diff --git a/skills/council/SKILL.md b/skills/council/SKILL.md
index 48edbc0..3034272 100644
--- a/skills/council/SKILL.md
+++ b/skills/council/SKILL.md
@@ -120,13 +120,37 @@ Always multi-round (minimum rounds=2, default rounds=3 for this flow):
 
 ## Model Selection
 
-Pick model tier based on topic complexity:
+Pick model tier based on topic complexity and role mission:
 
-- **light** (casual brainstorm, simple pros/cons): default model for advisors and referee.
-- **medium** (architecture decisions, strategy): default model for advisors, stronger model for referee.
-- **heavy** (critical decisions, deep analysis): stronger model for all agents.
+- **light** (casual brainstorm, simple pros/cons)
+  - advisors / freethinkers / arbiters: use dedicated council role agents on `litellm/gpt-5-mini`
+  - referee / meta-arbiter: use dedicated synthesis agents on `openai-codex/gpt-5.4` (with configured fallbacks)
+- **medium** (architecture decisions, strategy)
+  - advisors / freethinkers / arbiters: still use dedicated council role agents on `litellm/gpt-5-mini`
+  - referee / meta-arbiter: use dedicated synthesis agents on `openai-codex/gpt-5.4`
+- **heavy** (critical decisions, deep analysis)
+  - keep dedicated council role agents by profile
+  - prefer stronger synthesis on `openai-codex/gpt-5.4`
+  - if a specific council run is unusually high-stakes, the caller may override the spawned session model upward, but should preserve the dedicated council agent identity/role
 
-The caller (main agent) determines tier before spawning.
+### Dedicated council agent roster
+
+Personality mode:
+- `council-pragmatist`
+- `council-visionary`
+- `council-skeptic`
+- `council-referee`
+
+D/P mode:
+- `council-d-freethinker`
+- `council-d-arbiter`
+- `council-p-freethinker`
+- `council-p-arbiter`
+- `council-meta-arbiter`
+
+### Spawn rule
+
+When invoking the council skill, explicitly target these dedicated agent ids via `sessions_spawn` instead of generic/default subagents. The caller (main agent) determines tier before spawning, but should preserve the role-specific council agent identity and use model overrides only when the mission/tier truly justifies it.
 
 ## Round-Specific Prompt Guidance
 
diff --git a/skills/council/scripts/council.sh b/skills/council/scripts/council.sh
index 8546a6e..609da4d 100644
--- a/skills/council/scripts/council.sh
+++ b/skills/council/scripts/council.sh
@@ -117,14 +117,48 @@
 # ─── SUBAGENT LABELING ────────────────────────────────────────────
 #
 # Personality mode:
-#   Labels: council-r{round}-{role}
-#   Examples: council-r1-pragmatist, council-r2-skeptic, council-referee
-#   Single-round: council-pragmatist, council-referee (no round prefix)
+#   Agent ids:
+#     council-pragmatist
+#     council-visionary
+#     council-skeptic
+#     council-referee
+#   Labels can still be council-r{round}-{role} for bookkeeping, but sessions should target
+#   the dedicated council agent ids above.
 #
 # D/P mode:
-#   Labels: council-r{round}-{group}-{role}
-#   Examples: council-r1-d-freethinker, council-r1-p-arbiter, council-meta
-#   Single-round: council-d-freethinker, council-p-arbiter, council-meta
+#   Agent ids:
+#     council-d-freethinker
+#     council-d-arbiter
+#     council-p-freethinker
+#     council-p-arbiter
+#     council-meta-arbiter
+#   Labels can still be council-r{round}-{group}-{role} for bookkeeping, but sessions should target
+#   the dedicated council agent ids above.
+#
+# ─── ROLE-AWARE MODEL TIERING ─────────────────────────────────────
+#
+# light:
+#   advisors/freethinkers/arbiters -> dedicated council role agents on litellm/gpt-5-mini
+#   referee/meta-arbiter           -> dedicated synthesis agents on openai-codex/gpt-5.4
+#
+# medium:
+#   advisors/freethinkers/arbiters -> dedicated council role agents on litellm/gpt-5-mini
+#   referee/meta-arbiter           -> dedicated synthesis agents on openai-codex/gpt-5.4
+#
+# heavy:
+#   preserve dedicated role agent ids; prefer strongest synthesis on referee/meta-arbiter.
+#   Only override advisor models upward when the task is genuinely high-stakes or unusually complex.
+#
+# Principle: mission-specific synthesis gets the stronger model first; broad advisor diversity stays cheap
+# unless the topic quality/risk clearly justifies escalation.
+#
+# Example sessions_spawn targeting:
+#   sessions_spawn(... agentId="council-pragmatist" ...)
+#   sessions_spawn(... agentId="council-referee" ...)
+#   sessions_spawn(... agentId="council-meta-arbiter" ...)
+#
+# Single-round labels may still look like council-pragmatist / council-meta for readability,
+# but the runtime target should be the dedicated council agent id.
 #
 # ─── WORD COUNT GUIDANCE ──────────────────────────────────────────
 #