@cat-factory/executor-harness 1.145.1 → 1.149.0

package/src/docker-status.ts

@@ -1,6 +1,6 @@
-import { execFile } from 'node:child_process'
 import { readFile } from 'node:fs/promises'
-import { promisify } from 'node:util'
+import { probeDockerWorkload, type DockerWorkload } from './docker-capability.js'
+import { log, type Logger } from './logger.js'
 
 // What this container knows about its own Docker daemon, as recorded by `entrypoint.sh`.
 //
@@ -14,8 +14,12 @@ import { promisify } from 'node:util'
 // daemon reads it instead.
 //
 // The recorded verdict describes BOOT, and a container outlives its boot, so nothing refuses on
-// it unconfirmed: `resolveDockerVerdict` re-checks a recorded absence against a live daemon and
-// keeps the record for what only the record holds, the cause and the daemon's own log tail.
+// it unconfirmed: `resolveDockerVerdict` re-checks it against a live daemon and keeps the record
+// for what only the record holds, the cause and the daemon's own log tail. What it records is
+// also only that a SOCKET answered, which is a weaker fact than any caller wants, so the live
+// check RUNS A CONTAINER (docker-capability.ts) rather than settling for the daemon's word about
+// itself. It still reports that weaker fact alongside, since a check that could not be carried
+// out is what the boot record has to be read against, and nothing else establishes it.
 //
 // The three-valued shape is deliberate and is the point (CLAUDE.md, "Degrade loudly"): a daemon
 // that FAILED and a daemon nobody asked about are different facts with different correct
@@ -55,7 +59,20 @@ export type DockerSource =
 export interface DockerStatus {
   available: boolean | undefined
   source: DockerSource
-  /** Why, in the entrypoint's own closed vocabulary (`serving`/`failed`/`missing`/…). */
+  /**
+   * Why, in the entrypoint's own closed vocabulary (`serving`, `serving-without-nat`,
+   * `still-starting`, `failed`, `missing`, `unreachable`, `probing`).
+   *
+   * Reported and never branched on, which is what lets the entrypoint add a word without anything
+   * here having to know it. Two are worth knowing about. `serving-without-nat`: the daemon that
+   * manages its own firewall rules exited without serving, so the one that came up runs with
+   * `--iptables=false` and its NESTED containers have no egress. That is a CAUSE, and the only
+   * place one exists; what MEASURES the consequence is the egress half of the workload check
+   * (docker-capability.ts), from inside a container, with no way to learn why. `still-starting`
+   * is a daemon that had not answered when the boot budget ran out and is STILL RUNNING, which is
+   * the one absence that routinely stops being true: it is exactly what the live re-probe below
+   * exists to catch.
+   */
   reason: string
   /** A human detail for the failing cases: the dockerd log tail, or what was unreachable. */
   detail?: string
@@ -146,56 +163,149 @@ function unnamedSource(source: never): string {
   return `no Docker daemon answered in this container (unrecognised source ${JSON.stringify(source)})`
 }
 
-/** Whether a daemon is answering RIGHT NOW. Injected so the unit suite can state either answer. */
-export type DockerProbe = () => Promise<boolean>
-
-/** A live probe may not outlast the thing it is guarding; a hung socket is an absent daemon here. */
-const PROBE_TIMEOUT_MS = 10_000
+/**
+ * What a daemon can do RIGHT NOW. Injected so the unit suite can state every answer.
+ *
+ * It answers with a WORKLOAD rather than with a boolean, and that is the correction this type
+ * carries. It used to be `docker version`, which proves the daemon is serving; a stand-up needs
+ * a daemon that can materialise an image, and a sandboxed rootless daemon routinely serves while
+ * being unable to (issue #2120). Running compose against that one produced a mount error the
+ * agent had to interpret, from the one mechanism whose entire job is to say why infra did not
+ * come up.
+ *
+ * The weaker fact did not go away, though: it rides `daemonAnswered` on the `unknown` arm, since
+ * the workload check establishes it on its way past and {@link resolveDockerVerdict} still needs
+ * it. Takes the job's signal, because this is a live check on the critical path of a run that can
+ * be cancelled under it.
+ */
+export type DockerProbe = (signal?: AbortSignal) => Promise<DockerWorkload>
 
-const execFileAsync = promisify(execFile)
+/**
+ * The default {@link DockerProbe}: the process-wide workload probe, which loads a one-layer
+ * image and runs a container from it, memoised per container.
+ *
+ * Named for what it answers. It was `probeDockerServing`, which is the fact this module exists to
+ * say is not enough.
+ */
+export const probeLiveDockerCapability: DockerProbe = (signal) => probeDockerWorkload(signal)
 
 /**
- * The default {@link DockerProbe}: `docker version` talks to the SERVER, unlike the client-only
- * `docker --version`, which answers happily with no daemon at all.
+ * The sentence for a daemon that is serving and cannot run anything. It names what was tried,
+ * because "docker is unavailable" against a daemon the agent can see answering reads as a bug in
+ * the platform rather than as the sandbox limit it is.
  */
-export const probeDockerServing: DockerProbe = async () => {
-  try {
-    await execFileAsync('docker', ['version', '--format', '{{.Server.Version}}'], {
-      timeout: PROBE_TIMEOUT_MS,
-    })
-    return true
-  } catch {
-    return false
-  }
+export function describeDockerUnusable(workload: { detail: string }): string {
+  return `this container's Docker daemon is reachable but cannot run a container (${workload.detail})`
 }
 
 /** What a stand-up is entitled to conclude about the daemon at the moment it is about to run. */
 export interface DockerVerdict {
-  /** Three-valued exactly as {@link DockerStatus.available}, and read the same way. */
+  /**
+   * Whether a stand-up may PROCEED, three-valued exactly as {@link DockerStatus.available} and
+   * read the same way. It is the decision, not a description of the daemon: a daemon that is
+   * answering and cannot run a container is `false` here and `daemon: true` below, and a record
+   * that reported the first as the second would send an operator to restart a daemon that is
+   * already up.
+   */
   available: boolean | undefined
-  /** Set only for a CONFIRMED absence: the sentence to refuse with. Absent means proceed. */
+  /**
+   * Set only for a CONFIRMED negative: the sentence to refuse with. Absent means proceed.
+   *
+   * Two causes reach it and they read differently on purpose. Nothing is answering here, and
+   * something is answering here but cannot run a container: an operator sent to restart a daemon
+   * that is already up would find nothing wrong with it.
+   */
   refusal?: string
+  /**
+   * Whether a daemon ANSWERED the live check. Absent when nothing was checked (an undecided
+   * record probes nothing) or when nothing answered, so it is never read as a decided `false`.
+   */
+  daemon?: boolean
+  /** What a real container did on it, when one was tried. Absent when nothing was measured. */
+  workload?: DockerWorkload
 }
 
 /**
- * Resolve what to do now, from what boot recorded plus what a daemon says today.
+ * Resolve what to do now, from what boot recorded plus what the daemon can do today.
  *
- * `entrypoint.sh` probes once, at boot, within a bounded wait. A container outlives that: a warm
- * pool serves many jobs from one, and a sidecar daemon that took longer than the wait allows is
- * serving perfectly well by the second job. Refusing off the recorded verdict alone latches that
- * container into refusing local infra that in fact works, for its whole life, with a stale
- * sentence explaining why. So a recorded absence is a HYPOTHESIS here, and the live probe settles
- * it; the recorded verdict is still what supplies the cause and the daemon's own log tail, which
- * no probe can reconstruct.
+ * `entrypoint.sh` probes once, at boot, within a bounded wait, and it probes for a SOCKET. Two
+ * things follow, and the branches below are one each.
  *
- * Only a recorded `false` is re-confirmed. "Not decided" keeps attempting exactly as before: the
- * point of the third value is that nothing turns it into a refusal, and a probe here would.
+ * A recorded absence is a HYPOTHESIS. A container outlives its boot: a warm pool serves many jobs
+ * from one, and a sidecar daemon that took longer than the wait allows is serving perfectly well
+ * by the second job. Refusing off the record alone latches that container into refusing local
+ * infra that works, for its whole life, with a stale sentence explaining why. The record is still
+ * what supplies the cause and the daemon's own log tail, which no probe can reconstruct.
+ *
+ * A recorded PRESENCE is a hypothesis too, and that half was missing. `serving` is not `usable`:
+ * a rootless daemon in a sandbox answers while unable to mount any image layer, so compose ran
+ * and died on a mount error the agent then had to interpret. So the probe is consulted in both
+ * directions, and it runs a real container rather than asking the daemon about itself.
+ *
+ * A check that could not be CARRIED OUT settles nothing, and the cheap fact is what decides
+ * there. Falling straight back to the boot record would re-latch the very refusal the paragraph
+ * above rules out: the four ways the workload check can come back undeterminable (no payload in
+ * this image variant, an architecture it is not built for, a `docker load` the engine refuses, a
+ * timeout) have nothing to do with whether a daemon is up, so a warm container whose sidecar
+ * arrived late would be denied local infra for the rest of its life over a stale sentence. So a
+ * daemon that ANSWERED contradicts a recorded absence exactly as the old `docker version` probe
+ * did, and only a check that never reached a daemon at all leaves the record to decide.
+ *
+ * "Not decided" still keeps attempting, untouched. The point of the third value is that NOTHING
+ * turns it into a refusal: the entrypoint's bounded wait may still be running, and a workload
+ * probe against a daemon that has not finished starting fails for a reason that says nothing
+ * about what it will do a second later.
  */
 export async function resolveDockerVerdict(
   status: DockerStatus,
-  probe: DockerProbe = probeDockerServing,
+  opts: { probe?: DockerProbe; signal?: AbortSignal; logger?: Logger } = {},
 ): Promise<DockerVerdict> {
-  if (status.available !== false) return { available: status.available }
-  if (await probe()) return { available: true }
-  return { available: false, refusal: describeDockerAbsence(status) }
+  if (status.available === undefined) return { available: undefined }
+  const workload = await askTotally(
+    opts.probe ?? probeLiveDockerCapability,
+    opts.signal,
+    opts.logger,
+  )
+  if (workload.status === 'usable') return { available: true, daemon: true, workload }
+  if (workload.status === 'unusable') {
+    return {
+      available: false,
+      refusal: describeDockerUnusable(workload),
+      daemon: true,
+      workload,
+    }
+  }
+  if (workload.daemonAnswered) return { available: true, daemon: true, workload }
+  return status.available
+    ? { available: true, workload }
+    : { available: false, refusal: describeDockerAbsence(status), workload }
+}
+
+/**
+ * Call the probe and answer even if it throws.
+ *
+ * The default probe is total by construction and says so, but this is the seam an injected one
+ * arrives through, and the caller is a stand-up documented as best-effort: a throw here would
+ * fail a job over the mechanism whose whole purpose is to make a failure legible. A throw settles
+ * nothing about the daemon, so it becomes the same value as any other check that could not be
+ * carried out and the boot record decides, exactly as it did before the probe existed.
+ */
+async function askTotally(
+  probe: DockerProbe,
+  signal: AbortSignal | undefined,
+  logger: Logger | undefined,
+): Promise<DockerWorkload> {
+  try {
+    return await probe(signal)
+  } catch (err) {
+    const message = err instanceof Error ? err.message : String(err)
+    ;(logger ?? log).warn('docker: the live daemon check threw; falling back to the boot record', {
+      error: message,
+    })
+    return {
+      status: 'unknown',
+      reason: `the live Docker check could not be carried out (${message})`,
+      daemonAnswered: false,
+    }
+  }
 }

package/src/environment-inventory.ts

@@ -2,6 +2,11 @@ import { execFile } from 'node:child_process'
 import { promisify } from 'node:util'
 import { log, type Logger } from './logger.js'
 import { harnessListenPort } from './harness-port.js'
+import {
+  probeDockerWorkload,
+  type ContainerEgress,
+  type DockerWorkload,
+} from './docker-capability.js'
 
 // ---------------------------------------------------------------------------
 // What this machine actually has, probed ONCE per job and stated to the agent.
@@ -32,6 +37,18 @@ import { harnessListenPort } from './harness-port.js'
 //     a daemon this machine is CONFIGURED for but which has not answered yet is a fourth state
 //     that resolves to `unknown`, never to the absence a refused connection looks like: the
 //     image's daemon is started in the background and the job begins before it is ready.
+//   - A daemon that ANSWERS is still not a daemon that WORKS, which is the same mistake one level
+//     in. A rootless daemon nested in a sandbox serves while its snapshotter cannot mount any
+//     image layer, so `docker info` succeeds and `docker build` / `docker run` / `docker pull`
+//     all fail (issue #2120). Only a container that RAN settles that, so the reachable case is
+//     split by a real workload (docker-capability.ts) into `usable`, `unusable`, and a daemon
+//     that answered while the check itself could not be carried out.
+//   - And a daemon that runs containers is still not one whose containers have a NETWORK, which
+//     is the same mistake one level in again. Loading and running a local image needs no network,
+//     so a daemon started with `--iptables=false` passes the workload check while every nested
+//     container is cut off (issue #2174): what an agent then hits is a `docker build` whose
+//     `RUN npm ci` sits in retry backoff for about seven minutes before failing. So `usable`
+//     carries its own egress verdict, and the line below says something different for each.
 //
 // Deliberately NOT here: the agent's own tools (web search, file tools, MCP servers). Those are
 // the CLI's, they differ per harness, and each is already stated where it is true. Claiming one
@@ -68,6 +85,33 @@ export type ToolPresence =
   | { status: 'absent' }
   | { status: 'unknown'; reason: string }
 
+/**
+ * What this machine's Docker daemon is good for, which is FIVE answers and not three.
+ *
+ * `absent` and `unknown` mean for the daemon what they mean for any other tool. The three that
+ * are particular to Docker split the case where a daemon ANSWERED, because answering is not the
+ * question anyone is asking:
+ *
+ *   - `usable`:   a container was built and run on it here. `docker build` / `run` / `compose`
+ *                   work, and this is the only state that may say so.
+ *   - `unusable`: a container could NOT be run, with the daemon serving throughout. The state
+ *                   issue #2120 is about; stated as a prohibition, with the cause.
+ *   - `serving`:  it answered, and the workload check could not be carried out (no payload on
+ *                   this machine, an unmapped architecture, a timeout). Neither of the other two,
+ *                   and rendered as "try it if you need it".
+ *
+ * `usable` then carries what a nested container could REACH, which is a second fact and not a
+ * sixth state: what an agent may do with the daemon and what its containers can fetch are
+ * different questions, and each of the three egress answers changes the advice without changing
+ * the verdict on the daemon.
+ */
+export type DockerCapability =
+  | { status: 'usable'; server?: string; egress: ContainerEgress }
+  | { status: 'unusable'; server?: string; detail: string }
+  | { status: 'serving'; server?: string; reason: string }
+  | { status: 'absent' }
+  | { status: 'unknown'; reason: string }
+
 /** One probed entry: the name the agent would type, and what came back. */
 export interface ProbedTool {
   name: string
@@ -80,12 +124,13 @@ export interface ProbedTool {
 export interface EnvironmentInventory {
   tools: ProbedTool[]
   /**
-   * Whether a Docker daemon actually answered: the readiness fact, not the CLI's presence. The
-   * image's `entrypoint.sh` starts a rootless daemon BEST-EFFORT and execs the server without
-   * waiting for it, so at job start this probe is the only thing that knows how that went, and
-   * "has not answered yet" is one of its answers (see {@link probeDockerDaemon}).
+   * What the Docker daemon is good for: not the CLI's presence, and not merely whether the
+   * daemon answered. The image's `entrypoint.sh` starts a rootless daemon BEST-EFFORT and execs
+   * the server without waiting for it, so at job start this probe is the only thing that knows
+   * how that went; "has not answered yet" is one of its answers (see {@link probeDockerDaemon})
+   * and "answered, but cannot run a container" is another (see {@link DockerCapability}).
    */
-  dockerDaemon: ToolPresence
+  dockerDaemon: DockerCapability
   /**
    * The port the harness's own job server holds in this network namespace. Not probed: the
    * process reads its own {@link harnessListenPort}, which is the only honest answer when a
@@ -310,6 +355,18 @@ export interface ProbeEnvironmentOptions {
    * only so the suite can assert the rendered line without an ambient `PORT` deciding its text.
    */
   harnessPort?: number
+  /**
+   * Whether the daemon can actually RUN a container, defaulting to the process-wide probe
+   * (docker-capability.ts). Asked only once a daemon has answered, since there is nothing to run
+   * a workload on otherwise, and memoised per container so a warm pool pays for it once.
+   */
+  workload?: (signal?: AbortSignal) => Promise<DockerWorkload>
+  /**
+   * The job's signal, forwarded to the probes that spawn something. The workload check starts a
+   * CONTAINER, so a cancelled job must stop paying for it rather than hold the daemon for the
+   * rest of its budget.
+   */
+  signal?: AbortSignal
 }
 
 /**
@@ -346,11 +403,35 @@ export async function probeEnvironment(
         presence: toolPresence(await run(probe.command, probe.args)),
       })),
     ),
-    probeDockerDaemon(run, opts),
+    probeDockerCapability(run, opts),
   ])
   return { tools, dockerDaemon, harnessPort: opts.harnessPort ?? harnessListenPort() }
 }
 
+/**
+ * The daemon's full answer: whether one is reachable, and then whether it can run a container.
+ *
+ * The two steps are kept apart because they fail for unrelated reasons and only the FIRST has a
+ * cheap answer. A daemon nobody can reach has no workload to run, so the check that costs a
+ * container start is asked only where there is something to ask it of; a daemon that answered
+ * carries its server version into every one of the three states that follow it, because the
+ * agent reading the line is entitled to know which daemon the verdict is about.
+ */
+async function probeDockerCapability(
+  run: ProbeRunner,
+  opts: ProbeEnvironmentOptions,
+): Promise<DockerCapability> {
+  const daemon = await probeDockerDaemon(run, opts)
+  if (daemon.status === 'absent') return { status: 'absent' }
+  if (daemon.status === 'unknown') return { status: 'unknown', reason: daemon.reason }
+  const server = daemon.version ? { server: daemon.version } : {}
+  const workload = await (opts.workload ?? probeDockerWorkload)(opts.signal)
+  if (workload.status === 'usable') return { status: 'usable', ...server, egress: workload.egress }
+  if (workload.status === 'unusable')
+    return { status: 'unusable', ...server, detail: workload.detail }
+  return { status: 'serving', ...server, reason: workload.reason }
+}
+
 /**
  * Ask the daemon itself, and do not mistake a daemon that is STARTING for one that is not there.
  *
@@ -482,26 +563,124 @@ function harnessPortLine(port: number): string {
   )
 }
 
-/** The Docker line, which says something different in each of the three cases. */
-function dockerDaemonLine(daemon: ToolPresence): string {
-  if (daemon.status === 'present') {
-    const server = daemon.version ? ` (server ${daemon.version})` : ''
-    return (
-      `A Docker daemon is reachable${server}: \`docker build\`, \`docker run\` and ` +
-      '`docker compose up` work here.'
-    )
+/**
+ * The Docker line, which says something different in each of the five cases, and is TOTAL over
+ * them: adding a state without deciding what an agent should do about it stops the build.
+ *
+ * Only `usable` may claim the commands work, and it may only be reached by having RUN one. The
+ * line that used to stand here made that claim off `docker info` alone, which is how every agent
+ * in a run was told, as fact, that a daemon which could not mount a single image layer would
+ * build and run one.
+ */
+function dockerDaemonLine(daemon: DockerCapability): string {
+  const server = 'server' in daemon && daemon.server ? ` (server ${daemon.server})` : ''
+  switch (daemon.status) {
+    case 'usable':
+      return `A Docker daemon is reachable${server} and the platform ran a container on it: ${usableCommands(daemon.egress)} work here. ${egressSentence(daemon.egress)}`
+    case 'unusable':
+      return (
+        `A Docker daemon is reachable${server} but it CANNOT run a container: the platform ` +
+        `built a one-layer image and tried to run it here, and that failed (${daemon.detail}). ` +
+        '`docker build`, `docker run`, `docker pull` of a multi-layer image and ' +
+        '`docker compose up` all fail for the same reason, so there is nothing to retry and no ' +
+        'flag that works around it. Produce the Dockerfile or compose file you were asked for, ' +
+        'say in one line that it could not be built or run here, and move on.'
+      )
+    case 'serving':
+      return (
+        `A Docker daemon is reachable${server}, but whether it can actually build or run an ` +
+        `image was NOT established (${daemon.reason}). Reaching the daemon is not the same fact: ` +
+        'a sandboxed one answers while being unable to mount any image layer. Try it if you need ' +
+        'it, and do not read a failure as a defect in the work.'
+      )
+    case 'unknown':
+      return (
+        'Whether a Docker daemon is reachable could not be determined ' +
+        `(${daemon.reason}): try it if you need it, and do not read a failure as a defect in the work.`
+      )
+    case 'absent':
+      return (
+        'NO Docker daemon is reachable: `docker build`, `docker run` and `docker compose up` ' +
+        'will fail here whatever the CLI reports. Produce the Dockerfile or compose file you ' +
+        'were asked for, say in one line that you could not build it here, and move on.'
+      )
+    default:
+      return unnamedCapability(daemon)
   }
-  if (daemon.status === 'unknown') {
-    return (
-      'Whether a Docker daemon is reachable could not be determined ' +
-      `(${daemon.reason}): try it if you need it, and do not read a failure as a defect in the work.`
-    )
+}
+
+function unnamedCapability(daemon: never): string {
+  return `Whether a Docker daemon is reachable could not be determined (the platform reported an unrecognised verdict ${JSON.stringify(daemon)}): try it if you need it.`
+}
+
+/**
+ * Which commands the `usable` line may claim, which is the EGRESS verdict's business and not the
+ * daemon's.
+ *
+ * Split out because the line used to open with the full list and then, one sentence later, tell
+ * the agent that every `RUN` line which fetches anything fails. A block that also says not to
+ * spend turns re-checking it cannot afford to state and then retract the same fact: an agent
+ * reading the first sentence has already been told `docker build` works here, which is the exact
+ * shape of the lie issue #2174 is about. TOTAL over {@link ContainerEgress}, like its sibling.
+ */
+function usableCommands(egress: ContainerEgress): string {
+  switch (egress.status) {
+    case 'blocked':
+      // Deliberately omits `docker build`: it is the one the missing NAT rule actually breaks,
+      // and the sentence that follows explains which part of it and why.
+      return '`docker run` and `docker compose up` of images that are already built'
+    case 'reachable':
+    case 'undetermined':
+      return '`docker build`, `docker run` and `docker compose up`'
+    default:
+      return unnamedEgressCommands(egress)
   }
-  return (
-    'NO Docker daemon is reachable: `docker build`, `docker run` and `docker compose up` will ' +
-    'fail here whatever the CLI reports. Produce the Dockerfile or compose file you were asked ' +
-    'for, say in one line that you could not build it here, and move on.'
-  )
+}
+
+/** The commands claimed for an egress verdict this build does not know: none of them. */
+function unnamedEgressCommands(egress: never): string {
+  return `the docker commands the platform could name for its verdict ${JSON.stringify(egress)}`
+}
+
+/**
+ * The second half of the `usable` line: what a container started HERE can reach, and what to do
+ * about it. TOTAL over {@link ContainerEgress} for the same reason the line above is over the
+ * daemon's states.
+ *
+ * The `blocked` arm is the one this exists for, and it is precise about WHICH commands break,
+ * because "docker has no network" is not true and an agent that believed it would skip work it
+ * could have done. The daemon has a network: it pulls base images and compose pulls its services
+ * normally. What has none is the container each `RUN` line executes in. The seven minutes are
+ * named because the failure does not look like a failure from inside the agent's loop: npm turns
+ * "no route" into `EAI_AGAIN` only once its retry backoff gives up, so the build reads as a hang
+ * and the natural response is to wait longer.
+ */
+function egressSentence(egress: ContainerEgress): string {
+  switch (egress.status) {
+    case 'reachable':
+      return 'A container started here also reaches the network, so a build that installs dependencies works.'
+    case 'blocked':
+      return (
+        'A container started here could reach NOTHING the platform tried, and that is a fact ' +
+        `about this sandbox rather than about your work (${egress.detail}). The daemon itself is ` +
+        'fine: it pulls base images, and `docker compose up` of pre-built images works. What ' +
+        'fails is every `RUN` line in a `docker build` that fetches from the public internet ' +
+        '(`npm ci`, `apk add`, `pip install`), and it fails SLOWLY: npm reports `EAI_AGAIN` only ' +
+        'after some seven minutes of retry backoff, so it reads as a hang. Do not wait it out ' +
+        'and do not retry. Vendor what you need, skip the image build, verify some other way, or ' +
+        'say in one line that you could not verify it here. If this project already builds ' +
+        'against a mirror inside this network, that is not one of the addresses tried above and ' +
+        'is worth one attempt.'
+      )
+    case 'undetermined':
+      return `Whether a container started here can reach the network was NOT established (${egress.reason}), so try it if you need it and do not read a failure as a defect in the work.`
+    default:
+      return unnamedEgress(egress)
+  }
+}
+
+function unnamedEgress(egress: never): string {
+  return `Whether a container started here can reach the network could not be determined (the platform reported an unrecognised verdict ${JSON.stringify(egress)}).`
 }
 
 /**
@@ -520,21 +699,28 @@ function dockerDaemonLine(daemon: ToolPresence): string {
  */
 export async function appendEnvironmentInventory(
   systemPrompt: string,
-  opts: { signal?: AbortSignal; log?: Logger; run?: ProbeRunner } & ProbeEnvironmentOptions = {},
+  opts: { log?: Logger; run?: ProbeRunner } & ProbeEnvironmentOptions = {},
 ): Promise<string> {
   const logger = opts.log ?? log
+  // Everything that is not this function's OWN is forwarded by construction, rather than key by
+  // key. The list of copied keys silently dropped `workload`, whose whole point is that a suite
+  // can inject one: a test driving THIS entry point (the only one `handleAgent` uses) got the
+  // real probe instead, which starts a container on whatever machine the suite runs on.
+  const { log: _log, run, ...probeOptions } = opts
   try {
-    const inventory = await probeEnvironment(opts.run ?? spawnProbeRunner(opts.signal), {
-      ...(opts.sleep ? { sleep: opts.sleep } : {}),
-      ...(opts.daemonExpected === undefined ? {} : { daemonExpected: opts.daemonExpected }),
-      ...(opts.harnessPort === undefined ? {} : { harnessPort: opts.harnessPort }),
-    })
+    const inventory = await probeEnvironment(run ?? spawnProbeRunner(opts.signal), probeOptions)
     logger.info('agent: probed the environment', {
       installed: inventory.tools
         .filter((t) => t.presence.status === 'present')
         .map((t) => t.name)
         .join(','),
       dockerDaemon: inventory.dockerDaemon.status,
+      // Present only when a nested container was actually asked, which is the one state that has
+      // an egress answer at all. A word here for every other state would report "no measurement"
+      // and "measured, and it cannot get out" in the same field.
+      ...(inventory.dockerDaemon.status === 'usable'
+        ? { dockerEgress: inventory.dockerDaemon.egress.status }
+        : {}),
       // The unknowns, by NAME: the block tells the agent a probe failed, and this is the only place
       // an operator can see WHICH, since the reason the agent reads is deliberately wordy prose.
       unknown: inventory.tools

package/src/harness-server.ts

@@ -6,6 +6,7 @@ import { handleAgent } from './agent.js'
 import { handleInline } from './inline.js'
 import { redactSecrets } from './git.js'
 import { readDockerStatus } from './docker-status.js'
+import { reportedDockerWorkload } from './docker-capability.js'
 import { harnessListenPort } from './harness-port.js'
 import { JobRegistry, loadRunnerLimits, type JobResultBase, type RunOptions } from './runner.js'
 import { log } from './logger.js'
@@ -141,11 +142,16 @@ const server = createServer((req, res) => {
       // would spawn a process per poll to answer a question this endpoint is not the one to act
       // on; the stand-up re-confirms a recorded absence at the moment it matters
       // (`resolveDockerVerdict`), so a stale negative here never becomes a stale refusal there.
+      //
+      // `workload` is the other half, and the reason the block used to mislead: what the record
+      // says is `serving`, and serving is not usable. It reports the last measurement any job
+      // took (docker-capability.ts) and NEVER takes one itself, for the same polling reason,
+      // which is why `unmeasured` is one of the words it can answer.
       return send(res, 200, {
         status: 'ok',
         ...(HARNESS_VERSION ? { version: HARNESS_VERSION } : {}),
         capabilities: HARNESS_BODY_CAPABILITIES,
-        docker: await readDockerStatus(),
+        docker: { ...(await readDockerStatus()), workload: reportedDockerWorkload() },
       })
     }
     // All non-health endpoints are gated by the optional shared secret.