Claude Code. Any Model. The most powerful AI coding agent now speaks every language.
docs: fold Jack's answers into the backend reply and the work plan
Jack Rudenkob41efe87d4a2
diff --git a/ai-docs/reports/BACKEND_TASKS_models_index-20260920.md b/ai-docs/reports/BACKEND_TASKS_models_index-20260920.mdindex b259666f..f20fb70f 100644--- a/ai-docs/reports/BACKEND_TASKS_models_index-20260920.md+++ b/ai-docs/reports/BACKEND_TASKS_models_index-20260920.md@@ -13,18 +13,12 @@`authoritative_roster_identity_unresolved`) are inclusion-resolution reasons, not `routeReason`values. The rename request is unchanged; only our label for them was wrong.-**Design question 1: how to publish an unknown modality.** claudish asks for an explicit `null`, and-treats a missing field the same way. The three states must be distinguishable:--| Published | Meaning | What claudish does |-|---|---|---|-| `["text"]`, `["text","image"]` … | verified, from the owner | uses it: a model whose output list has no text is not a chat model |-| `null` (or the field absent) | no evidence | unknown: claudish keeps the model, and falls back to its own rule (a `videoOutput` flag, then a name-shaped guess) |-| `[]` | never publish this | ambiguous between "none" and "unknown" |--Please do not infer a list from a name or a probe, exactly as you propose. An explicit unknown is-more useful to us than a guess, because claudish can say "unknown" to the user and keep the model-available, while a wrong `["text"]` would silently route an image model into a chat.+**Design question 1: how to publish an unknown modality.** Publish what you verified, and leave the+rest unknown in whichever way suits the schema: `null`, the field absent, or an empty list. claudish+treats all three as unknown, because **no model produces nothing**, so an empty output list cannot+mean "this model has no outputs" and can only mean "not established". Please do not infer a list from+a name or a probe, exactly as you propose: claudish can say "unknown" and keep the model available,+while a wrong `["text"]` would silently route an image model into a chat.**Design question 2: how to represent tiered prices.** claudish needs one comparable number perconnection, because Jack's routing rule of 2026-09-20 orders candidates inside a tier by the model's@@ -32,26 +26,47 @@ own vendor first, then by price. A shape that keeps the truth and still sorts:```jsonc"pricing": {- "shape": "flat" | "tiered" | "unavailable",+ "type": "flat" | "tiered" | "free" | "unavailable","input": 0.3, "output": 1.2, "cachedRead": 0.006, // flat: as today- "tiers": [ // tiered: in ascending order+ "tiers": [ // tiered: ascending, by input size{ "maxInputTokens": 32000, "input": 0.3, "output": 1.2 },{ "maxInputTokens": 128000, "input": 0.6, "output": 2.4 }]}```-- For `tiered`, claudish orders by the FIRST tier and labels the choice "tiered" where it shows a- price. It will not average, and will not pick a tier by guessing the request size.-- For `unavailable`, claudish treats the connection as unpriced and places it after priced ones- inside the same tier, which is what we already agreed for missing prices.-- Alibaba's 28 tiered models are exactly the case this shape is for; the 20 PAYG ids with no- representable price should be `"shape": "unavailable"` rather than omitted, so we can tell "no- price yet" from "not published".--**Alibaba PAYG, reconciled.** We agree the credential is good: it reads the official model API here-too (`GET /compatible-mode/v1/models` on `dashscope-intl`, HTTP 200, 169 models). The failure is on-chat calls only, and it is not about the model id:+`type` is a discriminator: it says how to read the rest of the object. Without it, a reader cannot+tell "no price published" from "the price is zero", nor whether the top-level numbers are the whole+price or only the first tier.++- `flat`: as today.+- `tiered`: claudish orders by the FIRST tier and labels the choice "tiered" where it shows a price.+ It will not average tiers and will not guess a tier from the request size.+- `free`: a genuinely free model. `flat` with `input: 0, output: 0` says the same thing; pick+ whichever is cleaner in your schema, as long as it is distinguishable from "unknown".+- `unavailable`: claudish treats the connection as unpriced and places it after priced ones inside+ the same tier, which is what we already agreed for missing prices.++Alibaba's 28 tiered models are the case `tiered` exists for; the 20 PAYG ids with no representable+price should be `"type": "unavailable"` rather than omitted, so we can tell "no price yet" from "not+published".++**Probe picks: no change requested.** You publish one pick per static route, chosen from the models+common to every account and recent, and client-selected routes are already marked. That is sound, and+claudish treats a pick as the first candidate rather than an entitlement, so a per-account denial+costs one attempt. Nothing for you to do here.++**Alibaba PAYG: deferred on our side, no action needed from you.** We are leaving it until after the+v3 release. Do not spend time reconciling the remaining PAYG ids: the denial does not depend on the+id. Measured directly with the key, outside claudish, on `dashscope-intl`: three request shapes+(Anthropic `/apps/anthropic/v1/messages`, OpenAI `/compatible-mode/v1/chat/completions`, DashScope+native `/api/v1/services/aigc/text-generation/generation`) times three models (`qwen3.8-max-0902`,+`qwen-plus`, `qwen-max`) are nine of nine `403 Model.AccessDenied`, while the same key lists 169+models; the China host rejects the key outright.++**Alibaba PAYG, earlier evidence.** We agree the credential is good: it reads the official model API+here too (`GET /compatible-mode/v1/models` on `dashscope-intl`, HTTP 200, 169 models). The failure is+on chat calls only, and it is not about the model id:- your new pick `qwen3.8-max-0902`: `403 Model.AccessDenied` through claudish, after we fixed our ownfault on that path (we were sending `output_config.effort: "minimal"`, which that host rejects bydiff --git a/ai-docs/reports/REMAINING_WORK_claudish-20260920.md b/ai-docs/reports/REMAINING_WORK_claudish-20260920.mdindex b992cfea..3a6cb60e 100644--- a/ai-docs/reports/REMAINING_WORK_claudish-20260920.md+++ b/ai-docs/reports/REMAINING_WORK_claudish-20260920.md@@ -88,7 +88,7 @@ has tests). One brief, one Codex run, each test verified to fail when its fix is| Item | What we do when it lands ||---|---|-| T1 modalities | consume `inputModalities`/`outputModalities`; treat `null` as unknown; drop the name-shaped guess |+| T1 modalities | consume `inputModalities`/`outputModalities`; treat `null`, an absent field AND an empty list as unknown, since no model produces nothing; drop the name-shaped guess || T2 naming | switch the reader on the generation they name, with no alias; we read no `roster*` field today, so only the reason mapping matters || T3 prices | activate price ordering inside a tier; `unavailable` sorts after priced || T6 Poe | prefer their verified pick; keep our account-list discovery as the fallback |@@ -109,7 +109,11 @@ has tests). One brief, one Codex run, each test verified to fail when its fix isrefuses to deny whenever `getDiscoveryFailure(provider)` is set, while the picker and the probe usethe rows — no type change, the failure channel already exists;**(c)** treat only a continuation marker as incomplete and tolerate dropped rows.- Recommendation: (b). **DISCUSS** if you want (c) instead, which is less safe but never hides rows.+ **Decided (Jack, 2026-09-20): (b), and the incompleteness must be RETURNED, not only recorded.** A+ partial list is still a partial-failure state, and the caller has to be able to see it: a future+ picker can then say "this list may be incomplete" instead of quietly showing fewer models. Shape to+ build: a detailed call (`{ models, incomplete?: { reason, droppedRows?, hasMore? } }`) with the+ existing array function kept as a thin wrapper, so no caller has to change at once.- Redact `readinessDetail` before anything displays it. Nothing displays it today.- A probe failure rendered `[object Object]` as its error (seen on the Mistral walk). **RESEARCH** where the object reaches the message.- Extend the vocabulary guard to the retired terms above once they are renamed.@@ -130,3 +134,8 @@ says so.- After 9.8.0 is installed: delete the `QWEN_CLOUD_PLAN_API_KEY` Keychain item.- `/Users/jack/mag/claudish/.env` holds 5 duplicate lines.- Jack's side, not claudish: the Alibaba Coding key is the Token Plan key, and PAYG needs model access granted in the Model Studio console.+- **Alibaba PAYG is deferred until after the v3 release (Jack, 2026-09-20).** Measured: nine of nine+ combinations denied on `dashscope-intl` (Anthropic, OpenAI and DashScope-native request shapes,+ each with `qwen3.8-max-0902`, `qwen-plus` and `qwen-max`), while the same key lists 169 models; the+ China host rejects the key. So the endpoint and the parameters are not the cause. Nobody should+ spend time reconciling ids until the account's model access is settled.