refactor(config): take the environment as an input at every configuration read (#554)

* fix(platform): accept absolute paths for any host and stop blaming the ceiling for a missing grant

Addresses #548 and #545.

#548: processScenarioSchema required a leading "/" for executable,
working_directory and filesystem_roots, so no process scenario could be
expressed on Windows at all. The report suggested node:path's isAbsolute, but
that answers for the host running the check: on Linux `C:\tools` is not
absolute, so a scenario captured on Windows would fail validation when it is
replayed, compared, or asserted in a test on another host. A boundary schema
must not vary with its evaluator, so the check accepts the POSIX form, a
Windows drive path and a UNC share everywhere. The rejection message now names
the constraint instead of leaving the caller to infer it from "invalid request".
The HTTP `path` field inside `checkpoints` is an URL path and is deliberately
left alone.

I checked whether this merely moved the failure downstream: containment is
enforced by isPathWithinRoot, which already uses host-aware path.relative, so
validation really was the only blocker.

#545: when a scope is inside the administrator ceiling but no grant was ever
issued, the denial still pointed at widening the ceiling -- advice that cannot
help, since the ceiling already permits the operation. The `missing` payload is
computed against candidate *grants*, so it describes what a grant must cover,
not a ceiling gap. `PermissionRemediation` gains a `grant` variant saying
exactly that, and PermissionAuthority selects it when the denial is not a
ceiling violation. An explicitly requested remediation is still honoured, and a
genuine ceiling violation still says "configure".

* refactor(config): take the environment as an input at every configuration read

Completes B-2. Every `parseConfig(process.env)` site now reads from an
injected environment, so configuration-driven behaviour is a function of an
argument rather than of the machine running it.

The twelve sites were not equivalent, and treating them as one number would
have hidden the difference:

- **src/application/ (3 sites) were genuine defects.** `runSessionStatus`
  already injected its `logger` but hardcoded `process.env` for config, so the
  MCP path's `dependencies.env` never reached it; `captureProcessScenarioFile`
  and `runAnalysis` had the same problem. `runAnalysis` gained `environment` on
  the options object it already had, which is where its other injected
  dependencies live.

- **The CLI command layer (9 sites) is a documented boundary,** where
  `process.env` is legitimate. Rather than 9 independent ambient reads, they
  now share one seam: `createCli(environment)` passes it to the seven
  registrars that need configuration, and the nested helpers (`electronContext`,
  `runtimeContext`, `loadConfiguredAuthority`, `inspectJavaScriptReplay`) take
  it explicitly.

Every new parameter defaults to `process.env`, so all existing call sites are
unchanged and production behaviour is identical. The point is that the ambient
read is now a documented default at a boundary rather than a hidden dependency
deep inside behaviour.

`process.env` now survives in `src/` only as those default parameters —
nothing reads it at the point of use.

* style(cli): apply prettier formatting to the runtime observation command
This commit is contained in:
morluto authored and GitHub committed 2026-10-05 20:00:16 +08:00
1 parent 778e464c20
commit 53fe7d6d3b
13 files changed
+131 -43

No files matched your search

+10 -3
View File
@@ -40,6 +40,7 @@ const observeOptionsSchema = z.object({
export const registerJavaScriptRuntimeObservationCommands = (
cli: ReturnType<typeof Cli.create>,
logger: Logger,
environment: Readonly<Record<string, string | undefined>> = process.env,
): void => {
cli.command(CLI_COMMANDS.listJavaScriptRuntimeTargets, {
description: "List Node/Electron V8 Inspector targets",
@@ -81,7 +82,10 @@ export const registerJavaScriptRuntimeObservationCommands = (
options: observeOptionsSchema,
run: ({ args, options }) =>
logCliCommand(logger, CLI_COMMANDS.observeJavaScriptRuntime, async () => {
const context = await runtimeContext("observe_javascript_runtime");
const context = await runtimeContext(
"observe_javascript_runtime",
environment,
);
if (!context.ok) return context.error;
const parsed = observeJavaScriptRuntimeInputSchema.safeParse({
inspector_endpoint: args.endpoint,
@@ -100,8 +104,11 @@ export const registerJavaScriptRuntimeObservationCommands = (
});
};
const runtimeContext = async (operation: string) => {
const config = parseConfig(process.env);
const runtimeContext = async (
operation: string,
environment: Readonly<Record<string, string | undefined>> = process.env,
) => {
const config = parseConfig(environment);
if (!config.ok) return { ok: false as const, error: cliError(config.error) };
const policy = config.value.v8InspectorObservationPolicy;
if (policy.status === "disabled")