Compare commits

..
20 changed files with 557 additions and 137 deletions
+2
View File
@@ -63,5 +63,7 @@ jobs:
NODE_ENV: test
ELECTRON_RUN_AS_NODE: "1"
- run: npm run build
# Check the fresh trace before any static copying/curation.
- run: node scripts/smoke-provider-runtime.mjs
- run: node scripts/build-server.mjs
- run: node scripts/smoke-server.mjs --electron
+21
View File
@@ -129,6 +129,8 @@ jobs:
NODE_ENV: test
ELECTRON_RUN_AS_NODE: "1"
- run: npm run build
# Check the fresh trace before any static copying/curation.
- run: node scripts/smoke-provider-runtime.mjs
- run: node scripts/build-server.mjs
- run: node scripts/smoke-server.mjs --electron
- name: Validate macOS release secrets
@@ -169,6 +171,25 @@ jobs:
- name: Windows/Linux rehearsal (no publish)
if: github.event_name == 'workflow_dispatch' && runner.os != 'macOS'
run: npx electron-builder ${{ matrix.builder_args }} --publish never
- name: Initialize packaged providers offline (host and packaged Electron Node)
run: |
case "$RUNNER_OS" in
macOS)
app="dist/mac-arm64/LearnHarness.app/Contents"
payload="$app/Resources/server/standalone"
runtime="$app/MacOS/LearnHarness"
;;
Windows)
payload="dist/win-unpacked/resources/server/standalone"
runtime="dist/win-unpacked/LearnHarness.exe"
;;
Linux)
payload="dist/linux-unpacked/resources/server/standalone"
runtime="dist/linux-unpacked/learnharness"
;;
esac
node scripts/smoke-provider-runtime.mjs --standalone "$payload"
node scripts/smoke-provider-runtime.mjs --standalone "$payload" --runtime "$runtime"
- name: Exercise packaged demo runtime (own Electron, no system Chrome)
run: |
case "$RUNNER_OS" in
+13 -35
View File
@@ -131,19 +131,6 @@ export function composeGraphTools(agent: AgentSettings | undefined, on: boolean)
return { ...agent, graphTools: { enabled: on } };
}
/** The research tool runs on the search port — the teacher registry DROPS
* it when no engine is live (chat-handler's searchAvailable, the same
* /api/search signal the section fetches). Pure — exported for tests.
* Teacher-stream flag 2026-09-22: the settings row grays on the SAME
* signal, so the row and the actual registry availability can never
* disagree. `search === null` (loading/failed) grays NOTHING. */
export function researchToolUnavailable(
toolId: string,
search: { available: boolean } | null,
): boolean {
return toolId === "research" && search !== null && !search.available;
}
/** Compose: one slot's model/thinking override. Empty model = absent
* override (Default — falls to the factory pin); empty thinking = the
* tier keeps its resolved thinking budget. Per-field independence, same
@@ -792,28 +779,19 @@ export function AgentSection({
</SettingsRow>
);
})()}
{TOGGLE_TOOLS.map((id) => {
// Research needs a live search engine; the row + toggle gray
// on the same signal the registry drops the tool on (see
// researchToolUnavailable). The reason rides as the hint.
const searchDown = researchToolUnavailable(id, search);
return (
<SettingsRow
key={id}
label={`Tool · ${id}`}
hint={searchDown ? search?.reason : undefined}
testid={`tool-row-${id}`}
dimmed={searchDown}
>
<Toggle
checked={agent?.tools?.[id] ?? true}
disabled={searchDown}
label={`tool ${id}`}
onChange={(on) => onSectionChange(composeTool(agent, id, on))}
/>
</SettingsRow>
);
})}
{TOGGLE_TOOLS.map((id) => (
<SettingsRow
key={id}
label={`Tool · ${id}`}
testid={`tool-row-${id}`}
>
<Toggle
checked={agent?.tools?.[id] ?? true}
label={`tool ${id}`}
onChange={(on) => onSectionChange(composeTool(agent, id, on))}
/>
</SettingsRow>
))}
{(() => {
const engine = agent?.search?.engine ?? null;
const anyKey = engineKeys.tavily === true || engineKeys.brave === true;
+16 -4
View File
@@ -17,7 +17,8 @@ const nextConfig: NextConfig = {
// into the DMG (the packaged app never reads it: the shell passes
// LH_AUTH_FILE → userData). Excluded here; build-server.mjs also REFUSES
// to package any standalone still containing it. .env* belt-and-braces
// the same class (not observed swept, same seam).
// the same class. Next also copies loaded bare .env after tracing;
// the build-server refusal remains required even with these excludes.
//
// ⚠ EXCLUDE PATTERNS ARE LOAD-BEARING IN WAYS THE DOCS DON'T DESCRIBE
// (empirically established 2026-10-08): adding dir excludes like
@@ -26,8 +27,11 @@ const nextConfig: NextConfig = {
// under their dist/ — the traced copy collapsed to a bare package.json
// and the packaged server died on ERR_MODULE_NOT_FOUND; openai/
// @anthropic-ai — pi-ai's deps — vanished from the trace entirely).
// Anchored "./"-prefixed forms behave the same. Only the three entries
// below are verified safe (complete externals trace + clean payload).
// Anchored "./"-prefixed forms behave the same. **/data/**/* fixes
// suffix matching in Next's JavaScript collector, but NOT its native
// Turbopack tracer: that still prunes gcp-metadata and its private deps.
// Keep the narrow closure workaround below AND the native runtime gate.
// Real SDK data/ directories still need their explicit includes too.
// The recursive-DMG payload (repo dist/ swept into the bundle, shipping
// the previous build inside the next) is instead prevented by the
// desktop:dist clean step AND targeted post-trace curation in
@@ -37,7 +41,7 @@ const nextConfig: NextConfig = {
// stay intact, and the script guards against leaked data/keys.
outputFileTracingExcludes: {
"*": [
"data/**/*",
"**/data/**/*",
"auth.json",
".env*",
],
@@ -71,6 +75,14 @@ const nextConfig: NextConfig = {
"node_modules/@resvg/resvg-js-*/**/*",
"node_modules/@earendil-works/pi-ai/dist/providers/data/**/*",
"node_modules/@earendil-works/pi-agent-core/dist/**/*.json",
// Next 16.3.2 native Turbopack applies unanchored excludes while
// traversing modules; **/data/**/* still prunes gcp-metadata and
// its private runtime chain. Includes are applied after traversal.
// This is the verified GoogleAuth closure, not a general dep copy;
// the offline native Pi gate must catch future SDK closure changes.
"node_modules/gcp-metadata/**/*",
"node_modules/json-bigint/**/*",
"node_modules/bignumber.js/**/*",
],
},
// Framework-free core lives in src/ (plain TS, imported by routes and CLI alike).
+2 -2
View File
@@ -1,12 +1,12 @@
{
"name": "learnharness",
"version": "0.3.4",
"version": "0.3.5-preview.1",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "learnharness",
"version": "0.3.4",
"version": "0.3.5-preview.1",
"license": "AGPL-3.0-only",
"dependencies": {
"@earendil-works/pi-agent-core": "1.4.0",
+2 -2
View File
@@ -1,6 +1,6 @@
{
"name": "learnharness",
"version": "0.3.4",
"version": "0.3.5-preview.1",
"license": "AGPL-3.0-only",
"author": {
"name": "Amos Blomqvist",
@@ -21,7 +21,7 @@
"typecheck": "tsc --noEmit",
"desktop:dev": "concurrently -k \"next dev\" \"LH_DEV_URL=http://localhost:3000 electron .\"",
"dev:fresh": "bash scripts/dev-fresh.sh",
"desktop:dist": "rm -rf dist .next/standalone && next build && node scripts/build-server.mjs && node scripts/smoke-server.mjs && DEVELOPER_DIR=${DEVELOPER_DIR:-/Applications/Xcode.app/Contents/Developer} APPLE_KEYCHAIN_PROFILE=LH_NOTARY APPLE_TEAM_ID=JS6XN22HY6 electron-builder --mac --arm64"
"desktop:dist": "rm -rf dist .next/standalone && next build && node scripts/smoke-provider-runtime.mjs && node scripts/build-server.mjs && node scripts/smoke-server.mjs && DEVELOPER_DIR=${DEVELOPER_DIR:-/Applications/Xcode.app/Contents/Developer} APPLE_KEYCHAIN_PROFILE=LH_NOTARY APPLE_TEAM_ID=JS6XN22HY6 electron-builder --mac --arm64"
},
"dependencies": {
"@earendil-works/pi-agent-core": "1.4.0",
+130
View File
@@ -0,0 +1,130 @@
/**
* OFFLINE native Pi Google initialization gate (no generation or provider calls).
* Unlike /api/models, this traverses Pi's lazy adapter and constructs GoogleGenAI.
* The real onPayload hook throws a unique sentinel BEFORE generateContentStream.
* All file modules must resolve inside the payload, including nested dependencies.
*
* node scripts/smoke-provider-runtime.mjs [--standalone <dir>] [--runtime <Node/Electron>]
*/
import assert from "node:assert/strict";
import { spawnSync } from "node:child_process";
import * as fs from "node:fs";
import { createRequire, registerHooks, syncBuiltinESMExports } from "node:module";
import * as os from "node:os";
import * as path from "node:path";
import { randomUUID } from "node:crypto";
import { fileURLToPath, pathToFileURL } from "node:url";
import { formatProbeError, probeEnv } from "./smoke-demo-runtime.mjs";
export function verifyProviderRuntime(standalone, runtime = process.execPath) {
const home = fs.mkdtempSync(path.join(os.tmpdir(), "lh-provider-runtime-"));
try {
fs.writeFileSync(path.join(home, "auth.json"), "{}\n", { mode: 0o600 });
const result = spawnSync(runtime, [fileURLToPath(import.meta.url), "--probe", path.resolve(standalone)], {
cwd: home,
env: { ...probeEnv(process.env, home), ELECTRON_RUN_AS_NODE: "1" },
encoding: "utf8", timeout: 30_000, maxBuffer: 128 * 1024,
});
if (result.status !== 0 || result.error) {
throw new Error(`provider runtime probe failed (${result.error?.message ?? result.status}):\n${result.stdout ?? ""}${result.stderr ?? ""}`);
}
assert.match(result.stdout, /native Pi Google initialization passed/, "probe exited without success marker");
process.stdout.write(result.stdout);
} finally { fs.rmSync(home, { recursive: true, force: true }); }
}
// Dedicated child only: nothing delegates to a live transport, even if the SDK
// changes its fetch path. Count attempts as well: a swallowed guard error fails.
function closeNetwork() {
let attempts = 0;
const deny = () => { attempts++; throw new Error("OFFLINE provider gate: network forbidden"); };
globalThis.fetch = deny;
globalThis.WebSocket = class { constructor() { deny(); } };
const require = createRequire(import.meta.url);
for (const [name, methods] of [
["http", ["request", "get"]], ["https", ["request", "get"]],
["net", ["connect", "createConnection"]], ["tls", ["connect"]],
["dgram", ["createSocket"]],
["dns", ["lookup", "resolve", "resolve4", "resolve6", "resolveAny", "reverse"]],
["dns/promises", ["lookup", "resolve", "resolve4", "resolve6", "resolveAny", "reverse"]],
]) {
const module = require(`node:${name}`);
for (const method of methods) module[method] = deny;
}
require("node:net").Socket.prototype.connect = deny;
syncBuiltinESMExports();
return () => assert.equal(attempts, 0, "SDK attempted network access");
}
async function probe(standalone) {
const noNetwork = closeNetwork();
const root = fs.realpathSync(standalone);
const inside = (file) => {
const relative = path.relative(root, fs.realpathSync(file));
assert.ok(relative !== ".." && !relative.startsWith(`..${path.sep}`) && !path.isAbsolute(relative),
`dependency escaped standalone payload: ${file}`);
};
const loaded = new Set();
const hooks = registerHooks({
resolve(specifier, context, nextResolve) {
const result = nextResolve(specifier, context);
if (result.url.startsWith("file:")) {
inside(fileURLToPath(result.url));
loaded.add(fileURLToPath(result.url));
} else assert.ok(result.url.startsWith("node:"), `unexpected dependency URL: ${result.url}`);
return result;
},
});
try {
// Pi is an explicit server external. Import its shipped public entry and
// factories, never the checkout's module or a direct/fake Google adapter.
const entry = pathToFileURL(path.join(root, "node_modules/@earendil-works/pi-ai/dist/index.js"));
const { createModels } = await import(entry.href);
const models = createModels();
for (const [name, factory] of [["google", "googleProvider"], ["opencode", "opencodeProvider"], ["opencode-go", "opencodeGoProvider"]]) {
const provider = (await import(new URL(`providers/${name}.js`, entry).href))[factory]();
models.setProvider(provider);
const model = provider.getModels().find((candidate) => candidate.api === "google-generative-ai");
assert.ok(model, `${name}: missing Google-backed catalog model`);
const sentinel = `LH_OFFLINE_ON_PAYLOAD_${randomUUID()}`;
let reached = 0;
const result = await models.streamSimple(model, { messages: [{ role: "user", content: "Offline initialization only", timestamp: 0 }] }, {
apiKey: "dummy-offline-provider-gate-not-a-key",
onPayload(payload, selected) {
assert.equal(selected.api, "google-generative-ai");
assert.ok(payload && typeof payload === "object");
reached++;
throw new Error(sentinel);
},
}).result();
assert.equal(reached, 1, `${name}: did not reach onPayload: ${result.errorMessage}`);
assert.equal(result.stopReason, "error");
assert.equal(result.errorMessage, sentinel, `${name}: unexpected initialization error`);
noNetwork();
console.log(`[smoke-provider-runtime] ${name}: native lazy Google adapter reached onPayload; generation stopped`);
}
// Resolution/evaluation, not folder-exists checks. This also covers the
// previously missing transitive edge and permits legitimate nested deps.
for (const name of ["@google/genai", "google-auth-library", "gcp-metadata", "json-bigint", "bignumber.js"]) {
assert.ok([...loaded].some((file) => file.includes(`${path.sep}node_modules${path.sep}${name.replaceAll("/", path.sep)}${path.sep}`)),
`Google runtime chain was not loaded: ${name}`);
}
console.log(`[smoke-provider-runtime] native Pi Google initialization passed; ${loaded.size} modules confined to ${root}; zero network attempts`);
} finally { hooks.deregister(); }
}
async function main() {
const args = process.argv.slice(2);
if (args[0] === "--probe") return probe(args[1]);
const value = (name, fallback) => {
const index = args.indexOf(name);
if (index < 0) return fallback;
assert.ok(args[index + 1] && !args[index + 1].startsWith("--"), `${name} requires a path`);
return path.resolve(args[index + 1]);
};
verifyProviderRuntime(value("--standalone", path.resolve(import.meta.dirname, "../.next/standalone")), value("--runtime", process.execPath));
}
if (process.argv[1] && path.resolve(process.argv[1]) === fileURLToPath(import.meta.url)) {
main().catch((error) => { console.error(`[smoke-provider-runtime] FAIL: ${formatProbeError(error)}`); process.exitCode = 1; });
}
+3
View File
@@ -26,6 +26,7 @@
*/
import { spawn } from "node:child_process";
import { probeEnv, verifyDemoRuntime } from "./smoke-demo-runtime.mjs";
import { verifyProviderRuntime } from "./smoke-provider-runtime.mjs";
import { createRequire } from "node:module";
import * as fs from "node:fs";
import * as http from "node:http";
@@ -80,6 +81,8 @@ async function main() {
// HTTP boot alone never imports the dynamically loaded demo engine.
// Fail before packaging if its standalone dependency graph is incomplete.
verifyDemoRuntime(STANDALONE, RUNTIME);
// /api/models only reads catalogs; initialize the real lazy SDK offline too.
verifyProviderRuntime(STANDALONE, RUNTIME);
const port = await freePort();
const url = `http://127.0.0.1:${port}`;
const home = fs.mkdtempSync(path.join(os.tmpdir(), "lh-smoke-"));
+3 -2
View File
@@ -49,6 +49,8 @@ Tool outputs may be error strings (e.g. search unconfigured, a page that failed
Adapt: try another URL or another search angle; if research is genuinely impossible, say so
in the brief instead of inventing facts.
If web_search reports that search is unavailable because credentials are missing, do not retry it with different queries. Use web_fetch when you have source URLs, and continue with any other available sources. If the missing search capability materially limits your work, state the limitation and the missing-key setup needed in your brief. Never invent search results or sources.
Your FINAL assistant message is your entire deliverable — it must stand alone, using this format:
## Summary
@@ -70,8 +72,7 @@ export interface ResearcherRunOptions {
/** The search seam (search port wave, 2026-09-20): resolves the LIVE
* provider per call, null = honestly unavailable (the web_search tool
* reports that in its output — no crash, no silent fallback). Absent
* seam reads as unavailable: the server wiring always injects one; the
* dispatch-level drop (jobTools) keeps doomed jobs from even starting. */
* seam reads as unavailable; web_fetch remains usable without it. */
search?: () => SearchProvider | null;
/** The researcher's PDF dir (PDF wave, 2026-10-07): when present, the
* toolset gains web_fetch_pdf + pdf_info/pdf_extract/pdf_search scoped
+2 -3
View File
@@ -82,9 +82,8 @@ export interface JobRunnerDeps {
/** The search seam (search port wave, 2026-09-20): resolves the LIVE
* provider fresh per job — server/search.ts composition injected by
* the wiring; null = honestly unavailable (the researcher's web_search
* reports it). Absent seam (tests) reads the same. The dispatch-level
* drop (jobTools' searchAvailable) uses the same signal — jobs already
* in flight when a key disappears still fail honestly, never crash. */
* reports it as recoverable text). Absent seam (tests) reads the same;
* research dispatch and web_fetch do not require search credentials. */
searchProvider?: () => SearchProvider | null;
/** Key-presence gate (provider-key interlude, 2026-09-25): given the
* resolved engine's provider + the kind's slot tier, return the
+5 -13
View File
@@ -268,8 +268,7 @@ export interface ProberToolContext extends ProbeDoneContext {
jobStore: JobStore;
/** Fire the background runner for a freshly created job. */
spawn: (job: JobRecord, input: ResearchInput | DemoInput | VisualInput | PdfInput) => void;
/** Search availability — the research dispatch tool's honest-disable
* signal (same gate as the teacher's). Absent = available. */
/** Retained search-status seam; research dispatch is available keylessly. */
searchAvailable?: () => boolean;
/** The prober's graph-map reader (wave 7): present = the two map tools
* (list_sessions / list_propositions) are registered. The composition
@@ -331,13 +330,12 @@ export function createProberTools(
// research + pdf are ALWAYS available in every phase — the
// prober can verify facts or read learner PDFs while assessing.
// demo/visual stay teach-phase-only (they need the lesson's plan
// context). Same gating as the teacher: toggle-filtered here,
// research subject to the searchAvailable gate (honestly absent
// when no engine is live).
// context). Same user-toggle filtering as the teacher; research
// remains available without a search-provider key.
...(jobs.research ? { research: jobs.research } : {}),
pdf: jobs.pdf!,
// The map reader (wave 7): present only when the settings gate
// opened — same registration-time drop as the search gate.
// opened.
...(ctx.graphStore
? proberGraphTools({ chatStore: ctx.chatStore, graphStore: ctx.graphStore })
: {}),
@@ -400,9 +398,7 @@ export interface TeacherToolContext extends PlanToolContext {
* rides along — the spawn fires mid-stream, before the tool part is
* persisted to the transcript (a transcript scan would race it). */
spawn: (job: JobRecord, input: ResearchInput | DemoInput | VisualInput | PdfInput) => void;
/** Search availability (search port wave, 2026-09-20): the research
* dispatch tool's honest-disable signal — no live engine = the tool
* drops from this registry (JobToolContext.searchAvailable). */
/** Retained search-status seam; research dispatch is available keylessly. */
searchAvailable?: () => boolean;
}
@@ -462,10 +458,6 @@ export function createTeacherTools(
find: findTool,
grep: grepTool,
...planTools(ctx),
// Fix wave (2026-09-20): searchAvailable was RECEIVED here but
// DROPPED when building the inner JobToolContext — the drop-gate
// always took its ?? true default. Threaded now: the live signal
// reaches jobTools' research gate.
...jobTools({ chatId: ctx.chatId, jobStore: ctx.jobStore, spawn: ctx.spawn, searchAvailable: ctx.searchAvailable }),
},
withoutMechanics(enabled),
+14 -21
View File
@@ -23,11 +23,9 @@ export interface JobToolContext {
* input rides along because the spawn happens MID-STREAM, before
* onFinish persists the tool part — a transcript scan would race it. */
spawn: (job: JobRecord, input: ResearchInput | DemoInput | VisualInput | PdfInput) => void;
/** Search availability (search port wave, 2026-09-20): the honest
* disable signal for the research dispatch tool — no live engine
* means NO research tool in the registry (owner decision: registry-
* level drop, no doomed dispatch, no silent fallback). Absent =
* available (tests + any seam without search wiring). */
/** Search-provider availability, retained for existing composition seams.
* Research dispatch does not depend on it: web_fetch is keyless and
* web_search reports a missing provider as a recoverable tool result. */
searchAvailable?: () => boolean;
}
@@ -127,25 +125,20 @@ const PDF_DESCRIPTION =
* context (chatId + jobStore + the spawn seam), so concurrent chats never
* share state. chat-handler wires it each turn; tests wire their own.
*
* The research tool drops from the registry when ctx.searchAvailable reports
* no live engine (2026-09-20) — the teacher never sees it, so a keyless
* server can't dispatch doomed research jobs. Jobs ALREADY in flight when
* a key disappears still run to their honest failure (runner backstop).
* Research remains available without a search-provider key; the role's
* existing user-toggle filtering still applies outside this registry.
*/
export function jobTools(ctx: JobToolContext): Record<string, AgentTool<any>> {
const searchOn = ctx.searchAvailable?.() ?? true;
return {
...(searchOn && {
research: {
name: "research",
label: "research",
description: RESEARCH_DESCRIPTION,
parameters: researchInputSchema,
execute: serverExecute("research", researchInputSchema, (input, { toolCallId }) =>
dispatch(ctx, "research", toolCallId, RESEARCH_ACK, input),
),
},
}),
research: {
name: "research",
label: "research",
description: RESEARCH_DESCRIPTION,
parameters: researchInputSchema,
execute: serverExecute("research", researchInputSchema, (input, { toolCallId }) =>
dispatch(ctx, "research", toolCallId, RESEARCH_ACK, input),
),
},
demo: {
name: "demo",
label: "demo",
+4 -4
View File
@@ -40,8 +40,8 @@ export interface PracticeToolContext {
graphNodes?: () => ReadonlyArray<Pick<import("../../core/contracts/graph.ts").GraphNode, "id" | "name">>;
/** THE RARELY-NEEDED SET (owner round 10, 2026-10-05): the map readers
* behind the SAME settings gate as the prober's (graphToolsEnabled at
* the composition root); the research + pdf jobs with the honest
* searchAvailable drop. Each absent (tests) = absent from the registry. */
* the composition root); the research + pdf jobs (both keyless).
* Each absent context (tests) = absent from the registry. */
graph?: GraphToolContext;
jobs?: JobToolContext;
}
@@ -225,8 +225,8 @@ export function practiceTools(ctx: PracticeToolContext): Record<string, AgentToo
const gate = createPracticeGate(ctx);
// THE RARELY-NEEDED SET (owner round 10): read/ls/find/grep over the
// learner's files (always — the same read-only set the teacher and
// prober carry), the research + pdf jobs (research honestly absent
// when no search engine is live; demo/visual stay teach-only), and
// prober carry), the research + pdf jobs (available without search
// keys; demo/visual stay teach-only), and
// the prober's map readers when the settings gate opened. None are
// toggle-filtered: problem/conclude are the session; these are
// exceptions the doctrine says to reach for rarely.
+37
View File
@@ -610,6 +610,43 @@ describe("research runner (async protocol §3)", () => {
assert.deepEqual(search.queries, ["sigma.js latest version (5)"]);
});
it("a keyless fetch-only researcher still completes and delivers a sourced job result", async () => {
const url = "https://example.com/supplied-source";
const brief = `## Summary\nVerified supplied source. [Source](${url})\n\n## Sources\n- Source (${url})`;
const input = { question: `Verify ${url}` };
const chat = chatStore.createChat("Test: keyless fetch-only research");
trackChat(chat.id);
chatStore.saveMessages(chat.id, nativeToolMessages("verify supplied URL", "research", "call-keyless", input));
const job = jobStore.createJob({ kind: "research", chatId: chat.id, toolCallId: "call-keyless" });
trackJob(job.id);
const delivery = fakeDeliver();
const runner = createJobRunner({ chatStore, jobStore, deliver: delivery.deliver, searchProvider: () => null });
const fetched: string[] = [];
const realFetch = globalThis.fetch;
globalThis.fetch = (async (source) => {
fetched.push(String(source));
assert.equal(String(source), url);
return new Response("Synthetic verified source content", { headers: { "content-type": "text/plain" } });
}) as typeof fetch;
researchModelOverride.current = fauxEngine([
fauxAssistantMessage([fauxToolCall("web_fetch", { url }, { id: "keyless-fetch" })], { stopReason: "toolUse" }),
fauxAssistantMessage([fauxText(brief)]),
]);
try {
await runner.runJob(job.id);
} finally {
researchModelOverride.current = undefined;
globalThis.fetch = realFetch;
}
assert.deepEqual(fetched, [url]);
const done = jobStore.getJob(job.id)!;
assert.equal(done.status, "done");
assert.equal((done.result as ResearchResult).brief, brief);
assert.ok((done.result as ResearchResult).sources.some((s) => s.url === url));
assert.equal(delivery.steers.length, 1);
assert.equal(isSubagentFailure("research", done.result as ResearchResult), false);
});
it("a failing runner still steers (the teacher learns the outcome) and the job lands error", async () => {
const chat = chatStore.createChat("Test: runner failure");
trackChat(chat.id);
+5 -4
View File
@@ -208,14 +208,15 @@ describe("the practice agent", () => {
const jobs = { chatId: "c1", jobStore: null as never, spawn: (() => {}) as never, searchAvailable: () => false };
const graph = { chatStore: null as never, graphStore: null as never };
const tools = practiceTools({ ...ctxOf([]), jobs, graph });
// research honestly absent (no search engine live); pdf always; the
// prober's two map readers behind the settings gate.
// Both jobs remain available keylessly; the prober's two map
// readers stay behind their existing settings gate. Practice does
// not apply the teacher/prober user tool-toggle filter.
assert.deepEqual(
Object.keys(tools).sort(),
["conclude", "find", "grep", "list_propositions", "list_sessions", "ls", "pdf", "problem", "read"],
["conclude", "find", "grep", "list_propositions", "list_sessions", "ls", "pdf", "problem", "read", "research"],
);
const withSearch = practiceTools({ ...ctxOf([]), jobs: { ...jobs, searchAvailable: () => true }, graph });
assert.ok(withSearch.research, "a live engine adds research");
assert.ok(withSearch.research, "a live engine also keeps research");
});
it("constructs with exactly the two practice tools", () => {
+42
View File
@@ -0,0 +1,42 @@
/** Real installed Pi/Google SDK only; no provider stubs, keys, or generation. */
import assert from "node:assert/strict";
import { spawnSync } from "node:child_process";
import * as fs from "node:fs";
import * as os from "node:os";
import * as path from "node:path";
import { describe, it } from "node:test";
import { fileURLToPath } from "node:url";
const script = fileURLToPath(new URL("../scripts/smoke-provider-runtime.mjs", import.meta.url));
const checkout = fileURLToPath(new URL("..", import.meta.url));
const run = (payload: string) => spawnSync(process.execPath, [script, "--standalone", payload], { encoding: "utf8", timeout: 40_000 });
describe("offline native Pi Google initialization gate", () => {
it("reaches the sentinel for direct Google and Google-backed OpenCode adapters using the real SDK", () => {
// This is an unbundled positive control, NOT proof of standalone completeness.
const result = run(checkout);
assert.equal(result.status, 0, String(result.stderr));
for (const provider of ["google", "opencode", "opencode-go"]) assert.ok(String(result.stdout).includes(`${provider}: native lazy Google adapter reached onPayload`));
assert.match(String(result.stdout), /zero network attempts/);
});
it("rejects a missing Pi entry instead of treating any error as success", () => {
const root = fs.mkdtempSync(path.join(os.tmpdir(), "lh-provider-missing-"));
try {
const result = run(root);
assert.equal(result.status, 1);
assert.doesNotMatch(String(result.stdout), /initialization passed/);
assert.match(String(result.stderr), /MODULE_NOT_FOUND|ENOENT/);
assert.deepEqual(fs.readdirSync(root), [], "gate modified the payload");
} finally { fs.rmSync(root, { recursive: true, force: true }); }
});
it("rejects an otherwise complete SDK outside the payload (no host fallback)", () => {
const root = fs.mkdtempSync(path.join(os.tmpdir(), "lh-provider-escape-"));
try {
fs.symlinkSync(path.join(checkout, "node_modules"), path.join(root, "node_modules"), process.platform === "win32" ? "junction" : "dir");
const result = run(root);
assert.equal(result.status, 1);
assert.match(String(result.stderr), /dependency escaped standalone payload/);
assert.doesNotMatch(String(result.stdout), /initialization passed/);
} finally { fs.rmSync(root, { recursive: true, force: true }); }
});
});
+62
View File
@@ -0,0 +1,62 @@
/** JS matcher + config-policy regressions; native Turbopack needs the real payload gate. */
import assert from "node:assert/strict";
import { createRequire } from "node:module";
import { describe, it } from "node:test";
import nextConfig from "../next.config.ts";
const require = createRequire(import.meta.url);
// Use Next's actual bundled matcher, not a different glob implementation.
const picomatch = require("next/dist/compiled/picomatch");
const { makeIgnoreFn } = require("next/dist/build/collect-build-traces");
const patterns = nextConfig.outputFileTracingExcludes!["*"];
describe("standalone JavaScript tracing exclusions and native workaround policy", () => {
for (const windows of [false, true]) {
const root = windows ? "C:\\build\\learnharness" : "/build/learnharness";
const file = (relative: string) => root + (windows ? "\\" : "/") + (windows ? relative.replaceAll("/", "\\") : relative);
const ignored = picomatch(patterns, { contains: true, dot: true, windows });
it(`excludes genuine data directories and secrets (${windows ? "Windows" : "POSIX"})`, () => {
for (const relative of ["data/canary.json", "nested/data/private.json", "auth.json", ".env", ".env.local"]) {
assert.equal(ignored(file(relative)), true, relative);
}
// The explicit includes remain necessary: this raw ignore still matches SDK data.
assert.equal(ignored(file("node_modules/@earendil-works/pi-ai/dist/providers/data/models.json")), true);
assert.ok(nextConfig.outputFileTracingIncludes!["*"]!.includes("node_modules/@earendil-works/pi-ai/dist/providers/data/**/*"));
});
it(`does not confuse dependency suffixes with data directories (${windows ? "Windows" : "POSIX"})`, () => {
for (const relative of ["node_modules/gcp-metadata/build/src/index.js", "node_modules/form-data/lib/form_data.js", "node_modules/@swc/helpers/_/_ts_metadata.js", "node_modules/@swc/helpers/_/_ts_metadata/index.js", "node_modules/somedata/runtime.js"]) {
assert.equal(ignored(file(relative)), false, relative);
assert.equal(picomatch("data/**/*", { contains: true, dot: true, windows })(file("node_modules/gcp-metadata/build/src/index.js")), true, "old pattern must reproduce the bug");
}
});
}
it("uses the installed Next JavaScript early-ignore function on the host platform", () => {
const root = process.cwd();
const ignore = makeIgnoreFn(root, patterns);
assert.equal(ignore(`${root}/data/private.json`), true);
assert.equal(ignore(`${root}/node_modules/gcp-metadata/build/src/index.js`), false);
});
it("preserves all secret exclusions exactly", () => {
assert.deepEqual(patterns, ["**/data/**/*", "auth.json", ".env*"]);
});
it("retains exactly the narrow native-Turbopack closure workaround and existing runtime includes", () => {
assert.deepEqual(nextConfig.outputFileTracingIncludes!["*"], [
"node_modules/@puppeteer/browsers/lib/browser-data/**/*.js",
"node_modules/pdfjs-dist/legacy/**/*",
"node_modules/pdfjs-dist/standard_fonts/**/*",
"node_modules/@napi-rs/canvas-*/**/*",
"node_modules/@resvg/resvg-js-*/**/*",
"node_modules/@earendil-works/pi-ai/dist/providers/data/**/*",
"node_modules/@earendil-works/pi-agent-core/dist/**/*.json",
"node_modules/gcp-metadata/**/*",
"node_modules/json-bigint/**/*",
"node_modules/bignumber.js/**/*",
]);
});
it("limits the workaround to its three packages, never learner data or unrelated dependencies", () => {
const closure = nextConfig.outputFileTracingIncludes!["*"]!.slice(-3);
const included = picomatch(closure, { dot: true });
for (const file of ["node_modules/gcp-metadata/build/src/index.js", "node_modules/json-bigint/lib/parse.js", "node_modules/bignumber.js/bignumber.js"]) assert.equal(included(file), true, file);
for (const file of ["data/private.json", "nested/data/private.json", "auth.json", ".env.local", "node_modules/unrelated/runtime.js", "node_modules/gcp-metadata-lookalike/index.js"]) assert.equal(included(file), false, file);
});
});
+70
View File
@@ -0,0 +1,70 @@
import React, { act } from "react";
(globalThis as Record<string, unknown>).React = React;
import assert from "node:assert/strict";
import { it } from "node:test";
import { parseHTML } from "linkedom";
import { AgentSection } from "../components/settings/agent-section.tsx";
import type { AgentSettings } from "../src/core/index.ts";
it("keyless settings keep Research interactive while search-engine setup retains its own disabled state", async () => {
const dom = parseHTML("<html><body><div id='root'></div></body></html>");
const globals = globalThis as Record<string, unknown>;
const previous = new Map<string, PropertyDescriptor | undefined>();
const reason = "No search engine is configured.";
const bodies: Record<string, unknown> = {
"/api/search": { available: false, engine: null, reason },
"/api/keys": { keys: [{ id: "TAVILY_API_KEY", present: false }, { id: "BRAVE_API_KEY", present: false }] },
"/api/models": { models: [], slots: { strong: null, cheap: null, vision: null } },
"/api/settings/prompts": { prober: [], teacher: [] },
};
for (const [key, value] of Object.entries({
window: dom.window,
document: dom.document,
navigator: globals.navigator ?? { userAgent: "", platform: "" },
HTMLElement: dom.window.HTMLElement,
IS_REACT_ACT_ENVIRONMENT: true,
fetch: async (url: string) => {
assert.ok(url in bodies, `unexpected request: ${url}`);
return new Response(JSON.stringify(bodies[url]));
},
})) {
previous.set(key, Object.getOwnPropertyDescriptor(globalThis, key));
Object.defineProperty(globalThis, key, { configurable: true, writable: true, value });
}
const { createRoot } = await import("react-dom/client");
const root = createRoot(document.getElementById("root")!);
let written: AgentSettings | undefined;
try {
await act(async () => {
root.render(<AgentSection
agent={{ tools: { research: true } }}
models={undefined}
graph={undefined}
compaction={undefined}
onSectionChange={(next) => { written = next; }}
onModelsChange={() => {}}
onCompactionChange={() => {}}
/>);
});
const row = document.querySelector<HTMLElement>('[data-testid="tool-row-research"]')!;
const toggle = row.querySelector<HTMLButtonElement>('[data-testid="toggle-tool-research"]')!;
assert.ok(toggle);
assert.equal(toggle.hasAttribute("disabled"), false);
assert.equal(toggle.getAttribute("aria-checked"), "true");
assert.ok(!row.className.includes("opacity-40"));
assert.ok(!row.textContent?.includes(reason), "search setup hint is no longer a research-disable hint");
act(() => { toggle.click(); });
assert.equal(written?.tools?.research, false, "the keyless toggle emits the same stored boolean");
const engine = document.querySelector<HTMLButtonElement>('[data-testid="search-engine-trigger"]')!;
assert.ok(engine.className.includes("cursor-not-allowed"), "search engine preference retains its own key gate");
act(() => { engine.click(); });
assert.equal(document.querySelector('[data-testid="search-engine-menu"]'), null, "keyless engine preference does not open");
assert.ok(document.querySelector('[data-testid="search-engine-row"]')?.textContent?.includes(reason));
} finally {
act(() => { root.unmount(); });
for (const [key, descriptor] of previous) {
if (descriptor) Object.defineProperty(globalThis, key, descriptor);
else delete globals[key];
}
}
});
-16
View File
@@ -15,7 +15,6 @@ import {
composeSlotModel,
composeTool,
composeQuizMechanics,
researchToolUnavailable,
sectionedSegments,
} from "../components/settings/agent-section.tsx";
import { nextOverrides } from "../components/settings/keyboard-section.tsx";
@@ -328,21 +327,6 @@ describe("first-run gate — slotsConfigured + setupStage (model-slots wave)", (
});
});
describe("researchToolUnavailable — the row grays on the registry's own signal", () => {
it("research grays exactly when the live search status says unavailable", () => {
// the SAME /api/search signal chat-handler's searchAvailable uses —
// the row and the registry can never disagree (Teacher flag 2026-09-22)
assert.equal(researchToolUnavailable("research", { available: false }), true);
assert.equal(researchToolUnavailable("research", { available: true }), false);
// loading/failed fetch grays nothing (never a false-negative row)
assert.equal(researchToolUnavailable("research", null), false);
// only research — the other tools never touch the search port
for (const id of ["demo", "visual", "pdf"]) {
assert.equal(researchToolUnavailable(id, { available: false }), false);
}
});
});
describe("sectionedSegments — the prompts overhaul (2026-09-26)", () => {
// The server contract: BOTH per-agent lists include the shared
// segments (same paths) — the view must list them ONCE.
+124 -31
View File
@@ -1,20 +1,14 @@
/**
* Search rewiring (search port wave, 2026-09-20) — the web_search tool's
* port injection + the research dispatch tool's registry-level drop.
*
* Google CSE is dead: the tool runs on the injected search port
* (core/search.ts), the runner threads the seam (searchProvider — covered
* end-to-end in jobs.test.ts), and jobTools drops the research tool when
* ctx.searchAvailable reports no live engine (owner decision: registry
* drop, no doomed dispatch, no silent fallback).
*
/** Keyless research availability + the injected search port.
* No network, no shared data files — the jobStore fake below is in-memory.
* A missing search key is a recoverable tool result, not a dispatch gate.
*/
import assert from "node:assert/strict";
import { describe, it } from "node:test";
import { createWebTools } from "../src/agent/tools/web.ts";
import { jobTools } from "../src/agent/tools/jobs.ts";
import { runResearcherAgent } from "../src/agent/subagents/researcher.ts";
import { createFauxCore, fauxAssistantMessage, fauxText, fauxToolCall, type FauxResponseStep } from "@earendil-works/pi-ai";
import type { JobStore } from "../src/core/ports.ts";
import type { JobRecord } from "../src/core/contracts.ts";
import type { SearchHit, SearchProvider } from "../src/core/search.ts";
@@ -31,9 +25,13 @@ function stubProvider(hits: SearchHit[] = [{ title: "t", url: "https://e.com/",
describe("web_search on the injected port (web.ts)", () => {
it("null provider → the honest unavailability string, never a crash", async () => {
const { web_search } = createWebTools(() => null);
assert.ok(web_search, "web_search is still exposed without a provider");
const out = await web_search!.execute!("c1", { query: "anything" });
assert.match(String(out.details), /not available: no search engine is configured/);
assert.match(String(out.details), /Tavily or Brave/);
const missing = "web_search is not available: no search engine is configured. " +
"Set a Tavily or Brave search API key in settings. Report in your brief that web search is unavailable.";
assert.equal(out.details, missing);
assert.deepEqual(out.content, [{ type: "text", text: missing }]);
assert.notEqual(out.isError, true, "missing keys remain ordinary recoverable text, not a native error");
});
it("a live provider returns its hits (count defaults to 5)", async () => {
@@ -110,11 +108,86 @@ describe("web_search on the injected port (web.ts)", () => {
});
});
/** Minimal in-memory JobStore — the drop tests never touch the fs. */
describe("keyless researcher loop", () => {
it("sends the exact approved missing-credentials guidance in the researcher system prompt", async () => {
const approved = "If web_search reports that search is unavailable because credentials are missing, do not retry it with different queries. Use web_fetch when you have source URLs, and continue with any other available sources. If the missing search capability materially limits your work, state the limitation and the missing-key setup needed in your brief. Never invent search results or sources.";
const faux = createFauxCore({});
faux.setResponses([(context) => {
const system = context.messages.find((m) => m.role === "system");
assert.ok(system && system.role === "system");
assert.equal(typeof system.content, "string");
const prompt = system.content as string;
assert.ok(prompt.split("\n\n").includes(approved), "approved paragraph is passed verbatim");
assert.match(prompt, /Break the question into 2-4 searchable facets/);
assert.match(prompt, /## Gaps/);
assert.ok(system.toolsAdded?.some((t) => t.name === "web_search"));
assert.ok(system.toolsAdded?.some((t) => t.name === "web_fetch"));
return fauxAssistantMessage([fauxText("Prompt verified; no model behavior inferred.")]);
}]);
assert.equal(await runResearcherAgent("Check the prompt only", {
engine: { model: faux.getModel(), streamFn: faux.streamSimple },
}), "Prompt verified; no model behavior inferred.");
});
// These scripts prove tool/loop mechanics, not real-model compliance.
for (const searchFirst of [true, false]) {
it(searchFirst ? "recovers from missing search keys by fetching a supplied URL" : "fetch-only research needs no search credentials", async () => {
const url = "https://example.com/supplied-source";
const brief = `## Summary\nVerified supplied source. [Source](${url})`;
const steps: FauxResponseStep[] = [];
const fetchCall = () => fauxAssistantMessage(
[fauxToolCall("web_fetch", { url }, { id: "fetch1" })], { stopReason: "toolUse" },
);
if (searchFirst) {
steps.push(
() => {
return fauxAssistantMessage(
[fauxToolCall("web_search", { query: "supplied source" }, { id: "search1" })], { stopReason: "toolUse" },
);
},
(context) => {
const result = context.messages.find((m) => m.role === "toolResult" && m.toolCallId === "search1");
assert.ok(result && result.role === "toolResult");
assert.equal(result.isError, false, "missing keys do not abort the native loop");
assert.match(JSON.stringify(result.content), /no search engine is configured/);
return fetchCall();
},
);
} else steps.push(fetchCall());
steps.push((context) => {
const result = context.messages.find((m) => m.role === "toolResult" && m.toolCallId === "fetch1");
assert.ok(result && result.role === "toolResult");
assert.equal(result.isError, false);
assert.match(JSON.stringify(result.content), /Synthetic verified source content/);
return fauxAssistantMessage([fauxText(brief)]);
});
const faux = createFauxCore({});
faux.setResponses(steps);
const realFetch = globalThis.fetch;
const fetched: string[] = [];
globalThis.fetch = (async (input) => {
fetched.push(String(input));
assert.equal(String(input), url, "only the supplied URL is fetched");
return new Response("Synthetic verified source content", { headers: { "content-type": "text/plain" } });
}) as typeof fetch;
try {
assert.equal(await runResearcherAgent(`Verify ${url}`, {
engine: { model: faux.getModel(), streamFn: faux.streamSimple },
// No search seam: this is the keyless backstop, not a search stub.
}), brief);
assert.deepEqual(fetched, [url]);
} finally {
globalThis.fetch = realFetch;
}
});
}
});
/** Minimal in-memory JobStore — registry tests never touch the fs. */
function fakeJobStore(): JobStore {
const jobs = new Map<string, JobRecord>();
return {
createJob(init) {
createJob(init) {
const job: JobRecord = {
id: `job_${jobs.size + 1}`,
kind: init.kind,
@@ -149,13 +222,28 @@ function fakeCtx(available?: () => boolean) {
};
}
describe("the research registry drop (jobTools)", () => {
it("searchUnavailable: false → research drops from the registry (the other job tools stay)", () => {
describe("keyless research dispatch (jobTools)", () => {
it("searchAvailable: false → research remains in the registry", () => {
const tools = jobTools(fakeCtx(() => false));
assert.equal("research" in tools, false, "no research tool — no doomed dispatch");
assert.ok(tools.research, "research can fetch supplied URLs without a search provider");
assert.ok(tools.demo && tools.visual && tools.pdf, "the search-less server keeps its other sub-agents");
});
it("keyless dispatch preserves the async job acknowledgement and spawn contract", async () => {
const ctx = fakeCtx(() => false);
const spawned: Array<{ job: JobRecord; input: unknown }> = [];
const tools = jobTools({ ...ctx, spawn: (job, input) => { spawned.push({ job, input }); } });
const input = { question: "Verify the supplied URL", context: "https://example.com/source" };
const out = await tools.research!.execute!("keyless-call", input);
assert.match(String(out.details), /^Dispatched \(job job_1\)\./);
assert.match(String(out.details), /fire-and-forget/);
assert.equal(spawned.length, 1);
assert.equal(spawned[0]!.job.kind, "research");
assert.equal(spawned[0]!.job.toolCallId, "keyless-call");
assert.equal(spawned[0]!.job.status, "running");
assert.deepEqual(spawned[0]!.input, input);
});
it("searchAvailable: true → research present", () => {
const tools = jobTools(fakeCtx(() => true));
assert.ok(tools.research, "a keyed server exposes research");
@@ -167,12 +255,8 @@ describe("the research registry drop (jobTools)", () => {
});
});
// ─── the threaded gate (fix wave, 2026-09-20) ────────────────────────────────
//
// The fix: createTeacherTools/createProberTools used to DROP searchAvailable
// when building the inner JobToolContext, so jobTools' gate always took its
// ?? true default even when the live wiring said unavailable. These pins
// hold the THREADED path (the registry factories, not jobTools directly).
// Exercise live registry factories too: search status must not override
// the existing per-role/user-toggle policy.
import { createProberTools, createTeacherTools } from "../src/agent/tools.ts";
import type { ChatStore } from "../src/core/ports.ts";
@@ -186,19 +270,19 @@ const fakeChatStore = {
const ctxBase = { chatId: "chat-gate-test", chatStore: fakeChatStore };
describe("the threaded searchAvailable gate (registry factories)", () => {
it("createTeacherTools: unavailable drops research on the live wiring path", () => {
describe("keyless role registries", () => {
it("createTeacherTools: unavailable keeps research on the live wiring path", () => {
const tools = createTeacherTools({
...ctxBase,
jobStore: fakeJobStore(),
spawn: () => {},
searchAvailable: () => false,
});
assert.equal("research" in tools, false, "the threaded signal reached jobTools' gate");
assert.ok(tools.research, "no search key is required to dispatch research");
assert.ok(tools.demo && tools.visual && tools.pdf, "the other job tools survive");
});
it("createTeacherTools: available keeps research (the gate reads the signal both ways)", () => {
it("createTeacherTools: available keeps research", () => {
const tools = createTeacherTools({
...ctxBase,
jobStore: fakeJobStore(),
@@ -208,7 +292,7 @@ describe("the threaded searchAvailable gate (registry factories)", () => {
assert.ok(tools.research, "a keyed server exposes research through the teacher");
});
it("createProberTools: research + pdf are phase-omnipresent; the gate drops research only", () => {
it("createProberTools: research + pdf remain phase-omnipresent with or without search keys", () => {
const on = createProberTools({
...ctxBase,
deliver: () => {},
@@ -225,16 +309,25 @@ describe("the threaded searchAvailable gate (registry factories)", () => {
spawn: () => {},
searchAvailable: () => false,
});
assert.equal("research" in off, false, "no engine = no research dispatch in the prober either");
assert.ok(off.research, "keyless research dispatch is available in the prober too");
assert.ok(off.pdf && off.probe_done, "pdf is keyless — it survives; probe_done always-on");
});
it("createProberTools: the user toggle drops research independently of the gate", () => {
it("createProberTools: the user toggle still drops research without search keys", () => {
const tools = createProberTools(
{ ...ctxBase, deliver: () => {}, jobStore: fakeJobStore(), spawn: () => {}, searchAvailable: () => true },
{ ...ctxBase, deliver: () => {}, jobStore: fakeJobStore(), spawn: () => {}, searchAvailable: () => false },
{ research: false },
);
assert.equal("research" in tools, false, "toggle-filtered like the teacher side");
assert.ok(tools.pdf, "pdf unaffected by a research toggle");
});
it("createTeacherTools: the user toggle still drops research without search keys", () => {
const tools = createTeacherTools(
{ ...ctxBase, jobStore: fakeJobStore(), spawn: () => {}, searchAvailable: () => false },
{ research: false },
);
assert.equal("research" in tools, false);
assert.ok(tools.pdf && tools.demo && tools.visual);
});
});