@bacnh85/pi-subagent 0.10.1 → 0.12.2

package/CHANGELOG.md

@@ -1,5 +1,25 @@
 # Changelog
 
+## 0.12.0 (2026-07-24)
+
+### Model routing
+
+- Chains are now **free-first** to conserve the metered opencode-go budget: paid DeepSeek moves to the **last** position so free nvidia/openrouter fallbacks are tried first. Previously DeepSeek sat at position 2 and burned quota whenever the free GLM primary rate-limited.
+- **Removed two dead fallbacks** found by live-testing every free model on 2026-07-24: `nvidia/moonshotai/kimi-k2.6` (HTTP 404, NVCF not provisioned for the account) and `nvidia/z-ai/glm-5.2` (timeout). These are non-rate-limit failures, so the rate-limit retry loop did not advance past them — a subagent that reached them died instead of falling through to the working `:free` entries behind them.
+- New chains: scout/tester `glm-5-turbo` → `nvidia/gpt-oss-20b` → `deepseek-v4-flash`; worker/general-purpose `glm-5.1` → `nvidia/mistral-small-4-119b-2603` → `nemotron-3-super:free` → `deepseek-v4-flash`; planner/reviewer `glm-5.2` → `nemotron-3-ultra:free` → `deepseek-v4-pro`.
+- Verified-working free additions: `nvidia/openai/gpt-oss-20b` (1.2s), `nvidia/mistralai/mistral-small-4-119b-2603` (2.6s, 119B reasoning), `openrouter/nvidia/nemotron-3-nano-30b-a3b:free` (1.2s).
+
+## 0.11.0 (2026-07-23)
+
+### Model routing
+
+- Bundled roles now route through **zai-coding-cn** (GLM) as the primary provider, with a provider-diverse fallback chain: `zai-coding-cn` (free GLM) → `opencode-go` (cheap `deepseek-v4-flash`) → `nvidia` (free NIM) → `openrouter` (`:free` models, last resort).
+- Fast tier (scout, tester) uses `glm-5-turbo` with `thinking: off` (GLM reasoning is ~11× slower; these are mechanical roles).
+- Strong-coding tier (worker, general-purpose) uses `glm-5.1`; deep-reasoning tier (planner, reviewer) uses `glm-5.2` (the only GLM with reasoning-effort control).
+- opencode-go contributes one DeepSeek model per role, matched to strength: `deepseek-v4-flash` for scout/tester/worker/general-purpose, `deepseek-v4-pro` for planner/reviewer — never its GLM models, which cost ~$1.40/$4.40 per M versus zai-coding-cn's free GLM.
+- Chains are cost-ascending on failure and spread load 2/2/2 across the GLM tiers to respect GLM's low concurrency; the existing rate-limit retry walks the chain on 429s.
+- Free-model fallbacks verified live against the OpenRouter API: `nemotron-3-super:free` (worker/general-purpose) and `nemotron-3-ultra:free` (planner/reviewer) respond correctly with reasoning on. scout/tester omit a `:free` entry because their `thinking: off` conflicts with reasoning-mandatory `:free` models (`gpt-oss-20b:free` returns HTTP 400 when reasoning is disabled; `gemma-4-31b:free` rate-limits, `cohere/north-mini-code:free` returns empty).
+
 ## 0.9.2 (2026-07-16)
 
 ### Pi SDK compatibility

package/README.md

@@ -16,14 +16,14 @@ Requires Node.js >= 20.18.
 
 | Role | Ordered model preferences | Thinking | Tools |
 | --- | --- | --- | --- |
-| `scout` | `opencode-go/deepseek-v4-flash` → `openai-codex/gpt-5.6-luna` → `opencode-go/mimo-v2.5` | low | read, grep, find, ls |
-| `tester` | `openai-codex/gpt-5.6-luna` → `opencode-go/mimo-v2.5` → `opencode-go/deepseek-v4-flash` | low | read, bash, grep, find, ls |
-| `worker` | `openai-codex/gpt-5.6-terra` → `opencode-go/deepseek-v4-pro` | medium | read, bash, edit, write, grep, find, ls |
-| `general-purpose` | `openai-codex/gpt-5.6-terra` → `opencode-go/deepseek-v4-pro` | medium | read, bash, edit, write, grep, find, ls |
-| `planner` | `openai-codex/gpt-5.6-sol` → `opencode-go/deepseek-v4-pro` | high | read, grep, find, ls |
-| `reviewer` | `openai-codex/gpt-5.6-sol` → `opencode-go/deepseek-v4-pro` | high | read, grep, find, ls |
+| `scout` | `zai-coding-cn/glm-5-turbo` → `nvidia/openai/gpt-oss-20b` → `opencode-go/deepseek-v4-flash` | off | read, grep, find, ls |
+| `tester` | `zai-coding-cn/glm-5-turbo` → `nvidia/openai/gpt-oss-20b` → `opencode-go/deepseek-v4-flash` | off | read, bash, grep, find, ls |
+| `worker` | `zai-coding-cn/glm-5.1` → `nvidia/mistralai/mistral-small-4-119b-2603` → `openrouter/nvidia/nemotron-3-super-120b-a12b:free` → `opencode-go/deepseek-v4-flash` | medium | read, bash, edit, write, grep, find, ls |
+| `general-purpose` | `zai-coding-cn/glm-5.1` → `nvidia/mistralai/mistral-small-4-119b-2603` → `openrouter/nvidia/nemotron-3-super-120b-a12b:free` → `opencode-go/deepseek-v4-flash` | medium | read, bash, edit, write, grep, find, ls |
+| `planner` | `zai-coding-cn/glm-5.2` → `openrouter/nvidia/nemotron-3-ultra-550b-a55b:free` → `opencode-go/deepseek-v4-pro` | high | read, grep, find, ls |
+| `reviewer` | `zai-coding-cn/glm-5.2` → `openrouter/nvidia/nemotron-3-ultra-550b-a55b:free` → `opencode-go/deepseek-v4-pro` | high | read, grep, find, ls |
 
-Each role uses the first authenticated preference available through Pi's model registry, then falls back to the authenticated parent model. User/project agent files remain stronger overrides and may set legacy `model`, ordered `models`, and `thinking`.
+Each role uses the first authenticated preference available through Pi's model registry, then falls back to the authenticated parent model. Chains are **free-first** to conserve the metered opencode-go budget: **zai-coding-cn** (free GLM, primary) → free **nvidia** NIM and **openrouter** `:free` models → **opencode-go** (paid DeepSeek, last resort — one per role: `deepseek-v4-flash` for fast/strong-coding, `deepseek-v4-pro` for deep reasoning). opencode-go's GLM models cost ~$1.40/$4.40 per M versus zai-coding-cn's free GLM, so GLM stays on zai-coding-cn. Fallback models were live-verified on 2026-07-24; `nvidia/moonshotai/kimi-k2.6` and `nvidia/z-ai/glm-5.2` return 404/timeout on the user's account and were removed — non-rate-limit failures kill the subagent instead of advancing the chain. User/project agent files remain stronger overrides and may set legacy `model`, ordered `models`, and `thinking`.
 
 ## Agent files
 
@@ -34,8 +34,10 @@ Create `~/.pi/agent/agents/*.md` or `.pi/agents/*.md`:
 name: scout-fast
 description: Locate relevant files and symbols
 tools: read, grep, find, ls
-model: openai-codex/gpt-5.6-luna
-models: opencode-go/mimo-v2.5, opencode-go/deepseek-v4-flash
+models:
+  - zai-coding-cn/glm-5-turbo
+  - nvidia/openai/gpt-oss-20b
+  - opencode-go/deepseek-v4-flash
 ---
 
 Return concise evidence with file/symbol anchors.

package/agent-format.md

@@ -68,8 +68,10 @@ Read-only service execution (used by `pi-review`) restricts tools to the read-on
 Pi selects the first authenticated/configured model reported by the parent session's `ModelRegistry`. Resolution order is legacy `model`, then each `models` entry, then the authenticated parent model. Duplicate candidates are ignored. If no candidate is available, the error lists every attempted model.
 
 ```yaml
-model: openai-codex/gpt-5.6-terra
-models: opencode-go/deepseek-v4-pro, opencode-go/mimo-v2.5
+models:
+  - zai-coding-cn/glm-5.2
+  - opencode-go/deepseek-v4-flash
+  - nvidia/moonshotai/kimi-k2.6
 ```
 
 The registry includes OAuth subscriptions, API-key subscriptions such as OpenCode Go, environment/runtime credentials, and custom `models.json` providers. Use provider-qualified IDs for predictable routing.

package/agents/general-purpose.md

@@ -3,8 +3,10 @@ name: general-purpose
 description: General-purpose sub-agent for any delegated task. Use when no specialized agent fits. Good for complex research, multi-step operations, and code modifications.
 tools: read, bash, edit, write, grep, find, ls
 models:
-  - openai-codex/gpt-5.6-terra
-  - opencode-go/deepseek-v4-pro
+  - zai-coding-cn/glm-5.1
+  - nvidia/mistralai/mistral-small-4-119b-2603
+  - openrouter/nvidia/nemotron-3-super-120b-a12b:free
+  - opencode-go/deepseek-v4-flash
 thinking: medium
 color: yellow
 ---

package/agents/planner.md

@@ -3,7 +3,8 @@ name: planner
 description: Read-only planning and architecture specialist. Use for consequential design, tradeoff analysis, and implementation plans.
 tools: read, grep, find, ls
 models:
-  - openai-codex/gpt-5.6-sol
+  - zai-coding-cn/glm-5.2
+  - openrouter/nvidia/nemotron-3-ultra-550b-a55b:free
   - opencode-go/deepseek-v4-pro
 thinking: high
 color: blue

package/agents/reviewer.md

@@ -3,7 +3,8 @@ name: reviewer
 description: Code review specialist. Use for correctness, security, regression, and meaningful test-gap review.
 tools: read, grep, find, ls
 models:
-  - openai-codex/gpt-5.6-sol
+  - zai-coding-cn/glm-5.2
+  - openrouter/nvidia/nemotron-3-ultra-550b-a55b:free
   - opencode-go/deepseek-v4-pro
 thinking: high
 color: purple

package/agents/scout.md

@@ -3,10 +3,10 @@ name: scout
 description: Fast codebase recon that returns compressed context for handoff. Use for finding files, understanding structure, locating symbols.
 tools: read, grep, find, ls
 models:
+  - zai-coding-cn/glm-5-turbo
+  - nvidia/openai/gpt-oss-20b
   - opencode-go/deepseek-v4-flash
-  - openai-codex/gpt-5.6-luna
-  - opencode-go/mimo-v2.5
-thinking: low
+thinking: off
 color: cyan
 sandbox: read-only
 ---

package/agents/tester.md

@@ -3,10 +3,10 @@ name: tester
 description: Focused verification agent. Use for cheap routine test, typecheck, lint, build, and regression checks without editing files.
 tools: read, bash, grep, find, ls
 models:
-  - openai-codex/gpt-5.6-luna
-  - opencode-go/mimo-v2.5
+  - zai-coding-cn/glm-5-turbo
+  - nvidia/openai/gpt-oss-20b
   - opencode-go/deepseek-v4-flash
-thinking: low
+thinking: off
 color: orange
 ---
 

package/agents/worker.md

@@ -2,8 +2,10 @@
 name: worker
 description: General-purpose coding agent with full tool access. Use only when explicitly requested for isolated implementation.
 models:
-  - openai-codex/gpt-5.6-terra
-  - opencode-go/deepseek-v4-pro
+  - zai-coding-cn/glm-5.1
+  - nvidia/mistralai/mistral-small-4-119b-2603
+  - openrouter/nvidia/nemotron-3-super-120b-a12b:free
+  - opencode-go/deepseek-v4-flash
 thinking: medium
 color: green
 ---