Claude Code. Any Model. The most powerful AI coding agent now speaks every language.
refactor: retire "roster" for the thesaurus terms
Jack Rudenkoab1fc0953ed4
diff --git a/packages/cli/src/adapters/base-api-format.ts b/packages/cli/src/adapters/base-api-format.tsindex b858ca2b..37bed32f 100644--- a/packages/cli/src/adapters/base-api-format.ts+++ b/packages/cli/src/adapters/base-api-format.ts@@ -524,7 +524,7 @@ export abstract class BaseAPIFormat implements APIFormat, ModelDialect {* it, for every dialect.** The SHAPE of the knob is a PER-MODEL fact read from the slim catalog's- * `reasoning` record, never from a table here. Alibaba's Alibaba Token Plan roster is+ * `reasoning` record, never from a table here. The Alibaba Token Plan's membership is* why a fixed ladder is wrong: `qwen3.7-plus` is `control: "toggle"` (it* exposes no depth parameter at all, so a `budget_tokens` would be an* invented field), while `glm-5.2` and `deepseek-v4-pro` on the SAME endpoint@@ -1023,7 +1023,7 @@ export abstract class BaseAPIFormat implements APIFormat, ModelDialect {* `createAnthropicPassthroughStream` strips it and RE-INDEXES the remaining* content blocks to a contiguous 0,1,2… sequence.*- * Gating on the WIRE rather than a model roster is deliberate — a hardcoded+ * Gating on the WIRE rather than a model list is deliberate — a hardcoded* list would silently miss the next model added to a multi-vendor plan, which* is exactly how `[email protected]` and `qtoken@deepseek-v4-pro` kept leaking after* `[email protected]` was fixed. `wireFormat` is the composition hintdiff --git a/packages/cli/src/adapters/devin-api-format.ts b/packages/cli/src/adapters/devin-api-format.tsindex dc6c076d..05e3c984 100644--- a/packages/cli/src/adapters/devin-api-format.ts+++ b/packages/cli/src/adapters/devin-api-format.ts@@ -31,7 +31,7 @@* - **No effort field.** Devin encodes the reasoning tier in the model UID* suffix, not in a request field, so {@link applyNativeReasoning} only* RECORDS the level on the payload; the transport folds it into field 21 via- * `resolveDevinModelUid` against the LIVE roster.+ * `resolveDevinModelUid` against the dynamic models catalog.* - **No vision.** {@link supportsVision} returns false, which routes images* through ComposedHandler's existing vision-proxy path (described into text,* or stripped). That is a working degradation for free, and it is why Devin's@@ -143,7 +143,7 @@ export class DevinAPIFormat extends BaseAPIFormat {/*** Zero = "no opinion". The live per-uid window comes from the transport's* `getContextWindow()` (ComposedHandler step 5b, positive-only), which reads- * the backend's own `GetCliModelConfigs` roster — the only source that knows+ * the backend's own `GetCliModelConfigs` dynamic models catalog — the only source that knows* what THIS subscription serves.*/override getContextWindow(): number {@@ -161,7 +161,7 @@ export class DevinAPIFormat extends BaseAPIFormat {* Devin exposes no effort parameter — the tier lives in the uid suffix — so* there is nothing to set on the wire here. Writing the level onto the* payload is the whole job; `DevinProviderTransport.serializeBody` resolves- * it against the live roster.+ * it against the dynamic models catalog.*/protected override applyNativeReasoning(request: any, originalRequest: any): any {const effort = this.resolveEffortLevel(originalRequest);@@ -366,7 +366,7 @@ export class DevinAPIFormat extends BaseAPIFormat {const payload: DevinRequestPayload = {// The bare requested name. The transport resolves it to a SERVED uid- // (family + effort → tier) against the live roster before encoding.+ // (family + effort → tier) against the dynamic models catalog before encoding.modelUid: this.modelId,messages: devinMessages,};diff --git a/packages/cli/src/adapters/grok-effort-support.ts b/packages/cli/src/adapters/grok-effort-support.tsindex 7df42cb6..e7aaf5d3 100644--- a/packages/cli/src/adapters/grok-effort-support.ts+++ b/packages/cli/src/adapters/grok-effort-support.ts@@ -1,6 +1,6 @@/*** Which xAI models accept `reasoning_effort` — decided at RUNTIME, not from a- * pinned roster.+ * pinned model list.** ## Why this module exists*@@ -171,7 +171,7 @@ export function isReasoningEffortRejection(errorText: string): boolean {// "error":"This model does not support `reasoning_effort` value `none`."}//// That is a second axis, and enumerating it per model would recreate exactly the-// stale-roster problem this module exists to remove. So it is learned the same+// stale-model-list problem this module exists to remove. So it is learned the same// way: send the honest mapping, and on a value rejection remember the specific// (model, value) pair and fall back one rung.diff --git a/packages/cli/src/adapters/model-catalog.test.ts b/packages/cli/src/adapters/model-catalog.test.tsindex fc78a2a6..39e4088c 100644--- a/packages/cli/src/adapters/model-catalog.test.ts+++ b/packages/cli/src/adapters/model-catalog.test.ts@@ -128,11 +128,11 @@ describe("resolveSubscriptionRouting", () => {});});- test("preserves not-served when no Qwen plan includes the model and both rosters are catalog-authoritative", () => {+ test("preserves not-served when no Qwen plan includes the model and both memberships are catalog-authoritative", () => {writeCatalog([modelEntry("qwen3-coder-plus", [], "qwen-token-plan"),- modelEntry("qwen-roster-proof", [+ modelEntry("qwen-membership-proof", ["alibaba-token-plan-individual","alibaba-token-plan-team-edition",]),diff --git a/packages/cli/src/adapters/model-catalog.ts b/packages/cli/src/adapters/model-catalog.tsindex e3572b46..f23cbafd 100644--- a/packages/cli/src/adapters/model-catalog.ts+++ b/packages/cli/src/adapters/model-catalog.ts@@ -41,7 +41,7 @@ export interface ModelEntry {/*** Curated release date (ISO `YYYY-MM-DD`), when Firebase has one. The* authoritative freshness signal — pickers prefer it over any date a provider- * endpoint reports, which is a roster-added timestamp, not a release.+ * endpoint reports, which is a date-added timestamp, not a release.*/releaseDate?: string;}@@ -232,7 +232,7 @@ export function lookupFamilyDefaultVariant(* an error and must never block a request.** @param provider Only return variants recorded on this serving provider. A- * preset is an observation about ONE provider's roster, not a portable fact+ * preset is an observation about ONE provider's model list, not a portable fact* about the model — the same parameter may not exist on another host — so a* caller that cannot verify the parameter independently should pass the* provider it is actually routing to.@@ -347,13 +347,13 @@ export function resolveSubscriptionRouting(: { kind: "unknown" };}- // Client-selected and hybrid rosters are decided by the user's live session.+ // Client-selected and hybrid memberships are decided by the user's live session.if (plans.some((plan) => plan.modelDiscovery !== "catalog")) return { kind: "unknown" };const ids = new Set(plans.map((plan) => plan.id));- const hasRoster = cache.entries.some((candidate) =>+ const hasMembership = cache.entries.some((candidate) =>candidate.subscriptionPlanIds?.some((planId) => ids.has(planId)));- return hasRoster ? { kind: "not-served" } : { kind: "unknown" };+ return hasMembership ? { kind: "not-served" } : { kind: "unknown" };}/**@@ -422,7 +422,7 @@ function findCacheEntry(modelId: string, cachePath?: string): SlimModelEntry | u}/**- * The canonical catalog id a plan's exact roster id resolves to, or undefined.+ * The canonical catalog id an exact wire id in a plan's membership resolves to, or undefined.** Only a `described` resolution answers. `missing` and `ambiguous` are the* catalog stating that IT could not resolve the id, and inventing a lookup ondiff --git a/packages/cli/src/agent-availability.test.ts b/packages/cli/src/agent-availability.test.tsindex 6a7ce346..e226d432 100644--- a/packages/cli/src/agent-availability.test.ts+++ b/packages/cli/src/agent-availability.test.ts@@ -8,7 +8,7 @@ import {} from "./agent-availability.js";describe("parseAvailableAgents", () => {- test("extracts the roster from Claude Code's rejection message", () => {+ test("extracts the agent list from Claude Code's rejection message", () => {// Real message shape, measured 2026-08-22.const msg ="--agent 'zzz-bogus' not found. Available agents: claude, code-analysis:detective, " +@@ -23,22 +23,22 @@ describe("parseAvailableAgents", () => {]);});- test("returns null when the roster is absent", () => {+ test("returns null when the agent list is absent", () => {expect(parseAvailableAgents("some unrelated output")).toBeNull();expect(parseAvailableAgents("")).toBeNull();});- test("returns null rather than an empty roster", () => {+ test("returns null rather than an empty agent list", () => {// An empty list must not be cached as "no agents exist", which would reject// every name. Null means "could not determine" and fails open.expect(parseAvailableAgents("Available agents: ")).toBeNull();});});-describe("assertAgentAvailable — with a known roster", () => {+describe("assertAgentAvailable — with a known agent list", () => {const CWD = "/seeded/project";- test("accepts a name that is on the roster", async () => {+ test("accepts a name that is on the agent list", async () => {clearAgentCache();seedAgentCacheForTest(CWD, ["dev:reviewer", "dev:architect", "general-purpose"]);await assertAgentAvailable("dev:reviewer", CWD);@@ -66,7 +66,7 @@ describe("assertAgentAvailable — with a known roster", () => {await expect(assertAgentAvailable(" dev:reviewer ", CWD)).rejects.toThrow(/Unknown agent/);});- test("rosters are isolated per cwd", async () => {+ test("agent lists are isolated per cwd", async () => {// Measured: this repo resolves 24 agents, /tmp resolves 5. A global cache// would reject an agent that is valid in the directory the child runs in.clearAgentCache();@@ -78,9 +78,9 @@ describe("assertAgentAvailable — with a known roster", () => {);});- test("a seeded roster is served from cache, not re-probed", async () => {+ test("a seeded agent list is served from cache, not re-probed", async () => {// The seeded cwd does not exist, so a probe would fail open and accept- // anything. Rejection proves the cached roster was used.+ // anything. Rejection proves the cached agent list was used.clearAgentCache();seedAgentCacheForTest("/nonexistent-cached-dir", ["only-this-one"]);await expect(assertAgentAvailable("something-else", "/nonexistent-cached-dir")).rejects.toThrow(@@ -99,10 +99,10 @@ describe("assertAgentAvailable — degenerate inputs", () => {await assertAgentAvailable("", "/nonexistent-dir-for-test");});- test("fails OPEN when the roster cannot be determined", async () => {+ test("fails OPEN when the agent list cannot be determined", async () => {// Spawning in a directory that does not exist makes the probe error.// Blocking every session because a probe broke would be worse than the trap- // this guards, so an undeterminable roster must not throw.+ // this guards, so an undeterminable agent list must not throw.clearAgentCache();await assertAgentAvailable("whatever-agent", "/nonexistent-dir-for-test-xyz");});diff --git a/packages/cli/src/agent-availability.ts b/packages/cli/src/agent-availability.tsindex 2cca5422..34673334 100644--- a/packages/cli/src/agent-availability.ts+++ b/packages/cli/src/agent-availability.ts@@ -21,10 +21,10 @@ import { spawn } from "node:child_process";* both and keeps them from drifting.*/-/** A name no real agent can hold, used to make Claude Code print its roster. */+/** A name no real agent can hold, used to make Claude Code print its agent list. */const PROBE_AGENT = "__claudish_agent_probe__";-/** How long a discovered roster stays fresh. Plugins can be installed mid-session. */+/** How long a discovered agent list stays fresh. Plugins can be installed mid-session. */const TTL_MS = 5 * 60_000;interface Entry {@@ -33,14 +33,14 @@ interface Entry {}/**- * Keyed by cwd — NOT global. The roster is resolved relative to the working+ * Keyed by cwd — NOT global. The agent list is resolved relative to the working* directory: measured from this repo it is 24 names including every plugin agent,* and from /tmp it is 5. Caching globally would reject valid agents for one- * directory because another had a narrower roster.+ * directory because another had a narrower agent list.*/const cache = new Map<string, Entry>();-/** Extract the roster from Claude Code's rejection message. Null if absent. */+/** Extract the agent list from Claude Code's rejection message. Null if absent. */export function parseAvailableAgents(output: string): string[] | null {const m = /Available agents:\s*(.+)/.exec(output);if (!m) return null;@@ -51,7 +51,7 @@ export function parseAvailableAgents(output: string): string[] | null {return names.length > 0 ? names : null;}-/** Run the probe. Resolves to null when the roster cannot be determined. */+/** Run the probe. Resolves to null when the agent list cannot be determined. */async function probe(cwd: string): Promise<Set<string> | null> {return new Promise((resolveP) => {let out = "";@@ -97,7 +97,7 @@ async function probe(cwd: string): Promise<Set<string> | null> {});}-/** The roster for `cwd`, or null when it cannot be determined. Cached per cwd. */+/** The agent list for `cwd`, or null when it cannot be determined. Cached per cwd. */export async function discoverAvailableAgents(cwd: string): Promise<Set<string> | null> {const hit = cache.get(cwd);if (hit && Date.now() - hit.at < TTL_MS) return hit.agents;@@ -106,17 +106,17 @@ export async function discoverAvailableAgents(cwd: string): Promise<Set<string>return agents;}-/** Test seam — drop cached rosters. */+/** Test seam — drop cached agent lists. */export function clearAgentCache(): void {cache.clear();}/**- * Test seam — supply the roster for `cwd` instead of probing.+ * Test seam — supply the agent list for `cwd` instead of probing.** Exists because the accept/reject behaviour is the whole point of this module* and cannot otherwise be tested without spawning a real `claude`: a suite that- * only covers "no agent named" and "roster undeterminable" would stay green if+ * only covers "no agent named" and "agent list undeterminable" would stay green if* this file became a no-op. Seeding the cache is deliberately used rather than* mocking the module — a Bun `mock.module` on shared infrastructure bleeds into* sibling test files.@@ -128,7 +128,7 @@ export function seedAgentCacheForTest(cwd: string, agents: string[]): void {/*** Throw if `agent` is not a real agent in `cwd`.*- * FAILS OPEN when the roster cannot be determined (claude not on PATH, probe+ * FAILS OPEN when the agent list cannot be determined (claude not on PATH, probe* timed out, output shape changed). Blocking every session because a probe broke* would be a worse failure than the one this guards, and it mirrors the contract* `prehydrateCredentialsForSpawn` already documents: a resolution step is andiff --git a/packages/cli/src/auth/antigravity-oauth.ts b/packages/cli/src/auth/antigravity-oauth.tsindex 140a9ded..f6b2815f 100644--- a/packages/cli/src/auth/antigravity-oauth.ts+++ b/packages/cli/src/auth/antigravity-oauth.ts@@ -46,7 +46,7 @@ import { getServedAntigravityModels, setupAntigravityUser } from "./antigravity-/*** Resolve a concrete CURRENT served model id for the success message, LIVE from* the backend (its `defaultAgentModelId`, e.g. gemini-3.6-flash-high) — no- * hardcoded roster. Best-effort: any failure (no token, offline) → `<model>`.+ * hardcoded model list. Best-effort: any failure (no token, offline) → `<model>`.*/async function defaultSuggestModel(): Promise<string> {try {diff --git a/packages/cli/src/auth/antigravity-token.ts b/packages/cli/src/auth/antigravity-token.tsindex 9ef6877c..76efd735 100644--- a/packages/cli/src/auth/antigravity-token.ts+++ b/packages/cli/src/auth/antigravity-token.ts@@ -182,7 +182,7 @@ function defaultReadStore(): string | null {// That line is not an error. It is "not signed in", which this function// reports as `null` by design. Printing it corrupts the config TUI, which// shares its terminal with the Claude Code TUI, and it repeats forever- // because READ_STORE_TTL_MS is 3s and the provider roster is rebuilt on+ // because READ_STORE_TTL_MS is 3s and the provider list is rebuilt on// every render. Same policy as providers/keychain.ts — see the spawn// comment there, and terminal-isolation.ts for the display corruption.//diff --git a/packages/cli/src/auth/antigravity-user.test.ts b/packages/cli/src/auth/antigravity-user.test.tsindex 59c9f9fe..6c46a868 100644--- a/packages/cli/src/auth/antigravity-user.test.ts+++ b/packages/cli/src/auth/antigravity-user.test.ts@@ -85,7 +85,7 @@ describe("getServedAntigravityModels", () => {expect("contextWindow" in result.meta["string-window"]).toBe(false);});- test("returns an indexable empty meta object for an empty roster", async () => {+ test("returns an indexable empty meta object for an empty dynamic models catalog", async () => {globalThis.fetch = mock(async () =>Response.json({ models: {}, defaultAgentModelId: "unused-default" })) as unknown as typeof fetch;diff --git a/packages/cli/src/auth/antigravity-user.ts b/packages/cli/src/auth/antigravity-user.tsindex 3e4d0430..70ca3850 100644--- a/packages/cli/src/auth/antigravity-user.ts+++ b/packages/cli/src/auth/antigravity-user.ts@@ -34,7 +34,7 @@ import { log } from "../logger.js";* daily-cloudcode-pa.googleapis.com generate gemini-3.6-flash-high -> 200** On `cloudcode-pa` that account reads `currentTier: free-tier`, is served a- * roster of `gemini-2.5-*` and editor-internal `chat_*`/`tab_*` ids, reports+ * dynamic models catalog of `gemini-2.5-*` and editor-internal `chat_*`/`tab_*` ids, reports* every quota bucket at 100% forever (nothing is ever consumed), and can* generate with EXACTLY the two `tab_*` completion models — free Code Assist's* remaining entitlement — while all 28 chat models return a contentlessdiff --git a/packages/cli/src/auth/quota-command.ts b/packages/cli/src/auth/quota-command.tsindex 0c5b8c53..b6ae35cf 100644--- a/packages/cli/src/auth/quota-command.ts+++ b/packages/cli/src/auth/quota-command.ts@@ -225,8 +225,8 @@ function renderPlan(adapter: QuotaAdapter, plan: PlanUsage): void {// `claude-opus-4…` / `3.6-flash-med…` — losing exactly the tier suffix that// distinguishes one window from the next.//- // Clamped at both ends: 14 keeps short-roster providers from collapsing into- // a ragged narrow column, and 28 keeps the row inside 80 columns+ // Clamped at both ends: 14 keeps providers with only short window names from+ // collapsing into a ragged narrow column, and 28 keeps the row inside 80 columns// (2 indent + 2 gutter + 28 name + 24 bar + 2 + 4 pct + 2 + reset).//// The +1 is the column SEPARATOR, and it is not cosmetic padding: the bar isdiff --git a/packages/cli/src/auth/quota/sources/codex.ts b/packages/cli/src/auth/quota/sources/codex.tsindex a74caedc..978b7c56 100644--- a/packages/cli/src/auth/quota/sources/codex.ts+++ b/packages/cli/src/auth/quota/sources/codex.ts@@ -147,12 +147,12 @@ export function scrapeCodexHeaders(headers: Headers): PlanUsage | undefined {/*** Pick a model that actually exists on this account, from the Codex CLI's own- * cache of the roster it was served.+ * cache of the dynamic models catalog it was served.** This matters more than it looks: a stale model id is rejected with a 400,* and a 400 carries NO `x-codex-*` headers at all — so a wrong id here is* indistinguishable from "this provider does not report quota". Discovering- * the id rather than pinning one keeps a roster change from silently+ * the id rather than pinning one keeps a change in the dynamic models catalog from silently* presenting as a missing feature.*/function resolveProbeModel(): string | undefined {diff --git a/packages/cli/src/auth/quota/sources/grok.ts b/packages/cli/src/auth/quota/sources/grok.tsindex 6d682784..7f887009 100644--- a/packages/cli/src/auth/quota/sources/grok.ts+++ b/packages/cli/src/auth/quota/sources/grok.ts@@ -16,7 +16,7 @@* `xai-grok-shell/src/extensions/billing.rs` module whose error prose names the* path), then confirmed live. It is genuinely authenticated: a deliberately* bogus token answers 401, so a 200 here is evidence about the credential and- * not a public roster — the trap Alibaba's `coding-intl` list set.+ * not a public list — the trap Alibaba's `coding-intl` list set.** Full protocol write-up: `ai-docs/reports/grok-subscription/protocol-spec.md`.*/diff --git a/packages/cli/src/channel/e2e-channel.test.ts b/packages/cli/src/channel/e2e-channel.test.tsindex 101dc845..b03aa15e 100644--- a/packages/cli/src/channel/e2e-channel.test.ts+++ b/packages/cli/src/channel/e2e-channel.test.ts@@ -81,7 +81,7 @@ describe("Group 1: MCP Protocol — channel capability", () => {expect(instructions).toContain("completed");});- test("lists the exact public tool roster", async () => {+ test("lists the exact public tool list", async () => {const result = await client.listTools();const names = result.tools.map((t) => t.name).sort();expect(names).toEqual([diff --git a/packages/cli/src/handlers/composed-handler.ts b/packages/cli/src/handlers/composed-handler.tsindex 471e9527..6211eb14 100644--- a/packages/cli/src/handlers/composed-handler.ts+++ b/packages/cli/src/handlers/composed-handler.ts@@ -475,7 +475,7 @@ export class ComposedHandler implements ModelHandler {const behaviorSession = this.behaviorEngine.startSession({modelId: this.bareModelName,providerName: this.provider.name,- // A naming rule, not a pinned roster: Claude models are `claude-*`, and+ // A naming rule, not a pinned model list: Claude models are `claude-*`, and// they already follow these conventions (measured 87/87 on plan mode).//// `forceForeignModel` is the one escape hatch, and it exists because the@@ -1575,7 +1575,7 @@ export class ComposedHandler implements ModelHandler {// default `false` — which silently made the one dialect override of// it (MiniMaxModelDialect's `true`) unreachable dead code, and would// do the same to the base's wire-keyed `true` for every other model- // on this wire (qtoken@'s qwen / glm / deepseek rosters).+ // on this wire (qtoken@'s qwen / glm / deepseek models).adapter: (this.modelAdapter ?? adapter) as BaseAPIFormat,shouldBufferTool: (name) => behaviorSession?.interceptsTool(name) ?? false,repairToolArgs: (name, argsJson) =>diff --git a/packages/cli/src/handlers/native-handler-advisor.ts b/packages/cli/src/handlers/native-handler-advisor.tsindex 19db9ea4..9249a233 100644--- a/packages/cli/src/handlers/native-handler-advisor.ts+++ b/packages/cli/src/handlers/native-handler-advisor.ts@@ -1660,7 +1660,7 @@ function lookupOpenRouterId(name: string): OpenRouterIdLookup {/*** Is `vendor` a namespace OpenRouter actually publishes — the `x-ai` of* `x-ai/grok-4.6`? Answered from the LIVE catalog's own OpenRouter external- * ids, never from a provider roster in this repo (a pinned roster is exactly+ * ids, never from a provider's model list in this repo (a pinned model list is exactly* what goes stale, and CLAUDE.md forbids one).** A namespace match rather than an exact-id match on purpose: a model released@@ -1883,8 +1883,8 @@ const ADVISOR_MAX_OUTPUT_TOKENS = 2048;* is the existing per-format rule; its catalog lookup is the part that is* right, and it is reused here rather than restated. Its private name-based* fallback (`gpt-5`/`o1`/`o3`/`o4` → `max_completion_tokens`) is NOT copied —- * a second copy of a guess is exactly what CLAUDE.md's "no hardcoded roster"- * rule is about, and this endpoint has a better default (below).+ * a second copy of a guess is exactly what CLAUDE.md's "never hardcode a+ * provider's models" rule is about, and this endpoint has a better default (below).** ONE endpoint-level correction on the catalog's answer: `max_output_tokens` is* the RESPONSES-API spelling. The catalog says exactly that for gpt-5.6-soldiff --git a/packages/cli/src/handlers/shared/format/openai-messages.ts b/packages/cli/src/handlers/shared/format/openai-messages.tsindex 43296a86..1f410367 100644--- a/packages/cli/src/handlers/shared/format/openai-messages.ts+++ b/packages/cli/src/handlers/shared/format/openai-messages.ts@@ -283,7 +283,7 @@ function alignToolRound(assistant: any, collected: any[]): any[] {* whether to emit its "[image returned…]" marker.** The media type is validated for SHAPE (a non-empty string), not against a list- * of accepted image types. An allowlist here would be a second roster to keep+ * of accepted image types. An allowlist here would be a second list to keep* current, and a media type this converter has not heard of is the upstream* provider's judgement to make, not ours.*/diff --git a/packages/cli/src/handlers/shared/model-unsupported.ts b/packages/cli/src/handlers/shared/model-unsupported.tsindex 8464060d..e16126e5 100644--- a/packages/cli/src/handlers/shared/model-unsupported.ts+++ b/packages/cli/src/handlers/shared/model-unsupported.ts@@ -80,7 +80,7 @@ export function hasModelUnsupportedWording(errorBody: string): boolean {* https://opencode.ai/workspace/<id>/go"}}** The route is fine, the credential is fine, the model IS carried (it is in the- * live roster, so the availability filter correctly kept the provider) — the+ * dynamic models catalog, so the availability filter correctly kept the provider) — the* account simply has not opted into that region. claudish rendered it as* "Check API key / OAuth credentials.", sending the user to audit a working key* while the fix — a link — sat in the same sentence.diff --git a/packages/cli/src/handlers/shared/schema-validate.ts b/packages/cli/src/handlers/shared/schema-validate.tsindex 3e1167d7..13b91e92 100644--- a/packages/cli/src/handlers/shared/schema-validate.ts+++ b/packages/cli/src/handlers/shared/schema-validate.ts@@ -257,7 +257,7 @@ export function applySchemaDefaults(/*** Names a model reaches for when it does not use the one the schema declares.*- * This is a synonym vocabulary, not a roster in CLAUDE.md's sense: it is not+ * This is a synonym vocabulary, not a hardcoded list in CLAUDE.md's sense: it is not* per-user, not per-account, not time-varying, and it names no tool and no* agent. Every entry RENAMES A VALUE THE MODEL SUPPLIED; none invents one. The* gate in {@link renameToDeclaredKeys} is what makes that safe.diff --git a/packages/cli/src/handlers/shared/thinking-signature.ts b/packages/cli/src/handlers/shared/thinking-signature.tsindex 381220c7..4c224bd5 100644--- a/packages/cli/src/handlers/shared/thinking-signature.ts+++ b/packages/cli/src/handlers/shared/thinking-signature.ts@@ -33,7 +33,7 @@* It never reaches the client. On the Anthropic wire, thinking is stripped at* EMIT: `shouldFilterThinking()` is true for every dialect there* (base-api-format.ts — it keys on the wire, deliberately, not on a model- * roster), and `createAnthropicPassthroughStream` drops the blocks and+ * list), and `createAnthropicPassthroughStream` drops the blocks and* re-indexes what remains. Replaying that MiniMax fixture through the real* composition yields 0 thinking blocks and 0 signature frames. Same answer for* kimi-k2.5, glm-5.2 and deepseek-v4 — all `filterThinking: true`.diff --git a/packages/cli/src/handlers/shared/tool-call-recovery.ts b/packages/cli/src/handlers/shared/tool-call-recovery.tsindex f127cb60..07145a0f 100644--- a/packages/cli/src/handlers/shared/tool-call-recovery.ts+++ b/packages/cli/src/handlers/shared/tool-call-recovery.ts@@ -8,7 +8,7 @@* It does NOT invent arguments. It used to: `inferMissingParameters` filled a* missing `Glob.pattern` with `**\/*` (walks the whole tree), took `Task.prompt`* from "the last meaningful sentence" of the model's prose, and rewrote- * `subagent_type` against a hardcoded roster of eight agent names that claudish+ * `subagent_type` against a hardcoded list of eight agent names that claudish* cannot discover and that was already stale. A MISSING argument surfaces as a* visible harness error; a FABRICATED one is a confident wrong action with no* error anywhere. The codebase already prices that asymmetry the same way for@@ -88,7 +88,7 @@ export function hasExtractableFunctionTag(text: string): boolean {* a shape rule that was never about it.* 3. **Allowlist.** Reject a well-shaped name the client never advertised —* this is how a hallucinated tool is caught. The allowlist is the request's- * own tool list, so it is exact rather than a hardcoded roster.+ * own tool list, so it is exact rather than a hardcoded tool list.*/function keepOnlyRealTools(extracted: ExtractedToolCall[],@@ -301,7 +301,7 @@ function normalizeAgainstSchema(* carries no types, so every value arrives as a string.** `decodeToolName`, when supplied, turns the name the MODEL wrote back into the- * client's own. The model was given the encoded roster, so on a renaming wire+ * client's own. The model was given the encoded tool list, so on a renaming wire* every recovered call carries an encoded name that `knownToolNames` — built* from the client's originals — would otherwise reject in silence.*/@@ -468,7 +468,7 @@ export function extractToolCallsFromText(// Matches: "I'll use the Task tool with subagent_type=Explore"// Matches: "I will use the Read tool to read /path/to/file"// Matches: "Let me use the Bash tool to run ls -la"- // The caller's own tool list is exact; this roster is the fallback for callers+ // The caller's own tool list is exact; this fixed list is the fallback for callers// that have none in hand. Natural-language extraction is the loosest pattern// here, so it keeps its own guard even though `keepOnlyRealTools` runs after.const knownTools =diff --git a/packages/cli/src/index.ts b/packages/cli/src/index.tsindex d0bc8eff..c654eaf3 100644--- a/packages/cli/src/index.ts+++ b/packages/cli/src/index.ts@@ -575,12 +575,12 @@ async function runCli() {// Register the bundled endpoint catalog before ANYTHING enumerates or// validates providers. Two consumers below need it and both run long before// the proxy (which has always registered endpoints for the request path):- // the interactive picker enumerates the roster, and `validateApiKeysForModels`+ // the interactive picker enumerates the provider list, and `validateApiKeysForModels`// decides whether an explicit `vendor@model` has a credential — a provider// that is not registered yet reads as an unknown one.//// Placed AFTER parseArgs on purpose: terminal flags (`--version`, `--models`,- // `--probe`) exit inside it, so those paths pay nothing for a roster they+ // `--probe`) exit inside it, so those paths pay nothing for a provider list they// never show. Sync, config-only, idempotent; dynamically imported like every// other heavy module in runCli.await traceSpan("startup:endpoint-registration", async () => {diff --git a/packages/cli/src/keychain-command.ts b/packages/cli/src/keychain-command.tsindex 057722eb..567fd5a5 100644--- a/packages/cli/src/keychain-command.ts+++ b/packages/cli/src/keychain-command.ts@@ -94,7 +94,7 @@ function requireKeychain(): boolean {* Every API-key env var claudish knows about, primary names and aliases alike.** Built LIVE from the provider catalog rather than a hand-written list —- * a second roster would drift the moment a provider was added, and this repo+ * a second list would drift the moment a provider was added, and this repo* has already paid for that once (the config TUI's own provider array went* stale and made two working providers invisible).*/diff --git a/packages/cli/src/madbench/runner.ts b/packages/cli/src/madbench/runner.tsindex d0fc0bb1..8d722f5a 100644--- a/packages/cli/src/madbench/runner.ts+++ b/packages/cli/src/madbench/runner.ts@@ -74,7 +74,7 @@ function md5(p: string): string {/*** The task. Deliberately shaped so every assertion is model-agnostic:*- * - The model roster is NOT pinned. Claude Code is told to consult claudish's+ * - The model list is NOT pinned. Claude Code is told to consult claudish's* live catalog and pick the newest itself, so the benchmark keeps working* when vendors ship. What is checked is that it picked N distinct REAL ones.* - The question has exactly one correct answer, expressible as digits, sodiff --git a/packages/cli/src/madbench/types.ts b/packages/cli/src/madbench/types.tsindex c8df4416..3e57ad20 100644--- a/packages/cli/src/madbench/types.ts+++ b/packages/cli/src/madbench/types.ts@@ -10,7 +10,7 @@** The model names are deliberately NOT pinned. The prompt asks for "the latest* models" and Claude Code resolves that itself against claudish's live catalog.- * Pinning ids would make the test a snapshot of one afternoon's model roster —+ * Pinning ids would make the test a snapshot of one afternoon's model list —* it would go red every time a vendor ships, for reasons that have nothing to do* with claudish. What is asserted instead is the invariant: three DISTINCT* models, each one real (present in the catalog the session itself saw), eachdiff --git a/packages/cli/src/mcp-server.ts b/packages/cli/src/mcp-server.tsindex f62e727c..adb6b591 100644--- a/packages/cli/src/mcp-server.ts+++ b/packages/cli/src/mcp-server.ts@@ -1006,9 +1006,9 @@ function defineTools(tools.push({name: "preflight",description:- "DIAGNOSTIC. For a roster of models, report which provider would serve each, " ++ "DIAGNOSTIC. For a list of models, report which provider would serve each, " +"whether that hop is subscription or metered, and whether it is reachable right " +- "now. This is for a human investigating a roster. It is NOT a step before " ++ "now. This is for a human investigating a model list. It is NOT a step before " +"`team`, `create_session` or `run_prompt` — those resolve their own routing, and " +"a caller that hands them a bare model name never needs to know the route.",inputSchema: {@@ -1039,7 +1039,7 @@ function defineTools(group: "agentic",// N models each waiting on a live provider can sit well past the client's MCP// idle window, and this tool is specifically meant to be called with a big- // roster — exactly the shape that gets aborted without a keepalive.+ // model list — exactly the shape that gets aborted without a keepalive.heartbeat: true,handler: async (args, ctx) => {const models = Array.isArray(args.models) ? (args.models as string[]) : [];@@ -1053,7 +1053,7 @@ function defineTools(const doProbe = args.probe !== false;const timeoutMs = typeof args.timeout_ms === "number" ? args.timeout_ms : 20_000;// Start the proxy only if something will actually be probed. Native names- // never are (see the loop), so an all-native roster must neither wait on+ // never are (see the loop), so an all-native model list must neither wait on// proxy startup nor throw from it.const needsProxy = doProbe && models.some((m) => nativeRouteFor(m) === null);const proxy = needsProxy ? await getProxy() : null;@@ -1117,7 +1117,7 @@ function defineTools(if (plan.kind === "no-route") {// No credentialed provider can serve this at all. Reported as a FAILURE- // rather than omitted, because "this model is not in your roster" is the+ // rather than omitted, because "none of your credentials reaches this model" is the// single most useful thing to learn before the run rather than during it.failedModels.push(model);rows.push(`| \`${model}\` | — | — | ❌ no route | ${plan.reason} |`);@@ -1732,7 +1732,7 @@ function defineTools(// an option — it is process-global and races concurrent calls.const requestedModel = args.model as string;const workDir = args.work_dir as string | undefined;- // The roster is cwd-dependent, so validate against the directory this+ // The agent list is cwd-dependent, so validate against the directory this// session will actually run in, not the parent's.await assertAgentAvailable(args.agent as string | undefined, workDir ?? process.cwd());diff --git a/packages/cli/src/model-loader-v3.test.ts b/packages/cli/src/model-loader-v3.test.tsindex 2a4d3fa4..3b3e9cc4 100644--- a/packages/cli/src/model-loader-v3.test.ts+++ b/packages/cli/src/model-loader-v3.test.ts@@ -20,7 +20,7 @@ const envelope = (});describe("v3 provider-scoped pagination", () => {- test("returns the complete provider roster from one generation", async () => {+ test("returns the provider's complete model list from one generation", async () => {const requests: Array<{cursor: string | null;generation: string | null;diff --git a/packages/cli/src/model-selector.test.ts b/packages/cli/src/model-selector.test.tsindex 5711954e..9b24851e 100644--- a/packages/cli/src/model-selector.test.ts+++ b/packages/cli/src/model-selector.test.ts@@ -92,7 +92,7 @@ describe("warnDiscoveryFailure", () => {async () => new Response("not-json", { status: 200 })) as unknown as typeof fetch;break;- case "empty-roster":+ case "empty-models-catalog":globalThis.fetch = mock(async () => new Response(JSON.stringify({ data: [] }), { status: 200 })) as unknown as typeof fetch;@@ -125,10 +125,10 @@ describe("warnDiscoveryFailure", () => {}}- test("writes nothing when no failure is recorded or the roster is empty", async () => {+ test("writes nothing when no failure is recorded or the dynamic models catalog is empty", async () => {expect(captureWarning()).toEqual({ stderr: "", stdout: "", stderrCalls: 0, stdoutCalls: 0 });- await recordFailure("empty-roster");+ await recordFailure("empty-models-catalog");expect(captureWarning()).toEqual({ stderr: "", stdout: "", stderrCalls: 0, stdoutCalls: 0 });});@@ -204,9 +204,9 @@ describe("buildExplicitModelSpec", () => {});});-// ─── picker provider roster ──────────────────────────────────────────────────+// ─── picker provider list ────────────────────────────────────────────────────-describe("picker provider roster", () => {+describe("picker provider list", () => {const pickableBuiltins = BUILTIN_PROVIDERS.filter(isPickableProvider);const unpickableBuiltins = BUILTIN_PROVIDERS.filter((def) => !isPickableProvider(def));diff --git a/packages/cli/src/model-selector.ts b/packages/cli/src/model-selector.tsindex 38cc5e08..0527ba12 100644--- a/packages/cli/src/model-selector.ts+++ b/packages/cli/src/model-selector.ts@@ -36,10 +36,10 @@ import {discoverProviderModels,getDiscoveryFailure,rankDiscoveredModels,- toRosterEntry,+ toModelsCatalogEntry,} from "./providers/model-discovery.js";import { compareByReleaseDateDesc } from "./providers/model-ordering.js";-import { collapseRoster } from "./providers/model-resolvers/registry.js";+import { collapseModelsCatalog } from "./providers/model-resolvers/registry.js";import { type ModelOffer, offerIsLive } from "./providers/model-resolvers/types.js";import { PROVIDER_FILTER_ALIAS_EXTRA } from "./providers/picker-alias-extra.js";import {@@ -147,14 +147,14 @@ const LOCAL_OR_USER_DEPLOYED = new Set<string>(["litellm", "ollama", "lmstudio"]* RUNTIME-registered provider — a user's `customEndpoints` entry, a bundled* catalog row — belongs in the same class and is recognised by derivation* rather than by being listed, because the whole point of those is that- * claudish does not know their rosters.+ * claudish does not know their models.** That is not just "no data": for a name that happens to match a models-index* VENDOR slug, `modelsByVendor` answers with ids from the CREATOR namespace,* and the picker would then emit `vendor@<models-index id>` — an id the* vendor's own endpoint is not guaranteed to accept. Failing after the user* commits is worse than asking them to type a model name, and R7 forbids- * shipping a roster to check against.+ * shipping a model list to check against.*/export function isUserDeployedProvider(value: string): boolean {return LOCAL_OR_USER_DEPLOYED.has(value) || getRuntimeProviders().has(value);@@ -317,15 +317,16 @@ function resolveDiscoveredContextLength(m: DiscoveredModel): number {* Release date for a live-discovered model row, catalog first.** The Firebase slim catalog carries curated RELEASE dates, so it wins wherever- * it has an entry. The endpoint's date is a roster-added timestamp, which is+ * it has an entry. The endpoint's date is a date-added timestamp, which is* the only signal available for the models the slim catalog never listed —* `qwen3.8-max-preview` is exactly that, and without the fallback one of the* newest models on the plan would sort to the very bottom as undated.*/function resolveDiscoveredReleaseDate(m: DiscoveredModel): string | undefined {- // A roster of variants the catalog does not list gets NO catalog date: the- // catalog would be dating a different (older, base) model, and a partially- // dated roster sorts its undated — newest — half to the bottom.+ // A dynamic models catalog of variants the cloud models catalog does not list+ // gets NO catalog date: the catalog would be dating a different (older, base)+ // model, and a partially dated dynamic models catalog sorts its undated —+ // newest — half to the bottom.if (m.ignoreCatalogReleaseDate) return m.releaseDate;try {const catalogDate = lookupModel(m.id)?.releaseDate;@@ -530,7 +531,7 @@ function formatModelChoiceAsSpec(model: ModelInfo, spec: string, priceStr: strin* model vendors.** Derived, because the hand-written version was the same opt-in table the- * provider roster used to be: `devin` and `antigravity` were both missing from+ * provider list used to be: `devin` and `antigravity` were both missing from* it, so `@dv` silently matched nothing.** Rebuilt per call — see the note above on why it is not memoized.@@ -839,7 +840,7 @@ interface ProviderChoice {/*** Editorial overlay for the picker's provider rows — NOT a membership list.*- * The roster itself is DERIVED from `getAllProviders()` (see+ * The provider list itself is DERIVED from `getAllProviders()` (see* `buildProviderChoices`). This map only overrides copy where the picker's* wording beats the definition's `description`, which is written for the config* TUI's denser layout. A provider absent from this map still appears; it just@@ -932,7 +933,7 @@ const PICKER_ORDER = [* request time) and are both pickable. Everything else is offerable, and the* credential authority decides whether THIS user sees it.*- * A rule, not a roster — which is the point. The old hand-written+ * A rule, not a list — which is the point. The old hand-written* ALL_PROVIDER_CHOICES array made membership opt-in, so `devin` and* `antigravity` were both invisible here while working everywhere else* (the config TUI derives its list from the same definitions).@@ -1023,7 +1024,7 @@ async function getProviderChoices() {* this provider. The four provider rows here are longer aliases that are* equally valid and read better on a command line the user may copy —* `google@gemini-3-pro` over- * `g@gemini-3-pro`. `zen` is a legacy picker VALUE (the roster now uses the+ * `g@gemini-3-pro`. `zen` is a legacy picker VALUE (the provider list now uses the* definition name `opencode-zen`); kept so an old caller still resolves.** Do NOT add a row here just because a provider is new — the derived path@@ -1260,7 +1261,7 @@ async function pickModelFromList(* - **Not everything served is chat.** Alibaba's plan host answers with image* and TTS models alongside chat ones. Filtered with the SAME predicate the* probe path uses, so the picker and `--probe` agree on what is chat-capable- * — a name-based rule, never a model-id skip list. An all-non-chat roster+ * — a name-based rule, never a model-id skip list. An all-non-chat dynamic models catalog* returns [] so the caller falls through to the catalog / free-text path* instead of showing an empty list.* - **Neither price nor (always) context is reported.** Context comes from the@@ -1275,7 +1276,7 @@ async function pickModelFromList(* A live offer as a short badge, or undefined when there is nothing to say.** Evaluated against the clock on every render, never cached: two of the four- * promos on the measured Devin roster expired within days of being observed,+ * promos in the measured Devin dynamic models catalog expired within days of being observed,* and a stale "FREE" badge is a wrong-price bug — strictly worse than no badge.*/function describeOffer(offer: ModelOffer | undefined): string | undefined {@@ -1296,7 +1297,7 @@ function describeOffer(offer: ModelOffer | undefined): string | undefined {* inquirer on stdout, and a model spec is what this module ultimately RETURNS* — a notice must not be capturable as part of a piped answer.*- * `empty-roster` is deliberately silent: an endpoint that answers correctly+ * `empty-models-catalog` is deliberately silent: an endpoint that answers correctly* with nothing chat-capable is not an error, and the free-text prompt that* follows is the right affordance for it. Only actionable failures are named.*@@ -1308,7 +1309,7 @@ export function warnDiscoveryFailure(def: Pick<ProviderDefinition, "apiKeyEnvVar" | "apiKeyUrl">): void {const failure = getDiscoveryFailure(provider);- if (!failure || failure.kind === "empty-roster") return;+ if (!failure || failure.kind === "empty-models-catalog") return;process.stderr.write(`\n⚠ ${displayName} could not list its models: ${describeDiscoveryFailure(failure)}\n`@@ -1363,7 +1364,7 @@ export async function buildDiscoveredModelRows(// where 167 served uids are ~39 real choices multiplied out by reasoning tier// and speed premium. The chosen id is always a real wire id, so it still// round-trips through buildExplicitModelSpec and argv unchanged.- const choices = collapseRoster(provider, discovered.map(toRosterEntry));+ const choices = collapseModelsCatalog(provider, discovered.map(toModelsCatalogEntry));const discoveredById = new Map(discovered.map((m) => [m.id, m]));// The live endpoint decides WHICH models appear (entitlement) and overrides@@ -1411,7 +1412,7 @@ export async function buildDiscoveredModelRows(// Present newest-first, with the SAME comparator the catalog path uses, so// every picker in claudish orders identically. `rankDiscoveredModels` above// is only a stable, deterministic input order — its widest-window-first rule- // is meaningful for probe candidates, but for a plan whose entire roster is+ // is meaningful for probe candidates, but for a plan whose entire dynamic models catalog is// 1M it collapses to alphabetical, which is what made the picker look unsorted.return sortModelsNewestFirst(rows);}@@ -1432,7 +1433,7 @@ async function selectModelFromProvider(const prefix = pickerModelPrefix(provider) ?? `${provider}@`;const displayName = getPickerDisplayName(provider);- // Subscription providers (e.g. Kimi Coding) serve a roster that only their+ // Subscription providers (e.g. Kimi Coding) serve a dynamic models catalog that only their// own authenticated endpoint knows — the owner's public catalog lists models// the subscription can't serve, and the per-tier context windows aren't in it// at all. Ask the endpoint, and offer exactly what this user's plan allows.@@ -1449,7 +1450,7 @@ async function selectModelFromProvider(//// The fall-through itself is correct and deliberate; what was wrong was// doing it silently. A rejected API key and a provider that publishes no- // roster both ended up as the same free-text prompt, so a credential+ // dynamic models catalog both ended up as the same free-text prompt, so a credential// error was indistinguishable from normal behaviour.warnDiscoveryFailure(provider, displayName, def);}@@ -1458,7 +1459,7 @@ async function selectModelFromProvider(// Ollama and LM Studio used to be handled here, each its own way — Ollama by// an inline `/api/tags` branch, LM Studio not at all (free-text only). Both// now declare `modelDiscovery`, so the block above lists them like every- // other provider that knows its own roster. A local daemon is not a different+ // other provider that knows its own dynamic models catalog. A local daemon is not a different// KIND of thing; it is a provider whose endpoint happens to be on localhost.// Local / user-deployed providers: Firebase has no catalog, free-text only.diff --git a/packages/cli/src/providers-command.ts b/packages/cli/src/providers-command.tsindex 9d0bb27e..701558a4 100644--- a/packages/cli/src/providers-command.ts+++ b/packages/cli/src/providers-command.ts@@ -62,7 +62,7 @@ export async function collectProviderStatuses(// Called with NO argument on purpose. The `config` parameter here is a// deliberately narrow `{apiKeys?, localProviders?}` — it exists so the config// TUI can classify against an UNSAVED snapshot — and handing that to- // registration would mean deciding the endpoint roster from an object with no+ // registration would mean deciding the endpoint list from an object with no// `customEndpoints` and no `predefinedEndpoints` at all. Every user endpoint// would vanish and every opt-out would be ignored, silently, for any caller// that passed one. Registration reads the real config itself; the latch makesdiff --git a/packages/cli/src/providers/custom-endpoints-loader.ts b/packages/cli/src/providers/custom-endpoints-loader.tsindex 17481e54..9fb62644 100644--- a/packages/cli/src/providers/custom-endpoints-loader.ts+++ b/packages/cli/src/providers/custom-endpoints-loader.ts@@ -108,7 +108,7 @@ export interface EndpointDefinitionOverrides {apiKeyUrl?: string;/** Credential label shown in the TUI. Default: `<name> (custom endpoint)`. */apiKeyDescription?: string;- /** Roster/TUI description. Default: `Custom endpoint: <name>`. */+ /** Picker/TUI description. Default: `Custom endpoint: <name>`. */description?: string;}diff --git a/packages/cli/src/providers/devin/devin-models.ts b/packages/cli/src/providers/devin/devin-models.tsindex 2c244e93..a10672ac 100644--- a/packages/cli/src/providers/devin/devin-models.ts+++ b/packages/cli/src/providers/devin/devin-models.ts@@ -1,7 +1,7 @@/**- * The LIVE Devin model roster — capability ∩ entitlement.+ * Devin's dynamic models catalog — capability ∩ entitlement.*- * No model roster, context window, family name, or reasoning tier is ever+ * No model list, context window, family name, or reasoning tier is ever* written into claudish source: which models a subscription serves is* per-account and drifts, so it is discovered from the backend's own metadata* rpcs and cached briefly. The only literals here are endpoint paths, field@@ -65,8 +65,8 @@ const MODEL_CONFIGS_PATH = "/exa.api_server_pb.ApiServerService/GetCliModelConfi/** Entitlement rpc — note the DIFFERENT service, not ApiServerService. */const TEAM_SETTINGS_PATH = "/exa.seat_management_pb.SeatManagementService/GetCliTeamSettings";-/** Roster TTL. The served set moves on a daily cadence, not per request. */-const ROSTER_TTL_MS = 5 * 60 * 1000;+/** Dynamic models catalog TTL. It moves on a daily cadence, not per request. */+const MODELS_CATALOG_TTL_MS = 5 * 60 * 1000;/** Metadata rpcs are small and on interactive paths (picker, probe). */const UNARY_TIMEOUT_MS = 10_000;@@ -82,7 +82,7 @@ export interface DevinPromo {** Devin STATES what each uid is instead of leaving it to be parsed out of the* name, so this is the authority wherever it is present. Measured against a- * real roster:+ * real dynamic models catalog:** - `label` on the effort axis agrees with the uid suffix on all 130 uids that* carry both, and additionally expresses `Minimal`, which no suffix rule does.@@ -102,7 +102,7 @@ export interface DevinAxis {enabled: boolean;}-/** One row of the live roster. */+/** One row of the dynamic models catalog. */export interface DevinModelConfig {/** Routing key — request field 21 (e.g. `claude-opus-5-high`). */uid: string;@@ -334,7 +334,7 @@ export async function fetchDevinModelConfigs(apiKey: string): Promise<DevinModel/*** Pure decode of a `GetCliModelConfigs` response body.*- * Split out from the fetch so the roster decode can be replayed against a+ * Split out from the fetch so decoding the dynamic models catalog can be replayed against a* captured response byte-for-byte — the fixture is real wire data, never* hand-written.*/@@ -365,8 +365,8 @@ export async function fetchDevinAllowedUids(apiKey: string): Promise<string[]> {return uids;}-let rosterCache: DevinModelConfig[] | null = null;-let rosterCacheAt = 0;+let modelsCatalogCache: DevinModelConfig[] | null = null;+let modelsCatalogCacheAt = 0;/*** The models this subscription can actually run: configs ∩ allowed uids, minus@@ -375,7 +375,7 @@ let rosterCacheAt = 0;* That last filter drops `adaptive` (ctx 0), a server-side router pseudo-model* that is not a routable uid — it must never reach the picker or a probe.*- * Cached for {@link ROSTER_TTL_MS}. Never throws: a stale cache beats an empty+ * Cached for {@link MODELS_CATALOG_TTL_MS}. Never throws: a stale cache beats an empty* one, and an empty one beats a failed turn.*/export async function getServedDevinModels(opts?: {@@ -385,17 +385,18 @@ export async function getServedDevinModels(opts?: {apiKey?: string;}): Promise<DevinModelConfig[]> {const now = Date.now();- if (!opts?.force && rosterCache && now - rosterCacheAt < ROSTER_TTL_MS) return rosterCache;+ if (!opts?.force && modelsCatalogCache && now - modelsCatalogCacheAt < MODELS_CATALOG_TTL_MS)+ return modelsCatalogCache;const apiKey = opts?.apiKey ?? readDevinApiKey();- if (!apiKey) return rosterCache ?? [];+ if (!apiKey) return modelsCatalogCache ?? [];try {const [configs, allowed] = await Promise.all([fetchDevinModelConfigs(apiKey),fetchDevinAllowedUids(apiKey),]);- if (configs.length === 0) return rosterCache ?? [];+ if (configs.length === 0) return modelsCatalogCache ?? [];const entitled = new Set(allowed);const served = configs.filter(@@ -405,17 +406,17 @@ export async function getServedDevinModels(opts?: {log("[Devin] entitlement unknown — using the full config list (superset)");}- rosterCache = served;- rosterCacheAt = now;+ modelsCatalogCache = served;+ modelsCatalogCacheAt = now;return served;} catch (err) {log(`[Devin] served-model discovery error: ${err}`);- return rosterCache ?? [];+ return modelsCatalogCache ?? [];}}-/** Test seam: drop the cached roster. */+/** Test seam: drop the cached dynamic models catalog. */export function _resetDevinModelCache(): void {- rosterCache = null;- rosterCacheAt = 0;+ modelsCatalogCache = null;+ modelsCatalogCacheAt = 0;}diff --git a/packages/cli/src/providers/devin/model-id-resolver.test.ts b/packages/cli/src/providers/devin/model-id-resolver.test.tsindex bca3bd09..e44f94ac 100644--- a/packages/cli/src/providers/devin/model-id-resolver.test.ts+++ b/packages/cli/src/providers/devin/model-id-resolver.test.ts@@ -18,22 +18,29 @@ function servedModel(uid: string, family = "claude-opus-5"): DevinModelConfig {};}-// Synthetic and injected: the resolver tests never fetch or pin a real roster.-const effortRoster = (["none", "minimal", "low", "medium", "high", "xhigh", "max"] as const).map(- (effort) => servedModel(`claude-opus-5-${effort}`)-);+// Synthetic and injected: the resolver tests never fetch or pin a real dynamic models catalog.+const effortModelsCatalog = (+ ["none", "minimal", "low", "medium", "high", "xhigh", "max"] as const+).map((effort) => servedModel(`claude-opus-5-${effort}`));describe("resolveDevinModelUid", () => {test("passes an exact served uid through untouched", () => {- const roster = [servedModel("claude-opus-5-high"), servedModel("claude-opus-5-max-fast")];- expect(resolveDevinModelUid("claude-opus-5-high", "low", roster)).toBe("claude-opus-5-high");- expect(resolveDevinModelUid("claude-opus-5-max-fast", undefined, roster)).toBe(+ const modelsCatalog = [+ servedModel("claude-opus-5-high"),+ servedModel("claude-opus-5-max-fast"),+ ];+ expect(resolveDevinModelUid("claude-opus-5-high", "low", modelsCatalog)).toBe(+ "claude-opus-5-high"+ );+ expect(resolveDevinModelUid("claude-opus-5-max-fast", undefined, modelsCatalog)).toBe("claude-opus-5-max-fast");});test("resolves a family plus effort to the suffixed uid", () => {- expect(resolveDevinModelUid("claude-opus-5", "high", effortRoster)).toBe("claude-opus-5-high");+ expect(resolveDevinModelUid("claude-opus-5", "high", effortModelsCatalog)).toBe(+ "claude-opus-5-high"+ );});for (const effort of [@@ -46,20 +53,29 @@ describe("resolveDevinModelUid", () => {"max",] satisfies EffortLevel[]) {test(`supports the ${effort} effort vocabulary`, () => {- expect(resolveDevinModelUid("claude-opus-5", effort, effortRoster)).toBe(+ expect(resolveDevinModelUid("claude-opus-5", effort, effortModelsCatalog)).toBe(`claude-opus-5-${effort}`);});}test("never selects a fast variant implicitly", () => {- const roster = [servedModel("claude-opus-5-medium"), servedModel("claude-opus-5-max-fast")];+ const modelsCatalog = [+ servedModel("claude-opus-5-medium"),+ servedModel("claude-opus-5-max-fast"),+ ];- expect(resolveDevinModelUid("claude-opus-5", "max", roster)).toBe("claude-opus-5-medium");- expect(resolveDevinModelUid("claude-opus-5", undefined, roster)).toBe("claude-opus-5-medium");+ expect(resolveDevinModelUid("claude-opus-5", "max", modelsCatalog)).toBe(+ "claude-opus-5-medium"+ );+ expect(resolveDevinModelUid("claude-opus-5", undefined, modelsCatalog)).toBe(+ "claude-opus-5-medium"+ );});test("passes an unknown model through unchanged", () => {- expect(resolveDevinModelUid("unknown-model", "high", effortRoster)).toBe("unknown-model");+ expect(resolveDevinModelUid("unknown-model", "high", effortModelsCatalog)).toBe(+ "unknown-model"+ );});});diff --git a/packages/cli/src/providers/devin/model-id-resolver.ts b/packages/cli/src/providers/devin/model-id-resolver.tsindex dd39ee76..2a9b1c91 100644--- a/packages/cli/src/providers/devin/model-id-resolver.ts+++ b/packages/cli/src/providers/devin/model-id-resolver.ts@@ -4,12 +4,12 @@* This is now a THIN DELEGATE over the shared model-resolver seam* (`providers/model-resolvers/`). The rules, and the measurements behind them,* live in `model-resolvers/devin.ts`; this file exists only so the transport can- * keep passing a `DevinModelConfig[]` roster without knowing about the seam.+ * keep passing a `DevinModelConfig[]` dynamic models catalog without knowing about the seam.** ## Why the logic moved** The previous implementation parsed the reasoning tier out of the uid SPELLING- * and stripped only a `-fast` suffix. Measured against a real 170-uid roster,+ * and stripped only a `-fast` suffix. Measured against a real 170-uid dynamic models catalog,* that had two defects:** - **The 1M variants were unreachable.** `glm-5-2-max-1m` does not end in a@@ -26,12 +26,12 @@*/import type { EffortLevel } from "../../adapters/base-api-format.js";-import { devinRosterEntry } from "../model-resolvers/devin.js";+import { devinModelsCatalogEntry } from "../model-resolvers/devin.js";import { expandSelection } from "../model-resolvers/registry.js";import type { DevinModelConfig } from "./devin-models.js";/**- * Resolve `requested` to a served uid against the LIVE roster.+ * Resolve `requested` to a served uid against the dynamic models catalog.** Total by construction: an unrecognised request is returned unchanged so the* backend can answer for itself, which is what lets the served-set-aware error@@ -44,5 +44,5 @@ export function resolveDevinModelUid(): string {const trimmed = requested.trim();if (!trimmed || served.length === 0) return trimmed || requested;- return expandSelection("devin", trimmed, served.map(devinRosterEntry), { effort });+ return expandSelection("devin", trimmed, served.map(devinModelsCatalogEntry), { effort });}diff --git a/packages/cli/src/providers/endpoint-registration.ts b/packages/cli/src/providers/endpoint-registration.tsindex b84a903f..df382296 100644--- a/packages/cli/src/providers/endpoint-registration.ts+++ b/packages/cli/src/providers/endpoint-registration.ts@@ -1,7 +1,7 @@/*** The one place endpoints get registered, and the one place that decides when.*- * Six call sites need the provider roster to already contain the bundled+ * Six call sites need the provider list to already contain the bundled* catalog — the interactive picker, `--probe`, the proxy, the parent-side* spawn pre-hydration, the config TUI and `serve` — and each of them reaches it* by a different path. A latch here means the answer does not depend on whichdiff --git a/packages/cli/src/providers/missing-credential-hint.test.ts b/packages/cli/src/providers/missing-credential-hint.test.tsindex 5a554440..b36e19a0 100644--- a/packages/cli/src/providers/missing-credential-hint.test.ts+++ b/packages/cli/src/providers/missing-credential-hint.test.ts@@ -9,7 +9,7 @@* `SUBSCRIPTION_PROVIDERS`.** Everything here is derived from the live definitions rather than pinned to a- * roster: which providers are local, and which carry an `oauthFallback`, are+ * provider list: which providers are local, and which carry an `oauthFallback`, are* both read at run time, so adding a provider cannot leave a stale list behind.*/@@ -126,7 +126,7 @@ describe("describeMissingCredential — everything else is unchanged", () => {const signup = info?.url ? ` Get one at ${info.url}.` : "";// `siblingKeyEnvVars` appends one clause and is declared by almost nobody,// so the expectation is derived from the SAME field rather than pinned to- // a roster — a provider adopting it later must not have to edit this test,+ // a provider list — a provider adopting it later must not have to edit this test,// and one that quietly LOSES the declaration must still fail below.const expected = keyNames? `No API key for provider "${name}". Set ${keyNames} (env, config, or 1Password import).${signup}${siblingClause(name)}`diff --git a/packages/cli/src/providers/model-availability.test.ts b/packages/cli/src/providers/model-availability.test.tsindex 0740f4fd..7374cbe5 100644--- a/packages/cli/src/providers/model-availability.test.ts+++ b/packages/cli/src/providers/model-availability.test.ts@@ -59,7 +59,7 @@ function discoveryProvider(): ProviderDefinition {};}-function stubRoster(...ids: string[]): void {+function stubModelsCatalog(...ids: string[]): void {globalThis.fetch = mock(async () =>new Response(JSON.stringify({ data: ids.map((id) => ({ id })) }), {@@ -112,7 +112,7 @@ describe("providerServesModel catalog evidence", () => {// Catalog coverage is partial, so absence from one row cannot deny service:// openai-codex appears on only 1 of ~760 rows yet genuinely serves gpt-5.- // Only a complete live roster may turn absence into "not-served".+ // Only a complete dynamic models catalog may turn absence into "not-served".expect(await providerServesModel(CATALOG_PROVIDER, "model-one")).toBe("unknown");});@@ -168,21 +168,21 @@ describe("providerServesModel catalog evidence", () => {});describe("providerServesModel live discovery", () => {- test("returns serves for exact and case-insensitive roster membership", async () => {- stubRoster("MiniMax-M3");+ test("returns serves for exact and case-insensitive matches in the dynamic models catalog", async () => {+ stubModelsCatalog("MiniMax-M3");expect(await providerServesModel(DISCOVERY_PROVIDER, "MiniMax-M3")).toBe("serves");expect(await providerServesModel(DISCOVERY_PROVIDER, "minimax-m3")).toBe("serves");});- test("returns not-served when a non-empty roster omits the model", async () => {- stubRoster("available-model");+ test("returns not-served when a non-empty dynamic models catalog omits the model", async () => {+ stubModelsCatalog("available-model");expect(await providerServesModel(DISCOVERY_PROVIDER, "missing-model")).toBe("not-served");});- test("returns unknown when discovery returns an empty roster", async () => {- stubRoster();+ test("returns unknown when discovery returns an empty dynamic models catalog", async () => {+ stubModelsCatalog();// A temporarily unavailable listing endpoint must not silently switch the// user away from the provider they selected and may already pay for.@@ -191,7 +191,7 @@ describe("providerServesModel live discovery", () => {test("uses discovery before conflicting catalog evidence", async () => {_setCatalogEntriesForTest([catalogEntry("model-one", [{ provider: DISCOVERY_PROVIDER }])]);- stubRoster("different-model");+ stubModelsCatalog("different-model");expect(await providerServesModel(DISCOVERY_PROVIDER, "model-one")).toBe("not-served");});diff --git a/packages/cli/src/providers/model-availability.ts b/packages/cli/src/providers/model-availability.tsindex 200facf6..1d4efc02 100644--- a/packages/cli/src/providers/model-availability.ts+++ b/packages/cli/src/providers/model-availability.ts@@ -35,12 +35,12 @@ import { catalogRouteMatchesProvider } from "./catalog-route-bindings.js";* `deepseek-v4-pro-0813` on a bare chain reaches `opencode-zen-go` first. Zen Go* does not carry it and says so with **HTTP 401** — a status that reads as a* credential failure, and which masked the live OpenRouter hop further down the- * same chain. Its live roster answers the question before any request is sent:+ * same chain. Its dynamic models catalog answers the question before any request is sent:** Zen Go serves 26 models; 'deepseek-v4-pro-0813' is not among them* (it serves the undated `deepseek-v4-pro`), while 'kimi-k3' is.*- * Note the naming detail — the roster also settles what to SEND, which a+ * Note the naming detail — the dynamic models catalog also settles what to SEND, which a* membership test against catalog ids alone could not.*/@@ -56,8 +56,8 @@ import { getProviderByName } from "./provider-definitions.js";*/export type ModelAvailability = "serves" | "not-served" | "unknown";-/** Case-insensitive membership, since rosters disagree on casing (`MiniMax-M3`). */-function rosterHas(ids: string[], wireId: string): boolean {+/** Case-insensitive match, since dynamic models catalogs disagree on casing (`MiniMax-M3`). */+function modelsCatalogHas(ids: string[], wireId: string): boolean {const needle = wireId.trim().toLowerCase();return ids.some((id) => id.trim().toLowerCase() === needle);}@@ -82,15 +82,15 @@ export async function providerServesModel(if (def?.modelDiscovery) {const models = await discoverProviderModels(provider);if (models.length > 0) {- return rosterHas(+ return modelsCatalogHas(models.map((m) => m.id),wireId)? "serves": "not-served";}- // An empty roster is never a "no". `empty-roster` means the endpoint- // answered with nothing to offer; every other kind (unauthorized,+ // An empty dynamic models catalog is never a "no". `empty-models-catalog`+ // means the endpoint answered with nothing to offer; every other kind (unauthorized,// unreachable, malformed…) means we failed to ask. Both are `unknown` —// dropping a candidate because its listing endpoint was briefly down would// turn a transient blip into a silent provider switch.@@ -110,7 +110,7 @@ export async function providerServesModel(// but one. Caught by `route()`'s own tests, which is exactly what they are// for.//- // A live roster is different in kind: it is COMPLETE BY CONSTRUCTION, because+ // A dynamic models catalog is different in kind: it is COMPLETE BY CONSTRUCTION, because// the provider is enumerating everything the caller's key can reach. Only that// may deny. The catalog can still confirm, which is free and useful.const entries = getCatalogEntries();diff --git a/packages/cli/src/providers/model-catalog.ts b/packages/cli/src/providers/model-catalog.tsindex 8da0532e..7e3512e4 100644--- a/packages/cli/src/providers/model-catalog.ts+++ b/packages/cli/src/providers/model-catalog.ts@@ -111,7 +111,7 @@ const OWNER_PROVIDER_SLUGS = new Set<string>([]);/**- * Cold-start seed for the aggregator set. NOT the roster — see+ * Cold-start seed for the aggregator set. NOT the list itself — see* {@link aggregatorProviderSlugs}, which derives the real one from the catalog.** Used only when the slim cache is empty (first run, before the first fetch),diff --git a/packages/cli/src/providers/model-discovery-builtins.ts b/packages/cli/src/providers/model-discovery-builtins.tsindex c9232330..10454159 100644--- a/packages/cli/src/providers/model-discovery-builtins.ts+++ b/packages/cli/src/providers/model-discovery-builtins.ts@@ -1,5 +1,5 @@/**- * The three builtin roster fetchers that are not an HTTP GET.+ * The three builtin dynamic models catalog fetchers that are not an HTTP GET.** They live HERE, behind `registerModelDiscoveryFetcher`, rather than as* `if (descriptor.format === …)` branches inside `model-discovery.ts`. The@@ -22,20 +22,20 @@ import type { DiscoveredModel } from "./model-discovery.js";import { registerModelDiscoveryFetcher } from "./model-discovery.js";/**- * Devin's roster is capability ∩ entitlement over two protobuf rpcs.+ * Devin's dynamic models catalog is capability ∩ entitlement over two protobuf rpcs.** Carries the full variant metadata, not just id/name/window: the picker folds* ~170 uids into ~42 rows and needs the group label, the cost multiplier, the* promo and the vendor's own default flag to do it.*/-async function fetchDevinRoster(): Promise<DiscoveredModel[]> {+async function fetchDevinModelsCatalog(): Promise<DiscoveredModel[]> {const { getServedDevinModels } = await import("./devin/devin-models.js");const served = await getServedDevinModels();if (served.length === 0) return [];- const { devinRosterEntry } = await import("./model-resolvers/devin.js");+ const { devinModelsCatalogEntry } = await import("./model-resolvers/devin.js");return served.map((model) => {- const { wireId, ...rest } = devinRosterEntry(model);+ const { wireId, ...rest } = devinModelsCatalogEntry(model);return { id: wireId, ...rest };});}@@ -50,8 +50,9 @@ async function fetchDevinRoster(): Promise<DiscoveredModel[]> {* no images, `recommended: false`. They answer HTTP 200, which is precisely what* made a wrong-host problem look like a rate limit for an entire session.*- * Prefix-matched, because the ids carry build numbers that rot on each roster- * roll. This is the ONLY guess left in the filter; everything else is declared.+ * Prefix-matched, because the ids carry build numbers that rot on each roll of+ * the dynamic models catalog. This is the ONLY guess left in the filter; everything+ * else is declared.*/function isUndeclaredEditorInternal(id: string): boolean {return id.startsWith("tab_");@@ -66,7 +67,7 @@ function isUndeclaredEditorInternal(id: string): boolean {* in the catalog. `resolveDiscoveredContextLength` already prefers a discovered* window over the catalog; this is what gives it one to prefer.*/-async function fetchAntigravityRoster(): Promise<DiscoveredModel[]> {+async function fetchAntigravityModelsCatalog(): Promise<DiscoveredModel[]> {const { getValidAntigravityAccessToken } = await import("../auth/antigravity-token.js");const { setupAntigravityUser, getServedAntigravityModels } = await import("../auth/antigravity-user.js"@@ -97,7 +98,7 @@ async function fetchAntigravityRoster(): Promise<DiscoveredModel[]> {}/** Ollama's daemon speaks its own listing shape and carries capability data no OpenAI list has. */-async function fetchOllamaRoster(): Promise<DiscoveredModel[]> {+async function fetchOllamaModelsCatalog(): Promise<DiscoveredModel[]> {const { fetchOllamaModels } = await import("./ollama-discovery.js");const installed = await fetchOllamaModels({ enrichCapabilities: false });return installed.map((model) => ({@@ -107,6 +108,6 @@ async function fetchOllamaRoster(): Promise<DiscoveredModel[]> {}));}-registerModelDiscoveryFetcher("devin-connect", fetchDevinRoster);-registerModelDiscoveryFetcher("antigravity", fetchAntigravityRoster);-registerModelDiscoveryFetcher("ollama-tags", fetchOllamaRoster);+registerModelDiscoveryFetcher("devin-connect", fetchDevinModelsCatalog);+registerModelDiscoveryFetcher("antigravity", fetchAntigravityModelsCatalog);+registerModelDiscoveryFetcher("ollama-tags", fetchOllamaModelsCatalog);diff --git a/packages/cli/src/providers/model-discovery.test.ts b/packages/cli/src/providers/model-discovery.test.tsindex 156a3467..6dd0539d 100644--- a/packages/cli/src/providers/model-discovery.test.ts+++ b/packages/cli/src/providers/model-discovery.test.ts@@ -296,12 +296,12 @@ describe("discoverProviderModels", () => {expect(await discoverProviderModels("kimi-coding")).toEqual([]);});- test("classifies a 200 response with an empty data array as empty-roster", async () => {+ test("classifies a 200 response with an empty data array as empty-models-catalog", async () => {stubJsonResponse({ data: [] });expect(await discoverProviderModels("kimi-coding")).toEqual([]);expect(getDiscoveryFailure("kimi-coding")).toMatchObject({- kind: "empty-roster",+ kind: "empty-models-catalog",provider: "kimi-coding",});});@@ -455,9 +455,9 @@ describe("describeDiscoveryFailure", () => {substrings: ["not valid JSON", "https://example.test/models"],},{- kind: "empty-roster",+ kind: "empty-models-catalog",failure: {- kind: "empty-roster",+ kind: "empty-models-catalog",provider: "test",endpoint: "https://example.test/models",},@@ -488,7 +488,7 @@ describe("rankDiscoveredModels", () => {expect(ranked).not.toBe(models);});- test("ranks k3 first for the real Kimi roster", () => {+ test("ranks k3 first for the real Kimi dynamic models catalog", () => {const ranked = rankDiscoveredModels([{ id: "kimi-for-coding-highspeed", contextWindow: 262_144 },{ id: "k3-256k", contextWindow: 262_144 },@@ -501,7 +501,7 @@ describe("rankDiscoveredModels", () => {});describe("OpenAI-compatible probe discovery", () => {- test("uses an authenticated account roster and advances past excluded models", async () => {+ test("uses an authenticated dynamic models catalog and advances past excluded models", async () => {const provider: ProviderDefinition = {name: "account-scoped-openai-probe-test",displayName: "Account Scoped Probe Test",diff --git a/packages/cli/src/providers/model-discovery.ts b/packages/cli/src/providers/model-discovery.tsindex d8279369..1d940791 100644--- a/packages/cli/src/providers/model-discovery.ts+++ b/packages/cli/src/providers/model-discovery.ts@@ -1,7 +1,7 @@/*** Live, per-subscription model discovery.*- * Some providers serve a model roster — and per-model context windows — that+ * Some providers serve a dynamic models catalog — and per-model context windows — that* the cloud catalog CANNOT know statically, because the answer depends on the* caller's subscription tier. Kimi Code is the motivating case: `k3` advertises* a 1M context window, but only on Allegretto or higher; a Moderato subscriber@@ -26,7 +26,7 @@import { credentials } from "../auth/credentials/authority.js";import { log } from "../logger.js";import { VERSION } from "../version.js";-import type { ModelOffer, RosterAxis, RosterEntry } from "./model-resolvers/types.js";+import type { ModelOffer, ModelsCatalogAxis, ModelsCatalogEntry } from "./model-resolvers/types.js";import { getProviderByName } from "./provider-definitions.js";/** A model as reported by the provider's own live endpoint. */@@ -42,9 +42,9 @@ export interface DiscoveredModel {* when it reports a plausible one.** NOT an authoritative release date — it is whenever the provider added the- * model to this account's roster — but for a plan endpoint whose models the+ * model to this account's dynamic models catalog — but for a plan endpoint whose models the* public catalog may not list at all, it is the only freshness signal there- * is, and it orders a roster correctly. The catalog's own `releaseDate` wins+ * is, and it orders a dynamic models catalog correctly. The catalog's own `releaseDate` wins* wherever it exists; this covers the rest.*/releaseDate?: string;@@ -52,8 +52,8 @@ export interface DiscoveredModel {* Ignore the CATALOG's release date for this row, ordering it by the version* encoded in its id instead.*- * For a roster of tuned VARIANTS the catalog does not list, a catalog date is- * not a fresher signal — it is a date for a DIFFERENT model. Antigravity is+ * For a dynamic models catalog of tuned VARIANTS the cloud catalog does not list,+ * a catalog date is not a fresher signal — it is a date for a DIFFERENT model. Antigravity is* the case: measured 2026-08-24, 6 of its 19 served ids had a catalog date and* every one of those six was an OLD base model, while every new variant had* none:@@ -66,7 +66,7 @@ export interface DiscoveredModel {* `compareByReleaseDateDesc` puts undated rows after dated ones, so a 2025* model sat at the top of the picker and the newest model on the plan sat at* the bottom — the exact complaint that prompted this. With the catalog date- * suppressed the whole roster falls through to the version-parts rule and+ * suppressed the whole dynamic models catalog falls through to the version-parts rule and* orders 3.7 > 3.6 > 3.5 > 3.1 > 3 > 2.5, which is what the user meant by* "newest first".*/@@ -76,7 +76,7 @@ export interface DiscoveredModel {** Most endpoints report a flat list where each id is already the thing a* human would choose, and leave all of this undefined. Devin does not: its- * roster is ~33 models multiplied out by reasoning tier, speed premium, and+ * dynamic models catalog is ~33 models multiplied out by reasoning tier, speed premium, and* context window, and it publishes which is which. Carrying that here is what* lets `providers/model-resolvers/` fold 170 ids into 42 rows and unfold them* again at request time.@@ -87,7 +87,7 @@ export interface DiscoveredModel {costTier?: number;isFamilyDefault?: boolean;isRecommended?: boolean;- axes?: RosterAxis[];+ axes?: ModelsCatalogAxis[];offer?: ModelOffer;/*** Whether the model can call tools, when the endpoint says so.@@ -95,13 +95,13 @@ export interface DiscoveredModel {* Only Ollama reports it (via its own capability list / name heuristics), and* it genuinely varies there — a local embedding or vision-only pull cannot* drive Claude Code. Undefined means "not reported", which callers read as- * yes, since every hosted roster in claudish is tool-capable.+ * yes, since every hosted dynamic models catalog in claudish is tool-capable.*/supportsTools?: boolean;}/** A discovered model in the shape the model-resolver seam consumes. */-export function toRosterEntry(model: DiscoveredModel): RosterEntry {+export function toModelsCatalogEntry(model: DiscoveredModel): ModelsCatalogEntry {const { id, ...rest } = model;return { wireId: id, ...rest };}@@ -133,10 +133,10 @@ export interface ModelDiscoveryDescriptor {}/**- * A roster fetcher for a format this module does not know how to speak.+ * A dynamic models catalog fetcher for a format this module does not know how to speak.** Returning `[]` means "asked, got nothing" — the caller records an- * `empty-roster` failure. Throwing is also fine; the caller converts it.+ * `empty-models-catalog` failure. Throwing is also fine; the caller converts it.*/export type ModelDiscoveryFetcher = (providerName: string) => Promise<DiscoveredModel[]>;@@ -184,7 +184,7 @@ const _cache = new Map<string, { models: DiscoveredModel[]; expiresAt: number }>* launch or a picker — but for a long time "fail-soft" also meant "fail* SILENT": all five failure modes below collapsed into the same empty array* behind a `--debug`-only log line. Callers could not tell "your API key was- * rejected" from "this provider genuinely publishes no roster", so the picker+ * rejected" from "this provider genuinely publishes no models", so the picker* rendered a rejected credential as a free-text prompt, which reads as a* feature rather than an error. (Measured on qwen-token-plan: a 401 from Alibaba's* plan host was indistinguishable in the UI from a provider with no list.)@@ -192,7 +192,7 @@ const _cache = new Map<string, { models: DiscoveredModel[]; expiresAt: number }>* The kinds are ordered by what the user should do about them, not by HTTP* status: `unauthorized` and `no-credentials` are setup problems the user can* fix, `unreachable` and `http-error` are usually transient, and- * `empty-roster` means the endpoint answered correctly with nothing to offer —+ * `empty-models-catalog` means the endpoint answered correctly with nothing to offer —* the only kind where falling through to a catalog or free-text entry is the* genuinely right response.*/@@ -202,7 +202,7 @@ export type DiscoveryFailureKind =| "http-error"| "unreachable"| "malformed"- | "empty-roster";+ | "empty-models-catalog";export interface DiscoveryFailure {kind: DiscoveryFailureKind;@@ -285,7 +285,7 @@ export function describeDiscoveryFailure(failure: DiscoveryFailure): string {return `the model list was unreachable${at}${because}`;case "malformed":return `the model list was not valid JSON${at}`;- case "empty-roster":+ case "empty-models-catalog":return `the endpoint answered${at} but listed no models`;}}@@ -389,7 +389,7 @@ export async function discoverProviderModels(providerName: string): Promise<Discconst descriptor = def?.modelDiscovery;if (!def || !descriptor) return [];- // Devin's roster is not a GET — it is two protobuf rpcs (capability ∩+ // Devin's dynamic models catalog is not a GET — it is two protobuf rpcs (capability ∩// entitlement), owned by the module that speaks that wire. The import is// DYNAMIC to keep the codec off the cold-start path and out of a static// cycle. (If a third such provider ever appears, replace this branch with a@@ -405,13 +405,13 @@ export async function discoverProviderModels(providerName: string): Promise<Discfetcher = getModelDiscoveryFetcher(descriptor.format);}if (!fetcher) {- // A declared format nothing claims. Report it as a roster miss rather- // than falling through to the GET path, which would hit a bogus URL.- return recordFailure({ kind: "empty-roster", provider: providerName });+ // A declared format nothing claims. Report it as an empty dynamic models+ // catalog rather than falling through to the GET path, which would hit a bogus URL.+ return recordFailure({ kind: "empty-models-catalog", provider: providerName });}const models = await fetcher(providerName);if (models.length === 0) {- return recordFailure({ kind: "empty-roster", provider: providerName });+ return recordFailure({ kind: "empty-models-catalog", provider: providerName });}_failures.delete(providerName);log(`[model-discovery:${providerName}] discovered ${models.length} models`);@@ -464,7 +464,7 @@ export async function discoverProviderModels(providerName: string): Promise<Disc// Identify ourselves, and honour the definition's own headers.//- // `getRequestAuth` mints AUTH headers only, so a roster request went out with+ // `getRequestAuth` mints AUTH headers only, so a discovery request went out with// whatever User-Agent the runtime defaults to. Measured 2026-08-18: OpenCode// Zen Go answers such a request with `403 error code: 1010` — Cloudflare's// browser-integrity block, not an auth failure — while the identical request@@ -528,7 +528,7 @@ export async function discoverProviderModels(providerName: string): Promise<Discconst models = parseOpenAIModelsList(body);if (models.length === 0) {- return recordFailure({ kind: "empty-roster", provider: providerName, endpoint });+ return recordFailure({ kind: "empty-models-catalog", provider: providerName, endpoint });}_failures.delete(providerName);@@ -555,7 +555,7 @@ export async function discoverContextWindow(}/**- * Rank a discovered roster for presentation, largest context window first+ * Rank a dynamic models catalog for presentation, largest context window first* (ties broken alphabetically for determinism). The head of this list is the* picker's default.*@@ -566,7 +566,7 @@ export async function discoverContextWindow(** This is the PROBE-candidate ordering (`discoverProbeModel`), where widest* window first is the point. The model PICKER does not use it for presentation- * — a plan whose whole roster is 1M collapses to alphabetical — and sorts its+ * — a plan whose whole dynamic models catalog is 1M collapses to alphabetical — and sorts its* rows by release date instead; see `buildDiscoveredModelRows`.*/export function rankDiscoveredModels(models: DiscoveredModel[]): DiscoveredModel[] {diff --git a/packages/cli/src/providers/model-resolvers/devin.test.ts b/packages/cli/src/providers/model-resolvers/devin.test.tsindex 0377f497..e360a53b 100644--- a/packages/cli/src/providers/model-resolvers/devin.test.ts+++ b/packages/cli/src/providers/model-resolvers/devin.test.ts@@ -3,24 +3,24 @@ import { readFileSync } from "node:fs";import { join } from "node:path";import { EFFORT_LEVELS } from "../../adapters/base-api-format.js";import { decodeModelConfigs } from "../devin/devin-models.js";-import { DevinModelResolver, devinRosterEntry } from "./devin.js";-import type { ExpandContext, ModelChoice, RosterEntry } from "./types.js";+import { DevinModelResolver, devinModelsCatalogEntry } from "./devin.js";+import type { ExpandContext, ModelChoice, ModelsCatalogEntry } from "./types.js";import { groupKeyOf, nonEmpty, offerIsLive } from "./types.js";const fixturePath = join(import.meta.dir, "../../test-fixtures/devin/GetCliModelConfigs.res.bin");-const roster = decodeModelConfigs(readFileSync(fixturePath))+const modelsCatalog = decodeModelConfigs(readFileSync(fixturePath)).filter((config) => config.contextWindow > 0)- .map(devinRosterEntry);+ .map(devinModelsCatalogEntry);const resolver = new DevinModelResolver();-const choices = resolver.collapse(roster);-const rosterById = new Map(roster.map((entry) => [entry.wireId, entry]));+const choices = resolver.collapse(modelsCatalog);+const modelsCatalogById = new Map(modelsCatalog.map((entry) => [entry.wireId, entry]));-function groupKey(entry: RosterEntry): string {+function groupKey(entry: ModelsCatalogEntry): string {return JSON.stringify([groupKeyOf(entry), entry.contextWindow ?? 0]);}-function groupEntries(entries: RosterEntry[]): Map<string, RosterEntry[]> {- const groups = new Map<string, RosterEntry[]>();+function groupEntries(entries: ModelsCatalogEntry[]): Map<string, ModelsCatalogEntry[]> {+ const groups = new Map<string, ModelsCatalogEntry[]>();for (const entry of entries) {const key = groupKey(entry);const group = groups.get(key);@@ -30,9 +30,10 @@ function groupEntries(entries: RosterEntry[]): Map<string, RosterEntry[]> {return groups;}-function entryForChoice(choice: ModelChoice): RosterEntry {- const entry = rosterById.get(choice.id);- if (!entry) throw new Error(`Collapsed id is not in the fixture roster: ${choice.id}`);+function entryForChoice(choice: ModelChoice): ModelsCatalogEntry {+ const entry = modelsCatalogById.get(choice.id);+ if (!entry)+ throw new Error(`Collapsed id is not in the fixture dynamic models catalog: ${choice.id}`);return entry;}@@ -47,7 +48,7 @@ function seededPermutation<T>(values: T[], seed: number): T[] {return result;}-function exactNamesOf(entry: RosterEntry): string[] {+function exactNamesOf(entry: ModelsCatalogEntry): string[] {const names = new Map<string, string>();for (const value of [entry.groupLabel, entry.family]) {const name = nonEmpty(value);@@ -57,11 +58,11 @@ function exactNamesOf(entry: RosterEntry): string[] {return [...names.values()];}-function fixtureNameGroups(entries: RosterEntry[]): Array<{+function fixtureNameGroups(entries: ModelsCatalogEntry[]): Array<{selection: string;- entries: RosterEntry[];+ entries: ModelsCatalogEntry[];}> {- const byName = new Map<string, { selection: string; entries: RosterEntry[] }>();+ const byName = new Map<string, { selection: string; entries: ModelsCatalogEntry[] }>();for (const entry of entries) {for (const selection of exactNamesOf(entry)) {const key = selection.toLowerCase();@@ -76,7 +77,7 @@ function fixtureNameGroups(entries: RosterEntry[]): Array<{type DuplicateEffortCase = {choice: ModelChoice;effort: (typeof EFFORT_LEVELS)[number];- candidates: RosterEntry[];+ candidates: ModelsCatalogEntry[];};function fixtureDuplicateEffortCases(): DuplicateEffortCase[] {@@ -88,8 +89,8 @@ function fixtureDuplicateEffortCases(): DuplicateEffortCase[] {for (const effort of EFFORT_LEVELS) {const candidates = pool.filter((variant) => variant.effort === effort)- .map((variant) => rosterById.get(variant.wireId))- .filter((entry): entry is RosterEntry => entry !== undefined);+ .map((variant) => modelsCatalogById.get(variant.wireId))+ .filter((entry): entry is ModelsCatalogEntry => entry !== undefined);if (candidates.length > 1) cases.push({ choice, effort, candidates });}}@@ -98,11 +99,13 @@ function fixtureDuplicateEffortCases(): DuplicateEffortCase[] {function expectDuplicateEffortTieBreak(duplicate: DuplicateEffortCase,- permutations: RosterEntry[][]+ permutations: ModelsCatalogEntry[][]): void {const { choice, effort, candidates } = duplicate;const resolvedIds = new Set(- [roster, ...permutations].map((entries) => resolver.expand(choice.id, entries, { effort }))+ [modelsCatalog, ...permutations].map((entries) =>+ resolver.expand(choice.id, entries, { effort })+ ));expect(resolvedIds.size).toBe(1);@@ -125,13 +128,13 @@ function expectDuplicateEffortTieBreak(function exactNamePermutationFlaps(exactNames: string[],contexts: Array<{ label: string; ctx: ExpandContext }>,- permutations: RosterEntry[][]+ permutations: ModelsCatalogEntry[][]): string[] {const flaps: string[] = [];for (const selection of exactNames) {for (const { label, ctx } of contexts) {const resolvedIds = new Set(- [roster, ...permutations].map((entries) => resolver.expand(selection, entries, ctx))+ [modelsCatalog, ...permutations].map((entries) => resolver.expand(selection, entries, ctx)));if (resolvedIds.size > 1) {flaps.push(`${selection} @ ${label}: ${[...resolvedIds].sort().join(" | ")}`);@@ -139,21 +142,21 @@ function exactNamePermutationFlaps(}const [resolvedId] = resolvedIds;- expect(rosterById.has(resolvedId!)).toBe(true);+ expect(modelsCatalogById.has(resolvedId!)).toBe(true);}}return flaps;}-function deriveAmbiguousUidPrefixes(entries: RosterEntry[]): Array<{+function deriveAmbiguousUidPrefixes(entries: ModelsCatalogEntry[]): Array<{prefix: string;sourceUid: string;- matches: RosterEntry[];+ matches: ModelsCatalogEntry[];}> {const exactUids = new Set(entries.map((entry) => entry.wireId.toLowerCase()));const exactNames = new Set(entries.flatMap(exactNamesOf).map((name) => name.toLowerCase()));const usedRoots = new Set<string>();- const derived: Array<{ prefix: string; sourceUid: string; matches: RosterEntry[] }> = [];+ const derived: Array<{ prefix: string; sourceUid: string; matches: ModelsCatalogEntry[] }> = [];for (const source of entries) {const parts = source.wireId.split("-");@@ -176,7 +179,7 @@ function deriveAmbiguousUidPrefixes(entries: RosterEntry[]): Array<{return derived;}-function firstGroupForPrefix(entries: RosterEntry[], prefix: string): string {+function firstGroupForPrefix(entries: ModelsCatalogEntry[], prefix: string): string {const first = entries.find((entry) =>entry.wireId.toLowerCase().startsWith(`${prefix.toLowerCase()}-`));@@ -188,20 +191,20 @@ function expectPermutationInvariant(selection: string,passesThrough: boolean,ctx: ExpandContext,- permutations: RosterEntry[][]+ permutations: ModelsCatalogEntry[][]): void {- const expected = resolver.expand(selection, roster, ctx);+ const expected = resolver.expand(selection, modelsCatalog, ctx);if (passesThrough) expect(expected).toBe(selection);- else expect(rosterById.has(expected)).toBe(true);+ else expect(modelsCatalogById.has(expected)).toBe(true);for (const entries of permutations) {expect(resolver.expand(selection, entries, ctx)).toBe(expected);}}-const groups = groupEntries(roster);+const groups = groupEntries(modelsCatalog);const windowsByLabel = new Map<string, Set<number>>();-for (const entry of roster) {+for (const entry of modelsCatalog) {const label = groupKeyOf(entry);const windows = windowsByLabel.get(label) ?? new Set<number>();windows.add(entry.contextWindow ?? 0);@@ -209,11 +212,11 @@ for (const entry of roster) {}describe("DevinModelResolver collapse", () => {- test("collapses the captured roster to substantially fewer real, unique wire ids", () => {- expect(roster).toHaveLength(170);- expect(choices.length).toBeLessThan(roster.length / 2);+ test("collapses the captured dynamic models catalog to substantially fewer real, unique wire ids", () => {+ expect(modelsCatalog).toHaveLength(170);+ expect(choices.length).toBeLessThan(modelsCatalog.length / 2);expect(choices).toHaveLength(groups.size);- expect(choices.every((choice) => rosterById.has(choice.id))).toBe(true);+ expect(choices.every((choice) => modelsCatalogById.has(choice.id))).toBe(true);expect(new Set(choices.map((choice) => choice.id)).size).toBe(choices.length);});@@ -241,7 +244,7 @@ describe("DevinModelResolver collapse", () => {}});- test("makes every row's variants its exact group and partitions the roster", () => {+ test("makes every row's variants its exact group and partitions the dynamic models catalog", () => {const allVariantIds: string[] = [];for (const choice of choices) {@@ -254,7 +257,7 @@ describe("DevinModelResolver collapse", () => {allVariantIds.push(...actualIds);}- expect(allVariantIds.sort()).toEqual(roster.map((entry) => entry.wireId).sort());+ expect(allVariantIds.sort()).toEqual(modelsCatalog.map((entry) => entry.wireId).sort());});test("uses the provider-declared family default as the row id", () => {@@ -270,9 +273,11 @@ describe("DevinModelResolver collapse", () => {describe("DevinModelResolver expand", () => {test("reaches the 1M GLM effort variants", () => {- expect(resolver.expand("glm-5-2-1m", roster, { effort: "max" })).toBe("glm-5-2-max-1m");- expect(resolver.expand("glm-5-2-1m", roster, { effort: "none" })).toBe("glm-5-2-none-1m");- expect(resolver.expand("glm-5-2-1m", roster, {})).toBe("glm-5-2-1m");+ expect(resolver.expand("glm-5-2-1m", modelsCatalog, { effort: "max" })).toBe("glm-5-2-max-1m");+ expect(resolver.expand("glm-5-2-1m", modelsCatalog, { effort: "none" })).toBe(+ "glm-5-2-none-1m"+ );+ expect(resolver.expand("glm-5-2-1m", modelsCatalog, {})).toBe("glm-5-2-1m");});test("never auto-selects a premium variant when a plain variant exists", () => {@@ -280,7 +285,7 @@ describe("DevinModelResolver expand", () => {if (!choice.variants.some((variant) => variant.modifiers.length === 0)) continue;for (const effort of EFFORT_LEVELS) {- const resolved = resolver.expand(choice.id, roster, { effort });+ const resolved = resolver.expand(choice.id, modelsCatalog, { effort });const variant = choice.variants.find((candidate) => candidate.wireId === resolved);expect(variant).toBeDefined();if (resolved !== choice.id) expect(variant?.modifiers).toEqual([]);@@ -290,7 +295,7 @@ describe("DevinModelResolver expand", () => {test("round-trips every collapsed row id when effort is absent", () => {for (const choice of choices) {- expect(resolver.expand(choice.id, roster, {})).toBe(choice.id);+ expect(resolver.expand(choice.id, modelsCatalog, {})).toBe(choice.id);}});@@ -309,9 +314,9 @@ describe("DevinModelResolver expand", () => {const conflictingEffort = EFFORT_LEVELS.find((effort) => effort !== explicit.variant.effort);if (!conflictingEffort) throw new Error("No conflicting effort level is available");- expect(resolver.expand(explicit.variant.wireId, roster, { effort: conflictingEffort })).toBe(- explicit.variant.wireId- );+ expect(+ resolver.expand(explicit.variant.wireId, modelsCatalog, { effort: conflictingEffort })+ ).toBe(explicit.variant.wireId);});test("keeps a bare multi-window label inside one context group for every effort", () => {@@ -321,14 +326,14 @@ describe("DevinModelResolver expand", () => {expect(multiWindowLabels.length).toBeGreaterThan(0);for (const label of multiWindowLabels) {- const baselineId = resolver.expand(label, roster, {});- const baseline = rosterById.get(baselineId);+ const baselineId = resolver.expand(label, modelsCatalog, {});+ const baseline = modelsCatalogById.get(baselineId);if (!baseline) throw new Error(`Bare fixture label did not resolve: ${label}`);expect(groupKeyOf(baseline)).toBe(label);for (const effort of EFFORT_LEVELS) {- const resolvedId = resolver.expand(label, roster, { effort });- const resolved = rosterById.get(resolvedId);+ const resolvedId = resolver.expand(label, modelsCatalog, { effort });+ const resolved = modelsCatalogById.get(resolvedId);if (!resolved) throw new Error(`Bare fixture label did not resolve: ${label}`);expect(groupKeyOf(resolved)).toBe(label);expect(resolved.contextWindow).toBe(baseline.contextWindow);@@ -337,50 +342,50 @@ describe("DevinModelResolver expand", () => {});test("passes an unknown selection through unchanged", () => {- expect(resolver.expand("not-in-the-captured-roster", roster, { effort: "high" })).toBe(- "not-in-the-captured-roster"- );+ expect(+ resolver.expand("not-in-the-captured-models-catalog", modelsCatalog, { effort: "high" })+ ).toBe("not-in-the-captured-models-catalog");});- test("passes a selection through unchanged when the roster is empty", () => {+ test("passes a selection through unchanged when the dynamic models catalog is empty", () => {expect(resolver.expand("any-model", [], { effort: "high" })).toBe("any-model");});test("resolves vendor labels case-insensitively", () => {- const capitalisedLabel = roster+ const capitalisedLabel = modelsCatalog.map((entry) => nonEmpty(entry.groupLabel)).find((label) => label !== undefined && label !== label.toLowerCase());if (!capitalisedLabel) throw new Error("Fixture has no capitalised vendor label");- expect(resolver.expand(capitalisedLabel.toLowerCase(), roster, {})).toBe(- resolver.expand(capitalisedLabel, roster, {})+ expect(resolver.expand(capitalisedLabel.toLowerCase(), modelsCatalog, {})).toBe(+ resolver.expand(capitalisedLabel, modelsCatalog, {}));});test("prefers the vendor default, then lower cost, for duplicate effort levels", () => {const duplicateEffortCases = fixtureDuplicateEffortCases();- // The property is fixture-derived: a future roster with no duplicate+ // The property is fixture-derived: a future dynamic models catalog with no duplicate// declared levels has no tie-break case to exercise.if (duplicateEffortCases.length === 0) return;const permutations = Array.from({ length: 16 }, (_, index) =>- seededPermutation(roster, index + 1)+ seededPermutation(modelsCatalog, index + 1));for (const duplicate of duplicateEffortCases) {expectDuplicateEffortTieBreak(duplicate, permutations);}});- test("is invariant for every exact fixture name and effort across roster permutations", () => {+ test("is invariant for every exact fixture name and effort across dynamic models catalog permutations", () => {const permutations = Array.from({ length: 16 }, (_, index) =>- seededPermutation(roster, index + 1)+ seededPermutation(modelsCatalog, index + 1));expect(new Set(permutations.map((entries) => entries.map((entry) => entry.wireId).join("\0"))).size).toBe(permutations.length);- const exactNames = fixtureNameGroups(roster).map(({ selection }) => selection);+ const exactNames = fixtureNameGroups(modelsCatalog).map(({ selection }) => selection);expect(exactNames.length).toBeGreaterThan(0);const contexts: Array<{ label: string; ctx: ExpandContext }> = [@@ -393,7 +398,7 @@ describe("DevinModelResolver expand", () => {// (label/family, effort) pair and every wire id it alternated between.expect(flaps).toEqual([]);- const ambiguousPrefixes = deriveAmbiguousUidPrefixes(roster).slice(0, 2);+ const ambiguousPrefixes = deriveAmbiguousUidPrefixes(modelsCatalog).slice(0, 2);if (ambiguousPrefixes.length < 2) {throw new Error("Fixture has fewer than two independently derived ambiguous uid prefixes");}@@ -407,7 +412,7 @@ describe("DevinModelResolver expand", () => {const passThroughSelections = [...ambiguousPrefixes.map(({ prefix: selection }) => ({ selection, passesThrough: true })),- { selection: "not-in-the-captured-roster", passesThrough: true },+ { selection: "not-in-the-captured-models-catalog", passesThrough: true },];for (const choice of choices) {expectPermutationInvariant(choice.id, false, {}, permutations);@@ -420,26 +425,28 @@ describe("DevinModelResolver expand", () => {});test("passes independently fixture-derived multi-group uid prefixes through unchanged", () => {- const derived = deriveAmbiguousUidPrefixes(roster).slice(0, 2);+ const derived = deriveAmbiguousUidPrefixes(modelsCatalog).slice(0, 2);expect(derived).toHaveLength(2);expect(new Set(derived.map(({ sourceUid }) => sourceUid)).size).toBe(2);expect(new Set(derived.map(({ prefix }) => prefix.split("-")[0]!.toLowerCase())).size).toBe(2);- const exactUids = new Set(roster.map((entry) => entry.wireId.toLowerCase()));- const exactNames = new Set(roster.flatMap(exactNamesOf).map((name) => name.toLowerCase()));+ const exactUids = new Set(modelsCatalog.map((entry) => entry.wireId.toLowerCase()));+ const exactNames = new Set(+ modelsCatalog.flatMap(exactNamesOf).map((name) => name.toLowerCase())+ );for (const { prefix, sourceUid, matches } of derived) {expect(sourceUid.toLowerCase().startsWith(`${prefix.toLowerCase()}-`)).toBe(true);expect(new Set(matches.map(groupKey)).size).toBeGreaterThan(1);expect(exactUids.has(prefix.toLowerCase())).toBe(false);expect(exactNames.has(prefix.toLowerCase())).toBe(false);- expect(resolver.expand(prefix, roster, {})).toBe(prefix);- expect(resolver.expand(prefix, roster, { effort: "high" })).toBe(prefix);+ expect(resolver.expand(prefix, modelsCatalog, {})).toBe(prefix);+ expect(resolver.expand(prefix, modelsCatalog, { effort: "high" })).toBe(prefix);}});test("resolves fixture-derived exact names spanning context groups via the vendor default", () => {- const exactUids = new Set(roster.map((entry) => entry.wireId.toLowerCase()));- const multiWindowNames = fixtureNameGroups(roster).filter(+ const exactUids = new Set(modelsCatalog.map((entry) => entry.wireId.toLowerCase()));+ const multiWindowNames = fixtureNameGroups(modelsCatalog).filter(({ selection, entries }) =>!exactUids.has(selection.toLowerCase()) &&new Set(entries.map((entry) => entry.contextWindow)).size > 1@@ -452,8 +459,8 @@ describe("DevinModelResolver expand", () => {expect(defaults).toHaveLength(1);for (const ctx of [{}, { effort: "high" as const }]) {- const resolvedId = resolver.expand(selection, roster, ctx);- const resolved = rosterById.get(resolvedId);+ const resolvedId = resolver.expand(selection, modelsCatalog, ctx);+ const resolved = modelsCatalogById.get(resolvedId);if (!resolved) throw new Error(`Fixture name did not resolve to a real uid: ${selection}`);expect(entries.map((entry) => entry.wireId)).toContain(resolvedId);expect(groupKey(resolved)).toBe(groupKey(defaults[0]!));@@ -462,8 +469,8 @@ describe("DevinModelResolver expand", () => {});test("keeps exact-name effort resolution inside its selected context group", () => {- const exactUids = new Set(roster.map((entry) => entry.wireId.toLowerCase()));- const multiWindowNames = fixtureNameGroups(roster).filter(+ const exactUids = new Set(modelsCatalog.map((entry) => entry.wireId.toLowerCase()));+ const multiWindowNames = fixtureNameGroups(modelsCatalog).filter(({ selection, entries }) =>!exactUids.has(selection.toLowerCase()) &&new Set(entries.map((entry) => entry.contextWindow)).size > 1@@ -471,8 +478,8 @@ describe("DevinModelResolver expand", () => {expect(multiWindowNames.length).toBeGreaterThan(0);for (const { selection, entries } of multiWindowNames) {- const baselineId = resolver.expand(selection, roster, {});- const baseline = rosterById.get(baselineId);+ const baselineId = resolver.expand(selection, modelsCatalog, {});+ const baseline = modelsCatalogById.get(baselineId);if (!baseline) throw new Error(`Fixture name did not resolve to a real uid: ${selection}`);const selectedGroup = groups.get(groupKey(baseline));@@ -486,8 +493,8 @@ describe("DevinModelResolver expand", () => {expect(siblingWindows.size).toBeGreaterThan(0);for (const effort of EFFORT_LEVELS) {- const resolvedId = resolver.expand(selection, roster, { effort });- const resolved = rosterById.get(resolvedId);+ const resolvedId = resolver.expand(selection, modelsCatalog, { effort });+ const resolved = modelsCatalogById.get(resolvedId);if (!resolved) throw new Error(`Fixture name did not resolve to a real uid: ${selection}`);expect(selectedIds.has(resolvedId)).toBe(true);expect(resolved.contextWindow).toBe(baseline.contextWindow);diff --git a/packages/cli/src/providers/model-resolvers/devin.ts b/packages/cli/src/providers/model-resolvers/devin.tsindex b53ddbeb..0ddc24a2 100644--- a/packages/cli/src/providers/model-resolvers/devin.ts+++ b/packages/cli/src/providers/model-resolvers/devin.ts@@ -15,7 +15,7 @@** ## The three rules, and the measurements behind them*- * Measured against a real captured roster (see the session's `evidence.md`):+ * Measured against a real captured dynamic models catalog (see the session's `evidence.md`):** 1. **Group by `model_family_label` x `contextWindow`.** Not by uid spelling:* `-1m` appears on 7 uids while 97 natively-1M uids carry no suffix at all,@@ -52,13 +52,13 @@ import type { DevinModelConfig } from "../devin/devin-models.js";import type {ExpandContext,ModelChoice,+ ModelsCatalogEntry,+ ModelsCatalogVariant,ProviderModelResolver,- RosterEntry,- RosterVariant,} from "./types.js";import { groupKeyOf, nonEmpty } from "./types.js";-/** Axis keys Devin uses for the reasoning knob. Both appear in the live roster. */+/** Axis keys Devin uses for the reasoning knob. Both appear in the dynamic models catalog. */const EFFORT_AXIS_KEYS = ["effort", "reasoning effort"];/** Axis key marking a speed premium. Its flag covers `-fast` AND `-priority`. */@@ -110,7 +110,7 @@ export function devinHasSpeedPremium(config: DevinModelConfig): boolean {* Does this string NAME a reasoning tier?** Deliberately about the SPELLING of what the user asked for, not about what- * the roster says the uid is — it answers a question about intent. Typing+ * the dynamic models catalog says the uid is — it answers a question about intent. Typing* `claude-opus-5-xhigh` names a tier and must be honoured verbatim even at* effort `low`. Typing `glm-5-2-1m` names a *context* variant and leaves the* tier open, so it resolves against the effort of the turn — which is exactly@@ -135,8 +135,8 @@ function modifiersOf(config: DevinModelConfig): string[] {return devinHasSpeedPremium(config) ? ["premium"] : [];}-/** Normalise a Devin roster row into the seam's provider-neutral shape. */-export function devinRosterEntry(config: DevinModelConfig): RosterEntry {+/** Normalise a Devin dynamic models catalog row into the seam's provider-neutral shape. */+export function devinModelsCatalogEntry(config: DevinModelConfig): ModelsCatalogEntry {return {wireId: config.uid,displayName: config.displayName,@@ -152,8 +152,8 @@ export function devinRosterEntry(config: DevinModelConfig): RosterEntry {};}-/** Rebuild the Devin-shaped view a rule needs from a neutral roster entry. */-function asConfig(entry: RosterEntry): DevinModelConfig {+/** Rebuild the Devin-shaped view a rule needs from a neutral `ModelsCatalogEntry`. */+function asConfig(entry: ModelsCatalogEntry): DevinModelConfig {return {uid: entry.wireId,displayName: entry.displayName ?? entry.wireId,@@ -172,10 +172,10 @@ function asConfig(entry: RosterEntry): DevinModelConfig {/*** Order candidates for the default-of-last-resort: cheapest band, then cheapest* multiplier, then shortest id. Every tie is broken, so the answer is stable- * across roster reorderings — a picker whose default moves between runs reads- * as a bug even when both answers are defensible.+ * across reorderings of the dynamic models catalog — a picker whose default+ * moves between runs reads as a bug even when both answers are defensible.*/-function byCost(a: RosterEntry, b: RosterEntry): number {+function byCost(a: ModelsCatalogEntry, b: ModelsCatalogEntry): number {const tier = (a.costTier ?? Number.MAX_SAFE_INTEGER) - (b.costTier ?? Number.MAX_SAFE_INTEGER);if (tier !== 0) return tier;const cost =@@ -204,7 +204,7 @@ function byCost(a: RosterEntry, b: RosterEntry): number {* whose every member is premium, must still resolve to something — dropping it* from the picker would be worse than an imperfect representative.*/-function defaultOf(group: RosterEntry[]): RosterEntry {+function defaultOf(group: ModelsCatalogEntry[]): ModelsCatalogEntry {const declared = group.find((entry) => entry.isFamilyDefault);if (declared) return declared;const plain = group.filter((entry) => !devinHasSpeedPremium(asConfig(entry)));@@ -214,7 +214,7 @@ function defaultOf(group: RosterEntry[]): RosterEntry {}/** Group key: the vendor's row label, then the window — the two facts D1 rests on. */-function keyOf(entry: RosterEntry): string {+function keyOf(entry: ModelsCatalogEntry): string {return `${groupKeyOf(entry)} ${entry.contextWindow ?? 0}`;}@@ -228,9 +228,9 @@ function formatWindow(tokens: number): string {export class DevinModelResolver implements ProviderModelResolver {readonly provider = "devin";- collapse(roster: RosterEntry[]): ModelChoice[] {- const groups = new Map<string, RosterEntry[]>();- for (const entry of roster) {+ collapse(modelsCatalog: ModelsCatalogEntry[]): ModelChoice[] {+ const groups = new Map<string, ModelsCatalogEntry[]>();+ for (const entry of modelsCatalog) {const key = keyOf(entry);const bucket = groups.get(key);if (bucket) bucket.push(entry);@@ -241,7 +241,7 @@ export class DevinModelResolver implements ProviderModelResolver {// be distinguishable; a label with only one row does not, and adding it// there would be noise on 36 of 39 families.const windowsPerLabel = new Map<string, Set<number>>();- for (const entry of roster) {+ for (const entry of modelsCatalog) {const label = groupKeyOf(entry);const seen = windowsPerLabel.get(label) ?? new Set<number>();seen.add(entry.contextWindow ?? 0);@@ -255,7 +255,7 @@ export class DevinModelResolver implements ProviderModelResolver {const ambiguous = (windowsPerLabel.get(label)?.size ?? 1) > 1;const window = chosen.contextWindow ?? 0;- const variants: RosterVariant[] = group.map((entry) => {+ const variants: ModelsCatalogVariant[] = group.map((entry) => {const config = asConfig(entry);return {wireId: entry.wireId,@@ -294,12 +294,12 @@ export class DevinModelResolver implements ProviderModelResolver {* The served-set-aware error rewrite turns that into a message naming what* IS served; guessing here would hide it.*/- expand(selection: string, roster: RosterEntry[], ctx: ExpandContext): string {+ expand(selection: string, modelsCatalog: ModelsCatalogEntry[], ctx: ExpandContext): string {const requested = selection.trim();- if (!requested || roster.length === 0) return requested || selection;+ if (!requested || modelsCatalog.length === 0) return requested || selection;- const groups = new Map<string, RosterEntry[]>();- for (const entry of roster) {+ const groups = new Map<string, ModelsCatalogEntry[]>();+ for (const entry of modelsCatalog) {const key = keyOf(entry);const bucket = groups.get(key);if (bucket) bucket.push(entry);@@ -307,7 +307,7 @@ export class DevinModelResolver implements ProviderModelResolver {}const lower = requested.toLowerCase();- const exact = roster.find((entry) => entry.wireId.toLowerCase() === lower);+ const exact = modelsCatalog.find((entry) => entry.wireId.toLowerCase() === lower);if (exact) {if (!ctx.effort) return exact.wireId;const group = groups.get(keyOf(exact))!;@@ -322,7 +322,7 @@ export class DevinModelResolver implements ProviderModelResolver {// 2. An exact vendor label or family id. Unambiguous by construction — it// names one model, even when that model spans two context groups.- const named = roster.filter((entry) => {+ const named = modelsCatalog.filter((entry) => {const label = nonEmpty(entry.groupLabel)?.toLowerCase();const family = nonEmpty(entry.family)?.toLowerCase();return label === lower || family === lower;@@ -337,14 +337,14 @@ export class DevinModelResolver implements ProviderModelResolver {const pool =named.length > 0? named- : roster.filter((entry) => entry.wireId.toLowerCase().startsWith(`${lower}-`));+ : modelsCatalog.filter((entry) => entry.wireId.toLowerCase().startsWith(`${lower}-`));if (pool.length === 0) return requested;if (named.length === 0 && new Set(pool.map(keyOf)).size > 1) return requested;// One label, possibly several context groups (GLM-5.2 is 200K and 1M): take// the group the vendor itself defaults to, else the cheapest. Sorted, not// first-match — `byCost` breaks every tie, so the answer cannot depend on- // the order the roster happened to arrive in.+ // the order the dynamic models catalog happened to arrive in.const chosen =pool.filter((entry) => entry.isFamilyDefault).sort(byCost)[0] ?? pool.slice().sort(byCost)[0];return pickForEffort(groups.get(keyOf(chosen!))!, ctx.effort).wireId;@@ -358,7 +358,10 @@ export class DevinModelResolver implements ProviderModelResolver {* model is the worse failure. Premium variants are excluded entirely — they* cost ~2x and are reachable only by naming their uid.*/-function pickForEffort(group: RosterEntry[], effort: EffortLevel | undefined): RosterEntry {+function pickForEffort(+ group: ModelsCatalogEntry[],+ effort: EffortLevel | undefined+): ModelsCatalogEntry {if (!effort) return defaultOf(group);const plain = group.filter((entry) => !devinHasSpeedPremium(asConfig(entry)));@@ -366,7 +369,9 @@ function pickForEffort(group: RosterEntry[], effort: EffortLevel | undefined): Rconst tiered = pool.map((entry) => ({ entry, effort: devinEffortOf(asConfig(entry)) }))- .filter((row): row is { entry: RosterEntry; effort: EffortLevel } => row.effort !== undefined);+ .filter(+ (row): row is { entry: ModelsCatalogEntry; effort: EffortLevel } => row.effort !== undefined+ );if (tiered.length === 0) return defaultOf(group);const target = EFFORT_LEVELS.indexOf(effort);@@ -383,7 +388,7 @@ function pickForEffort(group: RosterEntry[], effort: EffortLevel | undefined): R// Real case — Claude Opus 4.6 declares `Effort=High` on both its plain and// its `-thinking` uid, which differ only in `Thinking` and in price (×6 vs// ×8). Falling back to first-encountered made the answer depend on the- // order the roster arrived in, across 21 (label, effort) pairs and three+ // order the dynamic models catalog arrived in, across 21 (label, effort) pairs and three// families, for swings up to 50%.const defaultA = a.entry.isFamilyDefault ? 0 : 1;const defaultB = b.entry.isFamilyDefault ? 0 : 1;diff --git a/packages/cli/src/providers/model-resolvers/registry.test.ts b/packages/cli/src/providers/model-resolvers/registry.test.tsindex 3121cd62..efc904d4 100644--- a/packages/cli/src/providers/model-resolvers/registry.test.ts+++ b/packages/cli/src/providers/model-resolvers/registry.test.ts@@ -1,6 +1,6 @@import { describe, expect, test } from "bun:test";-import { collapseRoster, expandSelection, getModelResolver } from "./registry.js";-import type { RosterEntry } from "./types.js";+import { collapseModelsCatalog, expandSelection, getModelResolver } from "./registry.js";+import type { ModelsCatalogEntry } from "./types.js";describe("model resolver registry", () => {test("registers only providers that opt in", () => {@@ -11,13 +11,13 @@ describe("model resolver registry", () => {});test("uses identity collapse for an unregistered provider", () => {- const roster: RosterEntry[] = [+ const modelsCatalog: ModelsCatalogEntry[] = [{ wireId: "first", displayName: "First", contextWindow: 100_000 },{ wireId: "second", displayName: "Second", costFactor: 2 },{ wireId: "third" },];- const choices = collapseRoster("openrouter", roster);+ const choices = collapseModelsCatalog("openrouter", modelsCatalog);expect(choices.map((choice) => choice.id)).toEqual(["first", "second", "third"]);expect(choices.map((choice) => choice.displayName)).toEqual(["First", "Second", "third"]);expect(choices.map((choice) => choice.variants)).toEqual([@@ -28,9 +28,9 @@ describe("model resolver registry", () => {});test("uses identity expand for an unregistered provider", () => {- const roster: RosterEntry[] = [{ wireId: "served-model" }];- expect(expandSelection("openrouter", "unlisted-selection", roster, { effort: "max" })).toBe(- "unlisted-selection"- );+ const modelsCatalog: ModelsCatalogEntry[] = [{ wireId: "served-model" }];+ expect(+ expandSelection("openrouter", "unlisted-selection", modelsCatalog, { effort: "max" })+ ).toBe("unlisted-selection");});});diff --git a/packages/cli/src/providers/model-resolvers/registry.ts b/packages/cli/src/providers/model-resolvers/registry.tsindex 7c1a2dda..1b0c1db1 100644--- a/packages/cli/src/providers/model-resolvers/registry.ts+++ b/packages/cli/src/providers/model-resolvers/registry.ts@@ -1,17 +1,22 @@/**- * Which providers collapse their roster, and which are 1:1.+ * Which providers collapse their dynamic models catalog, and which are 1:1.** Returning `undefined` is the DEFAULT and means identity — one picker row per- * roster entry, `expand` returning the selection unchanged. That is correct for+ * `ModelsCatalogEntry`, `expand` returning the selection unchanged. That is correct for* every provider whose live endpoint already lists exactly what a human would- * choose (Ollama's tags, LM Studio's models, Kimi Coding's plan roster), so+ * choose (Ollama's tags, LM Studio's models, Kimi Coding's dynamic models catalog), so* adding this seam changes nothing for them.** Only a provider that encodes knobs INTO its model ids needs an entry here.*/import { DevinModelResolver } from "./devin.js";-import type { ExpandContext, ModelChoice, ProviderModelResolver, RosterEntry } from "./types.js";+import type {+ ExpandContext,+ ModelChoice,+ ModelsCatalogEntry,+ ProviderModelResolver,+} from "./types.js";/*** Antigravity is a deliberate NON-member for now. Its resolution rule grew a@@ -31,15 +36,18 @@ export function getModelResolver(provider: string): ProviderModelResolver | unde}/**- * Roster -> picker rows, falling back to 1:1 for providers with no resolver.+ * Dynamic models catalog -> picker rows, falling back to 1:1 for providers with no resolver.** Callers should use this rather than branching on `getModelResolver`, so the* identity path is written once instead of at every call site.*/-export function collapseRoster(provider: string, roster: RosterEntry[]): ModelChoice[] {+export function collapseModelsCatalog(+ provider: string,+ modelsCatalog: ModelsCatalogEntry[]+): ModelChoice[] {const resolver = getModelResolver(provider);- if (resolver) return resolver.collapse(roster);- return roster.map((entry) => ({+ if (resolver) return resolver.collapse(modelsCatalog);+ return modelsCatalog.map((entry) => ({id: entry.wireId,displayName: entry.displayName ?? entry.wireId,contextWindow: entry.contextWindow,@@ -59,8 +67,8 @@ export function collapseRoster(provider: string, roster: RosterEntry[]): ModelChexport function expandSelection(provider: string,selection: string,- roster: RosterEntry[],+ modelsCatalog: ModelsCatalogEntry[],ctx: ExpandContext = {}): string {- return getModelResolver(provider)?.expand(selection, roster, ctx) ?? selection;+ return getModelResolver(provider)?.expand(selection, modelsCatalog, ctx) ?? selection;}diff --git a/packages/cli/src/providers/model-resolvers/types.ts b/packages/cli/src/providers/model-resolvers/types.tsindex 8fad0ba5..16c8869e 100644--- a/packages/cli/src/providers/model-resolvers/types.ts+++ b/packages/cli/src/providers/model-resolvers/types.ts@@ -1,10 +1,10 @@/**- * ProviderModelResolver — per-provider collapse/expand between a raw roster and- * the rows a human picks.+ * ProviderModelResolver — per-provider collapse/expand between a raw dynamic+ * models catalog and the rows a human picks.** ## Why this exists*- * A provider's live roster is not always the list a human wants to choose from.+ * A provider's dynamic models catalog is not always the list a human wants to choose from.* Devin is the motivating case: it encodes reasoning tier, speed modifier, and* context window in the model *uid* rather than in request parameters, so its* 170 served uids are really ~33 models with knobs on them. Listing all 170 is@@ -31,7 +31,7 @@* ## Default is identity** `getModelResolver()` returns undefined for every provider that has not opted- * in, and callers treat that as 1 row per roster entry with `expand` returning+ * in, and callers treat that as 1 row per `ModelsCatalogEntry` with `expand` returning* the selection unchanged. Adding this seam therefore changes nothing for a* provider that does not implement it.*/@@ -39,7 +39,7 @@import type { EffortLevel } from "../../adapters/base-api-format.js";/**- * A knob the provider DECLARES for one roster entry, with the value that entry+ * A knob the provider DECLARES for one `ModelsCatalogEntry`, with the value that entry* sits at.** The point of carrying this is that a provider which states what each variant@@ -48,7 +48,7 @@ import type { EffortLevel } from "../../adapters/base-api-format.js";* than parsing `-xhigh` off the end of the id, and it survives a vendor adding* a level whose spelling we never anticipated.*/-export interface RosterAxis {+export interface ModelsCatalogAxis {/** The provider's own name for the knob (`Effort`, `1M Context`, `Fast Mode`). */key: string;/** The value THIS entry sits at (`XHigh`, `Max`, `No Thinking`). */@@ -57,7 +57,7 @@ export interface RosterAxis {enabled: boolean;}-/** A time-boxed or plan-included price on one roster entry. */+/** A time-boxed or plan-included price on one `ModelsCatalogEntry`. */export interface ModelOffer {/*** - `promo` — a temporary price the vendor is advertising. ALWAYS carries an@@ -71,14 +71,14 @@ export interface ModelOffer {}/**- * One roster entry, exactly as the provider reports it.+ * One entry of a dynamic models catalog, exactly as the provider reports it.** Every field beyond `wireId` is optional because providers report wildly* different amounts: Ollama's `/api/tags` gives a name and little else, while* Devin publishes a label, a cost multiplier, a default flag, and the list of* knobs each model exposes.*/-export interface RosterEntry {+export interface ModelsCatalogEntry {/** Exactly what goes in the request's model field. The only required fact. */wireId: string;displayName?: string;@@ -100,12 +100,12 @@ export interface RosterEntry {/** The vendor marks this entry as recommended. */isRecommended?: boolean;/** The knobs the vendor says this entry exposes, in the vendor's own words. */- axes?: RosterAxis[];+ axes?: ModelsCatalogAxis[];offer?: ModelOffer;}/** One wire id a {@link ModelChoice} can resolve to. */-export interface RosterVariant {+export interface ModelsCatalogVariant {wireId: string;/** The reasoning level this variant represents, when it maps to one. */effort?: EffortLevel;@@ -130,7 +130,7 @@ export interface ModelChoice {displayName: string;contextWindow?: number;/** Every wire id this choice can expand into, including `id` itself. */- variants: RosterVariant[];+ variants: ModelsCatalogVariant[];/** Cost at the default variant. */costFactor?: number;offer?: ModelOffer;@@ -149,13 +149,13 @@ export interface ProviderModelResolver {readonly provider: string;/**- * Raw roster -> the rows a human chooses from.+ * Raw dynamic models catalog -> the rows a human chooses from.** Must be pure and total: any entry it cannot interpret still has to reach* the user somehow, because a model silently missing from the picker is* indistinguishable from a model the subscription does not serve.*/- collapse(roster: RosterEntry[]): ModelChoice[];+ collapse(modelsCatalog: ModelsCatalogEntry[]): ModelChoice[];/*** A chosen id + runtime signals -> the wire id to send.@@ -164,23 +164,23 @@ export interface ProviderModelResolver {* backend can answer for itself. Guessing here would replace an actionable* "that model is not served, here is what is" with a silent substitution.*/- expand(selection: string, roster: RosterEntry[], ctx: ExpandContext): string;+ expand(selection: string, modelsCatalog: ModelsCatalogEntry[], ctx: ExpandContext): string;}/*** A non-empty string, or undefined.** Providers report absent fields as `""` as often as they omit them, and `??`- * does not catch the empty string — a real bug found while grouping the Devin- * roster, where six entries with `family: ""` collapsed into one bogus group.+ * does not catch the empty string — a real bug found while grouping Devin's+ * dynamic models catalog, where six entries with `family: ""` collapsed into one bogus group.*/export function nonEmpty(value: string | undefined): string | undefined {const trimmed = value?.trim();return trimmed ? trimmed : undefined;}-/** The label to group a roster entry under: vendor label, else family, else its own id. */-export function groupKeyOf(entry: RosterEntry): string {+/** The label to group an entry under: vendor label, else family, else its own id. */+export function groupKeyOf(entry: ModelsCatalogEntry): string {return nonEmpty(entry.groupLabel) ?? nonEmpty(entry.family) ?? entry.wireId;}diff --git a/packages/cli/src/providers/picker-alias-extra.ts b/packages/cli/src/providers/picker-alias-extra.tsindex 5dd29484..eb0bf20e 100644--- a/packages/cli/src/providers/picker-alias-extra.ts+++ b/packages/cli/src/providers/picker-alias-extra.ts@@ -17,7 +17,7 @@*/export const PROVIDER_FILTER_ALIAS_EXTRA: Record<string, string> = {gem: "google",- // Legacy picker value for OpenCode Zen; the roster now uses the definition+ // Legacy picker value for OpenCode Zen; the provider list now uses the definition// name, and both still resolve downstream.zen: "opencode-zen",};diff --git a/packages/cli/src/providers/predefined-catalog.ts b/packages/cli/src/providers/predefined-catalog.tsindex 98f37cbd..e7980a10 100644--- a/packages/cli/src/providers/predefined-catalog.ts+++ b/packages/cli/src/providers/predefined-catalog.ts@@ -314,7 +314,7 @@ export const PREDEFINED_ENDPOINTS: readonly PredefinedEndpoint[] = [// service from the builtin `kimi` provider — `kimi@` / `moon@` /// `moonshot@` are that builtin's shortcuts and reach the international// Kimi endpoint on a different credential. Both can be configured at once;- // they do not share keys, rosters, or billing.+ // they do not share keys, model lists, or billing.name: "moonshot-cn",displayName: "Moonshot AI (China)",baseUrl: "https://api.moonshot.cn",diff --git a/packages/cli/src/providers/predefined-endpoint-schema.ts b/packages/cli/src/providers/predefined-endpoint-schema.tsindex 903d2e63..d67de3e2 100644--- a/packages/cli/src/providers/predefined-endpoint-schema.ts+++ b/packages/cli/src/providers/predefined-endpoint-schema.ts@@ -12,7 +12,7 @@** 1. **No model data, ever.** There is no `models`, `contextWindow`,* `maxOutputTokens`, `pricing`, `capabilities` or `modelDiscovery` field.- * A shipped roster is exactly the hardcoded model data this project forbids:+ * A shipped model list is exactly the hardcoded model data this project forbids:* it rots the moment a vendor adds a model, and the failure shape is* claudish refusing a model that actually works. Model metadata comes from* models-index or is absent. The schema is `.strict()`, so a future@@ -53,7 +53,7 @@ import { z } from "zod";* - `tier: "probe"` — a POST to the configured chat path with a deliberately* invalid key was answered by the vendor's auth layer, and differed from a* deliberately bogus sibling path. A `GET /v1/models` result is NOT evidence- * and must never be recorded here: Alibaba's `coding-intl` roster endpoint+ * and must never be recorded here: Alibaba's `coding-intl` model-list endpoint* answers a bogus key, and no key at all, with the full list.* - `verdict: "auth-realm"` — 401/403; credentials are checked at this path.* - `verdict: "model-gate"` — the route resolved and rejected the fake model.diff --git a/packages/cli/src/providers/predefined-endpoints.ts b/packages/cli/src/providers/predefined-endpoints.tsindex dfa68b45..00093e3d 100644--- a/packages/cli/src/providers/predefined-endpoints.ts+++ b/packages/cli/src/providers/predefined-endpoints.ts@@ -38,7 +38,7 @@* ── Why registration itself is what gets gated ──────────────────────────────** There is no "registered but hidden" state: `buildProviderDefinition` sets- * `shortcuts: [name]`, and both the picker roster (`isPickableProvider`) and+ * `shortcuts: [name]`, and both the picker's provider list (`isPickableProvider`) and* the `@prefix` alias table are DERIVED from that. The picker's own list is* credential-filtered, but `getProviderFilterAliases()` is not — so registering* every bundled vendor unconditionally would pollute the `@prefix` namespacediff --git a/packages/cli/src/providers/probe-live.ts b/packages/cli/src/providers/probe-live.tsindex 11fde5e4..6791394d 100644--- a/packages/cli/src/providers/probe-live.ts+++ b/packages/cli/src/providers/probe-live.ts@@ -385,7 +385,7 @@ export function classifyHttpError(status: number, body: string, latencyMs: numbe// Nor is it an auth failure when the provider handed back a link to act on.// Measured: Zen Go answers a model the account has not opted into with// `403 RegionError … requires explicit opt in: <url>` — the credential is- // fine, the model IS in its live roster, and the fix is a click. Reporting+ // fine, the model IS in its dynamic models catalog, and the fix is a click. Reporting// `auth-failed` there points the user at a working key. `error` keeps the// failure semantics (it is in `isFailureState`) without the false cause.if (hasActionableLink(body)) {diff --git a/packages/cli/src/providers/provider-definitions.ts b/packages/cli/src/providers/provider-definitions.tsindex 60562daf..d288d0c1 100644--- a/packages/cli/src/providers/provider-definitions.ts+++ b/packages/cli/src/providers/provider-definitions.ts@@ -177,7 +177,7 @@ export interface ProviderDefinition {/*** Opt in to live, per-subscription model discovery. When set, claudish calls* the provider's own authenticated model-listing endpoint to learn the real- * roster and per-model context windows for THIS user's plan, instead of+ * dynamic models catalog and per-model context windows for THIS user's plan, instead of* trusting a static list. Required for subscription endpoints whose context* window varies by tier (see providers/model-discovery.ts).*/@@ -319,8 +319,8 @@ export const BUILTIN_PROVIDERS: ProviderDefinition[] = [{ prefix: "antigravity/", stripPrefix: true },],// Not a GET — an OAuth POST to v1internal:fetchAvailableModels, so `path`- // is ignored. Declared so the picker prefers the LIVE per-subscription- // roster and, more importantly, its per-model `maxTokens`: the backend and+ // is ignored. Declared so the picker prefers the per-subscription dynamic+ // models catalog and, more importantly, its per-model `maxTokens`: the backend and// the shared catalog disagree by 4x on claude-sonnet-4-6 (250K vs 1M).modelDiscovery: { path: "", format: "antigravity" },isDirectApi: true,@@ -534,9 +534,9 @@ export const BUILTIN_PROVIDERS: ProviderDefinition[] = [// routing chain instead, where this provider sits FIRST — subscription// before metered, so a user holding both credentials is never silently// billed per token for a model their subscription already covers.- // The served roster is ACCOUNT-SCOPED and drifts, so it is discovered, never+ // The dynamic models catalog is ACCOUNT-SCOPED and drifts, so it is discovered, never// pinned. `/v1/models` is genuinely authenticated here (401 without a token,- // unlike Alibaba's coding-intl roster where a 200 proves nothing), and it+ // unlike Alibaba's coding-intl model list where a 200 proves nothing), and it// answers the standard OpenAI `{object, data:[{id}]}` shape. Discovery falls// back to the credential authority when `apiKeyEnvVar` is empty, which also// supplies the mandatory client-version headers.diff --git a/packages/cli/src/providers/routing-rules.test.ts b/packages/cli/src/providers/routing-rules.test.tsindex 6fc2536e..d696f029 100644--- a/packages/cli/src/providers/routing-rules.test.ts+++ b/packages/cli/src/providers/routing-rules.test.ts@@ -1157,7 +1157,7 @@ function routingCatalogEntry(modelId: string, providers: string[]): SlimModelEnt};}-function rosterProvider(name: string): ProviderDefinition {+function modelsCatalogProvider(name: string): ProviderDefinition {return {name,displayName: name,@@ -1187,16 +1187,16 @@ describe("route() model-availability filtering", () => {let cachePath = "";let cleanupCache: (() => void) | undefined;let credentialedProviders = new Set<string>();- let rosters = new Map<string, string[]>();+ let modelsCatalogs = new Map<string, string[]>();let fetchCalls: string[] = [];function allowCredentials(...providers: string[]): void {credentialedProviders = new Set(providers);}- function registerRoster(name: string, ...ids: string[]): void {- registerRuntimeProvider(rosterProvider(name));- rosters.set(name, ids);+ function registerModelsCatalog(name: string, ...ids: string[]): void {+ registerRuntimeProvider(modelsCatalogProvider(name));+ modelsCatalogs.set(name, ids);}beforeEach(() => {@@ -1206,7 +1206,7 @@ describe("route() model-availability filtering", () => {clearRuntimeRegistry();credentialedProviders = new Set();- rosters = new Map();+ modelsCatalogs = new Map();fetchCalls = [];const tempCatalog = makeTempCatalog({ modelId: AVAILABILITY_MODEL });@@ -1222,8 +1222,8 @@ describe("route() model-availability filtering", () => {typeof input === "string" ? input : input instanceof URL ? input.href : input.url;const provider = new URL(rawUrl).hostname.replace(/\.invalid$/, "");fetchCalls.push(provider);- const ids = rosters.get(provider);- if (!ids) throw new Error(`Unexpected roster request for ${provider}`);+ const ids = modelsCatalogs.get(provider);+ if (!ids) throw new Error(`Unexpected discovery request for ${provider}`);return new Response(JSON.stringify({ data: ids.map((id) => ({ id })) }), {status: 200,headers: { "Content-Type": "application/json" },@@ -1244,7 +1244,7 @@ describe("route() model-availability filtering", () => {test("removes not-served candidates and preserves the surviving order", async () => {const denied = "availability-denied";- registerRoster(denied, "some-other-model");+ registerModelsCatalog(denied, "some-other-model");allowCredentials(denied, "openai", "openrouter");_setCatalogEntriesForTest([routingCatalogEntry(AVAILABILITY_MODEL, ["openai", "openrouter"])]);@@ -1302,8 +1302,8 @@ describe("route() model-availability filtering", () => {test("returns no-route naming every checked provider when all are not-served", async () => {const first = "availability-denied-first";const second = "availability-denied-second";- registerRoster(first, "other-first");- registerRoster(second, "other-second");+ registerModelsCatalog(first, "other-first");+ registerModelsCatalog(second, "other-second");allowCredentials(first, second);const plan = await route(@@ -1322,7 +1322,7 @@ describe("route() model-availability filtering", () => {test("explicit not-served spec returns no-route without silent substitution", async () => {const denied = "availability-explicit-denied";- registerRoster(denied, "some-other-model");+ registerModelsCatalog(denied, "some-other-model");allowCredentials(denied, "openai");const plan = await route(@@ -1341,7 +1341,7 @@ describe("route() model-availability filtering", () => {test("explicit serves spec returns ok with the named provider", async () => {const serving = "availability-explicit-serving";- registerRoster(serving, AVAILABILITY_MODEL);+ registerModelsCatalog(serving, AVAILABILITY_MODEL);allowCredentials(serving);const plan = await route(`${serving}@${AVAILABILITY_MODEL}`, {}, undefined, cachePath);@@ -1355,8 +1355,8 @@ describe("route() model-availability filtering", () => {test("checks availability only after filtering providers without credentials", async () => {const noCredential = "availability-no-credential";const credentialed = "availability-credentialed";- registerRoster(noCredential, AVAILABILITY_MODEL);- registerRoster(credentialed, AVAILABILITY_MODEL);+ registerModelsCatalog(noCredential, AVAILABILITY_MODEL);+ registerModelsCatalog(credentialed, AVAILABILITY_MODEL);allowCredentials(credentialed);const plan = await route(@@ -1370,7 +1370,7 @@ describe("route() model-availability filtering", () => {if (plan.kind !== "ok") return;expect(plan.primary.provider).toBe(credentialed);// A provider the user cannot authenticate to must never incur the- // guaranteed-failing roster round-trip.+ // guaranteed-failing discovery round-trip.expect(fetchCalls).toEqual([credentialed]);expect(fetchCalls).not.toContain(noCredential);});diff --git a/packages/cli/src/providers/routing-rules.ts b/packages/cli/src/providers/routing-rules.tsindex 371c9a97..0a11bf72 100644--- a/packages/cli/src/providers/routing-rules.ts+++ b/packages/cli/src/providers/routing-rules.ts@@ -399,12 +399,12 @@ async function routeBare(// it does not carry this model.//// This runs AFTER the credential filter, not before, and the order is not- // cosmetic: `providerServesModel` may hit the provider's own roster endpoint,+ // cosmetic: `providerServesModel` may hit the provider's own discovery endpoint,// which needs that provider's credential. Asking about a provider the user// cannot authenticate to would be a guaranteed-failing round-trip.//// Only "not-served" removes anything. "unknown" — no source covers this- // provider, the catalog is cold, the roster endpoint was briefly down — keeps+ // provider, the catalog is cold, the discovery endpoint was briefly down — keeps// the candidate exactly where it was. That asymmetry is the whole safety// property: reading absence of evidence as denial would drop every provider// neither source covers, which is almost entirely the SUBSCRIPTION providers,@@ -460,7 +460,7 @@ async function routeBare(* whose ids are already vendor-qualified and are their own spec. Availability* must be asked about the wire id, never the name the user typed — comparing the* typed name would test the wrong side of an `externalId` mapping, and that- * mapping is exactly what a roster settles (OpenCode Zen Go serves+ * mapping is exactly what a dynamic models catalog settles (OpenCode Zen Go serves* `deepseek-v4-pro`, while the catalog id carries a date suffix).*/function wireIdOf(route: Route): string {diff --git a/packages/cli/src/providers/transport/anthropic-compat.ts b/packages/cli/src/providers/transport/anthropic-compat.tsindex b94321a7..175da325 100644--- a/packages/cli/src/providers/transport/anthropic-compat.ts+++ b/packages/cli/src/providers/transport/anthropic-compat.ts@@ -118,8 +118,8 @@ export class AnthropicProviderTransport implements ProviderTransport {* probe exercises the model the user would actually get.** Non-chat rows are dropped with the shared `isChatCapable()` filter rather- * than a per-provider skip list: qwen-token-plan's roster mixes image/TTS models- * (`wan2.7-image`, `qwen-audio-3.0-tts-plus`) in with the text models, and+ * than a per-provider skip list: qwen-token-plan's dynamic models catalog mixes+ * image/TTS models (`wan2.7-image`, `qwen-audio-3.0-tts-plus`) in with the text models, and* hardcoding model ids here would rot the moment Alibaba ships the next one.*/async discoverProbeModel(exclude?: ReadonlySet<string>): Promise<DiscoveryOutcome> {diff --git a/packages/cli/src/providers/transport/antigravity-modelid.test.ts b/packages/cli/src/providers/transport/antigravity-modelid.test.tsindex d92c549b..36854cf1 100644--- a/packages/cli/src/providers/transport/antigravity-modelid.test.ts+++ b/packages/cli/src/providers/transport/antigravity-modelid.test.ts@@ -4,7 +4,7 @@ import { resolveAntigravityModelId } from "./antigravity.js";/*** resolveAntigravityModelId operates ONLY on the LIVE served set (no hardcoded* model ids). Every case injects its own servedIds + defaultId, so the tests- * carry no pinned roster and touch no network.+ * carry no pinned model list and touch no network.*/describe("resolveAntigravityModelId", () => {test("returns an exactly-served id untouched", () => {diff --git a/packages/cli/src/providers/transport/devin.ts b/packages/cli/src/providers/transport/devin.tsindex 37150445..1d0745b7 100644--- a/packages/cli/src/providers/transport/devin.ts+++ b/packages/cli/src/providers/transport/devin.ts@@ -22,7 +22,7 @@** **2. The reasoning tier is part of the model id.** There is no effort* parameter — `claude-opus-5` + effort `high` must become the uid- * `claude-opus-5-high`. Resolution runs against the LIVE roster+ * `claude-opus-5-high`. Resolution runs against the dynamic models catalog* (`getServedDevinModels`), never a hardcoded table, so it tracks whatever this* subscription is entitled to today.*@@ -61,7 +61,7 @@ export class DevinProviderTransport implements ProviderTransport {/** The delegated per-request auth artifact, populated by refreshAuth(). */private cachedAuth: RequestAuth | null = null;- /** The LIVE roster this subscription serves, refreshed per request (TTL-cached). */+ /** The dynamic models catalog this subscription serves, refreshed per request (TTL-cached). */private served: DevinModelConfig[] = [];/** The uid actually sent in field 21, set by serializeBody(). */@@ -69,7 +69,7 @@ export class DevinProviderTransport implements ProviderTransport {constructor(modelName: string) {this.modelName = modelName;- // Resolved against the live roster once refreshAuth() has run; until then+ // Resolved against the dynamic models catalog once refreshAuth() has run; until then// the raw name is a safe placeholder (ComposedHandler always awaits// refreshAuth before any request).this.resolvedUid = modelName;@@ -88,13 +88,13 @@ export class DevinProviderTransport implements ProviderTransport {}/**- * Resolve credentials and the live roster before each request.+ * Resolve credentials and the dynamic models catalog before each request.** A missing credential throws with `terminal = true`, so ComposedHandler* answers HTTP 400 (surfaced inline) rather than 401 — a 401 sends the client* into ~11 retries of a condition that cannot self-heal.*- * Roster discovery is fail-soft by contract (`getServedDevinModels` returns+ * Model discovery is fail-soft by contract (`getServedDevinModels` returns* `[]` on any error), so a discovery outage degrades resolution to* pass-through instead of breaking the turn.*/@@ -102,7 +102,7 @@ export class DevinProviderTransport implements ProviderTransport {this.cachedAuth = await credentials.getRequestAuth("devin", { model: this.modelName });this.served = await getServedDevinModels();log(- `[Devin] auth refreshed, model: ${this.modelName}, served roster: ${this.served.length} models`+ `[Devin] auth refreshed, model: ${this.modelName}, dynamic models catalog: ${this.served.length} models`);}@@ -110,7 +110,7 @@ export class DevinProviderTransport implements ProviderTransport {* The logical request object → enveloped Connect-protobuf bytes.** This is where the uid is finally resolved: `modelUid` + `effort` against the- * live roster. Effort rides the PAYLOAD rather than the transport because a+ * dynamic models catalog. Effort rides the PAYLOAD rather than the transport because a* cached handler serves overlapping turns that may carry different levels.*/serializeBody(payload: any): { body: SerializedBody; contentType: string } {@@ -176,9 +176,9 @@ export class DevinProviderTransport implements ProviderTransport {}/**- * Pick a probe-friendly model from the live roster.+ * Pick a probe-friendly model from the dynamic models catalog.*- * The cloud catalog cannot know a Devin roster — it is per-subscription — so+ * The cloud catalog cannot know Devin's dynamic models catalog — it is per-subscription — so* the ranking is a RULE (widest context window first, ties alphabetical) and* never a pinned id. `getServedDevinModels` already drops rows with no context* window, which is what removes `adaptive`, the server-side router@@ -211,7 +211,7 @@ export class DevinProviderTransport implements ProviderTransport {}/**- * Rewrite a TERMINAL in-stream error when the live roster proves the uid is+ * Rewrite a TERMINAL in-stream error when the dynamic models catalog proves the uid is* not served. Follows Antigravity's `rewriteModelNotFound` doctrine exactly.** Only rewrites when the served set is non-empty AND does not contain the@@ -231,7 +231,7 @@ export class DevinProviderTransport implements ProviderTransport {const families = [...new Set(this.served.map((model) => model.family).filter(Boolean))].sort();// 167 models span ~40 families, and pasting all of them buries the one line// the user needs. Families sharing the request's leading token come first —- // someone who typed `claude-opus-99` wants the claude-opus rosters, not+ // someone who typed `claude-opus-99` wants the claude-opus families, not// an alphabetical wall starting at `code-fast`.const stem = this.resolvedUid.split("-")[0]?.toLowerCase() ?? "";const related = families.filter((f) => f.toLowerCase().startsWith(stem));diff --git a/packages/cli/src/providers/transport/openai.ts b/packages/cli/src/providers/transport/openai.tsindex 73480711..28122722 100644--- a/packages/cli/src/providers/transport/openai.ts+++ b/packages/cli/src/providers/transport/openai.ts@@ -96,7 +96,7 @@ export class OpenAIProviderTransport implements ProviderTransport {* Discover from the provider definition's authenticated model-list endpoint.** OpenAI-compatible subscriptions such as Sakana Subscription and Grok- * Build have account-scoped rosters that are not interchangeable with their+ * Build have account-scoped dynamic models catalogs that are not interchangeable with their* metered siblings. The shared helper uses the same credential authority and* model-discovery cache as the picker, so probe fallback never guesses a* cross-silo model id.diff --git a/packages/cli/src/providers/transport/opencode-zen.ts b/packages/cli/src/providers/transport/opencode-zen.tsindex d123dadc..71261e9a 100644--- a/packages/cli/src/providers/transport/opencode-zen.ts+++ b/packages/cli/src/providers/transport/opencode-zen.ts@@ -17,7 +17,7 @@** The same docs paragraph asks a client to "identify itself with its own user agent* … rather than a generic SDK or HTTP-library name", and this relay enforces that- * on its other route: `model-discovery.ts:465-481` records that a UA-less roster+ * on its other route: `model-discovery.ts:465-481` records that a UA-less discovery* request to Zen Go answers `403 error code: 1010` — Cloudflare's browser-integrity* block — while the identical request carrying one returns 200. The chat path sits* behind the same edge, so it sends the same `claudish/<version>` string. Credit to@@ -34,7 +34,7 @@ export class OpenCodeZenTransport extends OpenAIProviderTransport {/*** Both headers are written BEFORE the base transport's, so auth and any* provider-declared `headers` still win — the precedence `model-discovery.ts`- * uses for the roster request on this same relay. No builtin declares either+ * uses for the discovery request on this same relay. No builtin declares either* key today, so nothing changes in practice; it means a custom endpoint that* pins its own value keeps it.*diff --git a/packages/cli/src/providers/transport/provider-model-discovery.ts b/packages/cli/src/providers/transport/provider-model-discovery.tsindex 2e53d4bf..041105d9 100644--- a/packages/cli/src/providers/transport/provider-model-discovery.ts+++ b/packages/cli/src/providers/transport/provider-model-discovery.ts@@ -8,7 +8,7 @@ import { getProviderByName } from "../provider-definitions.js";import { type DiscoveryOutcome, isChatCapable } from "./probe-discovery.js";/**- * Pick a probe model from one provider's authenticated, account-scoped roster.+ * Pick a probe model from one provider's authenticated, account-scoped dynamic models catalog.** The cloud catalog can publish a useful default, but it cannot know which* models a subscription credential serves today. Keeping this rule shared bydiff --git a/packages/cli/src/serve-command.ts b/packages/cli/src/serve-command.tsindex 25642b9f..365ab984 100644--- a/packages/cli/src/serve-command.ts+++ b/packages/cli/src/serve-command.ts@@ -137,7 +137,7 @@ export async function serveCommand(args: string[]): Promise<void> {}// Before the slot map is read, so a slot naming a bundled endpoint resolves- // against a roster that already contains it. `createProxyServer` registers+ // against a provider list that already contains it. `createProxyServer` registers// too and this is idempotent — the call is here so the guarantee belongs to// `serve` rather than to an implementation detail of the proxy it happens to// start.diff --git a/packages/cli/src/session/baseline-pricing.ts b/packages/cli/src/session/baseline-pricing.tsindex 76c820ee..3ca0ca45 100644--- a/packages/cli/src/session/baseline-pricing.ts+++ b/packages/cli/src/session/baseline-pricing.ts@@ -9,7 +9,7 @@ import { catalogRouteMatchesProvider } from "../providers/catalog-route-bindings* WHY NOT A CONSTANT. Hardcoding `sonnet = 3/15` would be wrong within a release —* claude-sonnet-5 is 2/10 while claude-sonnet-4-5 is 3/15, and a pinned model id ages* into a lie the moment the next model ships. The rule claudish already follows for- * rosters and context windows applies verbatim to prices: defaults are RULES, not+ * model lists and context windows applies verbatim to prices: defaults are RULES, not* pinned ids. So the baseline is resolved through the catalog's own* `~anthropic/claude-<tier>-latest` aliases — a rule that keeps pointing at the current* flagship without this file knowing which model that is. Today they resolve todiff --git a/packages/cli/src/test-fixtures/devin/README.md b/packages/cli/src/test-fixtures/devin/README.mdindex 8627cc9d..f5852a46 100644--- a/packages/cli/src/test-fixtures/devin/README.md+++ b/packages/cli/src/test-fixtures/devin/README.md@@ -1,14 +1,14 @@-# Devin roster fixture+# Devin dynamic models catalog fixture`GetCliModelConfigs.res.bin` — a **real, unmodified** `GetCliModelConfigs` response body capturedfrom `server.codeium.com` on 2026-08-06 against the developer's own Devin subscription. Bareprotobuf (unary rpcs are not Connect-enveloped), 173 model configs, 170 of which have a non-zerocontext window.-It is the fixture for `providers/model-resolvers/devin.ts`, whose entire job is to fold this roster-into the rows a human picks and unfold a selection back into a wire id. Both directions are only-meaningful against a real roster: the resolver's rules exist because of properties this file has-and a hand-written fixture would not —+It is the fixture for `providers/model-resolvers/devin.ts`, whose entire job is to fold this dynamic+models catalog into the rows a human picks and unfold a selection back into a wire id. Both+directions are only meaningful against a real dynamic models catalog: the resolver's rules exist+because of properties this file has and a hand-written fixture would not —- 97 uids are natively 1M with no `-1m` suffix, while only 7 carry one- 32 uids carry a `-priority` speed premium at ~2x cost@@ -18,6 +18,6 @@ and a hand-written fixture would not —Tests must READ this file. Restating any of it as literals in a test would reintroduce exactly thestaleness the live-discovery design removes, and would stop catching the case the resolver exists-for: a roster shape nobody anticipated.+for: a dynamic models catalog shape nobody anticipated.Contains no credentials — the request metadata carrying the api key is not part of a response body.diff --git a/packages/cli/src/tui/App.tsx b/packages/cli/src/tui/App.tsxindex 2eb9c33b..91546f27 100644--- a/packages/cli/src/tui/App.tsx+++ b/packages/cli/src/tui/App.tsx@@ -436,11 +436,11 @@ export function App({ requestLogin }: AppProps = {}) {// is sorting — calling it again inside the comparator would make every// `indexOf` return -1 and collapse the catalog order the sort exists to// preserve.- const roster = getProviderDefs();- return [...roster].sort((a, b) => {+ const providerDefs = getProviderDefs();+ return [...providerDefs].sort((a, b) => {const aReady = providerIsReadyForDisplay(a, config, localLiveness);const bReady = providerIsReadyForDisplay(b, config, localLiveness);- if (aReady === bReady) return roster.indexOf(a) - roster.indexOf(b);+ if (aReady === bReady) return providerDefs.indexOf(a) - providerDefs.indexOf(b);return aReady ? -1 : 1;});}, [config, localLiveness]);@@ -961,7 +961,7 @@ export function App({ requestLogin }: AppProps = {}) {setKeychainEnabled(true);refreshKeychainVars();// Every provider whose key just changed has a stale authority memo,- // and the endpoint roster may now include vendors that were gated out+ // and the endpoint list may now include vendors that were gated out// for lack of a local credential.credentials.invalidate();invalidateProbeProxyHandlers();diff --git a/packages/cli/src/tui/components/ProfilesContent.tsx b/packages/cli/src/tui/components/ProfilesContent.tsxindex 78e9010d..10d637d1 100644--- a/packages/cli/src/tui/components/ProfilesContent.tsx+++ b/packages/cli/src/tui/components/ProfilesContent.tsx@@ -132,7 +132,7 @@ export function ProfilesContent({// aligns cleanly even for long prefixes like "gemini-codeassist@".// One call, two uses. `getProviderPrefixes()` builds a fresh array per call// (deliberately — see its doc), so asking twice per render allocates twice and,- // worse, would let the width be computed from a different roster than the rows+ // worse, would let the width be computed from a different prefix list than the rows// if a provider were ever registered between the two calls.const prefixes = getProviderPrefixes();const prefixColW = prefixes.reduce((max, p) => Math.max(max, p.prefix.length), 0) + 2;diff --git a/packages/cli/src/tui/constants.ts b/packages/cli/src/tui/constants.tsindex 2dec871f..ecc06398 100644--- a/packages/cli/src/tui/constants.ts+++ b/packages/cli/src/tui/constants.ts@@ -49,7 +49,7 @@ export function getProviderPrefixes(): Array<{}/**- * Chain selector — same roster as the prefix picker, so the two never disagree+ * Chain selector — same provider list as the prefix picker, so the two never disagree* about which providers exist. A function for the same reason* `getProviderPrefixes()` is: a const would freeze the list before runtime* registration.diff --git a/packages/cli/src/tui/index.tsx b/packages/cli/src/tui/index.tsxindex 17b66e71..2f13bb95 100644--- a/packages/cli/src/tui/index.tsx+++ b/packages/cli/src/tui/index.tsx@@ -33,7 +33,7 @@ import { invalidateProbeProxyHandlers, shutdownProbeProxy } from "./probe-proxy.export async function startConfigTui(): Promise<void> {// Register bundled endpoints before the first render: the Providers tab is// built from `getAllProviders()`, so a catalog vendor the user has a key for- // must already be in the roster or it simply is not there to configure.+ // must already be in the provider list or it simply is not there to configure.// Re-entry after an OAuth login calls this again — idempotent by design.ensureEndpointsRegistered();diff --git a/packages/cli/src/tui/providers.ts b/packages/cli/src/tui/providers.tsindex 8bbc19c7..8841dae4 100644--- a/packages/cli/src/tui/providers.ts+++ b/packages/cli/src/tui/providers.ts@@ -194,7 +194,7 @@ export function providerAuthCapabilities(}/**- * The provider roster, built LIVE from the catalog on every call.+ * The provider list, built LIVE from the catalog on every call.** It used to be a module-load-time `const`, and that made it a snapshot taken* before `ensureEndpointsRegistered()` had run — so a runtime provider (adiff --git a/scripts/daily-provider-healthcheck.ts b/scripts/daily-provider-healthcheck.tsindex 0aeeaed4..46fb769d 100755--- a/scripts/daily-provider-healthcheck.ts+++ b/scripts/daily-provider-healthcheck.ts@@ -20,13 +20,13 @@* Layer 2 — `claudish --team … --mode json`: real interactive Claude Code* sessions in a real magmux grid, `--batch` at a time. ~27k input tokens each* (Claude Code's own system prompt), so ~$0.03 a model, ~$0.21 for the current- * roster. This is the layer that catches what a probe cannot: a translation+ * model list. This is the layer that catches what a probe cannot: a translation* bug in tool-call or streaming shape that only appears once a real agent is* driving the model.** WHY THE MODEL LIST IS NOT PINNED: it is read from* ~/.claudish/recommended-models.json at run time. Pinning ids would turn this- * into a snapshot of one afternoon's roster that goes red every time a vendor+ * into a snapshot of one afternoon's model list that goes red every time a vendor* ships, for reasons that have nothing to do with claudish. Same argument as* madbench/types.ts. New models are picked up on the first run after they land,* and the report calls out which ones are new since the previous run.