@markii/host 0.0.0-stage → 0.17.0-next.2

package/dist/host/conformance-runner.js

@@ -0,0 +1,634 @@
+/**
+ * Executes one `conformance/host/<scenario>` case (batch 11 Phase 3,
+ * `tmp/W11-corpus-design.md`) against any real `HostAdapter`, so a single
+ * plain-data corpus proves the three apps' adapters behave the same.
+ *
+ * `conformance/host/` itself stays plain data (no TypeScript, matching the
+ * repo's existing corpora): a `note.mk.md`, a `scenario.json` (the ordered
+ * actions), an `expected.json` (the outcomes), and an optional `files/`
+ * directory. This module is the one place that reads that data, drives it
+ * through a `MarkiiHost`, and reports what happened.
+ *
+ * THE CAPABILITY RULE (the design's crux): `ACTION_CAPABILITY` is the one
+ * static map from an action's type to the capability it needs.
+ * `hostDeclaresCapability` checks the map against the ADAPTER UNDER TEST,
+ * never against the scenario data — the corpus itself never encodes
+ * per-host knowledge. `createMarkiiHost`'s own behaviors already return the
+ * shared `Unsupported` shape when a capability is missing; this module's
+ * job is to know that the fact should be a PASS (the correct thing for
+ * this host to say), not something to diff against `expected.json`.
+ *
+ * THE SCRIPTS-DISABLED SCENARIO (`10-scripts-disabled`) is not a
+ * `MarkiiHost` capability at all: the device-local execution switch is
+ * decided by each app's OWN call site (`preview-panel.ts`, `main.ts`)
+ * before it ever reaches `host.run`, using the shared, host-neutral
+ * wording in `./script-execution.js`. This module reproduces that same
+ * pre-check here, over the same shared functions and the SAME
+ * `adapter.labels` a real call site would use, so the scenario proves
+ * wording parity across hosts without inventing a capability the contract
+ * does not have. See `tmp/W11-phase3.md` for why this is a reported
+ * finding, not a silent workaround.
+ *
+ * SIMULATED EXECUTION. A `run` action's `simulate` field (plain data) tells
+ * this module what the isolate would have produced, instead of spawning a
+ * real Lua worker: whether the requested value needed a specific granted
+ * host, and whether it was a write blocked under the read-only tier. Real
+ * sandbox enforcement is `conformance/executor/`'s job; this corpus proves
+ * the HOST-LEVEL plumbing around it (grant prompts, wording, capability
+ * declarations, trigger forwarding) — see this file's `simulatedRunResult`.
+ * Every OTHER action runs through the real, unmocked shared behavior code.
+ */
+import { existsSync, readFileSync, readdirSync } from 'node:fs';
+import { dirname } from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { createHtmlExportEngine } from '../html-engine.js';
+import { createNodeArchiveExtractFs } from '../node-fs.js';
+import { usesPackTrust } from './pack-trust-behavior.js';
+import { join, relative, sep } from 'node:path';
+import { zipSync } from 'fflate';
+import { conformanceDir } from '@markii/core/corpus';
+import { createMarkiiHost, } from './create-host.js';
+import { scriptsDisabledDiagnosticLine, scriptsDisabledNotice, } from './script-execution.js';
+// --- corpus location ----------------------------------------------------
+/**
+ * `conformance/host/`, resolved relative to this module's own file, never
+ * `process.cwd()`. An installed package carries the corpus at
+ * `dist/conformance/host/`, a sibling of `dist/host/`; a repo checkout runs
+ * from `src/host/` and finds the repo-root corpus through
+ * `@markii/core/corpus`'s `conformanceDir`.
+ */
+export function hostConformanceDir() {
+    const here = dirname(fileURLToPath(import.meta.url));
+    const packaged = join(here, '..', 'conformance', 'host');
+    if (existsSync(packaged))
+        return packaged;
+    return join(conformanceDir(), 'host');
+}
+/** Every scenario directory name under `dir`, sorted. */
+export function listHostScenarioNames(dir = hostConformanceDir()) {
+    return readdirSync(dir, { withFileTypes: true })
+        .filter((entry) => entry.isDirectory())
+        .map((entry) => entry.name)
+        .sort();
+}
+/** Loads one scenario's three files. Never validates beyond `JSON.parse` — a malformed fixture is a corpus bug, and throwing loudly here is the right failure mode for it. */
+export function loadHostScenario(name, baseDir = hostConformanceDir()) {
+    const dir = join(baseDir, name);
+    const noteText = readFileSync(join(dir, 'note.mk.md'), 'utf8');
+    const scenario = JSON.parse(readFileSync(join(dir, 'scenario.json'), 'utf8'));
+    const expected = JSON.parse(readFileSync(join(dir, 'expected.json'), 'utf8'));
+    return { dir, noteText, scenario, expected };
+}
+// --- .mkp archive zipping (from a plain files/ directory, never a committed binary) ---
+function readDirRecursive(dir, base) {
+    const out = {};
+    for (const entry of readdirSync(dir, { withFileTypes: true })) {
+        const full = join(dir, entry.name);
+        if (entry.isDirectory()) {
+            Object.assign(out, readDirRecursive(full, base));
+        }
+        else {
+            const relativeName = relative(base, full).split(sep).join('/');
+            out[relativeName] = new Uint8Array(readFileSync(full));
+        }
+    }
+    return out;
+}
+/** Zips a scenario's `files/<name>` directory into `.mkp` archive bytes on the fly — this is how the corpus ships pack fixtures as plain files instead of a committed binary. */
+export function zipArchiveDirectory(archiveDir) {
+    return zipSync(readDirRecursive(archiveDir, archiveDir));
+}
+/** The one static map from an action's type to the capability it needs. `undefined` means every host supports it (there is nothing to decline). */
+export const ACTION_CAPABILITY = {
+    open: undefined,
+    run: 'isolate',
+    exportHtml: 'export-format',
+    exportAnsi: 'export-format',
+    installPack: 'packs',
+    loadPacks: 'packs',
+    removePack: 'packs',
+    placePack: 'pack-trust',
+    listPacks: 'pack-trust',
+    enablePack: 'pack-trust',
+    disablePack: 'pack-trust',
+    completeAt: 'editor',
+    hoverAt: 'editor',
+    insertComponent: 'editor',
+};
+/** An action's capability: the static map, except that a pack action flagged `requiresTrust` needs the device trust list. */
+function capabilityFor(action) {
+    if ((action.type === 'installPack' ||
+        action.type === 'loadPacks' ||
+        action.type === 'removePack') &&
+        action.requiresTrust === true) {
+        return 'pack-trust';
+    }
+    return ACTION_CAPABILITY[action.type];
+}
+function exportFormatFor(action) {
+    return action.type === 'exportAnsi' ? 'ansi' : 'html';
+}
+/** Whether `adapter` declares the capability `action` needs. Never consults the scenario or `expected.json` — only the adapter under test. */
+export function hostDeclaresCapability(adapter, action) {
+    const capability = capabilityFor(action);
+    if (capability === undefined)
+        return true;
+    if (capability === 'pack-trust')
+        return usesPackTrust(adapter);
+    if (capability === 'isolate')
+        return adapter.isolate !== undefined;
+    if (capability === 'editor')
+        return adapter.editor !== undefined;
+    if (capability === 'packs') {
+        if (adapter.packs === undefined)
+            return false;
+        if (action.type === 'installPack') {
+            return adapter.packs.installRoot !== undefined;
+        }
+        if (action.type === 'removePack') {
+            return (adapter.packs.installRoot !== undefined &&
+                adapter.packs.removeFolder !== undefined);
+        }
+        return true;
+    }
+    // 'export-format'
+    return adapter.exports?.formats.has(exportFormatFor(action)) ?? false;
+}
+// --- the scripted prompt queue --------------------------------------------
+/**
+ * Flattens every `run`/`installPack` action's `answers` into one queue, in
+ * document order — the same order the grant/consent prompts fire in, since
+ * `expected.json`'s `prompts` array is itself the ordered proof of that.
+ * `next()` THROWS once exhausted rather than returning a default: an
+ * unanswered prompt means the scenario and the implementation disagree
+ * about how many questions get asked, which must never be silently read
+ * as a deny (`tmp/W11-corpus-design.md`, "The prompt policy").
+ */
+export function createAnswerQueue(scenario) {
+    const answers = [];
+    for (const action of scenario.actions) {
+        const source = action.type === 'run' ||
+            action.type === 'installPack' ||
+            action.type === 'enablePack'
+            ? action.answers
+            : undefined;
+        for (const answer of source ?? [])
+            answers.push(answer === 'allow');
+    }
+    let index = 0;
+    return {
+        next() {
+            if (index >= answers.length) {
+                throw new Error(`scripted prompt queue exhausted after ${String(index)} answer(s): ` +
+                    'the scenario and the implementation disagree about how many ' +
+                    'questions get asked.');
+            }
+            return answers[index++];
+        },
+    };
+}
+// --- simulated execution ---------------------------------------------------
+function simulatedRunResult(sim, options) {
+    const name = sim?.name ?? 'value';
+    if (sim?.requiresManualTier === true && options.trigger !== 'manual') {
+        return {
+            values: {},
+            failures: [
+                {
+                    name,
+                    message: `${name}: a write was attempted under the read-only "${options.trigger ?? 'manual'}" tier.`,
+                    kind: 'tier-blocked',
+                },
+            ],
+            cacheSnapshot: options.cacheSnapshot,
+        };
+    }
+    if (sim?.requiresHost !== undefined &&
+        !options.netAllowlist.includes(sim.requiresHost)) {
+        return {
+            values: {},
+            failures: [
+                {
+                    name,
+                    message: `${name}: host "${sim.requiresHost}" was not granted.`,
+                    kind: 'capability-denied',
+                },
+            ],
+            cacheSnapshot: options.cacheSnapshot,
+        };
+    }
+    return {
+        values: { [name]: { status: 'fresh', value: sim?.value ?? 1 } },
+        failures: [],
+        cacheSnapshot: options.cacheSnapshot,
+    };
+}
+/** `unsupported()` in `./create-host.js` writes exactly this shape of line; scenarios describe the fully-capable host's behavior, so a decline line is not something `expected.json` should have to also state per host. */
+const UNSUPPORTED_DIAGNOSTIC_LINE = /^declined \([a-z-]+ unsupported on /;
+function decodeWritten(writtenFiles, path) {
+    if (path === undefined)
+        return undefined;
+    const bytes = writtenFiles.get(path);
+    return bytes === undefined ? undefined : new TextDecoder().decode(bytes);
+}
+/**
+ * Runs `scenario` against `adapter` end to end. `adapter` is the REAL,
+ * already-constructed host adapter under test (built by the caller from
+ * its own injected dependencies plus a scripted prompt source, e.g.
+ * `createAnswerQueue`); this function never builds one itself, so it works
+ * identically for any `HostAdapter`.
+ */
+export async function runHostScenario(adapter, scenario, noteText, scenarioDir, options = {}) {
+    const diagnostics = [];
+    const prompts = [];
+    const writtenFiles = new Map();
+    let promptQueueError;
+    const recording = {
+        ...adapter,
+        now: () => scenario.clock,
+        diagnostics: (line) => {
+            diagnostics.push(line);
+            adapter.diagnostics(line);
+        },
+        prompt: async (request) => {
+            prompts.push(request.message);
+            return adapter.prompt(request);
+        },
+        writeFile: async (path, bytes) => {
+            writtenFiles.set(path, bytes);
+            await adapter.writeFile(path, bytes);
+        },
+    };
+    let currentSimulation;
+    const host = createMarkiiHost(recording, {
+        // The runner is Node-only test tooling, so it supplies the real Node
+        // implementations of the seams the host layer takes by injection.
+        htmlEngine: createHtmlExportEngine(),
+        archiveExtractFs: createNodeArchiveExtractFs(),
+        spawnRun: (spawnOptions) => Promise.resolve(simulatedRunResult(currentSimulation, spawnOptions)),
+        ...(options.renderAnsi ? { renderAnsi: options.renderAnsi } : {}),
+    });
+    const notePath = join(scenarioDir, 'note.mk.md');
+    const outcomes = [];
+    for (const action of scenario.actions) {
+        try {
+            outcomes.push(await runOneAction(host, recording, action, {
+                noteText,
+                notePath,
+                scenarioDir,
+                writtenFiles,
+                setSimulation: (sim) => {
+                    currentSimulation = sim;
+                },
+            }));
+        }
+        catch (err) {
+            promptQueueError = err instanceof Error ? err.message : String(err);
+            break;
+        }
+    }
+    return {
+        outcomes,
+        diagnostics,
+        prompts,
+        ...(promptQueueError !== undefined ? { promptQueueError } : {}),
+    };
+}
+async function runOneAction(host, adapter, action, ctx) {
+    // A trust-scenario action on a host without the trust list never reaches
+    // the host: the declaration is what the corpus records.
+    if (capabilityFor(action) === 'pack-trust' && !usesPackTrust(adapter)) {
+        const detail = 'this host does not gate installed packs on a trust list.';
+        adapter.diagnostics(`declined (pack-trust unsupported on ${adapter.id}): ${detail}`);
+        return { kind: 'unsupported', capability: 'pack-trust', detail };
+    }
+    switch (action.type) {
+        case 'open':
+            return host.open({ path: ctx.notePath });
+        case 'run': {
+            if (action.scriptsDisabled === true) {
+                // The device-local scripts-off switch is not a `MarkiiHost`
+                // capability (see this file's top comment): it is decided at each
+                // app's own call site, before `host.run` is ever reached, using
+                // the shared wording in `./script-execution.js`. Reproduced here
+                // over the SAME functions and the adapter's own `labels` so the
+                // scenario proves wording parity without inventing a capability.
+                adapter.diagnostics(scriptsDisabledDiagnosticLine(action.trigger, adapter.labels));
+                return {
+                    kind: 'declined',
+                    reason: scriptsDisabledNotice(action.trigger, adapter.labels) ??
+                        scriptsDisabledDiagnosticLine(action.trigger, adapter.labels),
+                };
+            }
+            ctx.setSimulation(action.simulate);
+            return host.run({
+                documentKey: ctx.notePath,
+                text: ctx.noteText,
+                trigger: action.trigger,
+            });
+        }
+        case 'exportHtml':
+        case 'exportAnsi': {
+            const outcome = await host.exportNote({
+                format: exportFormatFor(action),
+                notePath: ctx.notePath,
+                text: ctx.noteText,
+            });
+            return augmentExportOutcome(outcome, ctx.writtenFiles);
+        }
+        case 'installPack': {
+            const archiveBytes = zipArchiveDirectory(join(ctx.scenarioDir, action.archive));
+            return host.installPack({ archiveBytes, archivePath: action.archive });
+        }
+        case 'loadPacks':
+            return host.loadPacks();
+        case 'removePack':
+            return host.removePack(action.namespace);
+        case 'placePack': {
+            const installRoot = adapter.packs?.installRoot?.();
+            if (installRoot === undefined)
+                return { kind: 'failed' };
+            const files = readDirRecursive(join(ctx.scenarioDir, action.archive), join(ctx.scenarioDir, action.archive));
+            for (const [name, bytes] of Object.entries(files)) {
+                await adapter.writeFile(join(installRoot, action.namespace, ...name.split('/')), bytes);
+            }
+            return { kind: 'placed', namespace: action.namespace };
+        }
+        case 'listPacks':
+            return host.installedPacks();
+        case 'enablePack':
+            return host.enablePack(action.namespace);
+        case 'disablePack':
+            return host.disablePack(action.namespace);
+        case 'completeAt':
+            return host.completeAt({ line: action.line, column: action.column });
+        case 'hoverAt':
+            return host.hoverAt({ line: action.line, column: action.column });
+        case 'insertComponent':
+            return host.insertComponent({ name: action.name });
+    }
+}
+function augmentExportOutcome(outcome, writtenFiles) {
+    if (outcome.kind !== 'exported')
+        return outcome;
+    return { ...outcome, content: decodeWritten(writtenFiles, outcome.path) };
+}
+// --- diffing against expected.json -----------------------------------------
+/** `expected[i]` compared against `actual[i]` for each index up to `expected`'s own length; a length mismatch is reported once rather than index-by-index noise. */
+export function containsInOrder(actual, expected) {
+    if (actual.length !== expected.length) {
+        return [
+            `count: expected ${String(expected.length)}, got ${String(actual.length)} ` +
+                `(actual: ${JSON.stringify(actual)})`,
+        ];
+    }
+    const mismatches = [];
+    expected.forEach((exp, i) => {
+        if (!(actual[i] ?? '').includes(exp)) {
+            mismatches.push(`[${String(i)}]: expected to contain "${exp}", got "${actual[i]}"`);
+        }
+    });
+    return mismatches;
+}
+/** Extracts `{name: value}` for every non-error `StoredValue` in a `'ran'` outcome's `values`, for a plain-data comparison against `expected.json`. */
+function landedValues(values) {
+    if (typeof values !== 'object' || values === null)
+        return {};
+    const out = {};
+    for (const [name, entry] of Object.entries(values)) {
+        if (typeof entry === 'object' &&
+            entry !== null &&
+            'status' in entry &&
+            entry.status !== 'error' &&
+            'value' in entry) {
+            out[name] = entry.value;
+        }
+    }
+    return out;
+}
+function failureKinds(failures) {
+    if (!Array.isArray(failures))
+        return [];
+    return failures.map((f) => typeof f === 'object' && f !== null && 'kind' in f
+        ? String(f.kind)
+        : 'unknown');
+}
+/**
+ * A coarse, per-`kind` comparison between one action's real outcome and its
+ * `expected.json` entry. Deliberately partial (matching the design's own
+ * "deliberately coarse" posture): a `'ran'` outcome compares landed values
+ * and failure KINDS, never timings or raw messages; an `'exported'` outcome
+ * compares `contains`/`excludes` substrings against the written file's
+ * content, never the whole document.
+ */
+export function outcomeMismatches(actual, expected) {
+    if (expected === undefined)
+        return [];
+    const actualRecord = typeof actual === 'object' && actual !== null
+        ? actual
+        : {};
+    if (actualRecord['kind'] !== expected['kind']) {
+        return [
+            `kind: expected "${String(expected['kind'])}", got "${String(actualRecord['kind'])}"`,
+        ];
+    }
+    const mismatches = [];
+    switch (expected['kind']) {
+        case 'ran': {
+            const wantValues = expected['values'] ?? {};
+            const gotValues = landedValues(actualRecord['values']);
+            if (JSON.stringify(gotValues) !== JSON.stringify(wantValues)) {
+                mismatches.push(`values: expected ${JSON.stringify(wantValues)}, got ${JSON.stringify(gotValues)}`);
+            }
+            const wantKinds = expected['failureKinds'] ?? [];
+            const gotKinds = failureKinds(actualRecord['failures']);
+            if (JSON.stringify(gotKinds) !== JSON.stringify(wantKinds)) {
+                mismatches.push(`failure kinds: expected ${JSON.stringify(wantKinds)}, got ${JSON.stringify(gotKinds)}`);
+            }
+            if (expected['traceOk'] !== undefined) {
+                const trace = actualRecord['trace'];
+                if (trace?.ok !== expected['traceOk']) {
+                    mismatches.push(`trace.ok: expected ${String(expected['traceOk'])}, got ${String(trace?.ok)}`);
+                }
+            }
+            break;
+        }
+        case 'exported': {
+            const content = String(actualRecord['content'] ?? '');
+            for (const needle of expected['contains'] ?? []) {
+                if (!content.includes(needle)) {
+                    mismatches.push(`export should contain "${needle}"`);
+                }
+            }
+            for (const needle of expected['excludes'] ?? []) {
+                if (content.includes(needle)) {
+                    mismatches.push(`export should not contain "${needle}"`);
+                }
+            }
+            break;
+        }
+        case 'installed': {
+            if (expected['namespace'] !== undefined &&
+                actualRecord['namespace'] !== expected['namespace']) {
+                mismatches.push(`namespace: expected "${String(expected['namespace'])}", got "${String(actualRecord['namespace'])}"`);
+            }
+            if (expected['replaced'] !== undefined &&
+                actualRecord['replaced'] !== expected['replaced']) {
+                mismatches.push(`replaced: expected ${String(expected['replaced'])}, got ${String(actualRecord['replaced'])}`);
+            }
+            break;
+        }
+        case 'declined': {
+            if (expected['step'] !== undefined &&
+                actualRecord['step'] !== expected['step']) {
+                mismatches.push(`step: expected "${String(expected['step'])}", got "${String(actualRecord['step'])}"`);
+            }
+            if (expected['reasonContains'] !== undefined &&
+                !String(actualRecord['reason'] ?? '').includes(String(expected['reasonContains']))) {
+                mismatches.push(`reason: expected to contain "${String(expected['reasonContains'])}", got "${String(actualRecord['reason'])}"`);
+            }
+            break;
+        }
+        case 'listed': {
+            const names = (list, pick) => (Array.isArray(list) ? list.map(pick) : []).map(String).sort();
+            const gotEnabled = names(actualRecord['enabled'], (entry) => entry.namespace);
+            const gotNotEnabled = names(actualRecord['notEnabled'], (entry) => entry);
+            for (const [key, got] of [
+                ['enabled', gotEnabled],
+                ['notEnabled', gotNotEnabled],
+            ]) {
+                const want = [...(expected[key] ?? [])].sort();
+                if (JSON.stringify(got) !== JSON.stringify(want)) {
+                    mismatches.push(`${key}: expected ${JSON.stringify(want)}, got ${JSON.stringify(got)}`);
+                }
+            }
+            break;
+        }
+        case 'enabled':
+        case 'disabled':
+        case 'not-enabled':
+        case 'placed':
+        case 'removed': {
+            if (expected['namespace'] !== undefined &&
+                actualRecord['namespace'] !== expected['namespace']) {
+                mismatches.push(`namespace: expected "${String(expected['namespace'])}", got "${String(actualRecord['namespace'])}"`);
+            }
+            break;
+        }
+        case 'loaded': {
+            const want = [...(expected['namespaces'] ?? [])].sort();
+            const got = [...(actualRecord['namespaces'] ?? [])].sort();
+            if (JSON.stringify(got) !== JSON.stringify(want)) {
+                mismatches.push(`namespaces: expected ${JSON.stringify(want)}, got ${JSON.stringify(got)}`);
+            }
+            // The component counts are engine-neutral on purpose: a pack with one
+            // view per engine has exactly one component unavailable on every host.
+            for (const [expectedKey, actualKey] of [
+                ['componentCount', 'components'],
+                ['unavailableCount', 'unavailable'],
+            ]) {
+                if (expected[expectedKey] === undefined)
+                    continue;
+                const list = actualRecord[actualKey];
+                const count = Array.isArray(list) ? list.length : 0;
+                if (count !== expected[expectedKey]) {
+                    mismatches.push(`${expectedKey}: expected ${String(expected[expectedKey])}, got ${String(count)}`);
+                }
+            }
+            break;
+        }
+        case 'completions': {
+            const got = actualRecord['labels'] ?? [];
+            for (const label of expected['labels'] ?? []) {
+                if (!got.includes(label)) {
+                    mismatches.push(`completions should include "${label}"`);
+                }
+            }
+            for (const key of ['replaceStart', 'replaceEnd', 'tokenStart']) {
+                if (expected[key] !== undefined &&
+                    actualRecord[key] !== expected[key]) {
+                    mismatches.push(`${key}: expected ${String(expected[key])}, got ${String(actualRecord[key])}`);
+                }
+            }
+            if (expected['fenceEditCount'] !== undefined) {
+                const got = Array.isArray(actualRecord['fenceEdits'])
+                    ? actualRecord['fenceEdits'].length
+                    : 0;
+                if (got !== expected['fenceEditCount']) {
+                    mismatches.push(`fenceEditCount: expected ${String(expected['fenceEditCount'])}, got ${String(got)}`);
+                }
+            }
+            break;
+        }
+        case 'hover': {
+            const contains = expected['contains'];
+            if (contains !== undefined &&
+                !String(actualRecord['text'] ?? '').includes(contains)) {
+                mismatches.push(`hover text should contain "${contains}"`);
+            }
+            for (const key of ['start', 'end']) {
+                if (expected[key] !== undefined &&
+                    actualRecord[key] !== expected[key]) {
+                    mismatches.push(`${key}: expected ${String(expected[key])}, got ${String(actualRecord[key])}`);
+                }
+            }
+            break;
+        }
+        default:
+            break; // kind equality is the whole check for every other outcome.
+    }
+    return mismatches;
+}
+/**
+ * The full assertion an app's conformance test makes for one scenario: for
+ * every action whose capability `adapter` declares, its outcome matches
+ * `expected`'s; for every other action, the outcome is exactly the shared
+ * `unsupported` record. Returns a list of human-readable problems — empty
+ * means the scenario passed.
+ */
+export function assertHostScenario(adapter, scenario, expected, result) {
+    const problems = [];
+    if (result.promptQueueError !== undefined) {
+        problems.push(`prompt queue: ${result.promptQueueError}`);
+    }
+    scenario.actions.forEach((action, i) => {
+        const actual = result.outcomes[i];
+        const declared = hostDeclaresCapability(adapter, action);
+        if (!declared) {
+            const capability = capabilityFor(action);
+            const actualRecord = typeof actual === 'object' && actual !== null
+                ? actual
+                : {};
+            if (actualRecord['kind'] !== 'unsupported' ||
+                actualRecord['capability'] !== capability) {
+                problems.push(`action ${String(i)} (${action.type}): expected {kind:"unsupported", capability:"${String(capability)}"}, got ${JSON.stringify(actual)}`);
+            }
+            return;
+        }
+        for (const problem of outcomeMismatches(actual, expected.outcomes[i])) {
+            problems.push(`action ${String(i)} (${action.type}): ${problem}`);
+        }
+    });
+    // `expected.json`'s `diagnostics`/`prompts` describe what a FULLY CAPABLE
+    // host does. When every action in this scenario needs a capability the
+    // adapter lacks (a whole scenario built around one capability domain,
+    // e.g. every action in `06-install-pack-then-render` needs `packs`),
+    // nothing meaningful ran and there is nothing to compare — the
+    // per-action `unsupported` check above already covers this host
+    // honestly. Comparing an unrelated host's empty prompt/diagnostics list
+    // against a capable host's expectation here would be double-counting
+    // the same capability gap as a second, spurious failure.
+    const capabilityGatedActions = scenario.actions.filter((action) => capabilityFor(action) !== undefined);
+    const anyActionDeclared = capabilityGatedActions.length === 0 ||
+        capabilityGatedActions.some((action) => hostDeclaresCapability(adapter, action));
+    if (anyActionDeclared) {
+        const filteredDiagnostics = result.diagnostics.filter((line) => !UNSUPPORTED_DIAGNOSTIC_LINE.test(line));
+        for (const problem of containsInOrder(filteredDiagnostics, expected.diagnostics)) {
+            problems.push(`diagnostics ${problem}`);
+        }
+        for (const problem of containsInOrder(result.prompts, expected.prompts)) {
+            problems.push(`prompts ${problem}`);
+        }
+    }
+    return problems;
+}