huskies/server/src/agents/pool/start/spawn.rs

//! Background async work spawned by `AgentPool::start_agent`.
//!
//! `start_agent` returns immediately after registering the agent as `Pending`;
//! this module runs the slow worktree creation, agent process launch, and
//! event streaming in the background (story 157).

use std::collections::HashMap;
use std::path::PathBuf;
use std::sync::{Arc, Mutex};

use portable_pty::ChildKiller;
use tokio::sync::broadcast;

use crate::agent_log::AgentLogWriter;
use crate::config::ProjectConfig;
use crate::http::context::AppContext;
use crate::io::watcher::WatcherEvent;
use crate::{slog, slog_error};

use super::super::super::runtime::{
    AgentRuntime, ClaudeCodeRuntime, GeminiRuntime, OpenAiRuntime, RuntimeContext,
};
use super::super::super::{
    AgentEvent, AgentStatus, PipelineStage, agent_config_stage, pipeline_stage,
};
use super::super::AgentPool;
use super::super::types::StoryAgent;

/// Maximum bytes of gate output to include in the `--append-system-prompt`
/// failure section.  The tail is preserved (where errors appear).
const GATE_OUTPUT_PROMPT_BYTES: usize = 3_000;

/// Truncate `output` to at most [`GATE_OUTPUT_PROMPT_BYTES`], keeping the tail.
#[allow(clippy::string_slice)] // adjusted is walked forward to a char boundary before slicing
fn truncate_for_system_prompt(output: &str) -> &str {
    if output.len() <= GATE_OUTPUT_PROMPT_BYTES {
        return output;
    }
    let start = output.len() - GATE_OUTPUT_PROMPT_BYTES;
    let mut adjusted = start;
    while !output.is_char_boundary(adjusted) {
        adjusted += 1;
    }
    &output[adjusted..]
}

/// Inject a gate-failure section into the `--append-system-prompt` CLI arg.
///
/// If a `--append-system-prompt` pair already exists, appends to its value.
/// Otherwise adds a new `--append-system-prompt <section>` pair.
fn inject_gate_failure_section(args: &mut Vec<String>, gate_output: &str) {
    let section = format!(
        "Your previous run's quality gates failed:\n{}",
        truncate_for_system_prompt(gate_output)
    );
    if let Some(pos) = args.iter().position(|a| a == "--append-system-prompt")
        && let Some(val) = args.get_mut(pos + 1)
    {
        val.push_str("\n\n");
        val.push_str(&section);
        return;
    }
    args.push("--append-system-prompt".to_string());
    args.push(section);
}

/// On retry spawns (retry_count > 0), read the stored gate_output from the DB
/// and inject it into `--append-system-prompt` so the agent always sees the
/// prior failure context, even when session-resuming (story 881).
pub(super) fn maybe_inject_gate_failure(args: &mut Vec<String>, story_id: &str) {
    let retry_count = crate::crdt_state::read_item(story_id)
        .map(|item| item.retry_count())
        .unwrap_or(0);
    if retry_count > 0
        && let Some(gate_output) = crate::db::read_content(&format!("{story_id}:gate_output"))
    {
        inject_gate_failure_section(args, &gate_output);
    }
}

/// Run the background worktree-creation + agent-launch flow.
///
/// Caller (`AgentPool::start_agent`) wraps this in `tokio::spawn` and stores
/// the resulting handle on the Pending entry so cancellation works.
#[allow(clippy::too_many_arguments)]
pub(super) async fn run_agent_spawn(
    project_root: PathBuf,
    config: ProjectConfig,
    resume_context: Option<String>,
    session_id_to_resume: Option<String>,
    story_id: String,
    agent_name: String,
    tx: broadcast::Sender<AgentEvent>,
    agents: Arc<Mutex<HashMap<String, StoryAgent>>>,
    key: String,
    event_log: Arc<Mutex<Vec<AgentEvent>>>,
    port: u16,
    log_writer: Option<Arc<Mutex<AgentLogWriter>>>,
    child_killers: Arc<Mutex<HashMap<String, Box<dyn ChildKiller + Send + Sync>>>>,
    watcher_tx: broadcast::Sender<WatcherEvent>,
    inactivity_timeout_secs: u64,
    // Formatted `<recent-events>` block drained from the previous session's
    // buffer. Prepended to the first agent turn so the agent sees what
    // happened while it was idle (story 736).  `None` when there were no
    // buffered events.
    buffered_events_block: Option<String>,
    app_ctx: Option<Arc<AppContext>>,
) {
    // Re-bind to the legacy `_clone` / `_owned` names so the body below remains
    // a verbatim copy of the original closure (story 157).
    let project_root_clone = project_root;
    let config_clone = config;
    let resume_context_owned = resume_context;
    let session_id_to_resume_owned = session_id_to_resume;
    let sid = story_id;
    let aname = agent_name;
    let tx_clone = tx;
    let agents_ref = agents;
    let key_clone = key;
    let log_clone = event_log;
    let port_for_task = port;
    let log_writer_clone = log_writer;
    let child_killers_clone = child_killers;
    let watcher_tx_clone = watcher_tx;
    let _ = inactivity_timeout_secs; // currently unused inside the closure body

    // Step 1: create the worktree (slow — git checkout, pnpm install, etc.)
    let wt_info = match crate::worktree::create_worktree(
        &project_root_clone,
        &sid,
        &config_clone,
        port_for_task,
    )
    .await
    {
        Ok(wt) => wt,
        Err(e) => {
            let error_msg = format!("Failed to create worktree: {e}");
            slog_error!("[agents] {error_msg}");
            let event = AgentEvent::Error {
                story_id: sid.clone(),
                agent_name: aname.clone(),
                message: error_msg,
            };
            if let Ok(mut log) = log_clone.lock() {
                log.push(event.clone());
            }
            let _ = tx_clone.send(event);
            if let Ok(mut agents) = agents_ref.lock()
                && let Some(agent) = agents.get_mut(&key_clone)
            {
                agent.status = AgentStatus::Failed;
            }
            AgentPool::notify_agent_state_changed(&watcher_tx_clone);
            return;
        }
    };

    // Step 1.5: Update the source map for changed files since master.
    // Non-blocking — failures are logged but do not gate the spawn.
    {
        let wt_path_for_map = wt_info.path.clone();
        let base_for_map = wt_info.base_branch.clone();
        let map_path = project_root_clone.join(".huskies").join("source-map.json");
        match tokio::task::spawn_blocking(move || {
            source_map_gen::update_for_worktree(&wt_path_for_map, &base_for_map, &map_path)
        })
        .await
        .unwrap_or_else(|e| Err(e.to_string()))
        {
            Ok(()) => {}
            Err(e) => slog_error!("[agents] source map update for {sid}: {e}"),
        }
    }

    // Step 2: store worktree info and render agent command/args/prompt.
    let wt_path_str = wt_info.path.to_string_lossy().to_string();
    {
        if let Ok(mut agents) = agents_ref.lock()
            && let Some(agent) = agents.get_mut(&key_clone)
        {
            agent.worktree_info = Some(wt_info.clone());
        }
    }

    let (command, mut args, mut prompt) = match config_clone.render_agent_args(
        &wt_path_str,
        &sid,
        Some(&aname),
        Some(&wt_info.base_branch),
    ) {
        Ok(result) => result,
        Err(e) => {
            let error_msg = format!("Failed to render agent args: {e}");
            slog_error!("[agents] {error_msg}");
            let event = AgentEvent::Error {
                story_id: sid.clone(),
                agent_name: aname.clone(),
                message: error_msg,
            };
            if let Ok(mut log) = log_clone.lock() {
                log.push(event.clone());
            }
            let _ = tx_clone.send(event);
            if let Ok(mut agents) = agents_ref.lock()
                && let Some(agent) = agents.get_mut(&key_clone)
            {
                agent.status = AgentStatus::Failed;
            }
            AgentPool::notify_agent_state_changed(&watcher_tx_clone);
            return;
        }
    };

    // On retry spawns (retry_count > 0), inject prior gate failure output into
    // --append-system-prompt so the agent always sees the failure context (story 881).
    maybe_inject_gate_failure(&mut args, &sid);

    // Append project-local prompt content (.huskies/AGENT.md) to the
    // baked-in prompt so every agent role sees project-specific guidance
    // without any config changes.  The file is read fresh each spawn;
    // if absent or empty, the prompt is unchanged and no warning is logged.
    if let Some(local) = crate::agents::local_prompt::read_project_local_prompt(&project_root_clone)
    {
        prompt.push_str("\n\n");
        prompt.push_str(&local);
    }

    // Prepend epic context when the story belongs to an epic (AC3, story 880).
    // Read the story's front matter to find the epic ID, then load the epic's
    // content and prepend it to the system prompt so the agent treats it as
    // authoritative context.
    //
    // Epic linkage has no CRDT register yet (story 933) — wrap the parse in
    // `yaml_residue` so the gap is grep-findable.
    if let Some(story_content) = crate::db::read_content(&sid)
        && let Ok(meta) = crate::db::yaml_legacy::yaml_residue(
            crate::db::yaml_legacy::parse_front_matter(&story_content),
        )
        && let Some(ref epic_id) = meta.epic
        && let Some(epic_content) = crate::db::read_content(epic_id)
    {
        let block = format!(
            "# Epic Context\n\nThis work item belongs to epic `{epic_id}`.\
             The following is the authoritative epic context you must respect:\n\n\
             ---\n{epic_content}\n---"
        );
        prompt = format!("{block}\n\n{prompt}");
    }

    // Append a reference to the source map if the file was written.
    let source_map_path = project_root_clone.join(".huskies").join("source-map.json");
    if source_map_path.exists() {
        prompt.push_str(
            "\n\nA source map of well-documented changed files is at `.huskies/source-map.json`.",
        );
    }

    // Build the effective prompt and determine resume session.
    //
    // When resuming a previous session, discard the full rendered prompt
    // (which would re-read CLAUDE.md and README) and send only the gate
    // failure context as a new message.  On a fresh start, append the
    // failure context to the original prompt as before.
    let (effective_prompt, fresh_prompt) = match &session_id_to_resume_owned {
        Some(_) => {
            // Keep the full rendered prompt as fallback if resume fails.
            let fallback = prompt;
            (resume_context_owned.unwrap_or_default(), Some(fallback))
        }
        None => {
            if let Some(ctx) = resume_context_owned {
                prompt.push_str(&ctx);
            }
            (prompt, None)
        }
    };

    // Prepend buffered pipeline events from the previous idle period so the
    // agent sees what happened while it was not running (story 736).
    let effective_prompt = match buffered_events_block {
        Some(block) if !block.is_empty() => format!("{block}\n\n{effective_prompt}"),
        _ => effective_prompt,
    };

    // Step 3: transition to Running now that the worktree is ready.
    {
        if let Ok(mut agents) = agents_ref.lock()
            && let Some(agent) = agents.get_mut(&key_clone)
        {
            agent.status = AgentStatus::Running;
        }
    }
    let _ = tx_clone.send(AgentEvent::Status {
        story_id: sid.clone(),
        agent_name: aname.clone(),
        status: "running".to_string(),
    });
    AgentPool::notify_agent_state_changed(&watcher_tx_clone);
    let _ = watcher_tx_clone.send(WatcherEvent::AgentStarted {
        story_id: sid.clone(),
        agent_name: aname.clone(),
    });

    // Step 4: launch the agent process via the configured runtime.
    let runtime_name = config_clone
        .find_agent(&aname)
        .and_then(|a| a.runtime.as_deref())
        .unwrap_or("claude-code");

    let run_result = match runtime_name {
        "claude-code" => {
            let runtime =
                ClaudeCodeRuntime::new(child_killers_clone.clone(), watcher_tx_clone.clone());
            let ctx = RuntimeContext {
                story_id: sid.clone(),
                agent_name: aname.clone(),
                command,
                args,
                prompt: effective_prompt,
                cwd: wt_path_str,
                inactivity_timeout_secs,
                app_ctx: app_ctx.clone(),
                session_id_to_resume: session_id_to_resume_owned.clone(),
                fresh_prompt: fresh_prompt.clone(),
            };
            runtime
                .start(ctx, tx_clone.clone(), log_clone.clone(), log_writer_clone)
                .await
        }
        "gemini" => {
            let runtime = GeminiRuntime::new();
            let ctx = RuntimeContext {
                story_id: sid.clone(),
                agent_name: aname.clone(),
                command,
                args,
                prompt: effective_prompt,
                cwd: wt_path_str,
                inactivity_timeout_secs,
                app_ctx: app_ctx.clone(),
                session_id_to_resume: session_id_to_resume_owned.clone(),
                fresh_prompt: fresh_prompt.clone(),
            };
            runtime
                .start(ctx, tx_clone.clone(), log_clone.clone(), log_writer_clone)
                .await
        }
        "openai" => {
            let runtime = OpenAiRuntime::new();
            let ctx = RuntimeContext {
                story_id: sid.clone(),
                agent_name: aname.clone(),
                command,
                args,
                prompt: effective_prompt,
                cwd: wt_path_str,
                inactivity_timeout_secs,
                app_ctx: app_ctx.clone(),
                session_id_to_resume: session_id_to_resume_owned,
                fresh_prompt,
            };
            runtime
                .start(ctx, tx_clone.clone(), log_clone.clone(), log_writer_clone)
                .await
        }
        other => Err(format!(
            "Unknown agent runtime '{other}'; check the 'runtime' field in project.toml. \
                     Supported: 'claude-code', 'gemini', 'openai'"
        )),
    };

    match run_result {
        Ok(result) => {
            // Persist token usage if the agent reported it.
            if let Some(ref usage) = result.token_usage
                && let Ok(agents) = agents_ref.lock()
                && let Some(agent) = agents.get(&key_clone)
                && let Some(ref pr) = agent.project_root
            {
                let model = config_clone
                    .find_agent(&aname)
                    .and_then(|a| a.model.clone());
                let record =
                    crate::agents::token_usage::build_record(&sid, &aname, model, usage.clone());
                if let Err(e) = crate::agents::token_usage::append_record(pr, &record) {
                    slog_error!(
                        "[agents] Failed to persist token usage for \
                                 {sid}:{aname}: {e}"
                    );
                }
            }

            // Persist session_id so respawns can resume prior reasoning.
            if let Some(ref sess_id) = result.session_id {
                let model = config_clone
                    .find_agent(&aname)
                    .and_then(|a| a.model.clone())
                    .unwrap_or_default();
                crate::agents::session_store::record_session(
                    &project_root_clone,
                    &sid,
                    &aname,
                    &model,
                    sess_id,
                );
            }

            // Mergemaster agents have their own completion path via
            // start_merge_agent_work / run_merge_pipeline and must NOT go
            // through server-owned gates.  When a mergemaster exits early
            // (e.g. rate-limited before calling start_merge_agent_work) the
            // feature-branch worktree compiles fine and post-merge tests on
            // master pass (nothing changed), which would wrongly advance the
            // story to 5_done/ without any squash merge having occurred.
            // Instead: just remove the agent from the pool and let
            // auto-assign restart a new mergemaster for the story.
            let stage = config_clone
                .find_agent(&aname)
                .map(agent_config_stage)
                .unwrap_or_else(|| pipeline_stage(&aname));

            // AC1/AC2 (bug 882): CLI crashed (SIGABRT) before establishing a
            // session.  Respawn immediately without running gates or incrementing
            // retry_count.  Cap consecutive crash-respawns at 5 to avoid
            // infinite loops; after the cap, block the story with a clear reason.
            if result.aborted_signal && stage != PipelineStage::Mergemaster {
                const ABORT_RESPAWN_CAP: u32 = 5;
                let db_key = format!("{sid}:abort_respawn_count");
                let count = crate::db::read_content(&db_key)
                    .and_then(|s| s.trim().parse::<u32>().ok())
                    .unwrap_or(0)
                    + 1;
                crate::db::write_content(&db_key, &count.to_string());

                // Remove the agent entry from the pool and emit Done so that
                // any caller blocked on wait_for_agent is unblocked.
                let tx_done = {
                    let mut lock = match agents_ref.lock() {
                        Ok(a) => a,
                        Err(_) => return,
                    };
                    if let Some(agent) = lock.remove(&key_clone) {
                        agent.tx
                    } else {
                        tx_clone.clone()
                    }
                };
                let _ = tx_done.send(AgentEvent::Done {
                    story_id: sid.clone(),
                    agent_name: aname.clone(),
                    session_id: None,
                });
                AgentPool::notify_agent_state_changed(&watcher_tx_clone);

                if count >= ABORT_RESPAWN_CAP {
                    let reason = format!(
                        "CLI crashed before establishing a session (signal=Aborted, no session) \
                         {count} times in a row. Stopping to avoid an infinite respawn loop."
                    );
                    slog_error!(
                        "[agents] Story '{sid}' blocked after {count} consecutive CLI crashes."
                    );
                    if let Err(e) = crate::agents::lifecycle::transition_to_blocked(&sid, &reason) {
                        slog_error!("[agents] Failed to block '{sid}' after abort cap: {e}");
                    }
                    let _ = watcher_tx_clone.send(WatcherEvent::StoryBlocked {
                        story_id: sid.clone(),
                        reason,
                    });
                } else {
                    // Prune session_store entries for this story so the next
                    // spawn starts cold (no `--resume` flag). The crash likely
                    // came from claude-code choking on the bloated stdio
                    // replay; resuming again would re-trigger the same abort.
                    crate::agents::session_store::remove_sessions_for_story(
                        &project_root_clone,
                        &sid,
                    );
                    slog!(
                        "[agents] CLI crashed before session for '{sid}:{aname}' \
                         (abort respawn {count}/{ABORT_RESPAWN_CAP}). \
                         Pruned session_store and respawning cold without \
                         consuming a retry slot."
                    );
                    let agents_for_respawn = Arc::clone(&agents_ref);
                    let watcher_for_respawn = watcher_tx_clone.clone();
                    let sid_r = sid.clone();
                    let aname_r = aname.clone();
                    let root_r = project_root_clone.clone();
                    let port_r = port_for_task;
                    tokio::spawn(async move {
                        let pool = AgentPool {
                            agents: agents_for_respawn,
                            port: port_r,
                            child_killers: Arc::new(Mutex::new(HashMap::new())),
                            watcher_tx: watcher_for_respawn,
                            status_broadcaster: Arc::new(
                                crate::service::status::StatusBroadcaster::new(),
                            ),
                        };
                        if let Err(e) = pool
                            .start_agent(&root_r, &sid_r, Some(&aname_r), None, None)
                            .await
                        {
                            slog_error!(
                                "[agents] Failed to respawn '{aname_r}' for '{sid_r}' \
                                 after CLI crash: {e}"
                            );
                        }
                    });
                }
                return;
            }

            // Reset the abort-respawn counter on any non-aborted exit so that
            // a single successful run clears the consecutive-crash history.
            crate::db::delete_content(&format!("{sid}:abort_respawn_count"));

            if stage == PipelineStage::Mergemaster {
                let (tx_done, done_session_id, merge_failure_reported) = {
                    let mut lock = match agents_ref.lock() {
                        Ok(a) => a,
                        Err(_) => return,
                    };
                    if let Some(agent) = lock.remove(&key_clone) {
                        (
                            agent.tx,
                            agent.session_id.or(result.session_id),
                            agent.merge_failure_reported,
                        )
                    } else {
                        (tx_clone.clone(), result.session_id, false)
                    }
                };
                // Clear any stale Running merge job so the next mergemaster
                // can call start_merge_agent_work without hitting "Merge
                // already in progress" (bug 498).
                if crate::crdt_state::read_merge_job(&sid)
                    .is_some_and(|job| job.status == "running")
                {
                    crate::crdt_state::delete_merge_job(&sid);
                }
                // Classify termination: genuine (report_merge_failure called, or
                // the transient-respawn budget is exhausted) vs transient
                // (watchdog / rate-limit / crash without an explicit give-up call).
                // Only mark mergemaster_attempted on a genuine give-up so that
                // transient exits can be re-spawned up to the cap (story 920).
                const MERGEMASTER_RESPAWN_CAP: u32 = 3;
                let spawn_count_key = format!("{sid}:mergemaster_spawn_count");
                let is_genuine = if merge_failure_reported {
                    slog!(
                        "[agents] Mergemaster '{aname}' for '{sid}' gave up genuinely \
                         (report_merge_failure called)."
                    );
                    true
                } else {
                    let count = crate::db::read_content(&spawn_count_key)
                        .and_then(|s| s.trim().parse::<u32>().ok())
                        .unwrap_or(0)
                        + 1;
                    crate::db::write_content(&spawn_count_key, &count.to_string());
                    if count >= MERGEMASTER_RESPAWN_CAP {
                        slog!(
                            "[agents] Mergemaster '{aname}' for '{sid}' exhausted \
                             respawn budget ({count}/{MERGEMASTER_RESPAWN_CAP}); \
                             marking as permanently blocked."
                        );
                        true
                    } else {
                        slog!(
                            "[agents] Mergemaster '{aname}' for '{sid}' terminated \
                             transiently (spawn {count}/{MERGEMASTER_RESPAWN_CAP}); \
                             will re-spawn."
                        );
                        false
                    }
                };
                if is_genuine {
                    crate::crdt_state::set_mergemaster_attempted(&sid, true);
                }
                let _ = tx_done.send(AgentEvent::Done {
                    story_id: sid.clone(),
                    agent_name: aname.clone(),
                    session_id: done_session_id,
                });
                AgentPool::notify_agent_state_changed(&watcher_tx_clone);
                // Send a WorkItem event so the auto-assign watcher loop
                // re-dispatches a new mergemaster if the story still needs
                // merging.  This avoids an async call to start_agent inside
                // a tokio::spawn (which would require Send).
                let _ = watcher_tx_clone.send(crate::io::watcher::WatcherEvent::WorkItem {
                    stage: "4_merge".to_string(),
                    item_id: sid.clone(),
                    action: "reassign".to_string(),
                    commit_msg: String::new(),
                    from_stage: None,
                });
            } else {
                // Server-owned completion: run acceptance gates automatically
                // when the agent process exits normally.
                super::super::pipeline::run_server_owned_completion(
                    &agents_ref,
                    port_for_task,
                    &sid,
                    &aname,
                    result.session_id,
                    watcher_tx_clone.clone(),
                )
                .await;
                AgentPool::notify_agent_state_changed(&watcher_tx_clone);
            }
        }
        Err(e) => {
            slog_error!("[agents] Agent process error for {aname} on {sid}: {e}");
            let event = AgentEvent::Error {
                story_id: sid.clone(),
                agent_name: aname.clone(),
                message: e,
            };
            if let Ok(mut log) = log_clone.lock() {
                log.push(event.clone());
            }
            let _ = tx_clone.send(event);
            if let Ok(mut agents) = agents_ref.lock()
                && let Some(agent) = agents.get_mut(&key_clone)
            {
                agent.status = AgentStatus::Failed;
            }
            AgentPool::notify_agent_state_changed(&watcher_tx_clone);
        }
    }
}

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

    /// AC1 + AC3 (story 881): when retry_count=1 and gate_output is stored in
    /// the DB, `maybe_inject_gate_failure` injects a failure section beginning
    /// `Your previous run's quality gates failed:` into `--append-system-prompt`.
    #[test]
    fn gate_failure_injected_into_system_prompt_on_retry() {
        crate::crdt_state::init_for_test();
        crate::db::ensure_content_store();

        let story_id = "9960_story_gate_injection_881";
        crate::db::write_item_with_content(
            story_id,
            "2_current",
            "---\nname: Test\n---\n",
            crate::db::ItemMeta::from_yaml("---\nname: Test\n---\n"),
        );
        crate::crdt_state::set_retry_count(story_id, 1);

        let gate_output =
            "error[E0308]: mismatched types\n --> src/lib.rs:5:10\n  = expected i32, found &str";
        crate::db::write_content(&format!("{story_id}:gate_output"), gate_output);

        let mut args: Vec<String> = vec!["--verbose".to_string()];
        maybe_inject_gate_failure(&mut args, story_id);

        let pos = args
            .iter()
            .position(|a| a == "--append-system-prompt")
            .expect("--append-system-prompt must be present after retry injection");
        let value = &args[pos + 1];
        assert!(
            value.starts_with("Your previous run's quality gates failed:"),
            "--append-system-prompt must begin with the failure marker; got: {value}"
        );
        assert!(
            value.contains("mismatched types"),
            "--append-system-prompt must contain a snippet of gate_output; got: {value}"
        );
    }

    /// AC2 (story 881): first-attempt spawn (retry_count == 0) must NOT add the
    /// failure section to `--append-system-prompt`.
    #[test]
    fn gate_failure_not_injected_on_first_attempt() {
        crate::crdt_state::init_for_test();
        crate::db::ensure_content_store();

        let story_id = "9961_story_no_gate_injection_881";
        crate::db::write_item_with_content(
            story_id,
            "2_current",
            "---\nname: Test\n---\n",
            crate::db::ItemMeta::from_yaml("---\nname: Test\n---\n"),
        );
        // retry_count is 0 (default — never bumped).

        crate::db::write_content(&format!("{story_id}:gate_output"), "some previous output");

        let mut args: Vec<String> = vec!["--verbose".to_string()];
        maybe_inject_gate_failure(&mut args, story_id);

        assert!(
            !args.iter().any(|a| a == "--append-system-prompt"),
            "no --append-system-prompt should be added when retry_count is 0; args: {args:?}"
        );
    }

    /// Injection appends to an existing `--append-system-prompt` value rather
    /// than adding a duplicate flag.
    #[test]
    fn gate_failure_appends_to_existing_system_prompt() {
        let gate_output = "test failure output";
        let mut args = vec![
            "--append-system-prompt".to_string(),
            "base prompt".to_string(),
        ];
        inject_gate_failure_section(&mut args, gate_output);

        // Only one --append-system-prompt flag.
        let count = args
            .iter()
            .filter(|a| a.as_str() == "--append-system-prompt")
            .count();
        assert_eq!(count, 1, "must not duplicate --append-system-prompt");

        let pos = args
            .iter()
            .position(|a| a == "--append-system-prompt")
            .unwrap();
        let value = &args[pos + 1];
        assert!(
            value.contains("base prompt"),
            "original prompt must be preserved"
        );
        assert!(
            value.contains("Your previous run's quality gates failed:"),
            "failure section must be appended"
        );
        assert!(
            value.contains("test failure output"),
            "gate_output must appear in value"
        );
    }

    /// AC3 (bug 882): simulates the abort-respawn counter mechanism to verify that
    /// retry_count is never bumped during consecutive aborted+no-session exits and
    /// that the abort counter reaches the cap (5) before blocking.
    #[test]
    fn abort_respawn_leaves_retry_count_unchanged_and_caps_at_five() {
        crate::crdt_state::init_for_test();
        crate::db::ensure_content_store();

        let story_id = "9962_story_abort_respawn_882";
        crate::db::write_item_with_content(
            story_id,
            "2_current",
            "---\nname: Test\n---\n",
            crate::db::ItemMeta::from_yaml("---\nname: Test\n---\n"),
        );

        let db_key = format!("{story_id}:abort_respawn_count");
        const CAP: u32 = 5;

        // Simulate CAP consecutive abort-before-session exits.
        for expected_count in 1u32..=CAP {
            // This is exactly the counter logic in run_agent_spawn's abort path.
            let count = crate::db::read_content(&db_key)
                .and_then(|s| s.trim().parse::<u32>().ok())
                .unwrap_or(0)
                + 1;
            crate::db::write_content(&db_key, &count.to_string());
            assert_eq!(
                count, expected_count,
                "abort counter must increment by 1 each time"
            );

            // retry_count must remain 0 — the abort path never calls bump_retry_count.
            let retry_count = crate::crdt_state::read_item(story_id)
                .map(|item| item.retry_count())
                .unwrap_or(0);
            assert_eq!(
                retry_count, 0,
                "retry_count must not be incremented by the abort-respawn path \
                 (got {retry_count} on cycle {expected_count})"
            );
        }

        // After CAP cycles the counter equals the cap — the story would be blocked.
        let final_count: u32 = crate::db::read_content(&db_key)
            .and_then(|s| s.trim().parse().ok())
            .unwrap_or(0);
        assert_eq!(
            final_count, CAP,
            "counter must equal {CAP} after {CAP} abort cycles"
        );
        assert!(
            final_count >= CAP,
            "a count of {final_count} triggers blocking (>= {CAP})"
        );
    }
}