feat(delegate): 清單加一欄記唯讀盤點的實際指令

切片交的那幾支，助理現在只會提醒不會動手，因為沒有東西記得下「那一段唯讀盤點到底要跑什麼」。清單加第十二欄記具體指令，種入那一支有值就拿它當動作、沒值才退回只提醒。

只認三種寫法。一行指令，路徑一律寫成代入點開頭，由種入那一支代進字面絕對根目錄；指令裡不可以出現金錢符號或波浪號，那兩種在無人值守那一輪解不出來也進不了允許清單，會被靜靜擋掉。要標未接線就寫理由，留白的話下一輪分不出是刻意還是漏填。沒有唯讀入口的寫減號。

觸發型與不交的列一律填減號。觸發的意思是呼叫整支技能，這一欄填了指令會讓種入那一支改拿指令當動作，於是整支交出降級成只跑一支腳本，該寫的頁一頁都不會寫，而且看起來完全正常。這條由檢核腳本擋。

逐項回各存放庫核對，不照抄既有盤點的措辭，因而抓到五處對不上實際腳本的地方。

最要緊的一處：既有盤點點名的那支工具腳本要三個參數，其中兩個無人值守那一輪根本拿不到，而且它是唯一會去問遠端的一方。改用同一件事的本機那一半，而那個子命令剛好在閘門腳本「免讀標準輸入」的清單裡。這一點非確認不可：同一支腳本的另一個子命令不在那份清單上，標準輸入是管線又沒人關閉時會一直等，實測會無限卡住。腳本自己的註解就記著這個坑，說工具腳本轉呼叫時曾經整支卡死——那正是先前心跳斷掉的同一種死法。

另一處：既有盤點點名的同步子命令會寫檔，不是唯讀。那一列剛好是觸發型所以填減號、衝突沒落地，但措辭本身是錯的。

七項標成未接線，理由都是要連網要金鑰。連網那一種一過期就讓那一項每輪失敗，或每輪靜靜回報沒事——後者更難查；而純本機讀取本來一輪都不會失敗。先填會連網的，等於用一批每輪報錯的項目把真的發現蓋掉。

還記下一件接線那天要注意的事：查遠端版本那一支永遠回成功，查不到就安靜放行，所以金鑰失效時它會靜靜回報沒有新版。接線要用另一個子命令，那個有「查不出來」這第三種結論。

檢核腳本從四項檢查改成五項。填錯欄位、指到不存在的腳本、用了認不得的代入點都算缺失；標未接線只印提示，那是判過知道還沒接，不是漏填；指到的 domain 本機沒裝也只提示，那是機器少裝一套不是清單填錯，算缺失會讓半套機器每輪亮紅。

五支技能異動流程一併改：寫清單時要問到並寫下這一欄，重判時要一併重問，刪除時要檢查別列有沒有指到一起刪掉的腳本，一次改多支時要注意跨存放庫搬腳本是這一欄最容易過期的地方。

skills/skill-new/SKILL.md

@@ -20,7 +20,9 @@ Single source of guidelines: [`../../references/guidelines.md`](../../references
       - Trigger (when to use, when not to, trigger keywords)
       - Input and output (can a standard input/output flow move down to `tools/`; does it need Gitea operations — if so, make the skill use `jsc-gitea/tools/gitea.sh` + token)
       - Owning domain (offer the domain list from the step 1.1 `domain<TAB>path` rows — the domains registered in the canonical marketplace)
-      - Delegation verdict — the five decision-tree questions of [`../../references/delegate-criteria.md`](../../references/delegate-criteria.md), in the order that file lists them, plus a sixth question for the `next` column: which skill should run after this one. Ask all six through this same `jsc-ask:ask` tree; never answer them from the model's own reading of the draft flow. Every option states its impact scope: a `full` verdict lets the background assistant run the skill unattended, a `slice` or `cond` verdict leaves the other half in the user's hands, `none` keeps the whole skill there. The `next` question applies to all four verdicts, `none` included — `none` says the assistant does not run this skill for the user, which says nothing about what should follow it — so offer the step 1.1 skill rows as its options and the answer then names a skill that exists.
+      - Delegation verdict — the five decision-tree questions of [`../../references/delegate-criteria.md`](../../references/delegate-criteria.md), in the order that file lists them, plus a sixth question for the `next` column: which skill should run after this one, and a seventh for the `probe` column: which read-only command the assistant actually runs for the delegated slice. Ask all seven through this same `jsc-ask:ask` tree; never answer them from the model's own reading of the draft flow. Every option states its impact scope: a `full` verdict lets the background assistant run the skill unattended, a `slice` or `cond` verdict leaves the other half in the user's hands, `none` keeps the whole skill there. The `next` question applies to all four verdicts, `none` included — `none` says the assistant does not run this skill for the user, which says nothing about what should follow it — so offer the step 1.1 skill rows as its options and the answer then names a skill that exists.
+
+        The `probe` question is asked only when the `way` answer holds no `invoke`; a `way` containing `invoke`, and a `none` verdict, both take `-` without asking, because the assistant's action there is the skill itself and a command in that column would silently downgrade the whole delegation to a bare script run. When it is asked, settle three things per that same reference file and never by copying another row: which script, whether it takes a read-only flag, and how a failure is reported. Verify the script, the sub-command and the flag name in the owning repo before writing them down — a description of a script is not evidence the script exists. The flag matters most: the patrol round is unattended, and a wrong flag turns a read-only stocktake into something that writes. An entry that needs the network gets `pending:{reason}` this round rather than a command, because an expired key then fails or silently reports nothing on every round while a local read fails on none.
 
       Completion condition: goal, trigger, input/output, owning domain and the delegation verdict each have a recorded answer, and the verdict carries a value for every column `delegate-criteria.md` marks mandatory for that verdict, `-` where it marks the column unused.
 2. If the domain does not exist (`tools/sync-domains.sh` clones every domain **registered in the marketplace**, so a missing directory means the domain is unregistered — the repository itself may already exist on Gitea):
@@ -37,15 +39,15 @@ Single source of guidelines: [`../../references/guidelines.md`](../../references
 3. Generate the skill per guidelines.md — this step MUST run as a sub agent:
    - `skills/{name}/SKILL.md`: entirely in English (description within either cap — ≤ 5 sentences or ≤ 5 steps — and stating when to use and when not to; body in STE100-style English)
    - Rules enforceable by hooks go to `jsc-hooks` (never scattered in this domain); standard input/output flows go to `tools/`
-   - `jsc-meta/tools/delegate-spec.tsv`: append this skill's row, built from the step 1.2 verdict — one skill one row, the eleven tab-separated columns in the order that file's header lists. A column the verdict does not use holds a single `-`; an empty cell and a cell holding a space both fail the checker. `origin` is `judged`, because the verdict came from the decision tree in this same run, and `version` is the version the three manifests carry after the `sync-skill-manifest.sh` run below. **A skill with no row is not created.** The row is the only thing that tells the background assistant this skill exists, so without it every later round is blind to it, and no later step recreates it. The file lives in `plugins/meta` whichever domain gained the skill, so a skill added to another domain changes two repos and step 5 opens the second Push Request for this one.
+   - `jsc-meta/tools/delegate-spec.tsv`: append this skill's row, built from the step 1.2 verdict — one skill one row, the twelve tab-separated columns in the order that file's header lists, `probe` last. A column the verdict does not use holds a single `-`; an empty cell and a cell holding a space both fail the checker. Write `probe` as the header describes: the script path starts at `{root}/jsc-{domain}/`, `{cli}` and `{repo}` are the only other substitution points, no dollar sign and no tilde, and any read-only flag goes in front of the command. `origin` is `judged`, because the verdict came from the decision tree in this same run, and `version` is the version the three manifests carry after the `sync-skill-manifest.sh` run below. **A skill with no row is not created.** The row is the only thing that tells the background assistant this skill exists, so without it every later round is blind to it, and no later step recreates it. The file lives in `plugins/meta` whichever domain gained the skill, so a skill added to another domain changes two repos and step 5 opens the second Push Request for this one.
    - `references/behaviors.md`: add one `## {name}` section for the new skill, placed in dictionary order among the existing sections, carrying the five rows the guidelines' 「技能行為清單」 section defines — 觸發時機、關鍵步驟、外部呼叫、完成條件、可驗證跡象. Write what the skill really does; do not copy the `description`. A read-only skill still fills 可驗證跡象 with 「無寫入跡象，只有回報內容」. The behavior list ships in this same PR — a skill added without its section leaves the domain's list out of sync the moment this PR merges. When the domain has no `references/behaviors.md` yet, create it with the header line `# jsc-{domain} 技能行為清單`.
 
    Then run `tools/sync-skill-manifest.sh {domain-path}` directly (no sub agent needed) to sync the domain README's 「Skills 目錄」 section and bump the version in all three manifests. Route each exit code: 0 — the README block and all three manifests are synced; 1 — the domain path, `skills/`, `README.md`, the `JSC-SKILLS` markers, a `SKILL.md`, a manifest, or a manifest `version` field is missing, so fix the named cause on stderr and rerun; 2 — usage error, the script takes exactly one argument; any other code — the script runs under `set -e`, so treat it as an environment fault and stop, never as a successful sync. Completion condition: `skills/{name}/SKILL.md` exists, `references/behaviors.md` holds a `## {name}` section with all five rows filled, `tools/delegate-spec.tsv` holds this skill's row with every mandatory column filled, the README lists the skill, and all three manifests show the same new version.
 4. Self-check every item of the guidelines.md audit checklist; fix anything that fails. Run `tools/check-behaviors.sh {domain-path}` for the behavior-list item instead of comparing by eye, and route each exit code: 0 — the list matches `skills/` and all five rows are filled; 1 — every mismatch is printed on stderr as `{檔案}:{技能名}:{說明}`, so fix each one and rerun; 2 — usage error, the tool takes exactly one argument; 3 — nothing was checked, because `references/behaviors.md` is missing, `skills/` is missing, or no `SKILL.md` was found, so create the missing file and rerun. **Exit 3 is never a pass.**
 
    Then run `tools/check-delegate.sh` for the delegation-list item. It takes the plugins root, not a domain path, and the list is one file covering every domain, so it runs **once for the whole flow** — a second run per domain checks the same file again and reports the same lines. Route each exit code:
-   - 0 — the list matches the skills on this machine and every mandatory column is filled. **A run that printed lines on stdout and exited 0 still passed.** Those lines are hints, not defects: `origin=seed` marks a row seeded from the earlier inventory and awaiting review, and a version-behind line marks a row whose `version` trails its domain's current one. The version number is per domain, so bumping one skill's domain marks every other skill in it — reading those lines as failures paints the whole domain red on every release until nobody reads them at all. Report the hints, fix nothing for them, and treat this item as passed.
-   - 1 — a missing row, a duplicate row, a row for a skill this machine does not have, an empty column, a column value outside its vocabulary, or a `next` pointing at a skill that does not exist. Every one is printed on stderr as `{清單路徑}:{domain}/{技能名}:{說明}`; fix each and rerun. The new skill's own missing row is the expected finding when step 3 skipped its write, and the fix is that write, not an edit here.
+   - 0 — the list matches the skills on this machine and every mandatory column is filled. **A run that printed lines on stdout and exited 0 still passed.** Those lines are hints, not defects: `origin=seed` marks a row seeded from the earlier inventory and awaiting review, and a version-behind line marks a row whose `version` trails its domain's current one, a `probe=pending:` line marks a delegated slice whose read-only entry point is not wired yet, and a line saying a `probe` domain is not installed here marks a script this machine cannot check. The version number is per domain, so bumping one skill's domain marks every other skill in it — reading those lines as failures paints the whole domain red on every release until nobody reads them at all. Report the hints, fix nothing for them, and treat this item as passed.
+   - 1 — a missing row, a duplicate row, a row for a skill this machine does not have, an empty column, a column value outside its vocabulary, a `next` pointing at a skill that does not exist, or a `probe` in the wrong shape — a command on a row whose `way` holds `invoke`, a `-` on a row whose `way` holds only `patrol` or `remind`, a dollar sign or tilde, an unknown substitution point, or a script that does not exist. Every one is printed on stderr as `{清單路徑}:{domain}/{技能名}:{說明}`; fix each and rerun. The new skill's own missing row is the expected finding when step 3 skipped its write, and the fix is that write, not an edit here.
    - 2 — usage error: the script takes at most one argument. Fix the call and rerun.
    - 3 — nothing was checked, because `tools/delegate-spec.tsv` is missing, the root could not be derived, or `list-skills.sh` listed no skill. Read stderr and fix the named cause; set `JSC_PLUGINS_ROOT` to the directory holding the domain repos for the root case, as in step 1.1. **Exit 3 is never a pass** — it means the check looked nowhere, so a new skill with no row would sail through it.