fix(command-code): surface reasoning-only output as content when no text-delta (#10986)

⭐5 — freeAccessPolicy "strict" opt-in (default off): verifica candidatos de auto-combo contra estado de quota ao vivo + segurança econômica por conexão antes do dispatch (fail-closed — estado desconhecido/stale/incompleto é excluído). Zero mudança de comportamento com o default "off". 82 testes focados, eslint/prettier/typecheck limpos, docs em docs/routing/STRICT_ZERO_COST.md.

CHANGELOG.md

@@ -2,6 +2,18 @@
 
 ## [Unreleased]
 
+### ✨ New Features
+
+- **feat(sse): STRICT_ZERO_COST** — opt-in, off-by-default `freeAccessPolicy: "strict"` setting
+  that hard-verifies every auto-combo candidate against live quota state and per-connection
+  economic safety before it can be dispatched, going beyond `hidePaidModels`'s static catalog
+  check. Adds curated `hardStopGuaranteed` metadata to `FREE_MODEL_BUDGETS`, a short-TTL quota
+  cache reusing `getUsageForProvider()`, and a connection-safety guarantee: a candidate backed
+  by multiple accounts has its `allowedConnectionIds` narrowed to exactly the connections
+  independently verified `SAFE`, so dispatch can never use an unverified account. An
+  `excludeTosAvoid` guard (default `false`) is available separately for contractual risk. See
+  `docs/routing/STRICT_ZERO_COST.md`.
+
 ---
 
 ## [3.8.50] — TBD

changelog.d/fixes/10986-reasoning-only-content.md

@@ -0,0 +1 @@
+- fix(command-code): surface reasoning-only output as content when a model emits no text-delta (#10986)

changelog.d/fixes/command-code-effort-capabilities.md

@@ -0,0 +1 @@
+- fix(combo): resolve effort-suffixed command-code variants (e.g. `deepseek-v4-flash-max`) to their base model for capability lookups, so tool-bearing combo requests keep the declared priority order instead of reordering behind models with confirmed capabilities

changelog.d/fixes/opencode-merge-provider-guard.md

@@ -0,0 +1 @@
+- **OpenCode config merge:** stop `mergeOpenCodeConfig` splaying a malformed `provider` block into index keys. The root was already guarded against a non-object; the `provider` branch it spreads one level down was not, so an existing `"provider": ["a", "b"]` merged to `{"0": "a", "1": "b", …}`. Its sibling `mergeOpenCodeConfigText` already refuses the same input.

config/quality/file-size-baseline.json

@@ -455,9 +455,10 @@
     "src/sse/handlers/chatHelpers.ts": 1019,
     "src/shared/middleware/chatBodyAdmission.ts": 1005,
     "_rebaseline_2026_08_20_10668_tabitoken_gateway": "#10668 (yawar-aquil) own catalog growth: src/shared/constants/providers/apikey/gateways.ts 1268->1283 (+15, entirely this PR diff -- one new tabitoken gateway entry, data lines only; base moved from 1255 to 1268 via other merges since the PR forked). Not combination drift: reproducible on the PR branch alone, so the WS5.5 release-captain rule does not apply. Extraction is not available -- the file is pure data (own header: \"Pure data; merged by apikey/index.ts via spread\") and already split into 6 family files under apikey/. Same precedent as _rebaseline_2026_08_14_imagetotext_servicekinds (#10275/#10291, gateways.ts 1250->1255, data lines only) and _rebaseline_2026_08_11_v3850_merge_storm_provider_registry (owner-authorized for this same file).",
-    "open-sse/executors/commandCode.ts": 1038,
+    "open-sse/executors/commandCode.ts": 1059,
     "_rebaseline_2026_08_21_10859_vision_bridge_catalog": "#10859 own growth (Vision Bridge fixes #10808/#10809): src/lib/modelCapabilities.ts 1006->1016 (+10, cmd/gpt-5.3-codex* text-only capability resolution) and open-sse/executors/commandCode.ts 988->1023 (+35, Command Code wire-model normalization for bare ids + reasoning field fallback for opencode-routed gateways). Cohesive bug fixes at the existing capability-resolution / executor chokepoints; not extractable mid-fix. Covered by tests/unit/model-capabilities-command-code-codex-textonly-10703.test.ts, tests/unit/command-code-vision.test.ts, tests/unit/opencode-mimo-reasoning-details-nonstream.test.ts. Pushed directly to release (own-session miss: the original rebaseline was made in a throwaway validation worktree and never landed on the PR branch or the release before merge).",
-    "_rebaseline_2026_08_21_10907_sticky_pin_clear": "#10907 own growth: open-sse/executors/commandCode.ts 1023->1038 (+15, effort-suffix sanitization threading for the sticky-pin-clear fix). Cohesive change at the existing executor chokepoint. Covered by tests/unit/command-code-executor.test.ts."
+    "_rebaseline_2026_08_21_10907_sticky_pin_clear": "#10907 own growth: open-sse/executors/commandCode.ts 1023->1038 (+15, effort-suffix sanitization threading for the sticky-pin-clear fix). Cohesive change at the existing executor chokepoint. Covered by tests/unit/command-code-executor.test.ts.",
+    "_rebaseline_2026_08_21_10986_reasoning_only_content": "#10986 own growth: open-sse/executors/commandCode.ts 1038->1059 (+21, reasoning-only content fallback — when upstream emits only reasoning-delta events and never a text-delta, surface the reasoning text as message.content in createJsonResponse and emit a synthetic content delta in createStreamResponse). Cohesive bug fix at the existing executor chokepoint (mirrors precedent style of #10907/#10859). Covered by tests/unit/command-code-executor.test.ts (2 new cases: non-stream + streaming)."
   },
   "_rebaseline_base_2026_08_10_proxyfetch": "Base-red fix (green-prs sweep, issue #9985): open-sse/utils/proxyFetch.ts 1207 > cap 1000 — new proxied-TLS fetch helper introduced by the Fal reference-image work. Owner-authorized quick rebaseline to green; structural slim tracked for v3.9.0.",
   "_rebaseline_2026_07_27_v3849_train2": "Merge-train 2 (7 PRs) — owner-approved 2026-07-27. Single entry: chatCore.ts 4955->5006 (#8595, Responses multi-turn image compaction before the context hard-reject). Genuine irreducible growth at the existing compaction chokepoint in handleChatCore — the PR adds a last-resort retry against the concrete budget plus the estimateFinalInputTokens helper, both wired at the pre-existing call site rather than a new branch. Covered by tests/unit/8560-responses-image-compaction.test.ts (4 tests).",

docs/routing/STRICT_ZERO_COST.md

@@ -0,0 +1,146 @@
+---
+title: "STRICT_ZERO_COST"
+version: 3.8.50
+lastUpdated: 2026-08-20
+---
+
+# STRICT_ZERO_COST
+
+> Opt-in, off by default (`settings.freeAccessPolicy !== "strict"` leaves every `auto/*`
+> candidate pool byte-identical). A stricter sibling of `hidePaidModels`
+> (`open-sse/services/autoCombo/paidModelFilter.ts`, #6512) for operators who need a hard
+> guarantee against ANY incremental monetary spend, not just "documented as free".
+
+## Why this exists, and why `hidePaidModels` alone isn't enough
+
+`hidePaidModels` answers "is this model classified free in `FREE_MODEL_BUDGETS` right now?" —
+a point-in-time catalog fact, checked via `isFreeModel()`/`providerHasFreeModels()`
+(`src/shared/utils/freeModels.ts`). It says nothing about two real risks:
+
+1. A `recurring-*`/`one-time-initial` free tier's allowance can be **exhausted** — the catalog
+   still lists the model as free, but the account behind it has no headroom left.
+2. Exceeding a free tier is not always a hard stop. Some providers document explicitly that no
+   payment method can ever be attached ("no credit card required"); others don't say, and a
+   handful bill automatically past the free allowance.
+
+`hidePaidModels` cannot distinguish these — it was never meant to. STRICT_ZERO_COST adds exactly
+these two checks, evaluated per candidate, **before** category/tier ranking and **before**
+dispatch — never after a request has already gone out.
+
+## Candidate classification
+
+For every candidate in the pool (`open-sse/services/autoCombo/virtualFactory.ts::buildPreparedPool`,
+right after `filterPaidOnlyCandidates`):
+
+1. **Not in `FREE_MODEL_BUDGETS` at all** → excluded. This covers genuinely paid models and any
+   provider/model OmniRoute hasn't classified yet — new candidates start excluded, not included.
+2. **`freeType: "keyless"`** → passes immediately, **but only for a candidate that genuinely
+   arrived via the no-auth path** (`connectionId === SYNTHETIC_NOAUTH_CONNECTION_ID`,
+   `open-sse/services/autoCombo/resilienceCandidateFilter.ts`). No credential exists for that
+   candidate, so no request against it can ever be billed — no runtime check is needed or
+   possible. The same catalogued `keyless` provider/model reached through a **real** DB
+   connection (`connectionId` is an actual connection id, or the candidate carries
+   `allowedConnectionIds`) does **not** get this shortcut — `keyless` metadata describes the
+   no-auth path specifically, not the provider in general, and never authorizes a real,
+   credentialed account. Such a candidate falls through to check 3 like any other, where it is
+   excluded unless the catalog entry separately carries `hardStopGuaranteed: true` (real
+   `keyless` entries never do — the shortcut was their only path to safety).
+3. **Any other `freeType`** (`recurring-daily`, `recurring-monthly`, `recurring-credit`,
+   `recurring-uncapped`, `one-time-initial`, and any future type this module doesn't
+   special-case) → passes only if **all** of the following hold:
+   - `hardStopGuaranteed: true` is set on the catalog entry (`FreeModelBudget.hardStopGuaranteed`,
+     `open-sse/config/freeModelCatalog.ts`) — a **curated, hand-set fact** about the provider's
+     own published terms (e.g. an explicit "no credit card required" claim), never derived from
+     `freeType` or from a live API response. Unset (`undefined`) and `false` are both treated as
+     "not guaranteed".
+   - A usage adapter exists for the provider in `USAGE_FETCHER_PROVIDERS`
+     (`open-sse/services/usage.ts`) — the same registry that already backs the quota dashboard and
+     `getUsageForProvider()`. No adapter → excluded, permanently, until one is added.
+   - The live, cached `FreeAccessState` for **the specific connection actually being
+     evaluated** is `status: "SAFE"`, was checked within
+     `settings.autoRefreshProviderQuotaInterval` (default 180s — the existing setting, not a new
+     number), and reports `remainingFreeAllowance` above a small safety margin.
+4. **`freeType: "discontinued"`** → always excluded.
+
+## Connection safety (per-connection verification, never per-candidate)
+
+A candidate in the auto-combo pool is not always tied to one connection. A "logical" candidate
+(`connectionId: null`) carries an `allowedConnectionIds` allowlist — one or more actual
+provider connections/accounts any of which could serve the request — and the account actually
+used is decided later, at dispatch time, by `open-sse/services/combo/autoStrategy.ts`
+(intersecting `allowedConnectionIds` against its own connection-selection logic, ~line 315-331).
+
+STRICT_ZERO_COST verifies the free-access state of **each connection in that allowlist
+individually** (`evaluateCandidateConnections()` in `strictZeroCostFilter.ts`) and rewrites
+`allowedConnectionIds` down to exactly the subset that came back `SAFE` — never the full
+original list, and never a single arbitrarily-chosen member. Concretely:
+
+- Account A `SAFE`, account B `UNKNOWN`/exhausted/billable → only A remains selectable.
+- All accounts `UNKNOWN` → the candidate is dropped entirely (empty safe set).
+- A single-connection candidate (`connectionId` set directly, no allowlist) that fails is
+  dropped outright, never returned with an empty `allowedConnectionIds`.
+
+Because `autoStrategy.ts` already enforces `allowedConnectionIds` as a hard allowlist before
+selecting a connection to dispatch to, rewriting it to the verified-SAFE subset is sufficient to
+guarantee the connection actually used at dispatch is always one this filter itself verified —
+never a different, unverified account on the same candidate. See
+`tests/unit/autoCombo/strict-zero-cost-connection-safety.test.ts` for the regression proof
+(keyless-bypass cases A/B/C, multi-account cases 1-5).
+
+`discovered automatically`: a provider/model shipped tomorrow with the right metadata (in the
+catalog, with a usage adapter, `hardStopGuaranteed: true`) is usable the moment OmniRoute knows
+about it — no code change, no whitelist entry, nothing to edit in this module. One removed from
+the catalog disappears the same way. See
+`tests/unit/autoCombo/strict-zero-cost-autodiscovery.test.ts` for the regression proof (via
+injectable fixtures, not by mutating the real catalog).
+
+## Quota caching (`open-sse/services/autoCombo/freeAccessQuota.ts`)
+
+Reuses `getUsageForProvider()` — no second quota system. A short, in-memory,
+process-lifetime cache sits in front of it (TTL equal to the default
+`autoRefreshProviderQuotaInterval`) so a Telegram-scale request rate never triggers a live
+billing-API call per candidate per request. Reads are synchronous: a cache miss returns
+`undefined` (→ excluded, fail-closed) and kicks off a background refresh for the _next_ read —
+nothing in the candidate-pool build path ever awaits a network call.
+
+`invalidateFreeAccessState(provider, connectionId)` is called from
+`src/sse/services/auth.ts::markAccountUnavailable()` the moment a connection fails for any
+reason, so the very next pool build reads a clean cache miss instead of a stale `SAFE` entry —
+no waiting out the TTL after a 402/403/quota-exhausted response.
+
+## ToS guard (independent of economic safety)
+
+`excludeTosAvoid` (default `false`) drops any candidate whose curated `tos` verdict
+(`FreeModelBudget.tos`) is `"avoid"` — reuses the same field `hidePaidModels`'s sibling docs
+(`docs/reference/FREE_TIERS.md`) already populate. Deliberately separate from
+`freeAccessPolicy`: a candidate can be economically `SAFE` and still excluded here for
+contractual reasons, or left in when this guard is off even with `freeAccessPolicy: "strict"` on.
+
+## What passes today
+
+Run `npx tsx scripts/ad-hoc/dry-run-strict-zero-cost.ts` against a live instance's
+`GET /v1/auto-combo/{channel}/candidates` output for a real before/after — the script now reads
+each candidate's real `connectionId`, so it also proves the connection-safety fix live, not just
+in unit tests. As of 2026-08-20, only `freeType: "keyless"` candidates pass in practice (7 of 29
+live candidates on this instance: `opencode/big-pickle`, `opencode/deepseek-v4-flash-free`, and
+5 `felo-web` models — all confirmed arriving with the genuine no-auth `connectionId`, never a
+real connection) — no currently-catalogued `recurring-*` provider both has a usage adapter
+registered in `USAGE_FETCHER_PROVIDERS` **and** `hardStopGuaranteed: true` declared (e.g. `groq`
+has neither the adapter registered here nor is fetched offline in this dry run; `kiro` lacks
+`hardStopGuaranteed`). This is not a bug: it's the honest state of two independently-curated
+metadata sets that happen not to overlap yet, not a limitation of the filter itself.
+
+With `excludeTosAvoid: true` added on top of the same live pool, the count drops from 7 to 0 —
+every one of the 7 surviving candidates is curated `tos: "avoid"` today (`felo-web`, `opencode`).
+This is a real, expected trade-off of turning the ToS guard on, not a bug: the guard is
+`false` by default for exactly this reason (see "ToS guard" above).
+
+## Enabling
+
+```json
+PUT /api/settings
+{ "freeAccessPolicy": "strict", "excludeTosAvoid": false }
+```
+
+Both new settings default to their pre-feature values (`"off"` / `false`) — enabling neither
+changes any existing `auto/*` routing behavior.

open-sse/config/freeModelCatalog.data.ts

@@ -108,8 +108,9 @@ export const FREE_MODEL_BUDGETS: FreeModelBudget[] = [
   { provider: "bytez", modelId: "meta-llama/Llama-3.3-70B-Instruct", displayName: "meta-llama/Llama-3.3-70B-Instruct", monthlyTokens: 0, creditTokens: 1000000, freeType: "recurring-credit", poolKey: "bytez", tos: "ambiguous" },
   { provider: "bytez", modelId: "mistralai/Mistral-7B-Instruct-v0.3", displayName: "mistralai/Mistral-7B-Instruct-v0.3", monthlyTokens: 0, creditTokens: 1000000, freeType: "recurring-credit", poolKey: "bytez", tos: "ambiguous" },
   { provider: "bytez", modelId: "Qwen/Qwen2.5-72B-Instruct", displayName: "Qwen/Qwen2.5-72B-Instruct", monthlyTokens: 0, creditTokens: 1000000, freeType: "recurring-credit", poolKey: "bytez", tos: "ambiguous" },
-  { provider: "cerebras", modelId: "zai-glm-4.7", displayName: "GLM 4.7", monthlyTokens: 30000000, creditTokens: 0, freeType: "recurring-daily", poolKey: "cerebras", tos: "caution" },
+  // hardStopGuaranteed: Cerebras pricing page states "Free Trial: 1M tokens/day... no credit card" (open-sse/services/../providers/apikey/inference-hosts.ts:74-84).
-  { provider: "cerebras", modelId: "gpt-oss-120b", displayName: "GPT OSS 120B", monthlyTokens: 30000000, creditTokens: 0, freeType: "recurring-daily", poolKey: "cerebras", tos: "caution" },
+  { provider: "cerebras", modelId: "zai-glm-4.7", displayName: "GLM 4.7", monthlyTokens: 30000000, creditTokens: 0, freeType: "recurring-daily", poolKey: "cerebras", tos: "caution", hardStopGuaranteed: true },
+  { provider: "cerebras", modelId: "gpt-oss-120b", displayName: "GPT OSS 120B", monthlyTokens: 30000000, creditTokens: 0, freeType: "recurring-daily", poolKey: "cerebras", tos: "caution", hardStopGuaranteed: true },
   // #8717: drop dead Workers AI ids (400/403/410). Keep Neurons/day budget on fp8-fast.
   { provider: "cloudflare-ai", modelId: "@cf/mistral/mistral-7b-instruct-v0.2-lora", displayName: "Mistral 7B (🆓)", monthlyTokens: 0, creditTokens: 0, freeType: "recurring-daily", poolKey: "cloudflare-ai", tos: "caution" },
   { provider: "cloudflare-ai", modelId: "@cf/qwen/qwen2.5-coder-32b-instruct", displayName: "Qwen 2.5 Coder 32B (🆓)", monthlyTokens: 0, creditTokens: 0, freeType: "recurring-daily", poolKey: "cloudflare-ai", tos: "caution" },
@@ -187,11 +188,12 @@ export const FREE_MODEL_BUDGETS: FreeModelBudget[] = [
   { provider: "glm-cn", modelId: "glm-4.5-flash", displayName: "GLM-4.5-Flash", monthlyTokens: 0, creditTokens: 0, freeType: "recurring-uncapped", poolKey: "zhipu-flash-free", tos: "ok" },
   { provider: "glm-cn", modelId: "glm-4.7-flash", displayName: "GLM-4.7-Flash", monthlyTokens: 0, creditTokens: 0, freeType: "recurring-uncapped", poolKey: "zhipu-flash-free", tos: "ok" },
   { provider: "glm-cn", modelId: "glm-signup-bonus", displayName: "Z.AI — 20M signup bonus", monthlyTokens: 0, creditTokens: 20000000, freeType: "one-time-initial", poolKey: "zhipu-signup", tos: "ok" },
-  { provider: "groq", modelId: "meta-llama/llama-4-scout-17b-16e-instruct", displayName: "Llama 4 Scout", monthlyTokens: 15000000, creditTokens: 0, freeType: "recurring-daily", poolKey: "groq", tos: "caution" },
+  // hardStopGuaranteed: Groq pricing page states "Free tier: 30 RPM / 14.4K RPD — no credit card" (open-sse/services/../providers/apikey/frontier-labs.ts:71-81).
-  { provider: "groq", modelId: "llama-3.3-70b-versatile", displayName: "Llama 3.3 70B", monthlyTokens: 15000000, creditTokens: 0, freeType: "recurring-daily", poolKey: "groq", tos: "caution" },
+  { provider: "groq", modelId: "meta-llama/llama-4-scout-17b-16e-instruct", displayName: "Llama 4 Scout", monthlyTokens: 15000000, creditTokens: 0, freeType: "recurring-daily", poolKey: "groq", tos: "caution", hardStopGuaranteed: true },
-  { provider: "groq", modelId: "openai/gpt-oss-120b", displayName: "GPT-OSS 120B", monthlyTokens: 15000000, creditTokens: 0, freeType: "recurring-daily", poolKey: "groq", tos: "caution" },
+  { provider: "groq", modelId: "llama-3.3-70b-versatile", displayName: "Llama 3.3 70B", monthlyTokens: 15000000, creditTokens: 0, freeType: "recurring-daily", poolKey: "groq", tos: "caution", hardStopGuaranteed: true },
-  { provider: "groq", modelId: "openai/gpt-oss-20b", displayName: "GPT-OSS 20B", monthlyTokens: 15000000, creditTokens: 0, freeType: "recurring-daily", poolKey: "groq", tos: "caution" },
+  { provider: "groq", modelId: "openai/gpt-oss-120b", displayName: "GPT-OSS 120B", monthlyTokens: 15000000, creditTokens: 0, freeType: "recurring-daily", poolKey: "groq", tos: "caution", hardStopGuaranteed: true },
-  { provider: "groq", modelId: "qwen/qwen3-32b", displayName: "Qwen3 32B", monthlyTokens: 15000000, creditTokens: 0, freeType: "recurring-daily", poolKey: "groq", tos: "caution" },
+  { provider: "groq", modelId: "openai/gpt-oss-20b", displayName: "GPT-OSS 20B", monthlyTokens: 15000000, creditTokens: 0, freeType: "recurring-daily", poolKey: "groq", tos: "caution", hardStopGuaranteed: true },
+  { provider: "groq", modelId: "qwen/qwen3-32b", displayName: "Qwen3 32B", monthlyTokens: 15000000, creditTokens: 0, freeType: "recurring-daily", poolKey: "groq", tos: "caution", hardStopGuaranteed: true },
   { provider: "hackclub", modelId: "meta-llama/llama-3.3-70b-instruct", displayName: "Llama 3.3 70B", monthlyTokens: 0, creditTokens: 0, freeType: "keyless", poolKey: "hackclub", tos: "caution" },
   { provider: "hackclub", modelId: "mistralai/mistral-7b-instruct", displayName: "Mistral 7B", monthlyTokens: 0, creditTokens: 0, freeType: "keyless", poolKey: "hackclub", tos: "caution" },
   { provider: "hackclub", modelId: "deepseek-ai/deepseek-coder-33b", displayName: "DeepSeek Coder 33B", monthlyTokens: 0, creditTokens: 0, freeType: "keyless", poolKey: "hackclub", tos: "caution" },

open-sse/config/freeModelCatalog.ts

@@ -26,6 +26,20 @@ export interface FreeModelBudget {
    * reports this per model as `mayTrainOnYourPrompts` on its public catalog.
    */
   trainsOnPrompts?: boolean;
+  /**
+   * True only when the provider's own published terms document that exceeding
+   * the free allowance is a hard stop (request refused / rate-limited) and NOT
+   * automatic pay-as-you-go billing — e.g. an explicit "no credit card
+   * required" claim on the provider's pricing page. This is a curated fact
+   * about the upstream provider, not something derivable from `freeType` or
+   * from any live API response, so it must be set by hand per entry with the
+   * source of the claim in a comment. Leave unset (undefined) whenever this
+   * isn't independently documented — `undefined` and `false` are both treated
+   * as "not guaranteed" by `strictZeroCostFilter.ts`; never default to `true`
+   * to grow the catalog. See STRICT_ZERO_COST in
+   * `open-sse/services/autoCombo/strictZeroCostFilter.ts`.
+   */
+  hardStopGuaranteed?: boolean;
 }
 
 export interface FreeModelTotals {
@@ -80,7 +94,7 @@ function fmt(n: number): string {
 function dedupedSum(
   models: FreeModelBudget[],
   pick: (m: FreeModelBudget) => number,
-  include: (m: FreeModelBudget) => boolean,
+  include: (m: FreeModelBudget) => boolean
 ): number {
   const poolMax = new Map<string, number>();
   let loose = 0;
@@ -100,30 +114,30 @@ export function computeFreeModelTotals(opts: { excludeTosAvoid?: boolean } = {})
   const steadyRecurringTokens = dedupedSum(
     models,
     (m) => m.monthlyTokens,
-    (m) => RECURRING.has(m.freeType),
+    (m) => RECURRING.has(m.freeType)
   );
   const recurringCredits = dedupedSum(
     models,
     (m) => m.creditTokens,
-    (m) => m.freeType === "recurring-credit",
+    (m) => m.freeType === "recurring-credit"
   );
   const oneTimeCredits = dedupedSum(
     models,
     (m) => m.creditTokens,
-    (m) => m.freeType === "one-time-initial",
+    (m) => m.freeType === "one-time-initial"
   );
 
   const steadyWithRecurringCreditsTokens = steadyRecurringTokens + recurringCredits;
   const firstMonthRealisticTokens = steadyWithRecurringCreditsTokens + oneTimeCredits;
 
   const poolCount = new Set(
-    models.filter((m) => RECURRING.has(m.freeType) && m.poolKey).map((m) => m.poolKey),
+    models.filter((m) => RECURRING.has(m.freeType) && m.poolKey).map((m) => m.poolKey)
   ).size;
 
   // Deposit-unlock boost: sum the FREE_TIER_BOOSTS whose pool still has a live
   // recurring model in the (optionally ToS-filtered) set.
   const livePools = new Set(
-    models.filter((m) => RECURRING.has(m.freeType) && m.poolKey).map((m) => m.poolKey),
+    models.filter((m) => RECURRING.has(m.freeType) && m.poolKey).map((m) => m.poolKey)
   );
   const boostMonthlyTokens = Object.entries(FREE_TIER_BOOSTS)
     .filter(([pool]) => livePools.has(pool))

open-sse/config/providerFieldStrips.ts

@@ -56,12 +56,21 @@ export function stripGroqUnsupportedFields<T extends Record<string, unknown>>(bo
   delete next.top_logprobs;
   if (Array.isArray(next.messages)) {
     next.messages = next.messages.map((m) => {
-      if (m && typeof m === "object" && "name" in m) {
+      if (m && typeof m === "object") {
-        const { name: _name, ...rest } = m as Record<string, unknown>;
+        const {
+          name: _name,
+          model: _model,
+          messageId: _msgId,
+          sender: _sender,
+          ...rest
+        } = m as Record<string, unknown>;
+
         return rest;
       }
+
       return m;
     });
   }
   return next as T;
 }
+

open-sse/executors/commandCode.ts

@@ -738,6 +738,7 @@ function createStreamResponse(
   const decoder = new TextDecoder();
   let buffer = "";
   let sentRole = false;
+  let sentContent = false;
   let closed = false;
   const state: AggregateState = {
     content: "",
@@ -772,7 +773,10 @@ function createStreamResponse(
         switch (event.type) {
           case "text-delta": {
             const text = stringValue(event.text) || "";
-            if (text) controller.enqueue(sse(chatCompletionChunk(id, model, { content: text })));
+            if (text) {
+              sentContent = true;
+              controller.enqueue(sse(chatCompletionChunk(id, model, { content: text })));
+            }
             state.content += text;
             break;
           }
@@ -810,6 +814,12 @@ function createStreamResponse(
             break;
           case "finish": {
             state.finishReason = mapFinishReason(event.finishReason);
+            // If the model only produced reasoning-delta events (no text-delta), the
+            // client-visible stream would otherwise end with no content. Emit one
+            // content delta carrying the accumulated reasoning text (#10986).
+            if (!sentContent && state.reasoning && state.toolCalls.length === 0) {
+              controller.enqueue(sse(chatCompletionChunk(id, model, { content: state.reasoning })));
+            }
             controller.enqueue(sse(chatCompletionChunk(id, model, {}, state.finishReason)));
             // Emit a standards-compliant usage-only chunk (choices: []) before
             // [DONE] when upstream reported usage. stream.ts's extractUsage
@@ -857,6 +867,9 @@ function createStreamResponse(
           if (!closed) {
             if (!sentRole)
               controller.enqueue(sse(chatCompletionChunk(id, model, { role: "assistant" })));
+            if (!sentContent && state.reasoning && state.toolCalls.length === 0) {
+              controller.enqueue(sse(chatCompletionChunk(id, model, { content: state.reasoning })));
+            }
             controller.enqueue(sse(chatCompletionChunk(id, model, {}, state.finishReason)));
             controller.enqueue(encoder.encode("data: [DONE]\n\n"));
             controller.close();
@@ -947,6 +960,14 @@ async function createJsonResponse(
   }
 
   const message: JsonRecord = { role: "assistant", content: state.content };
+  // Some Command Code models emit the whole answer as reasoning-delta events and
+  // never a text-delta. When that leaves content empty, surface the reasoning text
+  // as content too (#10986) so OpenAI-compatible clients get a usable answer. Keep
+  // reasoning_content populated as well for reasoning-aware clients, and do not
+  // override content when real text OR tool calls are present.
+  if (!state.content && state.reasoning && state.toolCalls.length === 0) {
+    message.content = state.reasoning;
+  }
   if (state.reasoning) message.reasoning_content = state.reasoning;
   if (state.toolCalls.length > 0) message.tool_calls = state.toolCalls;
 

open-sse/services/autoCombo/freeAccessQuota.ts

@@ -0,0 +1,209 @@
+/**
+ * Live wiring for STRICT_ZERO_COST's quota-based branch.
+ *
+ * Reuses the existing `getUsageForProvider()` (`open-sse/services/usage.ts`)
+ * instead of building a second quota system — this module only adds a short
+ * TTL cache in front of it (so a Telegram-scale request rate never triggers a
+ * live billing-API call per candidate per request) and an invalidation hook
+ * for the resilience layer to call the moment a 402/403/quota-exhausted
+ * response is observed (`accountFallback.ts`).
+ *
+ * The cache is intentionally synchronous to read: `resolveFreeAccessState()`
+ * never awaits. A cache miss returns `undefined` (→ UNKNOWN → excluded,
+ * fail-closed) and kicks off a background refresh for the *next* read —
+ * nothing here can make `strictZeroCostFilter.ts`'s pool build block on a
+ * network call.
+ */
+import {
+  getUsageForProvider,
+  USAGE_FETCHER_PROVIDERS,
+  type UsageFetcherProvider,
+} from "./../usage.ts";
+import { getCachedProviderConnections } from "@/lib/db/readCache";
+import { defaultLogger as log } from "@omniroute/open-sse/utils/logger";
+import type { FreeAccessState } from "./strictZeroCostFilter";
+
+const USAGE_FETCHER_PROVIDER_SET = new Set<string>(USAGE_FETCHER_PROVIDERS);
+
+/** Default cache TTL, reused verbatim from the already-shipped
+ * `settings.autoRefreshProviderQuotaInterval` (180s default,
+ * `src/lib/db/settings.ts`) instead of inventing a new number. */
+const FALLBACK_TTL_MS = 180_000;
+
+/** Cold-cache thundering-herd guard: after a process restart every candidate
+ * in a pool build is a simultaneous cache miss, which without a cap would
+ * fire one `getUsageForProvider()` call per distinct (provider, connection)
+ * pair in the same tick. Capping concurrent background refreshes spreads
+ * that burst out — a skipped refresh here just means this candidate stays
+ * UNKNOWN (excluded, fail-closed) until a later pool build tries again, never
+ * a correctness issue. */
+const MAX_CONCURRENT_REFRESHES = 4;
+
+/** Entries older than this are pruned outright even if nothing ever triggers
+ * a fresh refresh for that exact key again (e.g. the connection was deleted
+ * and no candidate references it anymore, so a normal stale-triggered
+ * refresh — which self-heals inside one TTL window — never fires). A sweep,
+ * not a timer: piggybacks on `resolveFreeAccessState` calls so this module
+ * never owns its own background interval. */
+const HARD_EVICTION_AGE_MS = FALLBACK_TTL_MS * 20; // 1 hour at the default TTL
+const SWEEP_EVERY_N_CALLS = 200;
+
+interface CacheEntry {
+  state: FreeAccessState;
+  fetchedAtMs: number;
+}
+
+// Keyed by `${provider}::${connectionId}` — module-level, process-lifetime
+// cache. Cleared per-entry by `invalidateFreeAccessState`, on a failed
+// refresh, or by the periodic sweep below; never wholesale.
+const cache = new Map<string, CacheEntry>();
+const inFlight = new Set<string>();
+let resolveCallCount = 0;
+
+function cacheKey(provider: string, connectionId: string): string {
+  return `${provider}::${connectionId}`;
+}
+
+// `getSettings()` is async (DB-backed); reading it synchronously here isn't
+// possible without changing `resolveFreeAccessState`'s synchronous contract.
+// Using the fallback unconditionally is equivalent in practice: it's the same
+// number as `settings.autoRefreshProviderQuotaInterval`'s own default
+// (`src/lib/db/settings.ts`), and `strictZeroCostFilter.ts`'s own
+// `maxStateAgeMs` (passed the real, live setting from `virtualFactory.ts`)
+// is the check that actually gates staleness for the STRICT_ZERO_COST
+// decision — this cache TTL only bounds how long a background refresh is
+// skipped, a looser, non-safety-critical concern.
+function ttlMs(): number {
+  return FALLBACK_TTL_MS;
+}
+
+/** Opportunistic sweep of very stale entries, run every N calls instead of on
+ * a timer. Cheap (a single Map iteration) and only ever removes entries no
+ * live candidate can plausibly still be waiting on. */
+function sweepIfDue(): void {
+  resolveCallCount += 1;
+  if (resolveCallCount % SWEEP_EVERY_N_CALLS !== 0) return;
+  const now = Date.now();
+  for (const [key, entry] of cache) {
+    if (now - entry.fetchedAtMs > HARD_EVICTION_AGE_MS) cache.delete(key);
+  }
+}
+
+/**
+ * Best-effort, provider-agnostic extraction of "how much free allowance is
+ * left" from whatever shape `getUsageForProvider()` returns for this
+ * provider today. Adapters were written for a human-readable quota display,
+ * not for this filter, so their payloads are heterogeneous; this function
+ * recognizes the two shapes already used by other read paths in this
+ * codebase (`quotas.*.remainingPercentage` / `.total`+`.remaining`, mirroring
+ * `quota_omniroute.py`'s own parsing) and returns `null` — never a guess —
+ * for anything else. `null` is treated as "not proven safe" by the filter.
+ */
+function extractRemainingAllowance(usage: unknown): number | null {
+  if (!usage || typeof usage !== "object") return null;
+  const quotas = (usage as Record<string, unknown>).quotas;
+  if (!quotas || typeof quotas !== "object") return null;
+
+  let worstPercent: number | null = null;
+  for (const raw of Object.values(quotas as Record<string, unknown>)) {
+    if (!raw || typeof raw !== "object") continue;
+    const q = raw as Record<string, unknown>;
+    if (q.unlimited === true) continue;
+    let pct: number | null =
+      typeof q.remainingPercentage === "number" ? q.remainingPercentage : null;
+    if (
+      pct === null &&
+      typeof q.total === "number" &&
+      typeof q.remaining === "number" &&
+      q.total > 0
+    ) {
+      pct = (100 * q.remaining) / q.total;
+    }
+    if (pct === null) continue;
+    worstPercent = worstPercent === null ? pct : Math.min(worstPercent, pct);
+  }
+  return worstPercent; // percentage points; the filter's threshold is compared against this unit
+}
+
+async function refresh(provider: string, connectionId: string): Promise<void> {
+  const key = cacheKey(provider, connectionId);
+  if (inFlight.has(key)) return;
+  if (inFlight.size >= MAX_CONCURRENT_REFRESHES) return; // thundering-herd guard — see const doc above
+  inFlight.add(key);
+  try {
+    const connections = await getCachedProviderConnections();
+    const connection = connections.find(
+      (c): c is Record<string, unknown> =>
+        !!c &&
+        typeof c === "object" &&
+        (c as Record<string, unknown>).id === connectionId &&
+        (c as Record<string, unknown>).provider === provider
+    );
+    if (!connection) {
+      cache.delete(key);
+      return;
+    }
+    const usage = await getUsageForProvider(
+      connection as unknown as Parameters<typeof getUsageForProvider>[0],
+      { forceRefresh: false }
+    );
+    const remaining = extractRemainingAllowance(usage);
+    const state: FreeAccessState = {
+      status: remaining === null ? "UNKNOWN" : remaining > 0 ? "SAFE" : "EXHAUSTED",
+      remainingFreeAllowance: remaining,
+      resetAt:
+        usage &&
+        typeof usage === "object" &&
+        typeof (usage as Record<string, unknown>).resetAt === "string"
+          ? ((usage as Record<string, unknown>).resetAt as string)
+          : null,
+      checkedAt: new Date().toISOString(),
+    };
+    cache.set(key, { state, fetchedAtMs: Date.now() });
+  } catch (err) {
+    // A failed lookup must never leave a stale SAFE entry behind — drop it so
+    // the next read is a clean cache miss (UNKNOWN), not a lucky reuse.
+    cache.delete(key);
+    log.warn("AUTO", "STRICT_ZERO_COST: usage refresh failed, treating as UNKNOWN", {
+      provider,
+      err: err instanceof Error ? err.message : String(err),
+    });
+  } finally {
+    inFlight.delete(key);
+  }
+}
+
+/**
+ * Synchronous read for `strictZeroCostFilter.ts`. Returns `undefined` when
+ * there's no usage adapter for this provider at all (a permanent UNKNOWN, no
+ * point ever refreshing), or on a cold/stale cache — in both cases a
+ * background refresh is kicked off (fire-and-forget, subject to the
+ * concurrency cap above) so a *later* read can benefit, but this call itself
+ * never blocks or throws.
+ */
+export function resolveFreeAccessState(
+  provider: string,
+  connectionId: string | undefined
+): FreeAccessState | undefined {
+  sweepIfDue();
+  if (!USAGE_FETCHER_PROVIDER_SET.has(provider as UsageFetcherProvider)) return undefined;
+  if (!connectionId) return undefined;
+
+  const key = cacheKey(provider, connectionId);
+  const entry = cache.get(key);
+  const fresh = entry && Date.now() - entry.fetchedAtMs <= ttlMs();
+  if (!fresh) {
+    void refresh(provider, connectionId);
+  }
+  return fresh ? entry.state : undefined;
+}
+
+/** Called by `accountFallback.ts` the moment a 402/403/quota-exhausted
+ * response is classified for a connection — drops the cached entry
+ * immediately instead of waiting out the TTL, so the very next candidate-pool
+ * build reads a clean cache miss (UNKNOWN) rather than a stale SAFE. */
+export function invalidateFreeAccessState(provider: string, connectionId: string): void {
+  cache.delete(cacheKey(provider, connectionId));
+}
+
+export const __testing = { cache, extractRemainingAllowance, sweepIfDue };

open-sse/services/autoCombo/strictZeroCostFilter.ts

@@ -0,0 +1,283 @@
+/**
+ * STRICT_ZERO_COST — an opt-in, stricter sibling of `hidePaidModels`
+ * (`paidModelFilter.ts`) for operators who need a hard guarantee against ANY
+ * incremental monetary spend, not just "documented as free".
+ *
+ * `hidePaidModels` answers "is this model classified free in FREE_MODEL_BUDGETS
+ * right now?" — a point-in-time catalog fact. It says nothing about whether a
+ * `recurring-*`/`one-time-initial` candidate's allowance has since been
+ * consumed, and nothing about whether exceeding it is a hard stop or silent
+ * pay-as-you-go billing. STRICT_ZERO_COST adds exactly those two checks,
+ * before ranking, before dispatch — never after.
+ *
+ * Design, kept deliberately close to `filterPaidOnlyCandidates`'s own stated
+ * goal: "a pure, dependency-light function so the filter is unit-testable in
+ * isolation". The live quota lookup (`getUsageForProvider`, cached with a TTL)
+ * lives in `freeAccessQuota.ts` and is injected here as a plain function —
+ * this file never imports the DB or makes a network call itself.
+ *
+ * No provider or model name appears anywhere in this file. A candidate passes
+ * or fails purely on the metadata it carries (`freeType`, `tos`,
+ * `hardStopGuaranteed`) plus, for quota-based types, a `FreeAccessState`
+ * resolved elsewhere. A future provider that ships correct metadata is
+ * handled automatically; one that doesn't is excluded automatically — see
+ * `docs/routing/STRICT_ZERO_COST.md`.
+ *
+ * ## Connection safety (fixed after code review, see `docs/routing/STRICT_ZERO_COST.md`)
+ *
+ * A candidate from `virtualFactory.ts`'s connection-based pool represents ONE
+ * provider/model pair with a set of *eligible* connections
+ * (`allowedConnectionIds`) — the actual connection used at dispatch is chosen
+ * later (session stickiness/LKGP), not by this filter. Two invariants follow:
+ *
+ *   1. The `keyless` shortcut (no live check needed, because no credential
+ *      exists) is valid ONLY for candidates that genuinely came from the
+ *      no-auth path — identified by `connectionId === SYNTHETIC_NOAUTH_CONNECTION_ID`
+ *      (`resilienceCandidateFilter.ts`). A `keyless`-catalogued model reached
+ *      through a real DB connection (the same provider also has a
+ *      credentialed connection) does NOT get the shortcut — it falls through
+ *      to the normal quota-based check like any other freeType, and is
+ *      excluded unless that specific connection independently proves SAFE.
+ *   2. For a multi-account candidate (`connectionId: null`,
+ *      `allowedConnectionIds: [...]`), each connection is checked
+ *      INDIVIDUALLY. The returned candidate's `allowedConnectionIds` is
+ *      REWRITTEN to exactly the subset proven SAFE — never the full original
+ *      list. `autoStrategy.ts` (`open-sse/services/combo/autoStrategy.ts:315-331`)
+ *      already intersects further routing against `allowedConnectionIds`
+ *      before connection selection, so rewriting it here is enough to make
+ *      "verified this connection" and "dispatch used this connection" the
+ *      same set, by construction — no new enforcement point needed.
+ */
+import {
+  FREE_MODEL_BUDGETS,
+  type FreeModelBudget,
+} from "@omniroute/open-sse/config/freeModelCatalog.ts";
+import { SYNTHETIC_NOAUTH_CONNECTION_ID } from "./resilienceCandidateFilter";
+
+/** Types whose allowance needs no runtime verification: no credential exists
+ * for the candidate at all, so no request against it can ever be billed. */
+const KEYLESS_FREE_TYPES = new Set<FreeModelBudget["freeType"]>(["keyless"]);
+
+export type FreeAccessStatus = "SAFE" | "EXHAUSTED" | "UNKNOWN";
+
+/** Live-checked allowance state for one (provider, connection) pair. Resolved
+ * and cached by `freeAccessQuota.ts`; passed in here as plain data so this
+ * module stays free of DB/network dependencies. */
+export interface FreeAccessState {
+  status: FreeAccessStatus;
+  /** Remaining free allowance in the provider's own unit (tokens, requests, or
+   * USD-equivalent) — whatever `getUsageForProvider()` reports. `null` when
+   * the provider's usage payload doesn't expose a numeric remaining figure. */
+  remainingFreeAllowance: number | null;
+  /** When the allowance next resets, if the provider reports it. */
+  resetAt: string | null;
+  /** When this state was fetched (ISO 8601) — used to detect staleness. */
+  checkedAt: string;
+}
+
+/** A candidate as this module needs to see it — a structural subset of
+ * `VirtualAutoComboCandidate` (`virtualFactory.ts`) so this file has no
+ * dependency on that module's full type. */
+export interface StrictZeroCostCandidate {
+  provider: string;
+  model: string;
+  connectionId: string | null;
+  allowedConnectionIds?: string[];
+}
+
+export interface StrictZeroCostOptions {
+  /** Master switch — mirrors `hidePaidModels`'s own off-by-default shape. */
+  enabled: boolean;
+  /**
+   * Resolves the live allowance state for ONE specific (provider, connection)
+   * pair. Returns `undefined` when no usage capability exists for the
+   * provider at all (no adapter registered in `USAGE_FETCHER_PROVIDERS`), or
+   * when the cache has nothing fresh for this exact connection — both are a
+   * meaningful, terminal UNKNOWN for that connection, not an error to retry.
+   *
+   * Synchronous by design: the caller (`virtualFactory.ts`) resolves and
+   * caches state per candidate up front, once per pool build, so this filter
+   * itself never awaits a network call and stays trivially testable.
+   */
+  resolveFreeAccessState: (provider: string, connectionId: string) => FreeAccessState | undefined;
+  /** Minimum remaining allowance (in the unit `resolveFreeAccessState` reports
+   * — percentage points for the built-in `freeAccessQuota.ts` resolver) a
+   * quota-based connection must exceed to pass. Must be >= 0; a fully-exhausted
+   * account (`remainingFreeAllowance === 0`) fails at any non-negative
+   * threshold via the strict `>` comparison below. */
+  minRemainingAllowance: number;
+  /** Maximum age, in ms, a `FreeAccessState.checkedAt` may have before it's
+   * treated as stale (→ UNKNOWN, excluded). */
+  maxStateAgeMs: number;
+  /** `now` injection for deterministic tests; defaults to `Date.now`. */
+  now?: () => number;
+  /**
+   * The free-model catalog to look candidates up against. Defaults to the
+   * real, live `FREE_MODEL_BUDGETS` — overridable so tests can prove the
+   * autodiscovery contract (a provider/model that appears in the catalog is
+   * automatically considered; one that's removed automatically disappears)
+   * with synthetic fixtures instead of mutating global state. Production
+   * callers should never pass this. Threaded through by
+   * `filterStrictZeroCostCandidates` (previously accepted but silently
+   * ignored — fixed alongside the connection-safety review).
+   */
+  catalog?: readonly FreeModelBudget[];
+}
+
+export function findBudgetEntry(
+  candidate: Pick<StrictZeroCostCandidate, "provider" | "model">,
+  catalog: readonly FreeModelBudget[] = FREE_MODEL_BUDGETS
+): FreeModelBudget | undefined {
+  return catalog.find((m) => m.provider === candidate.provider && m.modelId === candidate.model);
+}
+
+function isConnectionStateSafe(
+  provider: string,
+  connectionId: string,
+  resolveFreeAccessState: StrictZeroCostOptions["resolveFreeAccessState"],
+  options: Pick<StrictZeroCostOptions, "minRemainingAllowance" | "maxStateAgeMs" | "now">
+): boolean {
+  const state = resolveFreeAccessState(provider, connectionId);
+  if (!state) return false; // no usage adapter for this provider, or lookup never ran/is stale
+  if (state.status !== "SAFE") return false;
+
+  const now = (options.now ?? Date.now)();
+  const checkedAtMs = Date.parse(state.checkedAt);
+  if (!Number.isFinite(checkedAtMs) || now - checkedAtMs > options.maxStateAgeMs) return false;
+
+  if (state.remainingFreeAllowance === null) return false;
+  // A negative threshold would let a negative/garbage reading pass; a caller
+  // that genuinely wants "any allowance greater than zero" should pass 0.
+  if (options.minRemainingAllowance < 0) return false;
+  return state.remainingFreeAllowance > options.minRemainingAllowance;
+}
+
+/**
+ * Decide which of a candidate's connections satisfy STRICT_ZERO_COST. Pure —
+ * `resolveFreeAccessState` is the only injected side-effecting dependency,
+ * and it's a synchronous cache read (see `StrictZeroCostOptions` above).
+ *
+ * Returns the list of connection ids proven SAFE right now:
+ *   - `[SYNTHETIC_NOAUTH_CONNECTION_ID]` for a genuine no-auth candidate whose
+ *     catalog entry is `keyless` — no live check needed or possible.
+ *   - a (possibly empty) subset of the candidate's real connection id(s) for
+ *     every other case, each individually verified.
+ * An empty array means the caller must exclude the candidate entirely.
+ */
+export function evaluateCandidateConnections(
+  candidate: StrictZeroCostCandidate,
+  budgetEntry: FreeModelBudget | undefined,
+  resolveFreeAccessState: StrictZeroCostOptions["resolveFreeAccessState"],
+  options: Pick<StrictZeroCostOptions, "minRemainingAllowance" | "maxStateAgeMs" | "now">
+): string[] {
+  if (!budgetEntry) return []; // not in the catalog at all → paid, or genuinely unknown
+
+  const isGenuineNoAuthCandidate = candidate.connectionId === SYNTHETIC_NOAUTH_CONNECTION_ID;
+  if (KEYLESS_FREE_TYPES.has(budgetEntry.freeType)) {
+    // The keyless shortcut is trustworthy ONLY when this specific candidate
+    // instance actually has no credential behind it. A `keyless`-catalogued
+    // model reached through a real DB connection (connectionId is a real id,
+    // or the candidate carries allowedConnectionIds at all) must NOT take
+    // this shortcut — it falls through to the quota-based check below like
+    // any other freeType, and is excluded there unless hardStopGuaranteed is
+    // also set for it (which the curated catalog does not do for keyless
+    // entries today, so it will correctly exclude).
+    if (isGenuineNoAuthCandidate) return [SYNTHETIC_NOAUTH_CONNECTION_ID];
+  }
+  if (budgetEntry.freeType === "discontinued") return [];
+  if (isGenuineNoAuthCandidate) return []; // no-auth path but a non-keyless catalog entry: contradictory metadata, fail closed
+
+  // Every remaining freeType (recurring-*, one-time-initial, a keyless entry
+  // reached via a real connection, and any future type this module doesn't
+  // special-case) requires a documented hard stop before any live check even
+  // runs — no point burning a quota lookup on a connection we could never
+  // trust regardless of its answer.
+  if (budgetEntry.hardStopGuaranteed !== true) return [];
+
+  const candidateConnectionIds = candidate.connectionId
+    ? [candidate.connectionId]
+    : (candidate.allowedConnectionIds ?? []);
+
+  const safe: string[] = [];
+  for (const connectionId of candidateConnectionIds) {
+    if (connectionId === SYNTHETIC_NOAUTH_CONNECTION_ID) continue; // never reachable here, defensive
+    if (isConnectionStateSafe(candidate.provider, connectionId, resolveFreeAccessState, options)) {
+      safe.push(connectionId);
+    }
+  }
+  return safe;
+}
+
+/**
+ * Pool-level filter, same off-by-default identity contract as
+ * `filterPaidOnlyCandidates`. For a candidate that survives with a NARROWED
+ * connection set (the multi-account case), the returned object has
+ * `allowedConnectionIds` rewritten to exactly the SAFE subset — dispatch can
+ * then never select a connection this filter didn't verify, because
+ * `autoStrategy.ts` already enforces `allowedConnectionIds` as a hard
+ * allowlist downstream (see the module docstring above).
+ */
+export function filterStrictZeroCostCandidates<T extends StrictZeroCostCandidate>(
+  pool: T[],
+  options: StrictZeroCostOptions
+): T[] {
+  if (!options.enabled) return pool;
+
+  const kept: T[] = [];
+  let changed = false;
+  for (const candidate of pool) {
+    const budgetEntry = findBudgetEntry(candidate, options.catalog);
+    const safeConnectionIds = evaluateCandidateConnections(
+      candidate,
+      budgetEntry,
+      options.resolveFreeAccessState,
+      options
+    );
+    if (safeConnectionIds.length === 0) {
+      changed = true;
+      continue;
+    }
+
+    const isGenuineNoAuthCandidate = candidate.connectionId === SYNTHETIC_NOAUTH_CONNECTION_ID;
+    const isSingleConnectionCandidate = candidate.connectionId !== null;
+    if (isGenuineNoAuthCandidate || isSingleConnectionCandidate) {
+      // Nothing to narrow — either the no-auth sentinel, or a candidate that
+      // already pointed at exactly one connection which proved safe.
+      kept.push(candidate);
+      continue;
+    }
+
+    // Multi-account candidate: only rewrite if the safe subset is actually
+    // narrower than what was there before, to preserve the same
+    // identity-when-nothing-changed contract as `filterPaidOnlyCandidates`.
+    const original = candidate.allowedConnectionIds ?? [];
+    const isSameSet =
+      original.length === safeConnectionIds.length &&
+      safeConnectionIds.every((id) => original.includes(id));
+    if (isSameSet) {
+      kept.push(candidate);
+    } else {
+      changed = true;
+      kept.push({ ...candidate, allowedConnectionIds: safeConnectionIds });
+    }
+  }
+  return changed ? kept : pool;
+}
+
+/**
+ * Separate, optional ToS guard — kept independent from economic safety on
+ * purpose (Marco's requirement): a model can be economically SAFE and still
+ * excluded here for ToS reasons, or left in when this guard is off even if
+ * STRICT_ZERO_COST is on. Reuses the same curated `tos` field, no new data.
+ */
+export function filterTosAvoidCandidates<T extends StrictZeroCostCandidate>(
+  pool: T[],
+  excludeTosAvoid: boolean,
+  catalog?: readonly FreeModelBudget[]
+): T[] {
+  if (!excludeTosAvoid) return pool;
+  return pool.filter((candidate) => {
+    const budgetEntry = findBudgetEntry(candidate, catalog);
+    return budgetEntry?.tos !== "avoid";
+  });
+}

open-sse/services/autoCombo/virtualFactory.ts

@@ -26,6 +26,8 @@ import { buildFamilyCandidateFilter, type ModelFamily } from "./modelFamily";
 import { getHiddenModelsByProvider } from "@/models";
 import { getSyncedAvailableModelsByConnection, getCustomModels } from "@/lib/db/models";
 import { filterPaidOnlyCandidates } from "./paidModelFilter";
+import { filterStrictZeroCostCandidates, filterTosAvoidCandidates } from "./strictZeroCostFilter";
+import { resolveFreeAccessState } from "./freeAccessQuota";
 import { isModelExcludedByConnection } from "@/domain/connectionModelRules";
 import { resolveProviderAlias } from "../model.ts";
 import { filterExcludedCandidates } from "./candidateOverrides";
@@ -590,6 +592,29 @@ export async function prepareVirtualAutoComboInputs(
     // exclude paid-only backends from EVERY `auto/*` candidate pool.
     const paidFilteredPool = filterPaidOnlyCandidates(pool, settings.hidePaidModels === true);
     if (paidFilteredPool !== pool) pool = paidFilteredPool;
+
+    // STRICT_ZERO_COST: opt-in, off by default (`settings.freeAccessPolicy !== "strict"`
+    // leaves `pool` byte-identical, same contract as `hidePaidModels`). See
+    // `strictZeroCostFilter.ts` for why this is stricter than `hidePaidModels` alone —
+    // including the connection-safety invariant it enforces per-connection, not just
+    // per-candidate: `resolveFreeAccessState` here is a raw pass-through of the real
+    // per-(provider,connectionId) resolver; the filter itself decides which connection(s)
+    // on each candidate to check and rewrites `allowedConnectionIds` to the SAFE subset.
+    const strictFilteredPool = filterStrictZeroCostCandidates(pool, {
+      enabled: settings.freeAccessPolicy === "strict",
+      resolveFreeAccessState,
+      // 1 percentage point of headroom, not 0: `freeAccessQuota.ts` reports
+      // remaining allowance as a percentage, and a raw ">0" comparison would
+      // let a reading of e.g. 0.3% (rounding noise, not real headroom) pass.
+      minRemainingAllowance: 1,
+      maxStateAgeMs: (settings.autoRefreshProviderQuotaInterval ?? 180) * 1000,
+    });
+    if (strictFilteredPool !== pool) pool = strictFilteredPool;
+
+    // Separate, optional ToS guard — independent of economic safety on purpose.
+    const tosFilteredPool = filterTosAvoidCandidates(pool, settings.excludeTosAvoid === true);
+    if (tosFilteredPool !== pool) pool = tosFilteredPool;
+
     return pool;
   };
 

open-sse/utils/streamHandler.ts

@@ -589,10 +589,7 @@ function resolveSilentCloseOutcome(input: {
   if (!input.bytesWereForwarded) return null;
 
   if (!input.clientTerminalSeen) {
-    if (
+    if (input.clientResponseFormat === FORMATS.CLAUDE && input.contentWatcher.sawContent()) {
-      input.clientResponseFormat === FORMATS.CLAUDE &&
-      input.contentWatcher.sawContent()
-    ) {
       // #7699 — upstream dropped after content reached the client on a Claude
       // stream. Keep the partial response: emit a clean max_tokens completion
       // instead of an error frame so Anthropic SDK / Claude Code don't report
@@ -611,6 +608,16 @@ function resolveSilentCloseOutcome(input: {
     if (input.clientResponseFormat === FORMATS.OPENAI && input.contentWatcher.sawContent()) {
       return { kind: "error", reason: "Upstream stream ended without a terminal marker" };
     }
+    // Responses-format clients (Codex CLI and other /v1/responses consumers):
+    // a healthy OpenAI Responses stream ALWAYS terminates with an explicit
+    // `response.completed` event — it is the format's only terminal marker and
+    // carries the final status/usage. Content forwarded without it is an
+    // upstream drop, the same class as #10443 for chat completions; surface a
+    // synthetic response.failed instead of a silent close so clients report
+    // the break instead of waiting on a completion event that never comes.
+    if (isResponsesClientFormat(input.clientResponseFormat) && input.contentWatcher.sawContent()) {
+      return { kind: "error", reason: "Upstream stream ended without a terminal marker" };
+    }
   }
 
   const watcher = input.contentWatcher;

scripts/ad-hoc/dry-run-strict-zero-cost.ts

@@ -0,0 +1,105 @@
+/**
+ * Ad-hoc, one-shot dry run of STRICT_ZERO_COST against the real candidate
+ * pools currently served by this OmniRoute instance (fetched via the
+ * existing read-only `GET /v1/auto-combo/{channel}/candidates` endpoint —
+ * no changes made, no billable calls). Not wired into any test suite.
+ *
+ * Simulates the filter offline: no live usage-quota state is available
+ * (that adapter only runs inside the deployed container), so
+ * `resolveFreeAccessState` always returns `undefined` here — meaning any
+ * quota-based candidate is reported UNKNOWN unless it lacks even a usage
+ * adapter, in which case it's reported UNKNOWN for that reason instead. This
+ * intentionally shows the current, honest ceiling of what's usable today.
+ *
+ * Uses each candidate's REAL `connectionId` from the live endpoint (rather
+ * than assuming) to also exercise the post-code-review connection-safety
+ * check: a `keyless`-catalogued model whose live `connectionId` is NOT the
+ * no-auth sentinel is correctly reported as excluded here too.
+ */
+import { readFileSync } from "node:fs";
+import {
+  evaluateCandidateConnections,
+  findBudgetEntry,
+} from "../../open-sse/services/autoCombo/strictZeroCostFilter.ts";
+import { SYNTHETIC_NOAUTH_CONNECTION_ID } from "../../open-sse/services/autoCombo/resilienceCandidateFilter.ts";
+import { USAGE_FETCHER_PROVIDERS } from "../../open-sse/services/usage.ts";
+
+const usageProviders = new Set<string>(USAGE_FETCHER_PROVIDERS);
+const OPTIONS = { minRemainingAllowance: 1, maxStateAgeMs: 180_000 };
+
+interface Candidate {
+  provider: string;
+  model: string;
+  connectionId: string;
+}
+
+function loadCandidates(path: string): Candidate[] {
+  const raw = JSON.parse(readFileSync(path, "utf8"));
+  const list = Array.isArray(raw) ? raw : raw.candidates;
+  // The candidates endpoint's `model` field is the FULL "<providerOrAlias>/<modelId>"
+  // string (`modelStr` — the leading segment is sometimes the provider id,
+  // e.g. "groq/...", sometimes its short alias, e.g. "oc/..." for opencode);
+  // FREE_MODEL_BUDGETS.modelId is always bare. Strip exactly the first "/"
+  // segment (whichever form it is) so e.g. "groq/meta-llama/llama-4-scout..."
+  // becomes "meta-llama/llama-4-scout..." and "oc/big-pickle" becomes
+  // "big-pickle", matching the catalog's modelId either way.
+  return list.map((c: { provider: string; model: string; connectionId?: string }) => {
+    const slash = c.model.indexOf("/");
+    return {
+      provider: c.provider,
+      model: slash === -1 ? c.model : c.model.slice(slash + 1),
+      connectionId: c.connectionId ?? SYNTHETIC_NOAUTH_CONNECTION_ID,
+    };
+  });
+}
+
+function run(label: string, path: string): void {
+  const candidates = loadCandidates(path);
+  console.log(`\n=== ${label} — ${candidates.length} candidati live ===`);
+
+  const kept: Candidate[] = [];
+  const excluded: { candidate: Candidate; reason: string }[] = [];
+
+  for (const c of candidates) {
+    const entry = findBudgetEntry(c);
+    if (!entry) {
+      excluded.push({ candidate: c, reason: "non presente nel catalogo free curato" });
+      continue;
+    }
+    const isNoAuthConnection = c.connectionId === SYNTHETIC_NOAUTH_CONNECTION_ID;
+    if (entry.freeType === "keyless") {
+      const safe = evaluateCandidateConnections(c, entry, () => undefined, OPTIONS);
+      if (safe.length > 0) {
+        kept.push(c);
+      } else if (!isNoAuthConnection) {
+        excluded.push({
+          candidate: c,
+          reason:
+            "keyless nel catalogo ma raggiunto tramite una connessione DB reale (non il sentinel noauth) — shortcut non applicato, richiederebbe hardStopGuaranteed",
+        });
+      } else {
+        excluded.push({ candidate: c, reason: "keyless ma valutazione fallita (inatteso)" });
+      }
+      continue;
+    }
+    const hasAdapter = usageProviders.has(entry.provider);
+    const reason = !hasAdapter
+      ? `nessun usage adapter per '${entry.provider}' in USAGE_FETCHER_PROVIDERS`
+      : entry.hardStopGuaranteed !== true
+        ? "hardStopGuaranteed non dichiarato per questo modello"
+        : "nessuno stato quota live disponibile in questo dry-run offline (richiederebbe il container reale)";
+    excluded.push({ candidate: c, reason });
+  }
+
+  console.log(`PRIMA (STRICT_ZERO_COST off): ${candidates.length} candidati`);
+  console.log(`DOPO  (STRICT_ZERO_COST on):  ${kept.length} candidati sopravvissuti`);
+  console.log("Sopravvissuti:");
+  for (const c of kept) console.log(`  OK   ${c.provider}/${c.model}`);
+  console.log("Esclusi (motivo):");
+  for (const { candidate: c, reason } of excluded) {
+    console.log(`  EXCL ${c.provider}/${c.model} — ${reason}`);
+  }
+}
+
+run("auto/coding:free", process.argv[2] ?? "/tmp/dryrun_coding_free.json");
+run("auto/best-free", process.argv[3] ?? "/tmp/dryrun_best-free.json");

src/lib/db/settings.ts

@@ -238,6 +238,11 @@ export async function getSettings() {
     // (`:free` suffix, zero-price pricing, or FREE_MODEL_BUDGETS membership). Default
     // false preserves prior behaviour; opt-in only.
     hidePaidModels: false,
+    // Opt-in, default off: same shape as hidePaidModels above, but requires a
+    // live hard-stop-guaranteed quota check for non-keyless free candidates.
+    // See open-sse/services/autoCombo/strictZeroCostFilter.ts.
+    freeAccessPolicy: "off",
+    excludeTosAvoid: false,
     // #9418: Opt-in filter that hides auto/* virtual combos from the /v1/models catalog.
     // User-defined combos are unaffected; routing still works for hidden ids sent explicitly.
     hideAutoCombos: false,

src/lib/modelCapabilities.ts

@@ -254,6 +254,39 @@ function leafModelId(modelId: string | null | undefined): string | null {
   return leaf && leaf !== modelId ? leaf : null;
 }
 
+/**
+ * Effort suffixes the catalog synthesizes as `<base>-<tier>` variant ids from a
+ * base model's `supportedThinkingEfforts` (mirrors REGISTERED_EFFORT_SUFFIXES
+ * in open-sse/utils/registeredEffortVariants.ts, plus `minimal` for muse).
+ */
+const EFFORT_VARIANT_SUFFIXES = [
+  "none",
+  "minimal",
+  "low",
+  "medium",
+  "high",
+  "xhigh",
+  "max",
+] as const;
+
+/**
+ * Strip a trailing effort-tier suffix off a model id (e.g.
+ * `deepseek-v4-flash-max` → `deepseek-v4-flash`). Longest token first so
+ * `xhigh` is matched before `high`. Returns null when no known suffix matches
+ * or the id would be left empty.
+ */
+function stripKnownEffortSuffix(modelId: string): string | null {
+  const normalized = String(modelId || "").trim();
+  if (!normalized) return null;
+  for (const suffix of EFFORT_VARIANT_SUFFIXES) {
+    const token = `-${suffix}`;
+    if (normalized.length > token.length && normalized.endsWith(token)) {
+      return normalized.slice(0, -token.length);
+    }
+  }
+  return null;
+}
+
 function getStaticSpec(modelId: string | null, rawModel: string | null): ModelSpec | undefined {
   if (modelId) {
     const byCanonical = getModelSpec(modelId);
@@ -704,15 +737,38 @@ export function getResolvedModelCapabilities(
   // persisted override never feeds back into the comparison that (re)writes it.
   const usePersistedOverrides = options?.persistedOverrides !== false;
   const resolved = resolveCapabilityInput(input);
-  const spec = getStaticSpec(resolved.model, resolved.rawModel);
+  let spec = getStaticSpec(resolved.model, resolved.rawModel);
-  const registryModel = getRegistryModel(resolved.provider, resolved.model);
+  let registryModel = getRegistryModel(resolved.provider, resolved.model);
-  const synced = getSyncedCapabilityForResolved(
+  let synced = getSyncedCapabilityForResolved(
     resolved.provider,
     resolved.model,
     resolved.rawModel,
     snapshot
   );
 
+  // Effort-suffixed variants (e.g. command-code `deepseek-v4-flash-max`,
+  // `meta/muse-spark-1.2-contributor-xhigh`) are synthesized in the catalog
+  // from the base model's `supportedThinkingEfforts`; they have no registry
+  // row, synced row, or static spec of their own. Without a base-model
+  // fallback the variant resolves with NULL tool/vision/context capabilities,
+  // so a tool-bearing combo request treats the target as incompatible and
+  // silently reorders it behind models with confirmed capabilities. Resolve
+  // the variant's capabilities from its base model when every direct source
+  // misses.
+  if (!spec && !registryModel && !synced && resolved.provider && resolved.model) {
+    const baseModelId = stripKnownEffortSuffix(resolved.model);
+    if (baseModelId && baseModelId !== resolved.model) {
+      spec = getStaticSpec(baseModelId, resolved.rawModel);
+      registryModel = getRegistryModel(resolved.provider, baseModelId);
+      synced = getSyncedCapabilityForResolved(
+        resolved.provider,
+        baseModelId,
+        resolved.rawModel,
+        snapshot
+      );
+    }
+  }
+
   const modalitiesInput = parseModalities(synced?.modalities_input);
   const modalitiesOutput = parseModalities(synced?.modalities_output);
   const lookupKey =

src/shared/services/opencodeConfig.ts

@@ -91,11 +91,23 @@ export const mergeOpenCodeConfig = (
       ? existingConfig
       : {};
 
+  // Same guard as the root above, one level down. Spreading a non-object here
+  // does not throw, it splays the value into index keys: an existing
+  // `"provider": ["a", "b"]` merged to `{"0": "a", "1": "b", omniroute: ... }`
+  // and a string was exploded one character per key. mergeOpenCodeConfigText
+  // refuses the same input outright, so the two disagreed on what to do with a
+  // malformed config.
+  const existingProvider = (safeConfig as Record<string, unknown>).provider;
+  const safeProvider =
+    existingProvider && typeof existingProvider === "object" && !Array.isArray(existingProvider)
+      ? (existingProvider as Record<string, unknown>)
+      : {};
+
   return {
     ...safeConfig,
     $schema: safeConfig.$schema || "https://opencode.ai/config.json",
     provider: {
-      ...((safeConfig as any).provider || {}),
+      ...safeProvider,
       omniroute: buildOpenCodeProviderConfig(input),
     },
   };

src/shared/validation/settingsSchemas.ts

@@ -128,6 +128,16 @@ export const updateSettingsSchema = z.object({
   blockedProviders: z.array(z.string().max(100)).optional(),
   noAuthFallbackDisabledProviders: z.array(z.string().max(100)).optional(),
   hidePaidModels: z.boolean().optional(),
+  // STRICT_ZERO_COST (opt-in, default "off"): stricter than hidePaidModels — a
+  // candidate must be keyless (no credential exists, so no request against it
+  // can ever be billed) OR pass a live, fresh, hard-stop-guaranteed quota
+  // check, per candidate, before ranking/dispatch. See
+  // open-sse/services/autoCombo/strictZeroCostFilter.ts.
+  freeAccessPolicy: z.enum(["off", "strict"]).optional(),
+  // Separate from freeAccessPolicy on purpose: excludes candidates whose
+  // curated `tos` verdict is "avoid" (proxy/self-hosted use conflicts with the
+  // provider's own terms) — a contractual concern, not an economic one.
+  excludeTosAvoid: z.boolean().optional(),
   hideHealthCheckLogs: z.boolean().optional(),
   hideEndpointCloudflaredTunnel: z.boolean().optional(),
   hideEndpointTailscaleFunnel: z.boolean().optional(),

src/sse/services/auth.ts

@@ -2453,6 +2453,18 @@ export async function markAccountUnavailable(
   try {
     await currentMutex;
 
+    // STRICT_ZERO_COST: this connection just failed (whatever the reason) —
+    // drop any cached "SAFE" free-allowance reading for it immediately rather
+    // than waiting out the TTL, so the very next candidate-pool build reads a
+    // clean cache miss (UNKNOWN → excluded) instead of a stale SAFE. Cheap,
+    // idempotent, and correct to over-invalidate on non-quota failures too —
+    // worst case is one extra background refresh.
+    if (provider) {
+      const { invalidateFreeAccessState } =
+        await import("@omniroute/open-sse/services/autoCombo/freeAccessQuota.ts");
+      invalidateFreeAccessState(provider, connectionId);
+    }
+
     const resourceBypass = getResource404Bypass(status, errorText, connectionId, log);
     if (resourceBypass) return resourceBypass;
 

tests/unit/autoCombo/strict-zero-cost-autodiscovery.test.ts

@@ -0,0 +1,192 @@
+/**
+ * STRICT_ZERO_COST — autodiscovery contract (Fase 5-8 of the design audit):
+ * the filter must never hardcode a provider/model allowlist. It reads
+ * whatever `catalog` (FREE_MODEL_BUDGETS-shaped) it's given, so a new SAFE
+ * entry becomes usable the moment it appears, and a removed one disappears
+ * the moment it's gone — no code change, no KITT change.
+ *
+ * Uses `evaluateCandidateConnections`'s injectable `catalog` argument (added
+ * for exactly this test) with synthetic fixtures instead of mutating the
+ * real, shared `FREE_MODEL_BUDGETS` — that constant is imported by
+ * production code and other test files; mutating it at runtime would be a
+ * global side effect this file has no business causing.
+ */
+import { test } from "vitest";
+import assert from "node:assert/strict";
+
+import {
+  evaluateCandidateConnections,
+  findBudgetEntry,
+  type FreeAccessState,
+  type StrictZeroCostCandidate,
+} from "../../../open-sse/services/autoCombo/strictZeroCostFilter.ts";
+import { SYNTHETIC_NOAUTH_CONNECTION_ID } from "../../../open-sse/services/autoCombo/resilienceCandidateFilter.ts";
+import type { FreeModelBudget } from "../../../open-sse/config/freeModelCatalog.ts";
+
+const NOW = "2026-08-20T00:00:00.000Z";
+const OPTIONS = { minRemainingAllowance: 1, maxStateAgeMs: 180_000, now: () => Date.parse(NOW) };
+
+function safeState(): FreeAccessState {
+  return { status: "SAFE", remainingFreeAllowance: 40, resetAt: null, checkedAt: NOW };
+}
+
+// `evaluateCandidateConnections` has exactly one source of truth for "is this
+// candidate documented as free": the `budgetEntry` its caller looked up via
+// `findBudgetEntry(candidate, catalog)`. These provider ids are otherwise
+// arbitrary — the fixtures below prove the behavior is driven entirely by
+// catalog membership, not by any hardcoded provider/model name.
+const KEYLESS_PROVIDER = "felo-web";
+const QUOTA_PROVIDER = "groq";
+const REAL_CONN = "conn-1";
+
+function keylessEntry(modelId: string): FreeModelBudget {
+  return {
+    provider: KEYLESS_PROVIDER,
+    modelId,
+    displayName: modelId,
+    monthlyTokens: 0,
+    creditTokens: 0,
+    freeType: "keyless",
+    poolKey: null,
+    tos: "ok",
+  };
+}
+
+function quotaEntry(modelId: string, hardStopGuaranteed = true): FreeModelBudget {
+  return {
+    provider: QUOTA_PROVIDER,
+    modelId,
+    displayName: modelId,
+    monthlyTokens: 1_000_000,
+    creditTokens: 0,
+    freeType: "recurring-daily",
+    poolKey: null,
+    tos: "ok",
+    hardStopGuaranteed,
+  };
+}
+
+// 14/18. New provider/model appears in the catalog → automatically a candidate.
+test("a brand-new keyless model appearing in the catalog is usable the instant it appears", () => {
+  const modelId = "synthetic-new-model-14";
+  const candidate: StrictZeroCostCandidate = {
+    provider: KEYLESS_PROVIDER,
+    model: modelId,
+    connectionId: SYNTHETIC_NOAUTH_CONNECTION_ID,
+  };
+  const catalogBefore: FreeModelBudget[] = []; // model does not exist yet
+  const catalogAfter: FreeModelBudget[] = [keylessEntry(modelId)]; // "OmniRoute" just added it
+
+  assert.deepEqual(
+    evaluateCandidateConnections(
+      candidate,
+      findBudgetEntry(candidate, catalogBefore),
+      () => undefined,
+      OPTIONS
+    ),
+    [],
+    "must be excluded before the catalog knows about it"
+  );
+  assert.deepEqual(
+    evaluateCandidateConnections(
+      candidate,
+      findBudgetEntry(candidate, catalogAfter),
+      () => undefined,
+      OPTIONS
+    ),
+    [SYNTHETIC_NOAUTH_CONNECTION_ID],
+    "must pass automatically the moment the SAME code sees it in the catalog — no code change made between these two calls"
+  );
+});
+
+// 15/19. Provider/model removed from the catalog → automatically disappears.
+test("a model removed from the catalog stops being usable, with no code change", () => {
+  const modelId = "synthetic-removed-model-15";
+  const candidate: StrictZeroCostCandidate = {
+    provider: KEYLESS_PROVIDER,
+    model: modelId,
+    connectionId: SYNTHETIC_NOAUTH_CONNECTION_ID,
+  };
+  const catalogBefore: FreeModelBudget[] = [keylessEntry(modelId)];
+  const catalogAfter: FreeModelBudget[] = []; // "OmniRoute" dropped the promo
+
+  assert.deepEqual(
+    evaluateCandidateConnections(
+      candidate,
+      findBudgetEntry(candidate, catalogBefore),
+      () => undefined,
+      OPTIONS
+    ),
+    [SYNTHETIC_NOAUTH_CONNECTION_ID]
+  );
+  assert.deepEqual(
+    evaluateCandidateConnections(
+      candidate,
+      findBudgetEntry(candidate, catalogAfter),
+      () => undefined,
+      OPTIONS
+    ),
+    [],
+    "must be excluded automatically once gone — no whitelist entry to delete anywhere"
+  );
+});
+
+// Case 2 of Fase 8: a quota-based model ships WITH a usage adapter and
+// hardStopGuaranteed=true and a SAFE live state — passes automatically.
+test("a new quota-based model with hardStopGuaranteed + SAFE state is usable automatically", () => {
+  const modelId = "synthetic-quota-model-safe";
+  const candidate: StrictZeroCostCandidate = {
+    provider: QUOTA_PROVIDER,
+    model: modelId,
+    connectionId: REAL_CONN,
+  };
+  const catalog: FreeModelBudget[] = [quotaEntry(modelId, true)];
+  assert.deepEqual(
+    evaluateCandidateConnections(
+      candidate,
+      findBudgetEntry(candidate, catalog),
+      () => safeState(),
+      OPTIONS
+    ),
+    [REAL_CONN]
+  );
+});
+
+// Case 3 of Fase 8: metadata incomplete (no hardStopGuaranteed) → UNKNOWN → excluded automatically.
+test("a new quota-based model WITHOUT hardStopGuaranteed is excluded automatically, even with a SAFE state", () => {
+  const modelId = "synthetic-quota-model-unguaranteed";
+  const candidate: StrictZeroCostCandidate = {
+    provider: QUOTA_PROVIDER,
+    model: modelId,
+    connectionId: REAL_CONN,
+  };
+  const entry = quotaEntry(modelId);
+  delete entry.hardStopGuaranteed; // simulate metadata that was never set, not defaulted to true
+  const catalog: FreeModelBudget[] = [entry];
+  assert.deepEqual(
+    evaluateCandidateConnections(
+      candidate,
+      findBudgetEntry(candidate, catalog),
+      () => safeState(),
+      OPTIONS
+    ),
+    []
+  );
+});
+
+// Production never overrides `catalog` — `findBudgetEntry(candidate)` with no
+// second argument must read the real, live FREE_MODEL_BUDGETS, so an entirely
+// invented provider id (one no real OmniRoute release has ever catalogued) is
+// excluded automatically, with zero whitelist to edit, when the default is used.
+test("an entirely unknown provider id is excluded automatically when the default (real) catalog is used", () => {
+  const candidate: StrictZeroCostCandidate = {
+    provider: "totally-unheard-of-provider-xyz",
+    model: "whatever",
+    connectionId: REAL_CONN,
+  };
+  assert.deepEqual(
+    evaluateCandidateConnections(candidate, findBudgetEntry(candidate), () => undefined, OPTIONS),
+    [],
+    "findBudgetEntry() with no catalog override reads the real FREE_MODEL_BUDGETS — nothing to add or remove for this provider to stay excluded"
+  );
+});

tests/unit/autoCombo/strict-zero-cost-connection-safety.test.ts

@@ -0,0 +1,303 @@
+/**
+ * STRICT_ZERO_COST — connection-safety regression guards for the two
+ * blockers found in code review of commit 5d4f881:
+ *
+ *   BLOCKER 1 (keyless bypass): the `keyless` shortcut must apply ONLY to a
+ *   candidate that genuinely came from the no-auth path
+ *   (`connectionId === SYNTHETIC_NOAUTH_CONNECTION_ID`), never to a
+ *   `keyless`-catalogued model reached through a real DB connection.
+ *
+ *   BLOCKER 2 (multi-account mismatch): a multi-account candidate's
+ *   `allowedConnectionIds` must be rewritten to exactly the SAFE subset, so
+ *   whatever `autoStrategy.ts` selects at dispatch time can never be a
+ *   connection this filter didn't itself verify.
+ *
+ * Both are pure-function tests against `evaluateCandidateConnections()` /
+ * `filterStrictZeroCostCandidates()` — no DB, no network.
+ */
+import { test } from "vitest";
+import assert from "node:assert/strict";
+
+import {
+  evaluateCandidateConnections,
+  filterStrictZeroCostCandidates,
+  type FreeAccessState,
+  type StrictZeroCostCandidate,
+} from "../../../open-sse/services/autoCombo/strictZeroCostFilter.ts";
+import { SYNTHETIC_NOAUTH_CONNECTION_ID } from "../../../open-sse/services/autoCombo/resilienceCandidateFilter.ts";
+import type { FreeModelBudget } from "../../../open-sse/config/freeModelCatalog.ts";
+
+const NOW = "2026-08-20T00:00:00.000Z";
+const OPTIONS = { minRemainingAllowance: 1, maxStateAgeMs: 180_000, now: () => Date.parse(NOW) };
+
+function safeState(overrides: Partial<FreeAccessState> = {}): FreeAccessState {
+  return {
+    status: "SAFE",
+    remainingFreeAllowance: 40,
+    resetAt: null,
+    checkedAt: NOW,
+    ...overrides,
+  };
+}
+
+// A synthetic keyless budget entry — hardStopGuaranteed deliberately absent,
+// matching the real catalog (keyless entries never set it; the shortcut is
+// what makes them safe, not this field).
+function keylessEntry(): FreeModelBudget {
+  return {
+    provider: "kp",
+    modelId: "kp-model",
+    displayName: "kp-model",
+    monthlyTokens: 0,
+    creditTokens: 0,
+    freeType: "keyless",
+    poolKey: null,
+    tos: "ok",
+  };
+}
+
+function quotaEntry(hardStopGuaranteed = true): FreeModelBudget {
+  return {
+    provider: "qp",
+    modelId: "qp-model",
+    displayName: "qp-model",
+    monthlyTokens: 1_000_000,
+    creditTokens: 0,
+    freeType: "recurring-daily",
+    poolKey: null,
+    tos: "ok",
+    hardStopGuaranteed,
+  };
+}
+
+const REAL_CONN = "real-connection-42";
+
+// ─── BLOCKER 1 — keyless bypass ─────────────────────────────────────────
+
+// A. keyless + SYNTHETIC_NOAUTH_CONNECTION_ID → PASS
+test("A: keyless + SYNTHETIC_NOAUTH_CONNECTION_ID passes with zero live checks", () => {
+  const candidate: StrictZeroCostCandidate = {
+    provider: "kp",
+    model: "kp-model",
+    connectionId: SYNTHETIC_NOAUTH_CONNECTION_ID,
+  };
+  const resolveFreeAccessState = () => {
+    throw new Error("must never be called for a genuine no-auth candidate");
+  };
+  assert.deepEqual(
+    evaluateCandidateConnections(candidate, keylessEntry(), resolveFreeAccessState, OPTIONS),
+    [SYNTHETIC_NOAUTH_CONNECTION_ID]
+  );
+});
+
+// B. same provider/model, catalogued keyless, but reached via a real DB
+// connectionId → must NOT take the keyless shortcut.
+test("B: keyless-catalogued model reached via a real connectionId does NOT get the keyless shortcut", () => {
+  const candidate: StrictZeroCostCandidate = {
+    provider: "kp",
+    model: "kp-model",
+    connectionId: REAL_CONN, // NOT the sentinel — a genuine DB connection
+  };
+  let called = false;
+  const resolveFreeAccessState = () => {
+    called = true;
+    return safeState(); // even if the connection WOULD report SAFE quota...
+  };
+  const result = evaluateCandidateConnections(
+    candidate,
+    keylessEntry(),
+    resolveFreeAccessState,
+    OPTIONS
+  );
+  // ...the entry has no hardStopGuaranteed (real keyless catalog rows never
+  // set it — the shortcut was their only path to safety), so falling through
+  // to the quota branch must still exclude it.
+  assert.deepEqual(
+    result,
+    [],
+    "must exclude — keyless metadata alone cannot authorize a real connection"
+  );
+  assert.equal(
+    called,
+    false,
+    "must not even bother resolving live state once hardStopGuaranteed is absent"
+  );
+});
+
+// C. a connection that later gains a credential still can't exploit the
+// keyless classification — same code path as B, demonstrated with a
+// multi-account (allowedConnectionIds) shape to also prove the "logical
+// candidate" branch, not just the single-connectionId branch, is covered.
+test("C: a keyless-catalogued model behind allowedConnectionIds (credentialed accounts) is never shortcut-eligible", () => {
+  const candidate: StrictZeroCostCandidate = {
+    provider: "kp",
+    model: "kp-model",
+    connectionId: null,
+    allowedConnectionIds: ["acct-with-credential-1", "acct-with-credential-2"],
+  };
+  const resolveFreeAccessState = () => safeState(); // both accounts report SAFE quota
+  const result = evaluateCandidateConnections(
+    candidate,
+    keylessEntry(),
+    resolveFreeAccessState,
+    OPTIONS
+  );
+  assert.deepEqual(
+    result,
+    [],
+    "keyless metadata must never authorize any real, credentialed account, regardless of what its quota says"
+  );
+});
+
+// Sanity: a non-keyless entry reached via the no-auth sentinel is contradictory
+// metadata (shouldn't happen in practice) and must fail closed, not silently
+// fall through to a quota check that can never resolve a state for "noauth".
+test("sanity: non-keyless freeType via the no-auth sentinel connection fails closed", () => {
+  const candidate: StrictZeroCostCandidate = {
+    provider: "qp",
+    model: "qp-model",
+    connectionId: SYNTHETIC_NOAUTH_CONNECTION_ID,
+  };
+  const result = evaluateCandidateConnections(
+    candidate,
+    quotaEntry(true),
+    () => safeState(),
+    OPTIONS
+  );
+  assert.deepEqual(result, []);
+});
+
+// ─── BLOCKER 2 — multi-account connection mismatch ─────────────────────
+
+// 1. account A SAFE, B UNKNOWN → dispatch can use ONLY A.
+test("1: A SAFE, B UNKNOWN -> safe connection set is exactly [A]", () => {
+  const candidate: StrictZeroCostCandidate = {
+    provider: "qp",
+    model: "qp-model",
+    connectionId: null,
+    allowedConnectionIds: ["A", "B"],
+  };
+  const resolve = (_provider: string, connectionId: string) =>
+    connectionId === "A" ? safeState() : undefined; // B: no state at all == UNKNOWN
+  assert.deepEqual(evaluateCandidateConnections(candidate, quotaEntry(true), resolve, OPTIONS), [
+    "A",
+  ]);
+});
+
+// 2. A EXHAUSTED, B SAFE → SOLO B.
+test("2: A EXHAUSTED, B SAFE -> safe connection set is exactly [B]", () => {
+  const candidate: StrictZeroCostCandidate = {
+    provider: "qp",
+    model: "qp-model",
+    connectionId: null,
+    allowedConnectionIds: ["A", "B"],
+  };
+  const resolve = (_provider: string, connectionId: string) =>
+    connectionId === "A"
+      ? safeState({ status: "EXHAUSTED", remainingFreeAllowance: 0 })
+      : safeState();
+  assert.deepEqual(evaluateCandidateConnections(candidate, quotaEntry(true), resolve, OPTIONS), [
+    "B",
+  ]);
+});
+
+// 3. A SAFE, B billable/unverifiable (UNKNOWN) → B not selectable.
+test("3: A SAFE, B UNKNOWN (billable/unverifiable) -> B never enters the safe set", () => {
+  const candidate: StrictZeroCostCandidate = {
+    provider: "qp",
+    model: "qp-model",
+    connectionId: null,
+    allowedConnectionIds: ["A", "B"],
+  };
+  const resolve = (_provider: string, connectionId: string) =>
+    connectionId === "A" ? safeState() : undefined;
+  const safe = evaluateCandidateConnections(candidate, quotaEntry(true), resolve, OPTIONS);
+  assert.equal(safe.includes("B"), false);
+  assert.deepEqual(safe, ["A"]);
+});
+
+// 4. tutte UNKNOWN → candidato escluso.
+test("4: all connections UNKNOWN -> candidate fully excluded (empty safe set)", () => {
+  const candidate: StrictZeroCostCandidate = {
+    provider: "qp",
+    model: "qp-model",
+    connectionId: null,
+    allowedConnectionIds: ["A", "B", "C"],
+  };
+  assert.deepEqual(
+    evaluateCandidateConnections(candidate, quotaEntry(true), () => undefined, OPTIONS),
+    []
+  );
+});
+
+// 5. account selezionato al dispatch è uno di quelli verificati SAFE — proven
+// at the pool-filter level: the returned candidate's `allowedConnectionIds`
+// is rewritten to exactly the safe subset, which is the same field
+// `autoStrategy.ts` (open-sse/services/combo/autoStrategy.ts:315-331)
+// already intersects against before connection selection — so whatever it
+// picks is provably a member of this set, by construction.
+test("5: filterStrictZeroCostCandidates rewrites allowedConnectionIds to exactly the verified-SAFE subset", () => {
+  const candidate: StrictZeroCostCandidate = {
+    provider: "qp",
+    model: "qp-model",
+    connectionId: null,
+    allowedConnectionIds: ["A", "B", "C"],
+  };
+  const resolve = (_provider: string, connectionId: string) =>
+    connectionId === "B" ? safeState() : undefined; // only B is SAFE
+  const result = filterStrictZeroCostCandidates([candidate], {
+    enabled: true,
+    resolveFreeAccessState: resolve,
+    catalog: [quotaEntry(true)],
+    ...OPTIONS,
+  });
+  assert.equal(result.length, 1);
+  assert.deepEqual(
+    result[0].allowedConnectionIds,
+    ["B"],
+    "dispatch (via autoStrategy.ts's own allowedConnectionIds intersection) can only ever select B — the one connection this filter actually verified"
+  );
+  assert.equal(
+    result[0].allowedConnectionIds?.includes("A"),
+    false,
+    "A (UNKNOWN) must never remain selectable"
+  );
+  assert.equal(
+    result[0].allowedConnectionIds?.includes("C"),
+    false,
+    "C (UNKNOWN) must never remain selectable"
+  );
+});
+
+test("multi-account candidate with an unchanged safe set is returned as the SAME reference (identity contract)", () => {
+  const candidate: StrictZeroCostCandidate = {
+    provider: "qp",
+    model: "qp-model",
+    connectionId: null,
+    allowedConnectionIds: ["A"],
+  };
+  const pool = [candidate];
+  const result = filterStrictZeroCostCandidates(pool, {
+    enabled: true,
+    resolveFreeAccessState: () => safeState(),
+    catalog: [quotaEntry(true)],
+    ...OPTIONS,
+  });
+  assert.equal(result, pool, "nothing was narrowed, so the pool array reference must be preserved");
+  assert.equal(result[0], candidate, "the candidate object itself must be preserved, not cloned");
+});
+
+test("single-connection candidate that fails is dropped, never returned with an empty allowedConnectionIds", () => {
+  const candidate: StrictZeroCostCandidate = {
+    provider: "qp",
+    model: "qp-model",
+    connectionId: REAL_CONN,
+  };
+  const result = filterStrictZeroCostCandidates([candidate], {
+    enabled: true,
+    resolveFreeAccessState: () => undefined,
+    catalog: [quotaEntry(true)],
+    ...OPTIONS,
+  });
+  assert.deepEqual(result, []);
+});

tests/unit/autoCombo/strict-zero-cost-filter.test.ts

@@ -0,0 +1,248 @@
+/**
+ * STRICT_ZERO_COST — regression guard for `strictZeroCostFilter.ts`, wired
+ * into `open-sse/services/autoCombo/virtualFactory.ts::buildPreparedPool`
+ * right after `filterPaidOnlyCandidates` (#6512). Mirrors the style/shape of
+ * `tests/unit/autoCombo/paid-model-filter-6512.test.ts`.
+ *
+ * Pure, dependency-light by design: `resolveFreeAccessState` is injected as a
+ * plain function, so no DB/network mocking is needed anywhere in this file.
+ *
+ * Connection-safety-specific cases (the keyless bypass and multi-account
+ * bugs found in code review) live in their own file,
+ * `strict-zero-cost-connection-safety.test.ts` — this file covers the
+ * single-connection / no-connection-ambiguity cases only.
+ */
+import { test } from "vitest";
+import assert from "node:assert/strict";
+
+import {
+  evaluateCandidateConnections,
+  filterStrictZeroCostCandidates,
+  filterTosAvoidCandidates,
+  type FreeAccessState,
+} from "../../../open-sse/services/autoCombo/strictZeroCostFilter.ts";
+import { SYNTHETIC_NOAUTH_CONNECTION_ID } from "../../../open-sse/services/autoCombo/resilienceCandidateFilter.ts";
+import { FREE_MODEL_BUDGETS } from "../../../open-sse/config/freeModelCatalog.ts";
+
+const NOW = "2026-08-20T00:00:00.000Z";
+const nowMs = () => Date.parse(NOW);
+
+function freshState(overrides: Partial<FreeAccessState> = {}): FreeAccessState {
+  return {
+    status: "SAFE",
+    remainingFreeAllowance: 50,
+    resetAt: null,
+    checkedAt: NOW,
+    ...overrides,
+  };
+}
+
+const BASE_OPTIONS = { minRemainingAllowance: 1, maxStateAgeMs: 180_000, now: nowMs };
+
+const REAL_CONN = "conn-real-1";
+
+// A real keyless entry from the catalog (felo-web, all models keyless/tos=avoid),
+// as a genuine no-auth candidate (the only shape that legitimately gets the shortcut).
+const KEYLESS = {
+  provider: "felo-web",
+  model: "felo-chat",
+  connectionId: SYNTHETIC_NOAUTH_CONNECTION_ID,
+};
+// A real quota-based entry with hardStopGuaranteed: true (added by this feature),
+// as a concrete single-connection candidate.
+const QUOTA_SAFE = { provider: "groq", model: "llama-3.3-70b-versatile", connectionId: REAL_CONN };
+// A real quota-based entry WITHOUT hardStopGuaranteed (agentrouter: one-time-initial,
+// no usage adapter, no documented "no credit card" claim — must never pass).
+const QUOTA_UNGUARANTEED = {
+  provider: "agentrouter",
+  model: "claude-opus-5",
+  connectionId: REAL_CONN,
+};
+// Not in FREE_MODEL_BUDGETS at all.
+const UNKNOWN_MODEL = {
+  provider: "groq",
+  model: "definitely-not-cataloged",
+  connectionId: REAL_CONN,
+};
+// A provider with no free models documented anywhere (mirrors paid-model-filter-6512's own PAID fixture).
+const PAID = { provider: "openai", model: "gpt-4o", connectionId: REAL_CONN };
+
+test("sanity: fixtures exist in the real catalog with the metadata these tests assume", () => {
+  const groqEntry = FREE_MODEL_BUDGETS.find(
+    (m) => m.provider === "groq" && m.modelId === "llama-3.3-70b-versatile"
+  );
+  assert.equal(groqEntry?.hardStopGuaranteed, true, "groq must carry hardStopGuaranteed: true");
+  const arEntry = FREE_MODEL_BUDGETS.find(
+    (m) => m.provider === "agentrouter" && m.modelId === "claude-opus-5"
+  );
+  assert.notEqual(
+    arEntry?.hardStopGuaranteed,
+    true,
+    "agentrouter must NOT carry hardStopGuaranteed: true (no documented hard-stop guarantee)"
+  );
+  const feloEntry = FREE_MODEL_BUDGETS.find(
+    (m) => m.provider === "felo-web" && m.modelId === "felo-chat"
+  );
+  assert.equal(feloEntry?.freeType, "keyless");
+  assert.equal(feloEntry?.tos, "avoid", "felo-web must be tos=avoid for the ToS-guard tests below");
+});
+
+// 1. keyless SAFE (genuine no-auth candidate) → PASS
+test("keyless candidate from the genuine no-auth path passes with no state at all", () => {
+  const entry = FREE_MODEL_BUDGETS.find(
+    (m) => m.provider === "felo-web" && m.modelId === "felo-chat"
+  );
+  assert.deepEqual(
+    evaluateCandidateConnections(KEYLESS, entry, () => undefined, BASE_OPTIONS),
+    [SYNTHETIC_NOAUTH_CONNECTION_ID]
+  );
+});
+
+// 2. paid → EXCLUDE
+test("paid candidate (no free-model documentation at all) is excluded", () => {
+  assert.deepEqual(
+    evaluateCandidateConnections(PAID, undefined, () => undefined, BASE_OPTIONS),
+    []
+  );
+});
+
+// 3 & 4. unknown model / unknown provider → EXCLUDE
+test("model absent from the free catalog is excluded even under a known provider", () => {
+  assert.deepEqual(
+    evaluateCandidateConnections(UNKNOWN_MODEL, undefined, () => undefined, BASE_OPTIONS),
+    []
+  );
+});
+
+// 5. quota SAFE + fresh + hardStop → PASS
+test("quota-based candidate with hardStopGuaranteed, fresh SAFE state above threshold passes", () => {
+  const entry = FREE_MODEL_BUDGETS.find(
+    (m) => m.provider === "groq" && m.modelId === "llama-3.3-70b-versatile"
+  );
+  assert.deepEqual(
+    evaluateCandidateConnections(
+      QUOTA_SAFE,
+      entry,
+      () => freshState({ remainingFreeAllowance: 40 }),
+      BASE_OPTIONS
+    ),
+    [REAL_CONN]
+  );
+});
+
+// 6. quota exhausted → EXCLUDE
+test("EXHAUSTED status excludes even with a fresh checkedAt", () => {
+  const entry = FREE_MODEL_BUDGETS.find(
+    (m) => m.provider === "groq" && m.modelId === "llama-3.3-70b-versatile"
+  );
+  const state = freshState({ status: "EXHAUSTED", remainingFreeAllowance: 0 });
+  assert.deepEqual(
+    evaluateCandidateConnections(QUOTA_SAFE, entry, () => state, BASE_OPTIONS),
+    []
+  );
+});
+
+// 7. usage adapter absent (no state resolvable) → EXCLUDE
+test("quota-based candidate with no resolvable state is excluded, not assumed safe", () => {
+  const entry = FREE_MODEL_BUDGETS.find(
+    (m) => m.provider === "groq" && m.modelId === "llama-3.3-70b-versatile"
+  );
+  assert.deepEqual(
+    evaluateCandidateConnections(QUOTA_SAFE, entry, () => undefined, BASE_OPTIONS),
+    []
+  );
+});
+
+// 8. usage API error → EXCLUDE (modeled as UNKNOWN status; freeAccessQuota.ts
+// deletes the cache entry on a failed refresh, which resolves to `undefined`
+// here — same assertion as #7, the important contract is "never falls back to SAFE").
+test("UNKNOWN status excludes", () => {
+  const entry = FREE_MODEL_BUDGETS.find(
+    (m) => m.provider === "groq" && m.modelId === "llama-3.3-70b-versatile"
+  );
+  const state = freshState({ status: "UNKNOWN", remainingFreeAllowance: null });
+  assert.deepEqual(
+    evaluateCandidateConnections(QUOTA_SAFE, entry, () => state, BASE_OPTIONS),
+    []
+  );
+});
+
+// 9. usage state stale → EXCLUDE
+test("stale checkedAt excludes even when status is SAFE", () => {
+  const entry = FREE_MODEL_BUDGETS.find(
+    (m) => m.provider === "groq" && m.modelId === "llama-3.3-70b-versatile"
+  );
+  const stale = freshState({ checkedAt: "2026-08-19T00:00:00.000Z" }); // >24h before NOW
+  assert.deepEqual(
+    evaluateCandidateConnections(QUOTA_SAFE, entry, () => stale, BASE_OPTIONS),
+    []
+  );
+});
+
+// 10 & 11. hardStopGuaranteed undefined / false → EXCLUDE
+test("hardStopGuaranteed undefined excludes even with a perfect SAFE state", () => {
+  const entry = FREE_MODEL_BUDGETS.find(
+    (m) => m.provider === "agentrouter" && m.modelId === "claude-opus-5"
+  );
+  assert.deepEqual(
+    evaluateCandidateConnections(QUOTA_UNGUARANTEED, entry, () => freshState(), BASE_OPTIONS),
+    []
+  );
+});
+
+test("hardStopGuaranteed explicitly false excludes", () => {
+  const entry = {
+    ...FREE_MODEL_BUDGETS[0],
+    hardStopGuaranteed: false,
+    freeType: "recurring-monthly" as const,
+  };
+  assert.deepEqual(
+    evaluateCandidateConnections(
+      { provider: entry.provider, model: entry.modelId, connectionId: REAL_CONN },
+      entry,
+      () => freshState(),
+      BASE_OPTIONS
+    ),
+    []
+  );
+});
+
+// 12 & 13. ToS guard, independent of economic evaluation
+test("tos=avoid + excludeTosAvoid=true excludes a keyless-safe candidate", () => {
+  const result = filterTosAvoidCandidates([KEYLESS], true);
+  assert.deepEqual(result, [], "felo-web (tos=avoid) must be dropped when the guard is on");
+});
+
+test("tos=avoid + excludeTosAvoid=false leaves normal economic evaluation untouched", () => {
+  const result = filterTosAvoidCandidates([KEYLESS, PAID], false);
+  assert.deepEqual(result, [KEYLESS, PAID], "guard off must be a pure identity pass-through");
+});
+
+// Pool-level filter: default-off regression guard, mirrors paid-model-filter-6512's own first test.
+test("freeAccessPolicy off (default) returns the pool UNCHANGED (identity, regression guard)", () => {
+  const pool = [KEYLESS, QUOTA_SAFE, PAID];
+  const result = filterStrictZeroCostCandidates(pool, {
+    enabled: false,
+    resolveFreeAccessState: () => undefined,
+    ...BASE_OPTIONS,
+  });
+  assert.equal(result, pool, "must return the exact same array reference when opt-in is off");
+});
+
+test("strict pool filter keeps only keyless(no-auth) + guaranteed-quota-SAFE candidates", () => {
+  const pool = [KEYLESS, QUOTA_SAFE, QUOTA_UNGUARANTEED, PAID, UNKNOWN_MODEL];
+  const result = filterStrictZeroCostCandidates(pool, {
+    enabled: true,
+    resolveFreeAccessState: (provider) =>
+      provider === "groq" ? freshState({ remainingFreeAllowance: 40 }) : undefined,
+    ...BASE_OPTIONS,
+  });
+  assert.deepEqual(result, [KEYLESS, QUOTA_SAFE]);
+});
+
+// Autodiscovery cases (14-19) are in
+// tests/unit/autoCombo/strict-zero-cost-autodiscovery.test.ts, which fabricates
+// synthetic FREE_MODEL_BUDGETS-shaped fixtures rather than real provider ids —
+// see that file for why a fixture-based test is the correct tool here.
+// Blocker 1 (keyless bypass) and Blocker 2 (multi-account) cases are in
+// tests/unit/autoCombo/strict-zero-cost-connection-safety.test.ts.

tests/unit/command-code-executor.test.ts

@@ -333,6 +333,68 @@ test("Command Code data: SSE lines aggregate into non-stream ChatCompletion JSON
   });
 });
 
+test("Command Code reasoning-only output falls back to reasoning as content (non-stream)", async () => {
+  globalThis.fetch = async () =>
+    commandCodeStream(
+      [
+        { type: "reasoning-delta", text: "The user wants 79874+93658. " },
+        { type: "reasoning-delta", text: "That equals 173532." },
+        {
+          type: "finish",
+          finishReason: "stop",
+          totalUsage: { inputTokens: 20, outputTokens: 64, outputTokenDetails: { reasoningTokens: 61 } },
+        },
+      ],
+      { sse: true }
+    );
+
+  const { response } = await getExecutor("command-code").execute({
+    model: "meta/muse-spark-1.2-contributor",
+    stream: false,
+    credentials: { apiKey: "cc_test_key" },
+    body: { messages: [{ role: "user", content: "Calculate 79874+93658, and reply with the result only." }] },
+  });
+
+  const json = await response.json();
+  const message = json.choices[0].message;
+  // Regression #10986: when the model emits only reasoning-delta events (never a
+  // text-delta), content must fall back to the reasoning text instead of "" (which
+  // OpenAI-compatible clients treat as null/no answer).
+  assert.equal(message.content, "The user wants 79874+93658. That equals 173532.");
+  // reasoning_content must STAY populated for reasoning-aware clients.
+  assert.equal(message.reasoning_content, "The user wants 79874+93658. That equals 173532.");
+});
+
+test("Command Code reasoning-only output emits a content delta chunk when streaming", async () => {
+  globalThis.fetch = async () =>
+    commandCodeStream(
+      [
+        { type: "reasoning-delta", text: "The result is 173532." },
+        { type: "finish", finishReason: "stop" },
+      ],
+      { sse: true }
+    );
+
+  const { response } = await getExecutor("command-code").execute({
+    model: "meta/muse-spark-1.2-contributor",
+    stream: true,
+    credentials: { apiKey: "cc_test_key" },
+    body: { messages: [{ role: "user", content: "Calcular 79874+93658" }] },
+  });
+
+  const sse = await response.text();
+  assert.match(sse, /data: \[DONE\]/);
+  const chunks = parseSsePayloads(sse);
+  assert.equal(chunks[0].choices[0].delta.role, "assistant");
+  // Regression #10986: the reasoning-only stream must emit a content delta when it
+  // otherwise ends with no content. reasoning_content stays present too.
+  const contentDelta = chunks.find((c) => c.choices[0].delta.content !== undefined);
+  assert.equal(contentDelta.choices[0].delta.content, "The result is 173532.");
+  const reasoningDelta = chunks.find((c) => c.choices[0].delta.reasoning_content !== undefined);
+  assert.equal(reasoningDelta.choices[0].delta.reasoning_content, "The result is 173532.");
+  assert.equal(chunks.at(-1).choices[0].finish_reason, "stop");
+});
+
 test("Command Code executor surfaces upstream and streamed errors", async () => {
   globalThis.fetch = async () =>
     new Response("bad key", { status: 401, statusText: "Unauthorized" });

tests/unit/command-code-registry-vision.test.ts

@@ -74,3 +74,53 @@ test("MiniMax M3 via command-code keeps existing vision capability (no regressio
   const caps = getResolvedModelCapabilities("command-code/MiniMaxAI/MiniMax-M3");
   assert.equal(caps.supportsVision, true);
 });
+
+test("command-code effort-suffixed variants resolve capabilities from their base model", async () => {
+  // Effort variants (e.g. `-max`, `-xhigh`) are synthesized from the base's
+  // supportedThinkingEfforts and have no registry/synced row of their own.
+  // Without the base-model fallback they resolve NULL tool/vision/context,
+  // which makes a tool-bearing combo request drop them behind confirmed
+  // targets (observed: orchestrator tried opencode-go/mimo-v2.5-max at
+  // position 2 while command-code deepseek sat unused at its declared
+  // priority position 2).
+  //
+  // Seed the models.dev capability store (the source getResolvedModelCapabilities
+  // reads for tool/vision/context) with the base models, then verify the
+  // effort-suffixed variants inherit those capabilities via the base fallback.
+  const modelsDevSync = await import("../../src/lib/modelsDevSync.ts");
+  modelsDevSync.saveModelsDevCapabilities({
+    "command-code": {
+      "deepseek/deepseek-v4-flash": {
+        tool_call: true,
+        reasoning: true,
+        attachment: false,
+        limit_context: 1000000,
+        limit_input: 1000000,
+        limit_output: 131072,
+      },
+      "meta/muse-spark-1.2-contributor": {
+        tool_call: true,
+        reasoning: true,
+        attachment: true,
+        limit_context: 1048576,
+        limit_input: 1048576,
+        limit_output: 1048576,
+      },
+    },
+  });
+
+  const cases: Array<[string, boolean]> = [
+    ["command-code/deepseek/deepseek-v4-flash-max", false], // text-only base
+    ["command-code/deepseek/deepseek-v4-flash-high", false],
+    ["command-code/meta/muse-spark-1.2-contributor-xhigh", true], // vision base
+    ["command-code/meta/muse-spark-1.2-contributor-high", true],
+  ];
+  for (const [modelId, vision] of cases) {
+    const caps = getResolvedModelCapabilities(modelId);
+    assert.equal(caps.provider, "command-code", `${modelId} provider`);
+    assert.equal(caps.supportsTools, true, `${modelId} must inherit tool support from base`);
+    assert.equal(caps.supportsVision, vision, `${modelId} must inherit vision from base`);
+    assert.equal(typeof caps.contextWindow, "number", `${modelId} must inherit a context window`);
+    assert.equal(caps.supportsThinking, true, `${modelId} must inherit reasoning from base`);
+  }
+});

tests/unit/opencode-merge-provider-guard.test.ts

@@ -0,0 +1,68 @@
+// `mergeOpenCodeConfig` guarded the root against a non-object but not the
+// `provider` branch it spreads one level down. Spreading a non-object does not
+// throw, it splays the value into index keys, so an existing config with a
+// malformed `provider` was rewritten into nonsense instead of being rejected:
+//
+//   provider: ["a", "b"]  ->  { "0": "a", "1": "b", omniroute: {...} }
+//   provider: "oops"      ->  { "0": "o", "1": "o", "2": "p", "3": "s", ... }
+//
+// Its sibling `mergeOpenCodeConfigText` throws on the same input
+// ("Can not add index to parent of type array"), so the two disagreed on what a
+// malformed config means.
+import test from "node:test";
+import assert from "node:assert/strict";
+
+const { mergeOpenCodeConfig, mergeOpenCodeConfigText } =
+  await import("../../src/shared/services/opencodeConfig.ts");
+
+const INPUT = {
+  baseUrl: "http://localhost:20128/v1",
+  apiKey: "sk_test_opencode",
+  model: "claude-sonnet-4-5-thinking",
+};
+
+for (const [label, provider] of [
+  ["an array", ["a", "b"]],
+  ["a string", "oops"],
+  ["a number", 7],
+  ["null", null],
+] as const) {
+  test(`mergeOpenCodeConfig drops a provider block that is ${label}`, () => {
+    const merged = mergeOpenCodeConfig({ provider } as never, INPUT);
+
+    assert.deepEqual(
+      Object.keys(merged.provider),
+      ["omniroute"],
+      `a provider block that is ${label} must not be splayed into index keys`
+    );
+    assert.equal(merged.provider.omniroute.options.baseURL, "http://localhost:20128/v1");
+  });
+}
+
+test("mergeOpenCodeConfig still preserves sibling providers", () => {
+  // The guard must only reject non-objects, never a legitimate provider map.
+  const merged = mergeOpenCodeConfig(
+    { provider: { custom: { name: "Custom Provider" }, other: { name: "Other" } } } as never,
+    INPUT
+  );
+
+  assert.deepEqual(Object.keys(merged.provider).sort(), ["custom", "omniroute", "other"]);
+  assert.equal(merged.provider.custom.name, "Custom Provider");
+});
+
+test("mergeOpenCodeConfig still guards the root itself", () => {
+  for (const existing of [["a"], "oops", 7, null, undefined]) {
+    const merged = mergeOpenCodeConfig(existing as never, INPUT);
+    assert.deepEqual(Object.keys(merged.provider), ["omniroute"]);
+    assert.equal(merged.$schema, "https://opencode.ai/config.json");
+  }
+});
+
+test("mergeOpenCodeConfigText keeps refusing the same malformed input", () => {
+  // Pinning the sibling's behaviour: the object variant now drops the bad
+  // block, the text variant still refuses to touch the file. Both are safe;
+  // neither corrupts.
+  for (const text of ['{"provider": ["a","b"]}', '{"provider": "oops"}']) {
+    assert.throws(() => mergeOpenCodeConfigText(text, INPUT));
+  }
+});

tests/unit/provider-field-strips.test.ts

@@ -53,3 +53,27 @@ test("stripGroqUnsupportedFields is immutable (does not mutate input)", () => {
   assert.equal(input.messages[0].name, "bob");
   assert.equal(input.logprobs, true);
 });
+
+test("stripGroqUnsupportedFields drops unsupported messages[].model and other metadata while keeping role and content", () => {
+  const out = stripGroqUnsupportedFields({
+    messages: [
+      { role: "user", content: "hello" },
+      {
+        role: "assistant",
+        content: "hello!",
+        model: "groq/openai/gpt-oss-20b",
+        messageId: "msg_123",
+        sender: "assistant",
+      },
+    ],
+  });
+  assert.equal(out.messages.length, 2);
+  assert.equal(out.messages[0].role, "user");
+  assert.equal(out.messages[0].content, "hello");
+  assert.equal(out.messages[1].role, "assistant");
+  assert.equal(out.messages[1].content, "hello!");
+  assert.equal("model" in out.messages[1], false);
+  assert.equal("messageId" in out.messages[1], false);
+  assert.equal("sender" in out.messages[1], false);
+});
+

tests/unit/silent-sse-close-responses-no-terminal.test.ts

@@ -0,0 +1,116 @@
+/**
+ * Regression test — silent SSE close on OpenAI Responses-format clients.
+ *
+ * Companion to #10443 (OpenAI chat completions) and #7699 (Claude). A healthy
+ * OpenAI Responses stream ALWAYS terminates with an explicit
+ * `response.completed` event; it is the format's only terminal marker and
+ * carries the final status/usage. When the upstream forwards content deltas
+ * and then closes without that event, Responses-format clients (Codex CLI and
+ * other /v1/responses consumers) previously received a silent mid-stream
+ * close — indistinguishable from a healthy end, so the client waits on a
+ * completion event that never arrives. The close must surface a synthetic
+ * `response.failed` instead, keeping everything already forwarded.
+ */
+import test from "node:test";
+import assert from "node:assert/strict";
+
+const { createDisconnectAwareStream, createStreamController } =
+  await import("../../open-sse/utils/streamHandler.ts");
+const { FORMATS } = await import("../../open-sse/translator/formats.ts");
+
+function createNoopAbortWritableStream(): { getWriter: () => { abort: () => Promise<void> } } {
+  return { getWriter: () => ({ abort: () => Promise.resolve() }) };
+}
+
+async function drainStream(stream: ReadableStream<Uint8Array>): Promise<string> {
+  const reader = stream.getReader();
+  const parts: Uint8Array[] = [];
+  for (;;) {
+    const { done, value } = await reader.read();
+    if (done) break;
+    parts.push(value);
+  }
+  return new TextDecoder().decode(
+    parts.reduce((acc, c) => {
+      const merged = new Uint8Array(acc.length + c.length);
+      merged.set(acc, 0);
+      merged.set(c, acc.length);
+      return merged;
+    }, new Uint8Array(0))
+  );
+}
+
+async function runClientStream(
+  upstreamChunks: string[],
+  clientResponseFormat: string | null
+): Promise<string> {
+  const upstream = new ReadableStream<Uint8Array>({
+    start(controller) {
+      const encoder = new TextEncoder();
+      for (const chunk of upstreamChunks) controller.enqueue(encoder.encode(chunk));
+      controller.close();
+    },
+  });
+
+  const transform = new TransformStream<Uint8Array, Uint8Array>({
+    transform(chunk, controller) {
+      controller.enqueue(chunk);
+    },
+  });
+  const transformedBody = upstream.pipeThrough(transform);
+
+  const sc = createStreamController({
+    provider: "opencode-go",
+    model: "muse-spark-1.2-contributor",
+    clientResponseFormat,
+  });
+
+  const wrapped = createDisconnectAwareStream(
+    { readable: transformedBody, writable: createNoopAbortWritableStream() },
+    sc
+  );
+
+  return drainStream(wrapped);
+}
+
+test("Responses format: content then bare close emits response.failed, not a silent close", async () => {
+  const text = await runClientStream(
+    [
+      'event: response.output_text.delta\ndata: {"type":"response.output_text.delta","delta":"partial"}\n\n',
+    ],
+    FORMATS.OPENAI_RESPONSES
+  );
+
+  // The forwarded content survives...
+  assert.match(text, /partial/);
+  // ...and the close must be flagged with the format's failure terminal.
+  assert.match(text, /response\.failed/);
+  assert.match(text, /Upstream stream ended without a terminal marker/);
+});
+
+test("Responses format: response.completed counts as terminal, no synthetic failure", async () => {
+  const text = await runClientStream(
+    [
+      'event: response.output_text.delta\ndata: {"type":"response.output_text.delta","delta":"done"}\n\n',
+      'event: response.completed\ndata: {"type":"response.completed","response":{"status":"completed"}}\n\n',
+    ],
+    FORMATS.OPENAI_RESPONSES
+  );
+
+  assert.match(text, /done/);
+  assert.match(text, /response\.completed/);
+  assert.doesNotMatch(text, /response\.failed/);
+  assert.doesNotMatch(text, /Upstream stream ended without a terminal marker/);
+});
+
+test("Responses format: OPENAI_RESPONSE alias gets the same bare-close verdict", async () => {
+  const text = await runClientStream(
+    [
+      'event: response.output_text.delta\ndata: {"type":"response.output_text.delta","delta":"partial"}\n\n',
+    ],
+    FORMATS.OPENAI_RESPONSE
+  );
+
+  assert.match(text, /partial/);
+  assert.match(text, /response\.failed/);
+});