stevenfackley · Oct 4, 2026
diff --git a/.opencode/agent/designer.md b/.opencode/agent/designer.md
@@ -1,8 +1,8 @@
 ---
 description: UI designer/builder. Commits to a bold aesthetic direction and implements production-grade, non-generic interfaces. Embodies the distinctive-ui-design skill. Use for any UI work where "looks generic" is the failure to avoid.
 mode: all
-model: openrouter/google/gemma-4-26b-a4b-it:free   # multimodal (sees screenshots), fast MoE; 31b sibling is the quality upgrade when its :free endpoint isn't saturated
-fallback: openrouter/nvidia/nemotron-3-nano-30b-a3b:free   # used by tools/council.mjs retry (default; --no-retry disables); doctor.mjs validates it
+model: openrouter/thinkingmachines/inkling-small:free   # VISION (sees screenshots) + 1M ctx; gemma-4-26b's :free endpoint is chronically saturated
+fallback: openrouter/minimax/minimax-m3:free   # also VISION — a designer fallback must still see screenshots; used by tools/council.mjs retry
 temperature: 0.7                    # higher than other agents — UI work needs creative range
 ---
 

diff --git a/.opencode/agent/pr-reviewer.md b/.opencode/agent/pr-reviewer.md
@@ -1,8 +1,8 @@
 ---
 description: Read-only staff-level reviewer for OTHER people's pull requests. Embodies reviewing-others-code — collaborative, severity-labeled, mentoring. Never edits the code under review.
 mode: all
-model: openrouter/google/gemma-4-26b-a4b-it:free     # strong free general reviewer, multimodal; 31b sibling = quality upgrade when its endpoint isn't saturated
-fallback: openrouter/nvidia/nemotron-3-nano-30b-a3b:free   # used by tools/council.mjs retry (default; --no-retry disables); doctor.mjs validates it
+model: openrouter/minimax/minimax-m3:free   # VISION + 1M ctx; MiniMax family, replaces chronically-saturated gemma-4-26b
+fallback: openrouter/nvidia/nemotron-3.5-lightning:free   # NVIDIA != MiniMax primary, 1M ctx; used by tools/council.mjs retry
 temperature: 0.2
 permission:
   edit: deny              # never modify someone else's branch; bash stays on to fetch diffs / run it

diff --git a/.opencode/agent/proposer-b.md b/.opencode/agent/proposer-b.md
@@ -2,7 +2,7 @@
 description: Blind council proposer, family B (poolside Laguna). Produces ONE independent, self-contained proposal or analysis from a context packet — never sees other proposers' output. Used by /deep-design, /deep-debug, /council.
 mode: all   # `all`, not `subagent`: headless `opencode run --agent` only accepts primaries — a subagent SILENTLY falls back to the default agent (zero decorrelation). `all` keeps it usable as an in-session subagent too.
 model: openrouter/poolside/laguna-s-2.1:free   # family B (poolside) — ≠ NVIDIA (judge), ≠ Google (proposer-c)
-fallback: openrouter/inclusionai/ling-3.0-tiny:free   # family-diverse retry; used by tools/council.mjs retry (default; --no-retry disables); doctor.mjs validates it
+fallback: openrouter/minimax/minimax-m3:free   # family-diverse retry; replaces ling-3.0-tiny (slug no longer exists)
 temperature: 0.3
 permission:
   edit: deny

diff --git a/.opencode/agent/proposer-c.md b/.opencode/agent/proposer-c.md
@@ -1,8 +1,8 @@
 ---
 description: Blind council proposer, family C (Google Gemma). Produces ONE independent, self-contained proposal or analysis from a context packet — never sees other proposers' output. Used by /deep-design, /deep-debug, /council.
 mode: all   # `all`, not `subagent`: headless `opencode run --agent` only accepts primaries — a subagent SILENTLY falls back to the default agent (zero decorrelation). `all` keeps it usable as an in-session subagent too.
-model: openrouter/google/gemma-4-26b-a4b-it:free   # family C — ≠ NVIDIA, ≠ inclusionAI
-fallback: openrouter/google/gemma-4-31b-it:free   # bigger sibling, same family; used by tools/council.mjs retry (default; --no-retry disables); doctor.mjs validates it
+model: openrouter/thinkingmachines/inkling-small:free   # family C (ThinkingMachines) != poolside proposer-b, != NVIDIA judge; 1M ctx. Replaces saturated gemma-4-26b
+fallback: openrouter/minimax/minimax-m2.7:free   # family-diverse retry; replaces gemma-4-31b (also saturated)
 temperature: 0.3
 permission:
   edit: deny

diff --git a/.opencode/agent/refuter.md b/.opencode/agent/refuter.md
@@ -1,8 +1,8 @@
 ---
 description: Adversarial finding-killer for /deep-review. Attacks every merged review finding and tries to prove it wrong; only findings that survive reach the user. Family-decorrelated from both reviewers.
 mode: all   # `all`, not `subagent`: headless `opencode run --agent` only accepts primaries — a subagent SILENTLY falls back to the default agent. `all` keeps it usable as an in-session subagent too.
-model: openrouter/openai/gpt-oss-20b:free   # ≠ reviewer (NVIDIA) and ≠ reviewer-cheap (Qwen) — correlated skepticism is a rubber stamp
-fallback: openrouter/inclusionai/ling-3.0-tiny:free   # used by tools/council.mjs retry (default; --no-retry disables); doctor.mjs validates it
+model: openrouter/thinkingmachines/inkling-small:free   # ThinkingMachines != reviewer (NVIDIA) and != reviewer-cheap (MiniMax). Replaces gpt-oss-20b, now paid-only
+fallback: openrouter/nvidia/nemotron-3.5-lightning:free   # at full fallback reviewer=Cohere / reviewer-cheap=MiniMax / this=NVIDIA -> still 3 families
 temperature: 0.1
 permission:
   edit: deny

diff --git a/.opencode/agent/reviewer-cheap.md b/.opencode/agent/reviewer-cheap.md
@@ -1,8 +1,8 @@
 ---
 description: Decorrelated second reviewer on a CHEAP, architecturally-different model. Different training distribution = different blind spots. Used in the /review consensus vote.
 mode: all   # `all`, not `subagent`: headless `opencode run --agent` only accepts primaries — a subagent SILENTLY falls back to the default agent. `all` keeps it usable as an in-session subagent too.
-model: ollama/qwen3.5:9b   # local = $0 + no rate limits; Qwen family ≠ reviewer's NVIDIA
-fallback: openrouter/google/gemma-4-26b-a4b-it:free   # fires ONLY on the headless tools/council.mjs retry (default; --no-retry disables) — in-session /review and /deep-review always use the primary, so no-Ollama users must re-pin the PRIMARY to a :free slug; Google ≠ reviewer's NVIDIA/inclusionAI; doctor.mjs validates it
+model: openrouter/minimax/minimax-m3:free   # MiniMax != reviewer's NVIDIA and != poolside executor; 1M ctx. Replaces gemma-4-26b (saturated)
+fallback: openrouter/minimax/minimax-m2.7:free   # same-family sibling; at full fallback reviewer=Cohere / this=MiniMax / refuter=NVIDIA stay distinct
 temperature: 0.1
 permission:
   edit: deny

diff --git a/.opencode/agent/tester.md b/.opencode/agent/tester.md
@@ -2,7 +2,7 @@
 description: Independent test author. Writes spec tests from the PLAN ONLY and must NOT read the implementation. Eliminates verification gaming — tests can't be reverse-engineered to pass buggy code.
 mode: all   # `all`, not `subagent`: headless `opencode run --agent` only accepts primaries — a subagent SILENTLY falls back to the default agent. `all` keeps it usable as an in-session subagent too.
 model: openrouter/cohere/north-mini-code:free   # agentic coding model, Cohere family ≠ poolside executor
-fallback: ollama/qwen3.5:9b   # local — used by tools/council.mjs retry (default; --no-retry disables); doctor.mjs validates it
+fallback: openrouter/minimax/minimax-m3:free   # MiniMax ≠ Cohere primary and ≠ poolside executor; 1M ctx — used by tools/council.mjs retry (default; --no-retry disables); doctor.mjs validates it
 temperature: 0.1
 permission:
   bash: deny

diff --git a/.opencode/command/a11y-review.md b/.opencode/command/a11y-review.md
@@ -1,6 +1,6 @@
 ---
 description: Audit a screen or diff for Section 508 / WCAG 2.1 AA conformance
-model: openrouter/google/gemma-4-26b-a4b-it:free     # criterion-by-criterion audit; multimodal helps with UI evidence
+model: openrouter/thinkingmachines/inkling-small:free   # VISION — a11y review reads screenshots; replaces saturated gemma-4-26b
 ---
 
 Audit the target below against the `accessibility-508` skill. Go criterion by criterion and

diff --git a/.opencode/command/commit.md b/.opencode/command/commit.md
@@ -1,6 +1,6 @@
 ---
 description: Write a Conventional Commit message from the staged changes (does not commit)
-model: ollama/qwen3.5:4b   # local + instant — mechanical task, costs nothing
+model: openrouter/poolside/laguna-xs-2.1:free   # XS coding model, 262k — mechanical task, don't spend a big model on it (262k matters: a large staged diff is the input)
 ---
 
 Staged summary:

diff --git a/.opencode/command/name.md b/.opencode/command/name.md
@@ -1,6 +1,6 @@
 ---
 description: Generate distinctive product/app/company names (creative-naming) with rationale + availability
-model: openrouter/google/gemma-4-26b-a4b-it:free     # creative range across naming techniques
+model: openrouter/minimax/minimax-m2.7:free   # fast (~13s); replaces saturated gemma-4-26b
 ---
 
 Follow the `creative-naming` skill. From the concept below: extract the essence, generate 15–20

diff --git a/.opencode/command/setup.md b/.opencode/command/setup.md
@@ -7,8 +7,10 @@ Set up opencode-amplifier for use. Walk through and report status at each step:
 1. **Place the kit:** ensure `.opencode/` (agent, command, skills, plugins), `AGENTS.md`,
    `PATTERNS.md`, `patterns/`, `memory/` are present — copy into the project, or merge into
    `~/.config/opencode/` for global use across repos.
-2. **Wire the providers:** in `opencode.jsonc`, confirm `provider.ollama.options.baseURL` and
-   `provider.openrouter`'s `OPENROUTER_API_KEY`. Run `opencode models` to list the real model IDs.
+2. **Wire the provider:** in `opencode.jsonc`, confirm `provider.openrouter`'s
+   `OPENROUTER_API_KEY` resolves. Run `opencode models` to list the real model IDs. (If you added
+   the optional local tier, confirm `provider.ollama.options.baseURL` reaches a live server too —
+   a dead endpoint resolves silently and its agents fail with no useful error.)
 3. **Confirm the shipped per-agent pins resolve:** every `model:`/`fallback:` in
    `.opencode/agent/*.md` (and the `model:` in the five pinned `.opencode/command/*.md`) appears
    in `opencode models`. Do NOT rewrite the decorrelated pins unless the user asks — they are

diff --git a/ONBOARDING.md b/ONBOARDING.md
@@ -4,8 +4,11 @@ For a teammate adopting the kit. Goal: from zero to a contract-governed, multi-m
 free, open-weights models, fast.
 
 ## 1. Prereqs
-- OpenCode CLI installed. Ollama running locally (or on a LAN box — see `docs/local-models.md`).
-- `export OPENROUTER_API_KEY=…` (free key from openrouter.ai/keys).
+- OpenCode CLI installed.
+- `export OPENROUTER_API_KEY=…` (free key from openrouter.ai/keys). That's the only requirement —
+  every shipped pin is an `openrouter/…:free` slug.
+- *Optional:* Ollama locally or on a LAN box, to move the chatty roles off the OpenRouter pool —
+  see `docs/local-models.md`.
 
 ## 2. Get the kit
 - Clone this repo, OR merge its `.opencode/`, `AGENTS.md`, `PATTERNS.md`, `patterns/`, `memory/`

diff --git a/README.md b/README.md
@@ -98,21 +98,41 @@ Discipline lives in the `deep-analysis` skill + AGENTS.md rule 10.
    file.
 
 ## Model tiering (free tier, preserved quota)
-Models are provider-prefixed (`ollama/…` local, `openrouter/…:free` hosted). Chatty roles run
-local (no rate limits); bursty council roles spend the 1,000/day OpenRouter pool. Defaults wired
-across 5+ families so the `/review` consensus vote has uncorrelated blind spots:
+Every shipped pin is `openrouter/…:free` — one API key, no local model server required. Defaults
+are wired across 5+ families so the `/review` consensus vote has uncorrelated blind spots. All of
+it spends the same 1,000/day OpenRouter pool, so cheap roles get small models on purpose:
 | Tier | Models | Use for |
 |---|---|---|
-| Reasoning | `openrouter/nvidia/nemotron-3-ultra-550b-a55b:free` (1M ctx) | architect, debugger, judge, mapper |
-| Executor | `openrouter/poolside/laguna-s-2.1:free` (manual swap to `ollama/ornith:9b` when rate-limited) | the Build agent, step edits |
-| Cross-family review | `openrouter/nvidia/nemotron-3-super-120b-a12b:free`, `openrouter/cohere/north-mini-code:free` | reviewer, tester |
-| Decorrelated review | `ollama/qwen3.5:9b` (local, $0) | reviewer-cheap (different blind spots) |
-| Council diversity | `openrouter/poolside/laguna-s-2.1:free`, `openrouter/google/gemma-4-26b-a4b-it:free`, `openrouter/openai/gpt-oss-20b:free` | proposers, refuter, designer |
-| Mechanical | `ollama/qwen3.5:4b` | commit messages, boilerplate, renames |
-
-⚠️ The `:free` roster churns monthly. `node tools/doctor.mjs` (live, no `--offline`) catches
-delisted slugs; swap in a live `:free` model or a pennies-tier one (see `opencode.jsonc`
-comments).
+| Reasoning | `nvidia/nemotron-3-ultra-550b-a55b:free` (1M ctx) | architect, debugger, judge, mapper, `/adr`, `/port-from-angular` |
+| Executor | `poolside/laguna-s-2.1:free` | the Build agent, step edits |
+| Cross-family review | `nvidia/nemotron-3-super-120b-a12b:free`, `cohere/north-mini-code:free` | reviewer, synthesizer, tester |
+| Decorrelated review | `minimax/minimax-m3:free` (1M ctx, vision) | reviewer-cheap, pr-reviewer |
+| Council diversity | `thinkingmachines/inkling-small:free`, `poolside/laguna-s-2.1:free`, `minimax/minimax-m2.7:free` | proposers, refuter, designer, `/a11y-review`, `/name` |
+| Mechanical | `poolside/laguna-xs-2.1:free` | commit messages, boilerplate, renames |
+
+The `/review` consensus vote runs four different families — poolside (executor), NVIDIA
+(reviewer), MiniMax (reviewer-cheap), ThinkingMachines (refuter) — and the fallbacks are chosen
+so they stay distinct even if every primary is rate-limited at once. Vision roles (`designer`,
+`pr-reviewer`, `/a11y-review`) are pinned to multimodal slugs on purpose.
+
+**Optional local tier ($0, unmetered).** Running Ollama lets you move the chatty roles
+(`reviewer-cheap`, `/commit`) off the OpenRouter pool entirely. It is opt-in, not assumed —
+see `docs/local-models.md` for the server and `examples/opencode.overlay.example.jsonc` for the
+provider block and which pins to move.
+
+⚠️ **The `:free` roster churns monthly, and the catalog lies.** `opencode models` keeps listing
+slugs OpenRouter has since moved to paid, so a static check passes and only a real call fails
+with *"This model is unavailable for free."* On 2026-08-29 five of nine declared slugs were
+broken this way — including a **primary**. Verify with a live round-trip, never the catalog:
+
+```bash
+node tools/doctor.mjs          # round-trips every primary
+node tools/doctor.mjs --deep   # ...and every fallback (2x the calls, catches the silent rot)
+```
+
+Fallbacks rot most — they only run when everything else broke, so nothing exercises them. Run
+`--deep` after any provider change. Swap in a live `:free` model or a pennies-tier one (see
+`opencode.jsonc` comments).
 
 ## Claude Code parity
 OpenCode is Claude-Code-compatible: it auto-discovers `.claude/skills/` and falls back to

diff --git a/docs/local-models.md b/docs/local-models.md
@@ -1,6 +1,11 @@
-# Local model hosting (Ollama)
+# Local model hosting (Ollama) — optional
 
-How to stand up the `ollama/…` tier this kit pins — on your own machine or a LAN box.
+How to stand up an **optional** `ollama/…` tier — on your own machine or a LAN box.
+
+> The kit ships with no local dependency: every pin is an `openrouter/…:free` slug and one API
+> key runs the whole thing. Nothing here is required. Add this tier only if you want the chatty
+> roles off your OpenRouter pool — and only if the server will actually be up, because a declared
+> provider pointing at a dead host resolves silently and its agents fail with no useful error.
 
 ## 1. Why local
 
@@ -147,9 +152,19 @@ ping -c 1 macmini.local
 curl -s http://macmini.local:11434/api/tags
 ```
 
-Client side is one line of config — copy `examples/opencode.overlay.example.jsonc` and change
-`provider.ollama.options.baseURL` to `http://macmini.local:11434/v1`. Nothing else in the kit
-changes; agent pins stay `ollama/qwen3.5:9b` etc.
+Client side: copy the `provider.ollama` block from `examples/opencode.overlay.example.jsonc` into
+your config and set `options.baseURL` to `http://macmini.local:11434/v1`. Then move the pins you
+want local — the two the tier is designed for are:
+
+| File | Ships as | Local swap |
+|---|---|---|
+| `.opencode/agent/reviewer-cheap.md` | `openrouter/google/gemma-4-26b-a4b-it:free` | `ollama/qwen3.5:9b` |
+| `.opencode/command/commit.md` | `openrouter/nvidia/nemotron-3-nano-30b-a3b:free` | `ollama/qwen3.5:4b` |
+
+Keep `reviewer-cheap` in a family that is neither the executor's (poolside) nor the primary
+reviewer's (NVIDIA) — that decorrelation is the whole point of the `/review` consensus vote.
+`node tools/doctor.mjs` round-trips any `ollama/*` primary over direct HTTP and will tell you if
+the endpoint is dead.
 
 ## 5. 16GB tuning
 

diff --git a/examples/opencode.overlay.example.jsonc b/examples/opencode.overlay.example.jsonc
@@ -8,7 +8,12 @@
   // Usage: merge into the project's opencode.jsonc, or keep as ~/.config/opencode/opencode.jsonc.
   "$schema": "https://opencode.ai/config.json",
 
-  // Point the local tier at a LAN model server instead of localhost.
+  // ADD BACK the optional local tier. The kit itself declares NO ollama provider — every shipped
+  // pin is an openrouter/…:free slug, so it works with one API key and no model server. Paste
+  // this block only if you actually run Ollama; a provider pointing at a host that isn't up
+  // resolves silently and its agents fail with no useful error.
+  //
+  // Then move the pins you want local — see the table in docs/local-models.md §4.
   // Stable addressing: give the box a real hostname (mDNS → macmini.local) or a DHCP
   // reservation — never pin a dynamic IP. See docs/local-models.md for the server side.
   "provider": {