Ship v0.1.1–v0.1.5: first-live-run fixes and dashboard viewer

- v0.1.1: config persistence — FileHelper paths made absolute
  (/media/startos/volumes/main/...); relative paths resolved into the JS
  runtime's ephemeral cwd so action saves never reached /data
- v0.1.2: preJobStopContainers (Configure Grading) — docker-stop resident
  vLLM containers on the head Spark at job start, no auto-restart
- v0.1.3: preflight auth (LiteLLM master_key gates /models),
  poll-until-loaded, crash fast-fail (restarting counts as dead)
- v0.1.4: HF_HUB_OFFLINE/TRANSFORMERS_OFFLINE in airgapped serving
  (--internal network has no DNS); grader _post timeout 600→1800s for
  ~3.6 tok/s GB10 generation
- v0.1.5: dashboard viewer survives the periodic background refresh;
  download buttons for deck reports, deck JSON, and SCORECARD.md
- .gitignore: .startos/ build workspace, start-technologies/

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

startos/actions/configure-grading.ts

@@ -84,6 +84,18 @@ const inputSpec = InputSpec.of({
       'land in the inbox. Off by default so you trigger grading explicitly.',
     default: false,
   }),
+  preJobStopContainers: Value.text({
+    name: 'Stop These Containers Before Grading (optional)',
+    description:
+      'Space- or comma-separated docker container names stopped on the HEAD ' +
+      'Spark at the start of every grading job, so the job gets the GPU to ' +
+      'itself (e.g. an always-on vLLM another service runs). They are NOT ' +
+      'restarted afterwards — the service that owns them reloads its own ' +
+      'models (e.g. the Gazette\'s Fleet job). Empty = stop nothing.',
+    required: false,
+    default: null,
+    placeholder: 'vllm-gemma4-clerk',
+  }),
   // --- Deterministic-scorer weights (composite = quant 60 + qual 40 - flags) ---
   profitabilityKpi: Value.number({
     name: 'Weight: Profitability KPIs',
@@ -204,6 +216,7 @@ export const configureGrading = sdk.Action.withInput(
       adjudicatorPersona: cfg.adjudicatorPersona || undefined,
       wipeRemoteDocs: cfg.wipeRemoteDocs,
       autoRunOnDrop: cfg.autoRunOnDrop,
+      preJobStopContainers: cfg.preJobStopContainers || undefined,
       profitabilityKpi: cfg.weights.profitabilityKpi,
       otherKpi: cfg.weights.otherKpi,
       forecastIntegrity: cfg.weights.forecastIntegrity,
@@ -228,6 +241,7 @@ export const configureGrading = sdk.Action.withInput(
       adjudicatorPersona: input.adjudicatorPersona ?? '',
       wipeRemoteDocs: input.wipeRemoteDocs,
       autoRunOnDrop: input.autoRunOnDrop,
+      preJobStopContainers: input.preJobStopContainers ?? '',
       weights: {
         profitabilityKpi: input.profitabilityKpi,
         otherKpi: input.otherKpi,

startos/file-models/config.ts

@@ -172,6 +172,10 @@ export const configShape = z.object({
   autoRunOnDrop: z.boolean().default(false),
   // Name of the per-job Docker network created on the head Spark.
   networkName: z.string().default('boardroom-net'),
+  // Space/comma-separated docker container names stopped on the head Spark at
+  // the start of every job to free GPU memory (e.g. an always-on vLLM another
+  // service runs). NOT restarted afterwards — their owner reloads them.
+  preJobStopContainers: z.string().default(''),
 
   // --- Portfolio companies ---
   // The authoritative source of pinned KPI targets and KPI-name aliases. Decks
@@ -212,4 +216,11 @@ export const configShape = z.object({
 
 export type Config = z.infer<typeof configShape>
 
-export const configFile = FileHelper.json('./config.json', configShape)
+// Absolute path into the `main` volume (mounted at /media/startos/volumes/main
+// in the JS runtime, /data in the orchestrator container). A relative path here
+// resolves against the JS runtime's EPHEMERAL working directory and silently
+// loses the config on restart.
+export const configFile = FileHelper.json(
+  '/media/startos/volumes/main/config.json',
+  configShape,
+)

startos/file-models/secrets.ts

@@ -5,9 +5,13 @@ import { FileHelper } from '@start9labs/start-sdk'
  * `main` volume. Kept out of config.json so it is never returned in plaintext
  * config reads. The container copies it to a 600 path at runtime.
  *
- * Volume path './ssh/id_spark' -> /data/ssh/id_spark inside the container.
+ * Volume path (absolute — a relative path would resolve against the JS
+ * runtime's ephemeral cwd, not the volume) -> /data/ssh/id_spark inside the
+ * orchestrator container.
  */
-export const sshKeyFile = FileHelper.string('./ssh/id_spark')
+export const sshKeyFile = FileHelper.string(
+  '/media/startos/volumes/main/ssh/id_spark',
+)
 
 /**
  * Optional Hugging Face token (for gated/private model pulls on the Sparks).
@@ -15,7 +19,7 @@ export const sshKeyFile = FileHelper.string('./ssh/id_spark')
  * network mode models are served from a pre-pulled cache, so this is only used
  * the first time you warm a model (or in local_services mode).
  *
- * Volume path './secrets/hf_token' -> /data/secrets/hf_token.
+ * Volume path (absolute, see above) -> /data/secrets/hf_token.
  *
  * NOTE: Boardroom Map has NO frontier/cloud key. There is deliberately no Anthropic
  * key and no Gitea token — the whole point is that confidential documents and

startos/versions/index.ts

@@ -1,8 +1,13 @@
 import { VersionGraph } from '@start9labs/start-sdk'
 import { v_0_1_0 } from './v_0_1_0'
+import { v_0_1_1 } from './v_0_1_1'
+import { v_0_1_2 } from './v_0_1_2'
+import { v_0_1_3 } from './v_0_1_3'
+import { v_0_1_4 } from './v_0_1_4'
+import { v_0_1_5 } from './v_0_1_5'
 
 /** The current version MUST be the first argument (`current`). */
 export const versions = VersionGraph.of({
-  current: v_0_1_0,
-  other: [],
+  current: v_0_1_5,
+  other: [v_0_1_4, v_0_1_3, v_0_1_2, v_0_1_1, v_0_1_0],
 })

startos/versions/v_0_1_1.ts

@@ -0,0 +1,16 @@
+import { VersionInfo } from '@start9labs/start-sdk'
+
+/**
+ * Packaging fix. ExVer form `<upstream>:<downstream>` — we track our own
+ * packaging revision since Boardroom Map has no separate upstream semver.
+ */
+export const v_0_1_1 = VersionInfo.of({
+  version: '0.1.1:0',
+  releaseNotes:
+    'Fix: persist config.json, the Spark SSH key, and the optional HF token ' +
+    'to the main volume (absolute volume paths in the file models). ' +
+    'Previously the Configure actions wrote to an ephemeral runtime directory, ' +
+    'so the orchestrator never saw /data/config.json and configuration was ' +
+    'lost on restart. Re-run the Configure actions after updating.',
+  migrations: {},
+})

startos/versions/v_0_1_2.ts

@@ -0,0 +1,15 @@
+import { VersionInfo } from '@start9labs/start-sdk'
+
+/**
+ * GPU co-residency support. ExVer form `<upstream>:<downstream>`.
+ */
+export const v_0_1_2 = VersionInfo.of({
+  version: '0.1.2:0',
+  releaseNotes:
+    'New "Stop These Containers Before Grading" setting (Configure Grading): ' +
+    'container names listed there are docker-stopped on the head Spark at the ' +
+    'start of every grading job, so the job gets the GPU to itself when ' +
+    'another service keeps an always-on vLLM resident. They are deliberately ' +
+    'not restarted afterwards — the owning service reloads its own models.',
+  migrations: {},
+})

startos/versions/v_0_1_3.ts

@@ -0,0 +1,13 @@
+import { VersionInfo } from '@start9labs/start-sdk'
+
+/** Serving-readiness fix. ExVer form `<upstream>:<downstream>`. */
+export const v_0_1_3 = VersionInfo.of({
+  version: '0.1.3:0',
+  releaseNotes:
+    'Fix wave preflight: the model-proxy probe now authenticates (the router ' +
+    'requires its master key even on /models) and both the proxy check and the ' +
+    'per-model completion check poll until ready instead of failing on a ' +
+    'single early attempt — a 31B model takes minutes to load into GPU ' +
+    'memory. Crashed containers fail fast with their log tail.',
+  migrations: {},
+})

startos/versions/v_0_1_4.ts

@@ -0,0 +1,13 @@
+import { VersionInfo } from '@start9labs/start-sdk'
+
+/** Air-gapped serving fix. ExVer form `<upstream>:<downstream>`. */
+export const v_0_1_4 = VersionInfo.of({
+  version: '0.1.4:0',
+  releaseNotes:
+    'Air-gapped mode now serves models with HF_HUB_OFFLINE/TRANSFORMERS_OFFLINE ' +
+    'set: the per-job --internal network has no DNS, so without offline mode ' +
+    'the HF hub client crashed on name resolution even with the model fully ' +
+    'cached. Preflight also treats a crash-looping (restarting) vLLM container ' +
+    'as dead and fails fast with its logs instead of waiting out the timeout.',
+  migrations: {},
+})

startos/versions/v_0_1_5.ts

@@ -0,0 +1,14 @@
+import { VersionInfo } from '@start9labs/start-sdk'
+
+/** Dashboard viewer fixes. ExVer form `<upstream>:<downstream>`. */
+export const v_0_1_5 = VersionInfo.of({
+  version: '0.1.5:0',
+  releaseNotes:
+    'Dashboard: an open deck report / JSON / scorecard viewer now stays open ' +
+    'across the periodic background refresh (previously it silently closed ' +
+    'within ~20s) — it closes only via its close button or when switching ' +
+    'companies. Added ⬇ download buttons: per-deck report downloads in the ' +
+    'deck-history table, plus a download link in the viewer header for ' +
+    'reports, deck JSON records, and SCORECARD.md.',
+  migrations: {},
+})