Revamp login guards into general per-firm page guards

Replace the single hardcoded signed-out check with a guards list on each
firm: any guard can trip mid-run via a URL substring (checked on every
step) or a selector probe (checked when a step fails). Each guard defines
recovery steps or a human-wait fallback, a per-run trigger cap, and a
resume policy — 'continue' retries the tripped step, 'restart' rewinds
the run to step 1.

Runner bumped to 0.20.0; the server still ships the legacy
signedOut/authSteps shape for stale runners, and a new runner synthesizes
a guard from a legacy payload.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

app/api/autobuyer/runs/claim/route.ts

@@ -1,5 +1,5 @@
 import { claimRun } from '@/lib/db';
-import { findAutomation, resolveSteps, resolveAuthSteps } from '@/lib/automations';
+import { findAutomation, resolveSteps, resolveGuards, resolveAuthSteps } from '@/lib/automations';
 import { corsJson, corsPreflight } from '../../cors';
 
 /** The Python runner polls this. Returns the run plus the steps to execute, so
@@ -20,6 +20,9 @@ export async function POST() {
             firm: found.firm.label,
             label: found.automation.label,
             steps: resolveSteps(found.firm, found.automation, safeInputs(row.inputs)),
+            guards: resolveGuards(found.firm),
+            // Pre-0.20 runners know nothing of guards; keep feeding them the
+            // old shape until they are restarted.
             authSteps: resolveAuthSteps(found.firm),
         },
     });

clicker/actions.py

@@ -15,7 +15,8 @@ class StepError(Exception):
     """A step refused to run, or ran and could not be verified.
 
     `code` mirrors the CLI's exit codes: 3 = refused before acting,
-    4 = acted but verification failed, 5 = the session is signed out.
+    4 = acted but verification failed, 5 = a page guard tripped (login
+    redirect, blocking modal, and the like).
     """
 
     def __init__(self, message: str, code: int = 3):
@@ -23,33 +24,44 @@ class StepError(Exception):
         self.code = code
 
 
-class SignedOutError(StepError):
-    """The tab is on the login page, so the session has expired.
+class GuardTripped(StepError):
+    """A page guard's condition is on the page — the run cannot proceed past it.
 
-    Distinct from an ordinary failure because nothing is wrong with the
-    automation — it just cannot proceed until a human signs in.
+    Distinct from an ordinary failure because nothing is wrong with the step:
+    the page has been taken over by something (a login redirect, an expired-
+    session modal) and every selector now resolves against the wrong document.
+    A click or a keystroke aimed at the old page would land on whatever the
+    guard page happens to render in the same place. The runner decides how to
+    recover; the CLI just stops with exit code 5.
     """
 
-    def __init__(self, url: str):
+    def __init__(self, guard: dict, url: str):
+        name = guard.get("label") or guard.get("id") or "page"
         super().__init__(
-            f"signed out — the tab is on {url}. Log in to the site, then run this again.",
+            f"{name} guard tripped — the tab is on {url}. "
+            "Deal with the page, then run this again.",
             code=5,
         )
+        self.guard = guard
+        self.url = url
 
 
-def check_signed_in(found: dict, signed_out: str) -> None:
-    """Abort if the page redirected to a login screen.
+def check_guards(found: dict, guards: list | None) -> None:
+    """Abort if the page has tripped any URL guard.
 
-    Worth doing before every action, not just at the start: a session can expire
-    mid-run, and once it does, every selector resolves against a login form. A
-    click or a keystroke aimed at the old page would land on whatever that form
-    happens to render in the same place.
+    Worth doing before every action, not just at the start: a session can
+    expire (or a guard page appear) mid-run. Only URL guards are checked here —
+    the URL rides along with every locate answer, so they are free. Selector
+    guards cost a locate round trip each and are probed by the runner only once
+    a step has already failed.
     """
-    if not signed_out:
+    if not guards:
         return
     url = found.get("url") or ""
-    if signed_out in url:
-        raise SignedOutError(url)
+    for guard in guards:
+        pattern = guard.get("urlContains") or ""
+        if pattern and pattern in url:
+            raise GuardTripped(guard, url)
 
 
 # Display scaling factors that actually exist. Anything else means the two sides
@@ -117,6 +129,7 @@ def perform(
     url: str = "",
     open_url: str = "",
     signed_out: str = "",
+    guards: list | None = None,
     frame: str = "",
     text: str | None = None,
     clear: bool = False,
@@ -142,6 +155,12 @@ def perform(
     rng = rng or random.Random()
     say = report or (lambda _m: None)
 
+    # `guards` is the runner's list of page guards; `signed_out` is the CLI's
+    # single --signed-out pattern. Same check either way, so fold them together.
+    guards = list(guards or [])
+    if signed_out:
+        guards.append({"id": "signed-out", "label": "signed out", "urlContains": signed_out})
+
     # `frame` names the <iframe> the element lives in; the extension resolves the
     # selector inside it and offsets the coordinates back. Absent for the common
     # top-document case, and then locate behaves exactly as before.
@@ -150,7 +169,7 @@ def perform(
     found = dash.locate(selector, index, url, timeout, open_url, options=frame_opts)
 
     # Before any check that assumes we are on the real page.
-    check_signed_in(found, signed_out)
+    check_guards(found, guards)
 
     # Give a transient overlay a chance to finish leaving before giving up on it.
     if found.get("covered") and not force:
@@ -160,7 +179,7 @@ def perform(
         while found.get("covered") and time.time() < deadline:
             time.sleep(COVERED_POLL_SECONDS)
             found = dash.locate(selector, index, url, timeout, open_url, options=frame_opts)
-            check_signed_in(found, signed_out)
+            check_guards(found, guards)
         if not found.get("covered"):
             say("it cleared")
 
@@ -231,7 +250,7 @@ def perform(
     # Typing into a field that never took focus, or that ignores the input, fails
     # silently and is otherwise indistinguishable from success.
     after = dash.locate(selector, index, url, timeout, open_url, options=frame_opts)
-    check_signed_in(after, signed_out)
+    check_guards(after, guards)
     got = after.get("value")
     if got is None:
         say("element exposes no value to verify against")

clicker/runner.py

@@ -34,7 +34,8 @@ HEARTBEAT_SECONDS = 2.0
 # Bumped whenever the step vocabulary or the locate protocol changes. Reported in
 # the heartbeat so the dashboard can say "restart your runner" instead of letting
 # a stale process fail on a step type it has never heard of.
-VERSION = "0.19.0"
+# 0.20.0: page guards — run-level `guards` replace per-step signedOut/authSteps.
+VERSION = "0.20.0"
 
 # Shared with the heartbeat thread: whether a run is currently executing.
 _busy = threading.Event()
@@ -57,8 +58,14 @@ def heartbeat_loop(dash: Dashboard, opts, stop: threading.Event) -> None:
         stop.wait(HEARTBEAT_SECONDS)
 
 
-def run_one_step(dash: Dashboard, step: dict, opts, rng, report, run_id: int | None = None) -> None:
-    """Dispatch a single step. Raises StepError / DashboardError on failure."""
+def run_one_step(dash: Dashboard, step: dict, opts, rng, report,
+                 run_id: int | None = None, guards: list | None = None) -> None:
+    """Dispatch a single step. Raises StepError / DashboardError on failure,
+    GuardTripped when a URL guard's condition shows up on the page.
+
+    `guards` is empty for recovery steps: they run while a guard's condition
+    holds, so checking it again would trip on the very page they exist to fix.
+    """
     if step["action"] == "wait":
         time.sleep(float(step.get("seconds", 1)))
         report(f"waited {step.get('seconds', 1)}s")
@@ -72,7 +79,7 @@ def run_one_step(dash: Dashboard, step: dict, opts, rng, report, run_id: int | N
                              navigate_url=target)
         # Navigating to a protected page is where an expired session shows up:
         # the site answers with a redirect to its login form.
-        actions.check_signed_in(landed, step.get("signedOut", "") or "")
+        actions.check_guards(landed, guards)
         report(f"tab is on {landed.get('url', target)}")
 
     elif step["action"] == "waitFor":
@@ -110,6 +117,10 @@ def run_one_step(dash: Dashboard, step: dict, opts, rng, report, run_id: int | N
                                    step.get("urlPattern", "") or opts.url,
                                    opts.timeout, step.get("openUrl", "") or "",
                                    options=loc_opts or None)
+                # A guard page appearing mid-gate would otherwise burn the whole
+                # gate timeout before anyone noticed. GuardTripped is not caught
+                # below, so it escapes the wait as it should.
+                actions.check_guards(seen, guards)
                 if by_sight:
                     # An extension too old for the probe quietly does an ordinary
                     # locate instead, whose result has no `visible` key — reading
@@ -197,7 +208,7 @@ def run_one_step(dash: Dashboard, step: dict, opts, rng, report, run_id: int | N
                 index=int(step.get("index", 0)),
                 url=step.get("urlPattern", "") or opts.url,
                 open_url=step.get("openUrl", "") or "",
-                signed_out=step.get("signedOut", "") or "",
+                guards=guards,
                 frame=step.get("frame", "") or "",
                 text=step.get("text"),
                 clear=bool(step.get("clear")),
@@ -224,73 +235,195 @@ def run_status(dash: Dashboard, run_id: int) -> str | None:
         return None
 
 
-def wait_for_sign_in(dash: Dashboard, step: dict, run_id: int, opts, deadline: float) -> bool:
-    """Block until the tab leaves the login page.
+def guard_present(dash: Dashboard, guard: dict, step: dict, opts,
+                  probe_timeout: float = 6.0) -> bool | None:
+    """Is the guard's condition on the page right now? None = could not tell.
 
-    The runner does not log in — entering credentials is the human's job, and a
-    keystroke aimed at the wrong field is not something to risk automating. All
-    this does is watch the tab and pick the run back up once you are through.
+    A URL guard reads the tab's URL off a cheap body locate; a selector guard
+    needs a real DOM probe, whose timeout is kept short — a guard element that
+    exists shows up immediately, so most of the timeout is only ever spent
+    confirming absence. A guard defining both trips when either is present.
     """
-    signed_out = step.get("signedOut", "") or ""
     pattern = step.get("urlPattern", "") or opts.url
     open_url = step.get("openUrl", "") or ""
 
-    while time.time() < deadline:
-        if run_status(dash, run_id) == "cancelled":
-            return False
-        time.sleep(3.0)
+    if guard.get("urlContains"):
         try:
             landed = dash.locate("body", 0, pattern, opts.timeout, open_url)
         except DashboardError:
-            continue          # tab busy mid-redirect; look again shortly
-        if signed_out not in (landed.get("url") or ""):
+            return None       # tab busy mid-redirect; the caller decides
+        if guard["urlContains"] in (landed.get("url") or ""):
             return True
+        if not guard.get("selector"):
+            return False
 
+    if guard.get("selector"):
+        try:
+            dash.locate(guard["selector"], 0, pattern, probe_timeout, open_url)
+            return True
+        except NotFoundError:
+            return False
+        except DashboardError:
+            return None
     return False
 
 
-def signed_in_now(dash: Dashboard, step: dict, opts, settle_seconds: float = 20.0) -> bool:
-    """Did we actually get back in?
+def guard_cleared(dash: Dashboard, guard: dict, step: dict, opts,
+                  settle_seconds: float = 20.0) -> bool:
+    """Did recovery actually clear the condition?
 
-    Auth steps can 'succeed' — every click landing — while the site still rejects
-    the sign-in, so this confirms rather than assumes before the failed step is
-    retried.
+    Recovery steps can 'succeed' — every click landing — while the page stays in
+    the guarded state, so this confirms rather than assumes before resuming.
 
     It polls rather than checking once: submitting a login form starts a network
-    round trip and then a redirect, so the tab is still sitting on the login URL
-    for a second or two afterwards. Checking immediately reports a failure for a
-    sign-in that is merely still in flight.
+    round trip and then a redirect, so the tab is still sitting on the guard
+    page for a second or two afterwards. Checking immediately reports a failure
+    for a recovery that is merely still in flight.
     """
-    signed_out = step.get("signedOut", "") or ""
-    if not signed_out:
-        return True
-
     deadline = time.time() + settle_seconds
     while True:
-        try:
-            landed = dash.locate("body", 0, step.get("urlPattern", "") or opts.url,
-                                 opts.timeout, step.get("openUrl", "") or "")
-            if signed_out not in (landed.get("url") or ""):
-                return True
-        except DashboardError:
-            pass          # mid-redirect the tab can be briefly un-injectable
-
+        if guard_present(dash, guard, step, opts) is False:
+            return True
         if time.time() >= deadline:
             return False
         time.sleep(1.0)
 
 
+def wait_for_clear(dash: Dashboard, guard: dict, step: dict, run_id: int, opts,
+                   deadline: float) -> bool:
+    """Block until a human deals with the guard's condition.
+
+    The runner does not act here — for the login guard, entering credentials is
+    the human's job, and a keystroke aimed at the wrong field is not something
+    to risk automating. All this does is watch the page and pick the run back
+    up once the condition is gone.
+    """
+    while time.time() < deadline:
+        if run_status(dash, run_id) == "cancelled":
+            return False
+        time.sleep(3.0)
+        if guard_present(dash, guard, step, opts) is False:
+            return True
+    return False
+
+
+def find_tripped_guard(dash: Dashboard, guards: list, step: dict, opts) -> dict | None:
+    """After a step failure: is a guard's condition what actually broke it?
+
+    URL guards are caught in flight by check_guards; selector guards cost a
+    locate round trip each, so they are only probed once a step has already
+    failed — a session-expired modal explains a NotFound far better than a
+    wrong selector does.
+    """
+    for guard in guards:
+        if not guard.get("selector"):
+            continue          # URL-only guards were already checked in flight
+        if guard_present(dash, guard, step, opts):
+            return guard
+    return None
+
+
+def legacy_guards(run: dict) -> list:
+    """Guards synthesized from the pre-guard payload shape, so this runner still
+    works against a dashboard that predates run-level guards."""
+    for step in run.get("steps") or []:
+        if step.get("signedOut"):
+            return [{
+                "id": "signed-out",
+                "label": "signed out",
+                "urlContains": step["signedOut"],
+                "resume": "continue",
+                "waitSeconds": float(step.get("signedOutWait", 0) or 0),
+                "recoverySteps": run.get("authSteps") or [],
+            }]
+    return []
+
+
+def handle_guard(dash: Dashboard, run_id: int, step_index: int, label: str,
+                 guard: dict, step: dict, opts, rng, report,
+                 trigger_counts: dict) -> str:
+    """A guard tripped mid-run: recover, then say where the run picks up.
+
+    Returns "retry" (redo the step that tripped), "restart" (back to step 1),
+    "cancelled", or "fail" — the terminal outcomes have already been reported
+    and the run finished by the time they are returned.
+    """
+    gid = guard.get("id") or "guard"
+    name = guard.get("label") or gid
+
+    def fail(msg: str) -> str:
+        post_progress(dash, run_id, step_index, label, False, msg)
+        finish(dash, run_id, msg)
+        print(f"✗ run #{run_id} stopped: {msg}", file=sys.stderr)
+        return "fail"
+
+    # A guard whose recovery "works" without clearing the cause would trip
+    # again on the next step, forever. Cap it per run, not per step: a restart
+    # revisits every step, so a per-step cap would never fill up.
+    trigger_counts[gid] = trigger_counts.get(gid, 0) + 1
+    limit = int(guard.get("maxTriggers") or 3)
+    if trigger_counts[gid] > limit:
+        return fail(f"{name} guard tripped {trigger_counts[gid]} times in one run — giving up")
+
+    recovery = guard.get("recoverySteps") or []
+    if recovery:
+        post_progress(dash, run_id, step_index,
+                      f"{name} guard tripped — running {len(recovery)} recovery step(s)",
+                      False, None)
+        try:
+            for rstep in recovery:
+                rlabel = rstep.get("label") or f"recovery: {rstep['action']}"
+                print(f"        [{name}] {rlabel}")
+                # No guards passed: recovery runs on the guard page itself.
+                run_one_step(dash, rstep, opts, rng, report, run_id)
+        except (actions.StepError, DashboardError) as exc:
+            return fail(f"{name} guard recovery failed: {exc}")
+
+        if not guard_cleared(dash, guard, step, opts):
+            return fail(f"{name} guard recovery ran but the condition is still on the page")
+        report(f"{name} guard cleared")
+
+    else:
+        wait_seconds = float(guard.get("waitSeconds") or 0)
+        if wait_seconds <= 0:
+            return fail(f"{name} guard tripped and defines no recovery — stopping")
+
+        mins = wait_seconds / 60
+        note = f"{name} guard tripped — waiting up to {mins:.0f} min for you to deal with it"
+        print(f"        {note}", file=sys.stderr)
+        post_progress(dash, run_id, step_index, note, False, None)
+
+        if not wait_for_clear(dash, guard, step, run_id, opts, time.time() + wait_seconds):
+            if run_status(dash, run_id) == "cancelled":
+                print(f"■ run #{run_id} cancelled while waiting on the {name} guard")
+                return "cancelled"
+            return fail(f"{name} guard still tripped after {mins:.0f} min")
+        report(f"{name} guard cleared")
+
+    return "restart" if (guard.get("resume") or "continue") == "restart" else "retry"
+
+
 def run_steps(dash: Dashboard, run: dict, opts) -> None:
     """Work through one automation. Reports every step; stops on the first
-    failure, because a half-completed purchase flow should not barrel on."""
+    failure, because a half-completed purchase flow should not barrel on.
+
+    Page guards can interrupt any step. Once one is handled, its `resume`
+    policy decides where the run picks up: "continue" retries the step that
+    tripped, "restart" goes back to the first step — hence a while loop over
+    the step index rather than a plain for."""
     run_id = run["id"]
     steps = run["steps"]
-    auth_steps = run.get("authSteps") or []
+    guards = run.get("guards")
+    if guards is None:
+        guards = legacy_guards(run)
     rng = random.Random(opts.seed) if opts.seed is not None else random.Random()
 
     print(f"\n▶ run #{run_id} — {run['label']} ({len(steps)} steps)")
 
-    for i, step in enumerate(steps):
+    trigger_counts: dict[str, int] = {}
+    i = 0
+    while i < len(steps):
+        step = steps[i]
         label = step.get("label") or f"{step['action']} {step.get('selector', '')}".strip()
         print(f"  [{i + 1}/{len(steps)}] {label}")
 
@@ -300,74 +433,46 @@ def run_steps(dash: Dashboard, run: dict, opts) -> None:
             detail_parts.append(message)
             print(f"        {message}")
 
-        # One retry after a sign-in: the session can only be expired once per step.
-        for attempt in range(2):
-            try:
-                run_one_step(dash, step, opts, rng, report, run_id)
-                break
+        outcome = None
+        try:
+            run_one_step(dash, step, opts, rng, report, run_id, guards=guards)
 
-            except actions.SignedOutError as exc:
-                wait_seconds = float(step.get("signedOutWait", 0) or 0)
+        except actions.GuardTripped as exc:
+            print(f"        {exc}", file=sys.stderr)
+            outcome = handle_guard(dash, run_id, i, label, exc.guard, step,
+                                   opts, rng, report, trigger_counts)
 
-                # Scripted auth first, if the firm defines any; otherwise fall
-                # back to pausing for a human. Only ever attempted once per step —
-                # auth that "succeeds" without signing in would otherwise loop.
-                if auth_steps and attempt == 0:
-                    post_progress(dash, run_id, i, f"signed out — running {len(auth_steps)} auth step(s)",
-                                  False, exc.args[0])
-                    try:
-                        for astep in auth_steps:
-                            alabel = astep.get("label") or f"auth: {astep['action']}"
-                            print(f"        [auth] {alabel}")
-                            run_one_step(dash, astep, opts, rng, report, run_id)
-                    except (actions.StepError, DashboardError) as aexc:
-                        msg = f"auth steps failed: {aexc}"
-                        post_progress(dash, run_id, i, label, False, msg)
-                        finish(dash, run_id, msg)
-                        print(f"✗ run #{run_id} stopped: {msg}", file=sys.stderr)
-                        return
+        except (actions.StepError, DashboardError) as exc:
+            if run_status(dash, run_id) == "cancelled":
+                print(f"■ run #{run_id} cancelled from the dashboard")
+                return
 
-                    if not signed_in_now(dash, step, opts):
-                        msg = "auth steps ran but the session is still signed out"
-                        post_progress(dash, run_id, i, label, False, msg)
-                        finish(dash, run_id, msg)
-                        print(f"✗ run #{run_id} stopped: {msg}", file=sys.stderr)
-                        return
-
-                    report("signed in via auth steps — retrying")
-                    continue
-
-                if wait_seconds <= 0 or attempt > 0:
-                    print(f"        FAILED: {exc}", file=sys.stderr)
-                    post_progress(dash, run_id, i, label, False, str(exc))
-                    finish(dash, run_id, str(exc))
-                    print(f"✗ run #{run_id} stopped at step {i + 1}")
-                    return
-
-                mins = wait_seconds / 60
-                note = f"signed out — waiting up to {mins:.0f} min for you to log in"
-                print(f"        {note}", file=sys.stderr)
-                post_progress(dash, run_id, i, note, False, exc.args[0])
-
-                if not wait_for_sign_in(dash, step, run_id, opts, time.time() + wait_seconds):
-                    if run_status(dash, run_id) == "cancelled":
-                        print(f"■ run #{run_id} cancelled while waiting for sign-in")
-                        return
-                    msg = f"still signed out after {mins:.0f} min"
-                    post_progress(dash, run_id, i, label, False, msg)
-                    finish(dash, run_id, msg)
-                    print(f"✗ run #{run_id} gave up waiting for sign-in")
-                    return
-
-                report("signed in — retrying the step")
-
-            except (actions.StepError, DashboardError) as exc:
+            # Before writing the failure off, see whether a selector guard's
+            # condition is sitting on the page — an expired-session modal
+            # explains a missing button better than a wrong selector does.
+            tripped = find_tripped_guard(dash, guards, step, opts) if guards else None
+            if tripped is not None:
+                name = tripped.get("label") or tripped.get("id")
+                print(f"        step failed with the {name} guard's condition on the page",
+                      file=sys.stderr)
+                outcome = handle_guard(dash, run_id, i, label, tripped, step,
+                                       opts, rng, report, trigger_counts)
+            else:
                 print(f"        FAILED: {exc}", file=sys.stderr)
                 post_progress(dash, run_id, i, label, False, str(exc))
                 finish(dash, run_id, str(exc))
                 print(f"✗ run #{run_id} stopped at step {i + 1}")
                 return
 
+        if outcome == "retry":
+            continue          # same step, another go
+        if outcome == "restart":
+            print(f"  ↺ restarting run #{run_id} from step 1")
+            i = 0
+            continue
+        if outcome in ("fail", "cancelled"):
+            return            # handle_guard already reported and finished
+
         status = post_progress(dash, run_id, i + 1, label, True, "; ".join(detail_parts) or None)
 
         # The dashboard's Stop button shows up here.
@@ -375,6 +480,8 @@ def run_steps(dash: Dashboard, run: dict, opts) -> None:
             print(f"■ run #{run_id} cancelled from the dashboard")
             return
 
+        i += 1
+
     finish(dash, run_id, None)
     print(f"✓ run #{run_id} complete")
 

lib/automations.ts

@@ -110,7 +110,11 @@ export interface Automation {
 }
 
 /** A step as handed to the runner: the firm's tab pattern and fallback URL
- *  filled in, so the runner never has to know which firm it is working on. */
+ *  filled in, so the runner never has to know which firm it is working on.
+ *
+ *  `signedOut`/`signedOutWait` are the pre-guard (runner < 0.20) shape of the
+ *  login guard, still stamped on so a stale runner keeps working — a 0.20+
+ *  runner ignores them in favour of the run's `guards`. */
 export type ResolvedStep = AutomationStep & {
     urlPattern?: string;
     openUrl?: string;
@@ -118,6 +122,59 @@ export type ResolvedStep = AutomationStep & {
     signedOutWait?: number;
 };
 
+/** A condition that can take over the page at any point in a run — a login
+ *  redirect, a session-expired modal, a maintenance page, a rate-limit
+ *  interstitial. Guards belong to the firm, not to any step: once the page is
+ *  in that state, every selector resolves against the wrong document, so the
+ *  run must deal with the guard before anything else can proceed.
+ *
+ *  How a guard trips:
+ *   - `urlContains` — the tab's URL carries this substring. Checked on every
+ *     step, for free: the URL rides along with every locate answer.
+ *   - `selector` — this element is on the page. A DOM probe costs a round
+ *     trip, so the runner checks these when a step fails, where a guard is
+ *     the likeliest explanation. A guard may define either or both (both =
+ *     tripped when either is present).
+ *
+ *  What happens then:
+ *   1. `recoverySteps` run, if there are any. They execute *while the guard
+ *      condition holds* — on the login form, under the modal — so they carry
+ *      no guard checks of their own; a wrong selector here is worse than
+ *      elsewhere. Anything written here lives in this file in plain text,
+ *      and this file is in the repo.
+ *   2. With no recovery steps, the run pauses up to `waitSeconds` for a human
+ *      to deal with it (sign in, dismiss the thing). No recovery steps and no
+ *      wait means tripping the guard fails the run.
+ *   3. Once the condition has verifiably cleared, `resume` decides where the
+ *      run picks up: `'continue'` retries the step that tripped and carries
+ *      on; `'restart'` goes back to the automation's first step — for
+ *      conditions that reset page state (a re-login that lands on the
+ *      dashboard, a modal that closes the checkout flow).
+ */
+export interface PageGuard {
+    /** Stable id — names the guard in run logs and counts its triggers per run. */
+    id: string;
+    /** Shown in the run log; falls back to `id`. */
+    label?: string;
+    /** Trip when the tab's URL contains this substring. */
+    urlContains?: string;
+    /** Trip when this element is on the page. */
+    selector?: string;
+    /** Steps that clear the condition. Run guard-free, on the guard page itself. */
+    recoverySteps?: AutomationStep[];
+    /** Where the run picks up once the condition clears. */
+    resume: 'restart' | 'continue';
+    /** Seconds to pause for a human when there are no `recoverySteps`. */
+    waitSeconds?: number;
+    /** Times this guard may trip in one run before the run fails; a guard whose
+     *  recovery "works" without clearing the cause would otherwise loop.
+     *  Default 3. */
+    maxTriggers?: number;
+}
+
+/** A guard as handed to the runner: recovery steps expanded and resolved. */
+export type ResolvedGuard = Omit<PageGuard, 'recoverySteps'> & { recoverySteps: ResolvedStep[] };
+
 export interface Firm {
     id: string;
     label: string;
@@ -127,27 +184,11 @@ export interface Firm {
     /** Concrete page to open when no tab matches `urlPattern`. A match pattern
      *  can't be navigated to, so this has to be spelled out separately. */
     url: string;
-    /** Substring identifying the signed-out page. When a session expires the site
-     *  redirects here, and every subsequent selector resolves against a login form
-     *  — so a run that lands on it must stop rather than click through it. */
-    signedOutPattern: string;
-    /** Seconds to pause and let a human sign in when a run hits the login page.
-     *  0 aborts instead. Used only when `authSteps` is empty. */
-    signedOutWaitSeconds: number;
-    /** Steps run when a run lands on the login page, before retrying the step
-     *  that hit it. Left empty here on purpose — fill it in yourself.
-     *
-     *  Two things to know if you do:
-     *   - These run *while on the signed-out page*, so unlike normal steps they
-     *     carry no signed-out guard. Nothing stops them clicking around a login
-     *     form; that is the point, and also why a wrong selector here is worse
-     *     than elsewhere.
-     *   - Anything written here lives in this file in plain text, and this file
-     *     is in the repo.
-     *
-     *  While empty, a run that hits the login page falls back to pausing for
-     *  `signedOutWaitSeconds` so you can sign in by hand. */
-    authSteps: AutomationStep[];
+    /** Conditions that can interrupt a run at any point, and how to get past
+     *  them. Every firm wants at least the signed-out guard: when a session
+     *  expires the site redirects to its login page, and a run that lands there
+     *  must stop rather than click through the form. See PageGuard. */
+    guards: PageGuard[];
     automations: Automation[];
 }
 
@@ -160,10 +201,17 @@ export const FIRMS: Firm[] = [
         label: 'Tradeify',
         urlPattern: 'https://app-f.tradeify.co/*',
         url: 'https://app-f.tradeify.co/',
-        signedOutPattern: '/auth/',
-        signedOutWaitSeconds: 300,
-        authSteps: [
-            { action: 'click', selector: 'form > div > div:last-child button', label: 'Login' },
+        guards: [
+            {
+                id: 'signed-out',
+                label: 'Signed out',
+                urlContains: '/auth/',
+                resume: 'continue',
+                waitSeconds: 300,
+                recoverySteps: [
+                    { action: 'click', selector: 'form > div > div:last-child button', label: 'Login' },
+                ],
+            },
         ],
         automations: [
             {
@@ -248,14 +296,20 @@ export const FIRMS: Firm[] = [
         // well as subdomains, confirmed against a live tab on lucidtrading.com.
         urlPattern: 'https://*.lucidtrading.com/*',
         url: 'https://dash.lucidtrading.com/',
-        // Signed out lands on the apex; signed in stays on dash. The leading `//`
-        // anchors this to the start of the host — without it, the substring also
-        // matches dash.lucidtrading.com and every step would abort while logged in.
-        signedOutPattern: '//lucidtrading.com/',
-        signedOutWaitSeconds: 300,
-        authSteps: [
-            { action: 'click', selector: 'button.lucid-login-btn', label: 'Login' },
-
+        guards: [
+            {
+                id: 'signed-out',
+                label: 'Signed out',
+                // Signed out lands on the apex; signed in stays on dash. The leading `//`
+                // anchors this to the start of the host — without it, the substring also
+                // matches dash.lucidtrading.com and every step would abort while logged in.
+                urlContains: '//lucidtrading.com/',
+                resume: 'continue',
+                waitSeconds: 300,
+                recoverySteps: [
+                    { action: 'click', selector: 'button.lucid-login-btn', label: 'Login' },
+                ],
+            },
         ],
         automations: [
             {
@@ -371,8 +425,17 @@ function labelled(step: ResolvedStep, iteration: number, total: number): Resolve
     return { ...step, label: `${step.label ?? describeStep(step)} (${iteration}/${total})` };
 }
 
+/** The guard a pre-0.20 runner knew as "the" login guard: the first URL guard.
+ *  Its pattern is still stamped onto steps (`signedOut`) and its recovery steps
+ *  still ship as `authSteps`, so a stale runner keeps working until restarted.
+ *  A 0.20+ runner ignores both in favour of the run-level `guards`. */
+function legacyLoginGuard(firm: Firm): PageGuard | undefined {
+    return firm.guards.find((g) => g.urlContains);
+}
+
 /** One non-repeat step, with the firm's tab pattern and URLs filled in. */
 function resolveOne(firm: Firm, step: AutomationStep): ResolvedStep {
+    const login = legacyLoginGuard(firm);
     if (step.action === 'wait') return step;
     if (step.action === 'navigate') {
         return {
@@ -380,8 +443,8 @@ function resolveOne(firm: Firm, step: AutomationStep): ResolvedStep {
             url: step.url ?? firm.url,
             urlPattern: firm.urlPattern,
             openUrl: firm.url,
-            signedOut: firm.signedOutPattern,
-            signedOutWait: firm.signedOutWaitSeconds,
+            signedOut: login?.urlContains,
+            signedOutWait: login?.waitSeconds,
         };
     }
     if (step.action === 'waitFor') {
@@ -404,8 +467,8 @@ function resolveOne(firm: Firm, step: AutomationStep): ResolvedStep {
         ...step,
         urlPattern: step.urlPattern ?? firm.urlPattern,
         openUrl: firm.url,
-        signedOut: firm.signedOutPattern,
-        signedOutWait: firm.signedOutWaitSeconds,
+        signedOut: login?.urlContains,
+        signedOutWait: login?.waitSeconds,
     };
 }
 
@@ -451,19 +514,34 @@ export function resolveSteps(firm: Firm, automation: Automation, inputs: RunInpu
     return expand(automation.steps, firm, automation, inputs);
 }
 
-/** Auth steps run on the login page itself, so they get the firm's tab pattern
- *  but deliberately no `signedOut` guard — that guard exists to stop ordinary
+/** Recovery steps run on the guard page itself, so they get the firm's tab
+ *  pattern but deliberately no `signedOut` stamp — that exists to stop ordinary
  *  steps acting on a login form, and these are the exception.
  *
  *  They share the expansion path, so a repeat works here too; there are no
- *  per-run inputs during auth, so `timesFrom` falls back to a single pass. */
-export function resolveAuthSteps(firm: Firm): ResolvedStep[] {
-    const stub: Automation = { id: '_auth', label: 'auth', description: '', steps: firm.authSteps };
-    return expand(firm.authSteps, firm, stub, {}).map((step) =>
+ *  per-run inputs during recovery, so `timesFrom` falls back to a single pass. */
+function resolveRecoverySteps(firm: Firm, steps: AutomationStep[]): ResolvedStep[] {
+    const stub: Automation = { id: '_guard', label: 'guard recovery', description: '', steps };
+    return expand(steps, firm, stub, {}).map((step) =>
         step.action === 'wait' ? step : { ...step, signedOut: undefined, signedOutWait: undefined }
     );
 }
 
+/** Guards as handed to the runner, recovery steps expanded. */
+export function resolveGuards(firm: Firm): ResolvedGuard[] {
+    return firm.guards.map((guard) => ({
+        ...guard,
+        recoverySteps: resolveRecoverySteps(firm, guard.recoverySteps ?? []),
+    }));
+}
+
+/** The pre-0.20 payload shape: the login guard's recovery steps as `authSteps`.
+ *  Kept so a stale runner still recovers from a signed-out redirect. */
+export function resolveAuthSteps(firm: Firm): ResolvedStep[] {
+    const login = legacyLoginGuard(firm);
+    return login ? resolveRecoverySteps(firm, login.recoverySteps ?? []) : [];
+}
+
 /** What a step is doing, for the run log and the dashboard. */
 export function describeStep(step: AutomationStep): string {
     if (step.label) return step.label;

lib/db.ts

@@ -609,7 +609,7 @@ export const RUNNER_TIMEOUT_MS = 7000;
 /** The runner version this server's step vocabulary requires. A running process
  *  doesn't reload when the source changes, so an older one silently fails on
  *  steps it predates — the dashboard warns instead. */
-export const RUNNER_EXPECTED_VERSION = '0.19.0';
+export const RUNNER_EXPECTED_VERSION = '0.20.0';
 
 export interface RunnerHeartbeat {
     at: number;