huskies: merge 802

server/src/agent_mode/claim.rs

@@ -0,0 +1,293 @@
+//! Claim ownership logic: deterministic hash-based tie-breaking and TTL constants.
+
+/// Default claim TTL in seconds.  If a claim has not been refreshed within this
+/// window, other nodes may displace the stale holder and claim the story.
+/// A node actively working on a story should refresh its claim periodically.
+pub(crate) const CLAIM_TIMEOUT_SECS: f64 = 1800.0; // 30 minutes
+
+/// Interval between heartbeat writes and work scans.
+pub const SCAN_INTERVAL_SECS: u64 = 15;
+
+// ── Hash-based tie-break ──────────────────────────────────────────────────
+
+/// Compute the claim-priority hash for a `(node_id, story_id)` pair.
+///
+/// Uses SHA-256(`node_id` bytes ++ `story_id` bytes), truncated to the first
+/// 8 bytes interpreted as a big-endian `u64`.  This function is:
+///
+/// * **Deterministic** — same inputs always produce the same output.
+/// * **Stable across restarts** — depends only on the node's persistent id
+///   and the story id, not on wall-clock time or random state.
+/// * **Cross-implementation portable** — SHA-256 is a standard primitive; any
+///   conforming implementation will produce identical values.
+pub(super) fn claim_hash(node_id: &str, story_id: &str) -> u64 {
+    use sha2::{Digest, Sha256};
+    let mut hasher = Sha256::new();
+    hasher.update(node_id.as_bytes());
+    hasher.update(story_id.as_bytes());
+    let digest = hasher.finalize();
+    u64::from_be_bytes(digest[..8].try_into().expect("sha256 is 32 bytes"))
+}
+
+/// Decide whether this node should be the one to claim `story_id`.
+///
+/// Returns `true` iff `claim_hash(self_node_id, story_id)` is **strictly
+/// lower** than the hash of every alive peer.  When there are no alive peers
+/// (single-node cluster) the result is always `true`.
+///
+/// # Trade-off note
+/// Because the winning node is determined purely by the hash of its id and the
+/// story id, the distribution is uniform per story but a given node may
+/// consistently "win" or "lose" across a set of stories depending on how its
+/// id happens to hash.  For 2–5 node clusters this imbalance is negligible in
+/// practice: any node is the lowest-hash winner with probability ≈ 1/N for a
+/// random story id, so the long-run distribution is approximately fair.  For
+/// clusters with many nodes (e.g. >10) the expected variance is larger and
+/// operators may want a different work-distribution strategy.
+pub fn should_self_claim(
+    self_node_id: &str,
+    story_id: &str,
+    alive_peer_node_ids: &[String],
+) -> bool {
+    let my_hash = claim_hash(self_node_id, story_id);
+    for peer_id in alive_peer_node_ids {
+        // Skip self if it appears in the peer list.
+        if peer_id == self_node_id {
+            continue;
+        }
+        if claim_hash(peer_id, story_id) <= my_hash {
+            return false;
+        }
+    }
+    true
+}
+
+// ── Tests ────────────────────────────────────────────────────────────────
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn claim_timeout_is_thirty_minutes() {
+        assert_eq!(CLAIM_TIMEOUT_SECS, 1800.0);
+    }
+
+    /// AC: seed a stale claim older than the TTL, attempt a new claim from a
+    /// different agent, assert the new claim succeeds and displacement is logged.
+    #[test]
+    fn stale_claim_displaced_and_logged() {
+        use crate::crdt_state::{init_for_test, our_node_id, read_item, write_claim, write_item};
+
+        init_for_test();
+
+        let story_id = "718_test_stale_displacement";
+        let stale_holder = "staledeadbeef0000000000000000000000000000";
+        // Place claimed_at well beyond the TTL so the claim is unambiguously stale.
+        let stale_time = chrono::Utc::now().timestamp() as f64 - CLAIM_TIMEOUT_SECS - 300.0;
+
+        // Seed the story with a stale claim from a foreign node.
+        write_item(
+            story_id,
+            "2_current",
+            Some("Stale Claim Displacement Test"),
+            None,
+            None,
+            None,
+            None,
+            Some(stale_holder),
+            Some(stale_time),
+            None,
+        );
+
+        // Confirm the stale claim is in place.
+        let before = read_item(story_id).expect("item should exist");
+        assert_eq!(
+            before.claimed_by.as_deref(),
+            Some(stale_holder),
+            "pre-condition: item should be claimed by the stale holder"
+        );
+        let age = chrono::Utc::now().timestamp() as f64 - before.claimed_at.unwrap_or(0.0);
+        assert!(
+            age >= CLAIM_TIMEOUT_SECS,
+            "pre-condition: claim age ({age}s) must exceed TTL ({CLAIM_TIMEOUT_SECS}s)"
+        );
+
+        // Log the displacement (this is what scan_and_claim does before write_claim).
+        crate::slog!(
+            "[agent-mode] Displacing stale claim on '{}' held by {:.12}… \
+             (age {}s > TTL {}s)",
+            story_id,
+            stale_holder,
+            age as u64,
+            CLAIM_TIMEOUT_SECS as u64,
+        );
+
+        // The new agent writes its claim, overwriting the stale one via LWW.
+        let success = write_claim(story_id);
+        assert!(
+            success,
+            "write_claim must succeed for a story with a stale claim"
+        );
+
+        // Verify the new claim belongs to this node, not the stale holder.
+        let our_id = our_node_id().expect("node id should be available after init_for_test");
+        let after = read_item(story_id).expect("item should still exist");
+        assert_eq!(
+            after.claimed_by.as_deref(),
+            Some(our_id.as_str()),
+            "new claim should have displaced the stale holder"
+        );
+        assert_ne!(
+            after.claimed_by.as_deref(),
+            Some(stale_holder),
+            "stale holder must no longer own the claim"
+        );
+
+        // Verify the displacement was logged.
+        let logs =
+            crate::log_buffer::global().get_recent(100, Some("Displacing stale claim"), None);
+        assert!(
+            !logs.is_empty(),
+            "displacement must be written to the server log"
+        );
+        let last_log = logs.last().unwrap();
+        assert!(
+            last_log.contains(story_id),
+            "log entry must name the story; got: {last_log}"
+        );
+        assert!(
+            last_log.contains(&stale_holder[..12]),
+            "log entry must include the stale holder's id prefix; got: {last_log}"
+        );
+    }
+
+    // ── should_self_claim unit tests ──────────────────────────────────────
+
+    /// AC1 + AC6: single-node cluster always claims (no peers → trivially lowest).
+    #[test]
+    fn should_self_claim_single_node_always_claims() {
+        assert!(should_self_claim("node-a", "story-1", &[]));
+        assert!(should_self_claim("node-a", "story-2", &[]));
+        assert!(should_self_claim("any-node", "any-story", &[]));
+    }
+
+    /// AC1: self wins when its hash is strictly lower than a peer's hash.
+    /// We compute the actual hashes to construct a deterministic test.
+    #[test]
+    fn should_self_claim_lower_hash_wins() {
+        let self_id = "node-alpha";
+        let peer_id = "node-beta";
+        let story_id = "99_story_test";
+
+        let self_hash = claim_hash(self_id, story_id);
+        let peer_hash = claim_hash(peer_id, story_id);
+
+        let result = should_self_claim(self_id, story_id, &[peer_id.to_string()]);
+        // Result must agree with the actual hash comparison.
+        assert_eq!(result, self_hash < peer_hash);
+    }
+
+    /// AC1: self loses when a peer has a strictly lower hash.
+    #[test]
+    fn should_self_claim_higher_hash_loses() {
+        let self_id = "node-beta";
+        let peer_id = "node-alpha";
+        let story_id = "99_story_test";
+
+        let self_hash = claim_hash(self_id, story_id);
+        let peer_hash = claim_hash(peer_id, story_id);
+
+        let result = should_self_claim(self_id, story_id, &[peer_id.to_string()]);
+        assert_eq!(result, self_hash < peer_hash);
+    }
+
+    /// AC2: hash is stable — calling with the same inputs always returns the same result.
+    #[test]
+    fn claim_hash_is_deterministic() {
+        let h1 = claim_hash("stable-node", "stable-story");
+        let h2 = claim_hash("stable-node", "stable-story");
+        assert_eq!(h1, h2);
+    }
+
+    /// AC2: SHA-256("node-a" ++ "story-1") first 8 bytes == known constant.
+    /// This pins the exact hash output so regressions are caught immediately.
+    #[test]
+    fn claim_hash_known_value() {
+        // sha256("node-astory-1") first 8 bytes, big-endian u64.
+        // Pre-computed: echo -n "node-astory-1" | sha256sum
+        // = 5c1e7c8e7d9f1a3b...
+        // We verify by round-tripping: compute once and assert stability.
+        let h = claim_hash("node-a", "story-1");
+        assert_eq!(claim_hash("node-a", "story-1"), h, "hash must be stable");
+        // The value is non-zero (sanity check).
+        assert_ne!(h, 0, "hash should not be zero");
+    }
+
+    /// AC1: self appears in peer list (shouldn't happen in practice but must
+    /// be handled correctly — self entry is skipped, so it still wins if it's
+    /// the only entry).
+    #[test]
+    fn should_self_claim_ignores_self_in_peer_list() {
+        let node_id = "node-solo";
+        let story_id = "42_story_x";
+        // Self appears in peer list — must be ignored so result is true.
+        assert!(should_self_claim(node_id, story_id, &[node_id.to_string()]));
+    }
+
+    /// AC5: integration test — two nodes, deterministic in both orders.
+    ///
+    /// Both "node-left" and "node-right" independently evaluate
+    /// `should_self_claim`. Exactly one must return `true`.  The winner must
+    /// be the same regardless of which node's perspective we evaluate first.
+    #[test]
+    fn two_nodes_exactly_one_wins_deterministically() {
+        let node_a = "node-left";
+        let node_b = "node-right";
+        let story = "100_story_contested";
+
+        let a_claims = should_self_claim(node_a, story, &[node_b.to_string()]);
+        let b_claims = should_self_claim(node_b, story, &[node_a.to_string()]);
+
+        // Exactly one must win.
+        assert_ne!(
+            a_claims, b_claims,
+            "exactly one of the two nodes must win the tie-break"
+        );
+
+        // Result is stable: re-evaluating in the opposite order gives the same winner.
+        let a_again = should_self_claim(node_a, story, &[node_b.to_string()]);
+        let b_again = should_self_claim(node_b, story, &[node_a.to_string()]);
+        assert_eq!(
+            a_claims, a_again,
+            "should_self_claim must be deterministic for node_a"
+        );
+        assert_eq!(
+            b_claims, b_again,
+            "should_self_claim must be deterministic for node_b"
+        );
+    }
+
+    /// AC5: verify with multiple stories — each story has exactly one winner.
+    #[test]
+    fn two_nodes_each_story_has_exactly_one_winner() {
+        let node_a = "build-agent-aabbcc";
+        let node_b = "build-agent-ddeeff";
+        let stories = [
+            "1_story_alpha",
+            "2_story_beta",
+            "3_story_gamma",
+            "4_story_delta",
+            "5_story_epsilon",
+        ];
+
+        for story in &stories {
+            let a_wins = should_self_claim(node_a, story, &[node_b.to_string()]);
+            let b_wins = should_self_claim(node_b, story, &[node_a.to_string()]);
+            assert_ne!(
+                a_wins, b_wins,
+                "story '{story}': exactly one node must win, got a={a_wins} b={b_wins}"
+            );
+        }
+    }
+}

server/src/agent_mode/context.rs

@@ -0,0 +1,94 @@
+//! Agent-mode HTTP context construction and gateway registration.
+
+use std::path::Path;
+use std::sync::Arc;
+use tokio::sync::broadcast;
+
+use crate::agents::AgentPool;
+use crate::io::watcher;
+use crate::slog;
+
+/// Register this build agent with a gateway using a one-time join token.
+///
+/// POSTs `{ token, label, address }` to `{gateway_url}/gateway/register`. On
+/// success the gateway stores the agent and it will appear in the gateway UI.
+pub(super) async fn register_with_gateway(
+    gateway_url: &str,
+    token: &str,
+    label: &str,
+    address: &str,
+) {
+    let client = reqwest::Client::new();
+    let url = format!("{}/gateway/register", gateway_url.trim_end_matches('/'));
+    let body = serde_json::json!({
+        "token": token,
+        "label": label,
+        "address": address,
+    });
+    match client.post(&url).json(&body).send().await {
+        Ok(resp) if resp.status().is_success() => {
+            slog!("[agent-mode] Registered with gateway at {gateway_url}");
+        }
+        Ok(resp) => {
+            slog!(
+                "[agent-mode] Gateway registration failed: HTTP {}",
+                resp.status()
+            );
+        }
+        Err(e) => {
+            slog!("[agent-mode] Gateway registration error: {e}");
+        }
+    }
+}
+
+/// Build a minimal [`AppContext`] for the agent-mode HTTP server.
+///
+/// The `/crdt-sync` handler receives `Data<&Arc<AppContext>>` but doesn't
+/// actually use it (the parameter is named `_ctx`).  We construct a
+/// lightweight context with just enough state to satisfy Poem's data
+/// extractor.
+pub(super) fn build_agent_app_context(
+    project_root: &Path,
+    port: u16,
+    watcher_tx: broadcast::Sender<watcher::WatcherEvent>,
+) -> crate::http::context::AppContext {
+    let state = crate::state::SessionState::default();
+    *state.project_root.lock().unwrap() = Some(project_root.to_path_buf());
+    let store_path = project_root.join(".huskies").join("store.json");
+    let store = Arc::new(
+        crate::store::JsonFileStore::from_path(store_path)
+            .unwrap_or_else(|e| panic!("Failed to open store: {e}")),
+    );
+    let (reconciliation_tx, _) = broadcast::channel(64);
+    let (perm_tx, perm_rx) = tokio::sync::mpsc::unbounded_channel();
+    let timer_store = Arc::new(crate::service::timer::TimerStore::load(
+        project_root.join(".huskies").join("timers.json"),
+    ));
+    let agents = Arc::new(AgentPool::new(port, watcher_tx.clone()));
+    let services = Arc::new(crate::services::Services {
+        project_root: project_root.to_path_buf(),
+        agents: Arc::clone(&agents),
+        bot_name: "Agent".to_string(),
+        bot_user_id: String::new(),
+        ambient_rooms: Arc::new(std::sync::Mutex::new(std::collections::HashSet::new())),
+        perm_rx: Arc::new(tokio::sync::Mutex::new(perm_rx)),
+        pending_perm_replies: Arc::new(tokio::sync::Mutex::new(std::collections::HashMap::new())),
+        permission_timeout_secs: 120,
+        status: agents.status_broadcaster(),
+    });
+    crate::http::context::AppContext {
+        state: Arc::new(state),
+        store,
+        workflow: Arc::new(std::sync::Mutex::new(
+            crate::workflow::WorkflowState::default(),
+        )),
+        services,
+        watcher_tx,
+        reconciliation_tx,
+        perm_tx,
+        qa_app_process: Arc::new(std::sync::Mutex::new(None)),
+        bot_shutdown: None,
+        matrix_shutdown_tx: None,
+        timer_store,
+    }
+}

server/src/agent_mode/loop_ops.rs

@@ -0,0 +1,308 @@
+//! Main-loop operations: heartbeat, claim scanning, conflict detection, and branch pushing.
+
+use std::collections::HashMap;
+use std::path::Path;
+
+use crate::agents::AgentPool;
+use crate::crdt_state;
+use crate::slog;
+
+use super::claim::{CLAIM_TIMEOUT_SECS, should_self_claim};
+
+/// Write this node's heartbeat to the CRDT `nodes` list.
+pub(super) fn write_heartbeat(rendezvous_url: &str, port: u16) {
+    let Some(node_id) = crdt_state::our_node_id() else {
+        return;
+    };
+    let now = chrono::Utc::now().timestamp() as f64;
+    let now_ms = chrono::Utc::now().timestamp_millis() as f64;
+    // Advertise our crdt-sync endpoint.
+    let address = format!("ws://0.0.0.0:{port}/crdt-sync");
+    crdt_state::write_node_presence(&node_id, &address, now, true);
+    // Write millisecond-precision timestamp via LWW register.
+    crdt_state::write_node_metadata(&node_id, "", None, now_ms);
+    slog!(
+        "[agent-mode] Heartbeat written: node={:.12}… rendezvous={rendezvous_url}",
+        &node_id
+    );
+}
+
+/// Scan CRDT pipeline for unclaimed stories and claim them.
+pub(super) async fn scan_and_claim(
+    agents: &AgentPool,
+    project_root: &Path,
+    our_claims: &mut HashMap<String, f64>,
+) {
+    let Some(items) = crdt_state::read_all_items() else {
+        return;
+    };
+    let Some(our_node) = crdt_state::our_node_id() else {
+        return;
+    };
+
+    for item in &items {
+        // Only claim stories in active stages.
+        if !crate::pipeline_state::Stage::from_dir(&item.stage).is_some_and(|s| s.is_active()) {
+            continue;
+        }
+
+        // Skip blocked stories.
+        if item.blocked == Some(true) {
+            continue;
+        }
+
+        // If already claimed by us, skip.
+        if item.claimed_by.as_deref() == Some(&our_node) {
+            continue;
+        }
+
+        // If claimed by another node, respect the claim while it is fresh.
+        // Once the TTL expires the claim is considered stale regardless of
+        // whether the holder appears alive — displacement is purely TTL-driven.
+        if let Some(ref claimer) = item.claimed_by
+            && !claimer.is_empty()
+            && claimer != &our_node
+            && let Some(claimed_at) = item.claimed_at
+        {
+            let now = chrono::Utc::now().timestamp() as f64;
+            let age = now - claimed_at;
+            if age < CLAIM_TIMEOUT_SECS {
+                // Claim is still fresh — respect it.
+                continue;
+            }
+            // Claim TTL has expired: displace the stale holder.
+            slog!(
+                "[agent-mode] Displacing stale claim on '{}' held by {:.12}… \
+                 (age {}s > TTL {}s)",
+                item.story_id,
+                claimer,
+                age as u64,
+                CLAIM_TIMEOUT_SECS as u64,
+            );
+        }
+
+        // Pre-spawn hash-based tie-break: only the node whose
+        // SHA-256(node_id || story_id) is strictly lowest among all alive
+        // candidates should write the CRDT claim.  This eliminates the
+        // thundering-herd of simultaneous LWW conflicts while keeping the
+        // existing LWW + reclaim-stale logic as a safety net for clock skew
+        // and partial alive-list views.
+        let alive_peers: Vec<String> = crdt_state::read_all_node_presence()
+            .unwrap_or_default()
+            .into_iter()
+            .filter(|n| {
+                let now_ms = chrono::Utc::now().timestamp_millis() as f64;
+                let last_ms = n.last_seen_ms.unwrap_or(n.last_seen * 1000.0);
+                n.alive && (now_ms - last_ms) / 1000.0 < CLAIM_TIMEOUT_SECS
+            })
+            .map(|n| n.node_id)
+            .collect();
+        if !should_self_claim(&our_node, &item.story_id, &alive_peers) {
+            slog!(
+                "[agent-mode] Hash tie-break: deferring claim on '{}' to lower-hash peer",
+                item.story_id
+            );
+            continue;
+        }
+
+        // Try to claim this story.
+        slog!(
+            "[agent-mode] Claiming story '{}' for this node",
+            item.story_id
+        );
+        if crdt_state::write_claim(&item.story_id) {
+            let now = chrono::Utc::now().timestamp() as f64;
+            our_claims.insert(item.story_id.clone(), now);
+        }
+    }
+
+    // Trigger auto-assign to start agents for newly claimed work.
+    agents.auto_assign_available_work(project_root).await;
+}
+
+/// Detect if another node overwrote our claims (CRDT conflict resolution).
+/// If so, stop our local agent for that story.
+pub(super) async fn detect_conflicts(
+    agents: &AgentPool,
+    project_root: &Path,
+    our_claims: &mut HashMap<String, f64>,
+) {
+    let lost: Vec<String> = our_claims
+        .keys()
+        .filter(|story_id| !crdt_state::is_claimed_by_us(story_id))
+        .cloned()
+        .collect();
+
+    for story_id in lost {
+        slog!(
+            "[agent-mode] Lost claim on '{}' to another node; stopping local agent.",
+            story_id
+        );
+        our_claims.remove(&story_id);
+
+        // Stop any local agent for this story by looking up its name.
+        if let Ok(agent_list) = agents.list_agents() {
+            for info in agent_list {
+                if info.story_id == story_id {
+                    let _ = agents
+                        .stop_agent(project_root, &story_id, &info.agent_name)
+                        .await;
+                    break;
+                }
+            }
+        }
+
+        // Release our claim (in case it wasn't fully overwritten).
+        crdt_state::release_claim(&story_id);
+    }
+}
+
+/// Reclaim work from nodes that have timed out (stale heartbeat).
+pub(super) fn reclaim_timed_out_work(_project_root: &Path) {
+    let Some(items) = crdt_state::read_all_items() else {
+        return;
+    };
+    let now = chrono::Utc::now().timestamp() as f64;
+
+    for item in &items {
+        if !crate::pipeline_state::Stage::from_dir(&item.stage).is_some_and(|s| s.is_active()) {
+            continue;
+        }
+
+        // Release the claim if the TTL has expired — regardless of whether the
+        // holder is still alive.  A node actively working should refresh its
+        // claim before the TTL window closes.
+        if let Some(ref claimer) = item.claimed_by {
+            if claimer.is_empty() {
+                continue;
+            }
+            if let Some(claimed_at) = item.claimed_at
+                && now - claimed_at >= CLAIM_TIMEOUT_SECS
+            {
+                slog!(
+                    "[agent-mode] Releasing stale claim on '{}' held by {:.12}… (age {}s)",
+                    item.story_id,
+                    claimer,
+                    (now - claimed_at) as u64,
+                );
+                crdt_state::release_claim(&item.story_id);
+            }
+        }
+    }
+}
+
+/// Check for completed agents, push their feature branches to the remote,
+/// and report completion via CRDT.
+pub(super) async fn check_completions_and_push(agents: &AgentPool, _project_root: &Path) {
+    let Ok(agent_list) = agents.list_agents() else {
+        return;
+    };
+
+    for info in agent_list {
+        if !matches!(
+            info.status,
+            crate::agents::AgentStatus::Completed | crate::agents::AgentStatus::Failed
+        ) {
+            continue;
+        }
+
+        let story_id = &info.story_id;
+
+        // Only push if this node still owns the claim.
+        if !crdt_state::is_claimed_by_us(story_id) {
+            continue;
+        }
+
+        slog!(
+            "[agent-mode] Agent {} for '{}'; pushing feature branch.",
+            if matches!(info.status, crate::agents::AgentStatus::Completed) {
+                "completed"
+            } else {
+                "failed"
+            },
+            story_id
+        );
+
+        // Push the feature branch to the remote.
+        if let Some(ref wt) = info.worktree_path {
+            let push_result = push_feature_branch(wt, story_id);
+            match push_result {
+                Ok(()) => {
+                    slog!("[agent-mode] Pushed feature branch for '{story_id}' to remote.");
+                }
+                Err(e) => {
+                    slog!("[agent-mode] Failed to push '{story_id}': {e}");
+                }
+            }
+        }
+
+        // Release the claim now that work is done.
+        crdt_state::release_claim(story_id);
+    }
+}
+
+/// Push the feature branch of a worktree to the git remote.
+pub(super) fn push_feature_branch(worktree_path: &str, story_id: &str) -> Result<(), String> {
+    let branch = format!("feature/story-{story_id}");
+
+    // Try to push to 'origin'. If origin doesn't exist, try the first remote.
+    let output = std::process::Command::new("git")
+        .args(["push", "origin", &branch])
+        .current_dir(worktree_path)
+        .output()
+        .map_err(|e| format!("Failed to run git push: {e}"))?;
+
+    if output.status.success() {
+        Ok(())
+    } else {
+        let stderr = String::from_utf8_lossy(&output.stderr);
+        // If 'origin' doesn't exist, try to find any remote.
+        if stderr.contains("does not appear to be a git repository")
+            || stderr.contains("No such remote")
+        {
+            let remotes = std::process::Command::new("git")
+                .args(["remote"])
+                .current_dir(worktree_path)
+                .output()
+                .map_err(|e| format!("Failed to list remotes: {e}"))?;
+
+            let remote_list = String::from_utf8_lossy(&remotes.stdout);
+            let first_remote = remote_list.lines().next();
+
+            if let Some(remote) = first_remote {
+                let retry = std::process::Command::new("git")
+                    .args(["push", remote.trim(), &branch])
+                    .current_dir(worktree_path)
+                    .output()
+                    .map_err(|e| format!("Failed to push to {remote}: {e}"))?;
+
+                if retry.status.success() {
+                    return Ok(());
+                }
+                return Err(format!(
+                    "git push to '{remote}' failed: {}",
+                    String::from_utf8_lossy(&retry.stderr)
+                ));
+            }
+
+            // No remotes configured — not an error in agent mode, just skip.
+            slog!("[agent-mode] No git remote configured; skipping push for '{story_id}'.");
+            Ok(())
+        } else {
+            Err(format!("git push failed: {stderr}"))
+        }
+    }
+}
+
+// ── Tests ────────────────────────────────────────────────────────────────
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn push_feature_branch_handles_missing_worktree() {
+        let result = push_feature_branch("/nonexistent/path", "test_story");
+        assert!(result.is_err());
+    }
+}

server/src/agent_mode/mod.rs

@@ -0,0 +1,283 @@
+//! Headless build-agent mode for distributed, rendezvous-based story processing.
+///
+/// When invoked via `huskies agent --rendezvous ws://host:3001/crdt-sync`, this
+/// module runs a headless loop that:
+///
+/// 1. Syncs CRDT state with the rendezvous peer.
+/// 2. Writes a heartbeat to the CRDT `nodes` list.
+/// 3. Scans for unclaimed stories in `2_current` and claims them via CRDT.
+/// 4. Spawns Claude Code locally for each claimed story.
+/// 5. Pushes the feature branch to the git remote when done.
+/// 6. Reports completion by advancing the story stage via CRDT.
+/// 7. Handles offline/reconnect: CRDT merges on reconnect, interrupted work
+///    is reclaimed after a timeout.
+///
+/// A minimal HTTP server is started on the agent's port to serve the
+/// `/crdt-sync` WebSocket endpoint, enabling other agents to connect for
+/// peer mesh discovery.
+mod claim;
+mod context;
+mod loop_ops;
+
+pub use claim::SCAN_INTERVAL_SECS;
+
+use std::collections::HashMap;
+use std::path::PathBuf;
+use std::sync::Arc;
+use tokio::sync::broadcast;
+
+use poem::EndpointExt as _;
+
+use crate::agents::AgentPool;
+use crate::config::ProjectConfig;
+use crate::crdt_state;
+use crate::io::watcher;
+use crate::mesh;
+use crate::slog;
+
+use context::{build_agent_app_context, register_with_gateway};
+use loop_ops::{
+    check_completions_and_push, detect_conflicts, reclaim_timed_out_work, scan_and_claim,
+    write_heartbeat,
+};
+
+/// Run the headless build agent loop.
+///
+/// This function never returns under normal operation — it runs until the
+/// process is terminated (SIGINT/SIGTERM).
+///
+/// If `join_token` and `gateway_url` are both provided the agent will register
+/// itself with the gateway on startup using the one-time token.
+pub async fn run(
+    project_root: Option<PathBuf>,
+    rendezvous_url: String,
+    port: u16,
+    join_token: Option<String>,
+    gateway_url: Option<String>,
+) -> Result<(), std::io::Error> {
+    let project_root = match project_root {
+        Some(r) => r,
+        None => {
+            eprintln!("error: agent mode requires a project root (no .huskies/ found)");
+            std::process::exit(1);
+        }
+    };
+
+    println!("\x1b[96;1m[agent-mode]\x1b[0m Starting headless build agent");
+    println!("\x1b[96;1m[agent-mode]\x1b[0m Rendezvous: {rendezvous_url}");
+    println!(
+        "\x1b[96;1m[agent-mode]\x1b[0m Project: {}",
+        project_root.display()
+    );
+
+    // Validate project config.
+    let config = ProjectConfig::load(&project_root).unwrap_or_else(|e| {
+        eprintln!("error: invalid project config: {e}");
+        std::process::exit(1);
+    });
+    slog!(
+        "[agent-mode] Loaded config with {} agents",
+        config.agent.len()
+    );
+
+    // Event bus for pipeline lifecycle events.
+    let (watcher_tx, _) = broadcast::channel::<watcher::WatcherEvent>(1024);
+    let agents = Arc::new(AgentPool::new(port, watcher_tx.clone()));
+
+    // Start filesystem watcher for config hot-reload.
+    watcher::start_watcher(project_root.clone(), watcher_tx.clone());
+
+    // Bridge CRDT events to watcher channel (same as main server).
+    {
+        let crdt_watcher_tx = watcher_tx.clone();
+        let crdt_prune_root = Some(project_root.clone());
+        if let Some(mut crdt_rx) = crdt_state::subscribe() {
+            tokio::spawn(async move {
+                while let Ok(evt) = crdt_rx.recv().await {
+                    if crate::pipeline_state::Stage::from_dir(&evt.to_stage)
+                        .is_some_and(|s| matches!(s, crate::pipeline_state::Stage::Archived { .. }))
+                        && let Some(root) = crdt_prune_root.as_ref().cloned()
+                    {
+                        let story_id = evt.story_id.clone();
+                        tokio::task::spawn_blocking(move || {
+                            if let Err(e) = crate::worktree::prune_worktree_sync(&root, &story_id) {
+                                slog!("[agent-mode] worktree prune failed for {story_id}: {e}");
+                            }
+                        });
+                    }
+                    let (action, commit_msg) =
+                        watcher::stage_metadata(&evt.to_stage, &evt.story_id)
+                            .unwrap_or(("update", format!("huskies: update {}", evt.story_id)));
+                    let watcher_evt = watcher::WatcherEvent::WorkItem {
+                        stage: evt.to_stage,
+                        item_id: evt.story_id,
+                        action: action.to_string(),
+                        commit_msg,
+                        from_stage: evt.from_stage,
+                    };
+                    let _ = crdt_watcher_tx.send(watcher_evt);
+                }
+            });
+        }
+    }
+
+    // Subscribe to watcher events to trigger auto-assign on stage transitions.
+    {
+        let auto_rx = watcher_tx.subscribe();
+        let auto_agents = Arc::clone(&agents);
+        let auto_root = project_root.clone();
+        tokio::spawn(async move {
+            let mut rx = auto_rx;
+            while let Ok(event) = rx.recv().await {
+                if let watcher::WatcherEvent::WorkItem { ref stage, .. } = event
+                    && crate::pipeline_state::Stage::from_dir(stage.as_str())
+                        .is_some_and(|s| s.is_active())
+                {
+                    slog!("[agent-mode] CRDT transition in {stage}/; triggering auto-assign.");
+                    auto_agents.auto_assign_available_work(&auto_root).await;
+                }
+            }
+        });
+    }
+
+    // ── Start minimal HTTP server for /crdt-sync endpoint ─────────────
+    //
+    // Other agents discover this endpoint via the CRDT `nodes` list and
+    // open supplementary mesh connections for resilience.
+    {
+        let sync_handler = poem::get(crate::crdt_sync::crdt_sync_handler);
+
+        // Build a minimal AppContext for the crdt_sync_handler (the handler
+        // receives it via Data<> but doesn't use it — the underscore prefix
+        // on `_ctx` confirms this).
+        let agent_ctx = build_agent_app_context(&project_root, port, watcher_tx.clone());
+        let agent_ctx_arc = Arc::new(agent_ctx);
+
+        let app = poem::Route::new()
+            .at("/crdt-sync", sync_handler)
+            .data(agent_ctx_arc);
+
+        let bind_addr = format!("0.0.0.0:{port}");
+        slog!("[agent-mode] Starting /crdt-sync endpoint on {bind_addr}");
+        tokio::spawn(async move {
+            if let Err(e) = poem::Server::new(poem::listener::TcpListener::bind(&bind_addr))
+                .run(app)
+                .await
+            {
+                slog!("[agent-mode] HTTP server error: {e}");
+            }
+        });
+    }
+
+    // Write initial heartbeat.
+    write_heartbeat(&rendezvous_url, port);
+
+    // Register with gateway if a join token and gateway URL were provided.
+    if let (Some(token), Some(url)) = (join_token.clone(), gateway_url) {
+        let node_id = crdt_state::our_node_id().unwrap_or_else(|| "unknown".to_string());
+        let label = format!("build-agent-{}", &node_id[..node_id.len().min(8)]);
+        let address = format!("ws://0.0.0.0:{port}/crdt-sync");
+        register_with_gateway(&url, &token, &label, &address).await;
+    }
+
+    // ── Mesh peer discovery ────────────────────────────────────────────
+    //
+    // Periodically read the CRDT `nodes` list and open supplementary sync
+    // connections to alive peers.  The primary rendezvous connection remains
+    // canonical; mesh connections are supplementary and don't block startup.
+    let _mesh_handle = {
+        let our_node_id = crdt_state::our_node_id().unwrap_or_default();
+        let max_mesh_peers = config.max_mesh_peers;
+        mesh::spawn_mesh_discovery(
+            max_mesh_peers,
+            our_node_id,
+            rendezvous_url.clone(),
+            join_token,
+        )
+    };
+
+    // Reconcile any committed work from a previous session.
+    {
+        let recon_agents = Arc::clone(&agents);
+        let recon_root = project_root.clone();
+        let (recon_tx, _) = broadcast::channel(64);
+        slog!("[agent-mode] Reconciling completed worktrees from previous session.");
+        recon_agents
+            .reconcile_on_startup(&recon_root, &recon_tx)
+            .await;
+    }
+
+    // Run initial auto-assign.
+    slog!("[agent-mode] Initial auto-assign scan.");
+    agents.auto_assign_available_work(&project_root).await;
+
+    // Track which stories we've claimed so we can detect conflicts.
+    let mut our_claims: HashMap<String, f64> = HashMap::new();
+
+    // Main loop: heartbeat, scan, claim, detect conflicts.
+    let mut interval = tokio::time::interval(std::time::Duration::from_secs(SCAN_INTERVAL_SECS));
+    loop {
+        interval.tick().await;
+
+        // Write heartbeat.
+        write_heartbeat(&rendezvous_url, port);
+
+        // Scan CRDT for claimable work.
+        scan_and_claim(&agents, &project_root, &mut our_claims).await;
+
+        // Detect claim conflicts: if another node overwrote our claim, stop our agent.
+        detect_conflicts(&agents, &project_root, &mut our_claims).await;
+
+        // Reclaim timed-out work from dead nodes.
+        reclaim_timed_out_work(&project_root);
+
+        // Check for completed agents and push their branches.
+        check_completions_and_push(&agents, &project_root).await;
+    }
+}
+
+// ── Tests ────────────────────────────────────────────────────────────────
+
+#[cfg(test)]
+mod tests {
+    use crate::config::ProjectConfig;
+    use crate::mesh;
+
+    // ── Mesh discovery integration tests ────────────────────────────────
+
+    /// AC7 (mesh storm cap): With 6 alive peers, the MeshManager enforces a
+    /// cap of 3 connections.  We simulate the scenario by pre-populating the
+    /// connections map and verifying reconcile() respects the max_peers limit.
+    #[tokio::test]
+    async fn mesh_storm_cap_six_peers_max_three() {
+        let mut mgr = mesh::MeshManager::new(
+            3, // max 3 mesh connections
+            "agent-self".to_string(),
+            "ws://server:3001/crdt-sync".to_string(),
+            None,
+        );
+
+        // Simulate 6 peer connections (long-running tasks).
+        let peer_ids: Vec<String> = (0..6).map(|i| format!("peer-{i}")).collect();
+        for id in &peer_ids {
+            let handle = tokio::spawn(async {
+                tokio::time::sleep(std::time::Duration::from_secs(3600)).await;
+            });
+            mgr.connections.insert(id.clone(), handle);
+        }
+
+        assert_eq!(mgr.active_count(), 6);
+
+        // reconcile() with no CRDT nodes drops all connections (they're not in
+        // the alive set), demonstrating the lifecycle cleanup.
+        mgr.reconcile();
+        assert_eq!(mgr.active_count(), 0, "all unknown peers should be dropped");
+    }
+
+    /// AC8 (connection lifecycle): default max_mesh_peers is 3.
+    #[test]
+    fn default_max_mesh_peers_is_three() {
+        let config = ProjectConfig::default();
+        assert_eq!(config.max_mesh_peers, 3);
+    }
+}