feat: add Zro provider (OpenAI-compatible gateway)

.github/workflows/ci-fixer.yml

@@ -160,7 +160,7 @@ jobs:
           Failed log excerpt:
           EOF
             cat "$LOG_FILE"
-          } | opencode run --agent ci-fixer -m opencode/grok-4.5 | tee "$RESPONSE_FILE"
+          } | opencode run --agent ci-fixer -m opencode/muse-spark-1.3 --variant xhigh | tee "$RESPONSE_FILE"
 
       - name: Check changed paths
         if: steps.budget.outputs.run == 'true' && steps.budget-cache.outputs.cache-hit != 'true'

.github/workflows/deploy.yml

@@ -11,7 +11,7 @@ concurrency: ${{ github.workflow }}-${{ github.ref }}
 jobs:
   deploy:
     if: github.repository == 'anomalyco/models.dev'
-    runs-on: ubuntu-latest
+    runs-on: blacksmith-4vcpu-ubuntu-2404
     steps:
       - name: Checkout code
         uses: actions/checkout@v4

.github/workflows/issue-fixer.yml

@@ -60,7 +60,7 @@ jobs:
             + "If it is a feature request, a request to track a new kind of information, a question, or any miscellaneous non-catalog-data request, do not edit files. Respond briefly that it needs maintainer review and no automated fix was opened."
           ' "$ISSUE_FILE" > "$PROMPT_FILE"
 
-          opencode run --agent issue-fixer -m opencode/grok-4.5 --format json < "$PROMPT_FILE" | tee "$EVENTS_FILE"
+          opencode run --agent issue-fixer -m opencode/muse-spark-1.3 --variant xhigh --format json < "$PROMPT_FILE" | tee "$EVENTS_FILE"
 
           if ! jq -ers 'map(select(.type == "text") | .part.text) | last | select(length > 0)' "$EVENTS_FILE" > "$RESPONSE_FILE"; then
             echo "Issue fixer did not produce a final response." >&2

.github/workflows/pr-reviewer.yml

@@ -78,7 +78,7 @@ jobs:
           export PR_REVIEW_READY_FILE
           rm -f "$PR_REVIEW_READY_FILE"
 
-          opencode run --agent pr-reviewer -m opencode/grok-4.5 --format json <<'EOF' | tee "$EVENTS_FILE"
+          opencode run --agent pr-reviewer -m opencode/muse-spark-1.3 --variant xhigh --format json <<'EOF' | tee "$EVENTS_FILE"
           Review this pull request using the trusted reviewer instructions. Start with `.pr-review/pull-request.json`, `.pr-review/diff.patch`, `AGENTS.md`, and the contributing guidance in `README.md`. Read `sync.md`, the reasoning-options audit guide, schema code, and nearby base-revision files when relevant to the changed files. Use only the read, glob, grep, and mark-pr-ready tools. Return only the final review comment in the agent's required output format. Never include progress narration or passed-check summaries.
           EOF
 

.github/workflows/retry-issue-fixer.yml

@@ -0,0 +1,102 @@
+name: Retry stalled issue fixes
+
+on:
+  schedule:
+    - cron: "41 */6 * * *"
+  workflow_dispatch:
+
+permissions:
+  contents: write
+  issues: read
+  pull-requests: read
+
+concurrency:
+  group: retry-stalled-issue-fixes
+  cancel-in-progress: false
+
+jobs:
+  retry:
+    if: github.repository == 'anomalyco/models.dev'
+    runs-on: ubuntu-latest
+    steps:
+      - name: Redispatch oldest issues without pull requests
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          set -euo pipefail
+
+          cutoff="$(date -u -d '1 hour ago' '+%Y-%m-%dT%H:%M:%SZ')"
+          owner="${GITHUB_REPOSITORY%/*}"
+          repo="${GITHUB_REPOSITORY#*/}"
+
+          issues="$({
+            gh issue list \
+              --repo "$GITHUB_REPOSITORY" \
+              --state open \
+              --label automation \
+              --label model-sync \
+              --label missing-model \
+              --limit 1000 \
+              --json number,createdAt,labels
+          } | jq --arg cutoff "$cutoff" '
+            map(select(.createdAt <= $cutoff)) | sort_by(.createdAt)
+          ')"
+
+          dispatched=0
+          dispatched_issues=()
+          while IFS= read -r issue; do
+            number="$(jq -r '.number' <<< "$issue")"
+            provider="$(jq -r '
+              [.labels[].name | select(startswith("provider:")) | ltrimstr("provider:")]
+              | first // empty
+            ' <<< "$issue")"
+
+            if [ -z "$provider" ] || [ "$provider" = "openai" ]; then
+              continue
+            fi
+
+            closing_prs="$(gh api graphql \
+              -F owner="$owner" \
+              -F repo="$repo" \
+              -F number="$number" \
+              -f query='query($owner: String!, $repo: String!, $number: Int!) {
+                repository(owner: $owner, name: $repo) {
+                  issue(number: $number) {
+                    closedByPullRequestsReferences(first: 1, includeClosedPrs: true) {
+                      totalCount
+                    }
+                  }
+                }
+              }' \
+              --jq '.data.repository.issue.closedByPullRequestsReferences.totalCount')"
+
+            if [ "$closing_prs" -gt 0 ]; then
+              continue
+            fi
+
+            jq -n \
+              --arg provider "$provider" \
+              --argjson issue_number "$number" \
+              '{
+                event_type: "missing-model",
+                client_payload: {
+                  provider: $provider,
+                  issue_number: $issue_number
+                }
+              }' \
+              | gh api "repos/$GITHUB_REPOSITORY/dispatches" --method POST --input -
+
+            dispatched=$((dispatched + 1))
+            dispatched_issues+=("#$number")
+            if [ "$dispatched" -eq 3 ]; then
+              break
+            fi
+          done < <(jq -c '.[]' <<< "$issues")
+
+          if [ "$dispatched" -eq 0 ]; then
+            echo "No stalled automated issues were eligible for retry." | tee -a "$GITHUB_STEP_SUMMARY"
+          else
+            printf 'Redispatched %d issue(s): %s\n' \
+              "$dispatched" "${dispatched_issues[*]}" \
+              | tee -a "$GITHUB_STEP_SUMMARY"
+          fi

.github/workflows/sync-models.yml

@@ -15,7 +15,7 @@ concurrency: ${{ github.workflow }}-${{ github.ref }}
 jobs:
   providers:
     if: github.repository == 'anomalyco/models.dev'
-    runs-on: ubuntu-latest
+    runs-on: blacksmith-4vcpu-ubuntu-2404
     outputs:
       matrix: ${{ steps.providers.outputs.matrix }}
     steps:
@@ -41,7 +41,7 @@ jobs:
   sync:
     needs: providers
     if: github.repository == 'anomalyco/models.dev'
-    runs-on: ubuntu-latest
+    runs-on: blacksmith-4vcpu-ubuntu-2404
     strategy:
       fail-fast: false
       matrix: ${{ fromJSON(needs.providers.outputs.matrix) }}
@@ -77,12 +77,15 @@ jobs:
           DEEPINFRA_API_KEY: ${{ secrets.DEEPINFRA_API_KEY }}
           DIGITALOCEAN_API_TOKEN: ${{ secrets.DIGITALOCEAN_API_TOKEN }}
           DIGITALOCEAN_ACCESS_TOKEN: ${{ secrets.DIGITALOCEAN_ACCESS_TOKEN }}
+          FIREWORKS_API_KEY: ${{ secrets.FIREWORKS_API_KEY }}
           HF_TOKEN: ${{ secrets.HF_TOKEN }}
           OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
           VENICE_API_KEY: ${{ secrets.VENICE_API_KEY }}
           LLMGATEWAY_API_KEY: ${{ secrets.LLMGATEWAY_API_KEY }}
           MERGE_GATEWAY_API_KEY: ${{ secrets.MERGE_GATEWAY_API_KEY }}
+          MISTRAL_API_KEY: ${{ secrets.MISTRAL_API_KEY }}
+          NOVITA_AI_API_KEY: ${{ secrets.NOVITA_AI_API_KEY }}
           KILO_API_KEY: ${{ secrets.KILO_API_KEY }}
           GOOGLE_API_KEY: ${{ secrets.GOOGLE_API_KEY }}
           GEMINI_API_KEY: ${{ secrets.GEMINI_API_KEY }}

.opencode/agent/issue-fixer.md

@@ -25,15 +25,15 @@ Do not make code, schema, UI, documentation, or workflow changes. If the issue i
 
 When you do make a fix:
 
-- Follow `AGENTS.md` exactly (lab vs provider, **When to use `base_model`**, **Model fields**, **Reasoning options**, override-only hosts).
+- Follow `AGENTS.md` exactly (lab vs provider, **When to use `base_model`**, **Reasoning controls**, **Cost**, override-only hosts). Every model file is v2 (`schema = 2`).
 - Prefer the smallest correct change.
 - Verify every changed factual value against authoritative sources. Prefer first-party provider documentation, pricing pages, API references, model cards, or live provider catalog responses. Treat the issue as a lead, not sufficient verification by itself.
 - Do not broaden the issue's scope unless the additional changes are required for internal consistency and each one is independently verified.
 - Edit only `models/` and `providers/` TOML files.
 - If the host did not create the model: identify the lab model, **add** `models/<lab>/<model>.toml` when missing, then use `base_model`. Provider files are override-only — never restate identical description/modalities/structured_output/etc. Full inline only for first-party lab hosts or unique-to-host aliases per `AGENTS.md`.
-- Reasoning: classify first-party lab vs multi-model relay (**not** by npm). Copy the **lab/peer option set** for that model — do not force `low`/`medium`/`high` onto DeepSeek-style `high`/`max` (or other native sets). On relays, do not use `[]` from uncertainty when lab/peers have controls. No `toggle` beside effort that includes `none`. `toggle` + graded effort without `none` OK with a **leading top-of-file** wire comment. `budget_tokens` only per `AGENTS.md`. New lab `models/` files for inheritance must include dates, capability booleans, `limit`, and `modalities`.
-- Preserve provider-specific fields in provider TOMLs (`cost`, `reasoning_options`, `interleaved`, `status`, `provider`).
-- Costs are USD per million tokens; convert other currencies and note rate/date in a leading comment. Context bands use `[[cost.tiers]]`, never authored `context_over_200k`.
+- Reasoning: classify first-party lab vs multi-model relay (**not** by npm). Copy the **lab/peer option set** for that model — do not force `low`/`medium`/`high` onto DeepSeek-style `low`/`high`/`max` (or other native sets). On relays, do not leave out `capabilities.reasoning` controls from uncertainty when lab/peers have controls. No `toggle` beside effort that includes `none`. `toggle` + graded effort without `none` OK with a **leading top-of-file** wire comment. `budget` only per `AGENTS.md`. New lab `models/` files for inheritance must be complete: `release_date`, `open_weights`, `[capabilities]` (`tools`, `reasoning.supported`, `temperature`, `top_p`, `top_k`), `[limit]`, and `[modalities]`.
+- Preserve provider-specific fields in provider TOMLs (`[cost.*]`, `capabilities.reasoning` controls, cache, `status`, `[api.*]`, `experimental`).
+- Costs are USD per million tokens; convert other currencies and note rate/date in a leading comment. Context bands use `[[cost.text.input.overrides]]` / `[[cost.text.output.overrides]]` with `when = { context = { gt = … } }`.
 - Put durable source URLs in a leading TOML comment block when adding or changing factual data. Never put source comments between TOML sections because sync serialization removes them.
 - Do not run shell commands or use Bash. The workflow handles commits and pull request creation after you finish. Do not claim validation unless you actually performed it.
 

.opencode/agent/pr-reviewer.md

@@ -38,20 +38,21 @@ For model catalog changes, enforce these review rules:
 
 - Treat a missing compliant logo for a new provider as a merge blocker. The SVG must use `currentColor`, have no fixed size or hardcoded color, and preferably use a square `viewBox`.
 - Treat missing `base_model` as a merge blocker when the provider **did not create** the model (third-party / gateway host of a lab model). If `models/<lab>/<model>.toml` is missing but the lab model is nameable, the PR must **add** that lab entry and point `base_model` at it — full inline third-party definitions are a violation except unique-to-host / private-alias / first-party lab exceptions in `AGENTS.md`.
-- Treat **redundant `base_model` overrides** as a merge blocker: after `base_model`, the file must keep only provider-specific fields and real deltas. Flag restated identical `description`, `structured_output`, `modalities`, `tool_call`, `temperature`, dates, `family`, full copied `[limit]`/`[modalities]`, etc. Allowed always when needed: `cost`, `reasoning_options`, `interleaved`, `status`, `provider`, `experimental`, and genuine overrides (different name, limits, modalities, reasoning).
-- Treat missing `reasoning_options` on `reasoning = true` provider models as a merge blocker.
-- Apply **`AGENTS.md` → Reasoning options** and `.opencode/skills/audit-reasoning-options/SKILL.md` exactly.
+- Treat **redundant `base_model` overrides** as a merge blocker: after `base_model`, the file must keep only provider-specific fields and real deltas. Flag restated identical `description`, `modalities`, `capabilities.tools`, `temperature`, dates, `family`, full copied `[limit]`/`[modalities]`, etc. Allowed always when needed: `[cost.*]`, `capabilities.reasoning` controls, `structured_output`, cache, `status`, `[api.*]`, `experimental`, and genuine overrides (different name, limits, modalities).
+- Treat any model file without `schema = 2` as a merge blocker.
+- Treat missing `capabilities.reasoning` controls on reasoning provider models as a merge blocker, unless the model is affirmatively always-on with no caller control.
+- Apply **`AGENTS.md` → Reasoning controls** and `.opencode/skills/audit-reasoning-options/SKILL.md` exactly.
   - **Classify by host role, not npm:** first-party lab (provider is the model creator) vs multi-model relay. `@ai-sdk/openai-compatible` is used by both (DeepSeek/Alibaba are labs). Do not treat every openai-compatible host as a GPT gateway.
-  - **Baseline = lab + same-surface peer option set for that model**, not a fixed `low`/`medium`/`high`. GPT-style relays often use L/M/H; DeepSeek V4 is `toggle` + `high`/`max`; some Qwen paths are toggle + budget. Flag inventing L/M/H when lab/peers are narrower or different. Flag `[]` on a relay only from uncertainty when lab/peers expose controls.
+  - **Baseline = lab + same-surface peer option set for that model**, not a fixed `low`/`medium`/`high`. GPT-style relays often use L/M/H; DeepSeek V4 is `toggle` + `low`/`high`/`max`; some Qwen paths are toggle + budget. Flag inventing L/M/H when lab/peers are narrower or different. Flag missing controls on a relay only from uncertainty when lab/peers expose controls.
   - **`none` vs `toggle`:** violation only when `toggle` is paired with effort that already includes `none`. `toggle` + graded effort without `none` is valid when off is a separate wire control. Every `toggle` needs a leading top-of-file wire comment.
-  - **`budget_tokens`:** only real reasoning budgets (legacy Anthropic extended thinking, some Alibaba/Qwen, some older Gemini). Not GPT-5.x effort-only, Claude 4.7+ adaptive effort, DeepSeek V4. No min/max from `limit.output`/context.
+  - **`budget`:** only real reasoning budgets (legacy Anthropic extended thinking, some Alibaba/Qwen, some older Gemini). Not GPT-5.x effort-only, Claude 4.7+ adaptive effort, DeepSeek V4. No min/max from `limit.output`/context.
   - Do not treat Anthropic Messages and OpenAI chat-completions (or lab vs relay) as interchangeable control surfaces.
 - Do not treat absence of a sync module as a blocker. Recommend one only when a context-rich provider API can authoritatively populate model data or delete models no longer served.
 - Data-changing PRs should cite direct provider pricing, model documentation, or API references in the PR body. Missing citations are not by themselves a merge blocker, but should be reported as a low-severity request for evidence when material factual changes otherwise cannot be reviewed. Prefer first-party sources and require each citation to state what it supports.
 - You cannot fetch citation URLs. Assess whether citations are present, direct, and mapped to claims, but never claim you opened a URL or verified its contents. A URL or PR assertion alone does not prove a disputed value.
-- Source citations or rationale added to TOML files must be in a leading comment block above the first key because sync serialization removes comments elsewhere. A short adjacent comment that documents the exact provider request syntax for a reasoning option is allowed by `AGENTS.md`; do not confuse it with a source citation.
+- Source citations or rationale added to TOML files must be in a leading comment block above the first key because sync serialization removes comments elsewhere. A short adjacent comment that documents the exact provider request syntax for reasoning controls is allowed by `AGENTS.md`; do not confuse it with a source citation.
 - Model IDs come from filenames and must not be authored as `id` fields. The schema is strict, and required model capabilities, costs, limits, and modalities must be present either locally or through a valid `base_model`.
-- Review inherited values using the documented deep-merge rules. Arrays and primitives replace inherited values; plain objects merge; `base_model_omit` applies after merging; provider-specific fields such as `cost`, `reasoning_options`, `interleaved`, and `status` must remain provider-authored when needed. Costs must be USD/MTok (convert non-USD with a noted rate/date).
+- Review inherited values using the documented deep-merge rules. Arrays and primitives replace inherited values; plain objects merge; `base_model_omit` applies after merging; provider-specific fields such as `[cost.*]`, `capabilities.reasoning` controls, `[api.*]`, and `status` must remain provider-authored when needed. Costs must be USD/MTok (convert non-USD with a noted rate/date).
 - For sync changes, check authoritative deletion behavior, preservation of hand-authored and `base_model` fields, provider registration, focused scope, idempotence expectations, and the validation steps documented in `sync.md`.
 - For workflow changes, require third-party actions in new automation to be pinned to full commit SHAs, as documented in `sync.md`.