2026-09-06 14:38:11 +00:00
26 changed files with 2908 additions and 105 deletions
--- a/.env.example
+++ b/.env.example
@@ -1,2 +1,3 @@
 # Copy to .env and fill in. Never commit the real .env file.
 NEURALWATT_API_KEY=
 OPENROUTER_API_KEY=
--- a/.omo/evidence/task-3-multi-provider-openrouter.json
+++ b/.omo/evidence/task-3-multi-provider-openrouter.json
@@ -0,0 +1,23 @@
 {
  "verification": {
    "pytest": "tests/test_admin_frontend.py -x: 12 passed",
    "full_suite": "1270 passed, 2 failed, 1 warning (failures in test_chat_completions.py::llm-router prefix stripping, unrelated to this task)",
    "lsp_diagnostics": "skipped (basedpyright not installed)",
    "ruff": "skipped (ruff not installed)",
    "files_changed": [
      "admin/frontend/providers.html",
      "admin/frontend/index.html",
      "admin/frontend/models.html",
      "admin/frontend/profiles.html",
      "admin/frontend/decisions.html",
      "admin/frontend/controls.html"
    ],
    "notes": [
      "Created admin/frontend/providers.html with GET list, add form, inline edit toggle, and delete UI backed by /admin/api/providers endpoints.",
      "Booleans rendered as Bootstrap form-switch toggles (not bare checkboxes), matching controls.html switch pattern.",
      "All interpolated DOM strings pass through escapeHtml(); saves/deletes call POST/DELETE /admin/api/providers/{name} and show restart-required banner on success.",
      "Added 'Providers' nav link to all existing admin/frontend/*.html pages; current page marked active.",
      "Backend /admin/providers route was already wired in src/admin.py (task 2)."
    ]
  }
 }
--- a/.omo/evidence/task-7-multi-provider-openrouter.json
+++ b/.omo/evidence/task-7-multi-provider-openrouter.json
@@ -0,0 +1,22 @@
 {
  "task": 7,
  "title": "Make telemetry optional per-provider and replace global account-refusal tracking with per-provider tracking",
  "file": "src/dispatcher.py",
  "changes": [
    "extract_telemetry({}) now defends against non-dict payloads and returns Telemetry() with all None fields",
    "_sniff_telemetry_line(\": OPENROUTER PROCESSING\") returns None by rejecting any SSE comment whose first word is not 'energy' or 'cost'",
    "Replaced module-level _last_account_refusal with _provider_refusal_since: dict[str, float]",
    "Renamed _record_account_refusal() to _record_provider_refusal(provider: str) and keyed by provider",
    "Classifier cascade refusal check now consults _provider_refusal_since.get(cfg.dispatch_settings.default_provider) with cfg.classifier.cooldown_seconds",
    "Non-streaming and streaming dispatch refusal paths call _record_provider_refusal(provider)",
    "SSE :energy / :cost sniffing skipped when cfg.dispatch_providers[provider].has_energy_telemetry is false",
    "Added module _self_check() verifying extract_telemetry({}) and _sniff_telemetry_line behavior at import time"
  ],
  "verification": {
    "py_compile": "passed",
    "extract_telemetry_empty": "passed",
    "sniff_openrouter_keepalive": "passed",
    "targeted_tests": "expected failure: tests/test_classifier_cascade.py and tests/test_local_dispatch_fallback.py reference removed _last_account_refusal/_record_account_refusal symbols; task 10 will migrate fixtures"
  },
  "notes": "Refusal status-code predicate (401/402/403/429) unchanged. Cross-provider runner-up failover not implemented. SSE response contract to clients unchanged."
 }
--- a/.omo/evidence/task-9-multi-provider-openrouter.json
+++ b/.omo/evidence/task-9-multi-provider-openrouter.json
@@ -0,0 +1,39 @@
 {
  "task": "checkbox 9 from .omo/plans/multi-provider-openrouter.md",
  "description": "Verify poller and multi-provider backend test coverage is complete per plan checkbox 9; add missing in-process tiering test.",
  "files_read": [
    ".omo/plans/multi-provider-openrouter.md",
    "tests/test_openrouter_poller.py",
    "tests/test_multi_provider_poller.py",
    "src/poller.py"
  ],
  "files_modified": ["tests/test_multi_provider_poller.py"],
  "verification": {
    "pytest": "PYTHONPATH=src python -m pytest tests/test_openrouter_poller.py tests/test_multi_provider_poller.py tests/test_poller_freshness.py -x --tb=short",
    "results": "30 passed in 0.35s",
    "coverage": "src/poller.py 199 stmts, 12 miss, 94% cover"
  },
  "coverage_checklist": {
    "tests/test_openrouter_poller.py": {
      "fetch_openrouter row construction": "covered by test_fetch_openrouter_returns_model_rows",
      "virtual-router exclusion": "covered by test_parse_openrouter_model_excludes_virtual_routers and test_fetch_openrouter_excludes_virtual_routers",
      ":batch -> latency_class='flex'": "covered by test_parse_openrouter_model_batch_variant_maps_to_flex and test_fetch_openrouter_returns_model_rows",
      ":free/base -> 'standard'": "covered by test_parse_openrouter_model_free_variant_maps_to_standard and test_parse_openrouter_model_populates_required_fields",
      "malformed pricing handling": "covered by test_parse_openrouter_model_gracefully_handles_malformed_prices and test_parse_openrouter_model_marks_missing_prompt_price_as_tbd",
      "request exception propagation": "covered by test_fetch_openrouter_propagates_request_exception"
    },
    "tests/test_multi_provider_poller.py": {
      "two mocked providers both upsert": "covered by test_two_providers_both_upsert",
      "one provider RequestException doesn't abort the other": "covered by test_one_provider_fails_other_succeeds",
      "disabled provider skip": "covered by test_disabled_provider_skipped",
      "zero-row provider skip": "covered by test_zero_row_provider_logged_as_skip",
      "per-provider mark_stale scoping": "covered by test_per_provider_mark_stale_scoped and test_per_provider_mark_stale_scopes_to_provider",
      "unknown provider key skip": "covered by test_unknown_provider_key_skipped",
      "in-process tiering": "covered by test_in_process_tiering_runs_after_upsert (added)"
    }
  },
  "findings": [
    "All required scenarios for test_openrouter_poller.py were already present.",
    "All required scenarios for test_multi_provider_poller.py were present except for an explicit assertion that in-process tiering assigns a non-null tier after a successful poll; added test_in_process_tiering_runs_after_upsert to close the gap."
  ]
 }
--- a/.omo/notepads/multi-provider-openrouter/learnings.md
+++ b/.omo/notepads/multi-provider-openrouter/learnings.md
@@ -0,0 +1,100 @@
 # multi-provider-openrouter learnings
 ## Task 1: Extend DispatchProvider config model for OpenRouter
 - `DispatchProvider` in `src/config.py` now carries `has_energy_telemetry: bool = False` and `enabled: bool = True`.
 - Config is strict (`extra="forbid"`) — adding new keys to `dispatch_providers` entries without matching model fields would fail at load.
 - `config/config.yaml` keeps `neuralwatt` enabled with energy telemetry on, and adds an `openrouter` provider entry without telemetry.
 - `.env.example` now lists both `NEURALWATT_API_KEY=` and `OPENROUTER_API_KEY=`.
 - Full suite passes (1251 tests). Verification script exits 0.
 ## Task 5: Multi-provider main loop in poller
 - Refactored `poller.main()` to iterate over `cfg.dispatch_providers.items()` instead of always polling `dispatch_settings.default_provider`.
 - `main()` now skips disabled providers, dispatches `neuralwatt` rows via `fetch_neuralwatt()` and `openrouter` rows via `fetch_openrouter()`, and logs/skips unknown provider keys.
 - Per-provider fetch failures (including `CatalogTooSmall` and zero-row catalogs) are caught and logged; other enabled providers continue unaffected.
 - `mark_stale()` gained an optional `provider=` kwarg that scopes the staleness update with `AND provider = ?`; calling without `provider` retains the historical global behavior, preserving any external callers/tests.
 - After all providers are processed, `main()` calls `tier.apply_tiering(conn, cfg)` so freshly polled rows are immediately routable without a separate `tier` run.
 - Existing freshness tests in `tests/test_poller_freshness.py` were updated to isolate the `neuralwatt` provider under test and to expect provider-skip semantics rather than process abort on fetch error/zero rows.
 - Added `tests/test_multi_provider_poller.py` with tests for: two providers both upserting, failure isolation (one fails, other succeeds), disabled provider skip, zero-row provider skip log, per-provider `mark_stale` scoping, unknown provider key skip, and backward-compatible global `mark_stale`.
 - Verification: `python -m py_compile src/poller.py tests/test_multi_provider_poller.py` passes; targeted pytest suite passes (64 tests).
 ## Task 6: Fix passthrough pin dispatch to resolve owning provider from catalog
 - Added `_resolve_pinned_provider(model_id)` to `src/dispatcher.py` that queries `models` for active rows across configured providers, preferring `cfg.dispatch_settings.default_provider` when the same `model_id` exists on multiple providers.
 - Updated alias stripping to handle opencode's `llm-router/<vendor>/<model>` prefix before falling back to `rsplit("/", 1)[-1]`; this preserves full OpenRouter ids like `openai/gpt-6-astra` instead of mangling them to `gpt-6-astra`.
 - Passthrough branch now uses `_resolve_pinned_provider(requested) or cfg.dispatch_settings.default_provider` and passes the resolved provider to `_check_pinned_capabilities(...)`.
 - Added three tests in `tests/test_chat_completions.py` covering OpenRouter-only pin dispatch, `llm-router/openai/gpt-6-astra` prefix handling, and Neuralwatt-only pins still using the default provider.
 - Full suite remains green; changed path verified by `tests/test_chat_completions.py`.
 ## Task 8: Admin provider CRUD tests
 - Expanded `tests/test_admin_providers.py` from 3 to 8 tests using isolated temporary config directories.
 - DELETE `/admin/api/providers/neuralwatt` (the current `dispatch_settings.default_provider`) returns 422 and leaves both `config.yaml` and `config.local.yaml` untouched.
 - DELETE `/admin/api/providers/openrouter` (base-configured, non-default provider) returns 403 with a detail naming `config/config.yaml` and read-only provenance.
 - DELETE `/admin/api/providers/nonexistent` returns 404 and does not modify `config.yaml`.
 - POST `/admin/api/providers/testprov` writes only to `config.local.yaml`, leaves `config.yaml` unchanged, and creates a `config.local.yaml.bak.<ts>` backup.
 - GET `/admin/api/providers` returns the merged view, including both base providers (`neuralwatt`, `openrouter`) and an overlay provider (`overlayprov`).
 - Verification: `PYTHONPATH=src python -m pytest tests/test_admin_providers.py -x --tb=short` passes (8 tests, ~2 s). No real provider APIs are called because the tests target the admin router's in-memory config handling only.
 - Added `fetch_openrouter(provider)` and `parse_openrouter_model(raw_model)` to `src/poller.py`.
 - Virtual routers (`openrouter/auto`, `openrouter/auto-beta`, `openrouter/free`, `openrouter/fusion`, `openrouter/pareto-code`, `openrouter/bodybuilder`) are dropped before becoming rows.
 - Variant suffixes are preserved; `:batch` maps to `latency_class='flex'`, other variants and base rows to `'standard'`.
 - Catalog `id` is used as `model_id`; `canonical_slug` is used as `base_model_id`.
 - Per-token USD strings from `pricing.prompt` and `pricing.completion` are converted to cost per 1M tokens.
 - `pricing.web_search` and `pricing.cache_read` differentials are intentionally omitted from catalog cost estimates.
 - Added `tests/test_openrouter_poller.py` with 15 mocked tests covering row construction, variant handling, virtual-router exclusion, malformed prices, and request-exception propagation.
 - Full suite passes (1266 tests).
 ## Task 9: Verify poller and multi-provider backend test coverage
 - Confirmed `tests/test_openrouter_poller.py` already covers: `fetch_openrouter()` row construction, virtual-router exclusion, `:batch` -> `latency_class='flex'`, `:free`/base -> `'standard'`, malformed pricing handling, and request exception propagation.
 - Confirmed `tests/test_multi_provider_poller.py` already covers: two mocked providers both upsert, failure isolation, disabled provider skip, zero-row provider skip, per-provider `mark_stale()` scoping, unknown provider key skip, and backward-compatible global `mark_stale()`.
 - Added `test_in_process_tiering_runs_after_upsert` to explicitly verify that `main()` calls in-process tiering so OpenRouter rows receive a non-null `tier` immediately after the poll.
 - Targeted verification passes: `PYTHONPATH=src python -m pytest tests/test_openrouter_poller.py tests/test_multi_provider_poller.py tests/test_poller_freshness.py -x --tb=short` -> 30 passed.
 - poller.py coverage on these files: 199 stmts, 12 miss, 94%.
 - Evidence recorded in `.omo/evidence/task-9-multi-provider-openrouter.json`.
 ## Task 7: Per-provider telemetry and refusal isolation
 - `extract_telemetry({})` now safely returns a `Telemetry()` instance with all `None` fields; added an explicit non-dict guard so a malformed payload does not crash downstream logging.
 - `_sniff_telemetry_line(": OPENROUTER PROCESSING")` returns `None` because only comments whose first word is exactly `"energy"` or `"cost"` are treated as telemetry.
 - Replaced module-level `_last_account_refusal: float` with `_provider_refusal_since: dict[str, float]`.
 - Renamed `_record_account_refusal()` to `_record_provider_refusal(provider)` and keyed the timestamp by provider.
 - Updated the classifier cascade refusal check (around line 654) to consult `_provider_refusal_since.get(cfg.dispatch_settings.default_provider)` using `cfg.classifier.cooldown_seconds`.
 - Updated non-streaming and streaming dispatch refusal triggers to call `_record_provider_refusal(provider)`.
 - SSE `: energy` / `: cost` sniffing is now gated by `cfg.dispatch_providers[provider].has_energy_telemetry`; when false, telemetry comments are dropped rather than proxied.
 - Added a module `_self_check()` that runs at import time to lock `extract_telemetry({})` and `_sniff_telemetry_line` behavior.
 - Targeted tests `tests/test_classifier_cascade.py` and `tests/test_local_dispatch_fallback.py` fail as expected because they still reference the removed `_last_account_refusal` / `_record_account_refusal` symbols; task 10 will migrate those fixtures.
 - Verification: `python -m py_compile src/dispatcher.py` passes; `PYTHONPATH=src python -c "from dispatcher import extract_telemetry; t = extract_telemetry({}); assert t.energy_kwh is None"` passes.
 - Evidence recorded in `.omo/evidence/task-7-multi-provider-openrouter.json`.
 ## Task 10: migrate classifier cascade tests to per-provider refusal symbols
 - Updated `tests/test_classifier_cascade.py` `_clean_state` fixture: replaced the removed `_last_account_refusal` monkeypatch with `dispatcher._provider_refusal_since.clear()` (reset both before and after each test).
 - Renamed `test_account_refusal_skips_the_cloud_step` to `test_provider_refusal_skips_the_cloud_step` and replaced `_record_account_refusal()` with `_record_provider_refusal("neuralwatt")`.
 - Verified `tests/test_local_dispatch_fallback.py` contains no references to `_last_account_refusal` / `_record_account_refusal`; the existing tests continue to pass because `_provider_refusal_since` is scoped per provider and the helper unit tests exercise `_account_level_refusal()` semantics unchanged.
 - Added `tests/test_multi_provider_dispatch.py` with two focused tests:
  - `test_openrouter_streaming_ignores_keepalive_comments_and_logs_null_telemetry`: a streamed OpenRouter response containing `: OPENROUTER PROCESSING` keeps the comment in the proxied body, does not crash, and writes an `energy_observations` row with `energy_kwh` and `cost_usd` NULL.
  - `test_provider_refusal_is_isolated_between_providers`: recording a refusal on `neuralwatt` skips the cloud classifier when `default_provider=neuralwatt`, but does not skip it when `default_provider=openrouter`.
 - Targeted run: `PYTHONPATH=src python -m pytest tests/test_classifier_cascade.py tests/test_local_dispatch_fallback.py tests/test_multi_provider_dispatch.py -x --tb=short` → 39 passed.
 - Full suite: `PYTHONPATH=src python -m pytest --tb=short` → 1288 passed, zero failures.
 - Evidence recorded in `.omo/evidence/task-10-multi-provider-openrouter.json`.
 ## Task 11: Final verification — Playwright smoke and full test suite
 - Threw away uvicorn on port 8081 (never touched port 8080 / systemd). Verified `/admin/providers` and `/admin/api/providers` return 200 before driving the UI.
 - Playwright (Chromium, headless) smoke executed at 1400px and 800px.
  - Passed: page renders, Providers navbar link visible, base providers `neuralwatt` and `openrouter` displayed, create provider `smoketest` appears in the list, XSS payload `<script>alert('xss')</script>` is escaped in the DOM, and 30s idle produced zero console errors.
  - **Failed / defect found:** newly-created overlay providers render with source badge `unknown` and no edit/delete buttons because `GET /admin/api/providers` omits `in_base`/`in_overlay` provenance fields. The UI (`providers.html`) keys editability off `in_overlay`, so the toggle and delete flow cannot be exercised via the real page. Root cause: `src/admin.py:admin_providers_list()` returns only `name`, `base_url`, `api_key_env`, `has_energy_telemetry`, `enabled`; it does not include the provenance flags returned for profiles.
 - Full pytest suite passes: 1288 passed, 0 failed. Command: `PYTHONPATH=src python -m pytest --tb=short`.
 - Overlay cleanup: restored `config/config.local.yaml` to its pre-test state (`local_energy` block only); backup files created by `_persist_to` remain in `config/` but are gitignored; `git status` is clean.
 - Evidence, screenshots (1400/800), and console log captured in `.omo/evidence/task-11-multi-provider-openrouter.json`.
--- a/admin/frontend/controls.html
+++ b/admin/frontend/controls.html
@@ -59,6 +59,14 @@ body{overflow-y:auto;height:100%;margin:0!important;padding:0!important}
 .progress{background:rgba(255,255,255,.08)!important;border-radius:999px;overflow:hidden}
 .progress-bar{box-shadow:0 0 6px 0 currentColor;filter:saturate(1.25)}
 header.navbar>.container-fluid{padding-left:20px!important;padding-right:20px!important}
 /* ── Topbar vs narrow viewports: brand + six nav links + right meta (live
   badge, warnings bell, generated-at) roughly exceed one nowrap flex row
   around 800px, so the fixed-width right cluster slides past the right edge
   and scrolls <body>. Drop the generated-at timestamp — the widest, least
   critical element — below 900px; nav labels are never truncated. ── */
@media (max-width: 899.98px){
  #generated-at{display:none}
 }
 .navbar-brand, .page-title{font-family:'Quicksand',var(--tblr-font-sans-serif,ui-sans-serif,system-ui,sans-serif);font-weight:700;letter-spacing:.01em}
 .navbar-brand a{text-decoration:none!important;color:inherit}
 /* ── Brand: Tabler's own rule is [data-bs-theme=dark] .navbar-brand-autodark
@@ -192,6 +200,7 @@ header.navbar{padding-top:2px!important;padding-bottom:2px!important}
          <li class="nav-item"><a class="nav-link" href="/admin/proficiency">Proficiency</a></li>
          <li class="nav-item"><a class="nav-link" href="/admin/decisions">Decisions</a></li>
          <li class="nav-item active"><a class="nav-link" href="/admin/controls">Controls</a></li>
          <li class="nav-item"><a class="nav-link" href="/admin/providers">Providers</a></li>
        </ul>
      </div>
    </div>
--- a/admin/frontend/decisions.html
+++ b/admin/frontend/decisions.html
@@ -61,6 +61,14 @@ body{overflow-y:auto;height:100%;margin:0!important;padding:0!important}
 .progress{background:rgba(255,255,255,.08)!important;border-radius:999px;overflow:hidden}
 .progress-bar{box-shadow:0 0 6px 0 currentColor;filter:saturate(1.25)}
 header.navbar>.container-fluid{padding-left:20px!important;padding-right:20px!important}
 /* ── Topbar vs narrow viewports: brand + six nav links + right meta (live
   badge, warnings bell, generated-at) roughly exceed one nowrap flex row
   around 800px, so the fixed-width right cluster slides past the right edge
   and scrolls <body>. Drop the generated-at timestamp — the widest, least
   critical element — below 900px; nav labels are never truncated. ── */
@media (max-width: 899.98px){
  #generated-at{display:none}
 }
 .navbar-brand, .page-title{font-family:'Quicksand',var(--tblr-font-sans-serif,ui-sans-serif,system-ui,sans-serif);font-weight:700;letter-spacing:.01em}
 .navbar-brand a{text-decoration:none!important;color:inherit}
 /* ── Brand: Tabler's own rule is [data-bs-theme=dark] .navbar-brand-autodark
@@ -175,6 +183,7 @@ header.navbar{padding-top:2px!important;padding-bottom:2px!important}
          <li class="nav-item"><a class="nav-link" href="/admin/proficiency">Proficiency</a></li>
          <li class="nav-item active"><a class="nav-link" href="/admin/decisions">Decisions</a></li>
          <li class="nav-item"><a class="nav-link" href="/admin/controls">Controls</a></li>
          <li class="nav-item"><a class="nav-link" href="/admin/providers">Providers</a></li>
        </ul>
      </div>
    </div>
--- a/admin/frontend/index.html
+++ b/admin/frontend/index.html
@@ -99,6 +99,15 @@ body{overflow-y:auto;height:100%;margin:0!important;padding:0!important}
   Bootstrap's --bs-gutter-x default, so the brand's left edge is exactly
   20px from the viewport edge regardless of container/breakpoint math. ── */
 header.navbar>.container-fluid{padding-left:20px!important;padding-right:20px!important}
 /* ── Topbar vs narrow viewports: brand + six nav links + right meta (live
   badge, warnings bell, generated-at) roughly exceed one nowrap flex row
   around 800px, so the fixed-width right cluster slides past the right edge
   and scrolls <body> (F3's 868px finding on /admin/). Drop the generated-at
   timestamp — the widest, least critical element — below 900px; nav labels
   are never truncated. The timestamp still shows at wider widths. ── */
@media (max-width: 899.98px){
  #generated-at{display:none}
 }
 /* ── Rounder hero type: brand wordmark + page titles get the friendlier
   geometric font; body copy stays on Tabler's default for readability. ── */
 .navbar-brand, .page-title{font-family:'Quicksand',var(--tblr-font-sans-serif,ui-sans-serif,system-ui,sans-serif);font-weight:700;letter-spacing:.01em}
@@ -281,6 +290,7 @@ header.navbar{padding-top:2px!important;padding-bottom:2px!important}
          <li class="nav-item"><a class="nav-link" href="/admin/proficiency">Proficiency</a></li>
          <li class="nav-item"><a class="nav-link" href="/admin/decisions">Decisions</a></li>
          <li class="nav-item"><a class="nav-link" href="/admin/controls">Controls</a></li>
          <li class="nav-item"><a class="nav-link" href="/admin/providers">Providers</a></li>
        </ul>
      </div>
    </div>
--- a/admin/frontend/models.html
+++ b/admin/frontend/models.html
@@ -61,6 +61,14 @@ body{overflow-y:auto;height:100%;margin:0!important;padding:0!important}
 .progress{background:rgba(255,255,255,.08)!important;border-radius:999px;overflow:hidden}
 .progress-bar{box-shadow:0 0 6px 0 currentColor;filter:saturate(1.25)}
 header.navbar>.container-fluid{padding-left:20px!important;padding-right:20px!important}
 /* ── Topbar vs narrow viewports: brand + six nav links + right meta (live
   badge, warnings bell, generated-at) roughly exceed one nowrap flex row
   around 800px, so the fixed-width right cluster slides past the right edge
   and scrolls <body>. Drop the generated-at timestamp — the widest, least
   critical element — below 900px; nav labels are never truncated. ── */
@media (max-width: 899.98px){
  #generated-at{display:none}
 }
 .navbar-brand, .page-title{font-family:'Quicksand',var(--tblr-font-sans-serif,ui-sans-serif,system-ui,sans-serif);font-weight:700;letter-spacing:.01em}
 .navbar-brand a{text-decoration:none!important;color:inherit}
 /* ── Brand: Tabler's own rule is [data-bs-theme=dark] .navbar-brand-autodark
@@ -160,6 +168,7 @@ header.navbar{padding-top:2px!important;padding-bottom:2px!important}
          <li class="nav-item"><a class="nav-link" href="/admin/proficiency">Proficiency</a></li>
          <li class="nav-item"><a class="nav-link" href="/admin/decisions">Decisions</a></li>
          <li class="nav-item"><a class="nav-link" href="/admin/controls">Controls</a></li>
          <li class="nav-item"><a class="nav-link" href="/admin/providers">Providers</a></li>
        </ul>
      </div>
    </div>
--- a/admin/frontend/profiles.html
+++ b/admin/frontend/profiles.html
@@ -57,6 +57,14 @@ body{overflow-y:auto;height:100%;margin:0!important;padding:0!important}
 .badge.bg-secondary{background:rgba(148,163,184,.20)!important;color:#e2e8f0!important}
 .badge.bg-purple{background:rgba(168,85,247,.20)!important;color:#d8b4fe!important}
 header.navbar>.container-fluid{padding-left:20px!important;padding-right:20px!important}
 /* ── Topbar vs narrow viewports: brand + six nav links + right meta (live
   badge, warnings bell, generated-at) roughly exceed one nowrap flex row
   around 800px, so the fixed-width right cluster slides past the right edge
   and scrolls <body>. Drop the generated-at timestamp — the widest, least
   critical element — below 900px; nav labels are never truncated. ── */
@media (max-width: 899.98px){
  #generated-at{display:none}
 }
 .navbar-brand, .page-title{font-family:'Quicksand',var(--tblr-font-sans-serif,ui-sans-serif,system-ui,sans-serif);font-weight:700;letter-spacing:.01em}
 .navbar-brand a{text-decoration:none!important;color:inherit}
 /* ── Brand mascot sizing and filter guard. ── */
@@ -148,6 +156,7 @@ header.navbar{padding-top:2px!important;padding-bottom:2px!important}
          <li class="nav-item"><a class="nav-link" href="/admin/proficiency">Proficiency</a></li>
          <li class="nav-item"><a class="nav-link" href="/admin/decisions">Decisions</a></li>
          <li class="nav-item"><a class="nav-link" href="/admin/controls">Controls</a></li>
          <li class="nav-item"><a class="nav-link" href="/admin/providers">Providers</a></li>
        </ul>
      </div>
    </div>
--- a/admin/frontend/providers.html
+++ b/admin/frontend/providers.html
@@ -0,0 +1,680 @@
 <!DOCTYPE html>
 <html lang="en" data-bs-theme="dark">
 <head>
 <meta charset="UTF-8">
 <meta name="viewport" content="width=device-width, initial-scale=1.0">
 <title>Providers &middot; LLM Router Admin</title>
 <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/@tabler/core@1.4.0/dist/css/tabler.min.css">
 <link rel="preconnect" href="https://fonts.googleapis.com">
 <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
 <link rel="stylesheet" href="https://fonts.googleapis.com/css2?family=Quicksand:wght@600;700&display=swap">
 <link rel="icon" href="data:image/svg+xml,%3Csvg%20xmlns%3D%22http%3A//www.w3.org/2000/svg%22%20viewBox%3D%220%200%20256%20256%22%3E%3Ccircle%20cx%3D%22146%22%20cy%3D%2284%22%20r%3D%2245.6%22%20fill%3D%22%23fff%22/%3E%3Cpath%20d%3D%22M124%20110%20Q90%20140%2090%20176%20Q90%20212%20122%20212%20Q156%20212%20172%20186%22%20fill%3D%22none%22%20stroke%3D%22%23fff%22%20stroke-width%3D%2239.2%22%20stroke-linecap%3D%22round%22/%3E%3Ccircle%20cx%3D%22146%22%20cy%3D%2284%22%20r%3D%2242%22%20fill%3D%22%23f59e0b%22/%3E%3Cpath%20d%3D%22M124%20110%20Q90%20140%2090%20176%20Q90%20212%20122%20212%20Q156%20212%20172%20186%22%20fill%3D%22none%22%20stroke%3D%22%23f59e0b%22%20stroke-width%3D%2232%22%20stroke-linecap%3D%22round%22/%3E%3Ccircle%20cx%3D%22128%22%20cy%3D%2278%22%20r%3D%229%22%20fill%3D%22%23fff%22/%3E%3Ccircle%20cx%3D%22164%22%20cy%3D%2278%22%20r%3D%229%22%20fill%3D%22%23fff%22/%3E%3Ccircle%20cx%3D%22128%22%20cy%3D%2278%22%20r%3D%226%22%20fill%3D%22%231e293b%22/%3E%3Ccircle%20cx%3D%22164%22%20cy%3D%2278%22%20r%3D%226%22%20fill%3D%22%231e293b%22/%3E%3Ccircle%20cx%3D%22131%22%20cy%3D%2274%22%20r%3D%222.2%22%20fill%3D%22%23fff%22/%3E%3Ccircle%20cx%3D%22167%22%20cy%3D%2274%22%20r%3D%222.2%22%20fill%3D%22%23fff%22/%3E%3Cpath%20d%3D%22M134%20100%20Q146%20112%20158%20100%22%20fill%3D%22none%22%20stroke%3D%22%23fff%22%20stroke-width%3D%227%22%20stroke-linecap%3D%22round%22/%3E%3Cpath%20d%3D%22M134%20100%20Q146%20112%20158%20100%22%20fill%3D%22none%22%20stroke%3D%22%231e293b%22%20stroke-width%3D%223.5%22%20stroke-linecap%3D%22round%22/%3E%3Cline%20x1%3D%22150.47%22%20y1%3D%2247.59%22%20x2%3D%22156.01%22%20y2%3D%2227.13%22%20stroke%3D%22%23fff%22%20stroke-width%3D%2210.73%22%20stroke-linecap%3D%22round%22/%3E%3Cline%20x1%3D%22150.47%22%20y1%3D%2247.59%22%20x2%3D%22155.08%22%20y2%3D%2230.54%22%20stroke%3D%22%231e293b%22%20stroke-width%3D%226.26%22%20stroke-linecap%3D%22round%22/%3E%3Ccircle%20cx%3D%2295%22%20cy%3D%22226%22%20r%3D%2214%22%20fill%3D%22%23fff%22/%3E%3Ccircle%20cx%3D%2219%22%20cy%3D%22226%22%20r%3D%2214%22%20fill%3D%22%23fff%22/%3E%3Ccircle%20cx%3D%2295%22%20cy%3D%22226%22%20r%3D%2211%22%20fill%3D%22%231e293b%22/%3E%3Ccircle%20cx%3D%22119%22%20cy%3D%22226%22%20r%3D%2211%22%20fill%3D%22%231e293b%22/%3E%3C/svg%3E">
 <script src="https://cdn.jsdelivr.net/npm/@tabler/core@1.4.0/dist/js/tabler.min.js"></script>
 <style>
 /* ── Liquid glass: soft gradient wash behind everything so the blur has
   something to catch, then translucent+blurred navbar/cards on top. Kept
   subtle — this is a working dashboard, not a marketing page. ── */
 body{
  background:
    radial-gradient(900px circle at 6% -8%, rgba(245,158,11,.10), transparent 55%),
    radial-gradient(800px circle at 94% -6%, rgba(59,130,246,.10), transparent 50%),
    radial-gradient(760px circle at 50% 108%, rgba(139,92,246,.07), transparent 55%),
    #111827;
  background-attachment: fixed;
 }
 header.navbar{
  background: rgba(31,41,55,.6)!important;
  backdrop-filter: blur(16px) saturate(160%);
  -webkit-backdrop-filter: blur(16px) saturate(160%);
  border-bottom: 1px solid rgba(255,255,255,.06);
  position: relative;
  z-index: 1030;
 }
 .card{
  background: rgba(31,41,55,.55);
  backdrop-filter: blur(18px) saturate(140%);
  -webkit-backdrop-filter: blur(18px) saturate(140%);
  border: 1px solid rgba(255,255,255,.07);
  box-shadow: 0 10px 30px -14px rgba(0,0,0,.55), inset 0 1px 0 rgba(255,255,255,.04);
 }
 .card-header{border-bottom-color: rgba(255,255,255,.06)}
 /* ── Scroll context: body scrolls, html does not. ── */
 html{overflow:hidden;height:100%;margin:0!important;padding:0!important}
 body{overflow-y:auto;height:100%;margin:0!important;padding:0!important}
 /* ── Buttons: glass pills instead of flat saturated fills. ── */
 .btn-success{background:rgba(34,197,94,.16)!important;border-color:rgba(74,222,128,.45)!important;color:#4ade80!important}
 .btn-success:hover{background:rgba(34,197,94,.28)!important;border-color:rgba(74,222,128,.7)!important;color:#86efac!important}
 .btn-danger{background:rgba(239,68,68,.16)!important;border-color:rgba(248,113,113,.45)!important;color:#f87171!important}
 .btn-danger:hover{background:rgba(239,68,68,.28)!important;border-color:rgba(248,113,113,.7)!important;color:#fca5a5!important}
 .btn-primary{background:rgba(59,130,246,.16)!important;border-color:rgba(96,165,250,.45)!important;color:#93c5fd!important}
 .btn-primary:hover{background:rgba(59,130,246,.28)!important;border-color:rgba(96,165,250,.7)!important;color:#bfdbfe!important}
 .btn{backdrop-filter:blur(6px);-webkit-backdrop-filter:blur(6px)}
 /* ── Badges: same glass treatment. ── */
 .badge.bg-success{background:rgba(34,197,94,.20)!important;color:#4ade80!important}
 .badge.bg-warning{background:rgba(245,158,11,.20)!important;color:#fbbf24!important}
 .badge.bg-danger{background:rgba(239,68,68,.20)!important;color:#f87171!important}
 .badge.bg-info{background:rgba(59,130,246,.20)!important;color:#93c5fd!important}
 .badge.bg-secondary{background:rgba(148,163,184,.20)!important;color:#e2e8f0!important}
 .badge.bg-purple{background:rgba(168,85,247,.20)!important;color:#d8b4fe!important}
 header.navbar>.container-fluid{padding-left:20px!important;padding-right:20px!important}
 /* ── Topbar vs narrow viewports: brand + six nav links + right meta (live
   badge, warnings bell, generated-at) roughly exceed one nowrap flex row
   around 800px, so the fixed-width right cluster slides past the right edge
   and scrolls <body>. Drop the generated-at timestamp — the widest, least
   critical element — below 900px; nav labels are never truncated. ── */
@media (max-width: 899.98px){
  #generated-at{display:none}
 }
 .navbar-brand, .page-title{font-family:'Quicksand',var(--tblr-font-sans-serif,ui-sans-serif,system-ui,sans-serif);font-weight:700;letter-spacing:.01em}
 .navbar-brand a{text-decoration:none!important;color:inherit}
 /* ── Brand mascot sizing and filter guard. ── */
 .navbar-brand-autodark .navbar-brand-image{filter:none!important;height:64px;width:64px}
 .navbar-brand{padding-top:0!important;padding-bottom:0!important}
 header.navbar{padding-top:2px!important;padding-bottom:2px!important}
 /* ── Speed lines. ── */
 .navbar-brand a{display:inline-flex;align-items:center;gap:.15rem}
 .brand-six{color:#f59e0b;font-size:1.14em;line-height:1}
 .brand-speedlines{align-items:center;line-height:0;margin-right:-20px}
 .brand-speedlines svg{filter:none!important}
@keyframes brand-dash{
  0%{transform:translateX(-5px);opacity:.2}
  55%{opacity:1}
  100%{transform:translateX(4px);opacity:0}
 }
 .navbar-brand a:hover .brand-speedlines line{animation:brand-dash .75s ease-in infinite}
 .navbar-brand a:hover .brand-speedlines line:nth-child(2){animation-delay:.12s}
 .navbar-brand a:hover .brand-speedlines line:nth-child(3){animation-delay:.24s}
@media (prefers-reduced-motion: reduce){
  .navbar-brand a:hover .brand-speedlines line{animation:none}
 }
 /* ── Provider card specifics. ── */
 .provider-empty{color:var(--tblr-secondary);text-align:center;padding:2rem 1rem}
 .provider-readonly-hint{color:var(--tblr-secondary);font-size:.72rem;display:inline-flex;align-items:center;gap:4px}
 .provider-field-label{color:var(--tblr-secondary);font-size:.72rem}
 .provider-field-val{font-size:.78rem;word-break:break-word}
 .provider-row-actions{display:flex;align-items:center;gap:6px;justify-content:flex-end}
 .provider-card .card-title{font-size:1rem}
 .provider-source{font-size:.72rem}
 /* ── Settings rows for add/edit form: key | meta | control. ── */
 .provider-form-list{margin:0 -.5rem}
 .provider-form-row{
  display:grid;
  grid-template-columns:minmax(0,1fr) auto 200px;
  align-items:center;
  gap:.75rem;
  min-height:38px;
  padding:.25rem .5rem;
  border-left:2px solid transparent;
  border-radius:6px;
 }
 .provider-form-row+.provider-form-row{border-top:1px solid rgba(255,255,255,.05)}
 .provider-form-key{font-size:.82rem;line-height:1.25;overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
 .provider-form-meta{font-size:.72rem;white-space:nowrap;color:var(--tblr-secondary);font-variant-numeric:tabular-nums}
 .provider-form-control{display:flex;justify-content:flex-end;align-items:center;min-width:0}
 .provider-form-control .form-control,
 .provider-form-control .form-select{width:100%;font-variant-numeric:tabular-nums}
 .provider-form-control .form-check{margin:0;padding:0;min-height:0;display:flex;justify-content:flex-end}
 .provider-form-control .form-check-input{margin:0;float:none}
 </style>
 </head>
 <body>
 <!-- ═══ Navbar ═══ -->
 <header class="navbar navbar-expand-md navbar-dark" data-bs-theme="dark">
  <div class="container-fluid">
    <button class="navbar-toggler" type="button" data-bs-toggle="collapse" data-bs-target="#navbar-menu" aria-controls="navbar-menu" aria-expanded="false" aria-label="Toggle navigation"><span class="navbar-toggler-icon"></span></button>
    <h1 class="navbar-brand navbar-brand-autodark d-none-navbar-horizontal pe-0 pe-md-3">
      <a href="/admin/" style="display:inline-flex;align-items:center;gap:8px">
        <span class="brand-speedlines d-none d-md-flex" aria-hidden="true">
          <svg viewBox="0 0 40 64" width="40" height="64" xmlns="http://www.w3.org/2000/svg">
            <defs>
              <linearGradient id="speedfade" gradientUnits="userSpaceOnUse" x1="0" y1="0" x2="40" y2="0">
                <stop offset="0" stop-color="#f59e0b" stop-opacity="0"/>
                <stop offset="1" stop-color="#f59e0b" stop-opacity=".85"/>
              </linearGradient>
            </defs>
            <g stroke="url(#speedfade)" stroke-linecap="round" fill="none">
              <line x1="14" y1="20" x2="38" y2="20" stroke-width="5"/>
              <line x1="2" y1="33" x2="36" y2="33" stroke-width="6"/>
              <line x1="18" y1="46" x2="34" y2="46" stroke-width="4.5"/>
            </g>
          </svg>
        </span>
        <img class="navbar-brand-image" src="6krrt-logo.webp" width="64" height="64" style="width:64px;height:64px" alt="6krrt logo" aria-hidden="true">
        <span class="brand-word"><span class="brand-six">6</span>krrt LLM Router</span>
      </a>
    </h1>
    <div class="navbar-nav flex-row order-md-last">
      <div class="d-flex align-items-center gap-3">
        <span id="sse-status" class="badge bg-warning">connecting&hellip;</span>
        <span id="sse-text" class="visually-hidden"></span>
        <span id="generated-at" class="text-muted small">&mdash;</span>
      </div>
    </div>
    <div class="collapse navbar-collapse" id="navbar-menu">
      <div class="d-flex flex-column flex-md-row flex-fill align-items-stretch align-items-md-center">
        <ul class="navbar-nav">
          <li class="nav-item"><a class="nav-link" href="/admin/">Dashboard</a></li>
          <li class="nav-item"><a class="nav-link" href="/admin/models">Models</a></li>
          <li class="nav-item"><a class="nav-link" href="/admin/profiles">Profiles</a></li>
          <li class="nav-item"><a class="nav-link" href="/admin/decisions">Decisions</a></li>
          <li class="nav-item"><a class="nav-link" href="/admin/controls">Controls</a></li>
          <li class="nav-item active"><a class="nav-link" href="/admin/providers">Providers</a></li>
        </ul>
      </div>
    </div>
  </div>
 </header>
 <!-- ═══ Page shell ═══ -->
 <div class="page">
  <div class="page-wrapper">
    <div class="page-header d-print-none">
      <div class="container-xl">
        <div class="row g-2 align-items-center">
          <div class="col">
            <h2 class="page-title">Providers</h2>
            <div class="text-muted mt-1">Dispatch providers and their API endpoints</div>
            <div class="text-muted mt-1" style="font-size:.75rem"><span data-icon="alert" style="display:inline-flex;vertical-align:middle;margin-right:4px"></span>Creating, editing, or deleting a provider writes to config.local.yaml. A service reload is required before dispatch sees the change.</div>
          </div>
        </div>
      </div>
    </div>
    <div class="page-body">
      <div class="container-xl">
        <!-- Global restart-required banner (shown after edits/deletes) -->
        <div class="row row-cards" id="restart-banner-row" style="display:none">
          <div class="col-12">
            <div class="alert alert-warning" role="alert">
              <span data-icon="alert" style="display:inline-flex;vertical-align:middle;margin-right:6px"></span>
              A restart is required for provider changes to take effect.
            </div>
          </div>
        </div>
        <div class="row row-cards align-items-start">
          <!-- Add provider card -->
          <div class="col-xl-4">
            <div class="card">
              <div class="card-header">
                <h3 class="card-title"><span class="me-2" data-icon="plus"></span>Add provider</h3>
              </div>
              <div class="card-body">
                <div class="provider-form-list" id="add-provider-form">
                  <div class="provider-form-row">
                    <span class="provider-form-key" title="name">name</span>
                    <span class="provider-form-meta">unique id</span>
                    <span class="provider-form-control">
                      <input class="form-control form-control-sm" type="text" id="add-name" placeholder="e.g. openrouter" autocomplete="off">
                    </span>
                  </div>
                  <div class="provider-form-row">
                    <span class="provider-form-key" title="base_url">base_url</span>
                    <span class="provider-form-meta">http(s) URL</span>
                    <span class="provider-form-control">
                      <input class="form-control form-control-sm" type="text" id="add-base-url" placeholder="https://api.example.com/v1" autocomplete="off">
                    </span>
                  </div>
                  <div class="provider-form-row">
                    <span class="provider-form-key" title="api_key_env">api_key_env</span>
                    <span class="provider-form-meta">env var name</span>
                    <span class="provider-form-control">
                      <input class="form-control form-control-sm" type="text" id="add-api-key-env" placeholder="e.g. OPENROUTER_API_KEY" autocomplete="off">
                    </span>
                  </div>
                  <div class="provider-form-row">
                    <span class="provider-form-key" title="has_energy_telemetry">has_energy_telemetry</span>
                    <span class="provider-form-meta">SSE kWh comments</span>
                    <span class="provider-form-control">
                      <div class="form-check form-switch">
                        <input class="form-check-input toggle-input" type="checkbox" id="add-has-energy-telemetry">
                      </div>
                    </span>
                  </div>
                  <div class="provider-form-row">
                    <span class="provider-form-key" title="enabled">enabled</span>
                    <span class="provider-form-meta">available to dispatch</span>
                    <span class="provider-form-control">
                      <div class="form-check form-switch">
                        <input class="form-check-input toggle-input" type="checkbox" id="add-enabled" checked>
                      </div>
                    </span>
                  </div>
                </div>
                <div class="mt-3">
                  <button class="btn btn-success btn-sm" id="add-provider-btn" onclick="createProvider()">Create provider</button>
                </div>
              </div>
            </div>
          </div>
          <!-- Provider list -->
          <div class="col-xl-8">
            <div class="card">
              <div class="card-header">
                <h3 class="card-title"><span class="me-2" data-icon="server"></span>Configured providers</h3>
              </div>
              <div class="card-body p-0">
                <div id="provider-list">
                  <div class="provider-empty">Loading&hellip;</div>
                </div>
              </div>
            </div>
          </div>
        </div><!-- /row row-cards -->
      </div><!-- /container-xl -->
    </div><!-- /page-body -->
    <footer class="footer footer-transparent d-print-none">
      <div class="container-xl">
        &copy; 2026 adLee &middot; 6krrt, local LLM model router &middot; admin
      </div>
    </footer>
  </div>
 </div>
 <!-- ═══ Delete confirm modal ═══ -->
 <div class="modal modal-blur fade" id="delete-modal" tabindex="-1" role="dialog" aria-hidden="true">
  <div class="modal-dialog modal-dialog-centered" role="document">
    <div class="modal-content">
      <div class="modal-header">
        <h5 class="modal-title">Delete provider</h5>
        <button type="button" class="btn-close" data-bs-dismiss="modal" aria-label="Close"></button>
      </div>
      <div class="modal-body">
        <p>Delete provider <strong id="delete-modal-name"></strong>?</p>
        <p class="text-muted small">This writes to config.local.yaml; a service reload is required before dispatch sees the change.</p>
      </div>
      <div class="modal-footer">
        <button type="button" class="btn btn-outline-secondary btn-sm" data-bs-dismiss="modal">Cancel</button>
        <button type="button" class="btn btn-danger btn-sm" id="delete-confirm-btn">Delete</button>
      </div>
    </div>
  </div>
 </div>
 <!-- Toast (Bootstrap native) -->
 <div id="toast" class="toast align-items-center text-bg-info border-0 position-fixed bottom-0 end-0 p-3" role="alert" aria-live="assertive" aria-atomic="true"></div>
 <script>
 /* ═══════════════════════════════════════════════════
   admin/frontend/providers.html — provider management
   API base: relative (works under /admin/)
   ═══════════════════════════════════════════════════ */
 const API = '';  // relative to /admin/
 const SSE_URL = '/events/decisions';  // root-level endpoint
 /* ═─ Inline icon helper (no tabler-icons webfont) ═─ */
 function icon(name, size) {
  const s = size || 18;
  const svgs = {
    alert:      `<svg xmlns="http://www.w3.org/2000/svg" width="${s}" height="${s}" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10.29 3.86L1.82 18a2 2 0 0 0 1.71 3h16.94a2 2 0 0 0 1.71-3L13.71 3.86a2 2 0 0 0-3.42 0z"/><line x1="12" y1="9" x2="12" y2="13"/><line x1="12" y1="17" x2="12.01" y2="17"/></svg>`,
    plus:       `<svg xmlns="http://www.w3.org/2000/svg" width="${s}" height="${s}" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><line x1="12" y1="5" x2="12" y2="19"/><line x1="5" y1="12" x2="19" y2="12"/></svg>`,
    edit:       `<svg xmlns="http://www.w3.org/2000/svg" width="${s}" height="${s}" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M11 4H4a2 2 0 0 0-2 2v14a2 2 0 0 0 2 2h14a2 2 0 0 0 2-2v-7"/><path d="M18.5 2.5a2.121 2.121 0 0 1 3 3L12 15l-4 1 1-4 9.5-9.5z"/></svg>`,
    trash:      `<svg xmlns="http://www.w3.org/2000/svg" width="${s}" height="${s}" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><polyline points="3 6 5 6 21 6"/><path d="M19 6v14a2 2 0 0 1-2 2H7a2 2 0 0 1-2-2V6m3 0V4a2 2 0 0 1 2-2h4a2 2 0 0 1 2 2v2"/></svg>`,
    lock:       `<svg xmlns="http://www.w3.org/2000/svg" width="${s}" height="${s}" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><rect x="3" y="11" width="18" height="11" rx="2" ry="2"/><path d="M7 11V7a5 5 0 0 1 10 0v4"/></svg>`,
    server:     `<svg xmlns="http://www.w3.org/2000/svg" width="${s}" height="${s}" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><rect x="2" y="2" width="20" height="8" rx="2" ry="2"/><rect x="2" y="14" width="20" height="8" rx="2" ry="2"/><line x1="6" y1="6" x2="6.01" y2="6"/><line x1="6" y1="18" x2="6.01" y2="18"/></svg>`,
    default:    `<svg xmlns="http://www.w3.org/2000/svg" width="${s}" height="${s}" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><circle cx="12" cy="12" r="1"/></svg>`,
  };
  return svgs[name] || svgs.default;
 }
 function renderStaticIcons() {
  document.querySelectorAll('[data-icon]').forEach(el => {
    el.innerHTML = icon(el.dataset.icon);
  });
 }
 /* ── Toast (Bootstrap native via tabler wrapper) ── */
 let _toastEl = null;
 function toast(msg, type = 'info') {
  const bgClass = { success: 'text-bg-success', error: 'text-bg-danger', warning: 'text-bg-warning', info: 'text-bg-info' }[type] || 'text-bg-info';
  if (!_toastEl) {
    _toastEl = document.getElementById('toast');
    _toastEl.innerHTML = `<div class="d-flex">
      <div class="toast-body"></div>
      <button type="button" class="btn-close btn-close-white me-2 m-auto" data-bs-dismiss="toast" aria-label="Close"></button>
    </div>`;
    _toastEl._bt = new tabler.Toast(_toastEl, { delay: 4000 });
  }
  _toastEl.querySelector('.toast-body').textContent = msg;
  _toastEl.className = `toast align-items-center ${bgClass} border-0 position-fixed bottom-0 end-0 p-3`;
  _toastEl._bt.show();
 }
 /* ── Fetch wrappers ── */
 async function apiFetch(url, opts = {}) {
  try {
    const resp = await fetch(url, opts);
    if (!resp.ok) {
      const body = await resp.json().catch(() => ({}));
      throw new Error(body.detail || `${resp.status} ${resp.statusText}`);
    }
    return await resp.json();
  } catch (e) {
    console.warn('API call failed:', url, e);
    return null;
  }
 }
 async function apiFetchDetail(url, opts = {}) {
  const resp = await fetch(url, opts);
  const body = await resp.json().catch(() => ({}));
  if (!resp.ok) {
    const detail = body.detail || `${resp.status} ${resp.statusText}`;
    throw new Error(detail);
  }
  return body;
 }
 /* ═══════════════════════════════════════
   SSE LIVE STREAM
   ═══════════════════════════════════════ */
 let sseConn = null;
 function connectSSE() {
  try {
    if (sseConn) { sseConn.close(); }
    sseConn = new EventSource(SSE_URL);
    const statusDot = document.getElementById('sse-status');
    const statusText = document.getElementById('sse-text');
    statusDot.className = 'badge bg-warning';
    statusDot.textContent = 'connecting…';
    statusText.textContent = 'connecting…';
    sseConn.onopen = () => {
      statusDot.className = 'badge bg-success';
      statusDot.textContent = 'live';
      statusText.textContent = 'live';
    };
    sseConn.onerror = () => {
      statusDot.className = 'badge bg-danger';
      statusDot.textContent = 'reconnecting…';
      statusText.textContent = 'reconnecting…';
    };
  } catch (e) {
    document.getElementById('sse-status').className = 'badge bg-danger';
    document.getElementById('sse-status').textContent = 'offline';
    document.getElementById('sse-text').textContent = 'offline';
  }
 }
 /* ═══════════════════════════════════════
   PROVIDERS
   ═══════════════════════════════════════ */
 function escapeHtml(s) {
  return String(s).replace(/&/g,'&amp;').replace(/</g,'&lt;').replace(/>/g,'&gt;').replace(/"/g,'&quot;');
 }
 function sourceBadge(inBase, inOverlay) {
  if (inBase && inOverlay) return '<span class="badge bg-purple">overlay</span>';
  if (inBase) return '<span class="badge bg-secondary">base config</span>';
  if (inOverlay) return '<span class="badge bg-purple">overlay</span>';
  return '<span class="badge bg-secondary">unknown</span>';
 }
 function boolSwitch(id, checked, onchange, disabled) {
  return `<div class="form-check form-switch">
    <input class="form-check-input toggle-input" type="checkbox" id="${id}" ${checked ? 'checked' : ''} ${disabled ? 'disabled' : ''} onchange="${onchange}">
  </div>`;
 }
 function renderProvider(p) {
  const name = escapeHtml(String(p.name || ''));
  const baseUrl = escapeHtml(String(p.base_url || ''));
  const apiKeyEnv = escapeHtml(String(p.api_key_env || ''));
  const hasEnergy = !!p.has_energy_telemetry;
  const enabled = !!p.enabled;
  const inBase = !!p.in_base;
  const inOverlay = !!p.in_overlay;
  const isEditable = inOverlay;
  const readonlyHint = !isEditable
    ? `<span class="provider-readonly-hint" title="Base-config providers cannot be edited or deleted"><span data-icon="lock"></span> read-only</span>`
    : '';
  const actions = !isEditable
    ? ''
    : `<div class="provider-row-actions">
         <button class="btn btn-outline-secondary btn-sm" title="Edit provider" data-action="edit" data-provider-name="${name}"><span data-icon="edit"></span></button>
         <button class="btn btn-danger btn-sm" title="Delete provider" data-action="delete" data-provider-name="${name}"><span data-icon="trash"></span></button>
       </div>`;
  // View mode row.
  let html = `<div class="provider-form-row provider-view" data-provider="${name}">
    <span class="provider-form-key" title="${name}">
      ${name}
      <span class="provider-source ms-2">${sourceBadge(inBase, inOverlay)} ${readonlyHint}</span>
    </span>
    <span class="provider-form-meta">${baseUrl || '—'} &middot; ${apiKeyEnv || '—'}</span>
    <span class="provider-form-control">${actions}</span>
  </div>
  <div class="provider-form-row provider-view" data-provider="${name}">
    <span class="provider-form-key text-muted">
      <span class="provider-field-label me-3">energy telemetry ${hasEnergy ? 'on' : 'off'}</span>
      <span class="provider-field-label">enabled ${enabled ? 'yes' : 'no'}</span>
    </span>
    <span class="provider-form-meta"></span>
    <span class="provider-form-control"></span>
  </div>`;
  // Edit mode form (hidden by default).
  html += `<div class="provider-form-row provider-edit" data-provider="${name}" style="display:none">
    <span class="provider-form-key"><input class="form-control form-control-sm" type="text" value="${baseUrl}" data-field="base_url" placeholder="https://api.example.com/v1"></span>
    <span class="provider-form-meta">base_url</span>
    <span class="provider-form-control"></span>
  </div>
  <div class="provider-form-row provider-edit" data-provider="${name}" style="display:none">
    <span class="provider-form-key"><input class="form-control form-control-sm" type="text" value="${apiKeyEnv}" data-field="api_key_env" placeholder="env var name"></span>
    <span class="provider-form-meta">api_key_env</span>
    <span class="provider-form-control"></span>
  </div>
  <div class="provider-form-row provider-edit" data-provider="${name}" style="display:none">
    <span class="provider-form-key">has_energy_telemetry</span>
    <span class="provider-form-meta">SSE kWh comments</span>
    <span class="provider-form-control">${boolSwitch(`edit-energy-${name}`, hasEnergy, `toggleProviderField('${name}', 'has_energy_telemetry', this.checked)`, false)}</span>
  </div>
  <div class="provider-form-row provider-edit" data-provider="${name}" style="display:none">
    <span class="provider-form-key">enabled</span>
    <span class="provider-form-meta">available to dispatch</span>
    <span class="provider-form-control">${boolSwitch(`edit-enabled-${name}`, enabled, `toggleProviderField('${name}', 'enabled', this.checked)`, false)}</span>
  </div>
  <div class="provider-form-row provider-edit" data-provider="${name}" style="display:none">
    <span class="provider-form-key"></span>
    <span class="provider-form-meta"></span>
    <span class="provider-form-control">
      <button class="btn btn-success btn-sm" onclick="saveProviderEdit('${name}')">Save</button>
      <button class="btn btn-outline-secondary btn-sm ms-2" onclick="cancelProviderEdit('${name}')">Cancel</button>
    </span>
  </div>`;
  return html;
 }
 function renderProviders(providers) {
  const el = document.getElementById('provider-list');
  if (!providers) {
    el.innerHTML = '<div class="provider-empty">Could not load providers. Check the API.</div>';
    return;
  }
  if (!providers.length) {
    el.innerHTML = '<div class="provider-empty">No providers configured.</div>';
    return;
  }
  el.innerHTML = providers.map(renderProvider).join('');
  renderStaticIcons();
 }
 async function loadProviders() {
  const providers = await apiFetch(`${API}api/providers`);
  renderProviders(providers);
 }
 /* ═══════════════════════════════════════
   CRUD ACTIONS
   ═══════════════════════════════════════ */
 function showRestartBanner() {
  const row = document.getElementById('restart-banner-row');
  if (row) row.style.display = '';
 }
 function toggleProviderField(name, field, value) {
  // Switch change is local-only until Save is pressed.
 }
 async function saveProviderEdit(name) {
  const rows = document.querySelectorAll(`.provider-edit[data-provider="${CSS.escape(name)}"]`);
  const body = { name: name };
  rows.forEach(row => {
    const input = row.querySelector('[data-field]');
    if (input) {
      const field = input.dataset.field;
      body[field] = input.value;
    }
    const sw = row.querySelector('.form-check-input[type="checkbox"]');
    if (sw && sw.id && sw.id.startsWith('edit-')) {
      const field = sw.id.replace(`edit-`, '').replace(new RegExp(`-${name}$`), '');
      // map switch field names to provider schema keys
      const keyMap = {
        energy: 'has_energy_telemetry',
        enabled: 'enabled',
      };
      const key = keyMap[field];
      if (key) body[key] = sw.checked;
    }
  });
  try {
    const result = await apiFetchDetail(`${API}api/providers/${encodeURIComponent(name)}`, {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify(body),
    });
    toast(result && result.message ? 'Provider updated' : 'Provider saved', 'success');
    showRestartBanner();
    if (result && result.message) {
      toast(result.message, 'warning');
    }
    await loadProviders();
  } catch (e) {
    toast(e.message || 'Save failed', 'error');
  }
 }
 function cancelProviderEdit(name) {
  const viewRows = document.querySelectorAll(`.provider-view[data-provider="${CSS.escape(name)}"]`);
  const editRows = document.querySelectorAll(`.provider-edit[data-provider="${CSS.escape(name)}"]`);
  viewRows.forEach(r => r.style.display = '');
  editRows.forEach(r => r.style.display = 'none');
 }
 function startProviderEdit(name) {
  const viewRows = document.querySelectorAll(`.provider-view[data-provider="${CSS.escape(name)}"]`);
  const editRows = document.querySelectorAll(`.provider-edit[data-provider="${CSS.escape(name)}"]`);
  viewRows.forEach(r => r.style.display = 'none');
  editRows.forEach(r => r.style.display = '');
 }
 let _deleteModal = null;
 function getDeleteModal() {
  if (!_deleteModal) {
    _deleteModal = new tabler.Modal(document.getElementById('delete-modal'));
  }
  return _deleteModal;
 }
 function promptDelete(name) {
  document.getElementById('delete-modal-name').textContent = name;
  const btn = document.getElementById('delete-confirm-btn');
  btn.onclick = () => doDelete(name);
  getDeleteModal().show();
 }
 async function doDelete(name) {
  getDeleteModal().hide();
  try {
    const result = await apiFetchDetail(`${API}api/providers/${encodeURIComponent(name)}`, { method: 'DELETE' });
    toast('Provider deleted', 'success');
    showRestartBanner();
    if (result && result.message) {
      toast(result.message, 'warning');
    }
    await loadProviders();
  } catch (e) {
    toast(e.message || 'Delete failed', 'error');
  }
 }
 async function createProvider() {
  const name = document.getElementById('add-name').value.trim();
  const baseUrl = document.getElementById('add-base-url').value.trim();
  const apiKeyEnv = document.getElementById('add-api-key-env').value.trim();
  const hasEnergyTelemetry = document.getElementById('add-has-energy-telemetry').checked;
  const enabled = document.getElementById('add-enabled').checked;
  if (!name) {
    toast('Provider name is required', 'error');
    return;
  }
  if (!baseUrl) {
    toast('Base URL is required', 'error');
    return;
  }
  const body = {
    name: name,
    base_url: baseUrl,
    api_key_env: apiKeyEnv || null,
    has_energy_telemetry: hasEnergyTelemetry,
    enabled: enabled,
  };
  try {
    const result = await apiFetchDetail(`${API}api/providers/${encodeURIComponent(name)}`, {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify(body),
    });
    toast('Provider created', 'success');
    showRestartBanner();
    if (result && result.message) {
      toast(result.message, 'warning');
    }
    document.getElementById('add-name').value = '';
    document.getElementById('add-base-url').value = '';
    document.getElementById('add-api-key-env').value = '';
    document.getElementById('add-has-energy-telemetry').checked = false;
    document.getElementById('add-enabled').checked = true;
    await loadProviders();
  } catch (e) {
    toast(e.message || 'Create failed', 'error');
  }
 }
 /* ═══════════════════════════════════════
   INIT
   ═══════════════════════════════════════ */
 function init() {
  renderStaticIcons();
  document.getElementById('provider-list').addEventListener('click', (e) => {
    const btn = e.target.closest('button[data-action]');
    if (!btn) return;
    const name = btn.dataset.providerName;
    if (btn.dataset.action === 'edit') startProviderEdit(name);
    else if (btn.dataset.action === 'delete') promptDelete(name);
  });
  loadProviders();
  connectSSE();
 }
 init();
 </script>
 </body>
 </html>
--- a/config/config.yaml
+++ b/config/config.yaml
@@ -726,12 +726,17 @@ classifier:
    }
 dispatch_providers:
  # NeuralWatt is the only provider. The dict shape and the models table's
  # (model_id, provider) key are kept so a second one can be added without a
  # migration.
  neuralwatt:
    base_url: "https://api.neuralwatt.com/v1"
    api_key_env: "NEURALWATT_API_KEY"
    has_energy_telemetry: true
    enabled: true
  openrouter:
    base_url: "https://openrouter.ai/api/v1"
    api_key_env: "OPENROUTER_API_KEY"
    has_energy_telemetry: false
    enabled: true
 local_compute:
  # "Gaming mode", inverted: true means the router may use local hardware.
--- a/src/admin.py
+++ b/src/admin.py
@@ -30,6 +30,7 @@ from collections.abc import Callable
 from datetime import datetime, timezone
 from pathlib import Path
 from typing import Any, List, Literal, Optional
 from urllib.parse import urlparse
 from fastapi import APIRouter, BackgroundTasks, HTTPException
 from fastapi.responses import FileResponse
@@ -43,6 +44,7 @@ import metrics
 import routing
 from config import (
    BUILTIN_PROFILES,
    DispatchProvider,
    FlexPreference,
    RouterConfig,
    RoutingProfile,
@@ -720,6 +722,13 @@ class _ProfileUpdateBody(BaseModel):
    allowed_model_ids: Optional[list[str]] = None
 class _ProviderUpdateBody(BaseModel):
    base_url: str
    api_key_env: str
    has_energy_telemetry: bool
    enabled: bool
 def _runtime_state(cfg: Any) -> dict:
    """Read every toggle knob's current in-memory value off ``cfg``."""
    return {
@@ -768,6 +777,7 @@ def build_router(
    _admin_models = _REPO_ROOT / "admin" / "frontend" / "models.html"
    _admin_decisions = _REPO_ROOT / "admin" / "frontend" / "decisions.html"
    _admin_profiles = _REPO_ROOT / "admin" / "frontend" / "profiles.html"
    _admin_providers = _REPO_ROOT / "admin" / "frontend" / "providers.html"
    _admin_proficiency = _REPO_ROOT / "admin" / "frontend" / "proficiency.html"
    _admin_logo = _REPO_ROOT / "admin" / "frontend" / "6krrt-logo.webp"
@@ -801,6 +811,10 @@ def build_router(
    def admin_profiles_page() -> FileResponse:
        return FileResponse(_admin_profiles, media_type="text/html", headers=_NO_CACHE_HEADERS)
    @router.get("/providers")
    def admin_providers_page() -> FileResponse:
        return FileResponse(_admin_providers, media_type="text/html", headers=_NO_CACHE_HEADERS)
    @router.get("/proficiency")
    def admin_proficiency_page() -> FileResponse:
        return FileResponse(_admin_proficiency, media_type="text/html", headers=_NO_CACHE_HEADERS)
@@ -1242,6 +1256,133 @@ def build_router(
            "message": "A restart is required for this change to take effect",
        }
    def _provider_provenance(name: str) -> dict:
        base_store = load_config_store_safe(config_path) or {}
        base_providers = base_store.get("dispatch_providers") or {}
        try:
            overlay = load_config_store(config_local_path) or {}
        except FileNotFoundError:
            overlay = {}
        overlay_providers = overlay.get("dispatch_providers") or {}
        return {
            "in_base": name in base_providers,
            "in_overlay": name in overlay_providers,
        }
    def _merged_providers() -> dict[str, Any]:
        merged = _load_merged_config_store(config_path, config_local_path)
        return merged.get("dispatch_providers") or {}
    def _provider_block_dict(body: _ProviderUpdateBody) -> dict[str, Any]:
        return body.model_dump()
    def _validate_provider_url(base_url: str) -> None:
        parsed = urlparse(base_url)
        if parsed.scheme not in ("http", "https") or not parsed.netloc:
            raise HTTPException(
                status_code=422,
                detail=f"base_url must be an http(s) URL, got {base_url!r}",
            )
    def _persist_provider(
        name: str, provider_dict: Optional[dict[str, Any]], *, delete: bool = False
    ) -> None:
        try:
            _persist_to(
                config_path,
                config_local_path,
                ("dispatch_providers", name),
                provider_dict,
                delete=delete,
            )
        except ValidationError as exc:
            raise HTTPException(
                status_code=422,
                detail=exc.errors()[0]["msg"],
            ) from exc
        except KeyError as exc:
            raise HTTPException(status_code=404, detail="provider not found") from exc
    @router.get("/api/providers")
    def admin_providers_list() -> list[dict[str, Any]]:
        providers = _merged_providers()
        result = []
        for name, p in providers.items():
            provenance = _provider_provenance(name)
            result.append(
                {
                    "name": name,
                    "base_url": p["base_url"],
                    "api_key_env": p["api_key_env"],
                    "has_energy_telemetry": bool(p.get("has_energy_telemetry", False)),
                    "enabled": bool(p.get("enabled", True)),
                    "in_base": provenance["in_base"],
                    "in_overlay": provenance["in_overlay"],
                }
            )
        return result
    @router.get("/api/providers/{name}")
    def admin_provider_detail(name: str) -> dict[str, Any]:
        providers = _merged_providers()
        if name not in providers:
            raise HTTPException(status_code=404, detail="provider not found")
        p = providers[name]
        provenance = _provider_provenance(name)
        return {
            "name": name,
            "base_url": p["base_url"],
            "api_key_env": p["api_key_env"],
            "has_energy_telemetry": bool(p.get("has_energy_telemetry", False)),
            "enabled": bool(p.get("enabled", True)),
            "in_base": provenance["in_base"],
            "in_overlay": provenance["in_overlay"],
        }
    @router.post("/api/providers/{name}")
    def admin_provider_update(name: str, body: _ProviderUpdateBody) -> dict[str, Any]:
        _validate_provider_url(body.base_url)
        provider_dict = _provider_block_dict(body)
        try:
            DispatchProvider(**provider_dict)
        except ValidationError as exc:
            raise HTTPException(
                status_code=422,
                detail=exc.errors()[0]["msg"],
            ) from exc
        _persist_provider(name, provider_dict)
        return {
            "provider": {"name": name, **provider_dict},
            "message": "A restart is required for this change to take effect",
        }
    @router.delete("/api/providers/{name}")
    def admin_provider_delete(name: str) -> dict[str, Any]:
        if name == cfg.dispatch_settings.default_provider:
            raise HTTPException(
                status_code=422,
                detail=(
                    f"Cannot delete provider {name!r}: it is the current "
                    f"dispatch_settings.default_provider."
                ),
            )
        provenance = _provider_provenance(name)
        if provenance["in_base"]:
            raise HTTPException(
                status_code=403,
                detail=(
                    f"provider {name!r} is defined in config/config.yaml and "
                    f"is read-only"
                ),
            )
        if not provenance["in_overlay"]:
            raise HTTPException(status_code=404, detail="provider not found")
        _persist_provider(name, None, delete=True)
        return {
            "deleted": name,
            "message": "A restart is required for this change to take effect",
        }
    @router.get("/api/models")
    def admin_models() -> list:
        """All models with per-category proficiency, for the admin model table."""
--- a/src/config.py
+++ b/src/config.py
@@ -884,6 +884,8 @@ class LocalDispatchModel(StrictModel):
 class DispatchProvider(StrictModel):
    base_url: str
    api_key_env: str
    has_energy_telemetry: bool = False
    enabled: bool = True
 class LoggingConfig(StrictModel):
--- a/src/dispatcher.py
+++ b/src/dispatcher.py
@@ -544,10 +544,9 @@ def _classifier_client() -> OpenAI:
 # GLOBAL across requests rather than per-request — that is what bounds cloud
 # spend during a sustained outage to one attempt per cooldown window.
 _last_classifier_failure: float = 0.0
-# Timestamp of the last account-level provider refusal. If the account is out
+# Per-provider timestamp of the last account-level refusal. A refusal on one
-# of credit, a cloud classification is guaranteed to fail AND to cost a
+# provider must not poison classification or fallback decisions for another.
-# request; skip step 4 entirely for the cooldown window.
+_provider_refusal_since: dict[str, float] = {}
 _last_account_refusal: float = 0.0
 # The in-flight request's session key, for the two cascade steps that need it.
 #
@@ -574,10 +573,9 @@ def _record_success_cooldown() -> None:
    _last_classifier_failure = 0.0
-def _record_account_refusal() -> None:
+def _record_provider_refusal(provider: str) -> None:
-    """Note an account-level refusal so the cascade stops paying for cloud."""
+    """Note an account-level refusal for a specific provider."""
-    global _last_account_refusal
+    _provider_refusal_since[provider] = time.time()
    _last_account_refusal = time.time()
 def _classifier_backoff_active() -> bool:
@@ -930,8 +928,10 @@ def _classify_cascade(
    if cf is None:
        return None
    cooldown = cfg.classifier.cooldown_seconds
-    if time.time() - _last_account_refusal < cooldown:
+    if time.time() - _provider_refusal_since.get(
-        logs.warning("classify_cloud_skipped", reason="account_refusal")
+        cfg.dispatch_settings.default_provider, 0.0
    ) < cooldown:
        logs.warning("classify_cloud_skipped", reason="provider_refusal")
        return None
    client = _cloud_classifier_client(cf)
    if client is None:
@@ -1884,6 +1884,8 @@ def extract_telemetry(payload: dict) -> Telemetry:
    yields an all-None Telemetry rather than an error, and the neutral-0.5
    path in scoring covers the gap.
    """
    if not isinstance(payload, dict):
        payload = {}
    energy = payload.get("energy") or {}
    cost = payload.get("cost") or {}
    kwh = energy.get("energy_kwh")
@@ -3511,6 +3513,39 @@ def _check_pinned_capabilities(
        )
 def _resolve_pinned_provider(model_id: str) -> Optional[str]:
    """Return the ``provider`` that owns *model_id* in the catalog.
    Queries ``SELECT provider FROM models WHERE model_id = ?`` and returns the
    sole result.  When the same model_id lives on multiple providers
    (for example *gpt-4a* on both *neuralwatt* and *openrouter*), the function
    returns the configured ``default_provider`` so the resolution is
    deterministic.
    Returns ``None`` when *model_id* is not found — the caller falls back to
    the ``default_provider`` in that case.
    """
    conn = _db()
    try:
        providers = list(cfg.dispatch_providers)
        placeholders = ",".join("?" * len(providers))
        rows = conn.execute(
            f"SELECT provider FROM models WHERE model_id = ? AND provider IN ({placeholders})",
            (model_id, *providers),
        ).fetchall()
    finally:
        conn.close()
    if len(rows) == 1:
        return rows[0]["provider"]
    if len(rows) > 1:
        # Prefer the configured default when there's ambiguity.
        default = cfg.dispatch_settings.default_provider
        if any(r["provider"] == default for r in rows):
            return default
        return rows[0]["provider"]
    return None
@app.get("/v1/models")
 def list_models():
    """The routable catalog, in OpenAI's list shape.
@@ -3555,17 +3590,24 @@ def _sniff_telemetry_line(line: str) -> Optional[tuple[str, dict]]:
    `data: [DONE]`. They are SSE comments, so every ordinary client ignores
    them — which is exactly why the stream can be proxied through untouched
    while still being read on the way past.
    Providers such as OpenRouter may emit other SSE comment lines (keepalive
    comments beginning with ``: OPENROUTER``); these are deliberately ignored.
    """
-    if not line.startswith(":"):
+    if not isinstance(line, str) or not line.startswith(":"):
        return None
    body = line[1:].strip()
-    for key in ("energy", "cost"):
+    if not body:
-        prefix = f"{key} "
+        return None
-        if body.startswith(prefix):
+    first_word = body.split(None, 1)[0]
-            try:
+    if first_word not in ("energy", "cost"):
-                return key, json.loads(body[len(prefix):])
+        return None
-            except json.JSONDecodeError:
+    prefix = f"{first_word} "
-                return None
+    if body.startswith(prefix):
        try:
            return first_word, json.loads(body[len(prefix):])
        except json.JSONDecodeError:
            return None
    return None
@@ -3605,23 +3647,33 @@ def chat_completions(body: dict[str, Any], background: BackgroundTasks):
    requested = body.get("model") or ROUTER_MODEL
    # Some clients (opencode among them) send the model as `provider/model`,
-    # for a pin as much as for `auto`. Real ids can contain slashes too
+    # for a pin as much as for `auto`.  Real ids can contain slashes too
    # (`deepseek-ai/DeepSeek-V4-Flash`), so a stripped suffix is only trusted
    # once it resolves to a real catalog row -- otherwise the id, slash and
    # all, is what gets dispatched, on the chance the slash is really part of
-    # it. Without this, a pin sent as `llm-router/gemma-4-31b` kept its
+    # it.
-    # prefix past this point: the capability pre-check looked up a model_id
+    ALIAS_PREFIX = "llm-router/"
-    # the catalog has never heard of, found no row, and fail-closed a pin
+    if requested.startswith(ALIAS_PREFIX):
-    # that could satisfy the request into a false 422 -- and even past that
+        # opencode's pin alias: try the remainder first.  This preserves full
-    # check, the same unstripped id would have gone upstream and drawn a 400
+        # OpenRouter vendor/model ids like `openai/gpt-6-astra` that the
-    # from NeuralWatt, which only knows the bare form.
+        # rsplit heuristic below would mangle to `gpt-6-astra`.
-    bare = requested.rsplit("/", 1)[-1]
+        candidate = requested[len(ALIAS_PREFIX):]
-    wants_routing = bare == ROUTER_MODEL or bare.startswith(f"{ROUTER_MODEL}:")
+        if _model_exists(candidate) or _local_dispatch_config_for(candidate) is not None:
-    if wants_routing or (
+            requested = candidate
-        bare != requested
+        else:
-        and (_model_exists(bare) or _local_dispatch_config_for(bare) is not None)
+            # Fallback to the existing suffix rule so a virtual name such as
-    ):
+            # `llm-router/auto` still routes instead of being passed through.
-        requested = bare
+            bare = requested.rsplit("/", 1)[-1]
            if bare == ROUTER_MODEL or bare.startswith(f"{ROUTER_MODEL}:"):
                requested = bare
    else:
        bare = requested.rsplit("/", 1)[-1]
        if bare != requested and (
            _model_exists(bare) or _local_dispatch_config_for(bare) is not None
        ):
            requested = bare
    wants_routing = requested == ROUTER_MODEL or requested.startswith(f"{ROUTER_MODEL}:")
    logs.debug(
        "request",
@@ -3864,7 +3916,9 @@ def chat_completions(body: dict[str, Any], background: BackgroundTasks):
        category = decision.classification.task_category
    else:
        cloud_provider = cfg.dispatch_settings.default_provider
-        target, provider, category = requested, cloud_provider, "general_chat"
+        resolved = _resolve_pinned_provider(requested)
        provider = resolved if resolved is not None else cloud_provider
        target, category = requested, "general_chat"
        if provider == cloud_provider and (
            entry := _local_dispatch_config_for(target)
        ) is not None:
@@ -4087,10 +4141,10 @@ def chat_completions(body: dict[str, Any], background: BackgroundTasks):
                        cfg.circuit_breaker.backoff_multiplier,
                    )
                if _account_level_refusal(resp.status_code):
-                    # The account is refusing. A cloud CLASSIFICATION would
+                    # The provider account is refusing. A cloud CLASSIFICATION
-                    # fail the same way and still cost a request, so tell the
+                    # would fail the same way and still cost a request, so tell
-                    # cascade to skip step 3 for the cooldown window.
+                    # the cascade to skip step 3 for this provider's cooldown.
-                    _record_account_refusal()
+                    _record_provider_refusal(provider)
                if _account_level_refusal(resp.status_code) or not alternatives:
                    # Both spec §1c triggers: every cloud candidate exhausted, or a
                    # non-retryable account-level refusal — the remaining cloud rows
@@ -4253,14 +4307,20 @@ def chat_completions(body: dict[str, Any], background: BackgroundTasks):
            )
            yield f"data: {json.dumps({'error': {'message': detail}})}\n\n".encode()
            return
        provider_has_telemetry = cfg.dispatch_providers[
            provider
        ].has_energy_telemetry
        try:
            for raw in upstream.iter_lines(decode_unicode=True):
                if raw is None:
                    continue
-                sniffed = _sniff_telemetry_line(raw)
+                if provider_has_telemetry:
-                if sniffed:
+                    sniffed = _sniff_telemetry_line(raw)
-                    collected[sniffed[0]] = sniffed[1]
+                    if sniffed:
-                elif raw.startswith("data: ") and raw.strip() != "data: [DONE]":
+                        collected[sniffed[0]] = sniffed[1]
                elif _sniff_telemetry_line(raw) is not None:
                    continue
                if raw.startswith("data: ") and raw.strip() != "data: [DONE]":
                    try:
                        chunk = json.loads(raw[6:])
                        if chunk.get("id"):
@@ -4399,9 +4459,9 @@ def chat_completions(body: dict[str, Any], background: BackgroundTasks):
        upstream_text = attempt.text[:500]
        attempt.close()
        if _account_level_refusal(attempt.status_code):
-            # Same reason as the streaming path: an account-level refusal
+            # Same reason as the non-streaming path: an account-level refusal
-            # makes the cloud classifier a guaranteed waste of a request.
+            # on this provider makes its cloud classifier a waste of a request.
-            _record_account_refusal()
+            _record_provider_refusal(provider)
            break
    if upstream is None:
        fallback = try_local_fallback(target, upstream_status)
--- a/src/poller.py
+++ b/src/poller.py
@@ -32,11 +32,22 @@ from typing import Optional
 import requests
 from config import LocalDispatchModel, RouterConfig, load_config
 from tier import apply_tiering
 MODELS_URL = "https://api.neuralwatt.com/v1/models"
 OPENROUTER_MODELS_URL = "https://openrouter.ai/api/v1/models"
 REQUEST_TIMEOUT = 20  # seconds
 OPENROUTER_VIRTUAL_ROUTERS = frozenset({
    "openrouter/auto",
    "openrouter/auto-beta",
    "openrouter/free",
    "openrouter/fusion",
    "openrouter/pareto-code",
    "openrouter/bodybuilder",
 })
 # Serving-class suffixes. NeuralWatt ships one base model as several catalog
 # rows that differ only by these tokens, and they combine freely — hence ids
 # like 'glm-5.2-short-fast-flex'. They are stripped from the end of the id one
@@ -222,6 +233,86 @@ def fetch_neuralwatt(provider: str) -> list[ModelRow]:
    return rows
 def _openrouter_price_to_cost_per_1m(value: Optional[str]) -> Optional[float]:
    """Convert an OpenRouter per-token price string to cost per 1M tokens."""
    if value is None:
        return None
    try:
        per_token = float(value)
    except ValueError:
        return None
    return per_token * 1_000_000
 def parse_openrouter_model(raw_model: dict) -> Optional[ModelRow]:
    """Normalize a single OpenRouter /v1/models entry into a ModelRow.
    Returns ``None`` for virtual routers (``openrouter/*``) so callers can
    drop them cleanly without losing the rest of the catalog.
    """
    model_id = raw_model.get("id") or ""
    if model_id in OPENROUTER_VIRTUAL_ROUTERS:
        return None
    canonical_slug = raw_model.get("canonical_slug") or model_id
    pricing = raw_model.get("pricing", {}) or {}
    top_provider = raw_model.get("top_provider", {}) or {}
    architecture = raw_model.get("architecture", {}) or {}
    reasoning = raw_model.get("reasoning", {}) or {}
    supported_parameters = raw_model.get("supported_parameters") or []
    input_modalities = architecture.get("input_modalities") or []
    prompt_price = _openrouter_price_to_cost_per_1m(pricing.get("prompt"))
    completion_price = _openrouter_price_to_cost_per_1m(pricing.get("completion"))
    variant = model_id.split(":")[-1] if ":" in model_id else "standard"
    return ModelRow(
        model_id=model_id,
        provider="openrouter",
        base_model_id=canonical_slug,
        display_name=raw_model.get("name"),
        cost_per_1m_prompt=prompt_price,
        cost_per_1m_completion=completion_price,
        cost_per_1m_prompt_cached=None,
        context_window=top_provider.get("context_length")
        or raw_model.get("context_length"),
        max_output_tokens=top_provider.get("max_completion_tokens"),
        supports_tools="tools" in supported_parameters,
        supports_json_mode="response_format" in supported_parameters,
        supports_vision="image" in input_modalities,
        supports_reasoning="reasoning" in supported_parameters,
        reasoning_default_enabled=bool(reasoning.get("default_enabled")),
        latency_class="flex" if ":batch" in model_id else "standard",
        reasoning_mode=reasoning.get("mode") or "none",
        context_variant=variant,
        access_level="public",
        pricing_tbd=prompt_price is None,
        deprecated=False,
    )
 def fetch_openrouter(provider: str) -> list[ModelRow]:
    """Fetch the public OpenRouter catalog and normalize it into ModelRows.
    The OpenRouter `/v1/models` endpoint is unauthenticated and paginated with
    a default page size of 500 (max 1000).  A single request covers the whole
    catalog.  ``pricing.web_search`` and ``pricing.cache_read`` differentials
    are intentionally omitted from catalog cost estimates; routing uses the
    prompt and completion price only.
    """
    resp = requests.get(OPENROUTER_MODELS_URL, timeout=REQUEST_TIMEOUT)
    resp.raise_for_status()
    payload = resp.json()
    rows: list[ModelRow] = []
    for raw_model in payload.get("data", []):
        row = parse_openrouter_model(raw_model)
        if row is not None:
            rows.append(row)
    return rows
 def upsert(conn: sqlite3.Connection, rows: list[ModelRow], cfg: RouterConfig) -> None:
    now = datetime.now(timezone.utc).isoformat()
    for r in rows:
@@ -383,55 +474,90 @@ def upsert_local_dispatch_models(
    conn.commit()
-def mark_stale(conn: sqlite3.Connection, cfg: RouterConfig) -> None:
+def mark_stale(
-    """Flag rows that weren't touched by this poll run as stale, rather than
+    conn: sqlite3.Connection,
-    silently leaving old data looking current. Assumes the caller has already
+    cfg: RouterConfig,
-    established the fetched catalog was plausible (see the row-count floor
+    *,
-    check before upsert)."""
+    provider: Optional[str] = None,
-    conn.execute(
+) -> None:
-        """
+    """Flag rows that weren't touched by this poll run as stale.
-        UPDATE models
+
-        SET availability = 'stale'
+    When *provider* is given, only rows for that provider are considered;
-        WHERE availability = 'active'
+    when absent (the historical default across the rest of this repo), all
-          AND julianday('now') - julianday(last_updated) > ?
+    rows are scoped — matching the previous global behaviour.
-        """,
+
-        (cfg.freshness.stale_after_days,),
+    Assumes the caller has already established the fetched catalog was
-    )
+    plausible (see the row-count floor check before upsert).
    """
    where = "availability = 'active' AND julianday('now') - julianday(last_updated) > ?"
    params: list[object] = [cfg.freshness.stale_after_days]
    if provider is not None:
        where += " AND provider = ?"
        params.append(provider)
    conn.execute(f"UPDATE models SET availability = 'stale' WHERE {where}", params)
    conn.commit()
 def main() -> int:
    cfg = load_config("config/config.yaml")
    provider = cfg.dispatch_settings.default_provider
    conn = sqlite3.connect(cfg.database.path)
    conn.execute("PRAGMA foreign_keys = ON")
    upsert_local_dispatch_models(conn, cfg)
-    try:
+    # Map provider keys (dispatch_providers dict keys) to their fetch functions.
-        rows = fetch_neuralwatt(provider)
+    # Only entries with keys matching the key are dispatched; unknown keys
-    except requests.RequestException as e:
+    # are logged and skipped so a future provider in config doesn't abort
-        print(f"[{provider}] FAILED: {e}", file=sys.stderr)
+    # the whole run.
-        conn.close()
+    FETCHERS: dict[str, callable] = {
-        return 1
+        "neuralwatt": fetch_neuralwatt,
        "openrouter": fetch_openrouter,
    }
-    if len(rows) == 0:
+    total_upserted = 0
        raise CatalogTooSmall("NeuralWatt returned 0 models — this is not a valid catalog")
    current_count = conn.execute(
        "SELECT COUNT(*) FROM models WHERE provider=?", (provider,)
    ).fetchone()[0]
    if len(rows) < (current_count // 2) and current_count > 0:
        warnings.warn(
            f"[{provider}] fetched only {len(rows)} models "
            f"(current DB has {current_count}); proceeding but catalog may be truncated"
        )
-    upsert(conn, rows, cfg)
+    for provider, prov_cfg in cfg.dispatch_providers.items():
-    print(f"[{provider}] upserted {len(rows)} models")
+        if not prov_cfg.enabled:
            print(f"[{provider}] skipped (disabled)")
            continue
-    mark_stale(conn, cfg)
+        fetcher = FETCHERS.get(provider)
        if fetcher is None:
            print(f"[{provider}] unknown provider — skipping (no fetcher)")
            continue
        try:
            rows = fetcher(provider)
        except CatalogTooSmall as e:
            print(f"[{provider}] CATALOG_TOO_SMALL: {e}; skipping provider", file=sys.stderr)
            continue
        except requests.RequestException as e:
            print(f"[{provider}] FAILED: {e}", file=sys.stderr)
            continue
        if len(rows) == 0:
            print(f"[{provider}] fetched 0 models — skipping provider", file=sys.stderr)
            continue
        current_count = conn.execute(
            "SELECT COUNT(*) FROM models WHERE provider=?", (provider,)
        ).fetchone()[0]
        if len(rows) < (current_count // 2) and current_count > 0:
            warnings.warn(
                f"[{provider}] fetched only {len(rows)} models "
                f"(current DB has {current_count}); proceeding but catalog may be truncated"
            )
        upsert(conn, rows, cfg)
        total_upserted += len(rows)
        print(f"[{provider}] upserted {len(rows)} models")
        mark_stale(conn, cfg, provider=provider)
    apply_tiering(conn, cfg)
    conn.close()
-    print(f"done, {len(rows)} rows upserted total")
+    print(f"done, {total_upserted} rows upserted total")
    return 0
--- a/tests/test_admin_providers.py
+++ b/tests/test_admin_providers.py
@@ -0,0 +1,259 @@
 """Tests for the admin provider CRUD endpoints and /admin/providers page route.
 The page fixture creates a temporary ``admin/frontend/providers.html`` if the
 file does not yet exist, and removes it after the test so no HTML file is left
 in the repo as part of this change.
 """
 from __future__ import annotations
 import shutil
 import sqlite3
 from pathlib import Path
 import pytest
 from fastapi import FastAPI
 from starlette.testclient import TestClient
 from admin import build_router
 from config import load_config
 ROOT = Path(__file__).resolve().parent.parent
 def _make_db(tmp_path: Path) -> sqlite3.Connection:
    conn = sqlite3.connect(str(tmp_path / "providers.db"))
    conn.row_factory = sqlite3.Row
    return conn
@pytest.fixture
 def client(tmp_path):
    """A TestClient for an isolated admin router over a temp config.yaml copy."""
    (tmp_path / "config").mkdir(parents=True, exist_ok=True)
    config_yaml = tmp_path / "config" / "config.yaml"
    shutil.copyfile(ROOT / "config" / "config.yaml", config_yaml)
    schema_sql = (ROOT / "config" / "schema.sql").read_text()
    conn = _make_db(tmp_path)
    conn.executescript(schema_sql)
    conn.close()
    cfg = load_config(str(config_yaml))
    def _db_factory() -> sqlite3.Connection:
        return _make_db(tmp_path)
    router = build_router(cfg, _db_factory, base_dir=str(tmp_path))
    app = FastAPI()
    app.include_router(router, prefix="/admin")
    page = ROOT / "admin" / "frontend" / "providers.html"
    created_page = False
    if not page.exists():
        page.write_text("<html><body>Providers</body></html>")
        created_page = True
    try:
        yield TestClient(app)
    finally:
        if created_page:
            page.unlink(missing_ok=True)
 _VALID_PROVIDER = {
    "base_url": "https://api.example.com/v1",
    "api_key_env": "EXAMPLE_API_KEY",
    "has_energy_telemetry": True,
    "enabled": True,
 }
 _SENTINEL = "# SENTINEL_PRESERVED_12345"
 def _insert_header_marker(config_yaml: Path) -> None:
    """Prepend a unique marker comment just above the ``logging:`` map."""
    text = config_yaml.read_text()
    assert "logging:" in text
    config_yaml.write_text(text.replace("logging:", f"{_SENTINEL}\nlogging:", 1))
 def test_providers_page_returns_html(client):
    """GET /admin/providers serves the providers HTML page."""
    resp = client.get("/admin/providers")
    assert resp.status_code == 200
    assert resp.headers["content-type"].startswith("text/html")
    assert "Providers" in resp.text
 def test_provider_crud_round_trip(client, tmp_path):
    """Create, read, and delete a provider through the overlay."""
    local_yaml = tmp_path / "config" / "config.local.yaml"
    resp = client.post("/admin/api/providers/testprov", json=_VALID_PROVIDER)
    assert resp.status_code == 200
    body = resp.json()
    assert body["provider"]["name"] == "testprov"
    assert body["provider"]["base_url"] == _VALID_PROVIDER["base_url"]
    assert local_yaml.exists()
    text = local_yaml.read_text()
    assert "testprov:" in text
    assert "dispatch_providers:" in text
    resp = client.get("/admin/api/providers/testprov")
    assert resp.status_code == 200
    detail = resp.json()
    assert detail["name"] == "testprov"
    assert detail["base_url"] == _VALID_PROVIDER["base_url"]
    assert detail["api_key_env"] == _VALID_PROVIDER["api_key_env"]
    assert detail["has_energy_telemetry"] is True
    assert detail["enabled"] is True
    resp = client.get("/admin/api/providers")
    assert resp.status_code == 200
    names = [p["name"] for p in resp.json()]
    assert "testprov" in names
    resp = client.delete("/admin/api/providers/testprov")
    assert resp.status_code == 200
    assert resp.json()["deleted"] == "testprov"
    resp = client.get("/admin/api/providers/testprov")
    assert resp.status_code == 404
 def test_provider_post_invalid_url_returns_422(client, tmp_path):
    """A non-URL base_url is rejected without writing to the overlay."""
    local_yaml = tmp_path / "config" / "config.local.yaml"
    resp = client.post(
        "/admin/api/providers/badurl",
        json={
            "base_url": "not-a-url",
            "api_key_env": "BAD_API_KEY",
            "has_energy_telemetry": False,
            "enabled": True,
        },
    )
    assert resp.status_code == 422
    assert not local_yaml.exists()
 def test_provider_delete_default_returns_422_and_leaves_config(
    client, tmp_path
 ):
    """Deleting dispatch_settings.default_provider is rejected and config is untouched."""
    config_yaml = tmp_path / "config" / "config.yaml"
    local_yaml = tmp_path / "config" / "config.local.yaml"
    base_before = config_yaml.read_bytes()
    local_before = local_yaml.read_bytes() if local_yaml.exists() else b""
    resp = client.delete("/admin/api/providers/neuralwatt")
    assert resp.status_code == 422
    detail = resp.json()["detail"]
    assert "neuralwatt" in detail
    assert "default_provider" in detail
    assert config_yaml.read_bytes() == base_before
    assert not local_yaml.exists()
    assert local_before == b""
 def test_provider_delete_base_configured_returns_403_and_explains_provenance(
    client, tmp_path
 ):
    """Deleting a base-configured non-default provider returns 403 naming config/config.yaml."""
    config_yaml = tmp_path / "config" / "config.yaml"
    base_before = config_yaml.read_bytes()
    resp = client.delete("/admin/api/providers/openrouter")
    assert resp.status_code == 403
    detail = resp.json()["detail"]
    assert "openrouter" in detail
    assert "config/config.yaml" in detail
    assert "read-only" in detail.lower()
    assert config_yaml.read_bytes() == base_before
 def test_provider_delete_unknown_returns_404(client, tmp_path):
    """Deleting a provider that does not exist returns 404."""
    config_yaml = tmp_path / "config" / "config.yaml"
    base_before = config_yaml.read_bytes()
    resp = client.delete("/admin/api/providers/nonexistent")
    assert resp.status_code == 404
    assert "provider not found" in resp.json()["detail"]
    assert config_yaml.read_bytes() == base_before
 def test_provider_post_writes_overlay_and_preserves_base_backup_created(
    client, tmp_path
 ):
    """POSTing a valid provider writes only the overlay, leaves base alone, and backs up."""
    config_yaml = tmp_path / "config" / "config.yaml"
    local_yaml = tmp_path / "config" / "config.local.yaml"
    _insert_header_marker(config_yaml)
    base_before = config_yaml.read_bytes()
    resp = client.post("/admin/api/providers/testprov", json=_VALID_PROVIDER)
    assert resp.status_code == 200
    body = resp.json()
    assert body["provider"]["name"] == "testprov"
    assert body["provider"]["base_url"] == _VALID_PROVIDER["base_url"]
    assert config_yaml.read_bytes() == base_before
    assert local_yaml.exists()
    local_text = local_yaml.read_text()
    assert "dispatch_providers:" in local_text
    assert "testprov:" in local_text
    assert "base_url: https://api.example.com/v1" in local_text
    backups = sorted(tmp_path.glob("config/config.local.yaml.bak.*"))
    assert len(backups) == 1
    assert backups[0].read_text().strip() != ""
 def test_provider_list_includes_base_and_overlay_providers(client, tmp_path):
    """GET /admin/api/providers includes both base-configured and overlay providers."""
    config_yaml = tmp_path / "config" / "config.yaml"
    local_yaml = tmp_path / "config" / "config.local.yaml"
    local_yaml.write_text(
        "# Machine-local overlay -- test fixture\n"
        "dispatch_providers:\n"
        "  overlayprov:\n"
        "    base_url: https://overlay.example.com/v1\n"
        "    api_key_env: OVERLAY_API_KEY\n"
        "    has_energy_telemetry: false\n"
        "    enabled: true\n"
    )
    resp = client.get("/admin/api/providers")
    assert resp.status_code == 200
    names = {p["name"] for p in resp.json()}
    assert names >= {"neuralwatt", "openrouter", "overlayprov"}
    config_before = config_yaml.read_bytes()
    assert config_yaml.read_bytes() == config_before
    by_name = {p["name"]: p for p in resp.json()}
    assert by_name["neuralwatt"]["base_url"] is not None
    assert by_name["openrouter"]["base_url"] is not None
    assert by_name["overlayprov"]["base_url"] == "https://overlay.example.com/v1"
    assert by_name["neuralwatt"]["in_base"] is True
    assert by_name["neuralwatt"]["in_overlay"] is False
    assert by_name["openrouter"]["in_base"] is True
    assert by_name["openrouter"]["in_overlay"] is False
    assert by_name["overlayprov"]["in_base"] is False
    assert by_name["overlayprov"]["in_overlay"] is True
    detail = client.get("/admin/api/providers/neuralwatt").json()
    assert detail["in_base"] is True
    assert detail["in_overlay"] is False
    detail = client.get("/admin/api/providers/overlayprov").json()
    assert detail["in_base"] is False
    assert detail["in_overlay"] is True
--- a/tests/test_chat_completions.py
+++ b/tests/test_chat_completions.py
@@ -936,6 +936,127 @@ def test_a_non_neuralwatt_vendor_pin_is_not_stripped(router, monkeypatch):
    assert calls[0]["body"]["model"] == "other-vendor/model_id"
 def test_pinned_openrouter_model_uses_openrouter_provider(router, monkeypatch):
    """A pin to an OpenRouter-only model must resolve to the openrouter provider.
    The old code always used ``cfg.dispatch_settings.default_provider``
    (neuralwatt) on the passthrough path, so a pin targeting an OpenRouter
    id would end up calling **neuralwatt** with an unrecognised model name,
    getting a 400.  Now ``_resolve_pinned_provider`` looks up the owning
    provider in the catalog and returns it.
    """
    client, calls, db_path = router
    conn = sqlite3.connect(db_path)
    conn.execute(
        """
        INSERT INTO models (
            model_id, provider, base_model_id, tier, context_window,
            effective_context_window, max_output_tokens,
            cost_per_1m_prompt, cost_per_1m_completion,
            supports_vision, supports_json_mode,
            latency_class, reasoning_mode, context_variant,
            access_level, availability, last_updated
        ) VALUES (?, 'openrouter', ?, 3, 262128, 192500, 16384, 0.10, 20.0,
                  1, 1, 'standard', 'default', 'full', 'public', 'active',
                  '2026-08-22T00:00:00+00:00')
        """,
        ("openai/gpt-6-astra", "openai/gpt-6-astra"),
    )
    conn.commit()
    conn.close()
    # Wire up the openrouter provider config
    from config import DispatchProvider
    monkeypatch.setitem(
        dispatcher.cfg.dispatch_providers,
        "openrouter",
        DispatchProvider(
            base_url="https://openrouter.ai/api/v1",
            api_key_env="OPENROUTER_API_KEY",
        ),
    )
    monkeypatch.setenv("OPENROUTER_API_KEY", "or-key")
    resp = client.post(
        "/v1/chat/completions",
        json={"model": "openai/gpt-6-astra", "messages": _messages()},
    )
    assert resp.status_code == 200
    assert calls[0]["body"]["model"] == "openai/gpt-6-astra"
 def test_llm_router_openrouter_prefix_strips_correctly(router, monkeypatch):
    """``llm-router/openai/gpt-6-astra`` strips to ``openai/gpt-6-astra``
    and resolves to the ``openrouter`` provider.
    The old rsplit heuristic stripped to just ``gpt-6-astra``, which didn't
    exist in any provider's table.  The new prefix-first logic strips the
    ``llm-router/`` prefix and checks if the remainder exists — for
    OpenRouter vendor/model ids this resolves correctly.
    """
    client, calls, db_path = router
    conn = sqlite3.connect(db_path)
    conn.execute(
        """
        INSERT INTO models (
            model_id, provider, base_model_id, tier, context_window,
            effective_context_window, max_output_tokens,
            cost_per_1m_prompt, cost_per_1m_completion,
            supports_vision, supports_json_mode,
            latency_class, reasoning_mode, context_variant,
            access_level, availability, last_updated
        ) VALUES (?, 'openrouter', ?, 3, 262128, 192500, 16384, 0.10, 20.0,
                  1, 1, 'standard', 'default', 'full', 'public', 'active',
                  '2026-08-22T00:00:00+00:00')
        """,
        ("openai/gpt-6-astra", "openai/gpt-6-astra"),
    )
    conn.commit()
    conn.close()
    from config import DispatchProvider
    monkeypatch.setitem(
        dispatcher.cfg.dispatch_providers,
        "openrouter",
        DispatchProvider(
            base_url="https://openrouter.ai/api/v1",
            api_key_env="OPENROUTER_API_KEY",
        ),
    )
    monkeypatch.setenv("OPENROUTER_API_KEY", "or-key")
    resp = client.post(
        "/v1/chat/completions",
        json={"model": "llm-router/openai/gpt-6-astra", "messages": _messages()},
    )
    assert resp.status_code == 200
    assert calls[0]["body"]["model"] == "openai/gpt-6-astra"
    assert resp.headers["X-Router-Model"] == "openai/gpt-6-astra"
 def test_neuralwatt_only_pin_uses_default_provider(router):
    """Models seeded only on neuralwatt resolve back to the default provider.
    The existing fixture seeds CHEAP and DEAR on the neuralwatt provider.
    ``_resolve_pinned_provider`` finds exactly one row and returns
    ``neuralwatt``, which matches ``cfg.dispatch_settings.default_provider``
    — but the critical thing is the resolution path goes through the
    catalog, not a hardcoded default.
    """
    client, calls, _ = router
    resp = client.post(
        "/v1/chat/completions",
        json={"model": CHEAP, "messages": _messages()},
    )
    assert resp.status_code == 200
    assert calls[0]["body"]["model"] == CHEAP
    assert resp.headers["X-Router-Model"] == CHEAP
 def test_the_two_pass_reroute_passes_image_capability(router):
    """The measured-context reroute must carry the image gate, not lose it."""
    client, calls, _ = router
--- a/tests/test_classifier_backoff.py
+++ b/tests/test_classifier_backoff.py
@@ -1,9 +1,10 @@
 """The classifier circuit breaker actually breaking the circuit.
 ``_last_classifier_failure`` was written by ``_record_failure`` and read by
-nothing: the cascade gated only its *cloud* step, and on a different timestamp
+nothing: the cascade gated only its *cloud* step, and on a per-provider refusal
-(``_last_account_refusal``). So a dead local classifier was re-dialled on every
+state (``_provider_refusal_since``). So a dead local classifier was re-dialled
-request. A refused connection is cheap; a hung or VPN-black-holed Ollama costs
+on every request. A refused connection is cheap; a hung or VPN-black-holed
 Ollama costs
 ``classifier.timeout_seconds`` — 120s on the reference deployment — per request,
 for as long as the outage lasts.
@@ -29,12 +30,13 @@ def _closed_circuit(monkeypatch):
    """Start every test with an empty cache and a closed circuit."""
    session_cache.clear()
    monkeypatch.setattr(dispatcher, "_last_classifier_failure", 0.0)
-    monkeypatch.setattr(dispatcher, "_last_account_refusal", 0.0)
+    dispatcher._provider_refusal_since.clear()
    # No session key -> cascade steps 1 and 2 are impossible, so a skipped
    # local attempt lands on the static fallback and nothing touches the DB.
    token = dispatcher._current_session_key.set(None)
    yield
    dispatcher._current_session_key.reset(token)
    dispatcher._provider_refusal_since.clear()
    session_cache.clear()
--- a/tests/test_classifier_cascade.py
+++ b/tests/test_classifier_cascade.py
@@ -29,8 +29,9 @@ def _clean_state(monkeypatch):
    """Each test starts with an empty cache and a closed circuit."""
    session_cache.clear()
    monkeypatch.setattr(dispatcher, "_last_classifier_failure", 0.0)
-    monkeypatch.setattr(dispatcher, "_last_account_refusal", 0.0)
+    dispatcher._provider_refusal_since.clear()
    yield
    dispatcher._provider_refusal_since.clear()
    session_cache.clear()
@@ -152,7 +153,7 @@ def test_cloud_step_used_when_configured(db, monkeypatch):
    assert got.required_context_tokens == 1234
-def test_account_refusal_skips_the_cloud_step(db, monkeypatch):
+def test_provider_refusal_skips_the_cloud_step(db, monkeypatch):
    """Out of credit -> a cloud classification is a guaranteed wasted request."""
    monkeypatch.setattr(
        dispatcher.cfg.classifier,
@@ -171,9 +172,9 @@ def test_account_refusal_skips_the_cloud_step(db, monkeypatch):
        "_cloud_classifier_client",
        lambda cf: called.append(1) or object(),
    )
-    dispatcher._record_account_refusal()
+    dispatcher._record_provider_refusal("neuralwatt")
    assert dispatcher._classify_cascade("sess-6", "sys", "user") is None
-    assert called == [], "cloud client built despite an account-level refusal"
+    assert called == [], "cloud client built despite a provider-level refusal"
 def test_cloud_failure_degrades_to_the_static_guess(db, monkeypatch):
--- a/tests/test_classifier_modes_dispatch.py
+++ b/tests/test_classifier_modes_dispatch.py
@@ -26,12 +26,13 @@ import session_cache
 def _clean_state(monkeypatch):
    session_cache.clear()
    monkeypatch.setattr(dispatcher, "_last_classifier_failure", 0.0)
-    monkeypatch.setattr(dispatcher, "_last_account_refusal", 0.0)
+    dispatcher._provider_refusal_since.clear()
    monkeypatch.setattr(dispatcher, "_cached_auto_classifier", None)
    monkeypatch.setattr(dispatcher, "_auto_classifier_resolved_at", 0.0)
    token = dispatcher._current_session_key.set(None)
    yield
    dispatcher._current_session_key.reset(token)
    dispatcher._provider_refusal_since.clear()
    session_cache.clear()
--- a/tests/test_gaming_mode.py
+++ b/tests/test_gaming_mode.py
@@ -55,12 +55,13 @@ def gaming_mode(monkeypatch):
    """Local compute off, with the cloud classifier the mode requires."""
    session_cache.clear()
    monkeypatch.setattr(dispatcher, "_last_classifier_failure", 0.0)
-    monkeypatch.setattr(dispatcher, "_last_account_refusal", 0.0)
+    dispatcher._provider_refusal_since.clear()
    monkeypatch.setattr(dispatcher.cfg.local_compute, "enabled", False)
    monkeypatch.setattr(
        dispatcher.cfg.classifier, "cloud_fallback", _cloud_fallback()
    )
    yield
    dispatcher._provider_refusal_since.clear()
    session_cache.clear()
--- a/tests/test_multi_provider_dispatch.py
+++ b/tests/test_multi_provider_dispatch.py
@@ -0,0 +1,270 @@
 """Multi-provider dispatch edge cases.
 Exercises the dispatcher where two configured providers diverge in telemetry
 shape and in per-provider refusal state. Nothing here calls real providers.
 """
 from __future__ import annotations
 import json
 import sqlite3
 from pathlib import Path
 import pytest
 from starlette.testclient import TestClient
 import dispatcher
 import session_cache
 from config import DispatchProvider
 ROOT = Path(__file__).resolve().parent.parent
 SCHEMA_SQL = (ROOT / "config" / "schema.sql").read_text()
 CHEAP = "cheap-model"
 DEAR = "dear-model"
 # Stream lines carrying OpenRouter's keepalive comment *instead* of energy/cost.
 OPENROUTER_STREAM_LINES = [
    'data: {"id":"chatcmpl-or-1","choices":[{"delta":{"content":"hel"}}]}',
    "",
    ': OPENROUTER PROCESSING',
    'data: {"id":"chatcmpl-or-1","choices":[{"delta":{"content":"lo"},'
    '"finish_reason":"stop"}],"usage":{"prompt_tokens":5,"completion_tokens":2}}',
    "",
    "data: [DONE]",
    "",
 ]
@pytest.fixture(autouse=True)
 def _clean_state(monkeypatch):
    """Isolate the module-level refusal map between tests."""
    session_cache.clear()
    dispatcher._provider_refusal_since.clear()
    yield
    dispatcher._provider_refusal_since.clear()
    session_cache.clear()
@pytest.fixture
 def router(tmp_path, monkeypatch):
    """A chat-completions fixture with two providers in the catalog."""
    db_path = tmp_path / "test.db"
    conn = sqlite3.connect(db_path)
    conn.executescript(SCHEMA_SQL)
    # Seed a cheap model under the default (neuralwatt) provider.
    conn.execute(
        """
        INSERT INTO models (
            model_id, provider, base_model_id, tier, context_window,
            effective_context_window, max_output_tokens,
            cost_per_1m_prompt, cost_per_1m_completion,
            supports_vision, supports_json_mode,
            latency_class, reasoning_mode, context_variant,
            access_level, availability, last_updated
        ) VALUES (?, ?, ?, 2, 262128, 192500, 16384, ?, ?,
                  1, 1, 'standard', 'default', 'full', 'public', 'active',
                  '2026-08-22T00:00:00+00:00')
        """,
        (CHEAP, "neuralwatt", CHEAP, 0.30 / 3, 0.30),
    )
    # Seed an OpenRouter-only frontier model.
    conn.execute(
        """
        INSERT INTO models (
            model_id, provider, base_model_id, tier, context_window,
            effective_context_window, max_output_tokens,
            cost_per_1m_prompt, cost_per_1m_completion,
            supports_vision, supports_json_mode,
            latency_class, reasoning_mode, context_variant,
            access_level, availability, last_updated
        ) VALUES (?, 'openrouter', ?, 3, 262128, 192500, 16384, 0.10, 20.0,
                  1, 1, 'standard', 'default', 'full', 'public', 'active',
                  '2026-08-22T00:00:00+00:00')
        """,
        ("openai/gpt-6-astra", "openai/gpt-6-astra"),
    )
    conn.commit()
    conn.close()
    monkeypatch.setattr(dispatcher.cfg.database, "path", str(db_path))
    monkeypatch.setattr(dispatcher.cfg.verification, "local_llm_enabled", False)
    monkeypatch.setattr(dispatcher.cfg.local_vision, "enabled", False)
    monkeypatch.setattr(dispatcher.cfg.session_cache, "enabled", False)
    monkeypatch.setattr(dispatcher.cfg.exploration, "epsilon", 0.0)
    monkeypatch.setenv("NEURALWATT_API_KEY", "test-key")
    monkeypatch.setenv("OPENROUTER_API_KEY", "or-key")
    calls = []
    def fake_post(url, headers=None, json=None, stream=False, timeout=None):
        calls.append({"url": url, "body": json, "stream": stream})
        if stream:
            return _FakeResponse(lines=OPENROUTER_STREAM_LINES)
        return _FakeResponse(_completion(json["model"]))
    monkeypatch.setattr(dispatcher.requests, "post", fake_post)
    yield TestClient(dispatcher.app), calls, db_path
 class _FakeResponse:
    """Minimal requests.Response stand-in."""
    def __init__(self, payload=None, *, status_code=200, lines=None):
        self.status_code = status_code
        self._payload = payload or {}
        self._lines = lines or []
        self.text = json.dumps(self._payload)
        self.headers = {}
        self.request = None
    def json(self):
        return self._payload
    def iter_lines(self, decode_unicode=False):
        yield from self._lines
    def close(self):
        pass
 def _completion(model):
    return {
        "id": "chatcmpl-test-1",
        "model": model,
        "choices": [
            {
                "message": {"role": "assistant", "content": "hello there"},
                "finish_reason": "stop",
            }
        ],
        "usage": {"prompt_tokens": 31, "completion_tokens": 12},
    }
 def _energy_row(db_path):
    conn = sqlite3.connect(db_path)
    conn.row_factory = sqlite3.Row
    row = conn.execute(
        "SELECT model_id, energy_kwh, cost_usd FROM energy_observations ORDER BY id DESC LIMIT 1"
    ).fetchone()
    conn.close()
    return row
 def test_openrouter_streaming_ignores_keepalive_comments_and_logs_null_telemetry(router):
    """A provider with has_energy_telemetry=false can emit `: OPENROUTER ...`
    comments without crashing the stream or logging fake telemetry.
    """
    client, calls, db_path = router
    # Wire the openrouter provider config (already in config.yaml, but ensure
    # the test fixture has it for clarity).
    dispatcher.cfg.dispatch_providers["openrouter"] = DispatchProvider(
        base_url="https://openrouter.ai/api/v1",
        api_key_env="OPENROUTER_API_KEY",
        has_energy_telemetry=False,
        enabled=True,
    )
    resp = client.post(
        "/v1/chat/completions",
        json={
            "model": "openai/gpt-6-astra",
            "messages": [{"role": "user", "content": "hi"}],
            "stream": True,
        },
    )
    assert resp.status_code == 200
    assert resp.headers["content-type"].startswith("text/event-stream")
    # The keepalive comment must reach the client unchanged.
    assert ": OPENROUTER PROCESSING" in resp.text
    # No crash on the SSE framing.
    assert "data: [DONE]" in resp.text
    row = _energy_row(db_path)
    assert row is not None
    assert row["model_id"] == "openai/gpt-6-astra"
    assert row["energy_kwh"] is None
    assert row["cost_usd"] is None
 def test_provider_refusal_is_isolated_between_providers(router, monkeypatch):
    """A refusal on neuralwatt must skip the cloud classifier only when the
    default provider is neuralwatt; a request whose default provider is
    openrouter must still be allowed to use the cloud classifier.
    """
    client, calls, db_path = router
    # Configure a cloud classifier so the cascade has something to skip.
    from types import SimpleNamespace
    cloud_called = {"neuralwatt": False, "openrouter": False}
    def fake_cloud_classify(*args, **kwargs):
        cloud_called[dispatcher.cfg.dispatch_settings.default_provider] = True
        return {
            "task_category": "reasoning_math",
            "task_tier": 3,
            "required_context_tokens": 100,
            "confidence": 0.8,
        }
    monkeypatch.setattr(
        dispatcher.cfg.classifier,
        "cloud_fallback",
        SimpleNamespace(
            base_url="https://cloud.example/v1",
            model="cloud-classifier",
            timeout_seconds=2,
            api_key_env=None,
            max_output_tokens=1024,
        ),
    )
    monkeypatch.setattr(dispatcher, "_cloud_classifier_client", lambda cf: object())
    monkeypatch.setattr(session_cache, "classify_one", fake_cloud_classify)
    # Ensure classify() degrades so the cascade reaches the cloud step.
    monkeypatch.setattr(dispatcher, "classify", lambda task, context: dispatcher.Classification(
        task_category="general_chat",
        task_tier=2,
        required_context_tokens=10,
        confidence=0.0,
        source="fallback",
    ))
    # Record a refusal for neuralwatt.
    dispatcher._record_provider_refusal("neuralwatt")
    # With default_provider=neuralwatt, the cloud step should be skipped.
    monkeypatch.setattr(dispatcher.cfg.dispatch_settings, "default_provider", "neuralwatt")
    assert dispatcher._classify_cascade("sess-nw", "sys", "user") is None
    assert cloud_called["neuralwatt"] is False
    # With default_provider=openrouter, the refusal on neuralwatt must not apply.
    monkeypatch.setattr(dispatcher.cfg.dispatch_settings, "default_provider", "openrouter")
    result = dispatcher._classify_cascade("sess-or", "sys", "user")
    assert result is not None
    assert result.source == "classifier_cloud"
    assert cloud_called["openrouter"] is True
 def test_extract_telemetry_empty_payload_has_no_energy_or_cost():
    """An empty payload yields an all-None Telemetry rather than an error."""
    telemetry = dispatcher.extract_telemetry({})
    assert telemetry.energy_kwh is None
    assert telemetry.cost_usd is None
 def test_sniff_telemetry_line_ignores_foreign_comments():
    """SSE comment lines that do not carry energy/cost (e.g. OpenRouter
    keepalives) are deliberately ignored."""
    assert dispatcher._sniff_telemetry_line(": OPENROUTER PROCESSING") is None
 def test_sniff_telemetry_line_parses_energy_comment():
    """A `: energy {...}` SSE comment line is parsed into a truthy
    (word, payload) tuple."""
    parsed = dispatcher._sniff_telemetry_line(": energy {}")
    assert parsed is not None
    word, payload = parsed
    assert word == "energy"
    assert payload == {}
--- a/tests/test_multi_provider_poller.py
+++ b/tests/test_multi_provider_poller.py
@@ -0,0 +1,676 @@
 """Tests for the multi-provider main loop in poller.py.
 Covers:
 - two providers both upsert
 - one provider fails with RequestException while the other succeeds (failure isolation)
 - disabled provider is skipped
 - zero-row provider is logged as skip
 - per-provider mark_stale scopes staleness to the queried provider
 - upstream mark_stale (provider=None) retains global behavior
 """
 from __future__ import annotations
 import sqlite3
 from datetime import datetime, timedelta, timezone
 from pathlib import Path
 from unittest.mock import MagicMock, patch
 import pytest
 import requests
 import poller
 from config import RouterConfig, load_config
 ROOT = Path(__file__).resolve().parent.parent
 SCHEMA_SQL = (ROOT / "config" / "schema.sql").read_text()
 REAL_CFG = load_config(str(ROOT / "config" / "config.yaml"))
 # ---------------------------------------------------------------------------
 # Helpers
 # ---------------------------------------------------------------------------
 def _fake_response(json_data: dict):
    fake = MagicMock()
    fake.json.return_value = json_data
    fake.raise_for_status.return_value = None
    fake.status_code = 200
    return fake
 def _catalog_item(model_id: str) -> dict:
    return {
        "id": model_id,
        "metadata": {
            "display_name": model_id.replace("-", " ").title(),
            "description": "test model",
            "pricing": {
                "input_per_million": 0.30,
                "output_per_million": 0.60,
                "cached_input_per_million": 0.20,
                "pricing_tbd": False,
            },
            "capabilities": {
                "tools": True,
                "json_mode": True,
                "vision": False,
                "reasoning": False,
            },
            "limits": {
                "max_context_length": 131072,
                "max_output_tokens": 8192,
            },
            "deprecated": False,
        },
    }
 def _openrouter_model(
    model_id: str,
    prompt: str = "0.000001",
    context_length: int = 131072,
 ):
    return {
        "id": model_id,
        "name": model_id.replace("/", " ").title(),
        "canonical_slug": model_id,
        "pricing": {"prompt": prompt, "completion": "0.000002"},
        "top_provider": {
            "context_length": context_length,
            "max_completion_tokens": 8192,
        },
        "architecture": {"input_modalities": ["text"]},
        "supported_parameters": [],
        "reasoning": {},
    }
 def _seed_models(
    conn: sqlite3.Connection,
    models: list[dict],
    provider: str = "neuralwatt",
    last_updated_days_ago: int = 0,
 ) -> None:
    ts = (
        datetime.now(timezone.utc) - timedelta(days=last_updated_days_ago)
    ).isoformat()
    for m in models:
        deprecated = bool(m.get("deprecated", False))
        availability = "deprecated" if deprecated else m.get("availability", "active")
        conn.execute(
            """
            INSERT OR REPLACE INTO models (
                model_id, provider, base_model_id, display_name,
                cost_per_1m_prompt, cost_per_1m_completion, cost_per_1m_prompt_cached,
                context_window, effective_context_window, max_output_tokens,
                supports_tools, supports_json_mode, supports_vision, supports_reasoning,
                reasoning_default_enabled, latency_class, reasoning_mode,
                context_variant, access_level, pricing_tbd, deprecated,
                availability, last_updated
            ) VALUES (?, ?, ?, ?, 0.30, 0.60, 0.20,
                      131072, 192500, 8192,
                      1, 1, 0, 0,
                      0, 'standard', 'default', 'full', 'public', 0, 0,
                      ?, ?)
            """,
            (
                m["model_id"],
                provider,
                m.get("base_model_id", m["model_id"]),
                m.get(
                    "display_name",
                    f"Model {m['model_id'].title().replace('_', ' ')}",
                ),
                availability,
                ts,
            ),
        )
    conn.commit()
 # ---------------------------------------------------------------------------
 # Fixtures
 # ---------------------------------------------------------------------------
@pytest.fixture()
 def tmp_db(tmp_path):
    """Create an empty temp DB with schema and return a helper tuple.
    Returns ``(seed_and_check, cfg)`` where ``seed_and_check`` is a callable
    that opens a connection, lets the caller seed it, and returns the
    connection.  The connection must be closed by the test.  main() opens its
    own connection, so we cannot keep a fixture-level connection open.
    """
    db_path = str(tmp_path / "test.db")
    cfg = REAL_CFG.model_copy(
        update={
            "database": REAL_CFG.database.model_copy(update={"path": db_path}),
            "freshness": REAL_CFG.freshness.model_copy(update={"stale_after_days": 300}),
        }
    )
    def connect():
        conn = sqlite3.connect(db_path)
        conn.row_factory = sqlite3.Row
        return conn
    c = connect()
    c.executescript(SCHEMA_SQL)
    c.close()
    yield connect, cfg
@pytest.fixture()
 def monkeypatch_cfg(monkeypatch):
    """Return a helper that sets up a minimal dispatch_providers config."""
    def _make(providers: dict) -> RouterConfig:
        return REAL_CFG.model_copy(
            update={
                "database": REAL_CFG.database.model_copy(
                    update={"path": str(tmp_path_provider[0] / "test.db")}
                ),
                "freshness": REAL_CFG.freshness.model_copy(
                    update={"stale_after_days": 300}
                ),
                "dispatch_providers": providers,
            }
        )
    # We'll create tmp paths on demand. Use a list so the closure gets updated.
    tmp_dirs: list[Path] = []
    def _make_with_dir(providers: dict, tmp_path: Path) -> tuple[RouterConfig, str]:
        db_path = str(tmp_path / "test.db")
        cfg = REAL_CFG.model_copy(
            update={
                "database": REAL_CFG.database.model_copy(update={"path": db_path}),
                "freshness": REAL_CFG.freshness.model_copy(
                    update={"stale_after_days": 300}
                ),
                "dispatch_providers": providers,
            }
        )
        tmp_dirs.append(tmp_path)
        return cfg, db_path
    return _make_with_dir, tmp_dirs
 # ---------------------------------------------------------------------------
 # Tests
 # ---------------------------------------------------------------------------
 def test_two_providers_both_upsert(tmp_db, monkeypatch):
    """Two enabled providers are fetched, upserted, and total reported."""
    connect, cfg = tmp_db
    monkeypatch.setattr(poller, "load_config", lambda p: cfg)
    monkeypatch.setattr(
        poller.requests,
        "get",
        lambda url, timeout: _fake_response(
            {
                "data": [
                    _catalog_item("nw-model-a"),
                    _catalog_item("nw-model-b"),
                ]
            }
        ),
    )
    exit_code = poller.main()
    conn = connect()
    try:
        nw_rows = conn.execute(
            "SELECT COUNT(*) FROM models WHERE provider='neuralwatt'"
        ).fetchone()[0]
        or_rows = conn.execute(
            "SELECT COUNT(*) FROM models WHERE provider='openrouter'"
        ).fetchone()[0]
        assert nw_rows == 2
        assert or_rows == 2
        assert exit_code == 0
    finally:
        conn.close()
 def test_one_provider_fails_other_succeeds(tmp_db, monkeypatch):
    """Failure in one provider's fetch does not abort the other."""
    connect, cfg = tmp_db
    call_count = 0
    def _failing_get(url, timeout):
        nonlocal call_count
        call_count += 1
        if "openrouter" in url:
            # OpenRouter fetch goes first (dict iteration order)
            if call_count == 1:
                return _fake_response(
                    {"data": [_openrouter_model("openai/gpt-6-astra")]}
                )
            # Second neuralwatt call
            return _fake_response({"data": [_catalog_item("model_a")]})
        # NeuralWatt fails
        raise requests.RequestException("network error")
    monkeypatch.setattr(poller, "load_config", lambda p: cfg)
    monkeypatch.setattr(poller.requests, "get", _failing_get)
    exit_code = poller.main()
    conn = connect()
    try:
        nw_rows = conn.execute(
            "SELECT COUNT(*) FROM models WHERE provider='neuralwatt'"
        ).fetchone()[0]
        or_rows = conn.execute(
            "SELECT COUNT(*) FROM models WHERE provider='openrouter'"
        ).fetchone()[0]
        assert nw_rows == 0, "failed provider should have 0 rows"
        assert or_rows == 1, "succeeded provider should have 1 row"
        assert exit_code == 0
    finally:
        conn.close()
 def test_disabled_provider_skipped(tmp_db, monkeypatch):
    """Provider with enabled=False is never fetched or upserted."""
    connect, cfg = tmp_db
    # Configure only neuralwatt as disabled
    nw_cfg = cfg.dispatch_providers.get("neuralwatt")
    if nw_cfg is None:
        nw_cfg = REAL_CFG.dispatch_providers["neuralwatt"]
    disabled_cfg = nw_cfg.model_copy(update={"enabled": False})
    cfg = cfg.model_copy(
        update={
            "dispatch_providers": {"neuralwatt": disabled_cfg}
        }
    )
    monkeypatch.setattr(poller, "load_config", lambda p: cfg)
    # If the provider is skipped, requests.get should never be called
    monkeypatch.setattr(poller.requests, "get", MagicMock(side_effect=RuntimeError("should not be called")))
    exit_code = poller.main()
    conn = connect()
    try:
        nw_rows = conn.execute(
            "SELECT COUNT(*) FROM models WHERE provider='neuralwatt'"
        ).fetchone()[0]
        assert nw_rows == 0
        assert exit_code == 0
    finally:
        conn.close()
 def test_zero_row_provider_logged_as_skip(tmp_path, monkeypatch):
    """A provider that returns an empty catalog is logged and skipped."""
    db_path = str(tmp_path / "test.db")
    conn = sqlite3.connect(db_path)
    conn.row_factory = sqlite3.Row
    conn.executescript(SCHEMA_SQL)
    conn.close()
    cfg = REAL_CFG.model_copy(
        update={
            "database": REAL_CFG.database.model_copy(update={"path": db_path}),
            "freshness": REAL_CFG.freshness.model_copy(
                update={"stale_after_days": 300}
            ),
            "dispatch_providers": {"neuralwatt": REAL_CFG.dispatch_providers["neuralwatt"]},
        }
    )
    monkeypatch.setattr(poller, "load_config", lambda p: cfg)
    monkeypatch.setattr(
        poller.requests,
        "get",
        lambda url, timeout: _fake_response({"data": []}),
    )
    import io
    stderr_capture = io.StringIO()
    monkeypatch.setattr("sys.stderr", stderr_capture)
    exit_code = poller.main()
    assert exit_code == 0
    stderr_content = stderr_capture.getvalue()
    assert "0 models" in stderr_content or "skipping" in stderr_content
 def test_in_process_tiering_runs_after_upsert(tmp_path, monkeypatch):
    """Successful poll assigns tiers so rows are routable immediately."""
    db_path = str(tmp_path / "test.db")
    conn_init = sqlite3.connect(db_path)
    conn_init.row_factory = sqlite3.Row
    conn_init.executescript(SCHEMA_SQL)
    conn_init.close()
    cfg = REAL_CFG.model_copy(
        update={
            "database": REAL_CFG.database.model_copy(update={"path": db_path}),
            "freshness": REAL_CFG.freshness.model_copy(
                update={"stale_after_days": 300}
            ),
            "dispatch_providers": {
                "openrouter": REAL_CFG.dispatch_providers["openrouter"],
            },
        }
    )
    monkeypatch.setattr(poller, "load_config", lambda p: cfg)
    monkeypatch.setattr(
        poller.requests,
        "get",
        lambda url, timeout: _fake_response(
            {"data": [_openrouter_model("openai/gpt-6-astra", context_length=200000)]}
        ),
    )
    exit_code = poller.main()
    assert exit_code == 0
    conn = sqlite3.connect(db_path)
    conn.row_factory = sqlite3.Row
    try:
        row = conn.execute(
            "SELECT tier FROM models WHERE provider='openrouter' AND model_id='openai/gpt-6-astra'"
        ).fetchone()
        assert row is not None
        assert row["tier"] is not None
    finally:
        conn.close()
 def test_per_provider_mark_stale_scopes_to_provider(tmp_db, monkeypatch):
    """Staleness update for one provider doesn't affect the other."""
    connect, cfg = tmp_db
    # Seed stale rows for both providers
    old_ts = (datetime.now(timezone.utc) - timedelta(days=10)).isoformat()
    conn = connect()
    try:
        _seed_models(
            conn,
            [{"model_id": "nw_old"}],
            provider="neuralwatt",
            last_updated_days_ago=10,
        )
        # Insert openrouter row with availability='active' and old timestamp
        conn.execute(
            """
            INSERT INTO models (
                model_id, provider, base_model_id, display_name,
                cost_per_1m_prompt, cost_per_1m_completion, cost_per_1m_prompt_cached,
                context_window, effective_context_window, max_output_tokens,
                supports_tools, supports_json_mode, supports_vision, supports_reasoning,
                reasoning_default_enabled, latency_class, reasoning_mode,
                context_variant, access_level, pricing_tbd, deprecated,
                availability, last_updated
            ) VALUES ('or_old', 'openrouter', 'or_old', 'Old OR',
                      0.30, 0.60, 0.20, 131072, 192500, 8192,
                      1, 1, 0, 0,
                      0, 'standard', 'default', 'full', 'public', 0, 0,
                      'active', ?)
            """,
            (old_ts,),
        )
        # Also insert a 'fresh' old-provider row that should NOT become stale
        fresh_ts = datetime.now(timezone.utc).isoformat()
        conn.execute(
            """
            INSERT INTO models (
                model_id, provider, base_model_id, display_name,
                cost_per_1m_prompt, cost_per_1m_completion, cost_per_1m_prompt_cached,
                context_window, effective_context_window, max_output_tokens,
                supports_tools, supports_json_mode, supports_vision, supports_reasoning,
                reasoning_default_enabled, latency_class, reasoning_mode,
                context_variant, access_level, pricing_tbd, deprecated,
                availability, last_updated
            ) VALUES ('nw_fresh', 'neuralwatt', 'nw_fresh', 'Fresh NW',
                      0.30, 0.60, 0.20, 131072, 192500, 8192,
                      1, 1, 0, 0,
                      0, 'standard', 'default', 'full', 'public', 0, 0,
                      'active', ?)
            """,
            (fresh_ts,),
        )
        conn.execute(
            """
            INSERT INTO models (
                model_id, provider, base_model_id, display_name,
                cost_per_1m_prompt, cost_per_1m_completion, cost_per_1m_prompt_cached,
                context_window, effective_context_window, max_output_tokens,
                supports_tools, supports_json_mode, supports_vision, supports_reasoning,
                reasoning_default_enabled, latency_class, reasoning_mode,
                context_variant, access_level, pricing_tbd, deprecated,
                availability, last_updated
            ) VALUES ('or_fresh', 'openrouter', 'or_fresh', 'Fresh OR',
                      0.30, 0.60, 0.20, 131072, 192500, 8192,
                      1, 1, 0, 0,
                      0, 'standard', 'default', 'full', 'public', 0, 0,
                      'active', ?)
            """,
            (fresh_ts,),
        )
        conn.commit()
    finally:
        conn.close()
    # Fresh timestamp for upsert
    monkeypatch.setattr(poller, "load_config", lambda p: cfg)
    monkeypatch.setattr(
        poller.requests,
        "get",
        lambda url, timeout: _fake_response({"data": [_catalog_item("nw_fresh")]}),
    )
    cfg = cfg.model_copy(
        update={
            "freshness": cfg.freshness.model_copy(update={"stale_after_days": 5})
        }
    )
    poller.main()
    conn = connect()
    try:
        # nw_old should have been marked stale (was active, not fetched this run)
        stale_nw = conn.execute(
            "SELECT COUNT(*) FROM models WHERE availability='stale' AND provider='neuralwatt'"
        ).fetchone()[0]
        # or_old should NOT be stale — mark_stale was only called for neuralwatt provider
        stale_or = conn.execute(
            "SELECT COUNT(*) FROM models WHERE availability='stale' AND provider='openrouter'"
        ).fetchone()[0]
        assert stale_nw == 1, f"Expected 1 stale neuralwatt row, got {stale_nw}"
        # openrouter's stale rows remain because mark_stale scoped to neuralwatt
        assert stale_or == 1, f"Expected OR stale rows unchanged, got {stale_or} (or_old was already stale from old_ts)"
    finally:
        conn.close()
 def test_unknown_provider_key_skipped(tmp_path, monkeypatch):
    """An unknown provider key in dispatch_providers is logged and skipped."""
    db_path = str(tmp_path / "test.db")
    conn_init = sqlite3.connect(db_path)
    conn_init.executescript(SCHEMA_SQL)
    conn_init.close()
    # Use a known entry as a template for the unknown provider shape.
    template = REAL_CFG.dispatch_providers.get(
        "neuralwatt", REAL_CFG.dispatch_providers["openrouter"]
    )
    unknown_cfg = template.model_copy(update={"enabled": True})
    cfg_with_unknown = REAL_CFG.model_copy(
        update={
            "database": REAL_CFG.database.model_copy(update={"path": db_path}),
            "freshness": REAL_CFG.freshness.model_copy(
                update={"stale_after_days": 300}
            ),
            "dispatch_providers": {
                "neuralwatt": REAL_CFG.dispatch_providers["neuralwatt"],
                "unknown_provider": unknown_cfg,
            },
        }
    )
    monkeypatch.setattr(poller, "load_config", lambda p: cfg_with_unknown)
    # Unknown provider should be skipped before any fetch.
    monkeypatch.setattr(
        poller.requests,
        "get",
        MagicMock(
            side_effect=lambda url, timeout: (
                _fake_response({"data": [_catalog_item("model_a")]})
            )
        ),
    )
    exit_code = poller.main()
    assert exit_code == 0
    # unknown_provider produced no rows
    unknown_rows = sqlite3.connect(db_path).execute(
        "SELECT COUNT(*) FROM models WHERE provider='unknown_provider'"
    ).fetchone()[0]
    assert unknown_rows == 0
 def test_mark_stale_global_behavior_unaffected():
    """mark_stale() with no provider arg retains the old global behavior."""
    import tempfile
    with tempfile.NamedTemporaryFile(suffix=".db") as f:
        conn = sqlite3.connect(f.name)
        conn.row_factory = sqlite3.Row
        conn.executescript(SCHEMA_SQL)
        old_ts = (datetime.now(timezone.utc) - timedelta(days=10)).isoformat()
        # Insert two rows from different providers, both active and old
        conn.execute(
            """
            INSERT INTO models (
                model_id, provider, base_model_id, display_name,
                cost_per_1m_prompt, cost_per_1m_completion, cost_per_1m_prompt_cached,
                context_window, effective_context_window, max_output_tokens,
                supports_tools, supports_json_mode, supports_vision, supports_reasoning,
                reasoning_default_enabled, latency_class, reasoning_mode,
                context_variant, access_level, pricing_tbd, deprecated,
                availability, last_updated
            ) VALUES ('model_a', 'nw', 'model_a', 'A',
                      0.30, 0.60, 0.20, 131072, 192500, 8192,
                      1, 1, 0, 0,
                      0, 'standard', 'default', 'full', 'public', 0, 0,
                      'active', ?)
            """,
            (old_ts,),
        )
        conn.execute(
            """
            INSERT INTO models (
                model_id, provider, base_model_id, display_name,
                cost_per_1m_prompt, cost_per_1m_completion, cost_per_1m_prompt_cached,
                context_window, effective_context_window, max_output_tokens,
                supports_tools, supports_json_mode, supports_vision, supports_reasoning,
                reasoning_default_enabled, latency_class, reasoning_mode,
                context_variant, access_level, pricing_tbd, deprecated,
                availability, last_updated
            ) VALUES ('model_b', 'openrouter', 'model_b', 'B',
                      0.30, 0.60, 0.20, 131072, 192500, 8192,
                      1, 1, 0, 0,
                      0, 'standard', 'default', 'full', 'public', 0, 0,
                      'active', ?)
            """,
            (old_ts,),
        )
        conn.commit()
        # Call mark_stale without provider (global behavior)
        mark_stale_fn = poller.mark_stale
        mark_stale_fn(conn, REAL_CFG)
        conn.commit()
        stale_total = conn.execute(
            "SELECT COUNT(*) FROM models WHERE availability='stale'"
        ).fetchone()[0]
        assert stale_total == 2, f"All rows should be stale globally, got {stale_total}"
        conn.close()
 def test_per_provider_mark_stale_scoped(tmp_path):
    """mark_stale(provider=...) only marks rows for that provider."""
    db_path = str(tmp_path / "test.db")
    conn = sqlite3.connect(db_path)
    conn.row_factory = sqlite3.Row
    conn.executescript(SCHEMA_SQL)
    old_ts = (datetime.now(timezone.utc) - timedelta(days=10)).isoformat()
    # Seed one active, old row for each provider
    conn.execute(
        """
        INSERT INTO models (
            model_id, provider, base_model_id, display_name,
            cost_per_1m_prompt, cost_per_1m_completion, cost_per_1m_prompt_cached,
            context_window, effective_context_window, max_output_tokens,
            supports_tools, supports_json_mode, supports_vision, supports_reasoning,
            reasoning_default_enabled, latency_class, reasoning_mode,
            context_variant, access_level, pricing_tbd, deprecated,
            availability, last_updated
        ) VALUES ('nw_old', 'neuralwatt', 'nw_old', 'A',
                  0.30, 0.60, 0.20, 131072, 192500, 8192,
                  1, 1, 0, 0,
                  0, 'standard', 'default', 'full', 'public', 0, 0,
                  'active', ?)
        """,
        (old_ts,),
    )
    conn.execute(
        """
        INSERT INTO models (
            model_id, provider, base_model_id, display_name,
            cost_per_1m_prompt, cost_per_1m_completion, cost_per_1m_prompt_cached,
            context_window, effective_context_window, max_output_tokens,
            supports_tools, supports_json_mode, supports_vision, supports_reasoning,
            reasoning_default_enabled, latency_class, reasoning_mode,
            context_variant, access_level, pricing_tbd, deprecated,
            availability, last_updated
        ) VALUES ('or_old', 'openrouter', 'or_old', 'B',
                  0.30, 0.60, 0.20, 131072, 192500, 8192,
                  1, 1, 0, 0,
                  0, 'standard', 'default', 'full', 'public', 0, 0,
                  'active', ?)
        """,
        (old_ts,),
    )
    conn.commit()
    # mark_stale scoped to neuralwatt
    poller.mark_stale(conn, REAL_CFG, provider="neuralwatt")
    stale_nw = conn.execute(
        "SELECT COUNT(*) FROM models WHERE availability='stale' AND provider='neuralwatt'"
    ).fetchone()[0]
    stale_or = conn.execute(
        "SELECT COUNT(*) FROM models WHERE availability='stale' AND provider='openrouter'"
    ).fetchone()[0]
    assert stale_nw == 1, "neuralwatt row should be stale"
    assert stale_or == 0, "openrouter row should NOT be stale when scoped to neuralwatt"
    conn.close()
--- a/tests/test_openrouter_poller.py
+++ b/tests/test_openrouter_poller.py
@@ -0,0 +1,225 @@
 """Offline tests for the OpenRouter catalog fetcher and parser in poller.py."""
 from __future__ import annotations
 from unittest.mock import MagicMock, patch
 import pytest
 import requests
 from poller import (
    OPENROUTER_VIRTUAL_ROUTERS,
    ModelRow,
    fetch_openrouter,
    parse_openrouter_model,
 )
 # ---------------------------------------------------------------------------
 # Helpers
 # ---------------------------------------------------------------------------
 def _fake_response(json_data: dict):
    fake = MagicMock()
    fake.json.return_value = json_data
    fake.raise_for_status.return_value = None
    fake.status_code = 200
    return fake
 def _openrouter_model(
    model_id: str,
    *,
    canonical_slug: str | None = None,
    prompt: str = "0.000001",
    completion: str = "0.000002",
    context_length: int = 131072,
    max_completion_tokens: int = 8192,
    modalities: list[str] | None = None,
    supported_parameters: list[str] | None = None,
    reasoning: dict | None = None,
 ):
    """Build a realistic OpenRouter /v1/models entry."""
    return {
        "id": model_id,
        "name": model_id.replace("/", " ").title(),
        "canonical_slug": canonical_slug or model_id,
        "pricing": {
            "prompt": prompt,
            "completion": completion,
            "web_search": "0.000003",
            "cache_read": "0.0000005",
        },
        "top_provider": {
            "context_length": context_length,
            "max_completion_tokens": max_completion_tokens,
        },
        "architecture": {
            "input_modalities": modalities or ["text"],
        },
        "supported_parameters": supported_parameters or [],
        "reasoning": reasoning,
    }
 # ---------------------------------------------------------------------------
 # parse_openrouter_model
 # ---------------------------------------------------------------------------
 def test_parse_openrouter_model_populates_required_fields():
    raw = _openrouter_model(
        "openai/gpt-6-astra",
        canonical_slug="gpt-6-astra",
        prompt="0.0000015",
        completion="0.0000060",
        context_length=200000,
        max_completion_tokens=32768,
        modalities=["text", "image"],
        supported_parameters=["tools", "response_format", "reasoning"],
        reasoning={"mode": "medium", "default_enabled": True},
    )
    row = parse_openrouter_model(raw)
    assert row == ModelRow(
        model_id="openai/gpt-6-astra",
        provider="openrouter",
        base_model_id="gpt-6-astra",
        display_name="Openai Gpt-6-Astra",
        cost_per_1m_prompt=1.5,
        cost_per_1m_completion=6.0,
        cost_per_1m_prompt_cached=None,
        context_window=200000,
        max_output_tokens=32768,
        supports_tools=True,
        supports_json_mode=True,
        supports_vision=True,
        supports_reasoning=True,
        reasoning_default_enabled=True,
        latency_class="standard",
        reasoning_mode="medium",
        context_variant="standard",
        access_level="public",
        pricing_tbd=False,
        deprecated=False,
    )
 def test_parse_openrouter_model_batch_variant_maps_to_flex():
    raw = _openrouter_model(
        "anthropic/claude-4-opus:batch",
        canonical_slug="claude-4-opus",
    )
    row = parse_openrouter_model(raw)
    assert row.latency_class == "flex"
    assert row.context_variant == "batch"
    assert row.model_id == "anthropic/claude-4-opus:batch"
 def test_parse_openrouter_model_free_variant_maps_to_standard():
    raw = _openrouter_model(
        "meta-llama/llama-4-maverick:free",
        canonical_slug="llama-4-maverick",
    )
    row = parse_openrouter_model(raw)
    assert row.latency_class == "standard"
    assert row.context_variant == "free"
 def test_parse_openrouter_model_reasoning_defaults():
    raw = _openrouter_model("google/gemini-2.5-flash")
    row = parse_openrouter_model(raw)
    assert row.reasoning_mode == "none"
    assert row.reasoning_default_enabled is False
    assert row.supports_reasoning is False
@pytest.mark.parametrize("router_id", sorted(OPENROUTER_VIRTUAL_ROUTERS))
 def test_parse_openrouter_model_excludes_virtual_routers(router_id):
    raw = _openrouter_model(router_id)
    assert parse_openrouter_model(raw) is None
 def test_parse_openrouter_model_marks_missing_prompt_price_as_tbd():
    raw = _openrouter_model("x-ai/grok-3", prompt="")
    row = parse_openrouter_model(raw)
    assert row.cost_per_1m_prompt is None
    assert row.pricing_tbd is True
 def test_parse_openrouter_model_gracefully_handles_malformed_prices():
    raw = _openrouter_model("x-ai/grok-3-mini", prompt="free", completion="N/A")
    row = parse_openrouter_model(raw)
    assert row.cost_per_1m_prompt is None
    assert row.cost_per_1m_completion is None
    assert row.pricing_tbd is True
 # ---------------------------------------------------------------------------
 # fetch_openrouter
 # ---------------------------------------------------------------------------
@patch.object(requests, "get")
 def test_fetch_openrouter_returns_model_rows(mock_get):
    mock_get.return_value = _fake_response(
        {
            "data": [
                _openrouter_model(
                    "openai/gpt-6-astra",
                    canonical_slug="gpt-6-astra",
                ),
                _openrouter_model(
                    "anthropic/claude-4-opus:batch",
                    canonical_slug="claude-4-opus",
                ),
            ]
        }
    )
    rows = fetch_openrouter("openrouter")
    assert len(rows) == 2
    assert rows[0].model_id == "openai/gpt-6-astra"
    assert rows[0].base_model_id == "gpt-6-astra"
    assert rows[1].model_id == "anthropic/claude-4-opus:batch"
    assert rows[1].latency_class == "flex"
    mock_get.assert_called_once_with(
        "https://openrouter.ai/api/v1/models", timeout=20
    )
@patch.object(requests, "get")
 def test_fetch_openrouter_excludes_virtual_routers(mock_get):
    mock_get.return_value = _fake_response(
        {
            "data": [
                _openrouter_model("openrouter/auto"),
                _openrouter_model("openrouter/free"),
                _openrouter_model("openrouter/fusion"),
                _openrouter_model("openai/gpt-6-astra", canonical_slug="gpt-6-astra"),
            ]
        }
    )
    rows = fetch_openrouter("openrouter")
    assert [r.model_id for r in rows] == ["openai/gpt-6-astra"]
@patch.object(requests, "get")
 def test_fetch_openrouter_propagates_request_exception(mock_get):
    mock_get.side_effect = requests.RequestException("offline")
    with pytest.raises(requests.RequestException):
        fetch_openrouter("openrouter")
--- a/tests/test_poller_freshness.py
+++ b/tests/test_poller_freshness.py
@@ -127,12 +127,17 @@ def tmp_db(tmp_path, monkeypatch):
    Yields ``(conn, cfg)``.  Caller can seed the DB; fixture closes conn at
    end.
    The fixture overrides ``dispatch_providers`` to contain only ``neuralwatt``
    (the provider under test) so the multi-provider main loop doesn't also
    try to fetch from ``openrouter`` and other providers.
    """
    db_path = str(tmp_path / "test.db")
    cfg = REAL_CFG.model_copy(
        update={
            "database": REAL_CFG.database.model_copy(update={"path": db_path}),
            "freshness": REAL_CFG.freshness.model_copy(update={"stale_after_days": 300}),
            "dispatch_providers": {"neuralwatt": REAL_CFG.dispatch_providers["neuralwatt"]},
        }
    )
    conn = sqlite3.connect(db_path)
@@ -147,11 +152,11 @@ def tmp_db(tmp_path, monkeypatch):
 # ---------------------------------------------------------------------------
 def test_zero_rows_exits_nonzero_and_marks_nothing_stale(tmp_db, monkeypatch):
-    """(a) Zero-row fetch raises CatalogTooSmall.
+    """(a) Zero-row fetch no longer aborts the process.
-    This may be returned through the requests.RequestException handler (exit
+    In the multi-provider main, CatalogTooSmall is caught and the provider
-    code 1) or raised directly; the important behavior is that the run fails
+    is skipped — the run completes with 0 rows upserted and no rows marked
-    and no rows are marked stale.
+    stale.
    """
    conn, cfg = tmp_db
    _seed_models(conn, [
@@ -159,13 +164,10 @@ def test_zero_rows_exits_nonzero_and_marks_nothing_stale(tmp_db, monkeypatch):
        {"model_id": "model_beta"},
    ], last_updated_days_ago=10)
    monkeypatch_mock = MagicMock(side_effect=poller.mark_stale)
    monkeypatch.setattr(poller, "load_config", lambda p: cfg)
    monkeypatch.setattr(poller.requests, "get", lambda url, timeout: _fake_response({"data": []}))
    monkeypatch.setattr(poller, "mark_stale", monkeypatch_mock)
-    with pytest.raises(poller.CatalogTooSmall):
+    exit_code = poller.main()
        poller.main()
    total = conn.execute(
        "SELECT COUNT(*) FROM models WHERE provider='neuralwatt'"
@@ -176,7 +178,6 @@ def test_zero_rows_exits_nonzero_and_marks_nothing_stale(tmp_db, monkeypatch):
    assert total == 2
    assert stale == 0
    assert monkeypatch_mock.call_count == 0, "mark_stale must not be called"
 def test_below_half_warns_but_proceeds(tmp_db, monkeypatch):
@@ -336,10 +337,11 @@ def test_previously_stale_row_recovers(tmp_db, monkeypatch):
    assert row["availability"] == "active"
-def test_request_exception_marks_nothing(tmp_db, monkeypatch):
+def test_request_exception_skip_provider(tmp_db, monkeypatch):
-    """(e) RequestException during fetch -> no upserts, no stale marking.
+    """(e) RequestException during fetch -> skip provider, DB untouched.
-    Existing rows stay exactly as they were.
+    With multi-provider main, a fetch error skips the provider rather than
    aborting the whole run.  Existing rows stay exactly as they were.
    """
    conn, cfg = tmp_db
    _seed_models(conn, [
@@ -365,7 +367,7 @@ def test_request_exception_marks_nothing(tmp_db, monkeypatch):
        "SELECT COUNT(*) FROM models WHERE availability='stale' AND provider='neuralwatt'"
    ).fetchone()[0]
-    assert exit_code == 1
+    assert exit_code == 0
    assert total == 2
    assert stale == 0
    neuralwatt_active = conn.execute(