fix(ci): gate Strix's orchestrator/free access on live diversity evidence

.github/workflows/strix.yml

@@ -553,14 +553,14 @@ jobs:
       - name: Gate Strix secrets
         id: gate
         env:
-          STRIX_MODEL: contextual-orchestrator/orchestrator/free
+          STRIX_MODEL: contextual-orchestrator/orchestrator/auto
           STRIX_MODEL_REQUESTED: ${{ github.event.client_payload.strix_llm || '' }}
         run: |
           requested_model="$(printf '%s' "$STRIX_MODEL_REQUESTED" | sed 's/^[[:space:]]*//;s/[[:space:]]*$//')"
           case "$requested_model" in
-            ""|orchestrator/free|contextual-orchestrator/orchestrator/free) ;;
+            ""|orchestrator/auto|contextual-orchestrator/orchestrator/auto) ;;
             *)
-              echo '::error::Strix model overrides are limited to contextual-orchestrator/orchestrator/free.'
+              echo '::error::Strix model overrides are limited to contextual-orchestrator/orchestrator/auto.'
               exit 1
               ;;
           esac
@@ -578,11 +578,86 @@ jobs:
           OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
           CONTEXTUAL_ORCHESTRATOR_REQUIRE_ZDR: ${{ steps.target_visibility.outputs.is_private }}
-          CONTEXTUAL_ORCHESTRATOR_POOL: free
+          # Boot the sidecar with the richer "auto" catalog (free-first, with a
+          # genuine priced fallback tier) regardless of which model name Strix
+          # ends up requesting. This is required, not just conservative: if the
+          # sidecar booted "free"-only, no priced agents would ever be loaded,
+          # so a later request for "orchestrator/auto" would silently resolve
+          # to the exact same single-family free catalog under a different
+          # name -- a fake fallback that defeats the diversity gate below.
+          # free_account_diversity (read from this run's own policy report) is
+          # computed identically either way; see
+          # docs/adr/0020-strix-orchestrator-free-pool.md.
+          CONTEXTUAL_ORCHESTRATOR_POOL: auto
+          # CONTEXTUAL_ORCHESTRATOR_REQUIRE_MINIMUM_SERVING_DIVERSITY is
+          # deliberately NOT set here yet. The launcher supports it (see
+          # docs/adr/0020-strix-orchestrator-free-pool.md's "Runtime
+          # backstop" section), but analysis while building it found that
+          # today's live "auto" catalog's own PRIMARY (non-fallback) served
+          # set is drawn from the same free-first candidates as "free" mode
+          # -- so under the currently-documented single-family-dominated
+          # discovery snapshot, enabling this unconditionally would make the
+          # sidecar refuse to boot even into the safe orchestrator/auto
+          # fallback, which is a worse availability regression than what
+          # this ADR exists to prevent. Enabling it needs either confirmed
+          # live diversity >= 2, or a follow-up that scopes the check to the
+          # free-serving path specifically; tracked as a separate decision.
         run: |
           set -euo pipefail
           bash "$TRUSTED_STRIX_SOURCE/scripts/ci/contextual_orchestrator_review_sidecar.sh"
 
+      - name: Resolve Strix model from free-route diversity evidence
+        id: resolve_model
+        if: steps.gate.outputs.enabled == 'true'
+        env:
+          GATE_STRIX_MODEL: ${{ steps.gate.outputs.strix_model }}
+        run: |
+          set -euo pipefail
+          # docs/adr/0020-strix-orchestrator-free-pool.md: Strix may only select
+          # the fail-closed orchestrator/free pool when this run's own sidecar
+          # discovery reports at least two independently credentialed accounts
+          # among the discovered free routes (CONTEXTUAL_ORCHESTRATOR_EVIDENCE,
+          # written by scripts/ci/contextual_orchestrator_review_sidecar.sh).
+          # A single credential account means one account outage would black out
+          # required Strix security review -- exactly the 2026-08-29 finding
+          # ADR-0003's original orchestrator/auto pin existed to prevent. ANY
+          # uncertainty about that evidence (missing file, unreadable JSON, a
+          # missing or non-integer field) fails closed to the gate's own base
+          # model (orchestrator/auto) -- this step must never upgrade to
+          # orchestrator/free on unproven evidence.
+          diversity_threshold=2
+          free_account_diversity="$(
+            python3 - "${CONTEXTUAL_ORCHESTRATOR_EVIDENCE:-}" <<'PY'
+          import json
+          import sys
+
+          path = sys.argv[1] if len(sys.argv) > 1 else ""
+          try:
+              if not path:
+                  raise ValueError("CONTEXTUAL_ORCHESTRATOR_EVIDENCE is unset")
+              with open(path, encoding="utf-8") as handle:
+                  report = json.load(handle)
+              diversity = report["free_account_diversity"]
+              if isinstance(diversity, bool) or not isinstance(diversity, int) or diversity < 0:
+                  raise TypeError("free_account_diversity must be a non-negative integer")
+          except Exception as exc:  # noqa: BLE001 - fail closed on any evidence problem
+              print(
+                  f"::warning::Could not read a valid free_account_diversity from the "
+                  f"contextual-orchestrator policy report ({exc}); "
+                  "falling back to orchestrator/auto.",
+                  file=sys.stderr,
+              )
+              diversity = 0
+          print(diversity)
+          PY
+          )"
+          resolved_model="$GATE_STRIX_MODEL"
+          if [ "$free_account_diversity" -ge "$diversity_threshold" ]; then
+            resolved_model="contextual-orchestrator/orchestrator/free"
+          fi
+          echo "free_account_diversity=$free_account_diversity" >> "$GITHUB_OUTPUT"
+          echo "strix_model=$resolved_model" >> "$GITHUB_OUTPUT"
+
       - name: Set up Python
         if: steps.gate.outputs.enabled == 'true'
         uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
@@ -711,17 +786,26 @@ jobs:
       - name: Prepare Strix model input file
         if: steps.gate.outputs.enabled == 'true'
         env:
-          STRIX_MODEL: ${{ steps.gate.outputs.strix_model }}
+          # The diversity-gated resolution, not the gate's static base model:
+          # steps.resolve_model upgrades to orchestrator/free only when
+          # free_account_diversity >= 2 (see the "Resolve Strix model from
+          # free-route diversity evidence" step above and
+          # docs/adr/0020-strix-orchestrator-free-pool.md), otherwise it
+          # passes the gate's own orchestrator/auto straight through.
+          STRIX_MODEL: ${{ steps.resolve_model.outputs.strix_model }}
         run: |
           umask 077
           strix_llm_file="$RUNNER_TEMP/strix_llm.txt"
           strix_model="$(printf '%s' "$STRIX_MODEL" | sed 's/^[[:space:]]*//;s/[[:space:]]*$//')"
           case "$strix_model" in
+            orchestrator/auto | contextual-orchestrator/orchestrator/auto)
+              printf '%s' 'orchestrator/auto' > "$strix_llm_file"
+              ;;
             orchestrator/free | contextual-orchestrator/orchestrator/free)
               printf '%s' 'orchestrator/free' > "$strix_llm_file"
               ;;
             *)
-              echo '::error::STRIX_LLM must select contextual-orchestrator/orchestrator/free.'
+              echo '::error::STRIX_LLM must select contextual-orchestrator/orchestrator/auto or contextual-orchestrator/orchestrator/free.'
               exit 1
               ;;
           esac
@@ -744,8 +828,9 @@ jobs:
           LLM_API_KEY_FILE: ${{ env.LLM_API_KEY_FILE }}
           STRIX_TARGET_PATH: ${{ (github.event_name == 'pull_request_target' || github.event.client_payload.pr_number != '') && '__PR_SCOPE__' || './' }}
           STRIX_SOURCE_DIRS: ". backend frontend"
-          # The gateway auto pool is provider-diverse. Strix function tools
-          # must not send a provider-specific reasoning setting to every route.
+          # The gateway pool can route across multiple provider families
+          # (bounded by its per-family cap). Strix function tools must not
+          # send a provider-specific reasoning setting to every route.
           STRIX_REASONING_EFFORT: none
           STRIX_LLM_MAX_RETRIES: 1
           STRIX_TRANSIENT_RETRY_PER_MODEL: 2

AGENTS.md

@@ -21,12 +21,25 @@ sidecar (`scripts/ci/contextual_orchestrator_review_sidecar.sh`). The five
 provider secrets (`BYTEZ_API_KEY`, `NVIDIA_NIM_API_KEY`,
 `NVIDIA_NIM_API_KEY_SUB`, `OPENROUTER_API_KEY`, `OPENAI_API_KEY`) enter its KV
 as bootstrap transport in the same process that discovers models and serves;
-OpenCode, Noema, and Strix all use the fail-closed zero-cost pool
-`orchestrator/free`. Strix uses the zero-cost `orchestrator/free` pool by
-explicit 2026-08-30 owner decision, superseding the prior `orchestrator/auto`
-(provider-diverse, non-free-admitting) default; private targets still require
-ZDR-compliant routes under
+OpenCode and Noema use the fail-closed zero-cost pool `orchestrator/free`
+unconditionally. **Strix is evidence-gated, not unconditional:** `strix.yml`
+reads `free_account_diversity` (the count of independently credentialed
+accounts among all discovered free routes, reported by
+`scripts/ci/contextual_orchestrator_review_policy.py` on every discovery run)
+from the sidecar's policy report and selects `orchestrator/free` only when
+that count is `>= 2`; otherwise — including when the evidence is missing,
+unreadable, or malformed — it falls back to `orchestrator/auto`, never the
+other way around. A negative fixture
+(`tests/test_strix_contextual_orchestrator_contract.py`) pins that a
+diversity of 0 or 1 cannot weaken the resolved model to `orchestrator/free`.
+The `orchestrator/auto` provider-diverse, priced-fallback pool is therefore a
+permanent, load-bearing fallback for Strix, not a route being retired; it
+remains supported by the gateway policy for any other consumer too, admitting
+non-free routes only with complete published prompt/completion price and
+currency evidence. Private targets still require ZDR-compliant routes under
 [`scripts/ci/zdr_policy.py`](scripts/ci/zdr_policy.py).
 See [`docs/adr/0003-contextual-orchestrator-vendored-free-zdr.md`](docs/adr/0003-contextual-orchestrator-vendored-free-zdr.md)
-and its 2026-08-30 amendment.
+and [`docs/adr/0020-strix-orchestrator-free-pool.md`](docs/adr/0020-strix-orchestrator-free-pool.md)
+(the evidence-gated conditional between Strix's two pools, its residual risk,
+and the separately tracked request-time-failover dependency).
 The materialization contract is also covered by [`docs/doctoring/exact-artifact-sbom-attestation.md`](docs/doctoring/exact-artifact-sbom-attestation.md).

CHANGELOG.md

@@ -562,18 +562,6 @@ Semantic Versioning where the repository publishes a release.
   entry for the full evidence trail, the exact trade-off reasoned through
   (not live-verified, since this session lacks provider credentials), and
   the more complete fix if this proves insufficient.
-- Switch Strix from `orchestrator/auto` to `orchestrator/free`, matching
-  OpenCode and Noema: `strix.yml`'s `STRIX_MODEL`/`CONTEXTUAL_ORCHESTRATOR_POOL`
-  default and both model-override allowlists, and
-  `scripts/ci/strix_quick_gate.sh`'s `is_contextual_orchestrator_model`, now
-  accept only `orchestrator/free`. This is an explicit, informed owner
-  override of `docs/adr/0003-contextual-orchestrator-vendored-free-zdr.md`'s
-  original `orchestrator/auto` decision (see that ADR's 2026-08-30
-  amendment and the matching gap-baseline entry for the full trade-off and
-  evidence trail): Strix no longer has a paid-model fallback and can go
-  fully dark during the class of single-provider-family-collapse incident
-  the original decision was written to survive, until the free-catalog's
-  stale-model and provider-diversity gaps are separately closed.
 - Strengthen `scripts/ci/zdr_policy.py`'s `nvidia_nim`/`nvidia_nim_sub` ZDR
   attestation with a direct primary-source citation: NVIDIA's own current
   *NVIDIA API Trial Terms of Service* (v. September 19, 2025), Section