Node-based terminal manager for AI coding agents — tmux-backed terminals and parallel agent sessions as draggable nodes on an infinite pan/zoom canvas. macOS, Linux, and a browser Server Edition.
TypeScript adhdagent-orchestrationai-agentscanvasclaude-codecodexcoding-agentsdeveloper-toolselectrongeminilinuxmacosmultiplexeropencodeparallel-agentsterminalterminal-emulatorterminal-multiplexertmuxworkspace-manager
test(ssh): guard that remote shims stay machine-neutral
freshjuicez06eb58281178
diff --git a/src/main/remote-ssh/remote-shim-neutrality.guard.test.ts b/src/main/remote-ssh/remote-shim-neutrality.guard.test.tsnew file mode 100644index 000000000..2761b8036--- /dev/null+++ b/src/main/remote-ssh/remote-shim-neutrality.guard.test.ts@@ -0,0 +1,107 @@+// The shim scripts copied onto an SSH host must stay MACHINE-NEUTRAL: no path that only means+// something on the desktop/server that generated them.+//+// Since the local shims gained the shared-Codex identity prelude, the two builders take an+// `identityRoot` — this machine's `<userDataDir>/codex-thread-nodes`. The LOCAL installers pass it;+// the SSH installer must not, because a record root under one machine's user-data dir is not a+// path on somebody else's server. The rule is stated in three places (CLAUDE.md, CONTRIBUTING.md,+// docs/shared-codex-node-identity.md) and was enforced by none of them.+//+// WHY IT NEEDS A GUARD RATHER THAN A REVIEWER: the failure is SILENT AND ONE-SIDED. A remote shim+// carrying the prelude still works — the outer `[ -n "$CODEX_THREAD_ID" ]` guard is false in a+// remote pane, and even if it were true the baked root does not exist there, so every candidate+// read fails and the block changes nothing. Nothing breaks, no test goes red, no user reports a+// bug. The only symptom is that this machine's user-data layout is now sitting in a file on+// someone else's server, written there by us, and refreshed on every connect.+//+// Two legs, because they catch the mistake at different moments:+// 1. the VALUE — the neutral exports the SSH installer writes carry nothing machine-specific;+// 2. the IMPORT — `remote-hooks.ts` cannot even reach a builder that would produce one.+// Leg 2 is the one that survives a refactor: a future author who reaches for the parameterised+// builder "so the remote shim gets the fix too" is stopped at the import, before there is a value+// to assert about.+//+// Every assertion has a POSITIVE CONTROL beside it. A guard whose subject has been renamed or+// removed matches nothing and reports clean — the same failure mode `fs-atomic.guard.test.ts`+// opens with — and a silent pass here would restore exactly the blindness this file exists to end.+import { describe, expect, it } from 'vitest'+import { readFileSync } from 'fs'+import { join } from 'path'++import { CONTROL_SHIM_SCRIPT, buildControlShimScript } from '../canvas-control-core'+import { CONTEXT_SHIM_SCRIPT, buildContextShimScript } from '../../core/context-link-core'++/** A root no real machine would produce, so a hit is unambiguously the one we baked in. */+const SENTINEL_ROOT = '/nodeterm-guard-sentinel/codex-thread-nodes'++const REMOTE_HOOKS = join(__dirname, 'remote-hooks.ts')++/**+ * Identifiers `remote-hooks.ts` must never name. Each one is a way to obtain a shim body carrying+ * a local path, or the local path itself.+ */+const BANNED_IN_REMOTE_HOOKS = [+ 'buildControlShimScript',+ 'buildContextShimScript',+ 'codexThreadIdentityResolverSh',+ 'codexThreadIdentityRoot'+] as const++/**+ * Identifiers `remote-hooks.ts` must still name — the neutral exports it writes to the host. If+ * these disappear the file has been restructured and leg 2's scan is no longer looking at the+ * thing it was written to look at, so it must go red rather than pass over an empty subject.+ */+const REQUIRED_IN_REMOTE_HOOKS = ['CONTROL_SHIM_SCRIPT', 'CONTEXT_SHIM_SCRIPT'] as const++/**+ * Strip comments before scanning, so this file's own rule can be EXPLAINED at the call site in+ * `remote-hooks.ts` without the explanation tripping the scan. Only comments are removed; a string+ * literal naming a banned identifier is still a hit, which is the conservative direction.+ */+function code(source: string): string {+ return source.replace(/\/\*[\s\S]*?\*\//g, '').replace(/(^|[^:])\/\/[^\n]*/g, '$1')+}++describe('SSH-installed shims stay machine-neutral', () => {+ describe('leg 1 — the exported neutral bodies carry no local path', () => {+ for (const [name, neutral, build] of [+ ['canvas control', CONTROL_SHIM_SCRIPT, buildControlShimScript],+ ['linked context', CONTEXT_SHIM_SCRIPT, buildContextShimScript]+ ] as const) {+ it(`${name}: the builder DOES bake a root when given one (positive control)`, () => {+ // Without this, the two assertions below would also pass if the prelude had been dropped+ // from the local shims altogether — a regression, reported as compliance.+ const local = build(SENTINEL_ROOT)+ expect(local).toContain(SENTINEL_ROOT)+ expect(local).toContain('CODEX_THREAD_ID')+ })++ it(`${name}: the neutral export names no record root`, () => {+ expect(neutral).not.toContain('codex-thread-nodes')+ })++ it(`${name}: the neutral export carries no identity prelude at all`, () => {+ // Broader than the path check on purpose. The path is what leaks today; the prelude is the+ // only thing that can carry one, so refusing the whole block refuses the next spelling of+ // the leak too (a root read from an env var, a relative path, a default argument).+ expect(neutral).not.toContain('CODEX_THREAD_ID')+ expect(build()).toBe(neutral)+ })+ }+ })++ describe('leg 2 — remote-hooks.ts cannot reach a parameterised builder', () => {+ const source = code(readFileSync(REMOTE_HOOKS, 'utf8'))++ it('is reading the real remote installer (positive control)', () => {+ for (const required of REQUIRED_IN_REMOTE_HOOKS) expect(source).toContain(required)+ })++ for (const banned of BANNED_IN_REMOTE_HOOKS) {+ it(`never names ${banned}`, () => {+ expect(source).not.toContain(banned)+ })+ }+ })+})