Skip to content

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.md
index 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 per
connection, 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 own
fault on that path (we were sending `output_config.effort: "minimal"`, which that host rejects by
diff --git a/ai-docs/reports/REMAINING_WORK_claudish-20260920.md b/ai-docs/reports/REMAINING_WORK_claudish-20260920.md
index 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 is
refuses to deny whenever `getDiscoveryFailure(provider)` is set, while the picker and the probe use
the 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.