@smthrs/kernel 0.0.0-stage → 1.0.0-rc.4

package/src/ProcessConfinement.ts

@@ -0,0 +1,156 @@
+/** Operating-system confinement derived conservatively from effective grants.
+ * @since 1.0.0-rc.1
+ */
+
+import {
+  type Action,
+  Capability,
+  CapabilityPattern,
+  matches,
+  maxResourceLength,
+  subsumes,
+  withinMatchBudget
+} from "@smthrs/capability/Capability"
+import type { GrantStoreError } from "@smthrs/capability/Permission"
+import { Context, Effect, Layer, type Path, type Scope } from "effect"
+import type { PlatformError } from "effect/PlatformError"
+import type * as ChildProcess from "effect/unstable/process/ChildProcess"
+import type { Policy, PolicyRule, Service as GrantService } from "./GrantStore.ts"
+
+/** Workspace-relative trees; `.` names the workspace. No path opens its parent.
+ * @category models
+ * @since 1.0.0-rc.1
+ */
+export interface Profile {
+  readonly workspaceRoot: string
+  readonly reads: ReadonlyArray<string>
+  readonly writes: ReadonlyArray<string>
+  readonly readOnly: ReadonlyArray<string>
+  readonly network: "none" | "open"
+}
+/** Host wrapper for one process stage. Failure must occur before spawning.
+ * @category models
+ * @since 1.0.0-rc.1
+ */
+export interface Service {
+  readonly confine: (
+    command: ChildProcess.StandardCommand,
+    profile: Profile
+  ) => Effect.Effect<ChildProcess.StandardCommand, PlatformError, Scope.Scope>
+}
+/** Host confinement seam; missing service refuses spawning.
+ * @category services
+ * @since 1.0.0-rc.1
+ */
+export class ProcessConfinement
+  extends Context.Service<ProcessConfinement, Service>()("@smthrs/kernel/ProcessConfinement")
+{}
+/** Explicit unconfined composition.
+ * @category constructors
+ * @since 1.0.0-rc.1
+ */
+export const makeNoop: Service = ProcessConfinement.of({ confine: (command) => Effect.succeed(command) })
+/** Explicit unconfined layer.
+ * @category layers
+ * @since 1.0.0-rc.1
+ */
+export const layerNoop: Layer.Layer<ProcessConfinement> = Layer.succeed(ProcessConfinement)(makeNoop)
+
+const decidable = (pattern: CapabilityPattern, action: Action) =>
+  withinMatchBudget(pattern, new Capability({ action, resource: "x".repeat(maxResourceLength) }))
+
+// The capability matcher permits a single star to cross separators. Only
+// this universal spelling is equivalent to **; arbitrary globs stay closed.
+const universal = (pattern: CapabilityPattern): CapabilityPattern =>
+  pattern.resource === "*" ? new CapabilityPattern({ action: pattern.action, resource: "**" }) : pattern
+
+const coversCeiling = (ceiling: PolicyRule["ceiling"], target: CapabilityPattern): boolean =>
+  ceiling.groups.every((group) =>
+    group.some((pattern) => decidable(pattern, target.action as Action) && subsumes(universal(pattern), target))
+  )
+const actionRelevant = (pattern: CapabilityPattern, action: Action): boolean =>
+  matches(new CapabilityPattern({ action: pattern.action, resource: "**" }), new Capability({ action, resource: "" }))
+
+// A covering rule resolves the complete region. A narrower/conditional refusal
+// makes it unknown until a later covering rule resolves it. Unknown configured
+// denial vetoes all later grant groups, just as evaluate's configured veto does.
+const prove = (policy: Policy, target: CapabilityPattern): boolean => {
+  if (!coversCeiling(policy.ceiling, target)) return false
+  let decision: "allow" | "deny" | "ask" | "unknown" = "ask"
+  let configured: "allow" | "deny" | "ask" | "unknown" = "ask"
+  for (const [index, group] of policy.groups.entries()) {
+    for (const { rule, ceiling } of group) {
+      if (!decidable(rule.pattern, target.action as Action)) return false
+      if (!actionRelevant(rule.pattern, target.action as Action)) continue
+      if (target.action === "net:private" && rule.effect === "allow" && rule.pattern.action !== "net:private") continue
+      if (coversCeiling(ceiling, target) && subsumes(universal(rule.pattern), target)) decision = rule.effect
+      else if (rule.effect !== "allow") decision = "unknown"
+      if (index === 0) configured = decision
+    }
+  }
+  return configured !== "deny" && configured !== "unknown" && decision === "allow"
+}
+
+/** Derives a subset of authority, never a widening. Literal filesystem grants,
+ * complex globs and unrepresentable restrictions conservatively open nothing.
+ * Writable trees also require read authority and write authority over their
+ * exact directory root: native mounts permit reads and directory metadata
+ * changes. Only whole-tree grants prove filesystem trees; unrestricted
+ * authority for each of get/post/private is required to open network access.
+ * @category resolution
+ * @since 1.0.0-rc.1
+ */
+export const profile = (
+  grants: GrantService,
+  workspaceRoot: string,
+  path: Path.Path
+): Effect.Effect<Profile, GrantStoreError> =>
+  Effect.gen(function*() {
+    const policy = yield* grants.policy
+    const root = path.resolve(workspaceRoot)
+    const resources = [
+      ...policy.groups.flatMap((group) =>
+        group.flatMap(({ rule, ceiling }) => [rule.pattern, ...ceiling.groups.flat()])
+      ),
+      ...policy.ceiling.groups.flat()
+    ]
+    const candidates = new Set<string>()
+    for (const { resource } of resources) {
+      if (resource === "**" || resource === "*") {
+        candidates.add(".")
+        continue
+      }
+      if (!resource.endsWith("/**") || /[*?]/.test(resource.slice(0, -3))) continue
+      const base = resource.slice(0, -3)
+      // Relative policy resources do not authorize absolute kernel resources.
+      if (!path.isAbsolute(base) || path.normalize(base) !== base) continue
+      const relative = path.relative(root, base)
+      if (path.isAbsolute(relative) || relative.split(path.sep)[0] === "..") continue
+      candidates.add(relative === "" ? "." : relative.split(path.sep).join("/"))
+    }
+    const opened = (action: Action) =>
+      [...candidates].filter((relative) => {
+        const directory = relative === "." ? root : path.resolve(root, relative)
+        const resource = `${directory}/**`
+        return prove(policy, new CapabilityPattern({ action, resource })) &&
+          (action !== "fs:write" || (
+            prove(policy, new CapabilityPattern({ action: "fs:read", resource })) &&
+            prove(policy, new CapabilityPattern({ action: "fs:write", resource: directory }))
+          ))
+      })
+        .filter((relative, _, all) =>
+          !all.some((parent) => parent !== relative && (parent === "." || relative.startsWith(`${parent}/`)))
+        )
+        .sort()
+    return Object.freeze({
+      workspaceRoot: root,
+      reads: Object.freeze(opened("fs:read")),
+      writes: Object.freeze(opened("fs:write")),
+      readOnly: Object.freeze([]),
+      network: (["net:get", "net:post", "net:private"] as const).every((action) =>
+          prove(policy, new CapabilityPattern({ action, resource: "**" }))
+        ) ?
+        "open" as const :
+        "none" as const
+    })
+  })

package/src/ProcessLedger.ts

@@ -0,0 +1,423 @@
+/**
+ * The host's ledger of the processes it spawned.
+ *
+ * A cancelled run must not leave a child agent running, and the engine cannot
+ * do anything about that on its own: cancellation interrupts fibers, and a
+ * fiber closing its scope only reaches the processes THIS incarnation of the
+ * host still holds a handle for. A host that crashed left its children with no
+ * one holding anything.
+ *
+ * The ledger is the missing half. Every spawn is recorded here in memory and
+ * written to the journal as an ownerless durable record. It is ownerless
+ * because it describes the host rather than a run, and first-writer-wins is
+ * exactly the semantics a spawn record wants. The next incarnation of the same
+ * `hostId` replays those records, subtracts the ones that reported an exit,
+ * and is left with {@link Service.orphans}: the process groups a dead owner
+ * abandoned, which a platform reaper can then signal.
+ *
+ * The kernel stays BELOW the engine: this module records pids and process
+ * groups and knows nothing about runs, attempts, or ownership fences. It also
+ * stays platform-neutral, because signalling a process group is a platform
+ * concern and lives in `@smthrs/platform-node`.
+ *
+ * Governing design:
+ * `docs/specs/Concepts/Host Adapters.md` and
+ * `docs/specs/Concepts/Journal Queue.md`.
+ *
+ * @since 1.0.0-rc.0
+ */
+
+import * as JournalModule from "@smthrs/journal/Journal"
+import type { JournalError } from "@smthrs/journal/Journal"
+import * as JournalEvent from "@smthrs/journal/JournalEvent"
+import * as Context from "effect/Context"
+import * as Effect from "effect/Effect"
+import * as Layer from "effect/Layer"
+import * as Schema from "effect/Schema"
+
+/**
+ * The event type recording that a host started a process.
+ *
+ * @category models
+ * @since 1.0.0-rc.0
+ */
+export const SpawnedEventType = "flows.host.process-spawned.v1"
+
+/**
+ * The event type recording that a recorded process ended.
+ *
+ * @category models
+ * @since 1.0.0-rc.0
+ */
+export const ExitedEventType = "flows.host.process-exited.v1"
+
+/**
+ * The event type recording that a later incarnation killed an abandoned
+ * process group.
+ *
+ * @category models
+ * @since 1.0.0-rc.0
+ */
+export const ReapedEventType = "flows.host.process-reaped.v1"
+
+/**
+ * The event type recording that a later incarnation retired a record WITHOUT
+ * signalling anything.
+ *
+ * A stored pid outlives the process that wrote it, so a reaper that cannot
+ * prove the number still names the process the record describes must not
+ * signal it. Retiring the record is still correct: nothing will ever be able
+ * to prove it again, and leaving it would make every later incarnation
+ * re-examine a number the operating system has moved on from.
+ *
+ * @category models
+ * @since 1.0.0-rc.0
+ */
+export const SkippedEventType = "flows.host.process-reap-skipped.v1"
+
+/**
+ * The producer id every ledger record is written under.
+ *
+ * Replay checks it: a record forged by another producer cannot make a host
+ * signal a process group.
+ *
+ * @category models
+ * @since 1.0.0-rc.0
+ */
+export const sourceId = "@smthrs/kernel/ProcessLedger"
+
+/**
+ * The journal run a host's process records live on.
+ *
+ * Process records describe a host, not a run, so they get a run of their own,
+ * named after the host. Two incarnations of one `hostId` therefore share a
+ * history, which is what makes an abandoned process discoverable.
+ *
+ * @category constructors
+ * @since 1.0.0-rc.0
+ */
+export const hostRunId = (hostId: string): JournalEvent.RunId => `flows.host:${hostId}` as JournalEvent.RunId
+
+/**
+ * What the spawner knows about a process the moment it starts.
+ *
+ * `pgid` is `null` when the child shares the host's own process group. Such a
+ * process cannot be group-signalled — signalling that group would signal the
+ * host itself — so the ledger records the absence rather than a number a
+ * reaper could misread.
+ *
+ * `commandDigest` names the executable and nothing else, as
+ * `CommandLine.executable` renders it. Arguments carry credentials
+ * (`curl -u user:password`, `mysql -phunter2`) and these records are permanent
+ * journal entries, so the ledger never writes an argument down. A reader that
+ * has to act on a record matches it by pid and process group; the program name
+ * is there to make the row legible.
+ *
+ * @category models
+ * @since 1.0.0-rc.0
+ */
+export interface Spawned {
+  readonly pid: number
+  readonly pgid: number | null
+  readonly commandDigest: string
+}
+
+/**
+ * Schema for a recorded process, as it round-trips through the journal.
+ *
+ * @category schemas
+ * @since 1.0.0-rc.0
+ */
+export const ProcessRecord = Schema.Struct({
+  pid: Schema.Int,
+  pgid: Schema.NullOr(Schema.Int),
+  hostId: Schema.String,
+  ownerPid: Schema.Int,
+  startedAtMs: Schema.Int,
+  commandDigest: Schema.String
+})
+
+/**
+ * One process a host started, with the identity of the incarnation that
+ * started it.
+ *
+ * @category models
+ * @since 1.0.0-rc.0
+ */
+export type ProcessRecord = typeof ProcessRecord.Type
+
+/**
+ * Why a ledger history could not be replayed.
+ *
+ * - `journal_unreadable`: a page of the host's history could not be read.
+ * - `cursor_stalled`: the journal reported more history but did not advance.
+ *
+ * @category models
+ * @since 1.0.0-rc.1
+ */
+export const ProcessLedgerReplayErrorCode = Schema.Literals(["journal_unreadable", "cursor_stalled"])
+
+/**
+ * The ledger could not read a previous incarnation's history to its end.
+ *
+ * An unread history is never reported as an empty one: a caller that treated
+ * it as empty would leave every abandoned child running with no owner.
+ *
+ * @category errors
+ * @since 1.0.0-rc.1
+ */
+export class ProcessLedgerReplayError extends Schema.TaggedError<ProcessLedgerReplayError>()(
+  "@smthrs/kernel/ProcessLedgerReplayError",
+  {
+    code: ProcessLedgerReplayErrorCode,
+    message: Schema.String,
+    cause: Schema.optional(Schema.Defect())
+  }
+) {}
+
+const decodeRecord = Schema.decodeUnknownResult(ProcessRecord)
+const encodeRecord = Schema.encodeSync(ProcessRecord)
+
+/**
+ * Ledger operations.
+ *
+ * Every write that has a durable half REPORTS whether that half committed.
+ * The in-memory set is still updated first, so a caller that decides to carry
+ * on has an accurate view of this incarnation, but the failure is not
+ * swallowed: a spawn whose record did not commit is a child that no later
+ * incarnation can discover, which is the exact hole this ledger exists to
+ * close. {@link ContainedSpawner} therefore refuses such a spawn, and the
+ * reaper leaves a record it could not retire in place so a later incarnation
+ * tries again.
+ *
+ * A ledger built with no journal ({@link makeMemory}) never fails: it promises
+ * nothing durable in the first place.
+ *
+ * @category services
+ * @since 1.0.0-rc.0
+ */
+export interface Service {
+  /** Records a started process and returns the durable record written for it. */
+  readonly record: (spawned: Spawned) => Effect.Effect<ProcessRecord, JournalError>
+  /** Records that a process ended, and drops it from {@link Service.live}. */
+  readonly release: (record: ProcessRecord) => Effect.Effect<void, JournalError>
+  /** Records that an abandoned process group was signalled by this host. */
+  readonly reaped: (record: ProcessRecord) => Effect.Effect<void, JournalError>
+  /**
+   * Retires a record this host refused to signal, naming why. The record
+   * leaves {@link Service.orphans} without anything having been killed.
+   */
+  readonly skipped: (record: ProcessRecord, reason: string) => Effect.Effect<void, JournalError>
+  /** The processes this incarnation started and has not yet released. */
+  readonly live: Effect.Effect<ReadonlyArray<ProcessRecord>>
+  /**
+   * The processes a previous incarnation of this host left running.
+   *
+   * Fails when that history cannot be read to its end, so an unreadable
+   * history is never mistaken for one with nothing left running.
+   */
+  readonly orphans: Effect.Effect<ReadonlyArray<ProcessRecord>, ProcessLedgerReplayError>
+}
+
+/**
+ * The process-ledger service tag.
+ *
+ * @category services
+ * @since 1.0.0-rc.0
+ */
+export class ProcessLedger extends Context.Service<ProcessLedger, Service>()("@smthrs/kernel/ProcessLedger") {}
+
+/**
+ * Ledger configuration.
+ *
+ * `hostId` is the durable identity two incarnations share; `ownerPid` is the
+ * process id of THIS incarnation, which is how a replay tells its own live
+ * processes from an abandoned one's.
+ *
+ * @category models
+ * @since 1.0.0-rc.0
+ */
+export interface Options {
+  readonly hostId: string
+  readonly ownerPid: number
+}
+
+/** Where records are appended, and where a previous incarnation's are read. */
+interface Sink {
+  readonly append: (
+    eventType: string,
+    record: ProcessRecord,
+    extra?: Record<string, unknown>
+  ) => Effect.Effect<void, JournalError>
+  readonly replay: Effect.Effect<ReadonlyArray<ProcessRecord>, ProcessLedgerReplayError>
+}
+
+/** The event types that retire a record from the inherited set. */
+const retires = (eventType: string): boolean =>
+  eventType === ExitedEventType || eventType === ReapedEventType || eventType === SkippedEventType
+
+/** A pid may be reused before the previous process's scope releases its record. */
+const identity = (record: ProcessRecord): string => `${record.pid}:${record.startedAtMs}`
+
+/**
+ * One entry folded onto the live set.
+ *
+ * Returns the entry when it is one of the ledger's own records but its payload
+ * does not decode, so replay can report it. Such an entry names no process
+ * that could be signalled, so failing the whole replay over it would stop
+ * every later incarnation from reaping the well-formed records beside it.
+ */
+const apply = (live: Map<string, ProcessRecord>, entry: JournalEvent.Entry): JournalEvent.Entry | undefined => {
+  if (entry.sourceId !== sourceId) return undefined
+  if (entry.eventType !== SpawnedEventType && !retires(entry.eventType)) return undefined
+  const decoded = decodeRecord(entry.payload)
+  if (decoded._tag === "Failure") return entry
+  if (entry.eventType === SpawnedEventType) {
+    live.set(identity(decoded.success), decoded.success)
+    return
+  }
+  live.delete(identity(decoded.success))
+  return undefined
+}
+
+const PAGE = 256
+
+const journalSink = (options: Options, journal: JournalModule.Service): Sink => ({
+  append: (eventType, record, extra) =>
+    journal.emitDurableUnfenced(
+      new JournalEvent.Input({
+        runId: hostRunId(options.hostId),
+        sourceId: sourceId as JournalEvent.SourceId,
+        eventType,
+        payload: extra === undefined ? encodeRecord(record) : { ...encodeRecord(record), ...extra }
+      })
+    ).pipe(
+      // The failure is REPORTED, not logged away. A swallowed write leaves a
+      // child that no incarnation of this host can discover, which is the one
+      // outcome containment exists to prevent; the caller decides what to do
+      // about it, and every caller in this repository declines to pretend the
+      // process was recorded.
+      Effect.tapCause((cause) => Effect.logWarning(`process ledger could not journal ${eventType}`, cause)),
+      Effect.asVoid
+    ),
+  replay: Effect.gen(function*() {
+    const live = new Map<string, ProcessRecord>()
+    let after: JournalEvent.Seq | undefined
+    for (;;) {
+      const page = yield* journal.entries({
+        runId: hostRunId(options.hostId),
+        limit: PAGE,
+        ...(after === undefined ? {} : { after })
+      }).pipe(
+        Effect.mapError((cause) =>
+          new ProcessLedgerReplayError({
+            code: "journal_unreadable",
+            message: `process ledger could not read the history of host ${options.hostId}`,
+            cause
+          })
+        )
+      )
+      for (const entry of page.entries) {
+        const malformed = apply(live, entry)
+        if (malformed !== undefined) {
+          yield* Effect.logWarning(
+            `process ledger skipped an undecodable ${malformed.eventType} record at seq ${malformed.seq}`
+          )
+        }
+      }
+      const last = page.entries.at(-1)
+      if (last === undefined || !page.hasMore) break
+      // A page that ends where the previous one did would be replayed
+      // forever, and stopping there would report a partial history as whole.
+      if (after !== undefined && last.seq <= after) {
+        return yield* Effect.fail(
+          new ProcessLedgerReplayError({
+            code: "cursor_stalled",
+            message: `process ledger history of host ${options.hostId} stopped advancing at seq ${last.seq}`
+          })
+        )
+      }
+      after = last.seq
+    }
+    return [...live.values()]
+  })
+})
+
+const memorySink: Sink = {
+  append: () => Effect.void,
+  replay: Effect.succeed([])
+}
+
+const makeWith = (options: Options, sink: Sink): Service => {
+  const live = new Map<string, ProcessRecord>()
+  return ProcessLedger.of({
+    record: (spawned) =>
+      Effect.gen(function*() {
+        const startedAtMs = yield* Effect.clockWith((clock) => clock.currentTimeMillis)
+        const record: ProcessRecord = {
+          pid: spawned.pid,
+          pgid: spawned.pgid,
+          hostId: options.hostId,
+          ownerPid: options.ownerPid,
+          startedAtMs,
+          commandDigest: spawned.commandDigest
+        }
+        live.set(identity(record), record)
+        // A failed append tells the caller the spawn was not recorded, and the
+        // caller kills the child. Keeping it live would lie to every later
+        // reader about a process this incarnation still holds.
+        yield* sink.append(SpawnedEventType, record).pipe(
+          Effect.tapCause(() => Effect.sync(() => live.delete(identity(record))))
+        )
+        return record
+      }),
+    release: (record) =>
+      Effect.suspend(() => {
+        live.delete(identity(record))
+        return sink.append(ExitedEventType, record)
+      }),
+    reaped: (record) => sink.append(ReapedEventType, record),
+    skipped: (record, reason) => sink.append(SkippedEventType, record, { reason }),
+    live: Effect.sync(() => [...live.values()]),
+    orphans: Effect.map(sink.replay, (records) => records.filter((record) => record.ownerPid !== options.ownerPid))
+  })
+}
+
+/**
+ * Builds a journal-backed ledger.
+ *
+ * @category constructors
+ * @since 1.0.0-rc.0
+ */
+export const make = (options: Options): Effect.Effect<Service, never, JournalModule.Journal> =>
+  Effect.map(JournalModule.Journal, (journal) => makeWith(options, journalSink(options, journal)))
+
+/**
+ * Builds a ledger that records nothing durably.
+ *
+ * A host without a journal still contains the processes it holds handles for —
+ * scope closure does that — but it inherits nothing from a previous
+ * incarnation, so {@link Service.orphans} is always empty.
+ *
+ * @category constructors
+ * @since 1.0.0-rc.0
+ */
+export const makeMemory = (options: Options): Effect.Effect<Service> => Effect.sync(() => makeWith(options, memorySink))
+
+/**
+ * Provides a journal-backed ledger.
+ *
+ * @category layers
+ * @since 1.0.0-rc.0
+ */
+export const layer = (options: Options): Layer.Layer<ProcessLedger, never, JournalModule.Journal> =>
+  Layer.effect(ProcessLedger, make(options))
+
+/**
+ * Provides a ledger that records nothing durably.
+ *
+ * @category layers
+ * @since 1.0.0-rc.0
+ */
+export const layerMemory = (options: Options): Layer.Layer<ProcessLedger> =>
+  Layer.effect(ProcessLedger, makeMemory(options))