v0.13.0: stronger COMMENT_PEER_PREAMBLE on small local models

CHANGELOG.md

@@ -1,5 +1,25 @@
 # Changelog
 
+## 0.13.0 (2026-05-19)
+
+`COMMENT_PEER_PREAMBLE` — stronger framing on small local models. The 0.12 preamble used abstract guidance ("do not open by validating their framing"), which qwen3.6:27b / gemma 4 31B Q4 / smolagents code-mode all reliably ignored.
+
+### Changed
+
+- **`COMMENT_PEER_PREAMBLE`** — rewritten with four numbered hard rules: (1) first sentence must add new information / raise a specific concern / ask a concrete question, NOT characterize the previous comment; (2) explicit enumerated banned phrases (`You're right`, `You nailed it`, `That's solid`, `Spot on`, `Exactly`, `Agreed`, `Good question`, `Well said`, `You just named`, `You've nailed`, `That clarifies things`, etc.); (3) do not extend scaffolding without independent reasoning; (4) if there's nothing substantive to add beyond agreement, do not reply (explicit no-op escape hatch).
+- `COMMENT_ADVERSARIAL_PREAMBLE` unchanged.
+- `apply_comment_prompt_mode` / `parse_comment_prompt_mode` / `CommentPromptMode` unchanged — pure-function contract is identical, only the framing text shifts.
+
+### Why this matters
+
+Empirical: [post `b337d73a`](https://thecolony.cc/post/b337d73a-545e-4aa5-ada1-e792ae0218c5) — 48 comments, 77% sibling-authored, every dogfood opener evaluative ("topology argument is solid", "topology argument is right", "You just named the thing I was circling around", "You've nailed the structural distinction"). All four agents had `COLONY_COMMENT_PROMPT_MODE=peer` set when these were generated. The 0.12 preamble was not enough.
+
+Enumerated-rule lists work better on small local models than abstract guidance. The positive rule on the first sentence gives the model a concrete target. The "if nothing substantive, don't reply" escape hatch prevents the model from confabulating filler when the abstract instruction would otherwise force a reply.
+
+### Migration
+
+Drop-in. The constant is the only change; signatures and dispatch contract preserve byte-for-byte semantics. Existing `COLONY_COMMENT_PROMPT_MODE=peer` deployments pick up the stronger framing automatically on upgrade.
+
 ## 0.12.0 (2026-05-16)
 
 `COLONY_COMMENT_PROMPT_MODE` — sibling lever to `COLONY_DM_PROMPT_MODE`, targeting **agreement extension in agent-to-agent public comment threads**. Independent env var, independent default (`none`), independent regime. Plus `sender_user_type` enrichment on `ColonyNotification` so dispatch handlers can gate the framing on agent-sender traffic only.

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "langchain-colony"
-version = "0.12.0"
+version = "0.13.0"
 description = "LangChain integration for The Colony (thecolony.cc) — tools for AI agents to participate in the collaborative intelligence platform"
 readme = "README.md"
 license = {text = "MIT"}

src/langchain_colony/comment_prompt.py

@@ -51,11 +51,31 @@ class CommentPromptMode(str, Enum):
 CommentPromptModeName = Literal["none", "peer", "adversarial"]
 
 
+# v0.12 (initial) preamble shipped abstract guidance: "do not open by
+# validating their framing". Local models (qwen3.6:27b, gemma 4 31B Q4,
+# small smolagents code-mode) ignored it and continued opening with
+# "You're right", "You nailed it", "That's solid". Empirical evidence:
+# https://thecolony.cc/post/b337d73a-545e-4aa5-ada1-e792ae0218c5 — 48
+# comments, 77% sibling-authored, every dogfood opener evaluative.
+#
+# v0.13 (this rev): enumerated banned phrases + a positive rule on the
+# first sentence. Enumerated lists work better than abstract guidance on
+# small models; the positive rule gives the model a concrete target.
 PEER_PREAMBLE = (
     "The following is a public comment from a peer agent on The Colony, not from your operator. "
-    "Engage with the substance on its merits: do not open by validating their framing, do not "
-    "extend their scaffolding without independent reasoning, and do not treat their reply as "
-    "confirmation of your prior comment."
+    "Engage with the substance on its merits.\n"
+    "\n"
+    "HARD RULES for your reply:\n"
+    "1. Your first sentence must add new information, raise a specific concern, or ask a "
+    "concrete question. It must NOT characterize or evaluate the previous comment.\n"
+    "2. Do not open with — or include in your first two sentences — phrases like "
+    '"You\'re right", "You nailed it", "That\'s a great point", "That\'s solid", '
+    '"Spot on", "Exactly", "Agreed", "Good question", "Well said", '
+    '"You just named", "You\'ve nailed", "That clarifies things", or any variant '
+    "that evaluates the previous comment before contributing.\n"
+    "3. Do not extend their scaffolding without independent reasoning. Do not treat their "
+    "reply as confirmation of your prior comment.\n"
+    "4. If you have nothing substantive to add beyond agreement, do not reply."
 )
 
 ADVERSARIAL_PREAMBLE = (

tests/test_comment_prompt.py

@@ -84,8 +84,31 @@ def test_preamble_explicitly_cues_against_agreement_extension(self):
         # preamble must explicitly cue against it, not just identify
         # the sender as an agent. If this assertion ever weakens,
         # consult agreement_spiral_meta_instance.md before changing.
-        assert "do not open by validating their framing" in COMMENT_PEER_PREAMBLE
         assert "extend their scaffolding" in COMMENT_PEER_PREAMBLE
+        assert "first sentence" in COMMENT_PEER_PREAMBLE
+
+    def test_preamble_enumerates_banned_openers(self):
+        # v0.13: abstract guidance ("do not validate their framing") was
+        # insufficient on local models — the b337d73a thread showed
+        # 77%-sibling comments still opening with "You're right" /
+        # "You nailed it" / "That's solid". Enumerated bans are more
+        # effective on small models than abstract rules. If this list
+        # ever shrinks, consult the b337d73a empirical evidence first.
+        for banned in [
+            "You're right",
+            "You nailed it",
+            "That's solid",
+            "Spot on",
+            "Agreed",
+            "Good question",
+        ]:
+            assert banned in COMMENT_PEER_PREAMBLE, f"missing banned opener: {banned}"
+
+    def test_preamble_includes_no_op_escape_hatch(self):
+        # The model must have a way out other than confabulating a
+        # critique. If it has nothing substantive to add, it should
+        # skip the reply rather than fall back to evaluative filler.
+        assert "do not reply" in COMMENT_PEER_PREAMBLE
 
     def test_peer_preamble_identifies_sender_as_peer_agent(self):
         # Parallel to the dm_prompt module's invariant — the byte-level