fix(managed-agent): #13163 follow-ups: cold-cache cancel and deferred review suggestions #13269

Follow-ups and delivery status for #13163, which addresses #13162 items 1–3. After multiple substantive review rounds, optional suggestions remain here under the AGENTS.md scope rule. The bounded cancellation fix and the remaining inherited ordering/recovery gaps are distinguished below.

1. Cold-cache cancellation: bounded fix delivered in #13163

The correction first delivered at 60030e4b0f9c and carried by merged head 52704a9de0c6 gives persisted, submitted cancellation its own attachment recovery, validates the frozen Session identity without mutable create grants, and passively reuses the original resident Harness connection. Independent H2 reproduction was red before the correction. The exact historical head 52704a9de0c6 passed the complete Java suite (562 entries: 561 executed passes, one Linux-only skip) and both macOS-executable real Hosted scenarios, including cancellation after Java cache loss with can_create revoked and the registry draining. This closes the cache-loss acceptance case that keeps the original Harness and Broker/worker alive. New API cancellation requests still require the creator's read grant; accepted cancellation continues retrying after subsequent read revocation.

After the required #13173 synchronization, parked passive recovery adopts the original Runtime and reads status without preparing/executing work. Repeated passive loads retain the owed cancellation lease until terminal success or teardown. The merged CLI session/recovery suites pass 205 tests; WebShell passes 61. The final merge preserves the concurrent automated update with all 1722 verification inputs unchanged. Exact original Runtime receiver identity is asserted in helper coverage; the route proves the serial repeated-load/cancel/release lifecycle. Concurrent load/cancel/close schedules remain optional follow-up coverage.

Previous synchronized delivery b27a5958a16e preserves concurrent cleanup f8e83fc7707e and main through 5ddfacc9d4c1. Build/typecheck and 91 focused synchronization tests passed; the 30 Java tests recompiled 94 production and 104 test sources, including the retained main non-Broker rename-retirement case. The current verification report separates these results from the earlier full suite.

Current-head SDK Java CI has now completed successfully. The run records PR head b27a5958a16e; both database jobs checked out merge 707910c9904b, whose parents are exactly this head and main 5ddfacc9d4c1. Managed Agent ran 561 unit tests and 52 non-Hosted database integration tests. The Hosted MySQL JUnit reports contain 18 executed scenarios across all seven expected classes, including all four public-Workspace scenarios; the updated cold-cache cancellation and separate generation/storage refusal assertions are in the executed files scenario. These counts have zero failures, errors or skips. Broker fault gates (44 tests), O4 MySQL gates (40 tests), latency checks and all three owner-failover E2E steps also passed. This is current Linux CI evidence, separate from the historical local H2 results; it does not expand the recovery or physical-tool interruption guarantee.

The historical WebShell Smoke job also completed successfully, checking out exact head b27a5958a16e. The transcript browser gate passed six tests. Of 208 Playwright cases, 207 passed on their first attempt and one mobile-WebKit /btw drawer case passed on configured retry, with zero final failures. Its first attempt expected the side-question menu item to disappear after selecting Shell mode. The test, mock daemon, Add menu, editor and ordinary Session route are byte-identical to main; the root cause is not established. The retry is retained in the verification record.

Remaining recovery boundary: this evidence does not establish Broker/worker process-death adoption, or an original prompt admission whose reply was lost before its event epoch was saved. The latter path can still attempt submission and be refused by new-work authority. It needs separate fenced recovery/cancellation coverage; do not treat the cold-cache result as resolving every restart/takeover case. #13083 and #13054 remain their own recovery work.

Completed suggestions in the same correction: R1-2 now independently varies storage and generation; removing only the storage comparison makes the new test fail (2 tests, 1 failure, 0 errors), and restoration passes. R1-3 now exercises those variants through the real Hosted integration; local H2 passed, and the current Hosted MySQL CI run passed the same updated scenario as recorded above. R1-17 now explicitly states that workspaceTurns advertises submit/rename authority, with cancellation under its separate rule; the generated mirror and equality test are synchronized.

2. Review suggestion status

Tests that cannot fail: R1-9 (HarnessCoordinator.java:158), R1-12 (ManagedWorkspaceAdmissionTest.java:1149), R1-14 (HostedPublicWorkspaceIT.java:325), R2-2 (HostedPublicWorkspaceIT.java:166), R2-5 (ManagedWorkspaceAdmissionTest.java:799), R2-9 (RuntimeHarnessDrainTest.java:282). Each assertion or guard passes with the clause it is meant to pin removed. Tighten each so it goes red under the named mutation.
Missing coverage: R1-3 (HostedPublicWorkspaceIT.java:346), R1-11 (HostedPublicWorkspaceIT.java:319). R1-3 is addressed by the real Hosted generation/storage scenarios described above. R1-11 diagnostics remain optional follow-up work.
Dead code and duplication: R2-3 (ApiModels.java:93), R2-4 (ManagedWorkspaceRegistry.java:52), R2-6 (ManagedWorkspaceAdmissionTest.java:1273). Completed in preserved concurrent head f8e83fc7707e: the unused constructors and wrapper are removed, and this PR's duplicate rename-retirement test is deleted while main's existing test remains. Both exceptions reach the same retirement branch in the same store state. Current test compilation and the retained main regression passed in the 30-test focused Java run; no behavior was added.
Published wording: R2-7 (managed-agent-public-api.openapi.json:5185), R2-10 (README.md:196). Align the contract and README sentences with what the code enforces.
Cost: R1-6 (ManagedAgentService.java:626). maySubmitWorkspaceTurn runs three JDBC round trips per row of the WebShell session list.

The inherited Harness wedge from @wenshao's round-2 verification is tracked in #13054 and is not repeated here.

3. Late rename may leave the private Harness title stale

Review discussion. Independent real service/H2 reproduction used a stateful Harness fixture and a deterministic held-call schedule: K1 renames to A; a same-key sibling fails and retires K1; fresh K2 completes B; releasing the original K1 writes A at the Harness, then returns 409 session_mutation_superseded from SQL. Final state: public SQL title B, private Harness title A. The probe ran 1 test, 1 failed equality assertion, 0 errors; it did not run live Hosted transport.

The duplicate delivery, sibling retirement and external-write-before-SQL-completion already exist at the PR base ba09794545. #13163 adds a useful SQL completion guard, but it cannot fence the earlier external side effect. README and both design languages now restrict that guarantee to the public SQL title and receipt. The temporary red regression was archived and removed from the shipped passing suite.

Acceptance: concurrent same-key delivery, failure, fresh-key success and delayed older completion must leave public and private persisted titles consistent. A transport variant must cover an accepted remote write finishing after client timeout. The ordering decision belongs at the final durable Harness write and must preserve an explicitly chosen same-key retry contract across server replicas; a process-local lock or an after-the-fact SQL check does not establish it.

4. Owner-assignment workflow checkout failure — resolved

The post-push owner-assignment job fails before running the assignment script. Its checkout uses base commit 2a6879e648e1, supplies three file paths and leaves cone-mode sparse checkout enabled; Git refuses the first path because it is not a directory. This was a base-workflow checkout configuration failure, not a product-test result from #13163. Upstream #13306 fixes cone-mode checkout and #13310 uses the REST assignees endpoint; both are included in b27a5958a16e. The new synchronization job completed successfully: its log checks out trusted base 5ddfacc9d4c1, invokes the assignment script, and deliberately reports Assignment: skipped — no area path matched. No new assignee is claimed. Managed Agent database CI has now passed. WebShell Smoke has also passed (207 first-attempt passes plus one retry pass); two fresh review approvals remain pending.

5. Round-4 non-blocking follow-ups

R4-1 is fixed in a18ea0e42249: early acquisition notification keeps the exact Runtime identity reachable during final-authorization refusal or exception, so resident retirement releases it. It is not deferred. This repair lands only the verified Critical; the remaining Suggestions below stay here.

6. Current-head local verification intermittency

Conflict merge 762f4a2d2b83 is MERGEABLE; build/typecheck and targeted style checks pass. Fresh Java verification passes 127 cases and Checkstyle/SpotBugs, with Java/manifest inputs byte-identical to this head. The independent complete CLI run is retained as exit 1: 213 of 219 passed, six failed, no skips or unhandled errors. All six R4-1 lease cases, four main lost-reply redrive cases and live resident identity/profile negatives passed.

The six failures were four pre-model Hook waiting variants (three also hit ENOTEMPTY during cleanup), unknown Hook fence reload=true (ENOENT during writer release), and takeover cancel admission when detached (cold-load 401). They occur before the changed resident branch. One bounded, original-source reproduction of the three families passed all 23 selected cases, including the six failed variants; 171 were deliberately filtered. No assertion or waiting budget changed. This establishes intermittency, not its cause or a green full run.

Remaining work: capture timing/terminal state during the Hook waits, ensure owner writes finish before deleting fixture storage, and capture the unexpected cold-load response body/listener identity. The untracked replacement listener and short default waits are hypotheses, not proven causes. With nine substantive review rounds and no verified defect in this merge increment, further test-lifecycle diagnosis is deferred here. No speculative timeout or production fix was landed. Historical full/native/UI/CI results elsewhere in this issue retain their named heads.

7. round-2026-10-07 suggestion batch deferred from #13163

The review lane's account changed, so its 2026-10-07 round restarted finding numbering at R1-*. Those ids collide with the R1-*/R2-*/R4-* ids in sections 2 and 5 above and are unrelated findings — cite this batch by discussion link, never by id. All 21 are Suggestion-severity; the same round's three Criticals were fixed on the PR and are not listed here. Each thread carries an individual reply stating the reason, and each was resolved as deferral recorded, not as fixed. Head at deferral time: f2864f6a1d.

sdk-java — no JDK 21 / Maven / ~/.m2 on the editing host, so the fix cannot be compiled or run where it is written; CI's SDK Java lane is the only oracle

published contract surface — the change is a maintainer decision, not a review fix

packages/cli/src/serve/hosted-harness-session.ts — production edits deferred because the file is under concurrent edit on the branch; R1-14 is the one worth picking up first

test quality — assertions that stay green under the mutation they exist to catch