codex-workflow-v2 2.0.0-beta.13.2 → 2.0.0-beta.13.3

package/plugins/codex-workflow-gateway/skills/codex-workflow-gateway/SKILL.md

@@ -329,9 +329,11 @@ A Milestone chat and its Task chats are separate user-visible execution contexts
 conversation. Create every Task chat as a brand-new Codex task with `create_thread`, an explicit
 Task-only prompt, and an explicit title. Never use `fork_thread`, thread handoff, or a continuation
 of the Milestone chat to create a Task chat: those paths carry parent history into the child. A
-failure to create a new task is a routing blocker. Check the thread list for an already-created
-exact Task ID, retry `create_thread` once sequentially when none exists, and then stop; never fall
-back to a fork.
+confirmed failure to create a new task is a routing blocker. A pending `clientThreadId`, timeout,
+lost response, or absence from `list_threads` does not establish failure and never permits a
+second creation. Persist the attempt before calling the App, resolve its actual ID and confirm
+it with `read_thread`; never fall back to a fork. Read and follow
+`../../references/chat-dispatch.md` for the executable dispatch/recovery and model-selection contract.
 
 Codex App may wrap the supplied prompt in a host-generated `codex_delegation` envelope containing
 only routing provenance such as `source_thread_id`. That envelope is permitted and is not parent
@@ -376,6 +378,22 @@ chat type, Milestone/Task/Step membership ordinals, semantic title, authoritativ
 review attempt. Use returned `requestedTitle` verbatim. Never derive its prefix from
 `list_threads`, sidebar order, a count, or `count + 1`: allocation atomically reserves the
 project-wide monotonically increasing `#NNN`, and an abandoned number is never reused.
+Allocation reuses an existing matching reservation. For a new dispatch, call `dispatch-begin`
+with the closed packet and explicit model request, then call the App exactly once only when
+`createAllowed=true`, using returned `createArgs` unchanged. Record the full result via
+`dispatch-result`, resolve/read back via `dispatch-discover` and `dispatch-observe`, and bind.
+After restart use `dispatch-status`; an unresolved creation cannot be abandoned to create another.
+For legacy reservations without a dispatch marker, first reconcile existing App/source evidence;
+absence of the new marker does not prove no earlier creation occurred.
+
+When the user authorizes model selection, use the versioned role/phase/complexity matrix rather
+than coordinator inheritance. Both `model` and App `thinking` must be explicit in create arguments.
+Use the destination host's current supported model/effort catalog; no guessed aliases or silent
+fallback. Before a materially different phase, `model-select` returns explicit follow-up settings
+for the same bound thread. Routine lifecycle work should not retain planning/review reasoning by
+accident; security-sensitive and complex work must not be classified as routine to save tokens.
+Without user authorization for model selection, do not manufacture it from a registry field;
+resolve that policy choice before starting a new dispatch.
 
 The builder owns these compact title families; callers must not hand-compose variants:
 
@@ -397,7 +415,7 @@ Put `#NNN` and membership ordinals first so clipped sidebar titles remain distin
 inherit the parent title, use a generic title such as `Milestone recovery` / `Task execution`, or
 reuse one title for different entity IDs. Maintain the registry-backed
 `Task ID -> thread ID -> title` mapping in
-the Milestone chat. Before dispatch, verify from the thread list that the title contains the exact
+the Milestone chat. Before supervision, verify from actual `read_thread` evidence that the title contains the exact
 ordinal and entity ID, is unique, and the new chat input contains only its TaskContextPacket plus
 the permitted host routing envelope. Never trust the `create_thread` title argument without a
 readback: pass the observed title to registry `readback`. If it returns `renameRequired=true`,
@@ -406,6 +424,12 @@ compact the role while preserving `#NNN`, membership ordinals, attempt, and the
 Never invent another fallback or abbreviate the ID. Bind the verified entry to `threadId` and
 `hostId`; if all candidates fail, mark it blocked and stop routing. Visual sidebar clipping is
 harmless only when exact API readback still matches. Do not dispatch while title/context is wrong.
+Persist the single supervisor, cursor and next action using `supervision-record`. Large tool/CLI
+JSON must be saved completely to a private local file and parsed there, with only bounded non-secret
+summaries printed. For an uncertain mutation outcome, run fresh `status -> next` before recovery;
+never replay a mutation merely because displayed output was truncated. Review the exact serialized
+external-review envelope passed to Core, not a separately reconstructed object. Public CREF expiry
+is independent of lease expiry and must be checked before every credential-consuming transition.
 
 After every significant Task Step boundary—completion, failure, block, skip, sealed-review result,
 or corrective decision—the Task chat prints a compact Task progress Markdown table from fresh