From 8588e7e7aa7c912f6080892a33f56cbfec8c0578 Mon Sep 17 00:00:00 2001
From: Santhi Prakash <b.santhiprakash@gmail.com>
Date: Tue, 4 Aug 2026 16:44:52 +0000
Subject: [PATCH] fix(acp): honor persona `thread_replies` in prompt builder

`thread_replies` was parsed and merged in `buzz-persona` but never read
in `buzz-acp`. The harness always injected an `IMPORTANT:` directive
forcing `--reply-to <event_id>` for new top-level channel mentions,
so a persona with `thread_replies: false` still opened a thread.

Plumb the resolved flag through `Config` / `PromptContext` /
`FormatPromptArgs` with a default of `true` (per PERSONA_PACK_SPEC.md).
When `thread_replies` is `false`, `format_prompt` skips the forced
reply anchor for new top-level channel mentions. Existing thread and
DM anchors continue to work, so the agent can still participate in an
active thread or DM.

- Add `--thread-replies` / `BUZZ_ACP_THREAD_REPLIES` to `CliArgs` and
  `Config`.
- Thread the value through `PromptContext` and `FormatPromptArgs`.
- Skip `resolve_reply_anchor` for human-facing top-level channel turns
  when `thread_replies == false`.
- Add `queue.rs` unit tests covering top-level suppression, existing
  threads, and DMs when `thread_replies: false`.

Fixes #4699.

Signed-off-by: Santhi Prakash <b.santhiprakash@gmail.com>
---
 crates/buzz-acp/src/config.rs |  16 ++++
 crates/buzz-acp/src/lib.rs    |   3 +
 crates/buzz-acp/src/pool.rs   |   8 ++
 crates/buzz-acp/src/queue.rs  | 147 +++++++++++++++++++++++++++++++++-
 4 files changed, 172 insertions(+), 2 deletions(-)

diff --git a/crates/buzz-acp/src/config.rs b/crates/buzz-acp/src/config.rs
index 35aaec188d..6f91e9672b 100644
--- a/crates/buzz-acp/src/config.rs
+++ b/crates/buzz-acp/src/config.rs
@@ -474,6 +474,15 @@ pub struct CliArgs {
     #[arg(long, env = "BUZZ_ACP_RELAY_OBSERVER", default_value_t = false)]
     pub relay_observer: bool,
 
+    /// Whether the agent should default to replying in the current thread.
+    ///
+    /// Mirrors the `thread_replies` behavioral config from persona packs.
+    /// When false, the harness does NOT append `--reply-to` instructions for
+    /// new top-level channel mentions, letting the agent post flat at channel root.
+    /// Existing thread and DM reply anchors are unaffected.
+    #[arg(long, env = "BUZZ_ACP_THREAD_REPLIES", default_value_t = true)]
+    pub thread_replies: bool,
+
     /// Exit after this many seconds with no dispatched events and no turn in flight.
     /// 0 disables inactivity self-termination.
     #[arg(long, env = "BUZZ_ACP_EXIT_AFTER_INACTIVITY", default_value_t = 0)]
@@ -544,6 +553,11 @@ pub struct Config {
     pub respond_to_allowlist: HashSet<String>,
     /// Allowed `respond_to` modes. Empty = all modes allowed.
     pub allowed_respond_to: Vec<String>,
+    /// Whether the agent should default to replying in the current thread.
+    ///
+    /// Mirrors the `thread_replies` behavioral config from persona packs.
+    /// Defaults to `true` per `PERSONA_PACK_SPEC.md`.
+    pub thread_replies: bool,
     /// Per-persona env vars to inject at agent spawn time (e.g., GOOSE_PROVIDER, GOOSE_MODEL, BUZZ_AGENT_MODEL).
     /// Populated from persona pack resolution. Empty when no pack is configured.
     pub persona_env_vars: Vec<(String, String)>,
@@ -1102,6 +1116,7 @@ impl Config {
             respond_to: args.respond_to,
             respond_to_allowlist,
             allowed_respond_to,
+            thread_replies: args.thread_replies,
             persona_env_vars,
             has_generated_codex_config,
             relay_observer: args.relay_observer,
@@ -1473,6 +1488,7 @@ mod tests {
             respond_to: RespondTo::Anyone,
             respond_to_allowlist: HashSet::new(),
             allowed_respond_to: Vec::new(),
+            thread_replies: true,
             persona_env_vars: vec![],
             has_generated_codex_config: false,
             relay_observer: false,
diff --git a/crates/buzz-acp/src/lib.rs b/crates/buzz-acp/src/lib.rs
index 811253e4ac..0857697e9b 100644
--- a/crates/buzz-acp/src/lib.rs
+++ b/crates/buzz-acp/src/lib.rs
@@ -1610,6 +1610,7 @@ async fn tokio_main() -> Result<()> {
             .as_deref()
             .and_then(|hex| nostr::PublicKey::from_hex(hex).ok()),
         memory_enabled: config.memory_enabled,
+        thread_replies: config.thread_replies,
         harness_name: crate::config::normalize_agent_command_identity(&config.agent_command),
         relay_url: config.relay_url.clone(),
     });
@@ -5129,6 +5130,7 @@ mod build_mcp_servers_tests {
             respond_to: config::RespondTo::Anyone,
             respond_to_allowlist: std::collections::HashSet::new(),
             allowed_respond_to: vec![],
+            thread_replies: true,
             persona_env_vars: vec![],
             has_generated_codex_config: false,
             relay_observer: false,
@@ -5351,6 +5353,7 @@ mod error_outcome_emission_tests {
             respond_to: config::RespondTo::Anyone,
             respond_to_allowlist: HashSet::new(),
             allowed_respond_to: vec![],
+            thread_replies: true,
             persona_env_vars: vec![],
             has_generated_codex_config: false,
             relay_observer: false,
diff --git a/crates/buzz-acp/src/pool.rs b/crates/buzz-acp/src/pool.rs
index 64edf68ee2..ee01b047ae 100644
--- a/crates/buzz-acp/src/pool.rs
+++ b/crates/buzz-acp/src/pool.rs
@@ -557,6 +557,12 @@ pub struct PromptContext {
     /// `[Agent Memory — core]` section. On by default; disabled via
     /// `--no-memory` / `BUZZ_ACP_NO_MEMORY`.
     pub memory_enabled: bool,
+    /// Whether the agent should default to replying in the current thread.
+    ///
+    /// Mirrors the `thread_replies` behavioral config from persona packs.
+    /// When false, `format_prompt` skips forced `--reply-to` anchors for new
+    /// top-level channel mentions. Existing threads and DMs are unaffected.
+    pub thread_replies: bool,
     /// Harness identity string for NIP-AM `harness` field. Derived from the
     /// configured `agent_command` at startup (e.g. `"goose"`, `"buzz-agent"`).
     pub harness_name: String,
@@ -1875,6 +1881,7 @@ pub async fn run_prompt_task(
                 system_prompt: ctx.system_prompt.as_deref(),
                 team_instructions: ctx.team_instructions.as_deref(),
                 agent_canvas: agent_canvas.as_deref(),
+                thread_replies: ctx.thread_replies,
             },
         )
     } else {
@@ -6455,6 +6462,7 @@ mod tests {
             agent_keys: agent_keys.clone(),
             agent_owner_pubkey: owner_pubkey,
             memory_enabled: false,
+            thread_replies: true,
             harness_name: "goose".to_string(),
             relay_url: "ws://127.0.0.1:3000".to_string(),
         }
diff --git a/crates/buzz-acp/src/queue.rs b/crates/buzz-acp/src/queue.rs
index 5c960de202..805c3186d5 100644
--- a/crates/buzz-acp/src/queue.rs
+++ b/crates/buzz-acp/src/queue.rs
@@ -1354,7 +1354,6 @@ fn format_conversation_context(
 }
 
 /// Arguments for [`format_prompt`] beyond the required [`FlushBatch`].
-#[derive(Default)]
 pub struct FormatPromptArgs<'a> {
     pub agent_core: Option<&'a str>,
     pub channel_info: Option<&'a PromptChannelInfo>,
@@ -1377,6 +1376,31 @@ pub struct FormatPromptArgs<'a> {
     /// For legacy agents it rides in the user message on every turn of the
     /// session, alongside `[Base]`/`[System]`/`[Agent Memory — core]`.
     pub agent_canvas: Option<&'a str>,
+    /// Whether the agent should default to replying in the current thread.
+    ///
+    /// Mirrors the `thread_replies` behavioral config from persona packs.
+    /// When false, the harness does NOT append a forced `--reply-to` anchor
+    /// for new top-level channel mentions, leaving the agent free to post flat
+    /// at the channel root. Existing thread and DM anchors are unaffected.
+    /// Defaults to `true` to preserve existing behavior.
+    pub thread_replies: bool,
+}
+
+impl<'a> Default for FormatPromptArgs<'a> {
+    fn default() -> Self {
+        Self {
+            agent_core: None,
+            channel_info: None,
+            conversation_context: None,
+            profile_lookup: None,
+            has_system_prompt_support: false,
+            base_prompt: None,
+            system_prompt: None,
+            team_instructions: None,
+            agent_canvas: None,
+            thread_replies: true,
+        }
+    }
 }
 
 /// Format the `[Base]` section for the base prompt.
@@ -1469,8 +1493,14 @@ pub fn format_prompt(batch: &FlushBatch, args: &FormatPromptArgs<'_>) -> Vec<Str
     //   - top-level     → anchor to the triggering event (it becomes the root)
     // Agent↔agent turns get no forced anchor — deep nesting is intentional
     // there. DMs are always 1:1 with a human, so they always anchor.
+    //
+    // Persona-level `thread_replies: false` opts out of forced anchors only for
+    // new top-level channel mentions. Existing thread and DM anchors continue to
+    // work so the agent can still participate in an active thread or DM.
     let sender_pubkey = last_event.event.pubkey.to_hex();
-    let reply_anchor = if is_dm {
+    let reply_anchor = if !args.thread_replies && !is_dm && thread_tags.root_event_id.is_none() {
+        None
+    } else if is_dm {
         thread_tags
             .root_event_id
             .is_some()
@@ -3976,6 +4006,119 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_reply_instruction_absent_for_top_level_when_thread_replies_false() {
+        let ch = Uuid::new_v4();
+        let event = make_event("hello world");
+        let event_id = event.id.to_hex();
+        let batch = FlushBatch {
+            channel_id: ch,
+            events: vec![BatchEvent {
+                event,
+                prompt_tag: "test".into(),
+                received_at: Instant::now(),
+            }],
+            cancelled_events: vec![],
+            cancel_reason: None,
+        };
+
+        // Top-level channel mention with thread_replies=false: the harness must
+        // NOT force a new thread anchor, so the agent can post flat at the
+        // channel root.
+        let prompt = format_prompt(
+            &batch,
+            &FormatPromptArgs {
+                thread_replies: false,
+                ..Default::default()
+            },
+        )
+        .join("\n\n");
+        assert!(
+            !prompt.contains(&format!("--reply-to {event_id}")),
+            "thread_replies=false must suppress the new top-level reply anchor"
+        );
+        assert!(
+            !prompt.contains("new top-level message"),
+            "thread_replies=false must not emit the new-thread instruction"
+        );
+    }
+
+    #[test]
+    fn test_reply_instruction_present_for_existing_thread_when_thread_replies_false() {
+        let ch = Uuid::new_v4();
+        let root_id = "a".repeat(64);
+        let event = make_event_with_tags(
+            "@bot help",
+            vec![vec!["e".into(), root_id.clone(), "".into(), "reply".into()]],
+        );
+        let batch = FlushBatch {
+            channel_id: ch,
+            events: vec![BatchEvent {
+                event,
+                prompt_tag: "@mention".into(),
+                received_at: Instant::now(),
+            }],
+            cancelled_events: vec![],
+            cancel_reason: None,
+        };
+
+        // thread_replies=false only opts out of *new* top-level threads. The
+        // agent should still anchor to the root when replying inside an
+        // existing thread.
+        let prompt = format_prompt(
+            &batch,
+            &FormatPromptArgs {
+                thread_replies: false,
+                ..Default::default()
+            },
+        )
+        .join("\n\n");
+        assert!(
+            prompt.contains(&format!("--reply-to {root_id}")),
+            "existing thread reply should still anchor to root when thread_replies=false"
+        );
+    }
+
+    #[test]
+    fn test_reply_instruction_present_for_dm_thread_reply_when_thread_replies_false() {
+        let ch = Uuid::new_v4();
+        let root_id = "b".repeat(64);
+        let event = make_event_with_tags(
+            "thanks",
+            vec![vec!["e".into(), root_id.clone(), "".into(), "reply".into()]],
+        );
+        let event_id = event.id.to_hex();
+        let batch = FlushBatch {
+            channel_id: ch,
+            events: vec![BatchEvent {
+                event,
+                prompt_tag: "@mention".into(),
+                received_at: Instant::now(),
+            }],
+            cancelled_events: vec![],
+            cancel_reason: None,
+        };
+        let ci = PromptChannelInfo {
+            name: "DM".into(),
+            channel_type: "dm".into(),
+        };
+
+        // DMs are always 1:1 and not governed by thread_replies.
+        let prompt = format_prompt(
+            &batch,
+            &FormatPromptArgs {
+                channel_info: Some(&ci),
+                thread_replies: false,
+                ..Default::default()
+            },
+        )
+        .join("\n\n");
+        assert!(
+            prompt.contains(&format!("--reply-to {event_id}")),
+            "DM thread reply should keep its reply anchor when thread_replies=false"
+        );
+    }
+
     #[test]
     fn test_reply_instruction_absent_for_dm_non_reply() {
         let ch = Uuid::new_v4();