_name,
                    "plugins.add: skill passed semantic scan"
                );
            }
            ScanVerdict::Warn(ref reason) => {
                tracing::warn!(
                    skill = %skill_name,
                    reason = %reason,
                    "plugins.add: skill passed with warning"
                );
            }
            ScanVerdict::Block(reason) => {
                return Ok(Some(format!(
                    "plugin add failed: skill {skill_name:?} rejected by semantic scan: {reason}"
                )));
            }
            _ => {}
        }
    }
    Ok(None)
}

impl<C: Channel + Send + 'static> AgentAccess for Agent<C> {
    // ----- /memory -----

    fn memory_tiers<'a>(
        &'a mut self,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(
            async move {
                let Some(memory) = self.services.memory.persistence.memory.clone() else {
                    return Ok("Memory not configured.".to_owned());
                };
                match memory.sqlite().count_messages_by_tier().await {
                    Ok((episodic, semantic)) => {
                        let mut out = String::new();
                        let _ = writeln!(out, "Memory tiers:");
                        let _ = writeln!(out, "  Working:  (current context window — virtual)");
                        let _ = writeln!(out, "  Episodic: {episodic} messages");
                        let _ = writeln!(out, "  Semantic: {semantic} facts");
                        Ok(out.trim_end().to_owned())
                    }
                    Err(e) => Ok(format!("Failed to query tier stats: {e}")),
                }
            }
            .instrument(tracing::info_span!("core.agent_access.memory_tiers")),
        )
    }

    fn memory_promote<'a>(
        &'a mut self,
        ids_str: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(
            async move {
                let Some(memory) = self.services.memory.persistence.memory.clone() else {
                    return Ok("Memory not configured.".to_owned());
                };
                let ids: Vec<MessageId> = ids_str
                    .split_whitespace()
                    .filter_map(|s| s.parse::<i64>().ok().map(MessageId))
                    .collect();
                if ids.is_empty() {
                    return Ok(
                        "Usage: /memory promote <id> [id...]\nExample: /memory promote 42 43 44"
                            .to_owned(),
                    );
                }
                match memory.sqlite().manual_promote(&ids).await {
                    Ok(count) => Ok(format!("Promoted {count} message(s) to semantic tier.")),
                    Err(e) => Ok(format!("Promotion failed: {e}")),
                }
            }
            .instrument(tracing::info_span!("core.agent_access.memory_promote")),
        )
    }

    // ----- /graph -----

    fn graph_stats<'a>(
        &'a mut self,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(
            async move {
                let (_, store) = match self.resolve_graph_store() {
                    Ok(pair) => pair,
                    Err(msg) => return Ok(msg),
                };

                let stats_future = async {
                    tokio::join!(
                        store.entity_count(),
                        store.active_edge_count(),
                        store.community_count(),
                        store.edge_type_distribution()
                    )
                };
                let Ok((entities, edges, communities, distribution)) =
                    tokio::time::timeout(Duration::from_secs(5), stats_future).await
                else {
                    tracing::warn!("graph store call timed out after 5s (Qdrant unreachable)");
                    return Ok("Graph store unavailable (Qdrant unreachable).".to_owned());
                };
                let mut msg = format!(
                    "Graph memory: {} entities, {} edges, {} communities",
                    entities.unwrap_or(0),
                    edges.unwrap_or(0),
                    communities.unwrap_or(0)
                );
                if let Ok(dist) = distribution
                    && !dist.is_empty()
                {
                    let dist_str: Vec<String> =
                        dist.iter().map(|(t, c)| format!("{t}={c}")).collect();
                    write!(msg, "\nEdge types: {}", dist_str.join(", ")).unwrap_or(());
                }
                Ok(msg)
            }
            .instrument(tracing::info_span!("core.agent_access.graph_stats")),
        )
    }

    fn graph_entities<'a>(
        &'a mut self,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(
            async move {
                let (_, store) = match self.resolve_graph_store() {
                    Ok(pair) => pair,
                    Err(msg) => return Ok(msg),
                };

                let entities = match tokio::time::timeout(
                    Duration::from_secs(5),
                    store.all_entities(),
                )
                .await
                {
                    Ok(Ok(v)) => v,
                    Ok(Err(e)) => return Err(CommandError::new(e.to_string())),
                    Err(_) => {
                        tracing::warn!("graph store call timed out after 5s (Qdrant unreachable)");
                        return Ok("Graph store unavailable (Qdrant unreachable).".to_owned());
                    }
                };
                if entities.is_empty() {
                    return Ok("No entities found.".to_owned());
                }

                let total = entities.len();
                let display: Vec<String> = entities
                    .iter()
                    .take(50)
                    .map(|e| {
                        format!(
                            "  {:<40}  {:<15}  {}",
                            e.name,
                            e.entity_type.as_str(),
                            e.last_seen_at.split('T').next().unwrap_or(&e.last_seen_at)
                        )
                    })
                    .collect();
                let mut msg = format!(
                    "Entities ({total} total):\n  {:<40}  {:<15}  {}\n{}",
                    "NAME",
                    "TYPE",
                    "LAST SEEN",
                    display.join("\n")
                );
                if total > 50 {
                    write!(msg, "\n  ...and {} more", total - 50).unwrap_or(());
                }
                Ok(msg)
            }
            .instrument(tracing::info_span!("core.agent_access.graph_entities")),
        )
    }

    fn graph_facts<'a>(
        &'a mut self,
        name: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(
            async move {
                let (_, store) = match self.resolve_graph_store() {
                    Ok(pair) => pair,
                    Err(msg) => return Ok(msg),
                };

                let entity = match resolve_entity_by_name(&store, name).await? {
                    EntityLookup::Found(e) => e,
                    EntityLookup::Message(msg) => return Ok(msg),
                };

                let edges =
                    match with_graph_store_timeout(store.edges_for_entity(entity.id.0)).await? {
                        StoreCallOutcome::Completed(v) => v,
                        StoreCallOutcome::TimedOut => {
                            return Ok("Graph store unavailable (Qdrant unreachable).".to_owned());
                        }
                    };
                if edges.is_empty() {
                    return Ok(format!("Entity '{}' has no known facts.", entity.name));
                }

                let entity_names = build_entity_name_map(&store, &entity, &edges).await;

                let lines: Vec<String> = edges
                    .iter()
                    .map(|e| {
                        let src = entity_names
                            .get(&e.source_entity_id)
                            .cloned()
                            .unwrap_or_else(|| format!("#{}", e.source_entity_id));
                        let tgt = entity_names
                            .get(&e.target_entity_id)
                            .cloned()
                            .unwrap_or_else(|| format!("#{}", e.target_entity_id));
                        format!(
                            "  {} --[{}/{}]--> {}: {} (confidence: {:.2})",
                            src, e.relation, e.edge_type, tgt, e.fact, e.confidence
                        )
                    })
                    .collect();
                Ok(format!(
                    "Facts for '{}':\n{}",
                    entity.name,
                    lines.join("\n")
                ))
            }
            .instrument(tracing::info_span!("core.agent_access.graph_facts")),
        )
    }

    fn graph_history<'a>(
        &'a mut self,
        name: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(
            async move {
                let (_, store) = match self.resolve_graph_store() {
                    Ok(pair) => pair,
                    Err(msg) => return Ok(msg),
                };

                let entity = match resolve_entity_by_name(&store, name).await? {
                    EntityLookup::Found(e) => e,
                    EntityLookup::Message(msg) => return Ok(msg),
                };

                let edges =
                    match with_graph_store_timeout(store.edge_history_for_entity(entity.id.0, 50))
                        .await?
                    {
                        StoreCallOutcome::Completed(v) => v,
                        StoreCallOutcome::TimedOut => {
                            return Ok("Graph store unavailable (Qdrant unreachable).".to_owned());
                        }
                    };
                if edges.is_empty() {
                    return Ok(format!("Entity '{}' has no edge history.", entity.name));
                }

                let entity_names = build_entity_name_map(&store, &entity, &edges).await;

                let n = edges.len();
                let lines: Vec<String> = edges
                    .iter()
                    .map(|e| {
                        let status = if e.valid_to.is_some() {
                            let date = e
                                .valid_to
                                .as_deref()
                                .and_then(|s| s.split('T').next().or_else(|| s.split(' ').next()))
                                .unwrap_or("?");
                            format!("[expired {date}]")
                        } else {
                            "[active]".to_string()
                        };
                        let src = entity_names
                            .get(&e.source_entity_id)
                            .cloned()
                            .unwrap_or_else(|| format!("#{}", e.source_entity_id));
                        let tgt = entity_names
                            .get(&e.target_entity_id)
                            .cloned()
                            .unwrap_or_else(|| format!("#{}", e.target_entity_id));
                        format!(
                            "  {status} {} --[{}/{}]--> {}: {} (confidence: {:.2})",
                            src, e.relation, e.edge_type, tgt, e.fact, e.confidence
                        )
                    })
                    .collect();
                Ok(format!(
                    "Edge history for '{}' ({n} edges):\n{}",
                    entity.name,
                    lines.join("\n")
                ))
            }
            .instrument(tracing::info_span!("core.agent_access.graph_history")),
        )
    }

    fn graph_communities<'a>(
        &'a mut self,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(
            async move {
                let (_, store) = match self.resolve_graph_store() {
                    Ok(pair) => pair,
                    Err(msg) => return Ok(msg),
                };

                let communities =
                    match tokio::time::timeout(Duration::from_secs(5), store.all_communities())
                        .await
                    {
                        Ok(Ok(v)) => v,
                        Ok(Err(e)) => return Err(CommandError::new(e.to_string())),
                        Err(_) => {
                            tracing::warn!(
                                "graph store call timed out after 5s (Qdrant unreachable)"
                            );
                            return Ok("Graph store unavailable (Qdrant unreachable).".to_owned());
                        }
                    };
                if communities.is_empty() {
                    return Ok("No communities detected yet. Run graph backfill first.".to_owned());
                }

                let lines: Vec<String> = communities
                    .iter()
                    .map(|c| format!("  [{}]: {}", c.name, c.summary))
                    .collect();
                Ok(format!(
                    "Communities ({}):\n{}",
                    communities.len(),
                    lines.join("\n")
                ))
            }
            .instrument(tracing::info_span!("core.agent_access.graph_communities")),
        )
    }

    #[allow(clippy::too_many_lines)]
    fn graph_backfill<'a>(
        &'a mut self,
        limit: Option<usize>,
        progress_cb: &'a mut (dyn FnMut(String) + Send),
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        let store = match self.resolve_graph_store() {
            Ok((_, s)) => s,
            Err(msg) => return Box::pin(async move { Ok(msg) }),
        };
        let graph_cfg = self.services.memory.extraction.graph_config.clone();
        let embed_timeout_secs = self
            .services
            .memory
            .persistence
            .memory
            .as_ref()
            .map_or(5, |m| m.embed_timeout().as_secs());
        let provider = if graph_cfg.extract_provider.as_str().is_empty() {
            self.provider.clone()
        } else {
            self.resolve_background_provider(graph_cfg.extract_provider.as_str())
        };
        Box::pin(
            async move {
                let total = store.unprocessed_message_count().await.unwrap_or(0);
                let cap = limit.unwrap_or(usize::MAX);

                progress_cb(format!(
                    "Starting graph backfill... ({total} unprocessed messages)"
                ));

                let batch_size = 50usize;
                let mut processed = 0usize;
                let mut total_entities = 0usize;
                let mut total_edges = 0usize;

                loop {
                    let remaining_cap = cap.saturating_sub(processed);
                    if remaining_cap == 0 {
                        break;
                    }
                    let batch_limit = batch_size.min(remaining_cap);
                    let messages = store
                        .unprocessed_messages_for_backfill(batch_limit)
                        .await
                        .map_err(|e| CommandError::new(e.to_string()))?;
                    if messages.is_empty() {
                        break;
                    }

                    let ids: Vec<zeph_memory::types::MessageId> =
                        messages.iter().map(|(id, _)| *id).collect();

                    // extraction_cfg is loop-invariant (derived only from graph_cfg /
                    // embed_timeout_secs, never from message content), so it is built once per
                    // batch and cloned per message below.
                    let extraction_cfg = GraphExtractionConfig {
                        max_entities: graph_cfg.max_entities_per_message,
                        max_edges: graph_cfg.max_edges_per_message,
                        extraction_timeout_secs: graph_cfg.extraction_timeout_secs,
                        community_refresh_interval: 0,
                        expired_edge_retention_days: graph_cfg.expired_edge_retention_days,
                        max_entities_cap: graph_cfg.max_entities,
                        community_summary_max_prompt_bytes: graph_cfg
                            .community_summary_max_prompt_bytes,
                        community_summary_concurrency: graph_cfg.community_summary_concurrency,
                        lpa_edge_chunk_size: graph_cfg.lpa_edge_chunk_size,
                        note_linking: zeph_memory::NoteLinkingConfig::default(),
                        link_weight_decay_lambda: graph_cfg.link_weight_decay_lambda,
                        link_weight_decay_interval_secs: graph_cfg.link_weight_decay_interval_secs,
                        belief_revision_enabled: graph_cfg.belief_revision.enabled,
                        belief_revision_similarity_threshold: graph_cfg
                            .belief_revision
                            .similarity_threshold,
                        conversation_id: None,
                        apex_mem_enabled: graph_cfg.apex_mem.enabled,
                        llm_timeout_secs: graph_cfg.llm_timeout_secs,
                        embed_timeout_secs,
                        turn_index: None,
                        write_gate_min_relevance: graph_cfg
                            .write_gate
                            .enabled
                            .then_some(graph_cfg.write_gate.min_edge_relevance),
                        benna_fast_rate: graph_cfg.spreading_activation.benna_fast_rate,
                        benna_slow_rate: graph_cfg.spreading_activation.benna_slow_rate,
                        provenance: None,
                        system_prompt: None,
                        recall_include_imported: graph_cfg.recall_include_imported,
                    };

                    // Extract concurrently, bounded to 4 in-flight — matches
                    // semantic_scan_plugin_add's existing batched-LLM-call bound. Safe because
                    // `extract_and_store` builds a fresh `EntityResolver` per call (its
                    // `lock_name` guard does not span calls), so the actual concurrency-safety
                    // mechanism is the DB-level `UNIQUE(canonical_name, entity_type)` constraint
                    // and `ON CONFLICT ... DO UPDATE ... RETURNING id` upsert in
                    // `GraphStore::upsert_entity` (plus `add_alias`'s `INSERT OR IGNORE`), which
                    // makes concurrent entity creation for the same name idempotent regardless of
                    // in-process locking.
                    {
                        use futures::stream::StreamExt as _;

                        let extraction_futs: Vec<_> = messages
                            .iter()
                            .filter_map(|(_id, content)| {
                                if content.trim().is_empty() {
                                    return None;
                                }
                                let content = content.clone();
                                let provider = provider.clone();
                                let pool = store.pool().clone();
                                let extraction_cfg = extraction_cfg.clone();
                                Some(extract_and_store(
                                    content,
                                    vec![],
                                    provider,
                                    pool,
                                    extraction_cfg,
                                    None,
                                    None,
                                ))
                            })
                            .collect();

                        let results: Vec<_> = futures::stream::iter(extraction_futs)
                            .buffer_unordered(4)
                            .collect()
                            .await;

                        for result in results {
                            match result {
                                Ok(result) => {
                                    total_entities += result.stats.entities_upserted;
                                    total_edges += result.stats.edges_inserted;
                                }
                                Err(e) => {
                                    tracing::warn!("backfill extraction error: {e:#}");
                                }
                            }
                        }
                    }

                    store
                        .mark_messages_graph_processed(&ids)
                        .await
                        .map_err(|e| CommandError::new(e.to_string()))?;
                    processed += messages.len();

                    progress_cb(format!(
                        "Backfill progress: {processed} messages processed, \
                     {total_entities} entities, {total_edges} edges"
                    ));
                }

                Ok(format!(
                    "Backfill complete: {total_entities} entities, {total_edges} edges \
                 extracted from {processed} messages"
                ))
            }
            .instrument(tracing::info_span!("core.agent_access.graph_backfill")),
        )
    }

    // ----- /knowledge -----

    fn knowledge_status<'a>(
        &'a mut self,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(
            async move {
                use zeph_memory::graph::ingest::IngestLedger;

                let Some(memory) = self.services.memory.persistence.memory.clone() else {
                    return Ok("Memory subsystem not available.".to_owned());
                };
                let pool = memory.sqlite().pool().clone();
                let ledger = IngestLedger::new(pool);

                let rows = match ledger.summary().await {
                    Ok(r) => r,
                    Err(e) => return Err(CommandError(e.to_string())),
                };

                if rows.is_empty() {
                    return Ok("No knowledge has been ingested yet. \
                         Run `zeph knowledge ingest --source <src>`."
                        .to_owned());
                }

                let mut out = format!("Knowledge ingest ledger ({} entries):\n\n", rows.len());
                let mut current_batch = String::new();
                for row in &rows {
                    let batch_short = &row.import_batch_id[..row.import_batch_id.len().min(8)];
                    if current_batch != row.import_batch_id {
                        if !current_batch.is_empty() {
                            out.push('\n');
                        }
                        current_batch.clone_from(&row.import_batch_id);
                    }
                    let uri_display = &row.source_uri[..row.source_uri.floor_char_boundary(40)];
                    let at_display = &row.ingested_at[..row.ingested_at.len().min(19)];
                    let _ = writeln!(
                        out,
                        "  {uri_display:<40} batch={batch_short} at={at_display} \
                         e={} edges={}",
                        row.entities, row.edges,
                    );
                }
                Ok(out.trim_end().to_owned())
            }
            .instrument(tracing::info_span!("core.agent_access.knowledge_status")),
        )
    }

    fn knowledge_rollback<'a>(
        &'a mut self,
        batch_id: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(
            async move {
                use zeph_memory::graph::ingest::IngestLedger;

                let Some(memory) = self.services.memory.persistence.memory.clone() else {
                    return Ok("Memory subsystem not available.".to_owned());
                };
                let pool = memory.sqlite().pool().clone();
                let ledger = IngestLedger::new(pool.clone());

                match ledger.batch_exists(batch_id).await {
                    Ok(false) => {
                        return Ok(format!("Batch '{batch_id}' not found in ledger."));
                    }
                    Err(e) => return Err(CommandError(e.to_string())),
                    Ok(true) => {}
                }

                let Some(graph_store) = memory.graph_store.clone() else {
                    return Ok(
                        "Graph store unavailable (Qdrant unreachable or graph not enabled)."
                            .to_owned(),
                    );
                };

                let mut tx = zeph_db::begin_write(&pool)
                    .await
                    .map_err(|e| CommandError(e.to_string()))?;

                let (edges, entities) = graph_store
                    .delete_batch_in_tx(batch_id, &mut tx)
                    .await
                    .map_err(|e| CommandError(e.to_string()))?;
                ledger
                    .delete_batch_in_tx(batch_id, &mut tx)
                    .await
                    .map_err(|e| CommandError(e.to_string()))?;

                tx.commit().await.map_err(|e| CommandError(e.to_string()))?;

                let mut msg = format!(
                    "Rolled back batch '{batch_id}': removed {edges} edge(s) and \
                     {entities} entity(ies)."
                );
                if edges == 0 && entities == 0 {
                    msg.push_str(
                        "\nNote: no graph rows found. Phase-1 ingest writes to Qdrant notes — \
                         Qdrant embeddings are NOT removed by this rollback.",
                    );
                }
                Ok(msg)
            }
            .instrument(tracing::info_span!("core.agent_access.knowledge_rollback")),
        )
    }

    // ----- /guidelines -----

    fn guidelines<'a>(
        &'a mut self,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(
            async move {
                const MAX_DISPLAY_CHARS: usize = 4096;

                let Some(memory) = &self.services.memory.persistence.memory else {
                    return Ok("No memory backend initialised.".to_owned());
                };

                let cid = self.services.memory.persistence.conversation_id;
                let sqlite = memory.sqlite();

                let (version, text) = sqlite
                    .load_compression_guidelines(cid)
                    .await
                    .map_err(|e: zeph_memory::MemoryError| CommandError::new(e.to_string()))?;

                if version == 0 || text.is_empty() {
                    return Ok("No compression guidelines generated yet.".to_owned());
                }

                let (_, created_at) = sqlite
                    .load_compression_guidelines_meta(cid)
                    .await
                    .unwrap_or((0, String::new()));

                let (body, truncated) = if text.len() > MAX_DISPLAY_CHARS {
                    let end = text.floor_char_boundary(MAX_DISPLAY_CHARS);
                    (&text[..end], true)
                } else {
                    (text.as_str(), false)
                };

                let mut output =
                    format!("Compression Guidelines (v{version}, updated {created_at}):\n\n{body}");
                if truncated {
                    output.push_str("\n\n[truncated]");
                }
                Ok(output)
            }
            .instrument(tracing::info_span!("core.agent_access.guidelines")),
        )
    }

    // ----- /caveman -----

    fn handle_caveman<'a>(
        &'a mut self,
        arg: &'a str,
    ) -> Pin<Box<dyn Future<Output = String> + Send + 'a>> {
        Box::pin(async move {
            let active = &mut self.services.session.caveman_active;
            match arg.trim() {
                "on" | "enable" => {
                    *active = true;
                    "caveman: on".to_owned()
                }
                "off" | "disable" => {
                    *active = false;
                    "caveman: off".to_owned()
                }
                "status" => {
                    if *active {
                        "caveman: on".to_owned()
                    } else {
                        "caveman: off".to_owned()
                    }
                }
                _ => {
                    *active = !*active;
                    if *active {
                        "caveman: on".to_owned()
                    } else {
                        "caveman: off".to_owned()
                    }
                }
            }
        })
    }

    // ----- /model, /provider -----

    fn handle_model<'a>(
        &'a mut self,
        arg: &'a str,
    ) -> Pin<Box<dyn Future<Output = String> + Send + 'a>> {
        Box::pin(async move {
            let input = if arg.is_empty() {
                "/model".to_owned()
            } else {
                format!("/model {arg}")
            };
            self.handle_model_command_as_string(&input).await
        })
    }

    fn handle_provider<'a>(
        &'a mut self,
        arg: &'a str,
    ) -> Pin<Box<dyn Future<Output = String> + Send + 'a>> {
        Box::pin(async move { self.handle_provider_command_as_string(arg).await })
    }

    // ----- /think-tokens, /reasoning-effort -----

    fn handle_think_tokens<'a>(
        &'a mut self,
        arg: &'a str,
    ) -> Pin<Box<dyn Future<Output = String> + Send + 'a>> {
        Box::pin(async move {
            let arg = arg.trim();
            let provider_name = self.provider.name().to_owned();
            if arg.is_empty() {
                return match self.provider.current_thinking_budget() {
                    Some(n) => format!("think-tokens: {n} (provider: {provider_name})"),
                    None => format!("think-tokens: off (provider: {provider_name})"),
                };
            }

            let budget = match zeph_commands::handlers::think_tokens::parse_token_budget(arg) {
                Ok(b) => b,
                Err(e) => return format!("think-tokens: {e}"),
            };

            // Captured before the mutation so the cross-override note (Claude's Extended and
            // Adaptive thinking share one config field) only fires when this call actually
            // cleared a previously active reasoning-effort level.
            let had_reasoning_effort = self.provider.current_reasoning_effort().is_some();
            match self.provider.set_thinking_budget(budget) {
                Ok(()) => {
                    let mut msg = match budget {
                        Some(n) => format!("think-tokens: set to {n} (provider: {provider_name})"),
                        None => format!("think-tokens: disabled (provider: {provider_name})"),
                    };
                    if had_reasoning_effort && self.provider.current_reasoning_effort().is_none() {
                        msg.push_str(
                            " Note: this overrides the previously set reasoning-effort level \
                             — Claude's Extended and Adaptive thinking share one config field.",
                        );
                    }
                    if let Some(advisory) = self.provider.capability_delegation_advisory() {
                        let _ = write!(msg, " Note: {advisory}.");
                    }
                    msg
                }
                Err(zeph_llm::LlmError::ModelCapabilityMismatch { provider, message }) => {
                    format!("provider `{provider}` {message}")
                }
                Err(e) => format!("think-tokens: {e}"),
            }
        })
    }

    fn handle_reasoning_effort<'a>(
        &'a mut self,
        arg: &'a str,
    ) -> Pin<Box<dyn Future<Output = String> + Send + 'a>> {
        Box::pin(async move {
            let arg = arg.trim();
            let provider_name = self.provider.name().to_owned();
            if arg.is_empty() {
                return match self.provider.current_reasoning_effort() {
                    Some(e) => format!("reasoning-effort: {e} (provider: {provider_name})"),
                    None => format!("reasoning-effort: off (provider: {provider_name})"),
                };
            }

            let effort: zeph_llm::any::ReasoningEffort = match arg.parse() {
                Ok(e) => e,
                Err(e) => return format!("reasoning-effort: {e}"),
            };

            // Captured before the mutation — see the matching comment in handle_think_tokens.
            let had_thinking_budget = self.provider.current_thinking_budget().is_some();
            match self.provider.apply_reasoning_effort(effort) {
                Ok(()) => {
                    let mut msg = format!(
                        "reasoning-effort: set to {} (provider: {provider_name})",
                        effort.as_str()
                    );
                    if had_thinking_budget && self.provider.current_thinking_budget().is_none() {
                        msg.push_str(
                            " Note: this overrides the previously set thinking-token budget \
                             — Claude's Extended and Adaptive thinking share one config field.",
                        );
                    }
                    if let Some(advisory) = self.provider.capability_delegation_advisory() {
                        let _ = write!(msg, " Note: {advisory}.");
                    }
                    msg
                }
                Err(zeph_llm::LlmError::ModelCapabilityMismatch { provider, message }) => {
                    format!("provider `{provider}` {message}")
                }
                Err(e) => format!("reasoning-effort: {e}"),
            }
        })
    }

    // ----- /policy -----

    fn handle_policy<'a>(
        &'a mut self,
        args: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(async move { Ok(self.handle_policy_command_as_string(args)) })
    }

    // ----- /scheduler -----

    #[cfg(feature = "scheduler")]
    fn list_scheduled_tasks<'a>(
        &'a mut self,
    ) -> Pin<Box<dyn Future<Output = Result<Option<String>, CommandError>> + Send + 'a>> {
        Box::pin(async move {
            let result = self
                .handle_scheduler_list_as_string()
                .await
                .map_err(|e| CommandError::new(e.to_string()))?;
            Ok(Some(result))
        })
    }

    #[cfg(not(feature = "scheduler"))]
    fn list_scheduled_tasks<'a>(
        &'a mut self,
    ) -> Pin<Box<dyn Future<Output = Result<Option<String>, CommandError>> + Send + 'a>> {
        Box::pin(async move { Ok(None) })
    }

    // ----- /lsp -----

    fn lsp_status<'a>(
        &'a mut self,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(async move {
            self.handle_lsp_status_as_string()
                .await
                .map_err(|e| CommandError::new(e.to_string()))
        })
    }

    // ----- /recap -----

    fn session_recap<'a>(
        &'a mut self,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(
            async move {
                match self.build_recap().await {
                    Ok(text) => Ok(text),
                    Err(e) => {
                        // /recap is an explicit user command — surface a fixed message so that
                        // LlmError internals (URLs with embedded credentials, response excerpts)
                        // are never forwarded to the user channel. Full detail goes to the log.
                        tracing::warn!("session recap command: {}", e.0);
                        Ok("Recap unavailable — see logs for details".to_string())
                    }
                }
            }
            .instrument(tracing::info_span!("core.agent_access.session_recap")),
        )
    }

    // ----- /compact -----

    fn compact_context<'a>(
        &'a mut self,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(
            self.compact_context_command()
                .instrument(tracing::info_span!("core.agent_access.compact_context")),
        )
    }

    // ----- /new -----

    fn reset_conversation<'a>(
        &'a mut self,
        keep_plan: bool,
        no_digest: bool,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(async move {
            match self.reset_conversation(keep_plan, no_digest).await {
                Ok((old_id, new_id)) => {
                    let old = old_id.map_or_else(|| "none".to_string(), |id| id.0.to_string());
                    let new = new_id.map_or_else(|| "none".to_string(), |id| id.0.to_string());
                    let keep_note = if keep_plan { " (plan preserved)" } else { "" };
                    Ok(format!(
                        "New conversation started. Previous: {old} → Current: {new}{keep_note}"
                    ))
                }
                Err(e) => Ok(format!("Failed to start new conversation: {e}")),
            }
        })
    }

    // ----- /cache-stats -----

    fn cache_stats(&self) -> String {
        self.tool_orchestrator.cache_stats()
    }

    // ----- /status -----

    fn session_status<'a>(
        &'a mut self,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(async move { Ok(self.handle_status_as_string()) })
    }

    // ----- /guardrail -----

    fn guardrail_status(&self) -> String {
        self.format_guardrail_status()
    }

    // ----- /focus -----

    fn focus_status(&self) -> String {
        self.format_focus_status()
    }

    // ----- /sidequest -----

    fn sidequest_status(&self) -> String {
        self.format_sidequest_status()
    }

    // ----- /image -----

    fn load_image<'a>(
        &'a mut self,
        path: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        use zeph_common::path_guard::{PathRejection, classify_relative_path};
        use zeph_llm::provider::{ImageData, MessagePart};

        match classify_relative_path(path) {
            PathRejection::Allowed => {}
            PathRejection::Absolute => {
                return Box::pin(async move {
                    Ok(
                        "Invalid image path: absolute paths are not supported, use a path \
                        relative to the working directory"
                            .to_owned(),
                    )
                });
            }
            PathRejection::Traversal => {
                return Box::pin(async move {
                    Ok("Invalid image path: path traversal ('..') is not allowed".to_owned())
                });
            }
        }

        let path_owned = path.to_owned();
        Box::pin(async move {
            let path_for_task = path_owned.clone();
            let read_result = tokio::task::spawn_blocking(move || std::fs::read(&path_for_task))
                .await
                .map_err(|e| CommandError::new(format!("spawn_blocking join error: {e}")))?;
            let data = match read_result {
                Ok(d) => d,
                Err(e) => return Ok(format!("Cannot read image {path_owned}: {e}")),
            };
            if data.len() > crate::agent::message_queue::MAX_IMAGE_BYTES {
                return Ok(format!(
                    "Image {path_owned} exceeds size limit ({} MB), skipping",
                    crate::agent::message_queue::MAX_IMAGE_BYTES / 1024 / 1024
                ));
            }
            let mime_type =
                crate::agent::message_queue::detect_image_mime(Some(&path_owned)).to_string();
            self.msg
                .pending_image_parts
                .push(MessagePart::Image(Box::new(ImageData { data, mime_type })));
            Ok(format!("Image loaded: {path_owned}. Send your message."))
        })
    }

    // ----- /mcp -----

    fn handle_mcp<'a>(
        &'a mut self,
        args: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        // Extract all owned data before the async block so no &mut self reference is
        // held across an .await point, satisfying the `for<'a>` Send bound.
        let args_owned = args.to_owned();
        let parts: Vec<String> = args_owned.split_whitespace().map(str::to_owned).collect();
        let sub = parts.first().cloned().unwrap_or_default();

        match sub.as_str() {
            "list" => {
                // Read-only: clone all data before async.
                let manager = self.services.mcp.manager.clone();
                let tools_snapshot: Vec<(String, String)> = self
                    .services
                    .mcp
                    .tools
                    .iter()
                    .map(|t| (t.server_id.clone(), t.name.clone()))
                    .collect();
                Box::pin(async move {
                    use std::fmt::Write;
                    let Some(manager) = manager else {
                        return Ok("MCP is not enabled.".to_owned());
                    };
                    let server_ids = manager.list_servers().await;
                    if server_ids.is_empty() {
                        return Ok("No MCP servers connected.".to_owned());
                    }
                    let mut output = String::from("Connected MCP servers:\n");
                    let mut total = 0usize;
                    for id in &server_ids {
                        let count = tools_snapshot.iter().filter(|(sid, _)| sid == id).count();
                        total += count;
                        let _ = writeln!(output, "- {id} ({count} tools)");
                    }
                    let _ = write!(output, "Total: {total} tool(s)");
                    Ok(output)
                })
            }
            "tools" => {
                // Read-only: collect tool info before async.
                let server_id = parts.get(1).cloned();
                let owned_tools: Vec<(String, String)> = if let Some(ref sid) = server_id {
                    self.services
                        .mcp
                        .tools
                        .iter()
                        .filter(|t| &t.server_id == sid)
                        .map(|t| (t.name.clone(), t.description.clone()))
                        .collect()
                } else {
                    Vec::new()
                };
                Box::pin(async move {
                    use std::fmt::Write;
                    let Some(server_id) = server_id else {
                        return Ok("Usage: /mcp tools <server_id>".to_owned());
                    };
                    if owned_tools.is_empty() {
                        return Ok(format!("No tools found for server '{server_id}'."));
                    }
                    let mut output =
                        format!("Tools for '{server_id}' ({} total):\n", owned_tools.len());
                    for (name, desc) in &owned_tools {
                        if desc.is_empty() {
                            let _ = writeln!(output, "- {name}");
                        } else {
                            let _ = writeln!(output, "- {name} — {desc}");
                        }
                    }
                    Ok(output)
                })
            }
            // add/remove require mutating self after async I/O.
            // handle_mcp_command is structured so the only .await crossing a &mut self
            // boundary goes through a cloned Arc<McpManager> — no &self fields are held
            // across that .await.  The subsequent state-change methods (rebuild_semantic_index,
            // sync_mcp_registry) are also async fn(&mut self), but they only hold owned locals
            // across their own .await points (cloned tools Vec, cloned Arcs).
            _ => Box::pin(async move {
                self.handle_mcp_command(&args_owned)
                    .await
                    .map_err(|e| CommandError::new(e.to_string()))
            }),
        }
    }

    // ----- /skill -----

    fn handle_skill<'a>(
        &'a mut self,
        args: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        let args_owned = args.to_owned();
        Box::pin(async move {
            self.handle_skill_command_as_string(&args_owned)
                .await
                .map_err(|e| CommandError::new(e.to_string()))
        })
    }

    // ----- /skills -----

    fn handle_skills<'a>(
        &'a mut self,
        args: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        let args_owned = args.to_owned();
        Box::pin(async move {
            self.handle_skills_as_string(&args_owned)
                .await
                .map_err(|e| CommandError::new(e.to_string()))
        })
    }

    // ----- /feedback -----

    fn handle_feedback_command<'a>(
        &'a mut self,
        args: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        let args_owned = args.to_owned();
        Box::pin(async move {
            self.handle_feedback_as_string(&args_owned)
                .await
                .map_err(|e| CommandError::new(e.to_string()))
        })
    }

    // ----- /plan -----

    #[cfg(feature = "scheduler")]
    fn handle_plan<'a>(
        &'a mut self,
        input: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(async move {
            self.dispatch_plan_command_as_string(input)
                .await
                .map_err(|e| CommandError::new(e.to_string()))
        })
    }

    #[cfg(not(feature = "scheduler"))]
    fn handle_plan<'a>(
        &'a mut self,
        _input: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(async move { Ok(String::new()) })
    }

    // ----- /experiment -----

    fn handle_experiment<'a>(
        &'a mut self,
        input: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(async move {
            self.handle_experiment_command_as_string(input)
                .await
                .map_err(|e| CommandError::new(e.to_string()))
        })
    }

    // ----- /agent, @mention -----

    fn handle_agent_dispatch<'a>(
        &'a mut self,
        input: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<Option<String>, CommandError>> + Send + 'a>> {
        Box::pin(async move {
            match self.dispatch_agent_command(input).await {
                Some(Err(e)) => Err(CommandError::new(e.to_string())),
                Some(Ok(())) | None => Ok(None),
            }
        })
    }

    // ----- /plugins -----

    fn handle_plugins<'a>(
        &'a mut self,
        args: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        let args_owned = args.to_owned();
        // Clone the fields needed by PluginManager before entering the async block.
        // spawn_blocking requires 'static, so we cannot borrow &self inside the closure.
        let managed_dir = self.services.skill.managed_dir.clone();
        let mcp_allowed = self.services.mcp.allowed_commands.clone();
        let base_shell_allowed = self.runtime.lifecycle.startup_shell_overlay.allowed.clone();
        // Collect manifest paths for ephemeral plugins. Reading the actual files is
        // deferred into the async block below to avoid blocking the tokio worker thread.
        let ephemeral_manifest_paths: Vec<std::path::PathBuf> = self
            .runtime
            .ephemeral_plugins
            .iter()
            .map(|tmp| tmp.path().join("plugin.toml"))
            .collect();

        // Resolve scanner once, before the async block captures `self`.
        // Fail-closed: if semantic_scan is enabled but no provider is configured, refuse
        // to proceed rather than silently falling back to the primary provider (#4706, #4709).
        let semantic_scan_enabled = self.services.skill.semantic_scan;
        let maybe_scanner: Option<zeph_skills::semantic_scanner::SkillSemanticScanner> =
            if semantic_scan_enabled {
                let provider_name = self.services.skill.semantic_scan_provider.as_str();
                if provider_name.trim().is_empty() {
                    return Box::pin(async move {
                        Err(CommandError::new(
                            "semantic_scan is enabled but semantic_scan_provider is not set; \
                             refusing plugin add to maintain fail-closed security posture",
                        ))
                    });
                }
                let provider_known = self
                    .runtime
                    .providers
                    .provider_pool
                    .iter()
                    .any(|e| e.effective_name().eq_ignore_ascii_case(provider_name));
                if !provider_known {
                    let name = provider_name.to_owned();
                    return Box::pin(async move {
                        Err(CommandError::new(format!(
                            "semantic_scan is enabled but semantic_scan_provider '{name}' \
                             is not configured in [[llm.providers]]; \
                             refusing plugin add to maintain fail-closed security posture",
                        )))
                    });
                }
                let provider = self.resolve_background_provider(provider_name);
                Some(zeph_skills::semantic_scanner::SkillSemanticScanner::new(
                    provider,
                ))
            } else {
                None
            };

        Box::pin(async move {
            let (subcmd, source) = args_owned
                .trim()
                .split_once(' ')
                .unwrap_or((args_owned.trim(), ""));

            // Stage-2 LLM semantic scan runs before the blocking add(), fail-closed.
            if subcmd == "add"
                && !source.trim().is_empty()
                && let Some(ref scanner) = maybe_scanner
                && let Some(err) = semantic_scan_plugin_add(
                    scanner,
                    source.trim(),
                    managed_dir.clone(),
                    mcp_allowed.clone(),
                    base_shell_allowed.clone(),
                )
                .instrument(tracing::info_span!("core.agent.scan_plugin", plugin = %source.trim()))
                .await?
            {
                return Ok(err);
            }

            // Resolve ephemeral plugin names asynchronously before entering the blocking task.
            let ephemeral_names: Vec<String> = {
                use futures::future::join_all;
                let futs = ephemeral_manifest_paths.into_iter().map(|p| async move {
                    tokio::fs::read_to_string(&p)
                        .await
                        .ok()
                        .and_then(|s| toml::from_str::<zeph_plugins::PluginManifest>(&s).ok())
                        .map(|m| m.plugin.name.to_string())
                });
                join_all(futs).await.into_iter().flatten().collect()
            };

            // PluginManager performs synchronous filesystem I/O (copy, remove_dir_all,
            // read_dir). Run on a blocking thread to avoid stalling the tokio worker.
            tokio::task::spawn_blocking(move || {
                Self::run_plugin_command(
                    &args_owned,
                    managed_dir,
                    mcp_allowed,
                    base_shell_allowed,
                    ephemeral_names,
                )
            })
            .await
            .map_err(|e| CommandError(format!("plugin task panicked: {e}")))
        })
    }

    // ----- /acp -----

    fn handle_acp<'a>(
        &'a mut self,
        args: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(async move {
            self.handle_acp_as_string(args)
                .map_err(|e| CommandError::new(e.to_string()))
        })
    }

    // ----- /cocoon -----

    #[cfg(feature = "cocoon")]
    fn handle_cocoon<'a>(
        &'a mut self,
        args: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(async move {
            self.handle_cocoon_as_string(args)
                .await
                .map_err(|e| CommandError::new(e.to_string()))
        })
    }

    #[cfg(not(feature = "cocoon"))]
    fn handle_cocoon<'a>(
        &'a mut self,
        _args: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(async {
            Ok("Cocoon support is not compiled in. Rebuild with `--features cocoon`.".to_owned())
        })
    }

    // ----- /loop -----

    fn handle_loop<'a>(
        &'a mut self,
        args: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        use zeph_commands::handlers::loop_cmd::parse_loop_args;

        let args_owned = args.trim().to_owned();
        Box::pin(async move {
            if args_owned == "stop" {
                return Ok(self.stop_user_loop());
            }
            if args_owned == "status" {
                return Ok(match &self.runtime.lifecycle.user_loop {
                    Some(ls) => format!(
                        "Loop active: \"{}\" (iteration {}, interval every {}s).",
                        ls.prompt,
                        ls.iteration,
                        ls.interval.period().as_secs(),
                    ),
                    None => "No active loop.".to_owned(),
                });
            }
            let (prompt, interval_secs) = parse_loop_args(&args_owned)?;

            if prompt.starts_with('/') {
                return Err(CommandError::new(
                    "Loop prompt must not start with '/'. Slash commands cannot be used as loop prompts.",
                ));
            }

            let min_secs = self.runtime.config.loop_min_interval_secs;
            if interval_secs < min_secs {
                return Err(CommandError::new(format!(
                    "Minimum loop interval is {min_secs}s. Got {interval_secs}s."
                )));
            }
            if self.runtime.lifecycle.user_loop.is_some() {
                return Err(CommandError::new(
                    "A loop is already active. Use /loop stop first.",
                ));
            }

            self.start_user_loop(prompt.clone(), interval_secs);
            Ok(format!(
                "Loop started: \"{prompt}\" every {interval_secs}s. Use /loop stop to cancel."
            ))
        })
    }

    fn notify_test<'a>(
        &'a mut self,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        let notifier = self.runtime.lifecycle.notifier.clone();
        Box::pin(async move {
            let Some(notifier) = notifier else {
                return Ok(
                    "Notifications are disabled. Set `notifications.enabled = true` in config."
                        .to_owned(),
                );
            };
            match notifier.fire_test().await {
                Ok(()) => Ok("Test notification sent.".to_owned()),
                Err(e) => Err(CommandError::new(format!("notification test failed: {e}"))),
            }
        })
    }

    fn handle_trajectory(&mut self, args: &str) -> String {
        self.handle_trajectory_command_as_string(args)
    }

    fn handle_scope(&self, args: &str) -> String {
        self.handle_scope_command_as_string(args)
    }

    // ----- /goal -----

    fn handle_goal<'a>(
        &'a mut self,
        args: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        // Extract all non-Send data synchronously before entering the async block.
        if self.services.goal_accounting.is_none() {
            if !self.runtime.config.goals.enabled {
                return Box::pin(async {
                    Ok("Goals are disabled. Set `[goals] enabled = true` in config.".to_owned())
                });
            }
            let pool = match self.services.memory.persistence.memory.as_ref() {
                Some(m) => std::sync::Arc::new(m.sqlite().pool().clone()),
                None => {
                    return Box::pin(async {
                        Ok("Goals require a database backend (memory not configured).".to_owned())
                    });
                }
            };
            let store = std::sync::Arc::new(crate::goal::GoalStore::new(pool));
            let accounting = std::sync::Arc::new(crate::goal::GoalAccounting::new(store));
            self.services.goal_accounting = Some(accounting);
        }

        let accounting =
            self.services.goal_accounting.clone().expect(
                "invariant: goal_accounting is always Some at this point (initialized above)",
            );
        let max_chars = self.runtime.config.goals.max_text_chars;
        let default_budget = self.runtime.config.goals.default_token_budget;
        let autonomous_enabled = self.runtime.config.goals.autonomous_enabled;
        let autonomous_max_turns = self.runtime.config.goals.autonomous_max_turns;
        let args_owned = args.to_owned();

        // S1: `goal_create` may need to arm `AutonomousDriver` with a new session.
        // We capture a clone of the pending_start Arc that lives on the driver.
        // The async block fills it; the main agent loop (which has `&mut self`) drains it
        // via `AutonomousDriver::flush_pending_start()` after each command handler returns.
        let pending_start_arc = std::sync::Arc::clone(&self.services.autonomous.pending_start_arc);

        Box::pin(async move {
            let _ = accounting.refresh().await;
            let store = accounting.get_store();
            let args = args_owned.as_str();

            match args {
                "" | "status" => goal_status(&accounting).await,
                "pause" => goal_pause(&accounting, &store).await,
                "resume" => goal_resume(&accounting, &store).await,
                "complete" => goal_complete(&accounting, &store).await,
                "clear" => goal_clear(&accounting, &store).await,
                "list" => goal_list(&store).await,
                _ if args.starts_with("create") => {
                    let (msg, auto_req) = goal_create(
                        args,
                        &accounting,
                        &store,
                        max_chars,
                        default_budget,
                        autonomous_enabled,
                        autonomous_max_turns,
                    )
                    .await?;
                    if let Some(req) = auto_req {
                        *pending_start_arc.lock() = Some(req);
                    }
                    Ok(msg)
                }
                _ => Ok(
                    "Unknown /goal subcommand. Try: create, pause, resume, complete, clear, status, list."
                        .to_owned(),
                ),
            }
        })
    }

    fn active_goal_snapshot(&self) -> Option<zeph_commands::GoalSnapshot> {
        let accounting = self.services.goal_accounting.as_ref()?;
        let snap = accounting.snapshot()?;
        Some(zeph_commands::GoalSnapshot {
            id: snap.id,
            text: snap.text,
            status: match snap.status {
                crate::goal::GoalStatus::Active => zeph_commands::GoalStatusView::Active,
                crate::goal::GoalStatus::Paused => zeph_commands::GoalStatusView::Paused,
                crate::goal::GoalStatus::Completed => zeph_commands::GoalStatusView::Completed,
                crate::goal::GoalStatus::Cleared => zeph_commands::GoalStatusView::Cleared,
            },
            turns_used: snap.turns_used,
            tokens_used: snap.tokens_used,
            token_budget: snap.token_budget,
        })
    }

    // ----- /undo, /redo -----

    fn handle_undo<'a>(
        &'a mut self,
        args: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        let executor = std::sync::Arc::clone(&self.tool_executor);
        let args_owned = args.trim().to_owned();
        Box::pin(async move {
            if args_owned == "list" {
                let result = executor.checkpoint_list_erased();
                if !result.supported {
                    return Ok(
                        "Checkpoints are not enabled. Set `[tools.shell] checkpoints_enabled = true` in config.".to_owned()
                    );
                }
                if result.entries.is_empty() {
                    return Ok("Undo stack is empty.".to_owned());
                }
                let mut lines = vec![format!("Undo stack ({} entries):", result.entries.len())];
                for e in &result.entries {
                    lines.push(format!(
                        "  [{}] {} ({} file(s))",
                        e.index, e.command, e.file_count
                    ));
                }
                if result.redo_depth > 0 {
                    lines.push(format!("Redo depth: {}", result.redo_depth));
                }
                return Ok(lines.join("\n"));
            }

            let n: usize = if args_owned.is_empty() {
                1
            } else {
                match args_owned.parse::<usize>() {
                    Ok(v) if v > 0 => v,
                    _ => {
                        return Err(CommandError::new(format!(
                            "Invalid argument: expected a positive integer or 'list', got '{args_owned}'"
                        )));
                    }
                }
            };

            let result = tokio::task::spawn_blocking(move || executor.checkpoint_undo_erased(n))
                .await
                .map_err(|e| CommandError::new(format!("undo task panicked: {e}")))?;
            if !result.supported {
                return Ok(
                    "Checkpoints are not enabled. Set `[tools.shell] checkpoints_enabled = true` in config.".to_owned()
                );
            }
            Ok(result.message)
        })
    }

    fn handle_redo<'a>(
        &'a mut self,
        args: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        let _ = args;
        let executor = std::sync::Arc::clone(&self.tool_executor);
        Box::pin(async move {
            let result = tokio::task::spawn_blocking(move || executor.checkpoint_redo_erased())
                .await
                .map_err(|e| CommandError::new(format!("redo task panicked: {e}")))?;
            if !result.supported {
                return Ok(
                    "Checkpoints are not enabled. Set `[tools.shell] checkpoints_enabled = true` in config.".to_owned()
                );
            }
            Ok(result.message)
        })
    }

    // ----- /agents -----

    fn handle_agents<'a>(
        &'a mut self,
        args: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        use zeph_commands::handlers::agents_fleet::{FleetEntry, format_fleet_section};
        use zeph_subagent::AgentsCommand;

        let args_owned = args.trim().to_owned();
        Box::pin(async move {
            // Fleet view: bare `/agents` or `/agents fleet` shows autonomous sessions + definitions.
            let show_fleet = args_owned.is_empty() || args_owned == "fleet";

            let fleet_section = if show_fleet {
                let snapshots = self.services.autonomous_registry.list();
                let entries: Vec<FleetEntry> = snapshots
                    .into_iter()
                    .map(|s| FleetEntry {
                        goal_id: s.goal_id,
                        goal_text_short: s.goal_text_short,
                        state: s.state,
                        turns_executed: s.turns_executed,
                        max_turns: s.max_turns,
                        elapsed: s.elapsed,
                    })
                    .collect();
                format_fleet_section(&entries)
            } else {
                String::new()
            };

            // Sub-agent definitions section.
            let definitions_section = if show_fleet || args_owned == "list" {
                self.handle_agents_definitions_list()
            } else {
                // CRUD subcommands: show, create, edit, delete.
                match AgentsCommand::parse(&format!("/agents {args_owned}")) {
                    Ok(cmd) => self.handle_agents_crud(cmd),
                    Err(e) => e.to_string(),
                }
            };

            let mut out = fleet_section;
            if !definitions_section.is_empty() {
                if !out.is_empty() {
                    out.push('\n');
                }
                out.push_str(&definitions_section);
            }

            if out.is_empty() {
                "No active autonomous sessions or sub-agent definitions found."
                    .clone_into(&mut out);
            }

            Ok(out)
        })
    }

    // ----- /conv -----

    fn handle_conv<'a>(
        &'a mut self,
        args: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        let args_owned = args.trim().to_owned();
        Box::pin(async move {
            // `resume`/`fork` need `&mut self` (mid-session live conversation swap, D-9) —
            // handled first so `self` isn't already borrowed by the `list`/`show` path below.
            if let Some(id) = args_owned.strip_prefix("resume ") {
                return self.handle_conv_resume(id.trim()).await;
            }
            if let Some(id) = args_owned.strip_prefix("fork ") {
                return self.handle_conv_fork(id.trim()).await;
            }

            let Some(memory) = self.services.memory.persistence.memory.clone() else {
                return Ok(
                    "Conversation-session persistence requires memory to be enabled ([memory] enabled = true)."
                        .to_owned(),
                );
            };
            let store = zeph_session::SessionStore::new(memory.sqlite().pool().clone());

            if let Some(id) = args_owned.strip_prefix("show ") {
                return handle_conv_show(&store, id.trim()).await;
            }
            if args_owned.is_empty() || args_owned == "list" {
                return handle_conv_list(&store).await;
            }
            Ok(format!(
                "Unknown /conv subcommand '{args_owned}'. Usage: /conv [list | show <id> | resume <id> | fork <id>]"
            ))
        })
    }

    // ----- /worktree -----

    fn list_worktrees<'a>(
        &'a mut self,
    ) -> Pin<Box<dyn Future<Output = Result<Option<String>, CommandError>> + Send + 'a>> {
        Box::pin(async move {
            self.handle_worktree_list_as_string()
                .await
                .map_err(|e| CommandError::new(e.to_string()))
        })
    }

    fn clean_worktrees<'a>(
        &'a mut self,
        force: bool,
    ) -> Pin<Box<dyn Future<Output = Result<Option<String>, CommandError>> + Send + 'a>> {
        Box::pin(async move {
            self.handle_worktree_clean_as_string(force)
                .await
                .map_err(|e| CommandError::new(e.to_string()))
        })
    }

    // ----- /cd -----

    fn change_working_directory<'a>(
        &'a mut self,
        path: &'a str,
    ) -> Pin<Box<dyn Future<Output = Result<String, CommandError>> + Send + 'a>> {
        Box::pin(
            async move {
                let path = path.trim();
                if path.is_empty() {
                    let cwd = std::env::current_dir().map_err(|e| {
                        CommandError::new(format!("failed to read current working directory: {e}"))
                    })?;
                    return Ok(format!("Current working directory: {}", cwd.display()));
                }
                // Reuses the same path-resolution + `set_current_dir` logic as the
                // LLM-invoked `set_working_directory` tool (#6032 FR-001/FR-011) — no
                // parallel implementation. `allowed_paths` empty means no build site has set
                // it yet; default to `[cwd]` rather than "allow every path", matching
                // `FileExecutor::new`/`SetCwdExecutor::new`'s convention (SEC-2).
                let allowed_paths: Vec<std::path::PathBuf> =
                    if self.services.tool_state.allowed_paths.is_empty() {
                        // Canonicalize for byte-for-byte parity with `FileExecutor::new`/
                        // `SetCwdExecutor::new`'s fallback, which both canonicalize their
                        // default `[cwd]` entry (`.map(|p| p.canonicalize().unwrap_or(p))`).
                        std::env::current_dir()
                            .map(|p| p.canonicalize().unwrap_or(p))
                            .into_iter()
                            .collect()
                    } else {
                        self.services.tool_state.allowed_paths.clone()
                    };
                let new_cwd = zeph_tools::resolve_and_set_cwd(path, &allowed_paths)
                    .map_err(|e| CommandError::new(format!("cannot change to '{path}': {e}")))?;
                // Drives the same post-change pipeline the tool-invoked path gets for free
                // after a tool batch (`tier_loop.rs`) — a bare slash command must call it
                // explicitly (FR-002/FR-003/FR-004): mirror-update, `cwd_changed` hooks,
                // repo-map invalidation, and (unless safe-mode) instruction re-discovery.
                self.check_cwd_changed().await;
                Ok(format!(
                    "Working directory changed to: {}",
                    new_cwd.display()
                ))
            }
            .instrument(tracing::info_span!("core.commands.cd")),
        )
    }
}

impl<C: Channel> Agent<C> {
    /// `/conv resume <id>` (spec-068, #5343, D-9): mid-session live swap onto an existing
    /// durable session. Resolves `conversation_id` via the `SessionId`<->`ConversationId`
    /// bijection (spec §5.2) — reuses the session's existing linked conversation if one exists,
    /// otherwise mints one and links it (a session created via the HTTP API's `POST /sessions`,
    /// or a legacy session, may not have one yet).
    async fn handle_conv_resume(&mut self, id: &str) -> Result<String, CommandError> {
        if id.is_empty() {
            return Ok("Usage: /conv resume <id>".to_owned());
        }
        // #5487 fix 3: `load_and_resume_conversation` now opens the target session's event log
        // exclusively (INV-D2). Resuming into the session already live in this agent would try
        // to acquire a second exclusive lock on the same directory this agent's own
        // `SessionSink` already holds open, deadlocking on `AlreadyLocked` — short-circuit with a
        // clear message instead of attempting a self-conflicting reopen.
        if let Some(sink) = &self.services.session.session_sink
            && sink.session_id().as_str() == id
        {
            return Ok(format!("Already in session '{id}'."));
        }
        let Some(memory) = self.services.memory.persistence.memory.clone() else {
            return Ok(
                "Conversation-session persistence requires memory to be enabled ([memory] enabled = true)."
                    .to_owned(),
            );
        };
        let store = zeph_session::SessionStore::new(memory.sqlite().pool().clone());
        let Some(metadata) = store
            .get(id)
            .await
            .map_err(|e| CommandError::new(e.to_string()))?
        else {
            return Ok(format!("Session '{id}' not found."));
        };

        let conversation_id = if let Some(cid) = metadata.conversation_id {
            zeph_memory::ConversationId(cid)
        } else {
            let cid = memory
                .sqlite()
                .create_conversation()
                .await
                .map_err(|e| CommandError::new(e.to_string()))?;
            store
                .link_conversation(id, cid.0)
                .await
                .map_err(|e| CommandError::new(e.to_string()))?;
            cid
        };

        let session_id = zeph_common::SessionId::new(id);
        self.load_and_resume_conversation(&session_id, conversation_id)
            .await
            .map_err(|e| CommandError::new(e.to_string()))?;

        Ok(format!(
            "Resumed session {id} ({} event(s) replayed).",
            metadata.event_count
        ))
    }

    /// `/conv fork <id>` (spec-068, #5343, D-9): eager-copies `id`'s durable log into a fresh
    /// child session via `ForkEngine::fork` (P2), then immediately live-swaps onto the child —
    /// same effect as `POST /sessions/:id/fork` (spec §9.4) but for the current CLI/TUI session
    /// instead of spawning a new `SessionActor`.
    async fn handle_conv_fork(&mut self, id: &str) -> Result<String, CommandError> {
        if id.is_empty() {
            return Ok("Usage: /conv fork <id>".to_owned());
        }
        let Some(memory) = self.services.memory.persistence.memory.clone() else {
            return Ok(
                "Conversation-session persistence requires memory to be enabled ([memory] enabled = true)."
                    .to_owned(),
            );
        };
        let Some(session_persistence_config) =
            self.services.session.session_persistence_config.clone()
        else {
            return Ok(
                "Conversation-session persistence is not enabled ([session] enabled = true)."
                    .to_owned(),
            );
        };
        let store = zeph_session::SessionStore::new(memory.sqlite().pool().clone());
        let data_dir = std::path::PathBuf::from(&session_persistence_config.data_dir);
        let new_id = zeph_common::SessionId::generate();

        let fork_result =
            zeph_session::ForkEngine::fork(&data_dir, id, new_id.as_str(), None, &store, None)
                .await
                .map_err(|e| CommandError::new(e.to_string()))?;

        let conversation_id = memory
            .sqlite()
            .create_conversation()
            .await
            .map_err(|e| CommandError::new(e.to_string()))?;

        self.load_and_resume_conversation(&new_id, conversation_id)
            .await
            .map_err(|e| CommandError::new(e.to_string()))?;

        Ok(format!(
            "Forked session {id} -> {} ({} event(s) copied); now the active conversation.",
            fork_result.new_session_id, fork_result.events_copied
        ))
    }
}

/// Formats `/conv list` — mirrors `sessions list`'s CLI table layout
/// (`src/commands/sessions.rs`) and `zeph serve-sessions`'s `GET /sessions`.
async fn handle_conv_list(store: &zeph_session::SessionStore) -> Result<String, CommandError> {
    use std::fmt::Write as _;

    let sessions = store
        .list(&zeph_session::SessionFilter::default())
        .await
        .map_err(|e| CommandError::new(format!("failed to list sessions: {e}")))?;

    if sessions.is_empty() {
        return Ok("No conversation-sessions found.".to_owned());
    }

    let mut out = format!(
        "{:<38} {:<30} {:<9} {:>6} {:<24}\n",
        "ID", "TITLE", "STATUS", "EVENTS", "UPDATED"
    );
    out.push_str(&"-".repeat(110));
    out.push('\n');
    for s in &sessions {
        let title = s.title.as_deref().unwrap_or("(untitled)");
        let _ = writeln!(
            out,
            "{:<38} {:<30} {:<9} {:>6} {:<24}",
            s.session_id,
            crate::text::truncate_to_chars(title, 30),
            s.status.as_str(),
            s.event_count,
            s.updated_at
        );
    }
    Ok(out.trim_end().to_owned())
}

/// Formats `/conv show <id>` — one session's metadata, mirroring `zeph serve-sessions`'s
/// `GET /sessions/:id` (metadata only; use `zeph sessions show --events <id>` on the CLI for a
/// full event-log dump).
async fn handle_conv_show(
    store: &zeph_session::SessionStore,
    id: &str,
) -> Result<String, CommandError> {
    if id.is_empty() {
        return Ok("Usage: /conv show <id>".to_owned());
    }
    let metadata = store
        .get(id)
        .await
        .map_err(|e| CommandError::new(format!("failed to read session metadata: {e}")))?;
    let Some(m) = metadata else {
        return Ok(format!("Session '{id}' not found."));
    };
    Ok(format!(
        "Session {}\n  title: {}\n  status: {}\n  events: {} (last_seq={})\n  forked_from: {}\n  created: {}\n  updated: {}",
        m.session_id,
        m.title.as_deref().unwrap_or("(untitled)"),
        m.status.as_str(),
        m.event_count,
        m.last_seq,
        m.forked_from.as_deref().unwrap_or("-"),
        m.created_at,
        m.updated_at
    ))
}

type GoalStore = crate::goal::GoalStore;
type GoalAccounting = crate::goal::GoalAccounting;

/// Hard cap on `--turns` to prevent runaway autonomous loops (Security Low).
const AUTONOMOUS_MAX_TURNS_CAP: u32 = 1000;

async fn goal_status(accounting: &GoalAccounting) -> Result<String, CommandError> {
    match accounting.get_active().await {
        Ok(Some(g)) => {
            let budget_line = g.token_budget.map_or_else(
                || format!("  tokens used: {}", g.tokens_used),
                |b| format!("  budget: {}/{b}", g.tokens_used),
            );
            Ok(format!(
                "Active goal [{}]: {}\n  status: {}\n  turns: {}\n{}",
                &g.id[..8],
                g.text,
                g.status,
                g.turns_used,
                budget_line
            ))
        }
        Ok(None) => Ok("No active goal. Use `/goal create <text>` to set one.".to_owned()),
        Err(e) => Ok(format!("Goal lookup failed: {e}")),
    }
}

/// Returns `(display_message, auto_start_request)`.
///
/// `auto_start_request` is `Some((goal_id, goal_text, max_turns))` when `--auto` was passed and
/// the goal was successfully created. The caller must relay this to `AutonomousDriver` via the
/// `pending_start_arc` side-channel before the future resolves.
async fn goal_create(
    args: &str,
    accounting: &GoalAccounting,
    store: &GoalStore,
    max_chars: usize,
    default_budget: Option<u64>,
    autonomous_enabled: bool,
    autonomous_max_turns: u32,
) -> Result<(String, Option<(String, String, u32)>), CommandError> {
    let rest = args.strip_prefix("create").unwrap_or("").trim();

    // Strip --auto / --turns before passing text to the budget parser.
    let (stripped, is_auto, explicit_turns) = parse_auto_flags(rest);
    let (text, explicit_budget) = parse_goal_create_args(&stripped);

    if text.is_empty() {
        return Ok((
            "Usage: /goal create <text> [--budget N] [--auto [--turns N]]".to_owned(),
            None,
        ));
    }
    if is_auto && !autonomous_enabled {
        return Ok((
            "Autonomous mode is disabled. Set `[goals] autonomous_enabled = true` in config."
                .to_owned(),
            None,
        ));
    }
    let budget = explicit_budget.or(default_budget.filter(|&b| b > 0));

    let max_turns = explicit_turns
        .unwrap_or(autonomous_max_turns)
        .min(AUTONOMOUS_MAX_TURNS_CAP);
    if explicit_turns.is_some_and(|t| t > AUTONOMOUS_MAX_TURNS_CAP) {
        tracing::warn!(
            requested = explicit_turns,
            capped = AUTONOMOUS_MAX_TURNS_CAP,
            "autonomous max_turns capped to {AUTONOMOUS_MAX_TURNS_CAP}"
        );
    }

    match store.create(text, budget, max_chars).await {
        Ok(g) => {
            let _ = accounting.refresh().await;
            let auto_start = if is_auto {
                Some((g.id.clone(), g.text.clone(), max_turns))
            } else {
                None
            };
            let auto_note = if is_auto {
                " Autonomous mode enabled — use `/goal clear` to stop."
            } else {
                ""
            };
            Ok((
                format!("Goal created [{}]: {}{auto_note}", &g.id[..8], g.text),
                auto_start,
            ))
        }
        Err(crate::goal::store::GoalError::TextTooLong { max }) => Ok((
            format!("Goal text exceeds {max} characters. Please shorten it."),
            None,
        )),
        Err(e) => Ok((format!("Failed to create goal: {e}"), None)),
    }
}

async fn goal_pause(
    accounting: &GoalAccounting,
    store: &GoalStore,
) -> Result<String, CommandError> {
    match accounting.get_active().await {
        Ok(Some(g)) => {
            match store
                .transition(&g.id, crate::goal::GoalStatus::Paused, g.updated_at)
                .await
            {
                Ok(_) => {
                    let _ = accounting.refresh().await;
                    Ok(format!("Goal [{}] paused.", &g.id[..8]))
                }
                Err(crate::goal::store::GoalError::StaleUpdate(_)) => {
                    let current = accounting.get_active().await.ok().flatten();
                    Ok(format!(
                        "Goal state changed concurrently. Current: {}",
                        current.map_or_else(|| "none".into(), |g| g.status.to_string())
                    ))
                }
                Err(e) => Ok(format!("Pause failed: {e}")),
            }
        }
        Ok(None) => Ok("No active goal to pause.".to_owned()),
        Err(e) => Ok(format!("Failed: {e}")),
    }
}

async fn goal_resume(
    accounting: &GoalAccounting,
    store: &GoalStore,
) -> Result<String, CommandError> {
    let goals = store.list(10).await.unwrap_or_default();
    let paused = goals
        .into_iter()
        .find(|g| g.status == crate::goal::GoalStatus::Paused);
    match paused {
        Some(g) => {
            match store
                .transition(&g.id, crate::goal::GoalStatus::Active, g.updated_at)
                .await
            {
                Ok(_) => {
                    let _ = accounting.refresh().await;
                    Ok(format!("Goal [{}] resumed: {}", &g.id[..8], g.text))
                }
                Err(crate::goal::store::GoalError::StaleUpdate(_)) => {
                    Ok("Goal state changed concurrently — please retry.".to_owned())
                }
                Err(e) => Ok(format!("Resume failed: {e}")),
            }
        }
        None => Ok("No paused goal to resume.".to_owned()),
    }
}

async fn goal_complete(
    accounting: &GoalAccounting,
    store: &GoalStore,
) -> Result<String, CommandError> {
    match accounting.get_active().await {
        Ok(Some(g)) => {
            match store
                .transition(&g.id, crate::goal::GoalStatus::Completed, g.updated_at)
                .await
            {
                Ok(_) => {
                    let _ = accounting.refresh().await;
                    Ok(format!("Goal [{}] marked complete.", &g.id[..8]))
                }
                Err(e) => Ok(format!("Complete failed: {e}")),
            }
        }
        Ok(None) => Ok("No active goal.".to_owned()),
        Err(e) => Ok(format!("Failed: {e}")),
    }
}

async fn goal_clear(
    accounting: &GoalAccounting,
    store: &GoalStore,
) -> Result<String, CommandError> {
    let goals = store.list(10).await.unwrap_or_default();
    let target = goals.into_iter().find(|g| {
        g.status == crate::goal::GoalStatus::Active || g.status == crate::goal::GoalStatus::Paused
    });
    match target {
        Some(g) => {
            match store
                .transition(&g.id, crate::goal::GoalStatus::Cleared, g.updated_at)
                .await
            {
                Ok(_) => {
                    let _ = accounting.refresh().await;
                    Ok(format!("Goal [{}] cleared.", &g.id[..8]))
                }
                Err(e) => Ok(format!("Clear failed: {e}")),
            }
        }
        None => Ok("No active or paused goal to clear.".to_owned()),
    }
}

async fn goal_list(store: &GoalStore) -> Result<String, CommandError> {
    let goals = store.list(20).await.unwrap_or_default();
    if goals.is_empty() {
        return Ok("No goals recorded.".to_owned());
    }
    let mut out = String::from("Goals:\n");
    for g in goals {
        let _ = std::fmt::Write::write_fmt(
            &mut out,
            format_args!(
                "  {} [{}] {} — {} turns\n",
                g.status.badge_symbol(),
                &g.id[..8],
                g.text,
                g.turns_used
            ),
        );
    }
    Ok(out.trim_end().to_owned())
}

fn parse_goal_create_args(args: &str) -> (&str, Option<u64>) {
    if let Some(pos) = args.find("--budget") {
        let text = args[..pos].trim();
        let rest = args[pos + "--budget".len()..].trim();
        let budget = rest
            .split_whitespace()
            .next()
            .and_then(|s| s.parse::<u64>().ok());
        (text, budget)
    } else {
        (args, None)
    }
}

/// Parse `--auto` and `--turns N` flags from the remainder of a `/goal create` argument string.
///
/// Returns `(text_without_auto_flags, is_auto, explicit_turns)`.
fn parse_auto_flags(args: &str) -> (String, bool, Option<u32>) {
    let mut is_auto = false;
    let mut turns: Option<u32> = None;
    let mut text_words: Vec<&str> = Vec::new();
    let mut words = args.split_whitespace();

    while let Some(w) = words.next() {
        if w == "--auto" {
            is_auto = true;
        } else if w == "--turns" {
            turns = words.next().and_then(|n| n.parse::<u32>().ok());
        } else {
            text_words.push(w);
        }
    }

    (text_words.join(" "), is_auto, turns)
}

/// Convert `AgentError` to `CommandError` for the trait boundary.
impl From<AgentError> for CommandError {
    fn from(e: AgentError) -> Self {
        Self(e.to_string())
    }
}

#[cfg(test)]
mod tests {
    use super::super::agent_tests::{
        MockChannel, MockToolExecutor, create_test_registry, mock_provider,
    };
    use super::*;
    use zeph_commands::traits::agent::AgentAccess;
    use zeph_memory::semantic::SemanticMemory;

    async fn memory_without_qdrant() -> SemanticMemory {
        SemanticMemory::new(
            ":memory:",
            "http://127.0.0.1:1",
            None,
            zeph_llm::any::AnyProvider::Mock(zeph_llm::mock::MockProvider::default()),
            "test-model",
        )
        .await
        .unwrap()
    }

    // R-CRIT-4111: when graph is enabled in config but graph_store is None
    // (Qdrant unreachable), graph command handlers must report
    // "unavailable" rather than "not enabled".
    #[tokio::test]
    async fn graph_stats_enabled_but_no_store_reports_unavailable() {
        let cfg = crate::config::GraphConfig {
            enabled: true,
            ..Default::default()
        };
        let memory = memory_without_qdrant().await;
        let cid = memory.sqlite().create_conversation().await.unwrap();
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_graph_config(cfg);

        let result = agent.graph_stats().await.unwrap();
        assert!(
            result.contains("unavailable"),
            "expected 'unavailable' but got: {result}"
        );
        assert!(
            !result.contains("not enabled"),
            "must not report 'not enabled' when graph is enabled: {result}"
        );
    }

    #[tokio::test]
    async fn graph_stats_disabled_reports_not_enabled() {
        let cfg = crate::config::GraphConfig {
            enabled: false,
            ..Default::default()
        };
        let memory = memory_without_qdrant().await;
        let cid = memory.sqlite().create_conversation().await.unwrap();
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_graph_config(cfg);

        let result = agent.graph_stats().await.unwrap();
        assert!(
            result.contains("not enabled"),
            "expected 'not enabled' but got: {result}"
        );
    }

    // R-CRIT-4136: graph_backfill must resolve extract_provider before entering the async block.
    // When extract_provider is set to an unknown name, resolve_background_provider falls back to
    // the primary provider — the backfill still completes (no messages to process).
    // This test confirms that the provider-resolution code path executes without panic or borrow
    // errors, which would occur if the old code tried to access `&mut self` inside `async move`.
    #[tokio::test]
    async fn graph_backfill_with_extract_provider_resolves_without_panic() {
        let cfg = crate::config::GraphConfig {
            enabled: true,
            extract_provider: zeph_config::providers::ProviderName::new("nonexistent-provider"),
            ..Default::default()
        };
        let mut memory = memory_without_qdrant().await;
        // Install a real SQLite-backed GraphStore so resolve_graph_store succeeds.
        let pool = memory.sqlite().pool().clone();
        memory.graph_store = Some(std::sync::Arc::new(zeph_memory::GraphStore::new(pool)));
        let cid = memory.sqlite().create_conversation().await.unwrap();
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_graph_config(cfg);

        let mut progress = vec![];
        let result = agent
            .graph_backfill(Some(10), &mut |msg| progress.push(msg))
            .await
            .unwrap();

        // With an empty store there are zero unprocessed messages → backfill completes immediately.
        assert!(
            result.contains("Backfill complete"),
            "expected 'Backfill complete' but got: {result}"
        );
    }

    // #6261: graph_backfill extracts each batch's unprocessed messages concurrently via
    // `futures::stream::iter(...).buffer_unordered(4)` instead of a sequential per-message
    // loop. buffer_unordered completes futures in an order that need not match input order, so
    // this test asserts on aggregate totals (immune to completion order) and on per-entity /
    // per-message presence, proving the concurrent rewrite neither drops nor double-counts
    // results relative to the pre-#6261 sequential behavior.
    #[tokio::test]
    async fn graph_backfill_concurrent_extraction_aggregates_stats_without_dropping_results() {
        let n = 6;
        let cfg = crate::config::GraphConfig {
            enabled: true,
            ..Default::default()
        };
        let mut memory = memory_without_qdrant().await;
        let store = install_graph_store(&mut memory);
        let cid = memory.sqlite().create_conversation().await.unwrap();

        for i in 0..n {
            sqlx::query(zeph_db::sql!(
                "INSERT INTO messages (conversation_id, role, content) VALUES (?1, 'user', ?2)"
            ))
            .bind(cid.0)
            .bind(format!("message body {i}"))
            .execute(memory.sqlite().pool())
            .await
            .unwrap();
        }

        // One canned extraction response per message, each yielding exactly one distinct
        // entity. MockProvider serves responses in call order (not message order), which
        // mirrors buffer_unordered's out-of-order completion.
        let responses: Vec<String> = (0..n)
            .map(|i| {
                format!(
                    r#"{{"entities":[{{"name":"Entity{i}","type":"concept","summary":""}}],"edges":[]}}"#
                )
            })
            .collect();

        let mut agent = Agent::new(
            mock_provider(responses),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_graph_config(cfg);

        let mut progress = vec![];
        let result = agent
            .graph_backfill(None, &mut |msg| progress.push(msg))
            .await
            .unwrap();

        assert!(
            result.contains(&format!("{n} entities")),
            "expected all {n} entities aggregated in the result, got: {result}"
        );
        assert!(
            result.contains(&format!("from {n} messages")),
            "expected all {n} messages counted as processed, got: {result}"
        );

        // No drops/double-counts at the store level: every entity must be present exactly once.
        for i in 0..n {
            let name = format!("entity{i}");
            let found = store
                .find_entity(&name, zeph_memory::EntityType::Concept)
                .await
                .unwrap();
            assert!(found.is_some(), "entity{i} must have been upserted");
        }

        // Every message in the batch must be marked processed — none left behind by a
        // buffer_unordered future that was dropped or never polled to completion.
        let remaining = store.unprocessed_message_count().await.unwrap();
        assert_eq!(remaining, 0, "all messages must be marked graph_processed");
    }

    // #6261 follow-up (impl-critic finding): the aggregation test above uses an in-memory
    // SQLite database, which `zeph-db`'s pool forces to a single connection
    // (`connect_sqlite`'s `effective_max = if path == ":memory:" { 1 }`, see
    // `crates/zeph-db/src/pool.rs`) — so it never actually exercises concurrent writers racing
    // for the SQLite write lock. This test uses a real file-backed database instead (default
    // pool_size = 5, WAL journal mode + 5s busy_timeout — see `DbConfig::connect_sqlite`) with
    // more unprocessed messages than the `buffer_unordered(4)` bound, so multiple pooled
    // connections genuinely contend for writes concurrently. It confirms `extract_and_store`'s
    // upserts — relying on WAL mode + busy_timeout + `EntityResolver`'s per-entity-name locking,
    // the same assumption `semantic_scan_plugin_add`'s existing `buffer_unordered(4)` usage
    // relies on — complete without a "database is locked" error under real multi-connection
    // write contention.
    #[tokio::test]
    async fn graph_backfill_concurrent_extraction_survives_real_sqlite_write_contention() {
        let n = 8;
        let tmp = tempfile::NamedTempFile::new().expect("tempfile");
        let path = tmp.path().to_str().expect("valid utf-8 path").to_owned();

        let cfg = crate::config::GraphConfig {
            enabled: true,
            ..Default::default()
        };
        let mut memory = SemanticMemory::new(
            &path,
            "http://127.0.0.1:1",
            None,
            zeph_llm::any::AnyProvider::Mock(zeph_llm::mock::MockProvider::default()),
            "test-model",
        )
        .await
        .unwrap();
        let store = install_graph_store(&mut memory);
        let cid = memory.sqlite().create_conversation().await.unwrap();

        for i in 0..n {
            sqlx::query(zeph_db::sql!(
                "INSERT INTO messages (conversation_id, role, content) VALUES (?1, 'user', ?2)"
            ))
            .bind(cid.0)
            .bind(format!("contention message body {i}"))
            .execute(memory.sqlite().pool())
            .await
            .unwrap();
        }

        // A small per-call delay forces the (up to 4) concurrently in-flight extraction futures
        // to genuinely overlap their subsequent SQLite writes, rather than happening to resolve
        // one at a time fast enough to never actually race.
        let responses: Vec<String> = (0..n)
            .map(|i| {
                format!(
                    r#"{{"entities":[{{"name":"ContentionEntity{i}","type":"concept","summary":""}}],"edges":[]}}"#
                )
            })
            .collect();
        let mut provider = zeph_llm::mock::MockProvider::with_responses(responses);
        provider.delay_ms = 15;
        let provider = zeph_llm::any::AnyProvider::Mock(provider);

        let mut agent = Agent::new(
            provider,
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_graph_config(cfg);

        let mut progress = vec![];
        let result = agent
            .graph_backfill(None, &mut |msg| progress.push(msg))
            .await
            .unwrap();

        assert!(
            result.contains(&format!("{n} entities")),
            "expected all {n} entities aggregated despite concurrent SQLite writers, got: {result}"
        );

        // The decisive assertion: if a concurrent writer had hit "database is locked"
        // (SQLITE_BUSY surfacing as an error instead of the busy_timeout retry succeeding),
        // extract_and_store logs a warning and skips that message's upsert (the `Err(e) =>
        // tracing::warn!(...)` arm in graph_backfill) rather than failing the whole batch — so a
        // missing entity here is the observable symptom of exactly the failure mode flagged.
        for i in 0..n {
            let name = format!("contentionentity{i}");
            let found = store
                .find_entity(&name, zeph_memory::EntityType::Concept)
                .await
                .unwrap();
            assert!(
                found.is_some(),
                "entity {i} must have been upserted; a missing entity indicates a dropped/failed \
                 concurrent write (e.g. a 'database is locked' error) under real multi-connection \
                 contention"
            );
        }

        let remaining = store.unprocessed_message_count().await.unwrap();
        assert_eq!(remaining, 0, "all messages must be marked graph_processed");
    }

    // R-4139: graph_entities with enabled graph but no store (Qdrant unreachable) must
    // report unavailable, not panic or hang.
    #[tokio::test]
    async fn graph_entities_enabled_but_no_store_reports_unavailable() {
        let cfg = crate::config::GraphConfig {
            enabled: true,
            ..Default::default()
        };
        let memory = memory_without_qdrant().await;
        let cid = memory.sqlite().create_conversation().await.unwrap();
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_graph_config(cfg);

        let result = agent.graph_entities().await.unwrap();
        assert!(
            result.contains("unavailable"),
            "expected 'unavailable' but got: {result}"
        );
    }

    // R-4139: graph_communities with enabled graph but no store must report unavailable.
    #[tokio::test]
    async fn graph_communities_enabled_but_no_store_reports_unavailable() {
        let cfg = crate::config::GraphConfig {
            enabled: true,
            ..Default::default()
        };
        let memory = memory_without_qdrant().await;
        let cid = memory.sqlite().create_conversation().await.unwrap();
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_graph_config(cfg);

        let result = agent.graph_communities().await.unwrap();
        assert!(
            result.contains("unavailable"),
            "expected 'unavailable' but got: {result}"
        );
    }

    // R-4139: verify that the tokio::time::timeout pattern used in graph handlers
    // correctly returns Err on a never-resolving future. This is a direct regression
    // guard for the fix introduced in #4139: before the fix, these calls had no
    // timeout guard and would block indefinitely when Qdrant was unreachable.
    #[tokio::test]
    async fn graph_store_timeout_pattern_fires_on_pending_future() {
        use std::future;
        let result = tokio::time::timeout(
            Duration::from_millis(10),
            future::pending::<Result<Vec<()>, String>>(),
        )
        .await;
        assert!(
            result.is_err(),
            "timeout must fire on a never-resolving future"
        );
    }

    // ── #5770: with_graph_store_timeout had zero coverage of its own timeout branch —
    // existing tests only reached the "no store configured" short-circuit in
    // resolve_graph_store(), never the 5s deadline shared by resolve_entity_by_name,
    // graph_facts, and graph_history. Exercise the extracted helper directly with a
    // paused clock so the deadline fires deterministically without a real wall-clock wait.

    #[tokio::test]
    async fn with_graph_store_timeout_completes_on_success() {
        let result = with_graph_store_timeout(async { Ok::<_, zeph_memory::MemoryError>(42) })
            .await
            .unwrap();
        assert!(matches!(result, StoreCallOutcome::Completed(42)));
    }

    #[tokio::test]
    async fn with_graph_store_timeout_maps_store_error_to_command_error() {
        let result = with_graph_store_timeout(async {
            Err::<i32, _>(zeph_memory::MemoryError::GraphStore("boom".to_owned()))
        })
        .await;
        assert!(result.is_err(), "store error must surface as CommandError");
    }

    #[tokio::test]
    async fn with_graph_store_timeout_times_out_on_pending_future() {
        tokio::time::pause();
        let fut = with_graph_store_timeout(std::future::pending::<
            Result<i32, zeph_memory::MemoryError>,
        >());
        let handle = tokio::spawn(fut); // EXEMPT: test-only tokio::time::pause harness
        tokio::time::advance(std::time::Duration::from_secs(6)).await;
        let result = handle.await.expect("task panicked");
        assert!(
            matches!(result, Ok(StoreCallOutcome::TimedOut)),
            "call must resolve to TimedOut once the 5s deadline elapses"
        );
    }

    // ── #5764: graph_facts / graph_history had zero dedicated test coverage ──────

    /// Installs a real SQLite-backed `GraphStore` on `memory` (mirrors
    /// `graph_backfill_with_extract_provider_resolves_without_panic`), returning an `Arc`
    /// clone so callers can seed entities/edges before handing `memory` to `with_memory`.
    fn install_graph_store(memory: &mut SemanticMemory) -> std::sync::Arc<zeph_memory::GraphStore> {
        let pool = memory.sqlite().pool().clone();
        let store = std::sync::Arc::new(zeph_memory::GraphStore::new(pool));
        memory.graph_store = Some(store.clone());
        store
    }

    #[tokio::test]
    async fn graph_facts_happy_path_returns_formatted_facts() {
        let mut memory = memory_without_qdrant().await;
        let store = install_graph_store(&mut memory);
        let cid = memory.sqlite().create_conversation().await.unwrap();

        let alice = store
            .upsert_entity(
                "Alice",
                "alice",
                zeph_memory::EntityType::Person,
                None,
                None,
            )
            .await
            .unwrap();
        let bob = store
            .upsert_entity("Bob", "bob", zeph_memory::EntityType::Person, None, None)
            .await
            .unwrap();
        store
            .insert_edge(alice.0, bob.0, "knows", "Alice knows Bob", 0.9, None, None)
            .await
            .unwrap();

        let cfg = crate::config::GraphConfig {
            enabled: true,
            ..Default::default()
        };
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_graph_config(cfg);

        let result = agent.graph_facts("Alice").await.unwrap();
        assert!(
            result.contains("Facts for 'Alice'"),
            "expected facts header, got: {result}"
        );
        assert!(
            result.contains("Bob"),
            "expected target entity name, got: {result}"
        );
        assert!(result.contains("knows"), "expected relation, got: {result}");
        assert!(
            result.contains("Alice knows Bob"),
            "expected fact text, got: {result}"
        );
    }

    #[tokio::test]
    async fn graph_facts_entity_not_found_returns_message() {
        let mut memory = memory_without_qdrant().await;
        install_graph_store(&mut memory);
        let cid = memory.sqlite().create_conversation().await.unwrap();

        let cfg = crate::config::GraphConfig {
            enabled: true,
            ..Default::default()
        };
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_graph_config(cfg);

        let result = agent.graph_facts("Nobody").await.unwrap();
        assert_eq!(result, "No entity found matching 'Nobody'.");
    }

    // Mirrors graph_entities_enabled_but_no_store_reports_unavailable (R-4139): when the graph
    // store is None (Qdrant unreachable) but graph is enabled, report unavailable rather than
    // hang or panic.
    #[tokio::test]
    async fn graph_facts_enabled_but_no_store_reports_unavailable() {
        let cfg = crate::config::GraphConfig {
            enabled: true,
            ..Default::default()
        };
        let memory = memory_without_qdrant().await;
        let cid = memory.sqlite().create_conversation().await.unwrap();
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_graph_config(cfg);

        let result = agent.graph_facts("Alice").await.unwrap();
        assert!(
            result.contains("unavailable"),
            "expected 'unavailable' but got: {result}"
        );
    }

    // Self-loop edges (source == target) are rejected both by `GraphStore::insert_edge_typed`
    // and by a DB-level trigger (migration 044_graph_edges_no_self_loops) — so a real one can
    // only arise from data written before that migration. Drop the trigger to simulate that
    // legacy row and confirm graph_facts' defensive `entity_names` bookkeeping (which already
    // knows the entity's own name before resolving edge endpoints) handles it without panicking
    // or falling back to a raw `#id` placeholder.
    #[tokio::test]
    async fn graph_facts_self_loop_edge_does_not_panic() {
        let mut memory = memory_without_qdrant().await;
        let store = install_graph_store(&mut memory);
        let cid = memory.sqlite().create_conversation().await.unwrap();

        let self_entity = store
            .upsert_entity("Self", "self", zeph_memory::EntityType::Concept, None, None)
            .await
            .unwrap();
        let pool = memory.sqlite().pool().clone();
        zeph_db::query(zeph_db::sql!(
            "DROP TRIGGER IF EXISTS graph_edges_no_self_loops"
        ))
        .execute(&pool)
        .await
        .unwrap();
        zeph_db::query(zeph_db::sql!(
            "INSERT INTO graph_edges (source_entity_id, target_entity_id, relation, fact, confidence) \
             VALUES (?, ?, ?, ?, ?)"
        ))
        .bind(self_entity.0)
        .bind(self_entity.0)
        .bind("refers_to")
        .bind("Self refers to itself")
        .bind(1.0_f64)
        .execute(&pool)
        .await
        .unwrap();

        let cfg = crate::config::GraphConfig {
            enabled: true,
            ..Default::default()
        };
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_graph_config(cfg);

        let result = agent.graph_facts("Self").await.unwrap();
        assert!(
            result.contains("Facts for 'Self'"),
            "expected facts header, got: {result}"
        );
        assert!(
            result.contains("refers_to"),
            "expected self-loop relation, got: {result}"
        );
        assert!(
            !result.contains('#'),
            "self-loop endpoint must resolve to the entity's own name, not a raw #id \
             placeholder: {result}"
        );
    }

    #[tokio::test]
    async fn graph_history_happy_path_returns_formatted_history() {
        let mut memory = memory_without_qdrant().await;
        let store = install_graph_store(&mut memory);
        let cid = memory.sqlite().create_conversation().await.unwrap();

        let alice = store
            .upsert_entity(
                "Alice",
                "alice",
                zeph_memory::EntityType::Person,
                None,
                None,
            )
            .await
            .unwrap();
        let bob = store
            .upsert_entity("Bob", "bob", zeph_memory::EntityType::Person, None, None)
            .await
            .unwrap();
        store
            .insert_edge(alice.0, bob.0, "knows", "Alice knows Bob", 0.9, None, None)
            .await
            .unwrap();

        let cfg = crate::config::GraphConfig {
            enabled: true,
            ..Default::default()
        };
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_graph_config(cfg);

        let result = agent.graph_history("Alice").await.unwrap();
        assert!(
            result.contains("Edge history for 'Alice'"),
            "expected history header, got: {result}"
        );
        assert!(
            result.contains("[active]"),
            "expected active tag, got: {result}"
        );
        assert!(
            result.contains("Bob"),
            "expected target entity name, got: {result}"
        );
    }

    #[tokio::test]
    async fn graph_history_entity_not_found_returns_message() {
        let mut memory = memory_without_qdrant().await;
        install_graph_store(&mut memory);
        let cid = memory.sqlite().create_conversation().await.unwrap();

        let cfg = crate::config::GraphConfig {
            enabled: true,
            ..Default::default()
        };
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_graph_config(cfg);

        let result = agent.graph_history("Nobody").await.unwrap();
        assert_eq!(result, "No entity found matching 'Nobody'.");
    }

    #[tokio::test]
    async fn graph_history_enabled_but_no_store_reports_unavailable() {
        let cfg = crate::config::GraphConfig {
            enabled: true,
            ..Default::default()
        };
        let memory = memory_without_qdrant().await;
        let cid = memory.sqlite().create_conversation().await.unwrap();
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_graph_config(cfg);

        let result = agent.graph_history("Alice").await.unwrap();
        assert!(
            result.contains("unavailable"),
            "expected 'unavailable' but got: {result}"
        );
    }

    // See graph_facts_self_loop_edge_does_not_panic for why the DB trigger must be dropped.
    #[tokio::test]
    async fn graph_history_self_loop_edge_does_not_panic() {
        let mut memory = memory_without_qdrant().await;
        let store = install_graph_store(&mut memory);
        let cid = memory.sqlite().create_conversation().await.unwrap();

        let self_entity = store
            .upsert_entity("Self", "self", zeph_memory::EntityType::Concept, None, None)
            .await
            .unwrap();
        let pool = memory.sqlite().pool().clone();
        zeph_db::query(zeph_db::sql!(
            "DROP TRIGGER IF EXISTS graph_edges_no_self_loops"
        ))
        .execute(&pool)
        .await
        .unwrap();
        zeph_db::query(zeph_db::sql!(
            "INSERT INTO graph_edges (source_entity_id, target_entity_id, relation, fact, confidence) \
             VALUES (?, ?, ?, ?, ?)"
        ))
        .bind(self_entity.0)
        .bind(self_entity.0)
        .bind("refers_to")
        .bind("Self refers to itself")
        .bind(1.0_f64)
        .execute(&pool)
        .await
        .unwrap();

        let cfg = crate::config::GraphConfig {
            enabled: true,
            ..Default::default()
        };
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_graph_config(cfg);

        let result = agent.graph_history("Self").await.unwrap();
        assert!(
            result.contains("Edge history for 'Self'"),
            "expected history header, got: {result}"
        );
        assert!(
            result.contains("refers_to"),
            "expected self-loop relation, got: {result}"
        );
        assert!(
            !result.contains('#'),
            "self-loop endpoint must resolve to the entity's own name, not a raw #id \
             placeholder: {result}"
        );
    }

    // R-4706/R-4709: when semantic_scan is enabled but semantic_scan_provider is empty,
    // `plugin add` must return a CommandError immediately (fail-closed). Before this fix
    // the code fell through to resolve_background_provider which silently used the primary
    // provider, bypassing the intent that an unconfigured scanner means "do not proceed".
    #[tokio::test]
    async fn plugin_add_semantic_scan_enabled_empty_provider_returns_error() {
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_semantic_scan(true, "");

        let result = agent.handle_plugins("add some-plugin").await;
        assert!(
            result.is_err(),
            "expected CommandError for missing semantic_scan_provider, got: {result:?}"
        );
        let msg = result.unwrap_err().to_string();
        assert!(
            msg.contains("semantic_scan_provider"),
            "error message must mention semantic_scan_provider, got: {msg}"
        );
    }

    // R-4706/R-4709: when semantic_scan is disabled, plugin subcommands must proceed
    // normally regardless of whether semantic_scan_provider is set.
    #[tokio::test]
    async fn plugin_list_semantic_scan_disabled_succeeds() {
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_semantic_scan(false, "");

        // "list" does not trigger scan logic; it should succeed without error.
        let result = agent.handle_plugins("list").await;
        assert!(
            result.is_ok(),
            "plugin list must succeed when semantic_scan is disabled, got: {result:?}"
        );
    }

    // R-4706/R-4709: "plugin add" with semantic_scan disabled must reach the install path
    // rather than return a scan-related error. The install itself may fail (no real plugin
    // source), but it must NOT fail with the fail-closed error message.
    #[tokio::test]
    async fn plugin_add_semantic_scan_disabled_no_scan_error() {
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_semantic_scan(false, "");

        let result = agent.handle_plugins("add some-plugin").await;
        // The call may succeed or fail for unrelated reasons (no real plugin source),
        // but must NOT fail with the fail-closed error about semantic_scan_provider.
        if let Err(ref e) = result {
            assert!(
                !e.to_string().contains("semantic_scan_provider"),
                "must not fail with scan error when semantic_scan is disabled, got: {e}"
            );
        }
    }

    // R-4705: semantic_scan_plugin_add must scan all skills concurrently and return
    // None when every scanner call returns Allow. Verifies buffer_unordered path
    // processes N inputs without sequential bottleneck.
    #[tokio::test]
    async fn semantic_scan_plugin_add_concurrent_all_allow_returns_none() {
        use zeph_llm::any::AnyProvider;
        use zeph_llm::mock::MockProvider;
        use zeph_skills::semantic_scanner::SkillSemanticScanner;

        // MockProvider returns `{"verdict":"allow","reason":"ok"}` for every call.
        let allow_json = r#"{"verdict":"allow","reason":"ok"}"#.to_owned();
        let provider = AnyProvider::Mock(MockProvider::with_responses(vec![
            allow_json.clone(),
            allow_json,
        ]));
        let scanner = SkillSemanticScanner::new(provider);

        // Build a minimal plugin layout with two skills so scan_targets returns
        // two SkillScanInput entries.
        let tmp = tempfile::tempdir().unwrap();
        let plugin_toml = r#"
[plugin]
name = "test-plugin"
version = "0.1.0"
description = "test"

[[skills]]
path = "skill-a"

[[skills]]
path = "skill-b"
"#;
        std::fs::write(tmp.path().join("plugin.toml"), plugin_toml).unwrap();
        for name in ["skill-a", "skill-b"] {
            let skill_dir = tmp.path().join(name);
            std::fs::create_dir_all(&skill_dir).unwrap();
            std::fs::write(
                skill_dir.join("SKILL.md"),
                format!("# {name}\n\n## Purpose\nTest skill.\n"),
            )
            .unwrap();
        }

        let result =
            semantic_scan_plugin_add(&scanner, tmp.path().to_str().unwrap(), None, vec![], vec![])
                .await;

        // All skills allowed → no error message returned.
        assert!(result.is_ok(), "expected Ok, got: {result:?}");
        assert!(
            result.unwrap().is_none(),
            "expected None (all passed) but got Some(err)"
        );
    }

    // R-4705 regression: buffer_unordered yields in completion order, not input order.
    // A Block verdict on the *second* skill (index 1) must name that second skill, not the
    // first. Before the fix, the code zipped verdicts against scan_inputs by position and
    // discarded the tuple's skill_name, so the wrong skill was reported.
    #[tokio::test]
    async fn semantic_scan_plugin_add_block_names_correct_skill() {
        use zeph_llm::any::AnyProvider;
        use zeph_llm::mock::MockProvider;
        use zeph_skills::semantic_scanner::SkillSemanticScanner;

        // First call returns Allow, second returns Block — only the second skill is rejected.
        let allow_json = r#"{"verdict":"allow","reason":"ok"}"#.to_owned();
        let block_json = r#"{"verdict":"block","reason":"malicious"}"#.to_owned();
        let provider =
            AnyProvider::Mock(MockProvider::with_responses(vec![allow_json, block_json]));
        let scanner = SkillSemanticScanner::new(provider);

        let tmp = tempfile::tempdir().unwrap();
        let plugin_toml = r#"
[plugin]
name = "test-plugin-block"
version = "0.1.0"
description = "test"

[[skills]]
path = "skill-first"

[[skills]]
path = "skill-second"
"#;
        std::fs::write(tmp.path().join("plugin.toml"), plugin_toml).unwrap();
        for name in ["skill-first", "skill-second"] {
            let skill_dir = tmp.path().join(name);
            std::fs::create_dir_all(&skill_dir).unwrap();
            std::fs::write(
                skill_dir.join("SKILL.md"),
                format!("# {name}\n\n## Purpose\nTest skill.\n"),
            )
            .unwrap();
        }

        let result =
            semantic_scan_plugin_add(&scanner, tmp.path().to_str().unwrap(), None, vec![], vec![])
                .await;

        assert!(result.is_ok(), "expected Ok(_), got: {result:?}");
        let msg = result
            .unwrap()
            .expect("expected Some(err) for blocked skill");
        assert!(
            msg.contains("skill-second"),
            "rejection must name the blocked skill 'skill-second', got: {msg}"
        );
        assert!(
            !msg.contains("skill-first"),
            "rejection must NOT name the allowed skill 'skill-first', got: {msg}"
        );
    }

    // R-4706/R-4709: unknown provider name must also fail-closed rather than silently
    // falling back to the primary provider via resolve_background_provider.
    #[tokio::test]
    async fn plugin_add_semantic_scan_unknown_provider_returns_error() {
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_semantic_scan(true, "nonexistent_provider");

        let result = agent.handle_plugins("add some-plugin").await;
        assert!(
            result.is_err(),
            "expected CommandError for unknown semantic_scan_provider, got: {result:?}"
        );
        let msg = result.unwrap_err().to_string();
        assert!(
            msg.contains("semantic_scan_provider"),
            "error message must mention semantic_scan_provider, got: {msg}"
        );
    }

    // #5487 fix 3: `handle_conv_resume` had zero prior test coverage. Resuming into the
    // session already live in this agent must short-circuit before attempting to re-acquire
    // the exclusive lock this agent's own SessionSink already holds (a guaranteed
    // `AlreadyLocked` self-deadlock, since flock conflicts are per open-file-description, not
    // per-process).
    #[tokio::test]
    async fn handle_conv_resume_same_session_short_circuits() {
        let memory = memory_without_qdrant().await;
        let cid = memory.sqlite().create_conversation().await.unwrap();
        let dir = tempfile::tempdir().unwrap();
        let data_dir = dir.path().to_path_buf();
        let session_id = zeph_common::SessionId::new("s1");
        let session_path = zeph_session::session_dir(&data_dir, session_id.as_str());
        let log = zeph_session::SessionEventLog::open_exclusive(&session_path)
            .await
            .unwrap();
        let store = zeph_session::SessionStore::new(memory.sqlite().pool().clone());
        let sink = zeph_agent_persistence::SessionSink::new(
            std::sync::Arc::new(log),
            store,
            session_id.clone(),
        );
        let session_config = zeph_config::SessionConfig {
            enabled: true,
            data_dir: data_dir.to_string_lossy().into_owned(),
            ..Default::default()
        };

        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_session_sink(Some(std::sync::Arc::new(sink)))
        .with_session_persistence_config(Some(session_config));

        let result = agent.handle_conv("resume s1").await.unwrap();
        assert_eq!(
            result, "Already in session 's1'.",
            "resuming into the currently-active session must short-circuit, not attempt \
             hydration/lock acquisition"
        );
    }

    // Regression check for the guard above: resuming into a genuinely different session (not
    // the one already live in this agent) must still hydrate normally.
    #[tokio::test]
    async fn handle_conv_resume_different_session_still_hydrates() {
        let memory = memory_without_qdrant().await;
        let cid = memory.sqlite().create_conversation().await.unwrap();
        let dir = tempfile::tempdir().unwrap();
        let data_dir = dir.path().to_path_buf();

        // Agent is currently "in" session s1, whose own lock is held by its SessionSink.
        let active_session_id = zeph_common::SessionId::new("s1");
        let active_session_path = zeph_session::session_dir(&data_dir, active_session_id.as_str());
        let active_log = zeph_session::SessionEventLog::open_exclusive(&active_session_path)
            .await
            .unwrap();
        let active_store = zeph_session::SessionStore::new(memory.sqlite().pool().clone());
        let active_sink = zeph_agent_persistence::SessionSink::new(
            std::sync::Arc::new(active_log),
            active_store,
            active_session_id,
        );

        // Target session s2 exists in the store (unlocked directory) — this is what should be
        // resumed into.
        let store = zeph_session::SessionStore::new(memory.sqlite().pool().clone());
        store.create("s2").await.unwrap();

        let session_config = zeph_config::SessionConfig {
            enabled: true,
            data_dir: data_dir.to_string_lossy().into_owned(),
            ..Default::default()
        };

        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_session_sink(Some(std::sync::Arc::new(active_sink)))
        .with_session_persistence_config(Some(session_config));

        let result = agent.handle_conv("resume s2").await.unwrap();
        assert!(
            result.starts_with("Resumed session s2"),
            "resuming into a different, unlocked session must still hydrate normally, got: {result}"
        );
    }

    // #5764: `/conv fork` had zero test coverage — only `/conv resume` was tested above.
    // Forks session "s1" into a fresh child and confirms the agent live-swaps onto it.
    #[tokio::test]
    async fn handle_conv_fork_creates_child_session_and_switches_to_it() {
        let memory = memory_without_qdrant().await;
        let cid = memory.sqlite().create_conversation().await.unwrap();
        let dir = tempfile::tempdir().unwrap();
        let data_dir = dir.path().to_path_buf();

        let store = zeph_session::SessionStore::new(memory.sqlite().pool().clone());
        store.create("s1").await.unwrap();
        let src_dir = zeph_session::session_dir(&data_dir, "s1");
        let log = zeph_session::SessionEventLog::open(&src_dir).await.unwrap();
        log.append(
            None,
            None,
            zeph_session::SessionEvent::SessionStarted {
                session_id: "s1".to_owned(),
                cwd: "/repo".to_owned(),
                provider_name: "claude".to_owned(),
                model: "opus".to_owned(),
                forked_from: None,
            },
        )
        .await
        .unwrap();
        store
            .update_seq("s1", log.last_seq().unwrap(), 1)
            .await
            .unwrap();
        drop(log);

        let session_config = zeph_config::SessionConfig {
            enabled: true,
            data_dir: data_dir.to_string_lossy().into_owned(),
            ..Default::default()
        };

        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
        .with_memory(std::sync::Arc::new(memory), cid, 50, 5, 100)
        .with_session_persistence_config(Some(session_config));

        let result = agent.handle_conv("fork s1").await.unwrap();
        assert!(
            result.starts_with("Forked session s1 ->"),
            "expected fork confirmation message, got: {result}"
        );
        assert!(
            result.contains("event(s) copied"),
            "expected copied-event count in confirmation, got: {result}"
        );
    }

    // `AgentAccess::load_image` had zero direct coverage — only
    // `Agent::handle_image_as_string` (slash_commands.rs) and `cli.rs`'s inline check
    // were tested. These exercise the real `Agent<C>` impl via the trait.

    #[tokio::test]
    async fn load_image_rejects_absolute_path() {
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        );

        let result = AgentAccess::load_image(&mut agent, "/etc/passwd")
            .await
            .unwrap();
        assert!(result.contains("absolute paths are not supported"));
    }

    #[tokio::test]
    async fn load_image_rejects_parent_dir_traversal() {
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        );

        let result = AgentAccess::load_image(&mut agent, "../../etc/passwd")
            .await
            .unwrap();
        assert!(result.contains("path traversal") && result.contains("not allowed"));
    }

    // ── /think-tokens, /reasoning-effort (#3098) ─────────────────────────

    fn claude_agent() -> Agent<MockChannel> {
        let provider = zeph_llm::any::AnyProvider::Claude(zeph_llm::claude::ClaudeProvider::new(
            "key".into(),
            "claude-sonnet-5".into(),
            4096,
        ));
        Agent::new(
            provider,
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        )
    }

    #[tokio::test]
    async fn handle_think_tokens_empty_arg_displays_off_by_default() {
        let mut agent = claude_agent();
        let out = agent.handle_think_tokens("").await;
        assert!(out.contains("off"), "{out}");
        assert!(out.contains("claude"), "{out}");
    }

    #[tokio::test]
    async fn handle_think_tokens_sets_and_displays_budget() {
        let mut agent = claude_agent();
        let set = agent.handle_think_tokens("8k").await;
        assert!(set.contains("8000"), "{set}");

        let show = agent.handle_think_tokens("").await;
        assert!(show.contains("8000"), "{show}");
    }

    #[tokio::test]
    async fn handle_think_tokens_off_disables() {
        let mut agent = claude_agent();
        agent.handle_think_tokens("8k").await;
        let out = agent.handle_think_tokens("off").await;
        assert!(out.contains("disabled"), "{out}");
        assert!(agent.provider.current_thinking_budget().is_none());
    }

    #[tokio::test]
    async fn handle_think_tokens_invalid_parse_returns_error_no_mutation() {
        let mut agent = claude_agent();
        let out = agent.handle_think_tokens("1.2.3k").await;
        assert!(out.contains("think-tokens"), "{out}");
        assert!(agent.provider.current_thinking_budget().is_none());
    }

    #[tokio::test]
    async fn handle_think_tokens_unsupported_provider_returns_explicit_message() {
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        );
        let out = agent.handle_think_tokens("8k").await;
        assert!(out.contains("does not support"), "{out}");
        assert!(out.contains("mock"), "{out}");
    }

    #[tokio::test]
    async fn handle_think_tokens_cross_override_note_when_reasoning_effort_was_active() {
        let mut agent = claude_agent();
        agent.handle_reasoning_effort("high").await;
        let out = agent.handle_think_tokens("8k").await;
        assert!(
            out.contains("overrides the previously set reasoning-effort"),
            "{out}"
        );
    }

    #[tokio::test]
    async fn handle_reasoning_effort_empty_arg_displays_off_by_default() {
        let mut agent = claude_agent();
        let out = agent.handle_reasoning_effort("").await;
        assert!(out.contains("off"), "{out}");
    }

    #[tokio::test]
    async fn handle_reasoning_effort_sets_and_displays_level() {
        let mut agent = claude_agent();
        let set = agent.handle_reasoning_effort("high").await;
        assert!(set.contains("high"), "{set}");

        let show = agent.handle_reasoning_effort("").await;
        assert!(show.contains("high"), "{show}");
    }

    #[tokio::test]
    async fn handle_reasoning_effort_invalid_parse_returns_error_no_mutation() {
        let mut agent = claude_agent();
        let out = agent.handle_reasoning_effort("minimal").await;
        assert!(out.contains("reasoning-effort"), "{out}");
        assert!(agent.provider.current_reasoning_effort().is_none());
    }

    #[tokio::test]
    async fn handle_reasoning_effort_unsupported_provider_returns_explicit_message() {
        let mut agent = Agent::new(
            mock_provider(vec![]),
            MockChannel::new(vec![]),
            create_test_registry(),
            None,
            5,
            MockToolExecutor::no_tools(),
        );
        let out = agent.handle_reasoning_effort("high").await;
        assert!(out.contains("does not support"), "{out}");
    }

    #[tokio::test]
    async fn handle_reasoning_effort_cross_override_note_when_think_tokens_was_active() {
        let mut agent = claude_agent();
        agent.handle_think_tokens("8k").await;
        let out = agent.handle_reasoning_effort("high").await;
        assert!(
            out.contains("overrides the previously set thinking-token budget"),
            "{out}"
        );
    }
}