release: PilotSwarm 0.5.18

Ship model-visible image attachments across the portal, TUI, Web API,
SDK, and MCP server. Images paste/drop/pick into the portal composer
(desktop + mobile), Ctrl+V-paste into the TUI from the OS clipboard,
and ride sendMessage as artifact refs; workers fetch bytes, gate on the
model's vision capability, and deliver true multimodal blob attachments.
Orchestration 1.0.65 (1.0.64 frozen). Also: statusVersion-based stale-
status guard (fixes wedged 'Working..'), actionable 4xx artifact errors,
8 MB ingress body limit, configurable worker turn timeout, and the
durable-signals design proposal.

Author: affandar

.env.example

@@ -49,6 +49,7 @@ DATABASE_URL=postgresql://postgres:postgres@localhost:5432/durable_copilot
 # PILOTSWARM_FACTS_PG_POOL_MAX=3
 # PILOTSWARM_ORCHESTRATION_CONCURRENCY=2
 # PILOTSWARM_WORKER_CONCURRENCY=2
+# PILOTSWARM_TURN_TIMEOUT_MS=1200000
 
 # ── Optional provider overlays ──────────────────────────────────────────────
 # The default facts provider is plain Postgres (PgFactStore) and is configured

CHANGELOG.md

@@ -1,5 +1,66 @@
 # Changelog
 
+## 0.5.18 — 2026-07-21
+
+### Model-visible image attachments across every surface
+
+Operators can now show images to the model: paste (Ctrl/Cmd+V), drag-drop,
+or pick images in the portal composer — desktop and mobile — attach from the
+TUI with Ctrl+V (OS clipboard) or Ctrl+A (file path), or send attachment refs
+through the Web API, SDK, and MCP server. Images persist as session binary
+artifacts (upload-first, reference-after); vision-capable models receive them
+as true multimodal blob attachments with per-model capability gating. New
+sessions use orchestration `1.0.65`.
+
+### SDK
+
+- `sendMessage`/`send` accept `attachments: [{filename}]` refs, validated at
+  the API edge against the session's artifact store (png/jpeg/gif/webp, 4 MB
+  per image, 4 images / 8 MB per message) and carried on the durable queue as
+  resolved refs — bytes never enter orchestration history.
+- Orchestration `1.0.65` threads attachment refs through drain, merge, and
+  continue-as-new (`1.0.64` frozen); the runTurn activity fetches bytes from
+  the artifact store, gates on the live model catalog's vision capability
+  (`SessionManager.getModelVisionInfo`), and forwards blob attachments to the
+  Copilot session. Drops are explicit: prompt omission notes plus
+  `runtime.attachment_dropped` events.
+- Attachment refs ride the `user.message` event for transcript rendering.
+
+### Portal
+
+- Composer image intake: clipboard paste, drag-drop (with drop highlight),
+  and a paperclip picker (`accept=image/*` offers camera/photo library on
+  phones); staged images render as thumbnail chips with size and remove.
+- Transcript user messages render authenticated thumbnail strips; the Files
+  pane previews raster image artifacts inline.
+- Stale-status guard now trusts the server's monotonic `statusVersion`,
+  fixing sessions wedged on "Working.." after turn completion.
+- Artifact/attachment validation errors return actionable 4xx codes
+  (`ARTIFACT_CONTENT_TYPE_MISMATCH`, `ARTIFACT_TOO_LARGE`,
+  `INVALID_ATTACHMENT`) instead of generic 500s.
+- Ingress body limit raised to 8 MB for image uploads (deploy manifest).
+
+### TUI
+
+- Ctrl+V pastes an image from the OS clipboard (pngpaste/osascript on macOS,
+  wl-paste/xclip on Linux, PowerShell on Windows) and stages it on the next
+  message; Ctrl+A path uploads auto-stage raster images the same way.
+
+### MCP
+
+- `send_message` and `send_and_wait` accept `attachments` (web mode);
+  `get_capabilities` reports `prompt.imageAttachments`.
+
+### Worker
+
+- Turn timeout is deployment-configurable.
+
+### Docs
+
+- `docs/proposals/image-attachments-in-chat.md` rewritten to match the
+  shipped architecture; new draft proposal
+  `docs/proposals/durable-signals-and-webhooks.md` (`wait_for_signal`).
+
 ## 0.5.17 — 2026-07-20
 
 ### Reliable finite delegation and cleaner shared UI rendering

README.md

@@ -2,7 +2,7 @@
 
 > **Experimental** — This project is under active development and not yet ready for production use. APIs may change without notice.
 
-> **Latest release: v0.5.17** — Finite delegated work now wakes parents on substantive results, child contracts survive durable session creation, compact GFM table delimiters render correctly, and portal chat cards use the full pane width. New sessions use orchestration `1.0.64`.
+> **Latest release: v0.5.18** — Model-visible image attachments everywhere: paste, drop, or pick images in the portal (desktop + mobile), Ctrl+V clipboard paste in the TUI, and attachment refs through the Web API, SDK, and MCP server — persisted as session artifacts and vision-gated per model. New sessions use orchestration `1.0.65`.
 
 A durable execution runtime for [GitHub Copilot SDK](https://github.com/github/copilot-sdk) agents. Crash recovery, durable timers, session dehydration, and multi-node scaling — powered by [duroxide](https://github.com/microsoft/duroxide). Just add a connection string.
 

deploy/envs/template.env

@@ -169,6 +169,7 @@ INGRESS_CLASS=azure-application-gateway
 # own K8s secret keeps the connection-string flow intact.
 PILOTSWARM_USE_MANAGED_IDENTITY=1
 AZURE_STORAGE_CONTAINER=copilot-sessions
+PILOTSWARM_TURN_TIMEOUT_MS=1200000
 
 # ─── Azure AI Foundry (optional) ───
 # When true, base-infra provisions a Microsoft.CognitiveServices/accounts

deploy/gitops/worker/overlays/default/.env

@@ -25,6 +25,8 @@ AZURE_TENANT_ID=00000000-0000-0000-0000-000000000000
 # custom Azure OpenAI / Anthropic-direct endpoint can add a stamp-local
 # kustomize patch that adds them to the worker-env ConfigMap.
 AZURE_STORAGE_CONTAINER=copilot-sessions
+# Wall-clock cap for one Copilot turn. 0 disables the cap.
+PILOTSWARM_TURN_TIMEOUT_MS=1200000
 # Storage managed-identity flag + account URL (consumed by
 # packages/sdk/src/blob-store.ts → createSessionBlobStore). Setting the
 # flag to "1" tells the worker to use DefaultAzureCredential (federated

deploy/k8s/portal-ingress.yaml

@@ -6,6 +6,10 @@ metadata:
   annotations:
     cert-manager.io/cluster-issuer: "letsencrypt-prod"
     nginx.ingress.kubernetes.io/ssl-redirect: "true"
+    # Image attachment uploads are base64 JSON up to ~5.4 MB (4 MB decoded
+    # cap) on the artifact PUT route; nginx's 1 MB default rejects them with
+    # 413 before the app's own 8 MB route limit can apply.
+    nginx.ingress.kubernetes.io/proxy-body-size: "8m"
     nginx.ingress.kubernetes.io/proxy-read-timeout: "3600"
     nginx.ingress.kubernetes.io/proxy-send-timeout: "3600"
     nginx.ingress.kubernetes.io/proxy-http-version: "1.1"

deploy/k8s/worker-deployment.yaml

@@ -60,6 +60,8 @@ spec:
               value: "4"
             - name: PILOTSWARM_WORKER_CONCURRENCY
               value: "8"
+            - name: PILOTSWARM_TURN_TIMEOUT_MS
+              value: "1200000"
           envFrom:
             - secretRef:
                 name: copilot-runtime-secrets

deploy/scripts/test/foundry-substitute.test.mjs

@@ -47,6 +47,7 @@ test("__FOUNDRY_ENDPOINT__ in model_providers.json is substituted from FOUNDRY_E
       WORKER_IMAGE_TAG: "test",
       IMAGE: "acrtest.azurecr.io/pilotswarm-worker:test",
       PILOTSWARM_USE_MANAGED_IDENTITY: "1",
+      PILOTSWARM_TURN_TIMEOUT_MS: "1200000",
       PILOTSWARM_CMS_FACTS_DATABASE_URL: "postgres://test/cms",
       PILOTSWARM_DB_AAD_USER: "uami",
       LOCATION: "westus3",
@@ -99,6 +100,7 @@ test("__FOUNDRY_ENDPOINT__ stays unresolved when FOUNDRY_ENDPOINT is empty/unset
       WORKER_IMAGE_TAG: "test",
       IMAGE: "acrtest.azurecr.io/pilotswarm-worker:test",
       PILOTSWARM_USE_MANAGED_IDENTITY: "1",
+      PILOTSWARM_TURN_TIMEOUT_MS: "1200000",
       PILOTSWARM_CMS_FACTS_DATABASE_URL: "postgres://test/cms",
       PILOTSWARM_DB_AAD_USER: "uami",
       LOCATION: "westus3",

docs/developer/building/examples/blog-travel-event-scanner-sdk.md

@@ -680,7 +680,7 @@ const worker = new PilotSwarmWorker({
 
   // Performance tuning
   maxSessionsPerRuntime: 50,         // concurrent sessions per worker (default: 50)
-  turnTimeoutMs: 120_000,            // LLM turn timeout (default: 2 minutes)
+  turnTimeoutMs: 1_200_000,          // explicit LLM turn cap (default: 20 minutes)
   sessionIdleTimeoutMs: 3_600_000,   // idle session cleanup (default: 1 hour)
 
   // Management agents — auto-started system agents for monitoring and maintenance

docs/developer/deploy/aks.md

@@ -613,6 +613,7 @@ kubectl create secret generic copilot-runtime-secrets \
     --from-literal=PILOTSWARM_FACTS_PG_POOL_MAX="3" \
     --from-literal=PILOTSWARM_ORCHESTRATION_CONCURRENCY="2" \
     --from-literal=PILOTSWARM_WORKER_CONCURRENCY="2" \
+    --from-literal=PILOTSWARM_TURN_TIMEOUT_MS="1200000" \
     --from-literal=AZURE_STORAGE_CONNECTION_STRING="DefaultEndpointsProtocol=https;AccountName=..." \
     --from-literal=AZURE_STORAGE_CONTAINER="copilot-sessions"
 ```
@@ -624,6 +625,7 @@ Worker sizing is env-driven:
 - `PILOTSWARM_FACTS_PG_POOL_MAX` — facts `pg.Pool` max size. Default: `3`.
 - `PILOTSWARM_ORCHESTRATION_CONCURRENCY` — Duroxide orchestration concurrency. Default: `2`.
 - `PILOTSWARM_WORKER_CONCURRENCY` — Duroxide activity/worker concurrency. Default: `2`.
+- `PILOTSWARM_TURN_TIMEOUT_MS` — wall-clock cap for one Copilot turn. Default: `1200000` (20 minutes); `0` disables it.
 
 Provider availability in selectors is env-driven at worker startup. If you add or remove a provider key, refresh the secret and restart the workers; changing the checked-in template alone is not enough, and changing the real `.model_providers.json` only takes effect after the updated file is present in the runtime environment.
 
@@ -717,6 +719,7 @@ kubectl create secret generic copilot-runtime-secrets \
     --from-literal=PILOTSWARM_FACTS_PG_POOL_MAX="3" \
     --from-literal=PILOTSWARM_ORCHESTRATION_CONCURRENCY="2" \
     --from-literal=PILOTSWARM_WORKER_CONCURRENCY="2" \
+    --from-literal=PILOTSWARM_TURN_TIMEOUT_MS="1200000" \
     --from-literal=AZURE_STORAGE_CONNECTION_STRING="..." \
     --from-literal=AZURE_STORAGE_CONTAINER="copilot-sessions" \
     --dry-run=client -o yaml | kubectl apply -f -