Add WorkBuddy desktop setup recipe and 2026-09-12 thinking deviations

AGENTS.md

@@ -12,6 +12,7 @@ These instructions apply to maintaining this repository, not consuming apps.
 - `references/api_reference.md`: official endpoint links, not copied field tables.
 - `references/models.md`: selection defaults and current account-source pointers.
 - `references/examples.md`: ECNU-specific integration recipes.
+- `references/workbuddy_setup.md`: WorkBuddy desktop custom-model configuration recipe.
 - `references/workflows.md`: symptom-driven, opt-in diagnostics.
 - `references/known_deviations.md`: dated live evidence; preserve dates and scope.
 - `references/agent_development.md`: historical guidance cited by old observations.

README.md

@@ -38,6 +38,7 @@ credentials should not prevent the Agent from writing or reviewing code.
 | Start a normal integration | [SKILL.md](SKILL.md) |
 | Find the relevant official endpoint contract | [Endpoint map](references/api_reference.md) |
 | Image edits, multimodal retrieval, LangChain embeddings, thinking/tool history, Anthropic SDK | [Integration recipes](references/examples.md) |
+| WorkBuddy desktop custom-model setup | [WorkBuddy recipe](references/workbuddy_setup.md) |
 | Model choice or current account facts | [Model/account pointers](references/models.md) |
 | Diagnose a failure or select a live probe | [Targeted diagnosis](references/workflows.md) |
 | Inspect dated evidence | [Known deviations](references/known_deviations.md) |

SKILL.md

@@ -59,6 +59,7 @@ Its full messages URL ends in `/open/api/anthropic/v1/messages`.
 | Image/text embeddings or rerank | [Multimodal retrieval](references/examples.md#multimodal-retrieval): choose the model and preserve vector-space compatibility |
 | Anthropic SDK setup | [Anthropic recipe](references/examples.md#anthropic-sdk) |
 | Thinking and tool continuation | [Tool-history recipe](references/examples.md#thinking-and-tool-history) plus the linked ECNU contract |
+| WorkBuddy desktop custom-model setup | [WorkBuddy recipe](references/workbuddy_setup.md) |
 | Model choice, prices, quotas, or deployment questions | [Model and account pointers](references/models.md) |
 | An error, unexpected result, or requested live check | [Targeted diagnosis](references/workflows.md); consult only the matching dated deviation |
 

references/examples.md

@@ -172,6 +172,15 @@ when thinking is enabled; thinking defaults to off. Pass ECNU extensions through
 `extra_body` with the OpenAI SDK. Do not substitute upstream template switches
 or silently map an unsupported tier between models.
 
+For clients that cannot send `thinking`, the documented compatibility alias
+`ecnu-reasoner` defaults to enabled thinking on `ecnu-max`; see the
+[WorkBuddy recipe](workbuddy_setup.md) for one client-specific setup.
+
+The 2026-09-12 observations of [effort-only activation](known_deviations.md#ecnu-max-reasoning-effort-as-thinking-trigger)
+and [extra max effort tiers](known_deviations.md#unavailable-reasoning-effort-tiers)
+remain historical evidence. They do not override the current contract above or
+justify rejecting `medium` on `ecnu-plus`.
+
 For a tool exchange, append the complete actual assistant message before the
 matching tool results. ECNU documents preserving `reasoning_content` for
 thinking-mode tool calls through subsequent user turns. Keep it only in process

references/known_deviations.md

@@ -23,6 +23,17 @@ langchain-openai 0.3.35, and httpx 0.28.1. Requests used synthetic text,
 60-second timeouts, no retries, and the same cumulative 50-credit ceiling.
 Only response structure, field-preservation checks, and usage were retained.
 
+Test environment `live-2026-09-12-c` used the account owner's personal tokens
+(two distinct keys), Windows x64, Node.js 24.14.1 `fetch` direct HTTP, and
+Asia/Shanghai local dates 2026-09-12 through 2026-09-14, during an
+owner-authorized interactive integration session rather than the smoke runner.
+Reliability runs were serial with roughly four-second spacing; unspaced bursts
+are called out per entry. Repeated attempts in those runs were deliberate
+characterization samples, not uncertainty retries, and credit consumption was
+not metered against the runner's ceiling. Only synthetic arithmetic and
+vision-fixture inputs were used. Prompts, generated content, reasoning text,
+and credentials were not retained.
+
 ## Invalid bearer on model discovery
 
 - **Tested at:** 2026-08-23; reproduced 2026-09-12
@@ -333,6 +344,101 @@ estimated cumulative consumption was 3.92822 credits. Neither is verified
 account debit. This targeted run does not revalidate other historical cases,
 other SDK versions, streaming, or arbitrary framework adapters.
 
+## ecnu-reasoner alias default thinking
+
+- **Tested at:** 2026-09-12
+- **Environment:** live-2026-09-12-c, Node.js `fetch` direct HTTP
+- **Protocol and endpoint:** OpenAI-compatible `POST /chat/completions`
+- **Documented expectation:** The alias `ecnu-reasoner` equals `ecnu-max` with `thinking: {"type": "enabled"}`; a client must send the `thinking` parameter to activate thinking on `ecnu-max` itself.
+- **Observed behavior:** A bare `ecnu-reasoner` request with no `thinking` field returned `200` and a non-zero `usage.completion_tokens_details.reasoning_tokens` count (range 18-198 across samples). A matching bare `ecnu-max` control returned `reasoning_tokens: 0`. The alias activates thinking server-side without any client-side `thinking` parameter.
+- **Reproduction conditions:** Send two otherwise identical minimal requests, one with `model: "ecnu-reasoner"` and one with `model: "ecnu-max"`, both omitting `thinking`; compare `usage.completion_tokens_details.reasoning_tokens`.
+- **Impact:** A client that relies on sending `thinking` to detect whether reasoning is active will misclassify `ecnu-reasoner` responses. The absence of `message.reasoning_content` is not a reliable thinking indicator either (see the prior "Max thinking response fields" entry).
+- **Recommended fallback:** Detect active thinking via `usage.completion_tokens_details.reasoning_tokens`, not via `message.reasoning_content` or the presence of a client-side `thinking` parameter. Treat `ecnu-reasoner` as always-thinking for routing decisions.
+- **Status:** active
+
+## ecnu-max reasoning_effort as thinking trigger
+
+**Current-contract note (2026-10-02):** The [thinking page](https://developer.ecnu.edu.cn/vitepress/llm/thinking.html)
+still requires explicit `thinking` on `ecnu-max`. The observation below has
+not been rerun against the current deployment; use the documented switch for
+new integrations. Its original test date and status are retained.
+
+- **Tested at:** 2026-09-12
+- **Environment:** live-2026-09-12-c, Node.js `fetch` direct HTTP
+- **Protocol and endpoint:** OpenAI-compatible `POST /chat/completions`
+- **Documented expectation:** The thinking page states that `reasoning_effort` only takes effect when thinking mode is enabled via `thinking: {"type": "enabled"}`; without it, `reasoning_effort` is ignored.
+- **Observed behavior:** `ecnu-max` requests with `reasoning_effort` set but no `thinking` field returned `200` with non-zero `usage.completion_tokens_details.reasoning_tokens`. Dose-response was clean across `low` (median ~117), `high` (~155), and `max` (~343). Adding `thinking: {"type": "enabled"}` alongside `reasoning_effort` did not increase the token count beyond the same-effort baseline.
+- **Reproduction conditions:** Send `ecnu-max` requests with `reasoning_effort` set to `low`, `high`, and `max` but omit `thinking`; record `usage.completion_tokens_details.reasoning_tokens` for each.
+- **Impact:** Clients following the documented precondition may unnecessarily send a `thinking` parameter, or may wrongly conclude thinking is off when only `reasoning_effort` is sent. The documented gating did not match the sampled deployment.
+- **Recommended fallback:** Send the documented explicit `thinking` switch for `ecnu-max`; do not rely on this dated effort-only behavior. For a client that cannot send the switch, the documented `ecnu-reasoner` alias defaults to enabled thinking.
+- **Status:** active
+
+## Unavailable reasoning effort tiers
+
+**Current-contract note (2026-10-02):** Current documented tiers are
+`low` / `high` / `max` for `ecnu-max` and `low` / `medium` / `xhigh` for
+`ecnu-plus`. The September observation below covers max only and has not
+been rerun; it does not make `medium` invalid on plus or `xhigh` a supported
+max default. Its original test date and status are retained.
+
+- **Tested at:** 2026-09-12
+- **Environment:** live-2026-09-12-c, Node.js `fetch` direct HTTP
+- **Protocol and endpoint:** OpenAI-compatible `POST /chat/completions`
+- **Documented expectation:** The thinking page lists `low`, `high`, and `max` as the three supported `reasoning_effort` values for `ecnu-max`.
+- **Observed behavior:** `minimal` returned HTTP 500 on 5 of 7 attempts; `medium` returned HTTP 500 on 7 of 9 attempts. Both occasionally succeeded but cannot be relied on. `xhigh`, although absent from the documentation, returned `200` on all attempts with a reasoning-token median (~178) between `high` and `max`. `low`, `high`, and `max` were stable across all samples.
+- **Reproduction conditions:** Send `ecnu-max` requests with each of `minimal`, `medium`, `xhigh`, `low`, `high`, and `max` as `reasoning_effort`; use serial requests with roughly four-second spacing to separate this check from the rapid-request error observation below.
+- **Impact:** A client that sends `minimal` or `medium` (for example, via an SDK default or a UI selector that does not filter values) will hit intermittent 500 errors. `xhigh` is a usable but undocumented tier.
+- **Recommended fallback:** Use each model's current documented tiers, not the extra tiers accepted by this historical deployment. Reject an unsupported value locally rather than silently remapping it to a different effort.
+- **Status:** active
+
+## Rapid-request 401 metadata failure
+
+- **Tested at:** 2026-09-12
+- **Environment:** live-2026-09-12-c, Node.js `fetch` direct HTTP
+- **Protocol and endpoint:** OpenAI-compatible `POST /chat/completions` and `GET /models`
+- **Documented expectation:** `401` indicates an authentication failure; the official guidance only advises avoiding parallel calls.
+- **Observed behavior:** Rapid sequential requests (under roughly one second apart) to either endpoint returned `401` with `{"detail": "获取第三方元数据失败"}` even though the same bearer passed spaced requests. Spacing requests by roughly four seconds eliminated the failures. The error shape is identical to the unsupported-model case above, making them indistinguishable without a working control.
+- **Reproduction conditions:** Send the same valid bearer to the same documented model in a tight loop (sub-second spacing); then repeat with four-second spacing.
+- **Impact:** A client may mistake this spacing-correlated error for a credential failure and trigger unnecessary key rotation. These observations do not establish a server-side rate-limit cause or whether rejected requests were billed.
+- **Recommended fallback:** Inspect the complete error and chosen model first. If the request was explicitly rejected and a live check is authorized, one spaced control request can test whether the error persists. Do not automatically repeat a possibly accepted POST after a timeout or dropped connection, or rotate keys from this error shape alone.
+- **Status:** active
+
+## Verified coverage on 2026-09-12 through 2026-09-14
+
+**Current-contract note (2026-10-02):** This is historical coverage, not a
+current capability test. The model page now specifies max as text-only with
+512K context; use plus for image understanding and the current per-model
+thinking tiers. No live revalidation was performed for this note.
+
+Environment `live-2026-09-12-c` made characterization requests across two
+personal tokens on Windows x64, Node.js 24.14.1, over Asia/Shanghai dates
+2026-09-12 through 2026-09-14. Requests were serial with roughly four-second
+spacing except where unspaced bursts are noted. No credit metering was applied;
+this was an owner-authorized interactive session, not a smoke-runner batch.
+
+Successful coverage included:
+
+- Both `ecnu-max` and `ecnu-reasoner` returned `200` for `reasoning_effort`
+  values `low`, `high`, `xhigh`, and `max`, with a clean dose-response curve
+  (median reasoning tokens: low ~117, high ~155, xhigh ~178, max ~343).
+- `ecnu-reasoner` bare requests returned non-zero reasoning tokens; `ecnu-max`
+  bare requests returned zero. The alias activates thinking server-side.
+- `ecnu-max` with `reasoning_effort` but no `thinking` parameter activated
+  thinking, contradicting the documented gating.
+- `minimal` and `medium` efforts returned intermittent HTTP 500 and are
+  unreliable; `xhigh` is undocumented but stable.
+- Both models correctly recognized a synthetic 96x96 red-circle PNG and
+  returned the expected "circle red" description, with and without thinking.
+- Function calling with thinking coexisted normally on both models; a
+  two-turn tool exchange without preserved `reasoning_content` returned `200`.
+- Rapid sub-second request bursts returned `401`
+  `{"detail": "获取第三方元数据失败"}`; four-second spacing eliminated it.
+
+No SDK, streaming, long-context, embedding, rerank, TTS, or image-generation
+revalidation was performed in this run. The dose-response medians are
+characterization samples, not quality benchmarks. Older observations retain
+their original dates and statuses.
+
 ## Update rules
 
 1. Record the date, non-secret environment, protocol, endpoint, expectation,

references/models.md

@@ -55,6 +55,12 @@ Upstream model cards can explain model design, but do not establish ECNU's
 request fields, thinking defaults, context units, output limits, performance,
 or deployment path. Do not copy upstream serving flags into requests.
 
+For thinking controls, follow the [current recipe](examples.md#thinking-and-tool-history).
+The [2026-09-12 effort observations](known_deviations.md#ecnu-max-reasoning-effort-as-thinking-trigger)
+are scoped to that deployment; they are not current cross-model defaults.
+The [WorkBuddy recipe](workbuddy_setup.md) covers clients that cannot send the
+explicit thinking switch.
+
 For a cost calculation, fetch current prices and show the input/output and
 cache assumptions. Do not assume a cache-hit ratio or treat estimated usage as
 verified account debit. A code-only task does not require a cost calculation.