chore(opencode): migrate to released contextual-orchestrator (owner prerequisite)

AGENTS.md

@@ -98,3 +98,11 @@ This section applies to any agent (Claude, Codex, Cursor, opencode, ...) working
 - Do not justify risky debug defaults as temporary shortcuts.
 - Do not defer dependency review, SBOM generation, or supply-chain checks to a later phase.
 - Do not defer Windows or macOS build enforcement to a later phase.
+
+## Local OpenCode contract
+- Root `opencode.jsonc` is NVIDIA NIM only: `enabled_providers` is `["nvidia-nim"]`.
+- The provider binds `{env:NVIDIA_API_KEY}` to `https://integrate.api.nvidia.com/v1`. That env name is the NVIDIA/OpenCode client contract.
+- The organization GitHub secret is `NVIDIA_NIM_API_KEY`. CI maps that secret onto process env `NVIDIA_API_KEY`. Do not rename the local OpenCode binding to `{env:NVIDIA_NIM_API_KEY}`.
+- Never introduce `COPILOT_GITHUB_TOKEN`, `github-models`, or `STRIX_GITHUB_MODELS_TOKEN` in `opencode.jsonc`.
+- Do not set OpenAI-style `reasoningEffort` on `nvidia/llama-3.3-nemotron-super-49b-v1.5`. NIM uses `chat_template_kwargs` for that model (see `docs/doctoring/opencode-nvidia-nim-contract.md`).
+- Do not change existing review-agent credential names or routing. Central OpenCode Review and the PR review/merge scheduler stay in `ContextualWisdomLab/.github`.

CLAUDE.md

@@ -73,3 +73,4 @@ Supporting packages:
 - Lockfiles (`package-lock.json`, `uv.lock`, `Cargo.lock`) are committed and must stay in sync; GitHub Actions are SHA-pinned. Adding a direct dependency requires the admission rationale defined in `AGENTS.md` and `docs/security/dependency-policy.md`.
 - CI beyond quickcheck: `gate / ci / rust-check` (Tauri cargo check on macOS) and `build-baseline` Windows/macOS amd64+arm64 native builds are merge gates, alongside CodeQL, dependency-review, sbom, bandit, trivy, secret-scan, and security-audit workflows. Do not weaken or skip them.
 - Version metadata lives in `VERSION`, the root `package.json`, and `CHANGELOG.md`; release flow is tag-driven (see `docs/operations/deploy-runbook.md`).
+- Local developer OpenCode is NVIDIA NIM only. The binding, secret split, leftover-token deny-list, and `reasoningEffort` ban live in `AGENTS.md` (`Local OpenCode contract`) and `docs/doctoring/opencode-nvidia-nim-contract.md`.

docs/doctoring/opencode-nvidia-nim-contract.md

@@ -0,0 +1,53 @@
+# Local OpenCode NVIDIA NIM contract
+
+Next action: keep root `opencode.jsonc` on NVIDIA NIM only. Bind `{env:NVIDIA_API_KEY}` to
+`https://integrate.api.nvidia.com/v1`. Do not rename that local binding to the organization
+secret name `NVIDIA_NIM_API_KEY`. Do not put OpenAI-style `reasoningEffort` anywhere in the
+local OpenCode configuration.
+
+## Why this lock exists
+
+Local developer OpenCode is a separate trust boundary from central OpenCode Review and the
+PR review/merge scheduler in `ContextualWisdomLab/.github`. Those workflows keep their own
+credential names. This repository only records the local client allowlist.
+
+NVIDIA NIM reasoning models use `chat_template_kwargs` (for example `enable_thinking`) rather
+than the OpenAI `reasoning_effort` field. NVIDIA documents `reasoning_effort` as a Chat
+Completions knob for GPT-OSS models on multi-LLM NIM, not for Llama 3.3 Nemotron Super 49B
+v1.5 (NVIDIA, n.d.-a; NVIDIA, n.d.-b). Forwarding `reasoningEffort` from OpenCode can be
+ignored or rejected. The contract therefore forbids that OpenCode option across the whole
+repository config instead of checking only one model's nested `options` object.
+
+The organization GitHub secret remains `NVIDIA_NIM_API_KEY`. CI maps that secret onto process
+env `NVIDIA_API_KEY` because that is the NVIDIA/OpenCode client binding.
+
+## Held values
+
+- `enabled_providers`: `["nvidia-nim"]`
+- default model: `nvidia-nim/nvidia/llama-3.3-nemotron-super-49b-v1.5`
+- `small_model`: `nvidia-nim/meta/llama-3.3-70b-instruct`
+- `apiKey`: `{env:NVIDIA_API_KEY}`
+- leftover deny-list: `github-models`, `STRIX_GITHUB_MODELS_TOKEN`, `COPILOT_GITHUB_TOKEN`,
+  `openai/gpt-5`, `openai/o3`, `openai/o4-mini`, `models.github.ai`
+
+## References
+
+NVIDIA. (n.d.-a). *Use reasoning models with NVIDIA NIM for LLMs*.
+https://docs.nvidia.com/nim/large-language-models/latest/reasoning-model.html
+
+NVIDIA. (n.d.-b). *nvidia / llama-3.3-nemotron-super-49b-v1.5*.
+https://docs.api.nvidia.com/nim/reference/nvidia-llama-3_3-nemotron-super-49b-v1_5
+
+## Security Notes
+
+- Attack surface: local OpenCode HTTPS calls to `https://integrate.api.nvidia.com/v1` using a
+  process-env API key.
+- Trust boundary: untrusted prompts and repository contents sent to the provider; trusted
+  repo-controlled provider allowlist; secret name `NVIDIA_NIM_API_KEY` stays in GitHub and is
+  mapped onto `NVIDIA_API_KEY` only at process start.
+- Mitigations: single enabled provider, no GitHub Models or Copilot token fallback, leftover
+  string deny-list, and explicit whole-config bans on `reasoningEffort` and
+  `{env:NVIDIA_NIM_API_KEY}` inside `opencode.jsonc`.
+- Test points: `test_opencode_uses_nvidia_nim_only` owns the structured provider/model/secret
+  contract; `test_opencode_forbids_reasoning_effort_anywhere` independently prevents the
+  forbidden OpenAI-style option from being pasted at any config level.

docs/workflow/pr-review-merge-scheduler.md

@@ -16,6 +16,12 @@ workflow `GITHUB_TOKEN`, depending on which credential can perform the guarded r
 The local repository may keep product CI, security, release, and build workflows. It must not restore
 repo-local copies of `opencode-review.yml`, `pr-review-merge-scheduler.yml`, or their `scripts/ci` helper implementations.
 
+Local developer OpenCode (`opencode.jsonc`) is a separate trust boundary from those central
+review workflows. It uses NVIDIA NIM only, binds `{env:NVIDIA_API_KEY}` at
+`https://integrate.api.nvidia.com/v1`, and must not restore GitHub Models, Copilot tokens, or
+review-agent secrets. The organization secret name remains `NVIDIA_NIM_API_KEY`; CI maps that
+secret onto `NVIDIA_API_KEY` for the OpenCode client.
+
 ## Behavior
 
 - Inspect non-draft PRs targeting the repository default branch, currently `develop`.
@@ -44,14 +50,15 @@ repo-local copies of `opencode-review.yml`, `pr-review-merge-scheduler.yml`, or
 
 ## Security Notes
 
-- Attack surface: organization required workflows with write access to PR comments, PR branch updates, and normal merges.
-- Trust boundary touched: GitHub repository governance, PR review state, status checks, and CodeRabbit review requests.
+- Attack surface: organization required workflows with write access to PR comments, PR branch updates, and normal merges. Local developer OpenCode (`opencode.jsonc`) is a separate HTTPS client to NVIDIA NIM and is not invoked by this scheduler.
+- Trust boundary touched: GitHub repository governance, PR review state, status checks, and CodeRabbit review requests. The local OpenCode provider allowlist (`nvidia-nim` + `{env:NVIDIA_API_KEY}`) is recorded here so agents do not restore GitHub Models or review-agent secrets into this repository.
 - Realistic threats: spammed review comments, merging a PR with unresolved conversations, merging without required checks, or hiding conflicts behind automation.
 - Mitigations: central required workflow source pinning, idempotent per-head review comment marker,
   explicit unresolved-thread check, retry-bounded GitHub API reads, required-check verification
   through GitHub, conflict skip, guarded merge with `--match-head-commit`, and no admin bypass path.
 - Remaining risk: CodeRabbit and GitHub check state can be delayed or stale; the scheduler therefore only advances eligible PRs and leaves code-fix work to agents or maintainers.
 - Test points: organization ruleset inheritance, current-head OpenCode approval, unresolved review
   thread count, required-check rollup, approved behind PR, approved conflict-free PR, approved dirty PR,
-  external failed-check classification, provider/runtime failure summary, and Strix evidence lookup
-  scope diagnostics.
+  external failed-check classification, provider/runtime failure summary, Strix evidence lookup
+  scope diagnostics, and the separate local `opencode.jsonc` NIM-only contract (`small_model`,
+  `{env:NVIDIA_API_KEY}`, no `reasoningEffort`, no GitHub Models / Copilot leftovers).

opencode.jsonc

@@ -1,8 +1,8 @@
 {
   "$schema": "https://opencode.ai/config.json",
-  "model": "github-models/openai/gpt-5",
-  "small_model": "github-models/deepseek/deepseek-v3-0324",
-  "enabled_providers": ["github-models"],
+  "model": "nvidia-nim/nvidia/llama-3.3-nemotron-super-49b-v1.5",
+  "small_model": "nvidia-nim/meta/llama-3.3-70b-instruct",
+  "enabled_providers": ["nvidia-nim"],
   "mcp": {
     "codegraph": {
       "type": "local",
@@ -37,56 +37,29 @@
     }
   },
   "provider": {
-    "github-models": {
+    "nvidia-nim": {
       "npm": "@ai-sdk/openai-compatible",
-      "name": "GitHub Models",
+      "name": "NVIDIA NIM",
       "options": {
-        "baseURL": "https://models.github.ai/inference",
-        "apiKey": "{env:STRIX_GITHUB_MODELS_TOKEN}"
+        "baseURL": "https://integrate.api.nvidia.com/v1",
+        "apiKey": "{env:NVIDIA_API_KEY}"
       },
       "models": {
-        "openai/gpt-5": {
-          "name": "OpenAI GPT-5",
+        "nvidia/llama-3.3-nemotron-super-49b-v1.5": {
+          "name": "NVIDIA Llama 3.3 Nemotron Super 49B v1.5",
           "tool_call": true,
           "reasoning": true,
           "limit": {
-            "context": 200000,
-            "output": 100000
+            "context": 131072,
+            "output": 32768
           }
         },
-        "deepseek/deepseek-r1-0528": {
-          "name": "DeepSeek R1 0528",
+        "meta/llama-3.3-70b-instruct": {
+          "name": "Meta Llama 3.3 70B Instruct",
           "tool_call": true,
-          "reasoning": true,
-          "limit": {
-            "context": 128000,
-            "output": 4096
-          }
-        },
-        "deepseek/deepseek-v3-0324": {
-          "name": "DeepSeek V3 0324",
-          "tool_call": true,
-          "limit": {
-            "context": 128000,
-            "output": 4096
-          }
-        },
-        "openai/o3": {
-          "name": "OpenAI o3",
-          "tool_call": true,
-          "reasoning": true,
-          "limit": {
-            "context": 200000,
-            "output": 100000
-          }
-        },
-        "openai/o4-mini": {
-          "name": "OpenAI o4-mini",
-          "tool_call": true,
-          "reasoning": true,
           "limit": {
-            "context": 200000,
-            "output": 100000
+            "context": 131072,
+            "output": 16384
           }
         }
       }

services/analysis-engine/tests/test_opencode_nim_secret_contract.py

@@ -0,0 +1,13 @@
+"""Focused regression for forbidden OpenCode reasoning options."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+
+def test_opencode_forbids_reasoning_effort_anywhere() -> None:
+    """Reject OpenAI-style reasoningEffort anywhere in the local NIM config."""
+    repo_root = Path(__file__).resolve().parents[3]
+    opencode_text = (repo_root / "opencode.jsonc").read_text(encoding="utf-8")
+
+    assert '"reasoningEffort"' not in opencode_text

services/analysis-engine/tests/test_supply_chain_policy.py

@@ -5073,14 +5073,44 @@ def test_opencode_approval_write_failure_updates_overview_only() -> None:
     assert "source-backed repository findings" in policy
 
 
+def test_opencode_uses_nvidia_nim_only() -> None:
+    """Ensure local OpenCode is NVIDIA NIM only and does not use GitHub Models."""
+    repo_root = Path(__file__).resolve().parents[3]
+    opencode_text = (repo_root / "opencode.jsonc").read_text(encoding="utf-8")
+    opencode_config = json.loads(opencode_text)
+
+    assert opencode_config["model"] == "nvidia-nim/nvidia/llama-3.3-nemotron-super-49b-v1.5"
+    assert opencode_config["small_model"] == "nvidia-nim/meta/llama-3.3-70b-instruct"
+    assert opencode_config["enabled_providers"] == ["nvidia-nim"]
+    assert set(opencode_config["provider"]) == {"nvidia-nim"}
+
+    nim_provider = opencode_config["provider"]["nvidia-nim"]
+    assert nim_provider["options"]["baseURL"] == "https://integrate.api.nvidia.com/v1"
+    assert nim_provider["options"]["apiKey"] == "{env:NVIDIA_API_KEY}"
+    assert "nvidia/llama-3.3-nemotron-super-49b-v1.5" in nim_provider["models"]
+    assert "meta/llama-3.3-70b-instruct" in nim_provider["models"]
+
+    primary_model = nim_provider["models"]["nvidia/llama-3.3-nemotron-super-49b-v1.5"]
+    assert "reasoningEffort" not in primary_model.get("options", {})
+    assert "{env:NVIDIA_NIM_API_KEY}" not in opencode_text
+
+    leftover_tokens = (
+        "github-models",
+        "STRIX_GITHUB_MODELS_TOKEN",
+        "COPILOT_GITHUB_TOKEN",
+        "openai/gpt-5",
+        "openai/o3",
+        "openai/o4-mini",
+        "models.github.ai",
+    )
+    for leftover in leftover_tokens:
+        assert leftover not in opencode_text
+
+
 def test_pr_review_merge_scheduler_uses_central_mutation_credential() -> None:
     """Ensure mechanical PR queue handling uses the central mutation credential."""
-    repo_root = Path(__file__).resolve().parents[3]
     policy = central_required_workflow_policy_text()
 
-    opencode_config = (repo_root / "opencode.jsonc").read_text(encoding="utf-8")
-    assert '"openai/o3"' in opencode_config
-    assert '"openai/o4-mini"' in opencode_config
     assert_local_review_workflows_removed()
     assert "selected workflow mutation" in policy
     assert "credential, not by a maintainer's local `gh` session" in policy