Merge remote-tracking branch 'origin/master' into feat/web-message-feedback-ui

.agents/notes/implemented/architecture/2026-08-03-pi-ai-declared-provider-catalog.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-08-03-pi-ai-declared-provider-catalog.md
-2026-08-03-pi-ai-declared-provider-catalog.md: 45de54f29e718ed1c9364b03da762700e1282e7b
-2026-08-03-pi-ai-declared-provider-catalog.zh.md: 43206bc21dfa74f6ddf0583c72dc4e8be1dbb9e5
+2026-08-03-pi-ai-declared-provider-catalog.md: 571986a94c8547538b44b11510609e0000923537
+2026-08-03-pi-ai-declared-provider-catalog.zh.md: 917c19e77dc1db7fe4dc20acf1d9b280277a76d2

.agents/notes/implemented/architecture/2026-08-03-pi-ai-declared-provider-catalog.md

@@ -14,7 +14,7 @@ The adapter also streamed through `streamSimple` from `@earendil-works/pi-ai/com
 
 A provider route is a **declaration**, and the installed catalog is its default. `resolveProfiles` no longer checks route keys against `getBuiltinProviders()`. Instead each route resolves to a materialized model list plus the pi-ai `Provider` that serves it:
 
-- `catalog.ts` merges the installed catalog under the profile's own entries. A profile's `models` list *replaces* the route's catalog (an absent or empty list serves it unchanged), and each entry defaults its unset fields from the installed model of the same `id`. Only the fields the harness consumes are configurable — `id`, `name`, `contextWindow`, `maxTokens`; [[2026-08-08-pi-ai-per-model-reasoning-declarations]] later added `reasoningEfforts` and `compat`, which is also where the original "reasoning rides the installed entry or is absent" stance was revisited (a bare capability flag stays rejected; a full per-level declaration with wire spellings does not have its problem). Pricing and input modalities remain absent from the surface because nothing reads them: `replay.ts` zeroes pi-ai's cost metadata and `context.ts` keeps only text blocks. Materialization spreads the installed entry and overrides the configured fields, rather than enumerating the result: an enumerated rebuild silently drops every `Model` field this package does not model, which is how `headers` went missing from an nvidia route once already.
+- `catalog.ts` merges the installed catalog under the profile's own entries. A profile's `models` list *replaces* the route's catalog (an absent or empty list serves it unchanged), and each entry defaults its unset fields from the installed model of the same `id`. Only the fields the harness consumes are configurable — `id`, `name`, `contextWindow`, `maxTokens`; [[2026-08-08-pi-ai-per-model-reasoning-declarations]] later added `reasoningEfforts` and `compat`, which is also where the original "reasoning rides the installed entry or is absent" stance was revisited (a bare capability flag stays rejected; a full per-level declaration with wire spellings does not have its problem). Input modalities were opened later, as `input` on an entry plus a route-level `defaultInput`, once the image-admission points turned an unreported modality into a refusal a deployment could not lift ([[2026-08-12-pi-ai-route-default-input-modalities]]); the "nothing reads them" reasoning had described `llm-deepseek`'s serializer rather than this route, whose converter carries images. Pricing stays absent for the original reason: `replay.ts` zeroes pi-ai's cost metadata and no consumer reports spend. Materialization spreads the installed entry and overrides the configured fields, rather than enumerating the result: an enumerated rebuild silently drops every `Model` field this package does not model, which is how `headers` went missing from an nvidia route once already.
 - `provider.ts` builds the route's `Provider`. A catalog route that keeps its catalog protocol **reuses** the installed provider with `getModels()` replaced; every other route is built by `createProvider()` over a protocol table whose entries are the same `@earendil-works/pi-ai/api/*.lazy` factories pi-ai's own provider factories use. That table is narrower than pi-ai's full API set on purpose — it holds only protocols a profile can completely describe with a key, an endpoint, and headers, so Bedrock (SigV4 plus a region), Vertex (project, location, ADC), Azure (provider environment plus an api-version), and Codex (OAuth) are absent rather than offered as routes that cannot authenticate. Catalog routes still reach them through their own provider; only an explicit override is refused.
 - `adapter.ts` turns each resolution into an **immutable snapshot** — the profiles plus a `createModels()` collection holding those providers — and every operation captures a whole snapshot before its first `await`.
 - A model's **explicitly configured** `maxTokens` becomes the seam's `defaultMaxTokens`. The value inherited from the installed catalog does not: pi-ai requires `Model.maxTokens` as the model's output *capability*, while `defaultMaxTokens` is a cap the deployment chose to send on requests that name none, and materializing the former as the latter would start capping every request at a number nobody picked.
@@ -50,7 +50,7 @@ A route's auth follows from that. A catalog route keeps the installed provider's
 - **Keep `createProvider()` but skip the `Models` collection**, streaming through `provider.streamSimple(model, ctx, {apiKey})`. Smallest diff and the credential path is untouched, but `createProvider`'s `auth` is a required field that this path never invokes — a required-by-signature implementation with no caller. It also leaves `refreshModels` needing a hand-built `RefreshModelsContext`, and keeps the adapter off the runtime pi-ai actually supports.
 - **Reuse the installed provider for catalog routes and `createProvider()` only for declared ones**, with no shared resolution. Zero risk to catalog behavior, but catalog materialization, endpoint override, and per-model configuration would each exist twice, and a catalog route that repoints its protocol would have to jump paths mid-resolution. The chosen split confines the asymmetry to provider construction, where it is forced by pi-ai not exposing a built provider's API implementations.
 - **Rebuild every route through `createProvider()`**, including catalog ones. Fully symmetric, but a built `Provider` does not expose its `api`, so the protocol table would become the ceiling on which providers work — Bedrock loads its Smithy module through a separate entry point and would silently stop working.
-- **Expose pi-ai's whole `Model` shape** (cost, input modalities, `thinkingLevelMap`, `compat`). Maximum configurability, but no current consumer read those fields then, so a configured price or modality would change nothing while reading as supported. The consumer-driven half of this arrived later: [[2026-08-08-pi-ai-per-model-reasoning-declarations]] opened reasoning (as `reasoningEfforts`, not a raw `thinkingLevelMap`) and the two reasoning-dispatch `compat` switches once selectors and dispatch actually consumed them; cost and modalities stay closed for the original reason.
+- **Expose pi-ai's whole `Model` shape** (cost, input modalities, `thinkingLevelMap`, `compat`). Maximum configurability, but no current consumer read those fields then, so a configured price or modality would change nothing while reading as supported. The consumer-driven half of this arrived later, field by field as a consumer appeared: [[2026-08-08-pi-ai-per-model-reasoning-declarations]] opened reasoning (as `reasoningEfforts`, not a raw `thinkingLevelMap`) and the two reasoning-dispatch `compat` switches once selectors and dispatch actually consumed them, and [[2026-08-12-pi-ai-route-default-input-modalities]] opened modalities (as `input` and `defaultInput`, not a raw `Model.input` passthrough) once the image-admission points read them. Cost stays closed for the original reason.
 
 - **Keep one mutable `Models` collection and re-sync it.** Fewer allocations, and correct for every operation that resolves synchronously. It is exactly wrong for the one that does not: `stream()` awaits a credential between capturing its model and dispatching it.
 - **Simulate an atomic directory swap with dispose-then-register.** No seam change, and it works whenever the new set is valid — which is the case that never needed atomicity.

.agents/notes/implemented/architecture/2026-08-03-pi-ai-declared-provider-catalog.zh.md

@@ -14,7 +14,7 @@ Status: implemented
 
 提供方路由是一份**声明**，已安装 catalog 是它的默认值。`resolveProfiles` 不再拿路由键去核对 `getBuiltinProviders()`，而是把每条路由解析成一份物化模型列表，外加服务它的 pi-ai `Provider`：
 
-- `catalog.ts` 把已安装 catalog 合并到 profile 自身条目之下。profile 的 `models` 列表*替换*该路由的 catalog（列表缺席或为空则原样服务），每个条目从同 `id` 的已安装模型继承自身未设置的字段。只有 harness 会消费的字段可配置——`id`、`name`、`contextWindow`、`maxTokens`；[[2026-08-08-pi-ai-per-model-reasoning-declarations]] 之后加入了 `reasoningEfforts` 与 `compat`，当初「推理（reasoning）沿用已安装条目或直接缺席」的立场也在那里被重新审视（孤立的能力布尔量仍被拒绝；带 wire 拼写的逐档位完整声明没有它那个问题）。定价与输入模态仍不出现在配置面，因为没有任何读取方：`replay.ts` 把 pi-ai 的成本元数据清零，`context.ts` 只保留文本块。物化时以已安装条目铺底、再覆盖已配置的字段，而不是逐字段枚举结果：枚举式重建会静默丢弃本包未建模的每一个 `Model` 字段——`headers` 就是这样从某条 nvidia 路由上消失过一次。
+- `catalog.ts` 把已安装 catalog 合并到 profile 自身条目之下。profile 的 `models` 列表*替换*该路由的 catalog（列表缺席或为空则原样服务），每个条目从同 `id` 的已安装模型继承自身未设置的字段。只有 harness 会消费的字段可配置——`id`、`name`、`contextWindow`、`maxTokens`；[[2026-08-08-pi-ai-per-model-reasoning-declarations]] 之后加入了 `reasoningEfforts` 与 `compat`，当初「推理（reasoning）沿用已安装条目或直接缺席」的立场也在那里被重新审视（孤立的能力布尔量仍被拒绝；带 wire 拼写的逐档位完整声明没有它那个问题）。输入模态后来被开放，形态是条目上的 `input` 加路由级 `defaultInput`——图片准入点使得「未被报告的模态」变成部署无法解除的拒绝之后（[[2026-08-12-pi-ai-route-default-input-modalities]]）；当初「没有任何读取方」那句论证描述的其实是 `llm-deepseek` 的序列化器，而不是这条路由，它的转换器能携带图片。定价仍因原有理由不出现在配置面：`replay.ts` 把 pi-ai 的成本元数据清零，且没有任何消费方报告开销。物化时以已安装条目铺底、再覆盖已配置的字段，而不是逐字段枚举结果：枚举式重建会静默丢弃本包未建模的每一个 `Model` 字段——`headers` 就是这样从某条 nvidia 路由上消失过一次。
 - `provider.ts` 构造路由的 `Provider`。保持 catalog 协议不变的 catalog 路由会**复用**已安装提供方，只替换 `getModels()`；其余路由都由 `createProvider()` 基于一张协议表构造，表中条目正是 pi-ai 自己的提供方工厂所用的 `@earendil-works/pi-ai/api/*.lazy` factory。该表刻意窄于 pi-ai 的完整 API 集合——只保留 profile 能用密钥、端点与标头完整描述的协议，因此 Bedrock（SigV4 加 region）、Vertex（project、location、ADC）、Azure（提供方环境加 api-version）与 Codex（OAuth）不在其中，而不是被当作无法认证的路由提供出去。catalog 路由仍可经自己的 provider 抵达它们；被拒的只有显式覆盖。
 - `adapter.ts` 把每次解析变成一份**不可变快照**——profiles 加上持有这些 provider 的 `createModels()` 集合——每个操作都在自己第一个 `await` 之前整体捕获一份。
 - 模型**显式配置**的 `maxTokens` 会成为 seam 的 `defaultMaxTokens`；从已安装 catalog 继承来的那份不会：pi-ai 要求 `Model.maxTokens` 表示模型的输出*能力*，而 `defaultMaxTokens` 是部署选定、发给未点名上限的请求的那个值，把前者物化成后者会让每个请求都被一个无人选择的数字封顶。
@@ -50,7 +50,7 @@ pi-ai 的 `Models` 自带一套凭据概念——按提供方 id 索引的 `Cred
 - **保留 `createProvider()` 但不建 `Models` 集合**，改由 `provider.streamSimple(model, ctx, {apiKey})` 发起。改动最小且凭据路径原封不动，但 `createProvider` 的 `auth` 是必填字段，这条路上它永远不会被调用——一份因签名而必填、却没有调用方的实现。它还让 `refreshModels` 需要手工构造 `RefreshModelsContext`，并使适配器始终不在 pi-ai 真正支持的运行时上。
 - **catalog 路由复用已安装提供方，只有声明式路由走 `createProvider()`**，且两者不共享解析。对 catalog 行为零风险，但 catalog 物化、端点覆盖与每模型配置这三件事都要各写两遍，而改指协议的 catalog 路由还得在解析中途跳到另一条路径。已采纳的拆法把不对称收敛在提供方构造这一处——那里的不对称是 pi-ai 不暴露已构造提供方的 API 实现所强加的。
 - **让每条路由都经 `createProvider()` 重建**，包括 catalog 路由。完全对称，但已构造的 `Provider` 不暴露自己的 `api`，于是协议表会成为「哪些提供方能用」的天花板——Bedrock 经独立入口加载其 Smithy 模块，会因此静默失效。
-- **完整暴露 pi-ai 的 `Model` 形状**（成本、输入模态、`thinkingLevelMap`、`compat`）。可配置性最大，但这些字段当时没有任何读取方，因此配了价格或模态什么也不会改变，却看起来像是受支持的。这条否决里由消费方驱动的那一半后来兑现了：[[2026-08-08-pi-ai-per-model-reasoning-declarations]] 在选择器与分派真正消费之后开放了推理（以 `reasoningEfforts` 的形态，而非裸 `thinkingLevelMap`）和两个推理分派 `compat` 开关；成本与模态仍因原有理由保持关闭。
+- **完整暴露 pi-ai 的 `Model` 形状**（成本、输入模态、`thinkingLevelMap`、`compat`）。可配置性最大，但这些字段当时没有任何读取方，因此配了价格或模态什么也不会改变，却看起来像是受支持的。这条否决里由消费方驱动的那一半后来逐字段兑现了，每次都等到出现真实读取方：[[2026-08-08-pi-ai-per-model-reasoning-declarations]] 在选择器与分派真正消费之后开放了推理（以 `reasoningEfforts` 的形态，而非裸 `thinkingLevelMap`）和两个推理分派 `compat` 开关；[[2026-08-12-pi-ai-route-default-input-modalities]] 在图片准入点开始读取之后开放了模态（以 `input` 与 `defaultInput` 的形态，而非裸 `Model.input` 直通）。成本仍因原有理由保持关闭。
 
 - **保留单个可变 `Models` 集合并重新同步。** 分配更少，且对每个同步完成解析的操作都是正确的；唯独对那个不同步的操作恰恰是错的：`stream()` 会在捕获模型与派发模型之间 await 一次凭据。
 - **用「先 dispose 再注册」模拟目录原子替换。** 无需改 seam，且在新集合有效时确实可用——而那正是从不需要原子性的那种情形。

.agents/notes/implemented/architecture/2026-08-12-pi-ai-route-default-input-modalities.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-08-12-pi-ai-route-default-input-modalities.md
+2026-08-12-pi-ai-route-default-input-modalities.md: efd20b2cd73979208bb777fa42536bc5b918e29e
+2026-08-12-pi-ai-route-default-input-modalities.zh.md: 069a7916c8d4ffe738ff910a851e0a9cd1f66d0a

.agents/notes/implemented/architecture/2026-08-12-pi-ai-route-default-input-modalities.md

@@ -0,0 +1,53 @@
+# Agent Note: A pi-ai model declares its own input modalities, and undeclared means text
+
+Status: implemented
+
+English | [中文](2026-08-12-pi-ai-route-default-input-modalities.zh.md)
+
+## Problem
+
+Nothing in `settings.yaml` could describe a hand-declared pi-ai model as accepting images, and the adapter assumed text-only for every model the installed pi-ai catalog does not describe. Every model a deployment adds through the web UI's "add a custom provider" card is such a model, so an OpenAI-compatible gateway serving a vision model reported `inputModalities: ['text']` no matter what it actually served.
+
+The harness treats an omitted modality as negative capability, and three admission points act on it before any request is built: model selection refuses to switch into a session that already holds images, prompt admission refuses an image, and `read_image` refuses to read one. Their diagnostics tell the user to select an image-capable model — advice with no reachable referent, because no configuration key could make a hand-declared model image-capable. The route was closed at the metadata, not at the capability: the request converter and every pi-ai wire protocol carry images, and `llm-pi-ai`'s own stream guard is the only thing that would have stopped one.
+
+The assumption was justified in the source as the adapter's real capability rather than a deployment choice, and [[2026-08-03-pi-ai-declared-provider-catalog]] recorded the same reasoning when it decided which `Model` fields the configuration surface would expose ("nothing reads them: … `context.ts` keeps only text blocks"). That justification described the DeepSeek chat-completions adapter, whose serializer genuinely rejects image blocks, and had never been true of the pi-ai route. This note supersedes that one on modalities alone; pricing stays closed there for its own, still-current reason.
+
+## Decision
+
+**Modalities resolve entry `input` → installed catalog entry → route `defaultInput`, which itself defaults to `[text]`.** That is the chain `contextWindow` and `maxTokens` already use, field for field. pi-ai types `Model.input` required and per-model, so the entry field mirrors upstream directly: one route can serve a vision model beside a text-only one, and an override can correct a catalog model whose gateway serves other modalities than the catalog records. The route field spares a gateway whose *undescribed* models all take images from repeating itself on every entry.
+
+**The route value is a fallback, not an override — the catalog outranks it.** This is the `default*` ordering rather than `compat`'s, and the two are not interchangeable: `compat` shadows the catalog because a route-level protocol repoint invalidates the catalog's reasoning-dispatch facts wholesale, while a modality is a per-model property the catalog states accurately for the models it ships. Making the route value win would mean `defaultInput: [text]` silently strips images from every catalog vision model on the route — a footgun with no matching benefit, since narrowing one such model is what that model's own `input` is for.
+
+**Undeclared means `[text]`, and that is the absence of a declaration rather than a guess at the endpoint.** Nothing can interrogate a gateway for its modalities — no OpenAI-compatible listing endpoint reports them — so the only honest floor is the modality every supported protocol certainly carries. This is where the modality fallback parts company with the capacity ones: 262,144 tokens is merely plausible and wrong in both directions (a gateway serving 8k overflows, one serving 1M is wasted), while text is safe in one direction. The two wrong answers do not cost the same either. Under-claiming refuses the image before it is attached, naming the model, and the remedy is one documented line. Over-claiming admits an image the provider then rejects mid-turn, *after* prompt admission has committed the message durably, so the session keeps re-sending a request that cannot succeed and model selection refuses a switch to any text-only model. A cheap refusal at the earliest resolvable point beats an expensive one at the latest.
+
+**An entry's empty list means the same as an absent one; the route's is refused.** `[]` describes a model that accepts nothing and could serve no request, so it states no answer and resolution continues past it. That reading is not cosmetic: the config schema materializes `[]` for an absent array, so treating it as "accepts nothing" would silently strip images from every catalog vision model a `models` list happens to name. The route value has nothing below it to answer instead, so its empty list is refused where it is written. The route's `models` list already resolves absent-and-empty the same way for the same reason.
+
+**No configuration surface edits `input`.** It joins `compat`, `reasoningEfforts`, `thinkingBudgets`, and `headers` as a settings-document field, and the model-list editor stays a hand-written form over id, name, and the two capacities. This costs nothing durable because that card was already built to carry fields it does not edit: its row patch spreads the stored row before applying changes, and adoption keeps an existing row over a rediscovered candidate, so a hand-written `input` survives both.
+
+The DeepSeek chat-completions adapter is untouched. Its `['text']` is a fact about its serializer, not a missing declaration, and it keeps refusing before the send.
+
+## Alternatives considered
+
+- **An optimistic `[text, image]` default** — makes the motivating case work with zero configuration, and the web form writes no modality at all, so a conservative default leaves the remedy in a file a web-only user has no reason to open. Rejected on the severity of being wrong: a refused attachment is a speed bump with a documented fix, while a provider rejection poisons the session, presents as an unexplained repeating failure, and is escapable only by switching models or starting over. Documenting the remedy on the model-configuration page closes the discoverability gap; nothing closes the poisoned session.
+- **A route value that overrides the catalog** (`compat`'s ordering: entry → route → catalog) — lets a deployment that repoints a catalog route at its own gateway declare "no vision here" once. Rejected because the same sentence then silently disables every catalog vision model on a route where someone wrote it by analogy with the capacity fields, and the legitimate case is served by that model's own `input`. An override would also have to be named `input` at the route, since calling it `default*` beside two genuine fallbacks would misdescribe it.
+- **No route field at all, only the entry one** — closest to upstream, which has no route-level concept. Rejected on the bulk case the product's own flow produces: "fetch available models" adopts thirty ids with no modality, and an all-vision gateway would need `input` hand-written on each.
+- **A route-level `defaultInput` with no entry field** — cannot mix modalities on one route or correct a single catalog model, leaving "split the provider across two route keys" as the only workaround, at the cost of a second permanent provider id and a duplicate entry in every model selector.
+- **Probe the endpoint for its modalities** — no OpenAI-compatible listing endpoint reports them.
+- **Infer from the model id** (`*-vision`, `*-vl`) — a naming convention is not a capability, and a gateway renames freely.
+- **Keep refusing and improve the diagnostic** — the message was already accurate about the state and useless about the remedy; the missing thing was the remedy.
+
+## Consequences
+
+A vision model on a custom provider costs one line, `input: [text, image]`, written in the settings document — or one line at the route when every model it lists takes images. That is the whole of the fix: the three admission points then admit images on it and `read_image` works. A deployment that writes nothing keeps exactly the behavior it had, so no existing route changes what it reports.
+
+The image-admission gate keeps its meaning everywhere, because every modality it reads is now either recorded by the installed catalog or written by a person. Nothing claims a capability on a deployment's behalf.
+
+A model that declares images its endpoint does not serve is not caught locally — the claim is not verified — and the resulting failure is expensive. Prompt admission commits the user message durably (`agent/inbox/spliced`) before the request is built, so the rejected image stays in the session log: that model keeps re-sending it, and model selection refuses a switch to any text-only model. Recovery is to select a model that does serve images, fork before the image, or start a session. Making that failure non-destructive — rolling an unconsumed image message back out of the log when the send fails — is the change that would make an optimistic default reconsiderable, and is not attempted here.
+
+## Testing
+
+`packages/llm/llm-pi-ai/tests/catalog.spec.ts` covers each rung of the chain and both readings of an empty list at the resolver: one route mixing an undeclared model with entry-declared text-only and vision models, a route default answering an undeclared model while an entry still outranks it, a catalog vision model keeping its modalities under a narrower route default, an entry's `[]` inheriting rather than emptying, and the route's `[]` refused. A separate case re-asserts every rung end to end — a written settings section, the plugin's own registration, and `ctx.llm.listModels` / `resolveModelInfo` — so a break between the document and `LlmModelInfo` cannot pass.
+
+`config.spec.ts` holds the schema boundary: an unknown modality refused at both levels, the empty route list accepted by the schema and refused by the namespace validator that the settings seam actually runs, and the `[]` materialization for an absent array that the inheritance rule depends on.
+
+No keyless snapshot lane exercises a pi-ai route: the snapshot examples drive `dsh-llm-replay`, which declares modalities directly in its configuration, and a pi-ai route needs a live endpoint whose port a static `cordis.yml` cannot name. The admission points this change feeds are already covered there through that provider (`examples/acp-agent/image.cordis.snapshot.yml` and `image-text-route.cordis.snapshot.yml`) and are unaffected — what changed is what one adapter reports, not how a gate reads it.

.agents/notes/implemented/architecture/2026-08-12-pi-ai-route-default-input-modalities.zh.md

@@ -0,0 +1,53 @@
+# Agent Note：pi-ai 模型自行声明输入模态，未声明即为文本
+
+Status: implemented
+
+[English](2026-08-12-pi-ai-route-default-input-modalities.md) | 中文
+
+## 问题
+
+`settings.yaml` 里没有任何写法能把一个手写的 pi-ai 模型描述成接受图片，而适配器对已安装 pi-ai catalog 未描述的每个模型都假定纯文本。部署通过 Web UI 的“添加自定义提供商”卡片新增的模型统统属于这一类，因此一个提供视觉模型的 OpenAI 兼容网关，无论实际提供什么，报出来的都是 `inputModalities: ['text']`。
+
+Harness 把缺失的模态当作否定能力，并有三个准入点在构造任何请求之前就依据它行动：模型选择拒绝切换进已含图片的会话，prompt 准入拒绝图片，`read_image` 拒绝读取。它们的诊断让用户去选一个支持图片的模型——这条建议没有可达的指向对象，因为没有任何配置键能让手写模型变得支持图片。这条路封死在元数据上而非能力上：请求转换器和每种 pi-ai 线路协议都能携带图片，`llm-pi-ai` 自身的流前置检查才是唯一会拦下它的东西。
+
+源码中把这一假定论证为适配器的真实能力而非部署选择，而 [[2026-08-03-pi-ai-declared-provider-catalog]] 在决定配置面公开哪些 `Model` 字段时记录了同一套论证（「没有任何读取方：…… `context.ts` 只保留文本块」）。该论证描述的是 DeepSeek chat-completions 适配器——它的序列化器确实拒绝 image block——对 pi-ai 路由从来就不成立。本 note 仅在模态一项上取代那一条；定价在那里依然因其自身、仍然成立的理由保持关闭。
+
+## 决策
+
+**模态的解析顺序是：条目的 `input` → 已安装 catalog 条目 → 路由的 `defaultInput`，后者自身默认为 `[text]`。** 这与 `contextWindow`、`maxTokens` 已经在用的链条逐字段一致。pi-ai 把 `Model.input` 定义为必填且按模型给出，因此条目字段直接镜像上游：一条路由可以让视觉模型与纯文本模型并存，而 override 可以修正那些网关实际提供的模态与 catalog 记录不符的模型。路由字段则让那些**未被 catalog 描述的**模型全都接受图片的网关，不必在每个条目上重复一遍。
+
+**路由值是回退值而非覆盖值——catalog 的优先级更高。** 这采用的是 `default*` 的顺序而非 `compat` 的，两者不可互换：`compat` 之所以盖住 catalog，是因为路由级的协议改指会整体作废 catalog 关于推理分派的事实；而模态是按模型的属性，对 catalog 自己出货的那些模型，它记录得准确无误。让路由值获胜就意味着 `defaultInput: [text]` 会悄悄剥掉该路由上每一个 catalog 视觉模型的图片能力——一个没有对应收益的坑，因为收窄其中某个模型正是该模型自己的 `input` 要做的事。
+
+**未声明即 `[text]`，而这是「尚未声明」，不是对端点的猜测。** 没有任何环节能去询问网关的模态——没有任何 OpenAI 兼容的列表端点会报告它们——因此唯一诚实的底线是每个受支持协议都确定携带的那个模态。这也正是模态回退值与容量回退值分道扬镳之处：262,144 只是个说得过去的数字，且两个方向都会错（网关只给 8k 会溢出，给 1M 则被浪费），而 text 在一个方向上是安全的。两种猜错的代价同样并不对等。少声明会在图片被附加之前就拒绝并点名该模型，补救办法是一行有文档可依的配置。多声明会接纳一张图片、再由提供方在轮次中途拒绝——而此时 prompt 准入**早已**把消息持久化提交，于是会话会不断重发一个不可能成功的请求，且模型选择拒绝切换到任何纯文本模型。在最早可解析点付出一次廉价的拒绝，胜过在最晚点付出一次昂贵的。
+
+**条目的空列表与缺省同义；路由的空列表则被拒绝。** `[]` 描述的是一个什么都不接受、无法服务任何请求的模型，因此不作答，解析继续往下走。这个读法不是修辞：配置 schema 会为缺省数组物化出 `[]`，把它当作“什么都不接受”，会悄悄剥掉 `models` 列表恰好点到的每一个 catalog 视觉模型的图片能力。而路由值下面没有可以代为作答的层级，因此它的空列表在写入处即被拒绝。路由的 `models` 列表出于同样的理由，早已用同一种方式解析缺省与空。
+
+**没有任何配置界面编辑 `input`。** 它和 `compat`、`reasoningEfforts`、`thinkingBudgets`、`headers` 一样是 settings 文档字段，而模型列表编辑器仍是一张只覆盖 id、名称和两个容量的手写表单。这不会带来持久代价，因为那张卡片本来就是按“承载自己并不编辑的字段”建造的：它的行 patch 会先展开已存储的行再应用改动，而采纳候选时已有行优先于重新发现的候选，因此手写的 `input` 在两条路径上都能存活。
+
+DeepSeek chat-completions 适配器保持不动。它的 `['text']` 是关于其序列化器的事实，而不是一处缺失的声明，它继续在发送前拒绝。
+
+## 备选方案
+
+- **乐观的 `[text, image]` 默认值** —— 让触发本次变更的场景零配置即可工作；而且网页表单不会写入任何模态，因此保守默认值会把补救办法留在一个纯 Web 用户没有理由打开的文件里。被否决的理由是猜错时的严重程度：被拒绝的附件是一个有文档可依的减速带，而提供方拒绝会毒化整个会话、表现为一次无从解释的反复失败，且只能靠换模型或重开会话脱身。把补救办法写进配置模型页即可补上可发现性的缺口；而毒化的会话没有任何东西能补。
+- **让路由值盖住 catalog**（`compat` 的顺序：条目 → 路由 → catalog）—— 可以让把 catalog 路由改指到自家网关的部署，一句话声明「这里没有视觉能力」。被否决是因为同一句话也会在有人照着容量字段类比写下它的路由上，悄悄禁用每一个 catalog 视觉模型；而那个正当场景由该模型自己的 `input` 承担。覆盖值还必须在路由级改名叫 `input`，因为在两个货真价实的回退值旁边把它叫作 `default*` 是名不副实。
+- **完全不要路由字段，只要条目字段** —— 最贴近上游（上游没有路由级概念）。被否决的理由是产品自身流程会产生的批量场景：「获取可用模型」一次采纳三十个不带模态的 id，全是视觉模型的网关就得逐个手写 `input`。
+- **只要路由级 `defaultInput`，不要条目字段** —— 无法在一条路由上混合模态，也无法修正单个 catalog 模型，唯一的变通办法只剩「把该提供方拆成两个路由键」，代价是多一个永久的 provider id 和每个模型选择器里的一项重复。
+- **探测端点的模态** —— 没有任何 OpenAI 兼容的列表端点会报告它们。
+- **从模型 id 推断**（`*-vision`、`*-vl`）—— 命名约定不是能力，网关可以随意改名。
+- **继续拒绝，只改进诊断** —— 那条消息对状态的描述本来就准确，对补救办法则毫无用处；缺的正是补救办法。
+
+## 影响
+
+自定义提供方下的视觉模型只需在 settings 文档里写一行 `input: [text, image]`——或者当该路由列出的模型全都接受图片时，在路由上写一行。这就是修复的全部：三个准入点随即接纳它上面的图片，`read_image` 也可用。什么都不写的部署保持原有行为完全不变，因此没有任何既有路由改变它报出的内容。
+
+图片准入门禁在各处都保住了自己的意义，因为它读到的每一个模态，如今要么由已安装 catalog 记录，要么由人写下。没有任何环节会替部署宣称一项能力。
+
+声明了端点并不提供的图片能力的模型不会在本地被拦下——该断言不经验证——而由此产生的失败代价高昂。prompt 准入在构造请求之前就把用户消息持久化提交（`agent/inbox/spliced`），因此被拒绝的图片会留在会话日志里：该模型会不断重发它，而模型选择拒绝切换到任何纯文本模型。恢复途径是选择一个确实提供图片能力的模型、fork 到图片之前，或者开启新会话。让这次失败不具破坏性——发送失败时把尚未消费的图片消息从日志中回滚出去——才是能让乐观默认值重新可考虑的那项改动，本次未做尝试。
+
+## 测试
+
+`packages/llm/llm-pi-ai/tests/catalog.spec.ts` 在 resolver 层覆盖了这条链的每一级以及空列表的两种读法：一条路由上未声明的模型与条目声明的纯文本、视觉模型并存，路由默认值为未声明的模型作答而条目仍然压过它，catalog 视觉模型在更窄的路由默认值下保住自身模态，条目的 `[]` 走继承而非清空，以及路由的 `[]` 被拒绝。另有一条用例把每一级端到端复验一遍——写入的 settings 段、插件自身的注册、以及 `ctx.llm.listModels` / `resolveModelInfo`——因此文档与 `LlmModelInfo` 之间若有断点无法蒙混过关。
+
+`config.spec.ts` 负责 schema 边界：两个层级上的未知模态拒绝、路由空列表被 schema 接受而由 settings seam 真正运行的命名空间校验器拒绝，以及继承规则所倚赖的「缺省数组物化为 `[]`」这一事实。
+
+没有任何无密钥 snapshot 通道会跑 pi-ai 路由：snapshot 示例驱动的是 `dsh-llm-replay`，它在自己的配置里直接声明模态，而 pi-ai 路由需要一个真实端点，其端口是静态 `cordis.yml` 无法写出的。本次变更所供给的那些准入点已经通过该提供方在那里得到覆盖（`examples/acp-agent/image.cordis.snapshot.yml` 与 `image-text-route.cordis.snapshot.yml`）且不受影响——改变的是某个适配器报告什么，而非门禁如何读取它。