@reticlehq/server 3.3.0 → 3.5.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/LICENSE-ENTERPRISE +1 -1
- package/README.md +3 -3
- package/SKILL.md +12 -14
- package/dist/chromium-launch-options.d.ts +4 -3
- package/dist/chromium-launch-options.js +6 -2
- package/dist/command/cli/adhoc-suite.d.ts +22 -0
- package/dist/command/cli/adhoc-suite.js +90 -19
- package/dist/command/cli/adhoc-verdict.d.ts +35 -0
- package/dist/command/cli/adhoc-verdict.js +90 -3
- package/dist/command/cli/answers/gate-hook-message.d.ts +6 -0
- package/dist/command/cli/answers/gate-hook-message.js +6 -0
- package/dist/command/cli/auth/cloud-keystore.d.ts +0 -2
- package/dist/command/cli/auth/cloud-keystore.js +0 -3
- package/dist/command/cli/auth/cloud-session.d.ts +6 -2
- package/dist/command/cli/auth/cloud-session.js +2 -0
- package/dist/command/cli/auth/device-flow.d.ts +2 -2
- package/dist/command/cli/cli-doctor.js +7 -5
- package/dist/command/cli/cli-flow-commands.d.ts +3 -6
- package/dist/command/cli/cli-flow-commands.js +67 -4
- package/dist/command/cli/cli-parse.d.ts +8 -1
- package/dist/command/cli/cli-parse.js +24 -4
- package/dist/command/cli/cli-update-commands.js +8 -1
- package/dist/command/cli/cli-verify.d.ts +24 -3
- package/dist/command/cli/cli-verify.js +54 -11
- package/dist/command/cli/cloud-cli.js +64 -2
- package/dist/command/cli/cloud-kit.js +3 -5
- package/dist/command/cli/cloud-login.d.ts +1 -1
- package/dist/command/cli/cloud-login.js +27 -8
- package/dist/command/cli/demo/demo-page.js +6 -3
- package/dist/command/cli/doctor/browser/chromium-hint.d.ts +5 -0
- package/dist/command/cli/doctor/browser/chromium-hint.js +5 -0
- package/dist/command/cli/launch/cli-launch.d.ts +8 -0
- package/dist/command/cli/launch/cli-launch.js +21 -9
- package/dist/command/cli/launch/sdk-version-match.d.ts +84 -0
- package/dist/command/cli/launch/sdk-version-match.js +206 -0
- package/dist/command/cli/lifecycle/daemon-lifecycle.js +2 -16
- package/dist/command/cli/ports/resolve/cli-port.d.ts +25 -0
- package/dist/command/cli/ports/resolve/cli-port.js +125 -15
- package/dist/command/cli/setup-mcp-cli.js +21 -13
- package/dist/command/cli/tutorial.d.ts +2 -2
- package/dist/command/cli/tutorial.js +21 -7
- package/dist/command/cli.js +79 -11
- package/dist/command/daemon/binding/daemon-status-probe.d.ts +14 -2
- package/dist/command/daemon/binding/daemon-status-probe.js +34 -6
- package/dist/command/daemon/daemon-resolve.d.ts +9 -0
- package/dist/command/daemon/daemon-resolve.js +1 -1
- package/dist/command/daemon/lifetime/daemon-usefulness.d.ts +7 -1
- package/dist/command/daemon/lifetime/daemon-usefulness.js +30 -7
- package/dist/command/dev/stale-issue-guard.d.ts +3 -4
- package/dist/command/dev/stale-issue-guard.js +8 -8
- package/dist/command/setup/bringup/connect-lease.d.ts +17 -0
- package/dist/command/setup/bringup/connect-lease.js +38 -0
- package/dist/command/setup/bringup/dev-server-files.d.ts +68 -0
- package/dist/command/setup/bringup/dev-server-files.js +131 -0
- package/dist/command/setup/bringup/dev-server-log-cap.d.ts +35 -0
- package/dist/command/setup/bringup/dev-server-log-cap.js +121 -0
- package/dist/command/setup/bringup/dev-server-wait.d.ts +2 -0
- package/dist/command/setup/bringup/dev-server-wait.js +1 -1
- package/dist/command/setup/bringup/ensure-daemon.d.ts +65 -4
- package/dist/command/setup/bringup/ensure-daemon.js +135 -16
- package/dist/command/setup/bringup/static-page-server.d.ts +35 -0
- package/dist/command/setup/bringup/static-page-server.js +151 -0
- package/dist/command/setup/desktop-shape.d.ts +4 -0
- package/dist/command/setup/desktop-shape.js +18 -2
- package/dist/command/setup/init/init-port.d.ts +15 -0
- package/dist/command/setup/init/init-port.js +44 -0
- package/dist/command/setup/init/init-runtime.d.ts +1 -1
- package/dist/command/setup/init/init-runtime.js +91 -9
- package/dist/command/setup/node-effects.d.ts +27 -14
- package/dist/command/setup/node-effects.js +86 -45
- package/dist/command/setup/probe/loopback-probe-urls.d.ts +7 -0
- package/dist/command/setup/probe/loopback-probe-urls.js +40 -0
- package/dist/command/setup/probe/page-probe.d.ts +13 -2
- package/dist/command/setup/probe/page-probe.js +18 -3
- package/dist/command/setup/run-setup.d.ts +28 -1
- package/dist/command/setup/run-setup.js +103 -15
- package/dist/command/setup/session-pick.d.ts +13 -0
- package/dist/command/setup/session-pick.js +20 -0
- package/dist/command/setup/setup-command.d.ts +2 -14
- package/dist/command/setup/setup-command.js +105 -31
- package/dist/command/setup/setup-mcp.d.ts +2 -0
- package/dist/command/setup/setup-mcp.js +17 -8
- package/dist/command/setup/terminal/wait-progress.d.ts +8 -0
- package/dist/command/setup/terminal/wait-progress.js +17 -0
- package/dist/command/version/peer-announce.d.ts +6 -2
- package/dist/command/version/peer-announce.js +8 -6
- package/dist/command/version/version-nudge.d.ts +17 -3
- package/dist/command/version/version-nudge.js +25 -4
- package/dist/features/crawl/crawl-empty.js +1 -1
- package/dist/features/crawl/crawl-tools.d.ts +6 -0
- package/dist/features/crawl/crawl-tools.js +57 -0
- package/dist/features/crawl/crawl.js +171 -157
- package/dist/features/exhaust/app-coverage.d.ts +33 -0
- package/dist/features/exhaust/app-coverage.js +31 -0
- package/dist/features/exhaust/code-coverage.d.ts +47 -0
- package/dist/features/exhaust/code-coverage.js +77 -0
- package/dist/features/exhaust/exhaust-run.d.ts +33 -0
- package/dist/features/exhaust/exhaust-run.js +82 -0
- package/dist/features/exhaust/explorer.d.ts +103 -0
- package/dist/features/exhaust/explorer.js +195 -0
- package/dist/features/exhaust/failure-branches.d.ts +33 -0
- package/dist/features/exhaust/failure-branches.js +92 -0
- package/dist/features/exhaust/ledger.d.ts +202 -0
- package/dist/features/exhaust/ledger.js +252 -0
- package/dist/features/exhaust/session-fold.d.ts +31 -0
- package/dist/features/exhaust/session-fold.js +79 -0
- package/dist/features/exhaust/session-port.d.ts +36 -0
- package/dist/features/exhaust/session-port.js +102 -0
- package/dist/features/harness/drive-report.d.ts +0 -9
- package/dist/features/harness/drive-report.js +18 -0
- package/dist/features/harness/driver.js +4 -7
- package/dist/features/harness/fill-values.d.ts +10 -0
- package/dist/features/harness/fill-values.js +50 -11
- package/dist/features/harness/jev-driver.js +30 -13
- package/dist/features/harness/openai-driver.d.ts +12 -0
- package/dist/features/harness/openai-driver.js +21 -12
- package/dist/features/harness/platform-config.js +5 -6
- package/dist/features/license/license.d.ts +2 -2
- package/dist/index.d.ts +9 -0
- package/dist/index.js +56 -19
- package/dist/judgement/capsule/capsule-store.d.ts +4 -4
- package/dist/judgement/domain/domain-model.js +3 -1
- package/dist/judgement/oracles/self-instrument.js +4 -2
- package/dist/judgement/runs/artifact/replay-mapping.d.ts +2 -2
- package/dist/judgement/runs/artifact/replay-mapping.js +44 -9
- package/dist/judgement/runs/verification-sync.d.ts +3 -1
- package/dist/judgement/runs/verification-sync.js +1 -1
- package/dist/language/flows/change/affected.d.ts +9 -0
- package/dist/language/flows/change/affected.js +35 -0
- package/dist/language/flows/change/gate.d.ts +20 -0
- package/dist/language/flows/change/gate.js +12 -1
- package/dist/language/flows/flow-author.d.ts +7 -0
- package/dist/language/flows/flow-author.js +19 -0
- package/dist/language/flows/flow-expect-grammar.js +1 -1
- package/dist/language/flows/flow-journey.d.ts +54 -0
- package/dist/language/flows/flow-journey.js +119 -0
- package/dist/language/flows/flow-replay-run.d.ts +2 -3
- package/dist/language/flows/flow-replay-run.js +13 -47
- package/dist/language/flows/flow-replay.d.ts +22 -8
- package/dist/language/flows/flow-replay.js +55 -9
- package/dist/language/flows/flow-result.d.ts +2 -0
- package/dist/language/flows/flow-result.js +25 -0
- package/dist/language/flows/flow-tools.js +17 -5
- package/dist/language/flows/flows.d.ts +10 -1
- package/dist/language/flows/flows.js +90 -14
- package/dist/language/flows/recording/replay-halt.js +3 -1
- package/dist/language/flows/recording/tape/recordings.d.ts +30 -1
- package/dist/language/flows/recording/tape/recordings.js +55 -1
- package/dist/language/flows/replay.d.ts +19 -0
- package/dist/language/flows/replay.js +42 -6
- package/dist/language/flows/verify-change-tools.js +5 -2
- package/dist/launch-chromium.d.ts +80 -0
- package/dist/launch-chromium.js +218 -0
- package/dist/memory/cloud/cloud-config.d.ts +11 -2
- package/dist/memory/cloud/cloud-config.js +47 -7
- package/dist/memory/cloud/cloud-sync.d.ts +4 -1
- package/dist/memory/cloud/cloud-sync.js +6 -9
- package/dist/memory/cloud/harness-offer.js +5 -6
- package/dist/memory/cloud/harness-switch.js +5 -6
- package/dist/memory/cloud/sync-cycle.d.ts +77 -3
- package/dist/memory/cloud/sync-cycle.js +325 -55
- package/dist/memory/cloud/sync-daemon.js +24 -20
- package/dist/memory/cloud/sync-disk.js +16 -2
- package/dist/memory/intent/intent-shard.d.ts +10 -32
- package/dist/memory/intent/intent-shard.js +0 -10
- package/dist/memory/intent/intent-tools.js +8 -3
- package/dist/memory/journal/drive-flow.d.ts +18 -13
- package/dist/memory/journal/drive-flow.js +45 -37
- package/dist/memory/journal/envelope.d.ts +8 -8
- package/dist/memory/journal/journal-query.d.ts +14 -0
- package/dist/memory/journal/journal-query.js +15 -0
- package/dist/memory/journal/on-disk/workspace-tiers.js +2 -0
- package/dist/memory/journal/session-end.d.ts +9 -1
- package/dist/memory/journal/session-end.js +24 -12
- package/dist/memory/project/artifact-root-resolver.d.ts +10 -0
- package/dist/memory/project/artifact-root-resolver.js +46 -19
- package/dist/memory/project/dir/reticle-dir.d.ts +2 -0
- package/dist/memory/project/dir/reticle-dir.js +1 -0
- package/dist/memory/project/project-tools.js +3 -1
- package/dist/portal/bridge/auth-failure-reason.d.ts +1 -1
- package/dist/portal/bridge/auth-failure-reason.js +4 -2
- package/dist/portal/bridge/bridge.d.ts +12 -3
- package/dist/portal/bridge/bridge.js +49 -3
- package/dist/portal/input/js-coverage.d.ts +25 -0
- package/dist/portal/input/js-coverage.js +30 -0
- package/dist/portal/input/real-input.d.ts +11 -0
- package/dist/portal/input/real-input.js +29 -4
- package/dist/portal/input/viewport-tools.js +16 -6
- package/dist/portal/pool/playwright-launcher.d.ts +6 -0
- package/dist/portal/pool/playwright-launcher.js +20 -13
- package/dist/portal/session/dev-server/dev-command.d.ts +2 -2
- package/dist/portal/session/dev-server/dev-command.js +30 -12
- package/dist/portal/session/dev-server/dev-server-probe.js +17 -1
- package/dist/portal/session/dev-server/route-status-probe.d.ts +24 -0
- package/dist/portal/session/dev-server/route-status-probe.js +57 -0
- package/dist/portal/session/facts/observed-state.d.ts +11 -0
- package/dist/portal/session/facts/observed-state.js +34 -0
- package/dist/portal/session/no-session-diagnosis.d.ts +29 -0
- package/dist/portal/session/no-session-diagnosis.js +62 -5
- package/dist/portal/session/no-session-next-action.d.ts +13 -0
- package/dist/portal/session/no-session-next-action.js +48 -9
- package/dist/portal/session/no-session-watch.d.ts +18 -0
- package/dist/portal/session/no-session-watch.js +88 -2
- package/dist/portal/session/session-health.js +3 -2
- package/dist/portal/session/session-manager.d.ts +20 -6
- package/dist/portal/session/session-manager.js +19 -3
- package/dist/portal/session/session-reaper.d.ts +28 -0
- package/dist/portal/session/session-reaper.js +49 -1
- package/dist/portal/session/session.d.ts +10 -8
- package/dist/portal/session/session.js +24 -32
- package/dist/surface/http-server.d.ts +3 -2
- package/dist/surface/http-server.js +8 -5
- package/dist/surface/mcp/mcp-proxy.d.ts +2 -2
- package/dist/surface/mcp/mcp-proxy.js +2 -2
- package/dist/surface/mcp/mcp.js +5 -3
- package/dist/surface/mcp/proxy/proxy-daemon-probe.d.ts +13 -1
- package/dist/surface/mcp/proxy/proxy-daemon-probe.js +27 -2
- package/dist/surface/mcp/proxy/proxy-log.js +3 -5
- package/dist/surface/mcp/server-instructions.js +9 -7
- package/dist/surface/tools/act/after-match.d.ts +17 -0
- package/dist/surface/tools/act/after-match.js +30 -0
- package/dist/surface/tools/act/durable.d.ts +69 -0
- package/dist/surface/tools/act/durable.js +139 -0
- package/dist/surface/tools/act/navigation/navigate-arrival.d.ts +13 -6
- package/dist/surface/tools/act/navigation/navigate-arrival.js +47 -10
- package/dist/surface/tools/act/navigation/navigate-result.d.ts +1 -12
- package/dist/surface/tools/act/navigation/navigate-result.js +21 -3
- package/dist/surface/tools/act/proved-controls.d.ts +9 -0
- package/dist/surface/tools/act/proved-controls.js +11 -0
- package/dist/surface/tools/act/sequence-grade.d.ts +5 -0
- package/dist/surface/tools/act/sequence-grade.js +32 -1
- package/dist/surface/tools/act-capsule.d.ts +7 -0
- package/dist/surface/tools/act-capsule.js +46 -1
- package/dist/surface/tools/act-sequence-tool.js +25 -8
- package/dist/surface/tools/act-tools.js +28 -52
- package/dist/surface/tools/assert/assert-grade.js +12 -1
- package/dist/surface/tools/assert/assert-verdict.js +2 -2
- package/dist/surface/tools/browser-tools.js +14 -2
- package/dist/surface/tools/coverage-tools.js +27 -1
- package/dist/surface/tools/dynamic-tools.js +29 -2
- package/dist/surface/tools/error-recovery.js +5 -6
- package/dist/surface/tools/feedback-invite.js +6 -4
- package/dist/surface/tools/harness-explore.d.ts +3 -2
- package/dist/surface/tools/harness-explore.js +12 -7
- package/dist/surface/tools/invoke-tool.js +12 -8
- package/dist/surface/tools/lease-tools.js +15 -6
- package/dist/surface/tools/live-call-text.js +4 -1
- package/dist/surface/tools/observe-tools.js +9 -1
- package/dist/surface/tools/tool-kit.d.ts +10 -0
- package/dist/surface/tools/tools.js +4 -0
- package/dist/telemetry/argument-shape.d.ts +0 -2
- package/dist/telemetry/argument-shape.js +1 -5
- package/dist/telemetry/error-class.d.ts +2 -3
- package/dist/telemetry/error-class.js +2 -3
- package/dist/telemetry/harness-drive.d.ts +0 -2
- package/dist/telemetry/harness-drive.js +0 -4
- package/dist/telemetry/onboarding-firsts.d.ts +2 -2
- package/dist/telemetry/onboarding-firsts.js +2 -2
- package/dist/wire-journal.d.ts +3 -0
- package/dist/wire-journal.js +4 -0
- package/dist/workspace-packages.d.ts +0 -2
- package/dist/workspace-packages.js +1 -9
- package/docs/README.md +4 -2
- package/docs/actions.mdx +3 -4
- package/docs/agent-cheatsheet.md +4 -4
- package/docs/architecture.md +6 -6
- package/docs/best-practices.mdx +1 -1
- package/docs/capabilities.mdx +32 -1
- package/docs/cli/cloud.mdx +20 -8
- package/docs/cli/doctor.mdx +4 -0
- package/docs/cli/feedback.mdx +2 -2
- package/docs/cli/gate.mdx +9 -4
- package/docs/cli/update.mdx +10 -0
- package/docs/cli.mdx +22 -7
- package/docs/desktop.mdx +14 -10
- package/docs/environment.mdx +87 -0
- package/docs/faq.mdx +18 -16
- package/docs/first-drive.md +5 -13
- package/docs/flows.md +21 -2
- package/docs/frameworks.mdx +29 -10
- package/docs/getting-started.md +15 -5
- package/docs/http-transport.md +6 -3
- package/docs/human-control.md +2 -2
- package/docs/index.mdx +16 -6
- package/docs/install-agentic.mdx +2 -2
- package/docs/install-manual.mdx +1 -1
- package/docs/integration-patterns.md +3 -3
- package/docs/multi-agent-testing.md +1 -1
- package/docs/packages/babel-plugin.mdx +1 -1
- package/docs/packages/browser.mdx +2 -2
- package/docs/packages/next.mdx +3 -0
- package/docs/packages/react.mdx +2 -2
- package/docs/packages/tauri.mdx +1 -1
- package/docs/packages/vite-plugin.mdx +4 -2
- package/docs/packages.mdx +9 -9
- package/docs/predicates.mdx +50 -5
- package/docs/quickstart.mdx +12 -10
- package/docs/recipes.mdx +2 -2
- package/docs/state-management.mdx +2 -2
- package/docs/system-map.md +1 -1
- package/docs/telemetry.md +1 -1
- package/docs/token-efficiency.md +3 -3
- package/docs/tools/act-and-wait.mdx +1 -0
- package/docs/tools/act-sequence.mdx +1 -1
- package/docs/tools/act.mdx +1 -1
- package/docs/tools/console.mdx +1 -1
- package/docs/tools/inspect.mdx +1 -1
- package/docs/tools/network.mdx +1 -1
- package/docs/tools/query.mdx +1 -1
- package/docs/tools/sessions.mdx +1 -1
- package/docs/tools/snapshot.mdx +1 -1
- package/docs/tools/state.mdx +1 -1
- package/docs/tools/wait-for.mdx +1 -1
- package/docs/usage.md +7 -7
- package/docs/use-cases.mdx +157 -0
- package/docs/vs-screenshots.mdx +3 -3
- package/docs/what-is-recorded.md +2 -2
- package/docs/what-is-reticle.mdx +17 -20
- package/docs/why-reticle.mdx +14 -21
- package/package.json +10 -9
- package/dist/judgement/capsule/minimize.d.ts +0 -12
- package/dist/judgement/capsule/minimize.js +0 -22
- package/dist/judgement/phenomena/phenomena.d.ts +0 -27
- package/dist/judgement/phenomena/phenomena.js +0 -101
- package/dist/language/flows/flow-report.d.ts +0 -22
- package/dist/language/flows/flow-report.js +0 -106
|
@@ -0,0 +1,102 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* A live session, as the explorer's port.
|
|
3
|
+
*
|
|
4
|
+
* The one hard part is that a full page load REPLACES the session: the SDK reconnects under the
|
|
5
|
+
* same id (sessionStorage keeps it) but as a new object, and a command in flight on the old one is
|
|
6
|
+
* rejected. So every call reads the CURRENT session, and anything that may have navigated waits for
|
|
7
|
+
* the successor before it answers — the same wait `navigate { reload }` does.
|
|
8
|
+
*/
|
|
9
|
+
import { ActionType, ReticleCommand, ReticleTool, SnapshotMode, asRecord, asString, } from '@reticlehq/core';
|
|
10
|
+
import { RELOAD_RECONNECT_TIMEOUT_MS, waitForReconnect, } from '../../portal/session/session-reconnect.js';
|
|
11
|
+
import { isSessionReplacedError } from '../../portal/session/facts/session-replaced.js';
|
|
12
|
+
export function sessionPort(input) {
|
|
13
|
+
let current = input.start;
|
|
14
|
+
/** Wait for the document that replaced `before`. False when none arrived in time. */
|
|
15
|
+
const successor = async (before) => {
|
|
16
|
+
const back = await waitForReconnect({
|
|
17
|
+
current: () => input.sessions.get(before.id),
|
|
18
|
+
previous: before,
|
|
19
|
+
timeoutMs: RELOAD_RECONNECT_TIMEOUT_MS,
|
|
20
|
+
now: input.now,
|
|
21
|
+
sleep: input.sleep,
|
|
22
|
+
});
|
|
23
|
+
const next = input.sessions.get(before.id);
|
|
24
|
+
if (!back || next === undefined)
|
|
25
|
+
return false;
|
|
26
|
+
current = next;
|
|
27
|
+
return true;
|
|
28
|
+
};
|
|
29
|
+
const look = async () => {
|
|
30
|
+
for (let attempt = 0; attempt < 2; attempt += 1) {
|
|
31
|
+
try {
|
|
32
|
+
const res = await current.command(ReticleCommand.SNAPSHOT, {
|
|
33
|
+
mode: SnapshotMode.INTERACTIVE,
|
|
34
|
+
});
|
|
35
|
+
const body = asRecord(res.result);
|
|
36
|
+
const route = asString(asRecord(body['status'])['route']);
|
|
37
|
+
return { tree: asString(body['tree']) ?? '', ...(route === undefined ? {} : { route }) };
|
|
38
|
+
}
|
|
39
|
+
catch (error) {
|
|
40
|
+
// A click that followed a real link leaves the next read on a dead document.
|
|
41
|
+
if (!isSessionReplacedError(error) || !(await successor(current)))
|
|
42
|
+
break;
|
|
43
|
+
}
|
|
44
|
+
}
|
|
45
|
+
return { tree: '' };
|
|
46
|
+
};
|
|
47
|
+
return {
|
|
48
|
+
current: () => current,
|
|
49
|
+
look,
|
|
50
|
+
async reset() {
|
|
51
|
+
const before = current;
|
|
52
|
+
try {
|
|
53
|
+
const res = await before.command(ReticleCommand.NAVIGATE, { url: input.startUrl });
|
|
54
|
+
if (!res.ok)
|
|
55
|
+
return false;
|
|
56
|
+
}
|
|
57
|
+
catch (error) {
|
|
58
|
+
if (!isSessionReplacedError(error))
|
|
59
|
+
return false;
|
|
60
|
+
}
|
|
61
|
+
if (!(await successor(before)))
|
|
62
|
+
return false;
|
|
63
|
+
await input.sleep(input.settleMs);
|
|
64
|
+
return true;
|
|
65
|
+
},
|
|
66
|
+
async act(ref, action, value) {
|
|
67
|
+
const session = current;
|
|
68
|
+
const since = session.elapsed();
|
|
69
|
+
session.beginAction?.(ReticleTool.CRAWL, { ref, action });
|
|
70
|
+
let ok = false;
|
|
71
|
+
let replaced = false;
|
|
72
|
+
try {
|
|
73
|
+
const res = await session.command(ReticleCommand.ACT, {
|
|
74
|
+
ref,
|
|
75
|
+
action: 'fill' === action ? ActionType.FILL : ActionType.CLICK,
|
|
76
|
+
args: value === undefined ? {} : { value },
|
|
77
|
+
});
|
|
78
|
+
ok = res.ok && false !== asRecord(res.result)['dispatched'];
|
|
79
|
+
}
|
|
80
|
+
catch (error) {
|
|
81
|
+
// Following a real link replaces the document mid-command: the click WORKED.
|
|
82
|
+
if (!isSessionReplacedError(error))
|
|
83
|
+
throw error;
|
|
84
|
+
replaced = true;
|
|
85
|
+
ok = true;
|
|
86
|
+
}
|
|
87
|
+
finally {
|
|
88
|
+
session.finishAction?.();
|
|
89
|
+
}
|
|
90
|
+
await input.sleep(input.settleMs);
|
|
91
|
+
const events = session.eventsSince(since);
|
|
92
|
+
if (replaced && (await successor(session)))
|
|
93
|
+
return { ok, events: [...events, ...current.eventsSince(0)] };
|
|
94
|
+
return { ok, events };
|
|
95
|
+
},
|
|
96
|
+
...(input.setMocks === undefined
|
|
97
|
+
? {}
|
|
98
|
+
: {
|
|
99
|
+
mock: (rules) => input.setMocks?.(current.url, rules) ?? Promise.resolve(false),
|
|
100
|
+
}),
|
|
101
|
+
};
|
|
102
|
+
}
|
|
@@ -61,13 +61,4 @@ export declare function replayedFlows(toolCalls: readonly ToolOutcome[]): {
|
|
|
61
61
|
}[];
|
|
62
62
|
/** Reduce the raw call log to the actions that actually drove the app. */
|
|
63
63
|
export declare function drivenSteps(toolCalls: readonly ToolOutcome[]): DrivenStep[];
|
|
64
|
-
/**
|
|
65
|
-
* The account, as lines.
|
|
66
|
-
*
|
|
67
|
-
* Proved and not-proved are counted separately and both are stated, because the failure mode this
|
|
68
|
-
* whole product exists to prevent is a reader taking "we did twelve things" for "twelve things
|
|
69
|
-
* work". `unknown` is deliberately grouped under NOT proved and labelled as undecided evidence
|
|
70
|
-
* rather than as a failure — it calls for a better check, not a code change, and collapsing the two
|
|
71
|
-
* would send an agent to rewrite working code.
|
|
72
|
-
*/
|
|
73
64
|
export declare function describeDrive(toolCalls: readonly ToolOutcome[], savedFlows: readonly string[]): string;
|
|
@@ -134,6 +134,18 @@ export function drivenSteps(toolCalls) {
|
|
|
134
134
|
* rather than as a failure — it calls for a better check, not a code change, and collapsing the two
|
|
135
135
|
* would send an agent to rewrite working code.
|
|
136
136
|
*/
|
|
137
|
+
/** The pages the drive's own snapshots reported, in order, each distinct page once. */
|
|
138
|
+
function pagesReached(toolCalls) {
|
|
139
|
+
const pages = [];
|
|
140
|
+
for (const call of toolCalls) {
|
|
141
|
+
if (ReticleTool.SNAPSHOT !== call.name || call.isError)
|
|
142
|
+
continue;
|
|
143
|
+
const route = asString(asRecord(asRecord(call.result)['status'])['route']);
|
|
144
|
+
if (route !== undefined && !pages.includes(route))
|
|
145
|
+
pages.push(route);
|
|
146
|
+
}
|
|
147
|
+
return pages;
|
|
148
|
+
}
|
|
137
149
|
export function describeDrive(toolCalls, savedFlows) {
|
|
138
150
|
const steps = drivenSteps(toolCalls);
|
|
139
151
|
const replays = replayedFlows(toolCalls);
|
|
@@ -166,10 +178,16 @@ export function describeDrive(toolCalls, savedFlows) {
|
|
|
166
178
|
const proved = steps.filter((step) => Verified.YES === step.verified);
|
|
167
179
|
const failed = steps.filter((step) => Verified.NO === step.verified);
|
|
168
180
|
const undecided = steps.filter((step) => Verified.YES !== step.verified && Verified.NO !== step.verified);
|
|
181
|
+
const pages = pagesReached(toolCalls);
|
|
169
182
|
const lines = [
|
|
170
183
|
...(replayLine === undefined ? [] : [replayLine]),
|
|
171
184
|
`Drove ${String(steps.length)} action(s): ${String(proved.length)} proved, ` +
|
|
172
185
|
`${String(failed.length)} failed, ${String(undecided.length)} not decided.`,
|
|
186
|
+
// How DEEP it got. An action count reads the same for a checkout walked to its receipt and for
|
|
187
|
+
// eleven clicks on the home page; the pages reached are what tell them apart.
|
|
188
|
+
...(0 === pages.length
|
|
189
|
+
? []
|
|
190
|
+
: [`Reached ${String(pages.length)} page(s): ${pages.join(' → ')}.`]),
|
|
173
191
|
];
|
|
174
192
|
for (const step of steps.slice(0, MAX_LISTED)) {
|
|
175
193
|
const verdict = step.verified ?? 'nothing declared';
|
|
@@ -11,7 +11,7 @@
|
|
|
11
11
|
* purpose.
|
|
12
12
|
*/
|
|
13
13
|
import { z } from 'zod';
|
|
14
|
-
import { ReticleEnv,
|
|
14
|
+
import { ReticleEnv, platformCredentialFrom } from '@reticlehq/core';
|
|
15
15
|
/**
|
|
16
16
|
* The default driver model.
|
|
17
17
|
*
|
|
@@ -91,13 +91,10 @@ export function harnessOptionsFromEnv(env) {
|
|
|
91
91
|
*
|
|
92
92
|
* An explicit `ANTHROPIC_API_KEY` still wins: someone who exported one meant it.
|
|
93
93
|
*/
|
|
94
|
-
const
|
|
95
|
-
|
|
96
|
-
if (cloudKey === undefined || 0 === cloudKey.length)
|
|
94
|
+
const cloud = platformCredentialFrom(env);
|
|
95
|
+
if (cloud === undefined)
|
|
97
96
|
return undefined;
|
|
98
|
-
|
|
99
|
-
return undefined;
|
|
100
|
-
return withOptions(cloudKey, cloudUrl);
|
|
97
|
+
return withOptions(cloud.apiKey, cloud.url);
|
|
101
98
|
}
|
|
102
99
|
/** Build a driver backed by the Messages API. */
|
|
103
100
|
export function harnessDriver(options) {
|
|
@@ -31,6 +31,7 @@
|
|
|
31
31
|
* wrong verdict. Nothing here touches what is judged.
|
|
32
32
|
*/
|
|
33
33
|
import { type HarnessDriverOptions } from './driver.js';
|
|
34
|
+
import { type OpenAiDriverOptions } from './openai-driver.js';
|
|
34
35
|
/** What the caller must be able to do for values to survive past this drive. */
|
|
35
36
|
export interface FillCache {
|
|
36
37
|
get(label: string): string | undefined;
|
|
@@ -61,7 +62,16 @@ export declare function needsGeneration(label: string): boolean;
|
|
|
61
62
|
*/
|
|
62
63
|
export declare function fillValues(deps: {
|
|
63
64
|
cache?: FillCache;
|
|
65
|
+
/** GPT, the generator a Jev drive pairs with. Asked ahead of `generator` when both are set. */
|
|
66
|
+
openai?: OpenAiDriverOptions;
|
|
67
|
+
/** Anthropic, for a machine that holds only that key. */
|
|
64
68
|
generator?: HarnessDriverOptions;
|
|
65
69
|
/** One line about where the drive is, so a value suits the page it is typed on. */
|
|
66
70
|
context?: () => string;
|
|
71
|
+
/**
|
|
72
|
+
* A value the user SUPPLIED for this field — the same `RETICLE_SECRET_<FIELD>` replay reads for
|
|
73
|
+
* a redacted one. Checked first so a drive gets past a login, and never cached: the cache is a
|
|
74
|
+
* file on disk.
|
|
75
|
+
*/
|
|
76
|
+
secret?: (label: string) => string | undefined;
|
|
67
77
|
}): (label: string) => Promise<string>;
|
|
@@ -31,6 +31,7 @@
|
|
|
31
31
|
* wrong verdict. Nothing here touches what is judged.
|
|
32
32
|
*/
|
|
33
33
|
import { DEFAULT_HARNESS_BASE_URL, } from './driver.js';
|
|
34
|
+
import { DEFAULT_OPENAI_MODEL, OPENAI_REASONING_EFFORT, openAiCompletionsUrl, } from './openai-driver.js';
|
|
34
35
|
/** Deliberately small. A field value is a phrase; anything longer is the model misunderstanding. */
|
|
35
36
|
const MAX_TOKENS = 64;
|
|
36
37
|
const ANTHROPIC_VERSION = '2023-06-01';
|
|
@@ -87,6 +88,40 @@ const PROMPT = (label, context) => `A test harness is filling in a form field on
|
|
|
87
88
|
`Reply with ONLY the value to type. No quotes, no explanation, no markdown. It must be something a ` +
|
|
88
89
|
`real user would plausibly enter, and short. If the field wants an identifier or a code, invent a ` +
|
|
89
90
|
`well-formed one.`;
|
|
91
|
+
/**
|
|
92
|
+
* A model's one-line answer as a field value, or nothing. A model asked for one value occasionally
|
|
93
|
+
* answers with a sentence about the value; anything long enough to be prose is not a field value,
|
|
94
|
+
* and typing prose proves nothing about the app.
|
|
95
|
+
*/
|
|
96
|
+
function asValue(raw) {
|
|
97
|
+
const text = raw?.trim().split('\n')[0];
|
|
98
|
+
if (text === undefined || 0 === text.length)
|
|
99
|
+
return undefined;
|
|
100
|
+
return 80 < text.length ? undefined : text.replace(/^["']|["']$/g, '');
|
|
101
|
+
}
|
|
102
|
+
/** One string from GPT over Chat Completions — directly, or through the platform with its key. */
|
|
103
|
+
async function generateWithOpenAi(options, label, context) {
|
|
104
|
+
const call = options.fetch ?? ((url, init) => fetch(url, init));
|
|
105
|
+
try {
|
|
106
|
+
const res = await call(openAiCompletionsUrl(options), {
|
|
107
|
+
method: 'POST',
|
|
108
|
+
headers: { 'content-type': 'application/json', authorization: `Bearer ${options.apiKey}` },
|
|
109
|
+
body: JSON.stringify({
|
|
110
|
+
model: options.model ?? DEFAULT_OPENAI_MODEL,
|
|
111
|
+
max_completion_tokens: MAX_TOKENS,
|
|
112
|
+
reasoning_effort: OPENAI_REASONING_EFFORT,
|
|
113
|
+
messages: [{ role: 'user', content: PROMPT(label, context) }],
|
|
114
|
+
}),
|
|
115
|
+
});
|
|
116
|
+
if (!res.ok)
|
|
117
|
+
return undefined;
|
|
118
|
+
const parsed = JSON.parse(await res.text());
|
|
119
|
+
return asValue(parsed.choices?.[0]?.message?.content ?? undefined);
|
|
120
|
+
}
|
|
121
|
+
catch {
|
|
122
|
+
return undefined;
|
|
123
|
+
}
|
|
124
|
+
}
|
|
90
125
|
/** One string from a generating model, or nothing. Never throws: a drive must not fail over a field. */
|
|
91
126
|
async function generateOne(options, label, context) {
|
|
92
127
|
const call = options.fetch ?? ((url, init) => fetch(url, init));
|
|
@@ -109,17 +144,10 @@ async function generateOne(options, label, context) {
|
|
|
109
144
|
return undefined;
|
|
110
145
|
const parsed = JSON.parse(await res.text());
|
|
111
146
|
const content = parsed.content ?? [];
|
|
112
|
-
|
|
147
|
+
return asValue(content
|
|
113
148
|
.filter((block) => 'text' === block.type)
|
|
114
149
|
.map((block) => block.text ?? '')
|
|
115
|
-
.join('')
|
|
116
|
-
.trim()
|
|
117
|
-
.split('\n')[0];
|
|
118
|
-
if (text === undefined || 0 === text.length)
|
|
119
|
-
return undefined;
|
|
120
|
-
// A model asked for one value occasionally answers with a sentence about the value. Anything
|
|
121
|
-
// long enough to be prose is not a field value, and typing prose proves nothing about the app.
|
|
122
|
-
return 80 < text.length ? undefined : text.replace(/^["']|["']$/g, '');
|
|
150
|
+
.join(''));
|
|
123
151
|
}
|
|
124
152
|
catch {
|
|
125
153
|
// Offline, refused, rate-limited, malformed. All the same answer: use the heuristic.
|
|
@@ -135,14 +163,25 @@ async function generateOne(options, label, context) {
|
|
|
135
163
|
export function fillValues(deps) {
|
|
136
164
|
let spent = 0;
|
|
137
165
|
return async (label) => {
|
|
166
|
+
const supplied = deps.secret?.(label);
|
|
167
|
+
if (supplied !== undefined && supplied.length > 0)
|
|
168
|
+
return supplied;
|
|
138
169
|
const cached = deps.cache?.get(label);
|
|
139
170
|
if (cached !== undefined)
|
|
140
171
|
return cached;
|
|
141
172
|
const fallback = heuristicFillValue(label);
|
|
142
|
-
|
|
173
|
+
const { openai, generator } = deps;
|
|
174
|
+
if ((openai === undefined && generator === undefined) ||
|
|
175
|
+
!needsGeneration(label) ||
|
|
176
|
+
MAX_PER_DRIVE <= spent)
|
|
143
177
|
return fallback;
|
|
144
178
|
spent += 1;
|
|
145
|
-
const
|
|
179
|
+
const context = deps.context?.() ?? '';
|
|
180
|
+
const generated = openai !== undefined
|
|
181
|
+
? await generateWithOpenAi(openai, label, context)
|
|
182
|
+
: generator !== undefined
|
|
183
|
+
? await generateOne(generator, label, context)
|
|
184
|
+
: undefined;
|
|
146
185
|
if (generated === undefined)
|
|
147
186
|
return fallback;
|
|
148
187
|
// Cached under the LABEL, not the ref: refs expire with the page, labels are what the next drive
|
|
@@ -21,7 +21,7 @@
|
|
|
21
21
|
* exist because this product is the verdict, and a verdict sourced from a model's confidence is the
|
|
22
22
|
* thing Reticle sells the absence of.
|
|
23
23
|
*/
|
|
24
|
-
import { RETICLE_URL_PARAM, ReticleEnv, ReticleTool,
|
|
24
|
+
import { RETICLE_URL_PARAM, ReticleEnv, ReticleTool, platformCredentialFrom, asRecord, parseInteractive, } from '@reticlehq/core';
|
|
25
25
|
import { FINISH_TOOL } from './harness.js';
|
|
26
26
|
import { heuristicFillValue } from './fill-values.js';
|
|
27
27
|
export const DEFAULT_JEV_MODEL = 'jev-latest';
|
|
@@ -55,12 +55,15 @@ const Tool = {
|
|
|
55
55
|
RECORD: ReticleTool.RECORD,
|
|
56
56
|
FLOW_SAVE: ReticleTool.FLOW_SAVE,
|
|
57
57
|
FLOW_REPLAY: ReticleTool.FLOW_REPLAY,
|
|
58
|
+
NAVIGATE: ReticleTool.NAVIGATE,
|
|
58
59
|
FINISH: FINISH_TOOL.name,
|
|
59
60
|
};
|
|
60
61
|
/** The synthetic option meaning "stop driving". Not a tool arg — a key in the choice set. */
|
|
61
62
|
const FINISH_OPTION = 'finish';
|
|
62
63
|
/** The synthetic option meaning "I cannot tell from this page; look again". */
|
|
63
64
|
const RELOOK_OPTION = 'look_again';
|
|
65
|
+
/** Back to the page before this one — a dead end used to end the whole drive where it stood. */
|
|
66
|
+
const BACK_OPTION = 'go_back';
|
|
64
67
|
/**
|
|
65
68
|
* Tools the model may route to, beyond acting on a control.
|
|
66
69
|
*
|
|
@@ -244,14 +247,11 @@ export function jevOptionsFromEnv(env) {
|
|
|
244
247
|
};
|
|
245
248
|
}
|
|
246
249
|
// The ordinary path: the key the user minted on the platform, against the platform's own proxy.
|
|
247
|
-
// Reticle never ships a Jev key
|
|
248
|
-
const
|
|
249
|
-
|
|
250
|
-
if (cloudKey === undefined || 0 === cloudKey.length)
|
|
250
|
+
// Reticle never ships a Jev key; the URL is the configured platform, else the hosted service.
|
|
251
|
+
const cloud = platformCredentialFrom(env);
|
|
252
|
+
if (cloud === undefined)
|
|
251
253
|
return undefined;
|
|
252
|
-
|
|
253
|
-
return undefined;
|
|
254
|
-
return { apiKey: cloudKey, baseUrl: cloudUrl };
|
|
254
|
+
return { apiKey: cloud.apiKey, baseUrl: cloud.url };
|
|
255
255
|
}
|
|
256
256
|
/** Strip the snapshot's list marker, so a description reads as a phrase rather than a bullet. */
|
|
257
257
|
function cleanDesc(desc) {
|
|
@@ -287,6 +287,8 @@ function readState(history) {
|
|
|
287
287
|
saved: [],
|
|
288
288
|
tree: undefined,
|
|
289
289
|
route: undefined,
|
|
290
|
+
pages: [],
|
|
291
|
+
proved: [],
|
|
290
292
|
replayed: [],
|
|
291
293
|
acted: [],
|
|
292
294
|
blocked: [],
|
|
@@ -333,13 +335,22 @@ function readState(history) {
|
|
|
333
335
|
if ('string' === typeof tree)
|
|
334
336
|
state.tree = tree;
|
|
335
337
|
const route = asRecord(result['status'])['route'];
|
|
336
|
-
if ('string' === typeof route)
|
|
338
|
+
if ('string' === typeof route) {
|
|
337
339
|
state.route = appRoute(route);
|
|
340
|
+
if (state.pages.at(-1) !== route)
|
|
341
|
+
state.pages.push(route);
|
|
342
|
+
}
|
|
338
343
|
}
|
|
344
|
+
if (Tool.NAVIGATE === outcome.name)
|
|
345
|
+
state.tree = undefined;
|
|
339
346
|
if (Tool.ACT_AND_WAIT === outcome.name) {
|
|
340
347
|
const ref = asRecord(outcome.args)['ref'];
|
|
341
348
|
const failure = asRecord(outcome.result)['error'];
|
|
342
349
|
const refused = outcome.isError && 'string' === typeof failure && failure.includes(DANGEROUS_ARG);
|
|
350
|
+
const until = asRecord(asRecord(outcome.args)['until']);
|
|
351
|
+
const signal = until['name'];
|
|
352
|
+
if ('yes' === asRecord(outcome.result)['verified'] && 'string' === typeof signal)
|
|
353
|
+
state.proved.push(signal);
|
|
343
354
|
if ('string' === typeof ref) {
|
|
344
355
|
// A refused act never happened, so it is not something this drive has driven. Recording
|
|
345
356
|
// it as driven would retire the control after an attempt that never reached the app.
|
|
@@ -550,7 +561,7 @@ export function jevDriver(options) {
|
|
|
550
561
|
// ── the one real decision ────────────────────────────────────────────────────────────────
|
|
551
562
|
// What the drive is currently for: the first plan step that is not a replay. Replays are all
|
|
552
563
|
// done by the time we reach here, so whatever is left is the intent nobody has proved.
|
|
553
|
-
const objective = plan.find((step) => PLAN_REPLAY !== step.kind);
|
|
564
|
+
const objective = plan.find((step) => PLAN_REPLAY !== step.kind && !drive.proved.includes(step.target));
|
|
554
565
|
const candidates = candidatesFrom(drive.tree);
|
|
555
566
|
/**
|
|
556
567
|
* A page with nothing to act on ends the DRIVE, not just the journey.
|
|
@@ -567,11 +578,15 @@ export function jevDriver(options) {
|
|
|
567
578
|
return only(request(Tool.RECORD, { action: 'stop', recordingName: drive.open.name }), 'nothing interactive on this page; wind up');
|
|
568
579
|
}
|
|
569
580
|
const criteria = {
|
|
570
|
-
[FINISH_OPTION]: 'Stop driving:
|
|
581
|
+
[FINISH_OPTION]: 'Stop driving: every journey started here reached its end state, or nothing leads anywhere new.',
|
|
571
582
|
[RELOOK_OPTION]: 'Look at the page again without acting, because this reading looks incomplete.',
|
|
572
583
|
[OBSERVE_OPTION]: 'Read what the application DID since the last action — the requests it made and what it logged. Choose this when something was driven and it is not yet clear whether it worked.',
|
|
573
584
|
[STATE_OPTION]: "Read the application's own state, which is the strongest evidence there is and the one thing a screen cannot fake.",
|
|
574
585
|
};
|
|
586
|
+
const previous = drive.pages.at(-2);
|
|
587
|
+
if (previous !== undefined)
|
|
588
|
+
criteria[BACK_OPTION] =
|
|
589
|
+
`Go back to ${previous}: this page is a dead end, or a journey started there is not finished.`;
|
|
575
590
|
for (const candidate of candidates) {
|
|
576
591
|
const verb = actionFor(candidate.role);
|
|
577
592
|
const already = drive.acted.includes(candidate.ref) ? ' (already driven once)' : '';
|
|
@@ -580,12 +595,12 @@ export function jevDriver(options) {
|
|
|
580
595
|
const response = await callJev({ apiKey: options.apiKey, model, baseUrl, path, doFetch }, buildState(input.system, input.history, drive, objective), {
|
|
581
596
|
next_action: {
|
|
582
597
|
type: 'choice',
|
|
583
|
-
instructions: 'Which single action
|
|
598
|
+
instructions: 'Which single action carries the current user journey forward toward its END state — a submitted form, a saved record, a confirmation? Finish the journey this page started before opening a new one; prefer something not already driven only once it is finished.',
|
|
584
599
|
criteria,
|
|
585
600
|
},
|
|
586
601
|
journey_complete: {
|
|
587
602
|
type: 'noul',
|
|
588
|
-
instructions: '
|
|
603
|
+
instructions: 'Every journey started so far has been carried to its end state and checked there, and further driving would add nothing.',
|
|
589
604
|
},
|
|
590
605
|
});
|
|
591
606
|
const usage = {
|
|
@@ -602,6 +617,8 @@ export function jevDriver(options) {
|
|
|
602
617
|
}
|
|
603
618
|
if (RELOOK_OPTION === chosen)
|
|
604
619
|
return only(request(Tool.SNAPSHOT, { mode: 'interactive' }), 'look again', usage);
|
|
620
|
+
if (BACK_OPTION === chosen && previous !== undefined)
|
|
621
|
+
return only(request(Tool.NAVIGATE, { url: previous }), `back to ${previous}`, usage);
|
|
605
622
|
// Routing to a READ rather than an action. Neither changes the app, so neither is recorded as
|
|
606
623
|
// something driven — they are how the drive finds out whether what it drove worked.
|
|
607
624
|
if (OBSERVE_OPTION === chosen)
|
|
@@ -42,5 +42,17 @@ export interface OpenAiDriverOptions {
|
|
|
42
42
|
* OpenAI key, which is the entire point of the proxy.
|
|
43
43
|
*/
|
|
44
44
|
export declare function openAiOptionsFromEnv(env: Record<string, string | undefined>): OpenAiDriverOptions | undefined;
|
|
45
|
+
/**
|
|
46
|
+
* Where a Chat Completions call goes. The platform does not expose `/v1/chat/completions`; it
|
|
47
|
+
* exposes its own path and forwards. A base URL that is not OpenAI's is the platform's, which is
|
|
48
|
+
* the only other thing this talks to.
|
|
49
|
+
*/
|
|
50
|
+
export declare function openAiCompletionsUrl(options: Pick<OpenAiDriverOptions, 'baseUrl'>): string;
|
|
51
|
+
/**
|
|
52
|
+
* The small GPT models refuse function tools on Chat Completions while reasoning is on: measured,
|
|
53
|
+
* `gpt-5.6-luna` answers 400 to the first request that carries tools. A driving turn is a choice,
|
|
54
|
+
* not a plan, so no hidden reasoning is also the right depth and the cheapest one.
|
|
55
|
+
*/
|
|
56
|
+
export declare const OPENAI_REASONING_EFFORT = "none";
|
|
45
57
|
/** Build a driver backed by Chat Completions. */
|
|
46
58
|
export declare function openAiDriver(options: OpenAiDriverOptions): ModelDriver;
|
|
@@ -17,7 +17,7 @@
|
|
|
17
17
|
* it is reported below so the number is visible rather than assumed.
|
|
18
18
|
*/
|
|
19
19
|
import { z } from 'zod';
|
|
20
|
-
import { ReticleEnv,
|
|
20
|
+
import { ReticleEnv, platformCredentialFrom } from '@reticlehq/core';
|
|
21
21
|
/**
|
|
22
22
|
* The default model.
|
|
23
23
|
*
|
|
@@ -71,13 +71,10 @@ export function openAiOptionsFromEnv(env) {
|
|
|
71
71
|
const direct = env[ReticleEnv.HARNESS_OPENAI_KEY];
|
|
72
72
|
if (direct !== undefined && 0 < direct.length)
|
|
73
73
|
return withModel({ apiKey: direct });
|
|
74
|
-
const
|
|
75
|
-
|
|
76
|
-
if (cloudKey === undefined || 0 === cloudKey.length)
|
|
74
|
+
const cloud = platformCredentialFrom(env);
|
|
75
|
+
if (cloud === undefined)
|
|
77
76
|
return undefined;
|
|
78
|
-
|
|
79
|
-
return undefined;
|
|
80
|
-
return withModel({ apiKey: cloudKey, baseUrl: cloudUrl });
|
|
77
|
+
return withModel({ apiKey: cloud.apiKey, baseUrl: cloud.url });
|
|
81
78
|
}
|
|
82
79
|
function toWireTool(tool) {
|
|
83
80
|
return {
|
|
@@ -142,17 +139,28 @@ function parseArgs(raw) {
|
|
|
142
139
|
return {};
|
|
143
140
|
}
|
|
144
141
|
}
|
|
142
|
+
/**
|
|
143
|
+
* Where a Chat Completions call goes. The platform does not expose `/v1/chat/completions`; it
|
|
144
|
+
* exposes its own path and forwards. A base URL that is not OpenAI's is the platform's, which is
|
|
145
|
+
* the only other thing this talks to.
|
|
146
|
+
*/
|
|
147
|
+
export function openAiCompletionsUrl(options) {
|
|
148
|
+
const baseUrl = options.baseUrl ?? DEFAULT_OPENAI_BASE_URL;
|
|
149
|
+
return `${baseUrl}${DEFAULT_OPENAI_BASE_URL === baseUrl ? OPENAI_PATH : PLATFORM_PATH}`;
|
|
150
|
+
}
|
|
151
|
+
/**
|
|
152
|
+
* The small GPT models refuse function tools on Chat Completions while reasoning is on: measured,
|
|
153
|
+
* `gpt-5.6-luna` answers 400 to the first request that carries tools. A driving turn is a choice,
|
|
154
|
+
* not a plan, so no hidden reasoning is also the right depth and the cheapest one.
|
|
155
|
+
*/
|
|
156
|
+
export const OPENAI_REASONING_EFFORT = 'none';
|
|
145
157
|
/** Build a driver backed by Chat Completions. */
|
|
146
158
|
export function openAiDriver(options) {
|
|
147
159
|
const model = options.model ?? DEFAULT_OPENAI_MODEL;
|
|
148
|
-
const baseUrl = options.baseUrl ?? DEFAULT_OPENAI_BASE_URL;
|
|
149
160
|
const doFetch = options.fetch ?? ((url, init) => fetch(url, init));
|
|
150
|
-
// The platform does not expose `/v1/chat/completions`; it exposes its own path and forwards. A
|
|
151
|
-
// base URL that is not OpenAI's is the platform's, which is the only other thing this talks to.
|
|
152
|
-
const path = DEFAULT_OPENAI_BASE_URL === baseUrl ? OPENAI_PATH : PLATFORM_PATH;
|
|
153
161
|
return {
|
|
154
162
|
async turn(input) {
|
|
155
|
-
const response = await doFetch(
|
|
163
|
+
const response = await doFetch(openAiCompletionsUrl(options), {
|
|
156
164
|
method: 'POST',
|
|
157
165
|
headers: {
|
|
158
166
|
'content-type': 'application/json',
|
|
@@ -161,6 +169,7 @@ export function openAiDriver(options) {
|
|
|
161
169
|
body: JSON.stringify({
|
|
162
170
|
model,
|
|
163
171
|
max_completion_tokens: MAX_TOKENS,
|
|
172
|
+
reasoning_effort: OPENAI_REASONING_EFFORT,
|
|
164
173
|
tools: input.tools.map(toWireTool),
|
|
165
174
|
messages: [{ role: 'system', content: input.system }, ...toWireMessages(input.history)],
|
|
166
175
|
}),
|
|
@@ -16,7 +16,7 @@
|
|
|
16
16
|
* - it never overrides someone who was EXPLICIT. A driver named in the call, or set in the
|
|
17
17
|
* environment, wins over the stored preference. Precedence lives in the caller.
|
|
18
18
|
*/
|
|
19
|
-
import {
|
|
19
|
+
import { platformCredentialFrom } from '@reticlehq/core';
|
|
20
20
|
const CONFIG_PATH = '/v1/model/config';
|
|
21
21
|
/**
|
|
22
22
|
* How long the daemon will wait for a preference before driving without it.
|
|
@@ -68,12 +68,11 @@ function parse(body) {
|
|
|
68
68
|
* warning printed on a path that then works correctly is how people learn to ignore stderr.
|
|
69
69
|
*/
|
|
70
70
|
export async function fetchPlatformConfig(env, doFetch = (url, init) => fetch(url, init), timeoutMs = TIMEOUT_MS) {
|
|
71
|
-
const
|
|
72
|
-
|
|
73
|
-
if (key === undefined || 0 === key.length)
|
|
74
|
-
return undefined;
|
|
75
|
-
if (host === undefined || 0 === host.length)
|
|
71
|
+
const cloud = platformCredentialFrom(env);
|
|
72
|
+
if (cloud === undefined)
|
|
76
73
|
return undefined;
|
|
74
|
+
const key = cloud.apiKey;
|
|
75
|
+
const host = cloud.url;
|
|
77
76
|
const controller = new AbortController();
|
|
78
77
|
const timer = setTimeout(() => {
|
|
79
78
|
controller.abort();
|
|
@@ -39,14 +39,14 @@ declare const LicensePayloadSchema: z.ZodObject<{
|
|
|
39
39
|
exp: z.ZodNumber;
|
|
40
40
|
features: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
41
41
|
}, "strip", z.ZodTypeAny, {
|
|
42
|
-
lid: string;
|
|
43
42
|
org: string;
|
|
43
|
+
lid: string;
|
|
44
44
|
plan: string;
|
|
45
45
|
exp: number;
|
|
46
46
|
features?: string[] | undefined;
|
|
47
47
|
}, {
|
|
48
|
-
lid: string;
|
|
49
48
|
org: string;
|
|
49
|
+
lid: string;
|
|
50
50
|
plan: string;
|
|
51
51
|
exp: number;
|
|
52
52
|
features?: string[] | undefined;
|
package/dist/index.d.ts
CHANGED
|
@@ -15,6 +15,15 @@ export type { ToolDeps, ToolDef } from './surface/tools/tools.js';
|
|
|
15
15
|
export { createToolInvoker, UNKNOWN_TOOL_ERROR } from './surface/tools/tool-invoker.js';
|
|
16
16
|
export { runTool, SESSION_BOUND_TOOLS, SESSION_EXEMPT_TOOLS } from './surface/tools/invoke-tool.js';
|
|
17
17
|
export type { ToolInvoker } from './surface/tools/tool-invoker.js';
|
|
18
|
+
/**
|
|
19
|
+
* The port-presence check the CLI commands (`drive`, `verify`, `status`, `doctor`, `kill`) all share,
|
|
20
|
+
* re-exported so a consumer embedding `start()` in its own process can refuse a taken port the same
|
|
21
|
+
* honest way instead of letting a raw `EADDRINUSE` stack out — the gap `@reticlehq/test`'s
|
|
22
|
+
* `bootSession()` had (reticlehq/reticle#1141): it called `start()` directly with no pre-flight probe.
|
|
23
|
+
*/
|
|
24
|
+
export { probePresence, describePresence, PortPresence, } from './command/daemon/binding/port-presence.js';
|
|
25
|
+
export { probeDaemon } from './surface/mcp/proxy/proxy-daemon-probe.js';
|
|
26
|
+
export { fetchStatus } from './command/daemon/binding/daemon-status-probe.js';
|
|
18
27
|
export { BaselineStore, normalizeLines, diffLines } from './memory/project/baselines.js';
|
|
19
28
|
export { RecordingStore } from './language/flows/recording/tape/recordings.js';
|
|
20
29
|
export type { RecordedStep, CompiledProgram } from './language/flows/recording/tape/recordings.js';
|