​

Durable Tool Execution v1 — Deferred Findings

​

Open

• Storage-provider follow-up: the shared doctor inventories higher-owned schema domains such as jobs, runtime-store, and workgraph, but cannot certify their supported versions without inverting crate ownership. Normal store open and storage migrate --apply still refuse future versions typed; a future provider-aware certification registry should close the read-only doctor gap for all higher-owned domains together.
• Test-isolation follow-up: running every meerkat-rpc::router::tests case in one parallel test process can make seven unrelated live-session tests contend for the process-wide realtime projection budget. Phase 4’s job and runtime projection tests pass in that run; the live tests should receive explicit serial/resource-budget isolation without coupling that work to durable jobs.
• SDK live-provider follow-up: the Python and TypeScript SDK suites’ image scenarios fail when the local RPC subprocess has no image-generation executor configured (no image generation executor configured for provider gemini). Phase 4’s deterministic SDK builds, wrapper parity, and non-live tests pass; provider-backed image smoke remains an environment/release-lane concern outside the durable-jobs protocol slice.
• Machine-poster follow-up: DetachedJobMachine and RuntimeDeliveryMachine are explicitly listed in the poster generator’s shrink-only coverage-gap roster. Their canonical TLA, contracts, mappings, kernels, TLC verification, and authority docs are present; bespoke large-format SVG layouts are a separate documentation slice.
• Operation-timestamp persistence follow-up: recovered operation shell records currently reconstruct wall-clock timestamps at reopen. Durable DetachedJobWait identity, lifecycle, source, and job fencing remain exact, but the member-status activity.since_ms projection restarts at recovery. Preserving the original timestamp requires a versioned ops-snapshot format change and store-conformance migration, so it is deferred from Phase 6 rather than widening the composition slice.

​

Resolved in scope

• Phase 0–1: repeated runtime wake injection while AgentApplied lagged could duplicate an already accepted one-shot completion continuation. The runtime now injects once; boundary-level delivery owns Busy/cursor retry.
• Phase 0–1: macOS could return EPERM when probing a process-group identity after an accepted drop-time SIGKILL fence, preventing unreturned-operation rollback forever. The stale identity is retired only when the earlier kill fence is proven; ordinary probe failures remain armed.
• Phase 2A: external-wait writes could collapse back to Running, and per_phase use on terminal transitions could preserve the source phase. Explicit generated transitions now preserve waiting writes and enter each terminal phase correctly.
• Phase 2A: recovery-sensitive renewals and checkpoint/wait observations could regress lease facts or commit after the persisted lease. Renewals are now monotonic, lease-bounded writes carry observation time, and focused tests pin the behavior.
• Phase 2A: duplicate cancellation and delivery acknowledgements were rejected, which made crash/replay recovery brittle. Generated idempotent self-loops plus no-op CAS elision now preserve both state and revision.
• Phase 2A: revision saturation could reuse u64::MAX, and realm submission keys could alias malformed or cross-realm identities. Revision exhaustion now fails closed; realm identities and realm-scoped deduplication are validated.
• Phase 2B: the SQLite store now uses a versioned typed serialization envelope for job specifications, generated authority state, typed terminal results, and outbox entries. Pending outbox enumeration and exact reopen recovery are pinned by the SQLite store-conformance suite.
• Phase 2B: SQLite revision storage originally used SQLite’s signed integer range for a u64 domain, flattened shared SQLite error classification, and bypassed the shared TEXT/BLOB JSON codec. Revisions now use an eight-byte big-endian CAS token, SQLite errors stay typed, both JSON storage classes are accepted, and memory/SQLite share pending-outbox acknowledgement conformance.
• Phase 2C: job terminal delivery now crosses an explicit generated job_runtime_delivery composition. RuntimeDeliveryMachine alone mints/reuses the runtime sequence and advances the ordered application cursor; memory and SQLite stores only CAS opaque authority and inbox rows.
• Phase 2C: crash before the runtime insert leaves the job outbox pending, and crash after the runtime insert but before job acknowledgement reuses the original runtime sequence. Focused tests also pin concurrent replay, ordered one-time application, SQLite reopen, and unchanged job attempt/fence facts.
• Phase 2C: recovery validation no longer expands the numeric range implied by a corrupted delivery high-water mark. Persisted submissions re-run constructor invariants, and recovered inbox rows must match the generated source-sequence authority before application.
• Phase 3A: applied job-terminal inbox rows publish through the runtime-owned delivery boundary while operation completion remains only a read projection; the durable shell worker and store do not own lifecycle semantics.
• Phase 3A/3B: runner progress, typed artifacts, monitor notifications, and checkpoints are durable job facts without turning the worker shell or store into lifecycle authority.
• Phase 4 gate containment: the foreground-shell cancellation canary now waits for parseable parent/child PID files rather than mere file creation, and the two process-heavy durable-monitor stress tests have load-tolerant completion deadlines. Their focused tests remain sub-100-ms when the machine is idle.