Harden MCP transport and result handling

.papercuts/troubleshooting.md

@@ -1,5 +1,10 @@
 # Troubleshooting
 
+## 2026-09-12 — Listed upstream integration audit
+
+- This worktree has no `.memory/` or `node_modules/`. Read the main checkout's project memory as historical context, but use this worktree's exact HEAD/source as authority. The main checkout's `tsx` binary can execute dependency-free focused suites without installing packages here; suites with runtime package imports still fail module resolution (observed: `entities` in the subagent capability suite). Treat that as an environment limitation, not a product regression or passing test.
+- A repository in the earlier compaction reference table is not necessarily an installed integration. Verify runtime imports, vendored artifacts, implementation history, and explicit adoption decisions before proposing package upgrades.
+
 - `.papercuts/` is ignored even when its troubleshooting file is present in the PR branch, so persisting a required update needs an explicit `git add -f`.
 - Layout stabilization must race `animation.finished` against a short timeout because paused or infinite document animations never settle; keep geometry polling as the authoritative E2E readiness check.
 - Pi 0.80.10 can choose the oldest oversized user turn as `firstKeptEntryId`, leaving both summary inputs empty and producing a no-op checkpoint. When the journal has a newer turn, retry `prepareCompaction` with a minimal retained-tail budget; still refuse the checkpoint if both summary inputs remain empty.
@@ -429,3 +434,9 @@ symlink with this checkout's own npm ci. Full type-check and lint then passed.
 
 - E2E chat-title expectations assume the deterministic chat-model route. On a Mac where the native Foundation Models helper reports `ready`, automatic titles come from Apple Intelligence instead, so `chat-message-queue` sidebar-title lookups fail locally while passing in CI; probe the helper or move it aside before treating those failures as regressions.
 - `git add` on the tracked-but-ignored `.papercuts/troubleshooting.md` still needs `-f` after conflict resolution.
+
+## 2026-09-12 — MCP maintenance implementation
+
+- MCP SDK1.30.0 closes HTTP transports during OAuth redirection but expects finishAuth to reuse the same object and discovered metadata. Restart only its exchange request lifetime, retaining owner cancellation.
+- SDK OAuth metadata GETs also send MCP-Protocol-Version; a protocol header alone does not identify a timed MCP RPC. Classify actual request semantics and test the real SDK helpers. SSE per-frame limits apply only to successful requested streams, never arbitrary MIME-labeled JSON/error bodies.
+- Physical-iPhone native verification found three failures in unchanged RemoteClient fixture tests (catalog expectation, invalid JSON __SwiftValue, and legacy fallback invalidResponse). See the maintenance acceptance record; Android52 passed, iOS198 passed/3 skipped/3 failed.

docs/plans/README.md

@@ -32,7 +32,7 @@ This directory is the source of truth for Aiden's implementation plans. The engi
 | [rpiv-advisor integration](rpiv-advisor-integration-plan.md)                                       | Implemented | A bounded, tool-free, provider/auth-aware second opinion now uses an ephemeral per-consultation Ask User Question choice when the prompt does not name a reviewer, with no persistent Advisor settings or IPC. |
 | [rpiv-todo Integration](rpiv-todo-integration-plan.md)                                             | Partial | Attended desktop chats have a journal-replayed native todo tool, strict fail-closed snapshots, owner-fenced IPC, and a self-hiding floating progress chip with portal details; packaged visual/accessibility acceptance remains open. |
 | [rpiv-btw Integration](rpiv-btw-integration-plan.md)                                               | Partial | Attended desktop chats have bounded read-only side questions, ephemeral fingerprinted follow-ups, foreground-safe admission, exact provider dispatch, content-free usage accounting, and a native slash/card surface; packaged visual/accessibility acceptance remains open. |
-| [Subagent Orchestration Expansion](subagent-orchestration-expansion-plan.md)                       | Active  | Phases 0–6, Phase 7A durable lifecycle, the Phase 7B1 storage seam, reboot-safe V1/V2 migration checkpoints, startup/clone-boundary repairs, Pi-normalized foreground presentation/recovery, parent-only iOS/Android projection hardening, and the 2026-09-01 approval/capability/budget failure remediation are complete; app-lifetime coordinator activation is next.       |
+| [Subagent Orchestration Expansion](subagent-orchestration-expansion-plan.md)                       | Active  | Phases 0–6, Phase 7A durable lifecycle, the Phase 7B1 storage seam, reboot-safe V1/V2 migration checkpoints, startup/clone-boundary repairs, Pi-normalized foreground presentation/recovery, parent-only iOS/Android projection hardening, and the 2026-09-01 approval/capability/budget failure remediation are complete; foreground read-tool prompt truthfulness and refreshed background settlement gates are complete; app-lifetime coordinator activation is next.       |
 | [Taracodlab Learnings](taracodlab-learnings-plan.md)                                               | Partial | Phases A–B and D, plus core Phase E, are implemented; the remaining roadmap is open.                                                                                                                                      |
 | [Telegram First-Class Agent Parity](telegram-first-class-agent-parity-plan.md)                     | Active  | Controls, compaction, skills, rich inbound, drafts/activity, buttons, documents, Settings, and onboarding are green; Threaded Mode/profiles/extensions/TTS/live smoke remain.                                                               |
 | [Update, Microphone, and Computer Use Hardening](update-microphone-computer-use-hardening-plan.md) | Partial | Installed acceptance found a silent stalled download; observable full-download recovery is implemented, while repaired-build → next-release and clean-TCC acceptance remain.                                              |

docs/plans/subagent-orchestration-expansion-plan.md

@@ -746,6 +746,36 @@ Focused coordinator tests passed 3/3 with TypeScript clean.
 
 ## Verification matrix
 
+### 2026-09-12 upstream refresh: required before background activation
+
+The original August 5 design baseline remains historical. The latest comparison
+is `nicobailon/pi-subagents@940406c0d48890060d1f5f5280925c8f5a5c7389`
+(v0.67.0 plus 23 commits); see
+[the integrated-upstream audit](../testing/integrated-upstream-audit-2026-09-12.md).
+This refresh does not activate the coordinator or change the Pi runtime pin.
+
+Production activation must demonstrate all of the following in crash-point
+tests and the packaged nested-run soak:
+
+- Durable acceptance makes the run queryable before child allocation begins.
+- Result commit precedes terminal status publication and completion notification.
+- Steering distinguishes queued, consumed, and unconsumed-at-shutdown outcomes;
+  enqueue success alone must never be reported as consumption.
+- Descendants and owned tool/process work drain before ancestor completion and
+  before inference capacity is released. Exercise cancellation during a depth-two
+  handoff with one completed and one interrupted descendant.
+- Unresolved external effects remain `unknown`; incomplete inference remains
+  `interrupted`. Neither outcome causes an automatic effect retry on restart.
+- Corrupt status degrades to a bounded partial record without inventing completion.
+- Unchanged status does not repeatedly wake or notify the parent.
+- Desktop Remote, iOS, and Android preserve parent-only child projections.
+
+The foreground workspace-read prompt now follows the actual assembled tool
+intersection. An empty tool set remains valid for task/fork-only analysis; role
+names do not imply a missing capability or authorize widening access.
+
+### Existing final gates
+
 Every final gate includes:
 
 - strict TypeScript and lint;