docs(automation): 屏幕流 resume 契约与审批越权标记的文档 + 两份 changeset（#4502 跟进） (#4517)

* docs(changeset): changesets for the stranded-request inspection and the harness store (#4469, #4470)

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017gEHJN2NFpS9VMeURvakgD

* docs(automation): document the screen resume contract and the override marker (#4477, #4466, #4469)

flows.mdx — a new subsection on the durable-pause seam, beside the existing
`$`-namespace rule it mirrors: a `screen` node's `config.fields` is a contract,
`resume` enforces it server-side with 400 `INVALID_SCREEN_INPUT`, `visibleWhen`
is evaluated against the submitted values so a hidden field's `required` never
fires, the refusal precedes consuming the suspension, and the three shapes that
declare no contract keep the pass-through. The pausing-node table's `screen` row
now says the inputs are validated rather than merely posted.

approvals.mdx — the admin-override callout now states the rule it actually
implements ("the actor is an admin", not "the slate is unstaffed"), which is why
`sys_approval_action.via_override` records which door the decision came through,
including why an admin who IS an approver records `false` and why a legacy row's
absent value reads as "not recorded". The dead-run callout gains the terminal
request shape its pending-only scan cannot see, and why the new inspection
reports rather than rewrites.

content/docs/references/ is generated and untouched.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017gEHJN2NFpS9VMeURvakgD

* docs: point the "filed separately" references at #4518

Both the durable-suspend proof's KNOWN GAP note and the `@objectstack/verify`
changeset told the reader the harness persistence gap was "filed separately"
without naming it — a pointer to nothing, which is the shape of stale record
this branch exists to avoid. It is #4518.

Comment/changeset text only; no behaviour change.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017gEHJN2NFpS9VMeURvakgD

---------

Co-authored-by: Claude <noreply@anthropic.com>

Author: os-zhuang

.changeset/approvals-stranded-request-inspection.md

@@ -0,0 +1,48 @@
+---
+"@objectstack/plugin-approvals": patch
+---
+
+fix(approvals): find the zombie requests nothing was looking at (#4469)
+
+#4460 stopped new zombies being produced; the rows already stuck had no mechanism
+to find or release them. The failure shape (#4420) is a request flipped to
+`approved` / `rejected` / `returned` whose `flow_run_id` points at a run that no
+longer exists — the decision landed, the flow never moved. Any deployment on
+17.0.0-rc.1 that hit the wiring hole and crossed a restart mid-approval can be
+carrying these rows.
+
+`releaseDeadRunRequests` could not see them, and the reason is worth stating
+plainly: it scans `status: 'pending'`, and the very step that zombifies a request
+is the one that takes it OUT of `pending`. The act of breaking it removed it from
+the only sweeper's field of view — a large part of why this class of failure
+stayed silent. It could not have answered the question even if it had looked: its
+liveness oracle is `getRun`, which reads the execution LOG and returns `null` for
+a perfectly ALIVE suspended run after a restart. It treats `null` as alive
+(conservative, and correct for what it does) — which is exactly why it has no way
+to say "this run is really gone".
+
+Adds `ApprovalService.inspectStrandedRequests()`, which uses BOTH oracles and
+reports only rows that fail both:
+
+- `hasSuspendedRun(runId) === false` — the suspension store itself says no live
+  pause exists. It THROWS when the store cannot be read, and that case is
+  SKIPPED and counted as `undetermined`, never condemned: an unreadable store
+  means "unknown", and a storage outage must not be published as a lost run.
+- `getRun(runId) == null` — no terminal history row either. A run that merely
+  finished is not stranded; a request whose run neither waits nor ever completed
+  is.
+
+**It reports; it never rewrites.** No status is changed and no run is cancelled.
+The decision genuinely happened — a human approved or rejected — and silently
+rolling it back would make the audit trail disagree with the facts. The report
+carries what an operator needs to decide: which requests are stuck at which step,
+and what the mirrored status field on the business record still reads (usually
+the stale value the user is staring at). Whether to re-run the downstream actions
+or re-open the approval is a judgement call this cannot make.
+
+It rides the existing escalation/dead-run sweep clock, so the finding surfaces in
+the logs without an operator knowing to go looking for it. `recalled` is
+deliberately out of scope: a recall abandons its run on purpose, and reporting
+those would bury the real findings under expected ones.
+
+New export: `StrandedApprovalRequest` (the report row shape).

.changeset/verify-harness-durable-suspended-runs.md

@@ -0,0 +1,42 @@
+---
+"@objectstack/verify": minor
+---
+
+fix(verify): stop the harness pinning `suspendedRunStore: 'memory'` (#4470)
+
+`bootStack` hardcoded `suspendedRunStore: 'memory'` when it registered
+`@objectstack/service-automation`. That made the DB-backed suspended-run store
+**structurally unreachable** from every dogfood/e2e fixture — not under-tested,
+untestable. The coverage map had a clean seam nothing crossed:
+
+- unit tests covered ENGINE-side persistence (`suspended-run-store.test.ts`
+  drives suspend → restart → resume against a fake table);
+- e2e covered the BUSINESS chain (approvals), but single-process and wholly in
+  memory;
+- the ASSEMBLY between them — is `sys_automation_run` registered, is its table
+  created, is the store actually attached to the engine — was covered by
+  neither.
+
+#4420 grew in precisely that seam: the store hung off a table that was never
+created, every write failed into a `warn` nobody read, the pause reported
+success, and the run died at the next restart. #4460 added assembly unit tests;
+this makes the e2e half possible.
+
+The harness now boots the plugin's own `'auto'` default — the same wiring
+`objectstack dev` / `serve` get — so fixtures exercise the real assembly. Two
+new knobs:
+
+- `automation` accepts `{ suspendedRunStore: 'auto' | 'memory' }` as well as
+  `true`, so a fixture that wants the old in-memory behaviour asks for it
+  explicitly rather than getting it by default.
+- `databaseFile` backs the in-process SQLite database with a file instead of
+  `:memory:`, so state can outlive a kernel.
+
+Answering the question the issue raised — was `'memory'` pinned for speed or
+because persistence could not run there? **Speed/simplicity.** The durable path
+works in this harness: the accompanying dogfood proof boots with it, and the
+whole existing dogfood suite passes on it unchanged (38 files, 239 tests). Note
+`databaseFile` does not yet deliver a true cold boot: a second `bootStack` over
+the same file reads a database whose tables exist but whose rows are gone —
+ordinary records do not survive it either, so it is a harness/driver persistence
+gap rather than anything to do with suspended runs, and it is filed as #4518.

content/docs/automation/approvals.mdx

@@ -477,6 +477,19 @@ approver, or **recall** it — releasing the lock. An admin decision is
 authoritative: it finalizes the node even under `unanimous`/`quorum`/`per_group`,
 and is audited under the admin's own id. Prefer a guaranteed-staffed fallback
 approver so the set is never empty in the first place.
+
+Note the rule is **"the actor is an admin"**, not "the slate is unstaffed" — so
+an admin can also act on a request whose slate *is* properly staffed, bypassing
+the people on it. That is why the decision records **which door it came
+through**: `sys_approval_action.via_override` is `true` when the actor was
+admitted *only* by this privileged path, holding no slot themselves. An admin who
+is also a designated approver is approving normally and records `false` — the
+flag is about the branch that authorized the call, not about who holds admin
+rights. A row written before the column existed carries no value at all, which
+reads as *not recorded* rather than as *not an override*. Without it, an
+override and an ordinary approval were byte-for-byte identical, and the only
+trace that a slate had been bypassed was the designated approver's later
+`409 INVALID_STATE` — if they happened to try.
 </Callout>
 
 <Callout type="info">
@@ -491,6 +504,18 @@ The sweep only ever acts on a run it can positively confirm is terminal: a
 paused run (the normal state of a live approval), an unknown run, or an
 unreachable automation engine all count as *alive* and are left untouched. It
 frees orphaned records; it never cancels a live approval.
+
+That sweep scans **pending** requests, which leaves one shape outside it: a
+request already *decided* — `approved`, `rejected` or `returned` — whose run has
+since vanished. The decision landed and the flow never moved, and flipping the
+request out of `pending` is precisely what removed it from the sweep's view. A
+second, **read-only** inspection rides the same clock for those: it reports a
+terminal request only when the suspension store says no live pause exists **and**
+no terminal run record exists either, skipping (never condemning) any row whose
+store could not be read. It deliberately **does not rewrite** them — the decision
+really happened, and rolling it back automatically would put the audit trail at
+odds with the facts — so it names the stuck requests, their step, and the stale
+mirrored status for an operator to act on.
 </Callout>
 
 ### Progress and notification deep links

content/docs/automation/flows.mdx

@@ -468,7 +468,7 @@ POST /api/v1/automation/{flow}/runs/{runId}/resume
 | Pausing node | Suspends until… | Resumed by |
 | :--- | :--- | :--- |
 | `approval` | a human decision | the approvals service (`POST /api/v1/approvals/requests/:id/approve\|reject`) — resumes down the matching `approve` / `reject` edge. **Decide through the approvals API**; the resume route above **refuses** an approval pause outright (see below). |
-| `screen` | a user submits the form | the UI runner posting the collected `inputs`; a `paused` response carrying the next `screen` chains multi-step wizards under one stable `runId` |
+| `screen` | a user submits the form | the UI runner posting the collected `inputs` — **validated server-side against the screen's declared `fields`** (see below); a `paused` response carrying the next `screen` chains multi-step wizards under one stable `runId` |
 | `wait` (timer) | an ISO-8601 duration elapses | **automatically** — a one-shot job resumes the run; after a cold boot the engine re-arms pending timers from the durable store (overdue timers resume immediately) |
 | `wait` (signal) | a named external event | any caller invoking `resume(runId)` |
 
@@ -513,6 +513,48 @@ key. A reserved name answers **400**, nothing is applied (not even legitimate
 keys sent alongside it), and the run stays parked. Ordinary author variables are
 unaffected, `$` mid-name (`price$`) included.
 
+### A `screen` resume is checked against the declared fields
+
+A screen node's `config.fields` is a **contract**, not just a rendering hint:
+the author declares which keys are collected, which are `required`, and — via
+`visibleWhen` — when a field is even asked for. `resume` enforces all of it
+server-side, so skipping the dialog and posting to the route directly is not a
+way around what the author declared:
+
+```
+POST /api/v1/automation/{flow}/runs/{runId}/resume
+{ "inputs": { "kind": "escalate" } }
+
+400  Invalid screen input: Screen field "escalation_reason" is required
+     — declared fields: 'kind', 'escalation_reason'
+```
+
+Two conditions are refused, both reported at once and both with
+`code: 'INVALID_SCREEN_INPUT'`:
+
+- a **`required` field the caller was actually asked for** is missing (an empty
+  or blank string counts as missing);
+- a key the screen **never declared** was sent.
+
+`visibleWhen` is evaluated against the **submitted values** first, so a hidden
+field's `required` never fires — enforcing it would dead-end the run at a field
+the user was never shown. A predicate that cannot be evaluated is treated as
+hidden (and logged), because the client is the authority on what was rendered.
+
+Like the `$`-namespace rule above, the refusal happens **before** the suspension
+is consumed: nothing is applied, the run stays parked, and the corrected
+submission still lands.
+
+Three shapes declare no contract and so keep the pass-through — the same way an
+action with no `params` is untouched:
+
+- an **object-form** screen (`config.objectName`), whose flat `fields` list is
+  empty by construction; the client persists the record through the normal write
+  path, which enforces that object's own `required` fields;
+- a **message-only** screen (`waitForInput: true` with no fields);
+- `signal.output`, which is the node-*output* namespace of the approval-style
+  resume envelope rather than the screen's collected-values channel.
+
 Registering a pausing node of your own? Declare `resumeAuthority: 'service'` on
 its descriptor when the decision to continue belongs to your service rather
 than to whoever holds the run id.

packages/qa/dogfood/test/flow-durable-suspend.dogfood.test.ts

@@ -32,7 +32,7 @@
 // but whose ROWS are gone, so it fails for a reason with nothing to do with
 // suspended runs. Ordinary records do not survive it either, which is what
 // identifies it as a harness/driver persistence gap rather than a defect in the
-// suspended-run store. Filed separately; when it is fixed, the natural next test
+// suspended-run store. Filed as #4518; when it is fixed, the natural next test
 // here is the one this file was originally written around.
 
 import { describe, it, expect, beforeAll, afterAll } from 'vitest';