@foldrun/core 0.3.0 → 0.5.0

package/dist/src/evals.js

@@ -39,7 +39,7 @@ import { spawn } from "node:child_process";
 import { query } from "@anthropic-ai/claude-agent-sdk";
 import { workspaceDir, readRun, writeWorkspaceFile, resolveModel, resolveEffort, assertSafeName, } from "./store.js";
 import { startFlowRun, loadFlow } from "./runner.js";
-import { platform } from "./platform.js";
+import { platform, platformQueues } from "./platform.js";
 const ASSERTION_TYPES = ["contains", "not-contains", "matches", "file", "run", "judge"];
 /**
  * A scalar that may continue onto following lines.
@@ -138,6 +138,8 @@ export function parseEval(file, raw) {
         model: resolveModel(data.model ?? "fast"),
         effort: resolveEffort(data.effort ?? "low"),
         trigger: data.trigger === "manual" ? "manual" : "deploy",
+        live: data.live === true,
+        inputs: data.inputs === true,
         cases,
     };
 }
@@ -226,7 +228,9 @@ export function promoteRunToEval(tenant, workspace, runId, opts = {}) {
     let expect;
     if (opts.expect && opts.expect.length > 0) {
         expect = opts.expect.map((line) => {
-            const m = oneLine(line).match(/^([a-z-]+):\s*(.*)$/);
+            const m = oneLine(line).match(/^([a-z-]+):(.*)$/);
+            if (m)
+                m[2] = m[2].trim();
             if (!m || !ASSERTION_TYPES.includes(m[1]) || !m[2]) {
                 throw refuse(`"${line}" is not an assertion — expected one of ${ASSERTION_TYPES.map((t) => `${t}:`).join(", ")} followed by a value`, 400);
             }
@@ -258,7 +262,7 @@ export function promoteRunToEval(tenant, workspace, runId, opts = {}) {
         if (points.kind !== target.kind || points.name !== target.name) {
             throw refuse(`${rel} tests ${points.kind} ${points.name ?? "(nothing)"}, not ${target.kind} ${target.name} — pick another evalName`, 409);
         }
-        content = `${existing.replace(/\s+$/, "")}\n\n${caseText}`;
+        content = `${existing.trimEnd()}\n\n${caseText}`;
     }
     else {
         // A flow eval costs a whole run and touches real systems, so it waits
@@ -285,19 +289,23 @@ export function promoteRunToEval(tenant, workspace, runId, opts = {}) {
     });
     return { file: rel, created: !existing, caseName };
 }
-/** Wait for a run to finish, with a ceiling so a hung agent can't hang an eval. */
 /**
  * A web replica never drives a run: it has no route to the cluster API and
  * no right to create a pod, by design — so an eval started there (a push,
  * the Run button) queues its run for the worker like any other. From 29 Aug
  * every eval on the split platform died at "pod create failed: connection
  * refused" for exactly this reason. A single process — the CLI, a laptop,
- * FOLDRUN_ROLE unset — still drives it right here, as it always did.
+ * no platform registered — still drives it right here, as it always did.
+ *
+ * The test is whether a platform owns the queue, not which role this
+ * process was started as: on a worker the role read "worker", the eval
+ * started its run in-process, and the run never passed through the
+ * account cap or the lanes the platform enforces on every other run.
  */
-const runsViaQueue = () => process.env.FOLDRUN_ROLE === "web";
-async function beginEvalRun(tenant, workspace, steps, flowName, model, effort) {
+const runsViaQueue = () => platformQueues();
+async function beginEvalRun(tenant, workspace, steps, flowName, model, effort, test = true) {
     if (!runsViaQueue())
-        return startFlowRun(tenant, workspace, steps, flowName, model, [], effort);
+        return startFlowRun(tenant, workspace, steps, flowName, model, [], effort, null, { test });
     // The job carries a model override but no effort one, and driveRun cannot
     // re-read a flow called `eval:<name>`. Put the flow's defaults onto the
     // steps that have none — which is what nearest-wins resolves to anyway.
@@ -306,7 +314,7 @@ async function beginEvalRun(tenant, workspace, steps, flowName, model, effort) {
         ...(s.model || !model ? {} : { model }),
         ...(s.effort || !effort ? {} : { effort }),
     }));
-    return platform.enqueueFlowRun(tenant, workspace, withDefaults, flowName, model ?? null, []);
+    return platform.enqueueFlowRun(tenant, workspace, withDefaults, flowName, model ?? null, [], null, { test });
 }
 // Waits for as long as the run runs. There was a cap here (5 minutes, then
 // 30) and it marked a flow eval "run disappeared" at the cap while the flow
@@ -412,7 +420,10 @@ async function checkAssertion(assertion, output, agentDir, model, effort) {
 export async function runEval(tenant, workspace, info) {
     const startedAt = new Date().toISOString();
     const results = [];
-    for (const testCase of info.cases) {
+    // Saved inputs carry no assertions: running them would start and bill a
+    // run per case only to report every one failed. Nothing to grade.
+    const cases = info.inputs ? [] : info.cases;
+    for (const testCase of cases) {
         const target = info.flow ? null : info.agent;
         let runId = null;
         let output = "";
@@ -428,12 +439,14 @@ export async function runEval(tenant, workspace, info) {
                 const steps = flow.steps.map((s, i) => i === 0
                     ? { ...s, instruction: `${s.instruction}\n\n<run_task>\n${testCase.task}\n</run_task>` }
                     : s);
-                run = await beginEvalRun(tenant, workspace, steps, `eval:${info.name}`, flow.model, flow.effort);
+                // Test unless the eval OR the flow says live: an eval that runs a
+                // flow written `live: true` was written by someone who knew.
+                run = await beginEvalRun(tenant, workspace, steps, `eval:${info.name}`, flow.model, flow.effort, !(info.live || flow.live));
             }
             else {
                 if (!target)
                     throw new Error("this eval names neither an agent nor a flow");
-                run = await beginEvalRun(tenant, workspace, [{ agent: target, instruction: testCase.task, group: 1, optional: false }], `eval:${info.name}`);
+                run = await beginEvalRun(tenant, workspace, [{ agent: target, instruction: testCase.task, group: 1, optional: false }], `eval:${info.name}`, undefined, undefined, !info.live);
             }
             runId = run.id;
             const finished = await waitForRun(tenant, workspace, run.id);
@@ -592,7 +605,7 @@ commit = null) {
         return { ran: 0, skipped: "no revision to stamp the result with" };
     // Only the evals that asked to run on deploy — a `trigger: manual` eval is
     // one a person runs on purpose, because it costs a whole flow run.
-    const evals = listEvals(tenant, workspace).filter((e) => e.trigger !== "manual");
+    const evals = listEvals(tenant, workspace).filter((e) => e.trigger !== "manual" && !e.inputs);
     if (evals.length === 0)
         return { ran: 0, skipped: "no evals" };
     const dir = path.join(workspaceDir(tenant, workspace), "evals", ".results");

package/dist/src/flow-lint.d.ts

@@ -6,11 +6,23 @@ export interface FlowWarning {
     line?: number;
     message: string;
     detail: string;
+    /** "error" for a value the parser could not read — `timeout: abc` — which
+     *  `foldrun check` must fail on rather than note. Absent means advisory,
+     *  which every structural warning in this file is. */
+    level?: "error" | "warn";
+    /** A stable tag for warnings a caller wants to single out — the deploy
+     *  surfaces `kind: "outward"` as a non-blocking warning. */
+    kind?: string;
 }
 /** What exists in the workspace, when the caller knows. Optional: the lint
  *  is useful without it, and a caller that has the lists gets the reference
  *  checks too. */
 export interface KnownNames {
     agents: string[];
+    /** Agents that hold a tool declaring `outward: true` — one that sends,
+     *  posts, buys or publishes. The caller resolves the grants; this file only
+     *  needs the answer. Absent means the caller did not look, and the outward
+     *  rule is skipped rather than guessed at. */
+    outwardAgents?: string[];
 }
 export declare function lintFlow(flow: FlowInfo, known?: KnownNames): FlowWarning[];

package/dist/src/flow-lint.js

@@ -10,13 +10,68 @@
 // Every check here is *structural*: it reasons about what a step can possibly
 // have in context, never about whether the model did a good job. Nothing is
 // blocked — flows still run, warnings are advisory.
+import { whenRowsPath } from "./store.js";
+import { timezoneProblem } from "./clock.js";
+/** What a `[[flow:x]]` step's options are called in a flow file, for the
+ *  ones expandSubflows does not carry into the nested steps. */
+const SUBFLOW_DROPPED = [
+    ["approve", "an approval (!)"], ["when", "when:"], ["case", "case:"], ["else", "else:"],
+    ["retry", "retry:"], ["timeout", "timeout:"], ["verify", "verify:"], ["model", "model:"],
+    ["effort", "effort:"], ["onFail", "on-fail:"], ["waitSecs", "wait:"], ["waitFor", "wait:"],
+    ["ask", "ask:"], ["preview", "preview:"], ["delegate", "delegate:"], ["loop", "loop:"],
+    ["until", "until:"], ["each", "each:"], ["max", "max:"], ["output", "output:"],
+    ["schema", "schema:"], ["parallel", "parallel:"], ["maxTurns", "max_turns:"], ["limits", "limits:"],
+];
 // Anaphora: phrases that only mean something if earlier output exists.
 // Deliberately conservative — a false warning is worse than a missed one,
 // because noisy warnings get ignored and then the real one is invisible too.
 const REFERS_BACK = /\b(both|above|earlier|previous(ly)?|prior|the (changes|draft|research|results|findings|report|article|fixes|work)|what (they|the others?) (found|shipped|wrote|said))\b/i;
+/** One @, something either side, a dot in the domain, no whitespace. Written
+ *  as string checks rather than a regex: the obvious pattern backtracks
+ *  polynomially on a domain full of dots, and a lint that can be made slow
+ *  by a flow file is a lint someone will turn off. */
+function looksLikeAddress(a) {
+    const at = a.indexOf("@");
+    if (at < 1 || at !== a.lastIndexOf("@") || /\s/.test(a))
+        return false;
+    const domain = a.slice(at + 1);
+    const dot = domain.indexOf(".");
+    return dot > 0 && dot < domain.length - 1;
+}
+/** The assertion words a `verify:` may open with. Anything else is handed to
+ *  a shell — see `checkVerify` in step-exec.ts, which owns this list. */
+const VERIFY_ASSERTION = /^(contains|not-contains|matches|file|judge):\s/;
+/** What a shell command has that a sentence does not: an operator that only
+ *  means something to a shell, a path or bracket at the very start, or one
+ *  of the commands these flows actually reach for. Deliberately NOT any
+ *  punctuation a sentence might carry — the first version disqualified on
+ *  `;`, `,`, `'` and `"`, so "the reply opens with GOOD; a GOOD reply names
+ *  the SHA" read as shell and a guaranteed exit 127 sat in fix-desk unseen
+ *  by the very lint written to catch it. */
+const LOOKS_LIKE_SHELL = /&&|\|\||\s\|\s|\$\(|\$\{|`|\s[<>]\s|\d?>[>&]?\s|^\[|^[.~/]|^\s*\b(test|grep|egrep|find|head|tail|cat|ls|node|npm|npx|bash|sh|zsh|python3?|jq|awk|sed|curl|git|make|exit|n=|[A-Za-z_]+=)/;
+/** A `verify:` that is a claim in English rather than a command. Conservative
+ *  on purpose, the same bargain as REFERS_BACK: it must have no assertion
+ *  prefix, nothing shell-shaped anywhere in it, and enough words to be a
+ *  sentence — so a bare `./check.sh` or a one-word command is left alone and
+ *  the warning only fires where the author plainly wrote prose. */
+function looksLikeProse(verify) {
+    const v = verify.trim();
+    if (!v || VERIFY_ASSERTION.test(v))
+        return false;
+    if (LOOKS_LIKE_SHELL.test(v))
+        return false;
+    return v.split(/\s+/).length >= 4;
+}
 export function lintFlow(flow, known) {
     const warnings = [];
     const agentNames = known ? new Set(known.agents) : null;
+    // A frontmatter enum the parser could not read, or a step line whose
+    // `[[target]]` is malformed — recorded by parseFlow, reported here at
+    // error level so `foldrun check` fails on it, the way a bad step option
+    // does. A value nothing can read is never a silent default.
+    for (const fp of flow.frontProblems ?? []) {
+        warnings.push({ step: null, level: "error", message: fp, detail: "The line is written but its value cannot be read, so a run would use a silent default instead of what the author meant. Write one of the values the option names, or a target in lower case, digits and dashes." });
+    }
     // Cron's oldest trap, and an expensive one: when day-of-month AND
     // day-of-week are both restricted they are OR'd, not AND'd — the scheduler
     // does it too, correctly (`domRestricted && dowRestricted` → `domOk ||
@@ -25,6 +80,82 @@ export function lintFlow(flow, known) {
     // eleven runs where one was meant. Nothing about the expression looks wrong,
     // which is exactly why it needs saying here rather than being discovered on
     // the bill.
+    // An approvers list only one word can satisfy. `approvers: [editors]` is
+    // a role that is not `admins` and not an address; mayApprove would then
+    // refuse every person, and the gate could only ever be answered by an
+    // emailed link. Said here rather than discovered at the gate.
+    const odd = (flow.approvers ?? []).filter((a) => a !== "admins" && !looksLikeAddress(a));
+    if (odd.length) {
+        warnings.push({
+            step: null,
+            message: `approvers: ${odd.join(", ")} — not an email address, and the only role word is admins`,
+            detail: "Each entry is an address that may answer this flow's gates, or the word `admins`. Anything else matches nobody, and a list nobody matches locks every gate.",
+        });
+    }
+    // A flow with an approvers list and no gate at all is a rule the author
+    // believes is in force and which does nothing — the worst kind of safety
+    // control.
+    if (flow.approvers?.length && !flow.steps.some((s) => s.approve || s.ask)) {
+        warnings.push({
+            step: null,
+            message: "approvers: is set but no step in this flow is a gate",
+            detail: "Nothing here waits for a person, so the list has no effect. Mark the step that needs " +
+                "a second pair of eyes with `!` (or `approve: true`), or drop the approvers: line.",
+        });
+    }
+    // A zone nobody can read is two faults at once: the cron fires at the
+    // wrong hour, and every step's calendar quietly falls through to the level
+    // above. Both are invisible in a green run, so `foldrun check` refuses it.
+    const zoneProblem = timezoneProblem(flow.timezone);
+    if (zoneProblem) {
+        warnings.push({
+            step: null,
+            level: "error",
+            message: zoneProblem,
+            detail: "This is the zone the schedule fires in AND the calendar every step in this flow works " +
+                "to. Unreadable, the schedule falls back to UTC and each step inherits the workspace's " +
+                "zone instead of this one.",
+        });
+    }
+    // A deadline on a gate that does not exist reads as a safety net and is not
+    // one. Same reasoning, said separately because the fix is different.
+    if (flow.approveWithin && !flow.steps.some((s) => s.approve || s.ask)) {
+        warnings.push({
+            step: null,
+            message: "approve_within: is set but no step in this flow is a gate",
+            detail: "Nothing waits for a person, so nothing can expire. Mark a step with `!`, or drop the line.",
+        });
+    }
+    // A throttle no schedule can ever trip: the flow cannot run more often
+    // than its cron fires, so a floor below that interval never fires either.
+    // Cheap to write, silently pointless, and a reader will believe it.
+    if (flow.throttle && flow.trigger === "schedule" && flow.schedule) {
+        warnings.push({
+            step: null,
+            message: `throttle: on a scheduled flow — check it is shorter than the schedule, or it will drop runs you meant to happen`,
+            detail: "A schedule already decides how often this runs. A throttle longer than the gap between " +
+                "fires silently skips them; if that is what you want, say it in the cron instead.",
+        });
+    }
+    // `idempotency:` only has a delivery to read on a trigger that receives
+    // one. On a schedule or a chained flow there is no request and no field,
+    // so the key is inert.
+    if (flow.idempotency && flow.trigger !== "webhook" && flow.trigger !== "email") {
+        warnings.push({
+            step: null,
+            message: `idempotency: has nothing to read on trigger: ${flow.trigger}`,
+            detail: "A delivery id comes off an incoming request, so this only does something for " +
+                "trigger: webhook or trigger: email. On any other trigger it is ignored.",
+        });
+    }
+    // `catchup:` is about a fire that was missed, which only a clock can miss.
+    if (flow.catchup && flow.trigger !== "schedule") {
+        warnings.push({
+            step: null,
+            message: `catchup: has no effect on trigger: ${flow.trigger}`,
+            detail: "Only a schedule can miss a fire. Every other trigger happens when it happens.",
+        });
+    }
     if (flow.schedule) {
         const [, , dom, , dow] = flow.schedule.trim().split(/\s+/);
         if (dom && dow && dom !== "*" && dow !== "*") {
@@ -39,6 +170,9 @@ export function lintFlow(flow, known) {
     }
     // Trigger shapes that can never fire, said at check time rather than
     // discovered as a flow that "never runs".
+    if (flow.budgetProblem) {
+        warnings.push({ step: null, message: flow.budgetProblem, detail: "A flow's `budget:` is the most one run may spend — a number in USD, or `unlimited`. A cap over a day, week or month belongs in the workspace's or the account's AGENTS.md, where it means every run put together." });
+    }
     if (flow.trigger === "once" && !flow.at) {
         warnings.push({ step: null, message: "trigger: once needs an `at:` instant", detail: "Write `at: 2026-09-05T09:00:00+10:00` (ISO 8601). Without a readable instant this flow never fires." });
     }
@@ -62,6 +196,36 @@ export function lintFlow(flow, known) {
     const firstGroup = Math.min(...flow.steps.map((s) => s.group));
     flow.steps.forEach((step, i) => {
         const inFirstGroup = step.group === firstGroup;
+        // A value the parser refused. Not advisory: the run would go ahead on
+        // a number the author never wrote — `timeout: 15m` ran as one second
+        // for as long as this was a coercion — so the check fails instead.
+        for (const problem of step.problems ?? []) {
+            warnings.push({
+                step: i,
+                line: step.line,
+                level: "error",
+                message: problem,
+                detail: "The option is written but its value cannot be read, and a run would not do what the " +
+                    "line says. Durations are seconds or 90s / 30m / 4h / 3d; counts are whole numbers " +
+                    "within the key's cap (retry and loop at most 5, max at most 20).",
+            });
+        }
+        // An instruction written over more than one line. It is read in full
+        // now, but the parser had to decide where it stopped, and the line
+        // below a wrapped instruction is one edit away from being prose. Said
+        // out loud so an author who meant to wrap can see it, and one who did
+        // not can see what got joined on.
+        if (step.wrapped) {
+            warnings.push({
+                step: i,
+                line: step.line,
+                message: `"${step.subflow ?? step.agent}" has an instruction spread over more than one line`,
+                detail: "A step's instruction ends where the step ends: the indented lines directly under it " +
+                    "are joined on, and the first blank line, option, heading, bullet or unindented line " +
+                    "finishes it. Put the whole instruction on the step's own line — one line, however " +
+                    "long — so what the agent is told does not depend on where the paragraph below begins.",
+            });
+        }
         // The first group starts with context = null. An instruction that points
         // at earlier output therefore points at nothing.
         if (inFirstGroup && step.instruction && REFERS_BACK.test(step.instruction)) {
@@ -73,6 +237,29 @@ export function lintFlow(flow, known) {
                     "invent what it was asked to review rather than fail. Move it to a later group.",
             });
         }
+        // `../../storage/…` still works everywhere; `workspace/storage/…` is the
+        // spelling that also works in a shell, and the one the docs teach. Said
+        // once per step, advisory — never an error on a path that resolves.
+        {
+            const rowsOf = step.when ? whenRowsPath(step.when) : null;
+            const spots = [
+                step.eachPath ? ["each: rows of", step.eachPath] : null,
+                rowsOf ? ["when: rows of", rowsOf] : null,
+                step.verify ? ["verify:", step.verify] : null,
+                step.schemaPath ? ["schema:", step.schemaPath] : null,
+                step.instruction ? ["the instruction", step.instruction] : null,
+            ].filter((x) => !!x && /\.\.\/\.\.\/(storage|state)\b/.test(x[1]));
+            if (spots.length) {
+                warnings.push({
+                    step: i,
+                    line: step.line,
+                    kind: "spelling",
+                    message: `"${step.subflow ?? step.agent}" writes ../../storage or ../../state in ${spots.map((x) => x[0]).join(", ")}`,
+                    detail: "It still works. The documented spelling is workspace/storage/… and workspace/state/… — " +
+                        "the same place, and it also resolves in the shell and in scripts.",
+                });
+            }
+        }
         // A gate with no `preview:` shows only the previous step's reply and
         // whatever files that reply happens to name. The person deciding sees a
         // summary, not the thing — unless the agent remembered.
@@ -86,6 +273,41 @@ export function lintFlow(flow, known) {
                     "always sees the document, not a pointer, even when the agent forgets to name it.",
             });
         }
+        // A step that can act on the outside world, with nothing to say whether
+        // it did. The two answers are a check the runtime can fail on, or a
+        // person; either is fine, neither is not. This is the one lint here that
+        // is an error: every other structural warning costs a reread, and this
+        // one costs a send nobody can take back.
+        if (known?.outwardAgents?.includes(step.agent ?? "") && !step.verify && !step.approve && !step.ask) {
+            warnings.push({
+                step: step.group,
+                line: step.line,
+                level: "error",
+                kind: "outward",
+                message: `[[${step.agent}]] can act outside this workspace and this step has no verify: and no gate`,
+                detail: "A step that sends, posts, buys or publishes needs one of two things: a `verify:` the " +
+                    "runtime can fail the step on — a receipt file, a ledger row, the reply's own verdict " +
+                    "line — or a person, with `!` or `ask:`. Without either, a step that sent nothing and a " +
+                    "step that sent everything end the same way: reported as done.",
+            });
+        }
+        // retry: re-runs a failed step from the start. For a step that sends,
+        // a failure can come after the send — the verify: or the JSON check
+        // failed on a step whose email went — and the retry sends it again. The
+        // runner declines that one retry (see actedThenFailedCheck in runner.ts)
+        // but a crash mid-send still reruns; the author should know.
+        if (known?.outwardAgents?.includes(step.agent ?? "") && step.retry) {
+            warnings.push({
+                step: i,
+                line: step.line,
+                kind: "outward",
+                message: `[[${step.agent}]] can act outside this workspace and this step has retry: ${step.retry}`,
+                detail: "A retry runs the whole step again, sends included. The runner does not retry a failure " +
+                    "that came from the check after the step's tools ran, but an attempt that failed mid-way — " +
+                    "a timeout, a crash, a provider error after the first send — is retried and may repeat " +
+                    "what it already did. Make the send idempotent (check a ledger before sending), or drop retry:.",
+            });
+        }
         // `each: items` fans out over DATA, and only an `output: json` step in
         // an earlier group produces any. Without one the step expands to
         // nothing every time, quietly — the exact failure this file exists for.
@@ -102,6 +324,59 @@ export function lintFlow(flow, known) {
                 });
             }
         }
+        // Two options that read another option to mean anything.
+        if (step.parallel && !step.each) {
+            warnings.push({
+                step: i,
+                line: step.line,
+                message: `"${step.subflow ?? step.agent}" has parallel: but no each:`,
+                detail: "parallel: caps how many instances of a fan-out run at once. Without each: there is one instance, and the line does nothing.",
+            });
+        }
+        if ((step.schema !== undefined || step.schemaPath) && step.output !== "json") {
+            warnings.push({
+                step: i,
+                line: step.line,
+                message: `"${step.subflow ?? step.agent}" has schema: but no output: json`,
+                detail: "schema: describes the JSON value an output: json step returns. Without output: json nothing is parsed, so nothing is checked. Add `output: json`.",
+            });
+        }
+        // The step's own agent. `delegate:` and `on-fail:` are checked below and
+        // the step's target was not, which is the reference that stops the run
+        // rather than degrading it: the runner reaches the step, finds nothing to
+        // run and fails it with `agent "<name>" not found in workspace`. A
+        // scheduled flow does that at 3am, three times, before anyone looks.
+        if (agentNames && step.agent && !step.subflow && !agentNames.has(step.agent)) {
+            warnings.push({
+                step: i,
+                line: step.line,
+                level: "error",
+                message: `[[${step.agent}]] is not an agent in this workspace`,
+                detail: "This step has nothing to run. Every run of this flow fails here. Check the spelling " +
+                    "against `agents/<name>/agent.md` — the name is the folder's, not the file's heading — " +
+                    "or add the agent.",
+            });
+        }
+        // A `verify:` written as a sentence. Two dialects share the key: a line
+        // beginning `contains:`, `not-contains:`, `matches:`, `file:` or
+        // `judge:` is an assertion, and ANYTHING else is handed to a shell. So
+        // `verify: the reply names the clone it made` runs a program called
+        // `the`, exits 127, and fails a step that did its job perfectly — every
+        // run, with an error about bash that says nothing about the flow. The
+        // fix is one word: the same sentence behind `judge:` is graded by a
+        // model, which is what the author meant.
+        if (step.verify && looksLikeProse(step.verify)) {
+            warnings.push({
+                step: i,
+                line: step.line,
+                level: "error",
+                message: `verify: reads as a sentence, so it will be run as a shell command`,
+                detail: `\`${step.verify.slice(0, 60)}${step.verify.length > 60 ? "…" : ""}\` has no assertion ` +
+                    "prefix, so the runtime executes it: the first word becomes the program and the step " +
+                    "fails with `command not found`. Put `judge: ` in front to have a model grade the " +
+                    "claim, or write it as a shell command that exits 0 when the claim holds.",
+            });
+        }
         // A step's OTHER agent references — the ones no parser validates,
         // because they are options rather than the step's target. Both fail
         // silently and at the worst possible moment: a misspelled `delegate:`
@@ -141,7 +416,8 @@ export function lintFlow(flow, known) {
             });
         }
         // `when:` tests the previous results, which the first group doesn't have.
-        if (inFirstGroup && step.when) {
+        // `when: rows of` reads a file instead, which may well exist already.
+        if (inFirstGroup && step.when && whenRowsPath(step.when) === null) {
             warnings.push({
                 step: i,
                 line: step.line,
@@ -150,6 +426,23 @@ export function lintFlow(flow, known) {
                     "always be skipped. Move it after the step whose output it depends on.",
             });
         }
+        // A `[[flow:x]]` step is spliced in as the other flow's own steps, and
+        // only its `?` and its instruction survive the splice. Everything else
+        // written under it was dropped without a word — a `when:` that never
+        // gated, a `!` that never asked.
+        if (step.subflow) {
+            const dropped = SUBFLOW_DROPPED.filter(([field]) => step[field] !== undefined).map(([, name]) => name);
+            if (dropped.length) {
+                warnings.push({
+                    step: i,
+                    line: step.line,
+                    message: `"flow:${step.subflow}" has ${dropped.join(", ")}, which a nested flow ignores`,
+                    detail: "A nested flow runs as its own steps, with their own options; what is written under the " +
+                        "[[flow:…]] line is not applied to them. Put the option on the steps inside that flow, " +
+                        "or make this an agent step.",
+                });
+            }
+        }
     });
     // The same agent twice in one parallel group gets two identical contexts and
     // no way to divide the work — almost always a drag that landed wrong.
@@ -172,6 +465,43 @@ export function lintFlow(flow, known) {
                 });
             }
         }
+        // A gate parks its whole group, and on approval every step in it starts
+        // together. So `5. [[editor]]` beside `5! [[publisher]]` does not mean
+        // "edit, then ask, then publish": the editor waits for the approval too,
+        // and the publisher runs alongside it, reading files the editor has not
+        // written yet. medium-desk's publisher did exactly that on 2026-09-28
+        // (run-mukdjeh8-cr66): post.json was still empty when it started.
+        const gated = (s) => Boolean(s.approve || s.ask || s.waitFor === "event");
+        const gates = steps.filter(gated);
+        const others = steps.filter((s) => !gated(s));
+        if (gates.length && others.length) {
+            const name = (s) => (s.subflow ? `flow:${s.subflow}` : s.agent);
+            warnings.push({
+                step: flow.steps.indexOf(gates[0]),
+                line: gates[0].line,
+                message: `"${name(gates[0])}" waits for a person but shares number ${group} with "${others.map(name).join('", "')}"`,
+                detail: "Steps with the same number run together. The gate holds all of them until it is " +
+                    "approved, then they start at once — so the gated step cannot use what the others " +
+                    `produce, and the approver sees none of it. To run the gated step after them, number it ${group + 1}.`,
+            });
+        }
+    }
+    // A number with nothing under it is harmless to the runner, which only
+    // orders groups, but it is almost always one step numbered twice: 3, 3, 4,
+    // 5, 5!, 7 is a renumbering that missed the 5! that was meant to be 6.
+    const numbers = [...byGroup.keys()].sort((a, b) => a - b);
+    for (let i = 1; i < numbers.length; i++) {
+        if (numbers[i] - numbers[i - 1] < 2)
+            continue;
+        const missing = numbers[i - 1] + 1 === numbers[i] - 1 ? `${numbers[i - 1] + 1}` : `${numbers[i - 1] + 1} to ${numbers[i] - 1}`;
+        const shared = [...byGroup.entries()].filter(([, s]) => s.length > 1).map(([g]) => g);
+        warnings.push({
+            step: null,
+            message: `step numbers go from ${numbers[i - 1]} to ${numbers[i]}; nothing is numbered ${missing}`,
+            detail: "The same number means \"run together\", so a gap usually means a step kept an old number " +
+                "after a renumbering" +
+                (shared.length ? ` — ${shared.map((g) => `number ${g}`).join(", ")} has more than one step; check whether one of them belongs in the gap.` : "."),
+        });
     }
     return warnings;
 }