feat: add C++23 AWS Durable Execution SDK

[zhongkechen]

Summary

• add a dependency-free C++23 durable execution core with deterministic replay and a stable SDK-owned BLAKE2b operation-ID contract
• implement steps, waits, retries, child contexts, callbacks, chained invoke, polling, replay-safe values, recursion, parallel/map, typed DAG flows, and a deterministic local runner
• add a stable custom-operation extension SPI with one-shot reservations, local IDs, custom subtypes, stateful steps, callbacks, invokes, waits, and child contexts
• add versioned instrumentation plugins with invocation/operation/attempt/change hooks and fault isolation
• add optional AWS SDK for C++ and aws-lambda-cpp adapters plus reproducible Amazon Linux 2023 packaging
• define source, ABI, durable-history, wire, and serialized-payload compatibility contracts for the C++ SDK

Conformance

• maps all 171 requirements in the pinned official suite (148 core + 23 plugin)
• strict manifest generation reports 171/171 and zero NotImplemented
• official CloudWatch log matcher passes locally for all 23 plugin cases
• official validator parser reports zero uncovered requirements
• SAM basic validation, SAM lint, and offline SAM build pass
• deployed AWS validator report: 171 passed, 0 failed, 0 optional failed, 0 uncovered
• persistent concurrency-controlled stack avoids repeated resource creation between runs

Verification

• GitHub Actions Release job: passed
• GitHub Actions ASan/UBSan job: passed
• GitHub Actions deployed conformance job: passed
• Python-generated reference fixtures reproduce exactly as independent oracles
• shared library exposes SONAME 1 and preserves the original execution_state constructor symbol
• installed core and optional-adapter CMake package consumers pass
• Amazon Linux 2023 x86-64 runtime builds with GCC 14, AWS SDK 1.11.878, and aws-lambda-cpp 1.0.1
• current microbenchmarks remain near the established baseline (operation IDs ~350 ns, minimal invocation decode <1 microsecond)

[zhongkechen]

Codex goal execution metrics

The following values were captured from the completed goal telemetry and the final token-count event at the goal completion boundary.

Runtime configuration

• Model: openai.gpt-5.6-sol
• Reasoning effort: xhigh
• Model context window: 828,400 tokens
• Codex CLI version: 0.149.0
• Approval policy: on-request
• Approval reviewer: user
• Sandbox mode: workspace-write
• Model changes during the goal: none; every recorded turn used the same model and reasoning effort

Token usage

• Goal-accounted tokens: 5,087,315
• Uncached input tokens: 4,508,366
• Output tokens: 578,949
  • Reasoning output tokens: 208,872 (included in output tokens)
• Cached input tokens reused: 541,254,011
• Total input tokens processed, including cached input: 545,762,377
• Cache-write input tokens: 4,427,548

The goal-accounted total is:

uncached input + output = 4,508,366 + 578,949 = 5,087,315

The much larger total-input figure includes repeatedly reused cached context. Cached input is excluded from the goal-accounted token total above.

Time usage

• Goal start: 2026-08-25 16:46:41 UTC
• Goal completion: 2026-08-26 14:13:07 UTC
• Total wall-clock time: 77,186 seconds (21h 26m 26s)
• Clearly identifiable long approval-blocked intervals: approximately 52,177 seconds (14h 29m 37s)
• Estimated time excluding approval waits: approximately 25,009 seconds (6h 56m 49s)
• Reasonable active-time range: approximately 5–7 hours

Timing methodology and limitation

The system records exact goal wall-clock time, token usage, model metadata, and context-window size. It does not separately record command runtime and approval-wait duration for every approval-gated command. Short approval waits and command execution therefore overlap in the event log.

The approval-wait-excluded value is calculated by subtracting clearly identifiable long approval-blocked intervals from total wall-clock time. It is an event-log estimate, not an exact active-time measurement.