crew 0.14.0: opencode's current free models; Copilot and Cursor plan notes

bench/plugins/crew/results/2026-10-26-opencode-zen-refresh-CLAIMS.md

@@ -0,0 +1,20 @@
+# Batch 2026-10-26-opencode-zen-refresh — claims
+
+**Smoke notes, not run records.** n=1 per model, opencode v2.1.4.
+
+- `opencode models` listed eight free models: beam, deepseek-v4.1-flash,
+  flint-40b, glm-5.3-flash, mimo-v2.6-flash, nemotron-3-ultra,
+  nemotron-3.5-lightning and nemotron-3.5-super-vl. minimax-m3.1-flash, still in
+  crew's allowlist, was gone.
+- `crew.py ask opencode "In one sentence: what does the Unix command 'tee'
+  do?" --model opencode/<m>` answered correctly with `beam-free` (3 s) and
+  `flint-40b-free` (4 s). `nemotron-3-ultra-free` failed twice, 20 s apart:
+  `provider.invalid-request`, "Upstream request failed: Model is unavailable."
+  (400), so it stays out of the allowlist.
+- Zen's docs, read the same day: deepseek and glm are zero-retention and
+  no-training; beam does not train and keeps data 30 days; flint and mimo may
+  train while free; the Nemotron models are logged by NVIDIA ("do not submit
+  personal or confidential data").
+- GitHub's plans page: "Copilot Free and Copilot Student both have an
+  allowance of AI credits", and both offer only auto model selection. crew's
+  note that Free had no free quota was wrong.

plugins/crew/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "crew",
   "displayName": "Crew",
-  "version": "0.13.0",
+  "version": "0.14.0",
   "author": {
     "name": "4nrry",
     "url": "https://github.com/4nrry"

plugins/crew/plugin.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
   "name": "crew",
-  "version": "0.13.0",
+  "version": "0.14.0",
   "description": "Another vendor's agent CLI as an isolated subagent, from any host: ask a read-only question or hold a multi-turn conversation (agy, opencode, Codex, Grok, Hermes, Copilot, Cursor, Muse, Kimi, or Claude Code from another host), get a second opinion on a diff, or delegate one coding slice to agy, opencode, Copilot, Codex, Grok or Cursor under the worker CLI's own hooks and a mechanical gate. Tiers (cheap, mid, top) map to a model and effort per backend. The verdict never reads the CLI's exit 0 as success.",
   "author": {
     "name": "4nrry",

plugins/crew/skills/crew/references/copilot.md

@@ -10,8 +10,9 @@ python3 scripts/crew.py delegate /path/TASK.json --backend copilot
 
 ## When to use
 
-Rarely, and knowing it costs: the Free plan has no free quota, so every run
-spends AI credits (well under one per small run). It fits a slice where a model
+Rarely, and knowing it costs: every run spends AI credits (well under one per
+small run) from the plan's monthly allowance, which Free and Student plans also
+get; GitHub does not publish its size. It fits a slice where a model
 from another family is wanted and **the whole gate runs without the worker**:
 it has no shell, so it cannot run tests, formatters or git. The brief must be
 enough without running anything.
@@ -26,8 +27,9 @@ plugins, MCP servers and hooks (the verdict warns about hooks it finds in
 
 ## Model and credits
 
-- Only `auto` is accepted. On a paid plan `"copilot_model_override": true`
-  allows another slug — an untested path; say so.
+- Only `auto` is accepted: on Free and Student plans GitHub offers only auto
+  model selection. On a paid plan `"copilot_model_override": true` allows
+  another slug — an untested path; say so.
 - `max_ai_credits`: default and minimum 30 (the CLI refuses less). A brake on a
   runaway run, not a budget.
 - **Report `copilot.usage.ai_credits` to the user** after every run.

plugins/crew/skills/crew/references/cursor.md

@@ -9,6 +9,9 @@ python3 scripts/crew.py delegate /path/TASK.json --backend cursor
 ```
 
 Use it to spend Cursor plan quota on a slice. Never from Cursor itself.
+With an X Subscription (Grok), the shared pool pays for Grok models in Cursor
+only: the default `grok-4.7` spends it, while another vendor's model bills
+Cursor's own plan.
 
 ## Before the first run
 

plugins/crew/skills/crew/references/opencode.md

@@ -12,20 +12,25 @@ Use it when agy's quota is gone, or for extra slices at no cost.
 
 ## Models and data
 
-Only these six are accepted. Zen rotates its free models: on 2026-10-22
-minimax left (`provider.no-route`), so the default and the `cheap` tier are
-deepseek, `mid` is glm, and `top` is Nemotron Super VL (1M context, but NVIDIA
-logs the prompts). Terms from the Zen docs; re-read them before relying on this
-table, they change:
+Only these seven are accepted. Zen rotates its free models, so check
+`opencode models` before relying on the list: the default and the `cheap` tier
+are deepseek, `mid` is glm, and `top` is Nemotron Super VL (1M context, but
+NVIDIA logs the prompts). Terms from the Zen docs, checked 2026-10-26; re-read
+them, they change:
 
 | Model | Data terms |
 |---|---|
-| `opencode/minimax-m3.1-flash-free` (unavailable since 2026-10-22) | zero retention, no training |
 | `opencode/deepseek-v4.1-flash-free` (default) | zero retention, no training |
 | `opencode/glm-5.3-flash-free` | zero retention, no training |
+| `opencode/beam-free` | no training; kept 30 days for safety |
 | `opencode/nemotron-3.5-super-vl-free` | logged by NVIDIA to improve its products: no personal or confidential data |
 | `opencode/nemotron-3.5-lightning-free` | same as Super VL |
-| `opencode/mimo-v2.6-flash-free` | prompts may be used to improve the model while free |
+| `opencode/mimo-v2.6-flash-free` | may be used to improve the model while free |
+| `opencode/flint-40b-free` | may be used to improve the model while free; kept 30 days |
+
+Private code goes only to the two zero-retention models. Not accepted:
+minimax (left Zen on 2026-10-22), and `nemotron-3-ultra-free`, which `opencode
+models` lists but which answered "Model is unavailable" on 2026-10-26.
 
 ## The guard is weaker — tell the user
 

plugins/crew/skills/crew/scripts/backends.py

@@ -23,9 +23,10 @@
 LIVE = {"claude": "2.1.310", "agy": "1.4.7", "opencode": "2.1.4", "grok": "1.0.70",
         "codex": "0.164.0", "hermes": "0.22.5", "copilot": "1.0.103", "cursor": "2026.10.22-27dcf5f"}
 
-OPENCODE_MODELS = ("opencode/minimax-m3.1-flash-free", "opencode/deepseek-v4.1-flash-free",
-                   "opencode/glm-5.3-flash-free", "opencode/mimo-v2.6-flash-free",
-                   "opencode/nemotron-3.5-lightning-free", "opencode/nemotron-3.5-super-vl-free")
+OPENCODE_MODELS = ("opencode/deepseek-v4.1-flash-free", "opencode/glm-5.3-flash-free",
+                   "opencode/beam-free", "opencode/nemotron-3.5-super-vl-free",
+                   "opencode/nemotron-3.5-lightning-free", "opencode/mimo-v2.6-flash-free",
+                   "opencode/flint-40b-free")
 OPENCODE_DENY = {"permission": {"edit": "deny", "bash": "deny", "execute": "deny",
                                 "task": "deny", "external_directory": "deny"}}
 # --web: the same denials, web fetch and search allowed (fleet_worker's research runs).
@@ -350,7 +351,7 @@ def b_kimi(s, prompt, model, effort, session, cwd, opts=None):
                 read_only="--mode plan"),
     "opencode": dict(web=True, native_schema=False, vendor="opencode", bins=("opencode", "~/.opencode/bin/opencode"),
                      build=b_opencode, parse=p_opencode,
-                     # minimax (first) left Zen on 2026-10-22. top is the 1M-context
+                     # top is the 1M-context
                      # Nemotron, whose prompts NVIDIA logs: never for private material.
                      tiers={"cheap": ("opencode/deepseek-v4.1-flash-free", None),
                             "mid": ("opencode/glm-5.3-flash-free", None),

plugins/crew/skills/crew/scripts/crew.py

@@ -117,13 +117,13 @@
 BACKENDS = ("agy", "opencode", "copilot", "codex", "grok", "cursor", "claude")
 # review keeps crew's guard hooks on these; every other backend reviews through ask.
 REVIEW_GUARDED = ("agy", "opencode", "copilot")
-# minimax-m3.1-flash-free left Zen (provider.no-route, 2026-10-22); kept allowed should it return.
 OPENCODE_DEFAULT_MODEL = "opencode/deepseek-v4.1-flash-free"
-# The only opencode models a run may use (the user's choice, 2026-10-21); the
-# Zen data terms of each are in the opencode skill.
-OPENCODE_MODELS = ("opencode/minimax-m3.1-flash-free", "opencode/deepseek-v4.1-flash-free",
-                   "opencode/glm-5.3-flash-free", "opencode/nemotron-3.5-super-vl-free",
-                   "opencode/nemotron-3.5-lightning-free", "opencode/mimo-v2.6-flash-free")
+# The only opencode models a run may use: Zen's free list on 2026-10-26 (minimax
+# left on 2026-10-22). The Zen data terms of each are in references/opencode.md.
+OPENCODE_MODELS = ("opencode/deepseek-v4.1-flash-free", "opencode/glm-5.3-flash-free",
+                   "opencode/beam-free", "opencode/nemotron-3.5-super-vl-free",
+                   "opencode/nemotron-3.5-lightning-free", "opencode/mimo-v2.6-flash-free",
+                   "opencode/flint-40b-free")
 # opencode 2.x dropped `--dir` (the run uses cwd = repo) and by default attaches
 # to a shared background service, which would not see this process's
 # OPENCODE_CONFIG_CONTENT and so would run without the guard plugin.