@gobing-ai/spur 0.3.48 → 0.3.50

package/.claude-plugin/marketplace.json

@@ -6,7 +6,7 @@
     },
     "plugins": [
         {
-            "version": "0.3.48",
+            "version": "0.3.50",
             "name": "sp",
             "source": "./plugins/sp",
             "description": "Spur - my harness toolkits"

package/config/config.example.yaml

@@ -49,6 +49,28 @@ agent:
   # naming both accepted sets.
   default: coder
 
+  # Role → tier map (SSOT: DEFAULT_AGENT_ROLES in packages/config, task 0572 /
+  # ADR-061; plugins/sp/references/roles.md is a parity-gated projection).
+  # Resolution: role → tier → cheapest USABLE executor whose tier >= the role's
+  # tier, climbing the ladder across gaps. Declare at least one usable executor
+  # at or above each rung or that role fails loudly at dispatch:
+  #   scribe=cheap · coder=standard · reviewer=capable-1 · planner=capable-2
+  # Stage floors (stage-registry model_policy): changelog=cheap;
+  # implement/test/wrap/review/refine/brainstorm=standard;
+  # verify/dogfood=capable-1; plan=capable-2. A stage's fallback ladder
+  # (auth / resource-exhaustion / gate-fail) only advances when a higher rung
+  # has a usable executor — an empty rung turns a quota failure into a hang.
+  #
+  # Optional per-project override (0572): re-tier / re-stage a KNOWN role only —
+  # the vocabulary is closed; a config naming any other key fails at config load
+  # naming the accepted four. Per-field merge over the defaults (restate only
+  # what you override; a role you omit keeps its default wholesale):
+  # roles:
+  #   reviewer:
+  #     tier: capable-2        # e.g. pin review work one rung higher
+  #   coder:
+  #     stages: [implement]    # e.g. stop folding test/wrap into coder
+
   # Canonical coding-agent ids (ts-ai-runner DISPLAY_ORDER, 0.4.8+):
   #   claude, codex, gemini, pi, omp, opencode, antigravity-cli,
   #   openclaw, hermes, grok
@@ -56,32 +78,58 @@ agent:
   # auth-status verb).
 
   # Named executor profiles (ADR-033 / 0343). Each pairs a canonical agent with
-  # an optional opaque model override and an explicit capability tier.
+  # an explicit model override and an explicit capability tier.
+  # `model` is pinned on every entry: omitting it defers to the agent CLI's own
+  # default, which drifts with CLI upgrades and account state — pin what your
+  # CLI would resolve today (`grok models`, omp config.yml modelRoles.default,
+  # ~/.codex/config.toml, ~/.claude.json). The ids below are working examples,
+  # not vendor defaults — re-pin for your accounts.
+  #
+  # Pairing philosophy (operator guidance, 2026-08): a `native` pair — a vendor's
+  # model served by that vendor's own CLI (gpt-5.6-* via codex, claude-*-5 via
+  # claude, grok-4.6 via grok, gemini-3.x via agy) — consistently outperforms the
+  # same model routed through a third-party CLI, and belongs at the capable rungs.
+  # `portable` models (glm-5.x, deepseek-v4-*, …) are provider-agnostic and best
+  # carried by omp or pi at the cheap/standard rungs. The ladder below mixes both
+  # classes on purpose; measure pairings with the history plane before promoting
+  # a portable model into a capable rung.
   # Live tiers: cheap | standard | capable-1 | capable-2 | capable-3
   # (1=low quality within capable, 3=high). DECLARE tier — inference never
   # invents capable-2/3. Stage routing starts on the cheapest executor whose
   # tier >= min_tier; same exact tier ties break by array order.
   # Legacy bare `capable` is accepted at runtime as synonym for capable-1.
   executors:
+    # cheap rung is optional: with none declared, scribe-role work starts one
+    # rung up at the cheapest standard executor. Add one when transcript-heavy
+    # mechanical commands (changelog/gitmsg/handover) get noisy on cost.
+    # - name: minimax
+    #   agent: omp
+    #   model: minimax/MiniMax-M3
+    #   tier: cheap
     - name: omp
       agent: omp
+      model: opencode/deepseek-v4-flash
       tier: standard
     # - name: omp-zai
     #   agent: omp
     #   model: zai/glm-5.2
     #   tier: standard
+    - name: pi
+      agent: pi
+      model: "k3"
+      tier: capable-1
     - name: grok
       agent: grok
+      model: grok-4.6
       tier: capable-2
     - name: claude
       agent: claude
+      model: k3
       tier: capable-3
     - name: codex
       agent: codex
+      model: gpt-5.6
       tier: capable-3
-    # - name: pi
-    #   agent: pi
-    #   tier: capable-1
 
   # `default-by-phase` is a REMOVED (0452). Prefer
   # executor `tier` + stage-registry model_policy. When present, a matching