ContextualWisdomLab · cursor · Aug 16, 2026 · Aug 16, 2026 · Aug 16, 2026 · Aug 16, 2026
@@ -0,0 +1,103 @@
+name: OpenCode sidecar
+
+# Deploy/CI sidecar only. Does not run on pull_request so app tests and the
+# required Security workflow stay secret-free (see tests.yml / security.yml).
+on:
+  workflow_dispatch:
+  workflow_call:
+    secrets:
+      NVIDIA_NIM_API_KEY:
+        required: false
+      NVIDIA_NIM_API_KEY_SUB:
+        required: false
+      OPENAI_API_KEY:
+        required: false
+      OPENROUTER_API_KEY:
+        required: false
+      BYTEZ_API_KEY:
+        required: false
+  push:
+    branches: [main]
+
+permissions:
+  contents: read
+
+concurrency:
+  group: opencode-sidecar-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  seed_and_serve:
+    name: Seed KV and serve loopback sidecar
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # actions/checkout@v7
+        with:
+          persist-credentials: false
+
+      - name: Set up Python
+        uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # actions/setup-python@v6
+        with:
+          python-version: "3.12"
+
+      - name: Install package
+        run: |
+          python -m pip install --require-hashes -r requirements.lock
+          python -m pip install --no-deps -e .
+
+      - name: Seed org credentials into this-job report
+        env:
+          NVIDIA_NIM_API_KEY: ${{ secrets.NVIDIA_NIM_API_KEY }}
+          NVIDIA_NIM_API_KEY_SUB: ${{ secrets.NVIDIA_NIM_API_KEY_SUB }}
+          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
+          BYTEZ_API_KEY: ${{ secrets.BYTEZ_API_KEY }}
+        run: |
+          python -m contextual_orchestrator seed-provider-catalog \
+            --from-env --skip-missing \
+            --agents examples/agents.production.json \
+            --agents-db "$RUNNER_TEMP/agents.db" \
+            --discover-models | tee "$RUNNER_TEMP/seed-report.json"
+
+      - name: Serve loopback OpenAI-compatible API and smoke curl
+        env:
+          NVIDIA_NIM_API_KEY: ${{ secrets.NVIDIA_NIM_API_KEY }}
+          NVIDIA_NIM_API_KEY_SUB: ${{ secrets.NVIDIA_NIM_API_KEY_SUB }}
+          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
+          BYTEZ_API_KEY: ${{ secrets.BYTEZ_API_KEY }}
+        run: |
+          set -euo pipefail
+          CONTEXTUAL_ORCHESTRATOR_TOKEN="$(python -c 'import secrets; print(secrets.token_urlsafe(32))')"
+          export CONTEXTUAL_ORCHESTRATOR_TOKEN
+          python -m contextual_orchestrator --serve \
+            --seed-from-env \
+            --agents examples/agents.production.json \
+            --agents-db "$RUNNER_TEMP/agents.db" \
+            --host 127.0.0.1 \
+            --port 8000 \
+            --auth-token "$CONTEXTUAL_ORCHESTRATOR_TOKEN" &
+          server_pid=$!
+          trap 'kill "$server_pid" 2>/dev/null || true' EXIT
+          ready=0
+          for _ in 1 2 3 4 5 6 7 8 9 10; do
+            if curl -sf http://127.0.0.1:8000/healthz >/dev/null; then
+              ready=1
+              break
+            fi
+            sleep 1
+          done
+          if [ "$ready" -ne 1 ]; then
+            echo "sidecar did not become healthy"
+            exit 1
+          fi
+          registered="$(python -c 'import json,sys; print(len(json.load(open(sys.argv[1]))["registered_credentials"]))' "$RUNNER_TEMP/seed-report.json")"
+          if [ "$registered" = "0" ]; then
+            echo "no provider secrets in this job; skip live chat smoke (fail-closed, no GitHub Models fallback)"
+            exit 0
+          fi
+          curl -sS --fail http://127.0.0.1:8000/v1/chat/completions \
+            -H "authorization: Bearer $CONTEXTUAL_ORCHESTRATOR_TOKEN" \
+            -H "content-type: application/json" \
+            -d '{"model":"contextual-orchestrator","messages":[{"role":"user","content":"Write one sentence."}]}'
@@ -56,23 +56,30 @@ push or open a PR.
 - The reference implementation is xtrmLLMBatchPython's pgcrypto-encrypted
   Postgres credential registry (`get_credential(name)`); reuse that pattern (a
   DB-backed KV is fine) unless a dedicated KV is adopted.
-- **Known deviation to migrate:** this repo currently resolves provider API
-  keys from env — `ModelClient` reads `os.environ.get(agent.api_key_env)` in
-  `contextual_orchestrator/orchestrator.py` (and `CONTEXTUAL_ORCHESTRATOR_*`
-  tokens in `__main__.py`). Move these to KV-backed reads; keep env only as the
-  bootstrap path that seeds the KV.
+- Runtime provider keys resolve through `get_credential` / the KV registry.
+  Env is bootstrap transport only (`register-credential --from-env` /
+  `seed-provider-catalog --from-env` / `--seed-from-env` on serve). The five
+  org Actions secrets are `NVIDIA_NIM_API_KEY`, `NVIDIA_NIM_API_KEY_SUB`,
+  `OPENAI_API_KEY`, `OPENROUTER_API_KEY`, and `BYTEZ_API_KEY`. A missing secret
+  skips that upstream; it must not crash the pool.
+- ContextualWisdomLab **no longer uses GitHub Models**. Do not add
+  `COPILOT_GITHUB_TOKEN`, `models.github.ai`, `gpt-5.6-luna`, or `gpt-5.6-terra`.
 
 ### This repo: the org LLM gateway
 
 - `contextual-orchestrator` is the org's **LLM-communication hub** — the
-  OpenAI-compatible front door consumed by **gyeot** and **scopeweave**.
+  OpenAI-compatible front door consumed by **gyeot**, **scopeweave**,
+  **OpenCode**, and **Strix**.
 - **Direction:** grow it toward a **LiteLLM-class multi-provider gateway**. The
   org is open to a **Rust/Python hybrid** to cut overhead.
-- Its `ModelClient` currently reads `os.environ.get(agent.api_key_env)` — this
-  is the KV-principle deviation above. Resolve the API key (including the org
-  `OPENAI_API_KEY`) from the **KV / credential registry**, not env.
-- The **OpenCode review pipeline is separate** and stays on **GitHub Models** —
-  do not change it.
+- `ModelClient` resolves the API key from the **KV / credential registry** via
+  `get_credential` (including the org `OPENAI_API_KEY` and NIM / OpenRouter /
+  Bytez keys). Env is never the request-time source.
+- OpenCode/Strix should call this process as **one** OpenAI-compatible
+  provider: `baseURL http://127.0.0.1:8000/v1`, model `contextual-orchestrator`
+  (see `docs/opencode-sidecar.md`). The org-central review workflow lives in
+  `ContextualWisdomLab/.github` and should consume this sidecar — do not
+  reintroduce GitHub Models there from this repo.
 
 ### This repo's role in the ecosystem
 

@@ -0,0 +1,39 @@
+# Changelog
+
+## Unreleased
+
+### Added
+
+- Production agent catalog (`examples/agents.production.json`) for NVIDIA NIM
+  (primary + secondary Nemotron Super 49B / 120B), OpenAI, OpenRouter, and
+  Bytez. Capability tags cover coding, review, and reasoning so Fugu route vs
+  Conductor/TRINITY conduct can pick workers. GitHub Models, Copilot tokens,
+  `gpt-5.6-luna`, and `gpt-5.6-terra` are rejected.
+- `seed-provider-catalog` CLI and `--seed-from-env` serve flag register the five
+  org Actions secrets (`NVIDIA_NIM_API_KEY`, `NVIDIA_NIM_API_KEY_SUB`,
+  `OPENAI_API_KEY`, `OPENROUTER_API_KEY`, `BYTEZ_API_KEY`) into the KV. A
+  missing secret skips that upstream. Optional `GET /v1/models` discovery
+  appends chat models; providers without a list API keep the static seed
+  (`docs/doctoring/provider-catalog.md`).
+- OpenCode/Strix sidecar contract: loopback `http://127.0.0.1:8000/v1`, model
+  `contextual-orchestrator` (`.github/workflows/opencode-sidecar.yml`,
+  `docs/opencode-sidecar.md`). App tests and Security stay secret-free.
+
+### Changed
+
+- Fast-path routing is a cost-performance choose (quality per unit operator
+  cost), not deterministic keyword scoring and not a walk down the seed JSON.
+  429 / 5xx / timeout re-runs the same chooser on the remaining healthy pool.
+  Missing credentials drop that worker from the candidate set. An empty
+  healthy pool fail-closes (no GitHub Models). Deep `conduct` stays
+  Conductor-style and still requires a workflow hint.
+- Unconfigured remote workers are skipped at select/re-selection time. When every
+  provider credential is missing, routing raises `NotConfigured` and does not
+  fall back to GitHub Models.
+- Malformed upstream chat.completion bodies raise `ProviderResponseError` so
+  the gateway failovers or returns a JSON error instead of crashing.
+- Catalog `GET /v1/models` discovery uses the same public-HTTPS egress policy
+  as chat (no private/loopback/redirect leak of the KV Bearer token). HTTP is
+  opt-in via `allow_insecure_discovery`. Reseed replaces the sqlite agent pool
+  so stale discovered workers cannot linger. Sidecar CI fails when `/healthz`
+  never comes up.
@@ -8,7 +8,7 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 
 - **Security gate**: every PR to `main` runs the required Security workflow. A failing Trivy or pip-audit job is a real finding — remediate by bumping the dependency and regenerating `requirements.lock`; never weaken, `continue-on-error`, or disable the gate.
 - **KV, not env**: runtime config and provider secrets are resolved from the KV credential registry (`get_credential`), never `os.getenv` at request time. Env is only bootstrap transport into the KV (see `docs/kv-credentials.md`).
-- **Org role**: this repo is the org's LLM gateway (cost optimizer + sync/batch routing + upstream load balancing, LiteLLM-plus scope), consumed by `gyeot` and `scopeweave`. The OpenCode review pipeline is separate, stays on GitHub Models, and must not be changed.
+- **Org role**: this repo is the org's LLM gateway (cost optimizer + sync/batch routing + upstream load balancing, LiteLLM-plus scope), consumed by `gyeot`, `scopeweave`, OpenCode, and Strix. ContextualWisdomLab no longer uses GitHub Models; OpenCode should call `http://127.0.0.1:8000/v1` model `contextual-orchestrator` (see `docs/opencode-sidecar.md`).
 - **Research grounding**: substantive feature/process PRs should attach the relevant papers (PDF when redistribution is permissible, otherwise cite + link + summary) under `docs/papers/` with full citations.
 
 This file complements AGENTS.md with commands and architecture; where they differ, AGENTS.md wins.
@@ -55,6 +55,8 @@ python -m contextual_orchestrator --eval "prompt one" "prompt two"
 
 # Seed a provider credential into the KV at bootstrap
 echo "$OPENAI_API_KEY" | python -m contextual_orchestrator register-credential --name OPENAI_API_KEY --value-stdin
+python -m contextual_orchestrator seed-provider-catalog --from-env --skip-missing \
+  --agents examples/agents.production.json --agents-db /tmp/agents.db
 
 # Reproduce the Trivy security gate locally (against the merge result)
 trivy --download-db-only
@@ -80,7 +82,7 @@ A stdlib-Python lab implementing a single OpenAI-compatible API that routes, del
 2. `TaskOrchestrator.complete()` in `orchestrator.py` picks one of two paths:
    - **Fast path (`route`)**: select a single worker for simple or latency-sensitive requests.
    - **Deep path (`conduct`)**: build a natural-language workflow of `thinker → worker → verifier → synthesizer` steps. Each `WorkflowStep` carries an **access list** so a worker sees only the prior outputs deliberately exposed to it.
-3. `ModelClient` (infrastructure adapter) executes each step against `mock://` agents (offline, used by tests) or OpenAI-compatible HTTPS providers, with jittered retries for transient errors, failover to the next capability-matched agent, and a per-agent circuit breaker. Provider keys come from the KV via `get_credential`; egress to loopback/private/reserved addresses is blocked.
+3. `ModelClient` (infrastructure adapter) executes each step against `mock://` agents (offline, used by tests) or OpenAI-compatible HTTPS providers, with jittered retries for transient errors. Worker selection is a cost-performance choose (quality per unit cost); a 429/5xx/timeout **re-runs that chooser** on the remaining healthy pool rather than walking the seed list. A per-agent circuit breaker excludes a persistently failing provider until it cools down. Provider keys come from the KV via `get_credential`; egress to loopback/private/reserved addresses is blocked.
 4. The answer is framed as an OpenAI `chat.completion` (or SSE `chat.completion.chunk` stream). Full orchestration traces are only returned to trusted callers.
 
 ### Modules (`contextual_orchestrator/`)
@@ -91,7 +93,7 @@ A stdlib-Python lab implementing a single OpenAI-compatible API that routes, del
 - `credentials.py` / `kv_config.py` — the KV seam: `get_credential`/`register_credential` over pluggable backends (`InMemoryCredentialBackend` default; pgcrypto-encrypted `PostgresCredentialBackend`, selected via `CONTEXTUAL_ORCHESTRATOR_KV_BACKEND`).
 - `cost_ledger.py` / `cost_router.py` / `batch_routing.py` / `token_counting.py` — the cost-review + routing hub: prompt-safe usage ledger with seven attribution dimensions, `RoutingPolicy` (sync vs batch from request hints + KV thresholds), and the [pg-llm-batch](https://github.com/ContextualWisdomLab/pg-llm-batch) batch/embeddings backends (a local in-process backend keeps the standalone path working with no external service).
 - `api_contract.py` / `conventions.py` — API-shape and naming-rule enforcement helpers.
-- `__main__.py` — the single entry point: CLI completion, `--serve`, `--eval`, and the `register-credential` bootstrap subcommand.
+- `__main__.py` — the single entry point: CLI completion, `--serve`, `--eval`, `register-credential`, and `seed-provider-catalog` / `--seed-from-env`.
 
 Agent pools are **data, not code**: `examples/agents.mock.json` and `examples/agents.openai.json`. State is in-memory by default; `--state-db PATH` persists runs/audit/analytics to sqlite.
 

@@ -44,6 +44,13 @@ HTTP serving is hardened for local lab use:
 - Response caching is off by default. Pass `--cache-ttl SECONDS` to serve identical requests (same messages + mode) from an in-memory TTL+LRU cache and skip the provider calls; `0` disables it.
 - `ModelClient.batch_chat(agent, {custom_id: messages})` runs many requests through the provider's Batch API (async, 24h completion window, typically ~50% cheaper) — suited to evaluation/benchmark workloads, not latency-sensitive chat. The mock path answers synchronously.
 
+Production (not mock-only) seed: [examples/agents.production.json](examples/agents.production.json) — NVIDIA NIM primary/secondary (Nemotron Super 49B / 120B), OpenAI, OpenRouter, and Bytez. OpenCode/Strix should call `http://127.0.0.1:8000/v1` with model `contextual-orchestrator` ([docs/opencode-sidecar.md](docs/opencode-sidecar.md)). GitHub Models are not in the catalog.
+
+```bash
+python -m contextual_orchestrator seed-provider-catalog --from-env --skip-missing \
+  --agents examples/agents.production.json --agents-db /tmp/agents.db
+```
+
 Use real workers by replacing `mock://` agents with OpenAI-compatible endpoints. Provider secrets are resolved from a KV credential registry via `get_credential`, never from `os.getenv` at request time (see [docs/kv-credentials.md](docs/kv-credentials.md)):
 
 ```json
@@ -115,7 +122,7 @@ One fused orchestration loop:
 - Deep path: a natural-language workflow is built with planner, worker, verifier, and synthesizer steps.
 - Each step has an access list, so workers see only the prior outputs intentionally exposed to them.
 - Agent definitions are data, so provider preference, exclusions, privacy constraints, and mock testing do not require code changes.
-- Provider calls are resilient: transient failures (timeouts, 429, 5xx) retry with full-jitter exponential backoff, while caller errors (4xx) fail fast. If an agent still fails, the request fails over to the next capability-matched agent in the pool, and a per-agent circuit breaker skips a persistently failing provider until it cools down. Failover is recorded in the trace (`served_agent_id`, `failover_from`).
+- Fast-path selection is a cost-performance choose: one worker from the live pool that maximizes expected quality per unit cost (operator `price_per_million`, TRINITY role tags, measured circuit/latency when present). Seed JSON order and prompt keywords do not pick the winner. Transient failures (timeouts, 429, 5xx) retry with full-jitter exponential backoff on that worker; if it still fails, the **same chooser** runs again on the remaining healthy pool (circuit-open agents excluded). An empty healthy pool fail-closes. Re-selection is recorded in the trace (`served_agent_id`, `failover_from`).
 
 See [docs/architecture.md](docs/architecture.md) for the source-backed analysis.
 
@@ -252,6 +259,11 @@ python -m pip install --require-hashes -r requirements.lock
 python -m pip install --no-deps -e .
 python tests/test_self_check.py
 python tests/test_paper_contracts.py
+python tests/test_provider_catalog.py
+python tests/test_catalog_bootstrap.py
+python tests/test_provider_catalog_robustness.py
+python tests/test_cost_performance_chooser.py
+python tests/test_opencode_sidecar_contract.py
 python tests/test_admin_contract.py
 python tests/test_conventions.py
 python tests/test_api_contract.py

@@ -39,5 +39,7 @@ Provide one API and one domain model:
 - Training a learned coordinator.
 - Claiming compatibility with or ownership of any vendor model.
 - Adding provider SDKs before stdlib HTTP proves insufficient.
+- Reintroducing GitHub Models (`COPILOT_GITHUB_TOKEN`, `gpt-5.6-luna` / `terra`).
+  OpenCode and Strix call this process as one OpenAI-compatible provider.
 
 See `docs/product_planning.md` for the paper-grounded product plan.
@@ -4,3 +4,4 @@
 |---|---|---|
 | 001-paper-grounded-orchestrator | active | Implement the source-backed orchestration contract with TDD, DDD, and CDD |
 | 002-enterprise-design-foundation | active | Add paper-grounded screen design, user stories, REST API, code/DB conventions, and i18n |
+| 003-org-provider-catalog | active | Seed NIM/OpenAI/OpenRouter/Bytez into the KV, OpenCode sidecar, cost-performance choose + re-selection |