pi-freeflow 1.34.0 → 1.36.0

package/src/models.ts

@@ -1,10 +1,11 @@
 /**
  * Static model definitions and upstream routing catalogs for pi-freeflow
  *
- * Defines the 30 verified free models (live-verified 2026-10-09):
- * - 10 OpenCode Zen models (2 Responses API + 8 Chat Completions)
+ * Defines the 37 verified free models (live-verified 2026-10-13):
+ * - 11 OpenCode Zen models (2 Responses API + 9 Chat Completions)
  * - 15 KiloCode Keyless Gateway models (OpenRouter format)
- * - 5 Cline direct-only models (per-user pool)
+ * - 2 Cline direct-only models (per-user pool)
+ * - 9 OVHcloud AI Endpoints models (keyless, direct-only)
  */
 
 import type { ModelDef, ThinkingLevelMap, Upstream } from "./types.ts";
@@ -14,6 +15,27 @@ import type { ModelDef, ThinkingLevelMap, Upstream } from "./types.ts";
  * Endpoint: https://opencode.ai/zen/v1
  */
 export const OPENCODE_MODELS: ModelDef[] = [
+ {
+  id: "cat-jam-free",
+  name: "Cat Jam [OpenCode]",
+  reasoning: true,
+  // Context is 262_144, NOT the 1M an acceptance probe suggests: the gateway
+  // answers 200 to oversized prompts and silently truncates. A needle planted
+  // at the end of a 248,232-token prompt is recalled; at 250,272 it is not.
+  // Cross-checked against the published spec, which agrees.
+  contextWindow: 262_144,
+  maxTokens: 131_072,
+  input: ["text"],
+  thinkingLevelMap: {
+   off: null,
+   minimal: null,
+   low: "low",
+   medium: "medium",
+   high: "high",
+   xhigh: "xhigh",
+   max: null,
+  },
+ },
  {
   id: "muse-spark-1.2-contributor-free",
   name: "Muse Spark 1.2 [OpenCode]",
@@ -378,24 +400,10 @@ export const KILO_MODELS: ModelDef[] = [
 ];
 
 /**
- * Full-range effort map (Solar Pro 4 — reference lists none..max; off stays
- * null per repo convention). The proxy passes effort through verbatim and
- * upstream serves the levels it supports.
- */
-const CLINE_REASONING_MAP: ThinkingLevelMap = {
- off: null,
- minimal: "minimal",
- low: "low",
- medium: "medium",
- high: "high",
- xhigh: "xhigh",
- max: "max",
-};
-/**
- * Narrow effort map for DeepSeek V4 Flash + GLM Flash + Kimi K3 — reference
- * reasoningOptions list effort values [low, high, max] only (no toggle
- * on GLM); minimal/medium/xhigh are not offered so the host never sends
- * an unsupported level (the proxy forwards effort verbatim).
+ * Narrow effort map for the Cline flash models still in the catalog (GLM 5.3
+ * Flash) — the reference reasoningOptions list effort values [low, high, max]
+ * only (no toggle on GLM); minimal/medium/xhigh are not offered so the host
+ * never sends an unsupported level (the proxy forwards effort verbatim).
  */
 const CLINE_FLASH_REASONING_MAP: ThinkingLevelMap = {
  off: null,
@@ -415,15 +423,6 @@ const CLINE_FLASH_REASONING_MAP: ThinkingLevelMap = {
  * Endpoint: https://api.cline.bot/api/v1/chat/completions
  */
 export const CLINE_MODELS: ModelDef[] = [
- {
-  id: "cline-free/deepseek-v4.1-flash",
-  name: "DeepSeek V4.1 Flash [Cline]",
-  reasoning: true,
-  contextWindow: 1_000_000,
-  maxTokens: 384_000,
-  input: ["text", "image"],
-  thinkingLevelMap: CLINE_FLASH_REASONING_MAP,
- },
  {
   id: "cline-free/muse-spark-1.3-contributor",
   name: "Muse Spark 1.3 [Cline]",
@@ -451,30 +450,158 @@ export const CLINE_MODELS: ModelDef[] = [
  input: ["text", "image", "video"],
   thinkingLevelMap: CLINE_FLASH_REASONING_MAP,
  },
+];
+
+/**
+ * Narrow effort map for the OVH reasoning models — probed live 2026-10-13:
+ * `reasoning_effort: "low"` and `"high"` are accepted (HTTP 200 with a
+ * populated reasoning stream). minimal/medium/xhigh/max were not confirmed,
+ * so they map to null and the picker never offers them. `minimal` must be an
+ * explicit null, not an omitted key: omitting it leaves the level selectable
+ * in the host. Same shape as CLINE_FLASH_REASONING_MAP.
+ */
+const OVH_REASONING_MAP: ThinkingLevelMap = {
+ off: null,
+ minimal: null,
+ low: "low",
+ medium: null,
+ high: "high",
+ xhigh: null,
+ max: null,
+};
+
+/**
+ * Catalog-id namespace for the OVH lane. OVH model ids are case-sensitive on
+ * the wire and several carry capitals and dots, so the prefix namespaces the
+ * lane (and keeps the id recognized by isFreeCatalogId) while the suffix is
+ * sent upstream verbatim.
+ */
+export const OVH_MODEL_PREFIX = "ovh-free/";
+
+/**
+ * OVHcloud AI Endpoints free models — keyless, direct-only.
+ * Endpoint: https://oai.endpoints.kepler.ai.cloud.ovh.net/v1
+ *
+ * Every id is OVH-native with an `ovh-free/` prefix; upstream only knows the
+ * suffix (a prefixed id answers 404 `model_not_found`, verified 2026-10-13),
+ * so the proxy strips the prefix on the wire. `api` stays unset, i.e.
+ * openai-completions: chat is the native shape and responses is forwarded
+ * natively too (with store:false injected). Anthropic messages is not served
+ * (`/v1/messages` answers 404 unknown_url upstream).
+ *
+ * Specs (measured live 2026-10-13 and cross-checked against every twin
+ * provider carrying a spec): contextWindow is OVH's advertised value;
+ * maxTokens is deliberately the conservative FLOOR across twin providers —
+ * per-provider output caps for the same model range from 4K to 256K, and a
+ * cap that is too low never breaks a request while one that is too high lets
+ * a host ask for an output the provider refuses. OVH itself enforces only
+ * `context - prompt` (verified by binary search), so these are safe request
+ * budgets rather than a refusal boundary.
+ *
+ * Vision was probed with a generated solid-colour PNG sent as a data URL,
+ * twice with two different colours to defeat a memorised answer: the four
+ * Qwens and Mistral-Small-3.2 named the colour; Mistral-7B and Llama-3.3-70B
+ * answer HTTP 400 to an image part; gpt-oss-120b returns 200 while its own
+ * reasoning says it cannot see the image, so it is text-only. Reasoning is
+ * the presence of a populated `reasoning`/`reasoning_content` field on a
+ * non-tool answer.
+ */
+export const OVH_MODELS: ModelDef[] = [
  {
-  // Cline advertises this sixth free model only to its desktop client
-  // identity (see CLINE_CLIENT_HEADERS). Specs follow the vendored Cline
-  // catalog; maxTokens keeps the tighter chat value so a request can never
-  // overrun what the chat endpoint accepts.
-  id: "cline-free/kimi-k3",
-  name: "Kimi K3 [Cline]",
+  id: "ovh-free/gpt-oss-120b",
+  name: "GPT-OSS 120B [OVH]",
   reasoning: true,
-  contextWindow: 1_048_576,
-  maxTokens: 131_072,
- input: ["text", "image", "video"],
-  thinkingLevelMap: CLINE_FLASH_REASONING_MAP,
+  contextWindow: 131_072,
+  maxTokens: 32_768,
+  input: ["text"],
+  thinkingLevelMap: OVH_REASONING_MAP,
  },
  {
-  id: "cline-free/solar-pro4",
-  name: "Solar Pro4 [Cline]",
+  id: "ovh-free/gpt-oss-20b",
+  name: "GPT-OSS 20B [OVH]",
   reasoning: true,
-  contextWindow: 524_288,
-  maxTokens: 131_072,
+  contextWindow: 131_072,
+  maxTokens: 32_768,
+  input: ["text"],
+  thinkingLevelMap: OVH_REASONING_MAP,
+ },
+ {
+  id: "ovh-free/Qwen3.5-397B-A17B",
+  name: "Qwen3.5 397B A17B [OVH]",
+  reasoning: true,
+  contextWindow: 262_144,
+  maxTokens: 65_536,
+  input: ["text", "image"],
+  thinkingLevelMap: OVH_REASONING_MAP,
+ },
+ {
+  id: "ovh-free/Qwen3.5-9B",
+  name: "Qwen3.5 9B [OVH]",
+  reasoning: true,
+  contextWindow: 262_144,
+  maxTokens: 65_536,
+  input: ["text", "image"],
+  thinkingLevelMap: OVH_REASONING_MAP,
+ },
+ {
+  id: "ovh-free/Qwen3.6-27B",
+  name: "Qwen3.6 27B [OVH]",
+  reasoning: true,
+  contextWindow: 262_144,
+  maxTokens: 65_536,
+  input: ["text", "image"],
+  thinkingLevelMap: OVH_REASONING_MAP,
+ },
+ {
+  id: "ovh-free/Qwen3.8-27B",
+  name: "Qwen3.8 27B [OVH]",
+  reasoning: true,
+  contextWindow: 262_144,
+  maxTokens: 32_768,
+  input: ["text", "image"],
+  thinkingLevelMap: OVH_REASONING_MAP,
+ },
+ {
+  id: "ovh-free/Mistral-Small-3.2-24B-Instruct-2506",
+  name: "Mistral Small 3.2 24B [OVH]",
+  reasoning: false,
+  contextWindow: 131_072,
+  maxTokens: 16_384,
+  input: ["text", "image"],
+ },
+ {
+  id: "ovh-free/Mistral-7B-Instruct-v0.3",
+  name: "Mistral 7B Instruct v0.3 [OVH]",
+  reasoning: false,
+  contextWindow: 65_536,
+  maxTokens: 65_536,
+  input: ["text"],
+ },
+ {
+  id: "ovh-free/Meta-Llama-3_3-70B-Instruct",
+  name: "Meta Llama 3.3 70B Instruct [OVH]",
+  reasoning: false,
+  contextWindow: 131_072,
+  maxTokens: 16_384,
   input: ["text"],
-  thinkingLevelMap: CLINE_REASONING_MAP,
  },
 ];
 
+/**
+ * Resolve any OVH id to the bare id OVH knows on the wire. Canonicalises
+ * first, then strips: a short alias (`llama-3.3-70b`) must resolve to
+ * `ovh-free/Meta-Llama-3_3-70B-Instruct` before the prefix comes off, or it
+ * reaches OVH unrecognised and answers 404 `model_not_found` (live
+ * 2026-10-13, both for a prefixed id and for an alias sent verbatim). The
+ * remainder is case-sensitive, so it is returned unchanged.
+ */
+export function ovhUpstreamModelId(id: string): string {
+ const canonical = resolveCanonicalModelId(id);
+ return canonical.startsWith(OVH_MODEL_PREFIX)
+  ? canonical.slice(OVH_MODEL_PREFIX.length)
+  : canonical;
+}
+
 /**
  * Model ID Aliases — maps user-friendly / slash-free CLI IDs to canonical upstream model IDs.
  * Single clean alias per model (no :free duplicates). Wrong cross-lab aliases removed:
@@ -505,6 +632,16 @@ export const MODEL_ALIASES: Record<string, string> = {
  "laguna-xs-2.1:free": "poolside/laguna-xs-2.1:free",
  "kilo-auto": "kilo-auto/free",
  "openrouter": "openrouter/free",
+// OVHcloud AI Endpoints — slash-free spellings of the `ovh-free/` ids.
+"gpt-oss-120b": "ovh-free/gpt-oss-120b",
+"gpt-oss-20b": "ovh-free/gpt-oss-20b",
+"qwen3.5-397b": "ovh-free/Qwen3.5-397B-A17B",
+"qwen3.5-9b": "ovh-free/Qwen3.5-9B",
+"qwen3.6-27b": "ovh-free/Qwen3.6-27B",
+"qwen3.8-27b": "ovh-free/Qwen3.8-27B",
+"mistral-small-3.2": "ovh-free/Mistral-Small-3.2-24B-Instruct-2506",
+"mistral-7b-v0.3": "ovh-free/Mistral-7B-Instruct-v0.3",
+"llama-3.3-70b": "ovh-free/Meta-Llama-3_3-70B-Instruct",
 };
 
 /**
@@ -533,10 +670,20 @@ export const CLINE_MODEL_IDS = new Set<string>([
   .filter(([_, target]) => CLINE_MODELS.some((cm) => cm.id === target))
   .map(([alias]) => alias),
 ]);
+
 /**
- * Combined list of all 30 static free models (canonical)
+ * Set of all OVH model IDs (including aliases) for fast lookup
  */
-export const ALL_MODELS: ModelDef[] = [...OPENCODE_MODELS, ...KILO_MODELS, ...CLINE_MODELS];
+export const OVH_MODEL_IDS = new Set<string>([
+ ...OVH_MODELS.map((m) => m.id),
+ ...Object.entries(MODEL_ALIASES)
+  .filter(([_, target]) => OVH_MODELS.some((om) => om.id === target))
+  .map(([alias]) => alias),
+]);
+/**
+ * Combined list of all 36 static free models (canonical)
+ */
+export const ALL_MODELS: ModelDef[] = [...OPENCODE_MODELS, ...KILO_MODELS, ...CLINE_MODELS, ...OVH_MODELS];
 
 /**
  * Map of model ID -> ModelDef
@@ -576,10 +723,19 @@ export function isClineModel(id: string): boolean {
  return CLINE_MODEL_IDS.has(id) || CLINE_MODEL_IDS.has(canonical);
 }
 
+/**
+ * Check if a model ID belongs to OVHcloud AI Endpoints
+ */
+export function isOvhModel(id: string): boolean {
+ const canonical = resolveCanonicalModelId(id);
+ return OVH_MODEL_IDS.has(id) || OVH_MODEL_IDS.has(canonical);
+}
+
 /**
  * Determine the upstream provider for a given model ID
  */
 export function getModelUpstream(id: string): Upstream {
  if (isClineModel(id)) return "cline";
+ if (isOvhModel(id)) return "ovh";
  return isKiloModel(id) ? "kilo" : "opencode";
 }

package/src/opencode-fingerprint.ts

@@ -33,6 +33,7 @@ import {
  isOmpLikeCaller,
  restoreToolNameForCaller,
  retargetToolChoiceForUpstream,
+ sanitizeUpstreamToolName,
  translateToolsForPath,
  upstreamToolNameFor,
  type FindGlobRestore,
@@ -283,6 +284,133 @@ export function ensureMessagesFingerprintTools(body: Record<string, unknown>, om
  }
 }
 
+/**
+ * Drop `phase: "commentary"` from assistant message items in the replayed
+ * history.
+ *
+ * Upstream tags the lead-in text of a tool-calling turn `phase: "commentary"`
+ * and the closing answer `phase: "final_answer"`. When the model narrates a
+ * tool call and emits none (the muse-spark ATEM leak, upstream
+ * anomalyco/opencode#44659), that commentary item ends its own turn — and
+ * upstream then rejects every replay of the history with `400
+ * invalid_request_error: The request contains invalid parameters`.
+ *
+ * Measured on `muse-spark-1.3-contributor-free`, one field at a time: the
+ * phase alone decides it. A terminal commentary item 400s; the same item
+ * followed by a function call, or by a `final_answer` item, returns 200; the
+ * id prefix and the text are irrelevant. Only the phase field is removed, and
+ * `phase` is output metadata the model never reads as instruction, so dropping
+ * it cannot change the answer.
+ *
+ * Every assistant message is swept, not just the turn's last item: upstream
+ * validates the whole replayed history, and a host may seat `system`,
+ * `developer` or additional-tool items between a commentary item and the next
+ * user message, so a tail-only sweep would miss one. One sweep also covers a
+ * commentary item left behind by an earlier turn once later turns append to
+ * the same history.
+ */
+function stripCommentaryPhase(body: Record<string, unknown>): void {
+ const input = body.input;
+ if (!Array.isArray(input)) return;
+ for (const item of input) {
+  if (!item || typeof item !== "object" || Array.isArray(item)) continue;
+  const record = item as Record<string, unknown>;
+  if (record.role === "assistant" && record.phase === "commentary") delete record.phase;
+ }
+}
+
+/**
+ * Reduce `tool_choice` to what Zen accepts: `"auto"`, or absent.
+ *
+ * Upstream rejects every other value with `400 only "auto" is supported for
+ * tool_choice` — `"none"`, `"required"`, and named function choices alike
+ * (measured live 2026-10-12 on `muse-spark-1.3-contributor-free`, responses and
+ * chat). Absent means the same thing as `"auto"`, so an unsupported choice is
+ * dropped rather than translated. A choice is never imposed on a host that
+ * sent none.
+ *
+ * Dropping a choice can leave a host's native reasoning switched off: OMP
+ * forces its synthetic `think` tool with `reasoning: { effort: "none" }`
+ * (captured request 2026-10-12), so removing the forced call without
+ * restoring native effort would run the model with no reasoning at all.
+ */
+function enforceAutoToolChoice(body: Record<string, unknown>): void {
+ const choice = body.tool_choice;
+ if (choice === undefined) return;
+ if (typeof choice === "string" && choice.trim().toLowerCase() === "auto") return;
+ delete body.tool_choice;
+ restoreNativeEffortAfterChoiceDrop(body);
+}
+
+/**
+ * Put native reasoning back when the only thing enabling it was a dropped
+ * forced tool call.
+ *
+ * `"none"` is not a reasoning-off switch upstream, it is an unsupported value:
+ * `muse-spark-1.3-contributor-free` answers `400 reasoning_effort 'none' is
+ * not supported` (measured 2026-10-12). So the override is replaced with the
+ * model's own highest supported level rather than merely deleted, and a model
+ * whose declared map supports nothing but `off` gets the field dropped.
+ * Genuine reasoning-off (`"off"`) is the caller's real intent, never rewritten.
+ */
+function restoreNativeEffortAfterChoiceDrop(body: Record<string, unknown>): void {
+ const model = body.model;
+ if (typeof model !== "string") return;
+ const replacement = highestSupportedEffort(model);
+ if (replacement === null) return;
+ const nested = body.reasoning;
+ if (nested !== null && typeof nested === "object" && !Array.isArray(nested)) {
+  const rec = nested as Record<string, unknown>;
+  if (typeof rec.effort === "string" && rec.effort.trim().toLowerCase() === "none") {
+   rec.effort = replacement;
+   return;
+  }
+ }
+ for (const key of ["reasoning_effort", "reasoningEffort"] as const) {
+  const flat = body[key];
+  if (typeof flat === "string" && flat.trim().toLowerCase() === "none") {
+   body[key] = replacement;
+   return;
+  }
+ }
+}
+
+/**
+ * Highest non-null level a model's declared map supports, or null when the
+ * model declares no map or only supports `off` (nothing to fall back to).
+ */
+function highestSupportedEffort(model: string): string | null {
+ const map = getModelDef(model)?.thinkingLevelMap;
+ if (!map) return null;
+ for (const level of EFFORT_DESCENDING) {
+  if (map[level] !== null && map[level] !== undefined) return level;
+ }
+ return null;
+}
+
+/**
+ * Fold the name of a replayed tool-call item to a shape upstream accepts.
+ *
+ * A model can emit a malformed call and the host then replays it in `input[]`
+ * forever: upstream validates every replayed item, so one call named
+ * `id id=` (captured in a live session, emitted by `space-bunny-free`) 400s
+ * every subsequent turn of that conversation. Declarations were sanitized
+ * already; this covers the calls.
+ */
+function sanitizeReplayItemNames(body: Record<string, unknown>): void {
+ const input = body.input;
+ if (!Array.isArray(input)) return;
+ for (const item of input) {
+  if (!item || typeof item !== "object" || Array.isArray(item)) continue;
+  const record = item as Record<string, unknown>;
+  if (record.type !== "function_call" && record.type !== "custom_tool_call") continue;
+  const name = record.name;
+  if (typeof name !== "string") continue;
+  const folded = sanitizeUpstreamToolName(name, new Set());
+  if (folded !== name) record.name = folded;
+ }
+}
+
 /**
  * Normalize a Zen Responses body to what upstream accepts. OMP/Pi hosts
  * sometimes carry Chat Completions leftovers on the responses path and
@@ -294,9 +422,12 @@ export function ensureMessagesFingerprintTools(body: Record<string, unknown>, om
  * - `temperature` / `top_p` are dropped (Responses rejects them).
  * - `response_format` is dropped (Chat Completions field, Responses rejects it).
  * - `parallel_tool_calls: false` is dropped; absent-or-true rides verbatim.
- * - `tool_choice` is auto-or-absent: `"none"` (string or `{ type: "none" }`)
- *   is dropped, everything else rides (named choices still retarget find->glob).
+ * - `tool_choice` is auto-or-absent: `"auto"` rides verbatim and everything
+   else is dropped, because Zen supports no other value (see
+   enforceAutoToolChoice).
  * - `store` defaults to false (never persisted server-side).
+ * - Assistant message items lose `phase: "commentary"`; upstream 400s the whole
+   request otherwise (see stripCommentaryPhase).
  * - Reasoning effort is clamped per model: when the model's thinkingLevelMap
  *   declares the requested effort as null (unsupported), the effort is clamped
  *   to the highest supported level. Covers all OpenCode models generically
@@ -310,6 +441,10 @@ export function normalizeResponsesBody(body: Record<string, unknown>): void {
   body.input = body.messages;
  }
  delete body.messages;
+ stripCommentaryPhase(body);
+ // After input[] exists: a replayed call name is only folded once the array is
+ // the one upstream will read.
+ sanitizeReplayItemNames(body);
  if (body.max_output_tokens === undefined) {
   if (typeof body.max_tokens === "number") body.max_output_tokens = body.max_tokens;
   else if (typeof body.max_completion_tokens === "number") body.max_output_tokens = body.max_completion_tokens;
@@ -320,11 +455,6 @@ export function normalizeResponsesBody(body: Record<string, unknown>): void {
  delete body.top_p;
  delete body.response_format;
  if (body.parallel_tool_calls === false) delete body.parallel_tool_calls;
- const choice = body.tool_choice;
- if (typeof choice === "string" ? choice.toLowerCase() === "none" : (
-  choice !== null && typeof choice === "object" && !Array.isArray(choice)
-  && (choice as Record<string, unknown>).type === "none"
- )) delete body.tool_choice;
  if (typeof body.model === "string") {
   const reasoning = body.reasoning;
   if (reasoning !== null && typeof reasoning === "object" && !Array.isArray(reasoning)) {
@@ -342,6 +472,9 @@ export function normalizeResponsesBody(body: Record<string, unknown>): void {
    }
   }
  }
+ // After the effort clamp: dropping a forced tool call is what triggers the
+ // native-effort restore, so it runs once clamping is settled.
+ enforceAutoToolChoice(body);
  if (body.store === undefined) body.store = false;
 }
 
@@ -380,19 +513,13 @@ export function enforceOpencodeFingerprint(
 } {
  const clientRequestedStream = body.stream === true;
  const callerHadTools = Array.isArray(body.tools) && body.tools.length > 0;
- // Zen responses shape conformance first (Chat leftovers, tool_choice none,
- // spark effort clamp, store default): Kilo/Cline bodies never reach here.
+ // Zen responses shape conformance first (Chat leftovers, tool_choice, spark
+ // effort clamp, store default): Kilo/Cline bodies never reach here.
  if (pathname.endsWith("/responses")) normalizeResponsesBody(body);
  // Chat path mirrors the responses drop: Zen free tier accepts only auto, so a
  // caller "none" (string or { type: "none" }) would 400 upstream. Messages
  // path keeps its own choice semantics.
- if (!pathname.endsWith("/responses") && !pathname.endsWith("/messages")) {
-  const noneChoice = body.tool_choice;
-  if (typeof noneChoice === "string" ? noneChoice.toLowerCase() === "none" : (
-   noneChoice !== null && typeof noneChoice === "object" && !Array.isArray(noneChoice)
-   && (noneChoice as Record<string, unknown>).type === "none"
-  )) delete body.tool_choice;
- }
+ enforceAutoToolChoice(body);
  // Upstream Zen free tier mandates stream: true for all free requests
  body.stream = true;