feat(upstream): post-v0.9.4 upstream catch-up reliability batch

.pylon/features.yaml

@@ -527,3 +527,192 @@ decisions:
     upstream_support: Current upstream allocates a second timestamp when persisting custom messages, so reloading or compacting changes live transcript identities. Preserve the original message timestamp at every AgentSession persistence boundary.
     revisit_when:
       - Upstream preserves exact custom-message identities across live emission, persistence and reconstruction.
+
+  kernel-stderr-stream-errors:
+    area: kernel
+    state: candidate
+    owner: upstream
+    decision: adopt
+    pylon_refs:
+      - https://github.com/pylon-code/prime-agent/issues/1
+      - https://github.com/pylon-code/pylon/issues/114
+    upstream_refs:
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2383
+    fork_change: absorb-kernel-child-stderr-stream-errors
+    upstream_support: Absorb child stderr stream errors to prevent worker crashes on broken pipe or process death without altering process ownership.
+    revisit_when:
+      - Upstream redesigns kernel stdio streams.
+
+  repl-bounded-frames:
+    area: kernel
+    state: candidate
+    owner: upstream
+    decision: adopt
+    pylon_refs:
+      - https://github.com/pylon-code/prime-agent/issues/1
+    upstream_refs:
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2423
+    fork_change: bound-repl-protocol-frame-sizes
+    upstream_support: Bound REPL message sizes to prevent memory exhaustion and protocol desync.
+    revisit_when:
+      - Kernel REPL migrates to chunked framing.
+
+  owner-only-kernel-stderr-log:
+    area: kernel-security
+    state: candidate
+    owner: upstream
+    decision: adopt
+    pylon_refs:
+      - https://github.com/pylon-code/prime-agent/issues/1
+    upstream_refs:
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2425
+    fork_change: owner-only-kernel-stderr-log-permissions
+    upstream_support: Set mode 0o600 on kernel stderr log files.
+    revisit_when:
+      - Kernel logging subsystem is replaced.
+
+  heartbeats-list-bounded-live-state:
+    area: daemon-health
+    state: candidate
+    owner: upstream
+    decision: adopt
+    pylon_refs:
+      - https://github.com/pylon-code/prime-agent/issues/1
+    upstream_refs:
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2378
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2342
+    fork_change: heartbeats-list-from-live-state-with-deadline
+    upstream_support: Read active heartbeats from in-memory state with 10s deadline without queueing behind busy worker turns; must not fabricate liveness for quarantined sessions.
+    revisit_when:
+      - Heartbeat storage moves out-of-process.
+
+  session-open-restart-wait:
+    area: daemon-lifecycle
+    state: candidate
+    owner: upstream
+    decision: adopt
+    pylon_refs:
+      - https://github.com/pylon-code/prime-agent/issues/1
+    upstream_refs:
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2391
+    fork_change: session-open-wait-during-daemon-restart
+    upstream_support: Let session opens wait through daemon update restart instead of failing immediately; re-verifies daemon hello and supervisor generation on connect.
+    revisit_when:
+      - Daemon lifecycle management changes.
+
+  spawn-name-reservation-admission:
+    area: session-lifecycle
+    state: candidate
+    owner: upstream
+    decision: adopt
+    pylon_refs:
+      - https://github.com/pylon-code/prime-agent/issues/1
+    upstream_refs:
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2396
+    fork_change: hold-spawn-name-reservations-until-admission
+    upstream_support: Hold spawn name reservations until admission is durable, respecting caller-owned admission proofs and reconnect handles.
+    revisit_when:
+      - Session admission protocol is redesigned.
+
+  async-bash-notice-turn-boundary:
+    area: runtime-reliability
+    state: candidate
+    owner: shared
+    decision: hybridize
+    pylon_refs:
+      - https://github.com/pylon-code/prime-agent/issues/64
+      - https://github.com/pylon-code/prime-agent/issues/1
+    upstream_refs:
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2372
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2386
+    fork_change: async-bash-notice-re-parking-with-handle-withdrawal
+    upstream_support: Hybridize turn-boundary notice fixes with fork #72 per-handle withdrawal and single terminal settlement.
+    revisit_when:
+      - Upstream introduces handle-bound async command receipts.
+
+  compaction-tool-identity-and-anchoring:
+    area: context-compaction
+    state: candidate
+    owner: upstream
+    decision: adopt
+    pylon_refs:
+      - https://github.com/pylon-code/prime-agent/issues/1
+    upstream_refs:
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2424
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2385
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2463
+    fork_change: compaction-tool-identity-anchoring-and-validation
+    upstream_support: Preserve tool identity in summarizer input, anchor summaries to kept tail, validate refinement writes, and preserve custom message timestamps without altering cursor order.
+    revisit_when:
+      - Compaction architecture or prompt format changes.
+
+  codex-stale-chain-recovery:
+    area: provider-codex
+    state: candidate
+    owner: upstream
+    decision: adopt
+    pylon_refs:
+      - https://github.com/pylon-code/prime-agent/issues/1
+    upstream_refs:
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2374
+    fork_change: codex-stale-chain-recovery-after-metadata
+    upstream_support: Recover stale Codex chains after metadata while preserving connection binding.
+    revisit_when:
+      - Codex completion API changes.
+
+  prime-inference-glm-no-thinking:
+    area: provider-glm
+    state: candidate
+    owner: upstream
+    decision: adopt
+    pylon_refs:
+      - https://github.com/pylon-code/prime-agent/issues/1
+    upstream_refs:
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2459
+    fork_change: prime-inference-glm-disable-thinking
+    upstream_support: Stop sending enable_thinking to Prime Inference GLM models.
+    revisit_when:
+      - GLM adds reasoning support.
+
+  abort-and-send-queued:
+    area: session-input
+    state: candidate
+    owner: shared
+    decision: hybridize
+    pylon_refs:
+      - https://github.com/pylon-code/pylon/issues/114
+    upstream_refs:
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2426
+    fork_change: capability-gated-abort-and-send-queued-schema34
+    upstream_support: Renumber schema revision above fork 33 to 34 and gate behind abort_and_send_queued_v1 capability.
+    revisit_when:
+      - Upstream negotiates queued aborts via capability.
+
+  quota-park-and-auto-resume:
+    area: provider-quota
+    state: deferred
+    owner: shared
+    decision: redesign
+    pylon_refs:
+      - https://github.com/pylon-code/pylon/issues/114
+    upstream_refs:
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2375
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2280
+    fork_change: none
+    upstream_support: Quota waiting and auto-resume daemon behavior. Pending maintainer decision.
+    revisit_when:
+      - Maintainer guides quota park and auto-resume projection.
+
+  mcp-service-catalog-oauth:
+    area: mcp-ownership
+    state: deferred
+    owner: shared
+    decision: retain
+    pylon_refs:
+      - https://github.com/pylon-code/pylon/issues/114
+    upstream_refs:
+      - https://github.com/PrimeIntellect-ai/prime-agent/pull/2330
+    fork_change: none
+    upstream_support: Daemon MCP service catalog and OAuth engine. Retaining Pylon MCP ownership pending maintainer decision.
+    revisit_when:
+      - Maintainer guidance on MCP service catalog.

.pylon/upstream-review.md

@@ -389,3 +389,41 @@ The SDK token gates the new local API; daemon wire shapes and stock behavior do
 not change. Validation: focused malformed/private/path/registration cases and a real
 isolated supervisor/owned-worker test prove registered before cleanup and settled
 after exact owned cleanup and supervisor exit. Pylon delivery tracked in [pylon#557](https://github.com/pylon-code/pylon/issues/557).
+
+## 2026-09-22 — post-v0.9.4 upstream review (cfb2d3fe3..18a56bf35 and open Sept 14 triage)
+
+- Review range: upstream commits `cfb2d3fe35c7c23cf6678ec235412b0d3f286fae..18a56bf3557063c0cd0be12f99f0bcce2a46da32` (72 commits, including release v0.9.5 `a7d791bc1`) plus open Sept 14 triage items.
+- Merge-base and complete integration checkpoint remain frozen at v0.9.4 `1eee2938b4eeb7a4d72e17035adda669a89b63de`. Selective adoption does not advance the full integration checkpoint.
+- Fork product base: `72899e4e4fdaef52cfffca58152f769a6aef91b1`. Tracking: #1 and Pylon #114.
+
+### Dispositions and Revisit Triggers
+
+| PR / Commit | Subject / Area | Disposition | Notes & Revisit Triggers |
+| --- | --- | --- | --- |
+| #2383 (`b6b94c78a`) | Kernel child stderr stream error absorption | **Adopt** | Prevents worker crashes on unhandled child stderr errors (EPIPE). Does not touch process ownership, isolation, or cleanup receipts. Revisit if upstream redesigns kernel stdio streams. |
+| #2423 (`c91e6e991`) | Bound REPL protocol frame sizes in kernel and host | **Adopt** | Guards against memory exhaustion and protocol desync by bounding message frames. Revisit if kernel REPL switches to streaming chunked wire format. |
+| #2425 (`b6d955a9b`) | Keep kernel stderr log owner-only (mode 0o600) | **Adopt** | Security fix ensuring private process logs are not world-readable. Revisit if kernel logging subsystem is replaced. |
+| #2378 (`3dfb0276b`) + #2342 (`ad426c672`) | Heartbeats list from live state with 10s deadline | **Adopt** | Reads active heartbeats from in-memory state without blocking behind serialized session worker queues. Must not fabricate liveness for terminating or quarantined sessions. Revisit if heartbeat storage moves out-of-process. |
+| #2391 (`54da82e97`) | Session opens wait through daemon update restart | **Adopt** | Retries and waits for restart completion instead of immediately throwing connection errors. Wait must re-verify daemon hello, schema revision, and supervisor generation upon reconnect to prevent attaching across unauthorized generation boundaries. Revisit if daemon process manager lifecycle changes. |
+| #2396 (`126fe7010`) | Hold spawn name reservations until admission is durable | **Adopt** | Prevents concurrent duplicate session creation collision in the registry. Must respect caller-owned admission proofs and not reject reconnects or recovery handles under the same session ID. Revisit if session creation protocol is redesigned. |
+| #2372 (`66df06fc1`) + #2386 (`b3e04b58b`) | Async-bash completion notices and re-parking | **Hybridize** | Overlaps fork #72 (acknowledged per-handle background notice withdrawal). Preserves fork #72 random handle identities, host-acknowledged withdrawal, and single terminal turn settlement. Revisit if upstream introduces handle-bound async receipts. |
+| #2424 (`5fb3d9acb`) | Compaction: keep tool identity in serialized summarizer input | **Adopt** | Retains tool name and call parameters in summarizer prompt. Preserves custom message timestamps (#76) and does not alter event cursor ordering. Revisit if compaction serialization schema changes. |
+| #2385 (`27f32ddb7`) | Compaction: anchor summaries to kept-tail state and stop re-summarizing file lists | **Adopt** | Fixes repeated file listings and anchors summaries to recent state. Does not alter transcript disk format or break replay verification. Revisit if compaction prompt architecture is modified. |
+| #2463 (`41e4e0e18`) | Validate refinement writes and skip malformed entries in digest | **Adopt** | Bounds digest errors on corrupted refinement records. Logs malformed entries without corrupting the digest hash or causing divergence between replicas. Revisit if digest format evolves. |
+| #2374 (`e683fcb53`) | AI: recover stale Codex chains after metadata | **Adopt** | Follow-up to adopted #2314. Retains producing connection binding and one-shot retry before output. Revisit if Codex completion API changes. |
+| #2459 (`63d88319b`) | AI: stop sending enable_thinking to Prime Inference GLM models | **Adopt** | Fixes provider rejection for GLM models. Revisit if Prime Inference GLM adds reasoning support. |
+| Perf: #2414, #2416, #2433, #2389, #2393, #2399, #2379, #2387, #2405 | In-memory caching, frame decoding, kernel startup perf | **Adopt** | Pure performance improvements that do not mutate snapshot/catalog ownership boundaries or wire protocols. #2433 must not alter catalog disk schema, serialized format, or bypass immutable snapshot boundaries. Revisit if underlying subsystems are overhauled. |
+| Excluded: #2398, #2296, #2273, #2260, #2242, #2246, #2276 | On-demand catalog, transcript download cache, persistent SupervisorLink, orphan GC, session append | **Retain (Deferred)** | These touch snapshot/catalog/supervisor/ownership boundaries. Fork preserves immutable bounded snapshots, attachment generations, per-event recovery cursors, authenticated ownership, and cleanup receipts. Revisit in separate review with real-daemon and managed-artifact integration gates. |
+| #2426 (`e2fb7bfa1`) | Send queued messages after interrupt (`abort_and_send_queued`) | **Hybridize / Redesign** | Daemon wire change: renumber schema above fork 33 to schema 34 and gate behind `abort_and_send_queued_v1` capability. Pylon negotiates capability and falls back to interrupt-then-send. Revisit when upstream capability-gates queued aborts. |
+| #2282 (`ebe530d28`) | Child progress notes (`rlm_progress_note`) and child snapshots (`rlm.collect`) | **Redesign (Pylon issue)** | Stock 0.9.5 events are safely ignored in Pylon via `Ignored/unknown-event`. Pylon issue tracks Agents surface integration. Revisit when Agents UI exposes child progress. |
+| #2453 (`3b1aa5ff3`) | Image-model routing | **Redesign (Pylon issue)** | Dynamic model switching impacts model identity, usage attribution, and pricing. Pylon issue tracks provider routing design. Revisit with multimodal routing architecture. |
+| #2455 (`b09e37d84`) | ACP model and effort pickers | **Adopt (Pylon issue)** | Exposes model/effort options for ACP fallback. Pylon issue tracks UI pickers. Revisit during ACP settings review. |
+| #2334 (`dca77ecfe`) + #2465 (`b08f08efa`) | Human message priority and continuation holds while bash runs | **Redesign (Pylon issue)** | Evaluated against Pylon queue projection and occurrence ownership. Pylon issue tracks queue prioritization semantics. Revisit with turn scheduling audit. |
+| Stock 0.9.5 compiled binary without Node SDK | Runtime detection and fallback | **Retain (Pylon issue)** | Stock 0.9.5 installer defaults to Bun-compiled binary omitting Node SDK. Pylon issue tracks fallback to ACP with actionable error guidance. |
+| #2375 (`2883a7843`) + #2280 | Quota park and auto-resume (`providerBackupModel`) | **Maintainer Decision** | Daemon parks quota-blocked sessions until reset and auto-resumes. Options: disable for Pylon-owned sessions, or project unsolicited resumed turns as first-class runtime state. Pending maintainer decision. |
+| #2330 (`b6ac5d014`) | MCP service catalog and OAuth engine | **Maintainer Decision** | Overlaps Pylon-owned MCP server management. Retain Pylon MCP ownership pending maintainer confirmation. |
+| Context/Refinement: #2217, #2226, #2220, #2394, #2400, #2450, #2392 | Context summarization and harness ranking | **Adopt / Hybridize** | Improves compaction accuracy and prompt-cache stability. #2226 already hybridized in fork #71. Revisit if context architecture changes. |
+| Models/CLI: #2505, #2462, #2381, #2410, #2409, #2216, #2380, #2515, #2336 | Catalog updates, CLI diagnostics, test repairs | **Adopt** | Additive model options (Grok 4.7), diagnostics (`PI_REQUEST_TIMING`), flag validations, and test flakiness fixes. Revisit when affected providers change. |
+| TUI / CLI UI / CI: #2446, #2447, #2430, #2466, #2340, #2331, #2328, #2324, #2325, #2301, #2302, #2306, #2449, #2376, #2456, #2323 | Terminal UI interactions, upstream CI/benchmarks, install docs | **Skip** | Prime-specific TUI screens, terminal commands, and upstream internal CI are not consumed by Pylon web/desktop/mobile clients. No daemon protocol or SDK impact. Revisit if features move to daemon protocol. |
+| #2458 (`2e9ab77b8`) | Reconnect attached windows when daemon restarts | **Skip / Retain** | TUI client window reconnect is skipped because Pylon manages its own reconnection; supervisor update-restart phase closing reason is retained/deferred with supervisor lifecycle audit. |
+| Distribution: #2366, #2370, #2369, #2368, #2361, #2344, #2319 | Native binaries, updater manifests, musl packages | **Retain** | Retain Pylon recipe 2 deterministic SDK archives per fork #75 decision. Revisit if upstream provides importable Node SDK in binary archives. |

packages/ai/.changes/fix-codex-chain-reset-metadata.md

@@ -0,0 +1 @@
+- Fixed Codex stale-continuation recovery when WebSocket lifecycle or metadata events arrive before the provider error.

packages/ai/.changes/prime-inference-glm-thinking-compat.md

@@ -0,0 +1 @@
+- Fixed Prime Inference GLM models failing every request with an unsupported `enable_thinking` parameter.

packages/ai/scripts/generate-models.ts

@@ -117,6 +117,13 @@ const ZAI_THINKING_COMPAT: OpenAICompletionsCompat = {
 	thinkingFormat: "zai",
 };
 
+// Prime Inference rejects `enable_thinking` on GLM routes with a 400 on every
+// request. The routes think by default, so send no thinking parameter (and no
+// reasoning_effort, whose support is unverified).
+const PRIME_INFERENCE_ZAI_COMPAT: OpenAICompletionsCompat = {
+	supportsReasoningEffort: false,
+};
+
 const PRIME_INFERENCE_BASE_URL = "https://api.pinference.ai/api/v1";
 const PRIME_INFERENCE_COMPAT: OpenAICompletionsCompat = {
 	supportsStore: false,
@@ -452,6 +459,10 @@ function isPrimeInferenceReasoningModel(modelId: string, catalogReasoning?: bool
 	);
 }
 
+function isPrimeInferenceZaiModelId(modelId: string): boolean {
+	return modelId.toLowerCase().startsWith("z-ai/glm-");
+}
+
 function getPrimeInferenceCompat(modelId: string): OpenAICompletionsCompat {
 	const id = modelId.toLowerCase();
 	if (id.includes("deepseek-v4")) {
@@ -460,10 +471,10 @@ function getPrimeInferenceCompat(modelId: string): OpenAICompletionsCompat {
 			...DEEPSEEK_V4_COMPAT,
 		};
 	}
-	if (id.startsWith("z-ai/glm-")) {
+	if (isPrimeInferenceZaiModelId(id)) {
 		return {
 			...PRIME_INFERENCE_COMPAT,
-			...ZAI_THINKING_COMPAT,
+			...PRIME_INFERENCE_ZAI_COMPAT,
 		};
 	}
 
@@ -610,7 +621,11 @@ function createPrimeInferenceModel(
 			...(openRouter?.supportsReasoningEffort === false
 				? {
 						supportsReasoningEffort: false,
-						...(!compat.thinkingFormat ? { thinkingFormat: "openrouter" as const } : {}),
+						// GLM routes must not receive any thinking parameter; the
+						// `reasoning` object is not verified against the gateway.
+						...(!compat.thinkingFormat && !isPrimeInferenceZaiModelId(entry.id)
+							? { thinkingFormat: "openrouter" as const }
+							: {}),
 					}
 				: {}),
 		},