ContextualWisdomLab · seonghobae · Aug 30, 2026
@@ -5,6 +5,29 @@ this file. The format follows Keep a Changelog, and versioned releases follow
 Semantic Versioning where the repository publishes a release.
 
 ## [Unreleased]
+- Remove `scripts/ci/select_nvidia_nim_model.py`: an orphaned helper that
+  opened a direct `HTTPSConnection` to `integrate.api.nvidia.com` with a raw
+  provider API key to resolve a live NVIDIA NIM model id, predating the
+  ADR-0003 gateway migration and wired into nothing (no workflow or script
+  referenced it — only its own dedicated test, `test_scheduled_autofix_routes_through_contextual_orchestrator`
+  already pins `https://integrate.api.nvidia.com/v1` as a **forbidden**
+  string in `pr-review-autofix.yml`). Removed with its test
+  (`tests/test_select_nvidia_nim_model.py`). Not the same thing as
+  `contextual-orchestrator`'s own legitimate internal support for NVIDIA NIM
+  as one of its five backend providers, which is unmodified and out of this
+  repository's scope. Corrected four stale pre-ADR-0003 "single hardcoded
+  `NVIDIA_NIM_API_KEY`" descriptions of the scheduled autofix worker's model
+  credential in `docs/doctoring/hourly-nvidia-nim-autofix.md` (dated
+  addendum, extensive historical detail kept as history),
+  `docs/doctoring/originweave-hourly-review-caller.md`,
+  `docs/doctoring/nonnest2-hourly-review-caller.md` (one stale sentence
+  each), and `docs/automation/hourly-review-repair.md` (summary bullet and
+  the Clearfolio credential paragraph) to describe the current five-secret
+  `contextual-orchestrator` gateway routing instead. Split out of
+  `ContextualWisdomLab/.github#1437` per its exact-head governance review's
+  fifth acceptance criterion: this cleanup is orthogonal to that PR's
+  Strix pool-gating behavior repair and belongs in its own,
+  independently-reviewable change.
 - Raise `contextual_orchestrator_review_sidecar.sh`'s
   `ORCHESTRATOR_CATALOG_FAMILY_CAP` default from 4 to 8: root-caused the
   live "no provider route passed the Strix plain-chat preflight" outage

@@ -12,13 +12,15 @@ engine**.
   contextual-orchestrator, Inkspan, or another CWL service with an explicit
   repository and base branch.
 - `pr-review-autofix.yml` is the bounded write-capable worker. It uses OpenCode
-  with NVIDIA NIM and does not approve or merge pull requests.
+  routed through the vendored contextual-orchestrator gateway and does not
+  approve or merge pull requests.
 
-Orgmetra's caller remains provider-neutral. The intended model boundary is the
-contextual-orchestrator gateway: provider keys stay in its KV registry and
-automatic model discovery selects upstream models. A caller schedule is not
-evidence that gateway credentials, discovery, or a live OpenCode tool loop are
-available; those facts require exact worker-run evidence.
+Every caller, including Clearfolio's, is provider-neutral by design. The model
+boundary is the contextual-orchestrator gateway: provider keys stay in its KV
+registry and automatic model discovery selects upstream models. A caller
+schedule is not evidence that gateway credentials, discovery, or a live
+OpenCode tool loop are available; those facts require exact worker-run
+evidence.
 
 Merge eligibility remains owned by the separate merge scheduler, branch
 protection, required checks, independent review, and unresolved-thread policy.
@@ -44,8 +46,10 @@ not overlap its successor. At most one repair dispatch is created per run.
 
 The caller passes only the established `PR_REVIEW_MERGE_TOKEN` and
 `OPENCODE_APPROVE_TOKEN` scheduler credentials. It does not receive or forward
-`NVIDIA_NIM_API_KEY`; the model credential is scoped exclusively to the two
-OpenCode execution steps in the separately reviewed autofix worker.
+any of the five gateway provider secrets (`BYTEZ_API_KEY`, `NVIDIA_NIM_API_KEY`,
+`NVIDIA_NIM_API_KEY_SUB`, `OPENROUTER_API_KEY`, `OPENAI_API_KEY`); those stay
+scoped exclusively to the contextual-orchestrator sidecar provisioning step in
+the separately reviewed autofix worker.
 
 ## Orgmetra execution contract
 
@@ -199,7 +203,8 @@ organization-level queue inspection and bounded repair dispatch.
 When a scheduled run fails, classify the result before rerunning:
 
 - no actionable file-scoped feedback: expected no-op;
-- missing `NVIDIA_NIM_API_KEY`: central secret configuration failure;
+- missing gateway provider secrets or an unprovisioned contextual-orchestrator
+  sidecar: central secret/configuration failure;
 - head changed: safe optimistic-concurrency refusal; inspect the new head rather
   than retrying predecessor evidence;
 - out-of-scope or ignored-path change: treat as a security failure and preserve
@@ -225,7 +230,8 @@ Permanent tests prove:
 - the dispatch budget and same-head retry floor remain one;
 - caller and reusable-workflow secrets are explicit and never use
   `secrets: inherit`;
-- immutable source, NVIDIA-only model authentication, child-process credential
+- immutable source, gateway-only model authentication (no direct-provider
+  credential in the model child process), child-process credential
   stripping, live-head guards, and independent reviewer identity remain intact;
 - ordinary and conflict repair share the complete ignored-inclusive snapshot and
   NUL-delimited allowlist boundary;

@@ -1,5 +1,42 @@
 # Hourly NVIDIA NIM Review-Autofix Boundary
 
+## Update (2026-08-30): model execution migrated to the contextual-orchestrator gateway
+
+This record's "Decision", "Provider contract", and "Credential boundary"
+sections below describe the worker's design **as it stood before**
+[ADR-0003](../adr/0003-contextual-orchestrator-vendored-free-zdr.md)
+(2026-08-27): a single direct NVIDIA NIM provider
+(`https://integrate.api.nvidia.com/v1`), the `mistralai/mistral-small-4-119b-2603`
+writer model, and a lone `NVIDIA_NIM_API_KEY` credential bound to the two
+OpenCode execution steps. That description is now superseded, not deleted —
+kept below as the historical record of the write-scope/security work this
+file documents, most of which (the allowlist, snapshot/verification, hook
+suppression, and explicit push-destination controls) is provider-independent
+and unchanged.
+
+What actually changed: `pr-review-autofix.yml`'s OpenCode execution now
+routes through the vendored `contextual-orchestrator` gateway sidecar
+(`scripts/ci/contextual_orchestrator_review_sidecar.sh`), provisioned with
+all five provider secrets (`BYTEZ_API_KEY`, `NVIDIA_NIM_API_KEY`,
+`NVIDIA_NIM_API_KEY_SUB`, `OPENROUTER_API_KEY`, `OPENAI_API_KEY`) and the
+fail-closed zero-cost virtual model `contextual-orchestrator/orchestrator/free`
+for both the primary and small model. There is no direct
+`https://integrate.api.nvidia.com/v1` base URL, no hardcoded
+`mistralai/mistral-small-4-119b-2603` or `nvidia/nemotron-3-nano-30b-a3b`
+model id, and no bare `NVIDIA_API_KEY`/`NVIDIA_NIM_API_KEY` environment
+variable read anywhere in the workflow's execution steps — enforced by
+`tests/test_pr_review_autofix_nvidia_nim_contract.py::test_scheduled_autofix_routes_through_contextual_orchestrator`,
+which asserts the gateway strings are present and the direct-NVIDIA base URL
+and legacy model ids are absent. Separately,
+`scripts/ci/select_nvidia_nim_model.py` — a standalone helper that resolved a
+live NVIDIA NIM model id by calling `integrate.api.nvidia.com` directly —
+was found wired into nothing (no workflow or script referenced it, only its
+own test) and removed as orphaned dead code predating this migration.
+
+This does not change contextual-orchestrator's own internal use of NVIDIA NIM
+as one of its five auto-discovered backend providers; that remains entirely
+the gateway's concern, not this worker's.
+
 ## Decision
 
 Materialize accepts only exact SHA-256 pins or a bounded relative `-r` include; a lone `--require-hashes` line is not lock evidence.

@@ -74,9 +74,16 @@ OpenCode GitHub App token from GitHub OIDC when the mapped PAT is absent
 forbids executing the caller with write or model privileges it does not
 need (MITRE, 2026).
 
-Model execution remains inside the central worker. The model credential
-is the GitHub Secret `NVIDIA_NIM_API_KEY`; the caller does not receive or
-forward it.
+Model execution remains inside the central worker, which routes through the
+vendored `contextual-orchestrator` gateway sidecar
+(`scripts/ci/contextual_orchestrator_review_sidecar.sh`) using all five
+provider secrets (`BYTEZ_API_KEY`, `NVIDIA_NIM_API_KEY`,
+`NVIDIA_NIM_API_KEY_SUB`, `OPENROUTER_API_KEY`, `OPENAI_API_KEY`) and the
+fail-closed zero-cost `orchestrator/free` pool (see
+[ADR-0003](../adr/0003-contextual-orchestrator-vendored-free-zdr.md)); the
+caller receives none of them. (Updated 2026-08-30: this paragraph originally
+described a single hardcoded `NVIDIA_NIM_API_KEY` model credential, which
+predates the ADR-0003 gateway migration.)
 
 Before protected-master activation, the repository variable
 `OPENCODE_REPOSITORY_DISPATCH_TARGETS` must contain the exact

@@ -75,9 +75,16 @@ OpenCode GitHub App token from GitHub OIDC when the mapped PAT is absent
 forbids executing the caller with write or model privileges it does not
 need (MITRE, 2026).
 
-Model execution remains inside the central worker. The model credential
-is the GitHub Secret `NVIDIA_NIM_API_KEY`; the caller does not receive or
-forward it.
+Model execution remains inside the central worker, which routes through the
+vendored `contextual-orchestrator` gateway sidecar
+(`scripts/ci/contextual_orchestrator_review_sidecar.sh`) using all five
+provider secrets (`BYTEZ_API_KEY`, `NVIDIA_NIM_API_KEY`,
+`NVIDIA_NIM_API_KEY_SUB`, `OPENROUTER_API_KEY`, `OPENAI_API_KEY`) and the
+fail-closed zero-cost `orchestrator/free` pool (see
+[ADR-0003](../adr/0003-contextual-orchestrator-vendored-free-zdr.md)); the
+caller receives none of them. (Updated 2026-08-30: this paragraph originally
+described a single hardcoded `NVIDIA_NIM_API_KEY` model credential, which
+predates the ADR-0003 gateway migration.)
 
 Before protected-main activation, the repository variable
 `OPENCODE_REPOSITORY_DISPATCH_TARGETS` must contain the exact