feat: paginate candidate model recommendations

capabilities/am516-delivery-prediction/AGENTS.md

@@ -34,6 +34,12 @@ eRob90H100I-FHM-18CT[V6] 50
 
 If multiple models are provided, stop and ask AR51 to run one model at a time.
 
+After a valid query has produced an initial candidate batch, the same sender in
+the same Feishu chat may ask whether there are other candidates without
+repeating the model and quantity. This continuation is valid only while the
+runtime still holds that query context. It continues from the next unqueried
+candidate; it does not create a new model or quantity assumption.
+
 ## 4. API Rule
 
 Use only:
@@ -83,6 +89,7 @@ Include:
 
 - input model and quantity;
 - candidate summary;
+- current candidate batch and remaining-candidate count;
 - predicted lead time or API exception;
 - main factor;
 - model difference;

capabilities/am516-delivery-prediction/docs/acceptance_samples.md

@@ -28,15 +28,50 @@ Expected:
 - reducer shortages are the sole exception: when `缺料物料` starts with `20.30.`, output only the reducer code, inventory, inspection, in-transit, existing-demand quantities, and in-transit coverage conclusion in the Skill-defined order;
 - shortage items outside the `20.30.` reducer prefix never expose their material codes or quantity details;
 - result is marked as internal candidate;
-- all Skill-permitted brake, encoder, sensor, grease, and DZ/LF combinations are generated without changing strong constraints or version;
-- every generated candidate is queried through the controlled script; usable candidates are detailed and unsupported candidates are summarized under `不可用候选`;
-- when the full report exceeds one Feishu rich-text message, it is split into ordered continuation posts and no candidate is dropped;
+- all Skill-permitted brake, encoder, sensor, grease, and DZ/LF combinations are generated in a stable rule order without changing strong constraints or version;
+- the original model is queried as a baseline and does not count toward the recommendation batch;
+- the initial reply queries only candidate positions 1-4, details usable candidates, summarizes unsupported candidates under `不可用候选`, states the remaining count, and invites the human to ask `还有没有其他候选型号`;
+- candidate position 5 and later are not queried or displayed in the initial reply;
+- when the current four-candidate batch exceeds one Feishu rich-text message, it is split into ordered continuation posts and no candidate from that batch is dropped;
 - record is appended.
-- the appended record contains query metadata, model parsing, and a complete Markdown table with one row per queried candidate;
+- the appended record contains query metadata, model parsing, and a complete Markdown table with one row per model queried in the current batch;
 - each future appended candidate row records the same query-completion time in `YYYY-MM-DD HH:mm:ss +0800` form (with the actual local UTC offset), without rewriting or backfilling historical records;
 - record-table columns follow this order: 序号、记录时间、型号、推荐类型、差异项、API状态、AM516预测交期、可用库存、主导因素;
 - unavailable candidates remain in their original rule order and are recorded as `❌ 不存在` with unavailable result fields shown as `—`;
 
+## Sample 1A | Continue With Other Candidates
+
+Preconditions:
+
+- Sample 1 has completed in the same Feishu chat for the same sender;
+- more than four rule candidates exist;
+- the bridge has not restarted and the initial-query context is less than six hours old.
+
+Input:
+
+```text
+还有没有其他候选型号？
+```
+
+Equivalent inputs such as `还有吗`、`有没有别的`、`再推荐几个`、`继续`、`下一批`
+or `more candidates` must produce the same continuation behavior; exact wording
+is not required.
+
+Expected:
+
+- the bot reuses the active model and quantity only for that same sender and chat;
+- the bot queries candidate positions 5-8 in the unchanged rule order;
+- the original model and candidate positions 1-4 are not queried or displayed again;
+- unavailable candidates still consume their original positions and are summarized under `不可用候选`;
+- the reply states the current candidate range and the remaining count, or states that all rule candidates have been queried;
+- one new record block contains only the models queried in this continuation batch;
+- the fixed disclaimer appears once at the end.
+
+If the same continuation text comes from another sender or chat, after a bridge restart, after the six-hour expiry, or without a valid initial query, the bot stays silent and makes no API request or record append.
+
+Unrelated expressions such as `more interesting`、`继续查询英语例句`、`下一课` or
+`再来一首歌` do not qualify as candidate continuation, even while a query context exists.
+
 ## Sample 2 | Missing Quantity
 
 Input:

capabilities/am516-delivery-prediction/docs/deployment_checklist.md

@@ -44,6 +44,8 @@
 - [ ] After the approval gate is satisfied, at least one valid model + quantity query succeeds; otherwise the blocked outbound state is recorded without attempting the request.
 - [ ] Missing API key returns a structured error without leaking secrets.
 - [ ] Missing, invalid, multiple, or ambiguous model/quantity input stays silent and makes no API call or record append.
+- [ ] The initial response queries no more than four recommended candidates in addition to the original-model baseline; same-sender, same-chat continuation queries the next four without duplication.
+- [ ] Candidate continuation without matching active context stays silent; a bridge restart or six-hour expiry requires the complete model and quantity again.
 - [ ] Output includes the fixed internal-candidate disclaimer.
 - [ ] Records append to `records/delivery_prediction_records.md`.
 - [ ] A direct AR51 task in Mac Studio Codex can modify a test project file or perform the intended Git update without being blocked by the Feishu read-only rule.

capabilities/am516-delivery-prediction/docs/runbook.md

@@ -7,12 +7,22 @@
 3. Confirm AR51 has approved the functional test.
 4. Confirm a separate trusted AR51 approval record explicitly covers `api.zeroerr-agent.com`, the AM516 use case, and API-key use.
 5. If that domain-and-permission record is absent, stop before any outbound request; a general functional-test approval is insufficient by itself.
-6. Use model recommendation rules to generate internal candidates.
-7. Run the controlled script for the original model and approved candidates.
-8. Summarize the internal candidate result.
+6. Use model recommendation rules to generate the complete ordered internal candidate pool.
+7. Run the controlled script for the original model baseline and only the first four candidate models.
+8. Summarize the current batch, state how many candidates remain, and invite the same sender to ask naturally for more when candidates remain; exact wording is not required.
 9. Append a non-secret record.
 10. Return Feishu-friendly text with the fixed disclaimer.
 
+## Candidate Continuation
+
+1. Accept a clear request for more candidates only from the same sender in the same Feishu chat as the active initial query. Equivalent expressions such as `还有吗`、`有没有别的`、`再推荐几个`、`继续`、`下一批` or `more candidates` are valid; no exact phrase is required.
+2. The in-memory continuation context expires after six hours and is cleared by a bridge restart. If no matching context exists, stay silent and require a new complete model-and-quantity query.
+3. Query the next four unqueried rule candidates; do not re-query the original baseline or any earlier candidate.
+4. Missing or unsupported candidates still consume their rule-order positions and remain under `不可用候选`.
+5. Append one record block for the models queried in this continuation batch, then state whether more candidates remain.
+6. A new complete model-and-quantity request replaces the earlier continuation context for that sender and chat.
+7. Keep the dual gate of continuation intent plus valid context: unrelated text such as `more interesting` or `继续查询英语例句` must not trigger the old query.
+
 ## Approved Script
 
 ```text

capabilities/am516-delivery-prediction/skills/joint-module-leadtime-report/SKILL.md

@@ -49,6 +49,16 @@ Stop if:
 - An @mention by itself does not bypass this content gate.
 - After a valid message passes the gate, API or runtime failures may still return
   the bounded internal error response defined by this capability.
+- Exception for candidate continuation: after a valid initial query, the same
+  sender in the same Feishu chat may use any clear equivalent continuation
+  expression without repeating the model and quantity. Examples include `还有吗`、
+  `有没有别的`、`再推荐几个`、`继续`、`下一批` and `more candidates`; exact wording
+  is not required.
+- A continuation uses only the active in-memory query context, expires after six
+  hours or a bridge restart, and is never shared across senders or chats. Without
+  matching active context, the message stays silent.
+- A new complete model-and-quantity request replaces any earlier continuation
+  context for that sender and chat.
 
 ## Feishu File-Write Boundary
 
@@ -83,6 +93,26 @@ Do not use temporary curl, WebFetch, or any API method outside the controlled sc
 - If AM516 fields are absent, use original API prediction and mark it as original algorithm reference.
 - If only total days are available, state the total days and ask for human review.
 
+## Candidate Batch Rule
+
+- Generate and preserve the complete ordered rule-candidate pool, but query and
+  recommend at most four candidate models per reply.
+- The original input model is queried as the initial baseline and does not count
+  toward the four recommended candidate models.
+- The initial reply queries the baseline plus candidate positions 1-4. If more
+  candidates remain, state the remaining count and invite the human to ask for
+  more in natural language; do not require one fixed command phrase.
+- Recognize short or synonymous continuation expressions only together with the
+  valid same-sender, same-chat continuation context. Unrelated topic text such as
+  `more interesting` or `继续查询英语例句` must not reuse the old delivery query.
+- Each valid continuation queries the next four unqueried candidates in the same
+  rule order. Do not query or display an earlier candidate again, and do not
+  re-query the original baseline.
+- Continue in batches of four until the full candidate pool is exhausted. The
+  final batch explicitly states that all rule candidates have been queried.
+- Missing or unsupported models count as queried positions and remain summarized
+  under `不可用候选`; do not silently skip them to pull later candidates forward.
+
 ## Reducer Shortage Detail Rule
 
 - Treat a shortage item as a reducer only when `data.采购信息.缺料明细[].缺料物料` starts with `20.30.`.
@@ -105,6 +135,7 @@ Include:
 - title and generation time;
 - input model and quantity;
 - overall judgment;
+- current candidate batch and remaining-candidate count;
 - candidate details;
 - main factor;
 - API exception if any;
@@ -120,6 +151,7 @@ Use this field order for a successful Feishu report:
 数据时间戳：{API data timestamp or generation time}
 状态：试运行版输出 / 内部候选
 判断等级：AM516算法预测参考，需人工复核
+候选范围：本次先推荐4个候选型号（原型号另作基准）；尚有N个规则候选未查询
 
 总体判断：
 - 现货候选：{fully covered candidate or none}
@@ -148,12 +180,13 @@ Use Feishu rich-text `post` output. Bold only:
 
 - the report title;
 - `总体判断：`;
+- `本批次判断：` for a continuation reply;
 - each numbered candidate heading;
 - `不可用候选：`.
 
 Do not repeat a separate human-review line under each candidate. The global judgment level and fixed disclaimer carry the human-review boundary.
 
-If the full candidate report exceeds the Feishu rich-text size limit, split it into multiple ordered `post` replies without dropping any queried candidate. Use the original report title on the first part and append `（续）` to continuation titles.
+If the current candidate-batch report exceeds the Feishu rich-text size limit, split it into multiple ordered `post` replies without dropping any candidate queried in that batch. Use the original report title on the first part and append `（续）` to continuation titles.
 
 ## Fixed Disclaimer
 

capabilities/am516-delivery-prediction/skills/joint-module-model-recommendation/SKILL.md

@@ -62,8 +62,10 @@ Allowed upgrades:
 - Encoder options are directional: `S -> M/HS/HM`, `M -> HS/HM`, `HS -> HM`, and `HM` has no further automatic upgrade.
 - Include the original model separately as the baseline and deduplicate all generated models.
 - Never generate a brake, encoder, sensor, or grease downgrade.
-- Query every generated candidate through the controlled delivery-prediction script.
-- Show candidates with usable lead-time summaries in full; summarize missing or unsupported models under `不可用候选`.
+- Preserve the complete deduplicated candidate pool in rule order so the lead-time capability can page through it deterministically.
+- The original model is a baseline and does not count toward the four recommended candidate models in a batch.
+- When used by the lead-time capability, pass only the first four candidates on the initial query. A same-conversation human request for more candidates passes the next four, without repeating or reordering earlier candidates.
+- This recommendation Skill does not call the delivery-prediction API itself. The lead-time capability queries only the current candidate batch and summarizes missing or unsupported models under `不可用候选`.
 
 ## Output