boost-asio-pro

$npx mdskill add alirezarezvani/claude-skills/boost-asio-pro

Write async C++ networking code that compiles on the *user's* Boost, not the newest one. Asio's API changed shape three times (classic `io_service` → `io_context` → C++20 coroutines) and most Asio code on the internet is from the first era, so **pick the style from the toolchain first**, then follow that style's reference file.

SKILL.md

.github/skills/boost-asio-proView on GitHub ↗
---
name: "boost-asio-pro"
description: "Use when writing or reviewing asynchronous C++ networking code with Boost.Asio or standalone Asio — TCP/UDP servers and clients, SSL/TLS, timers, strands, io_context, co_spawn, awaitable, async_read/async_write, asio::spawn, yield_context, or pre-C++20 completion-handler callbacks."
---

# Boost.Asio / standalone Asio

## Overview

Write async C++ networking code that compiles on the *user's* Boost, not the newest one. Asio's API changed shape three times (classic `io_service` → `io_context` → C++20 coroutines) and most Asio code on the internet is from the first era, so **pick the style from the toolchain first**, then follow that style's reference file.

**References:** [Boost.Asio](https://www.boost.org/doc/libs/latest/doc/html/boost_asio.html) · [standalone Asio](https://think-async.com/Asio/)

Use this skill whenever async C++ networking code is being written or reviewed — and especially when the target toolchain is old, where coroutine examples simply will not compile. The three worked implementations it references are CI-verified from Boost 1.62 (2016) through 1.90.

## Step 1: pick the style (do this before writing code)

Determine the Boost (or Asio) version and the C++ standard actually in use — `find_package(Boost)` output, `dpkg -l libboost-dev`, `brew info boost`, `CMAKE_CXX_STANDARD`, or ask. Do not assume the newest.

| Boost | C++ std | Style | Read |
|-------|---------|-------|------|
| ≥ 1.77 | C++20 | Coroutines (`co_await` + `awaitable<T>`) — preferred | [references/coroutines.md](references/coroutines.md) |
| ≥ 1.74 | C++11–17 | Completion handlers (callbacks) — the portable baseline | [references/pre-cpp20.md](references/pre-cpp20.md) |
| ≥ 1.80 | C++11–17 | Stackful `asio::spawn` + `yield_context` (links Boost.Coroutine — not header-only) | [references/pre-cpp20.md](references/pre-cpp20.md) |
| 1.62–1.65 | C++11 | Classic `io_service` / `strand.wrap` / `expires_from_now` | [references/classic-boost.md](references/classic-boost.md) |

SSL/TLS in any style: [references/ssl.md](references/ssl.md). CMake for any style: [references/build.md](references/build.md).

`io_context`, `make_strand`, `bind_executor`, `steady_timer`, `signal_set`, `async_read`/`async_write`/`async_read_until`, buffers and `resolver` are **library** features — identical in the coroutine and callback styles. Only the suspension mechanism differs.

## Step 2: version floors (verified by compiling, not from docs)

Reach for one of these and the build breaks on older distros:

| Feature | Floor |
|---------|-------|
| `experimental/awaitable_operators.hpp` (the `\|\|` / `&&` operators) | **Boost ≥ 1.77** / Asio ≥ 1.20 |
| `as_tuple` completion token | **Boost ≥ 1.79** / Asio ≥ 1.21 |
| `co_composed` (custom composed ops) | **Boost ≥ 1.85** / Asio ≥ 1.30 |
| 3-arg `asio::spawn(ex, fn, token)` | **Boost ≥ 1.80** (older Boost has only `spawn(ex, fn)`) |
| `any_io_executor` (`strand<any_io_executor>`, `tcp::socket`'s default executor) | **Boost ≥ 1.74** — the floor for the callback style; below it, use legacy `io_context::strand` |
| `io_context`, `make_strand`, `expires_after` | **Boost ≥ 1.66** — below it, classic `io_service` |

Distro floors that bite: **Debian bookworm ships Boost 1.74** (no `awaitable_operators.hpp` — `#include` fails outright), Ubuntu 20.04 ships 1.71 (no `any_io_executor`), Debian 9 ships 1.62.

Language, not library: the chrono literals `250ms` / `30s` are **C++14**. For a true C++11 build write `std::chrono::milliseconds(250)`.

## Step 3: the rules that are actually easy to get wrong

**A strand does not serialize writes.** A strand serializes handler *execution*, not whole composed operations. Two `async_write`s in flight on the same strand still **interleave bytes on the wire**. Full-duplex (a read loop plus concurrent pushes/replies) needs a per-connection strand **and** an outbound queue with an in-flight flag, so at most one `async_write` exists at a time. This is the single most common wrong answer about Asio.

**Buffers do not own memory.** `asio::buffer()` is a view. Storage must outlive the operation: coroutine locals are fine across `co_await` in the same frame; in callback style the same data must become a **member**, not a local.

**Connections must outlive their handlers.** `enable_shared_from_this`, and capture `self` in *every* `co_spawn` / handler — read loop, write loop, and each timer.

**Frame with composed reads.** `async_read` (fills the buffer exactly) for a length prefix and then the body; never `async_read_some`, which returns short.

**Wrap `as_tuple`.** Always `as_tuple(use_awaitable)`. Bare `as_tuple` resolves against the operation's default token and compiles in some contexts, fails in others.

**`async_accept(make_strand(...))` changes two things**: it forces an explicit completion token back on the call, and the accepted socket is `basic_stream_socket<tcp, strand<...>>`, not `tcp::socket`. Take it **by value** or with `auto` — binding it to `tcp::socket&` will not compile.

**Re-arming a timer resolves the pending wait with `operation_aborted`.** In an idle-timeout loop that is the signal to keep waiting, not an error.

**GCC needs `-fcoroutines`** for the C++20 style, and header-only Boost needs `BOOST_ERROR_CODE_HEADER_ONLY` defined in exactly one place (CMake).

## Anti-Patterns

| Mistake | Fix |
|---------|-----|
| Buffer dangling (local goes out of scope during async op) | Ensure buffer lifetime ≥ operation lifetime; coroutine locals or members, not callback locals |
| Forgetting `io.run()` | No handlers dispatch without `run()` / `run_one()` |
| Concurrent socket access without strand | Wrap in `strand<>` or serialize via one coroutine chain |
| Assuming a strand prevents interleaved writes | Add a write queue — see Step 3 |
| Using `use_awaitable` where `deferred` suffices | Omit the token (default is `deferred`) unless using `\|\|` / `&&` |
| Ignoring short reads/writes | Use composed `async_read` / `async_write` / `async_read_until`, not `async_read_some` |
| Not setting `reuse_address` on the acceptor | Set before `bind`/`listen` or restarts hit "address in use" |
| SSL operations without a strand | *All* `ssl::stream` ops need strand synchronization |
| Blocking inside a handler | Never block in a completion handler |
| Accepting a socket with the wrong executor type | See `async_accept(make_strand(...))` in Step 3 |
| Requiring the `Boost::system` component | Header-only since 1.74: `Boost::headers` + `BOOST_ERROR_CODE_HEADER_ONLY`. Only classic (pre-1.66) needs the link |
| Missing `-fcoroutines` on GCC | Build fails — add `$<$<CXX_COMPILER_ID:GNU>:-fcoroutines>` |
| Writing coroutine code for a Boost that predates it | Do Step 1 first |

## Boost.Asio vs standalone Asio

Same author, same API — namespace and includes differ.

| Aspect | Boost.Asio | Standalone Asio |
|--------|-----------|-----------------|
| Namespace / include | `boost::asio` / `<boost/asio.hpp>` | `asio` / `<asio.hpp>` |
| Error code | `boost::system::error_code` | `asio::error_code` (or `std::error_code`) |
| Install (brew) | `brew install boost` | `brew install asio` |
| CMake | `Boost::headers` | manual include path |
| Version (2025) | 1.87–1.90 (with Boost) | 1.30–1.36 (independent) |
| Macro prefix | `BOOST_ASIO_` | `ASIO_` |

Support both with a shim, then use `net::` throughout:
```cpp
#ifdef USE_STANDALONE_ASIO
  #include <asio.hpp>
  namespace net = asio;
  using error_code = asio::error_code;
#else
  #include <boost/asio.hpp>
  namespace net = boost::asio;
  using error_code = boost::system::error_code;
#endif
namespace ssl = net::ssl;
using tcp = net::ip::tcp;
```

## Before you call it done

Check the code you just wrote against this list:

- [ ] Style matches the target Boost version and C++ standard (Step 1), and every API used clears its floor (Step 2).
- [ ] Every buffer passed to an async op outlives that op — no callback locals, no dangling `string_view`.
- [ ] At most one `async_write` per socket in flight, enforced by a queue + flag, if anything writes concurrently with reading.
- [ ] Every async chain on a shared object runs on the same strand; `self` captured in every handler and `co_spawn`.
- [ ] Framing / delimited reads use composed `async_read` / `async_read_until`.
- [ ] Errors are handled, not swallowed: `as_tuple(use_awaitable)` destructured, or the callback's `ec` checked, on every op.
- [ ] `operation_aborted` distinguished from real errors wherever a timer is re-armed or an op is cancelled.
- [ ] Acceptor sets `reuse_address`; shutdown path closes the acceptor and drains sessions.
- [ ] CMake has the standard, `-fcoroutines` for GCC (C++20 only), `BOOST_ERROR_CODE_HEADER_ONLY` in one place, and `Boost::coroutine` only if using stackful `spawn`.
- [ ] It compiles. Build it — most of the mistakes above are compile-time, and the version floors are only real once tested.

## Worked examples

Three CI-verified implementations of the same full-duplex framed-protocol server, one per style — copy from the one matching Step 1. All three live in the upstream repository and are built by CI on every push.

- [market-data-feed](https://github.com/alexprivalov/boost-asio-skill/tree/main/examples/market-data-feed) — C++20 coroutines (Boost 1.77+; verified 1.83–1.90)
- [market-data-feed-precpp20](https://github.com/alexprivalov/boost-asio-skill/tree/main/examples/market-data-feed-precpp20) — callbacks, C++11-clean (verified Boost 1.74+, incl. Windows/MSVC)
- [market-data-feed-classic](https://github.com/alexprivalov/boost-asio-skill/tree/main/examples/market-data-feed-classic) — classic `io_service` (verified back to Boost 1.62 / Debian 9)

## Official documentation

- Overview: https://www.boost.org/doc/libs/latest/doc/html/boost_asio/overview.html
- Reference: https://www.boost.org/doc/libs/latest/doc/html/boost_asio/reference.html
- Examples: https://www.boost.org/doc/libs/latest/doc/html/boost_asio/examples.html

## Cross-References

- `engineering/docker-development` — the old-Boost verification lanes this skill's floors come from are containerised builds (Debian 9 / bookworm, Fedora).
- `engineering/chaos-engineering` — for exercising the failure paths this skill tells you to handle: half-open sockets, idle timeouts, partial frames.
- `engineering-team/playwright-pro` — the client-side counterpart when the server built here is driven from browser-based integration tests.

More from alirezarezvani/claude-skills

SkillDescription
a11y-auditAccessibility audit skill for scanning, fixing, and verifying WCAG 2.2 Level A and AA compliance across React, Next.js, Vue, Angular, Svelte, and plain HTML codebases. Use when auditing accessibility, fixing a11y violations, checking color contrast, generating compliance reports, or integrating accessibility checks into CI/CD pipelines.
ab-test-setupWhen the user wants to plan, design, or implement an A/B test or experiment. Also use when the user mentions "A/B test," "split test," "experiment," "test this change," "variant copy," "multivariate test," "hypothesis," "conversion experiment," "statistical significance," or "test this." For tracking implementation, see analytics-tracking.
ad-creativeWhen the user needs to generate, iterate, or scale ad creative for paid advertising. Use when they say 'write ad copy,' 'generate headlines,' 'create ad variations,' 'bulk creative,' 'iterate on ads,' 'ad copy validation,' 'RSA headlines,' 'Meta ad copy,' 'LinkedIn ad,' or 'creative testing.' This is pure creative production — distinct from paid-ads (campaign strategy). Use ad-creative when you need the copy, not the campaign plan.
adversarial-reviewerAdversarial code review that breaks the self-review monoculture. Use when you want a genuinely critical review of recent changes, before merging a PR, or when you suspect Claude is being too agreeable about code quality. Forces perspective shifts through hostile reviewer personas that catch blind spots the author's mental model shares with the reviewer.
aeoAnswer Engine Optimization (AEO) skill — optimize content to be cited by AI language models (ChatGPT, Perplexity, Claude, Gemini, Mistral) as authoritative sources. Distinct from SEO — AEO optimizes for citation in LLM-generated responses, not search rankings. Use when planning content for AI-first search audiences, auditing existing content for E-E-A-T signals, tracking which pages get cited by which LLMs, or building a citation-friendly content strategy. Triggers — 'AEO audit', 'optimize for ChatGPT', 'get cited by Perplexity', 'LLM citation strategy', 'answer engine optimization', 'content for AI search', 'E-E-A-T audit'. Output is a markdown audit report (default) or JSON for pipeline integration. Stdlib-only Python tools.
agent-decision-receiptsMint a tamper-evident, post-quantum-signed receipt for a consequential agent action (deploy, delete, pay, grant-access, model decision) so it can be verified later from the certificate alone. Use when an autonomous agent takes a side-effecting action that may need to be proven later, or when satisfying EU AI Act Article 12 record-keeping. Three decisions: whether an action needs a receipt, minting it, verifying it. Signing is delegated to the open-source OpenAgentOntology package. Not after-the-fact log analysis; not a hosted notary; not a legal opinion.
agent-designerUse when the user asks to design a multi-agent system, pick an orchestration pattern (supervisor/swarm/pipeline), generate tool schemas for agents, or evaluate agent execution logs for cost, latency, and failure bottlenecks. Examples: 'design an agent architecture for research automation', 'generate Anthropic tool schemas from these tool descriptions', 'analyze these agent run logs for bottlenecks'. NOT for Claude Code workflow files (use workflow-builder) or single-agent prompt design (use agent-workflow-designer).
agent-harnessTurn any domain folder of skills into a bounded agentic loop: compile a goal into a verifiable task plan, execute tasks with the domain's own tools, verify every task with machine-run checks, retry with caps, escalate to a human when budgets exhaust, and refuse to close until everything is verified or explicitly waived. Use when you want an agent or subagent to pick up a goal and drive it to a verified close across one of this repo's 18 domains ('run this goal through the engineering harness', 'set up an agentic loop for marketing work', 'make the finance domain self-verifying'). NOT for authoring Claude Code Workflow-tool .js scripts (workflow-builder), N-agent tournaments on one task (agenthub), single-file metric optimization (autoresearch-agent), or discovering published loop recipes (loop-library).
agent-launcher-orchestratorUse when a user wants to build, launch, grade, or schedule a Claude Managed Agent (CMA) in their own Anthropic account — "build me an agent", "launch this as a managed agent", "run this on a schedule", "grade my agent against a rubric", "set up a nightly worker". Reads the per-session goal (./my-agent/goal.json), routes deterministically to one of five phase sub-skills (interview → stage-launch → grade-iterate → run-without-you → wrap-up) via goal_router.py, and compiles the goal+phase into an execution shape (single-pass workflow / bounded grade→iterate loop / recurring cron deployment loop) via loop_compiler.py. Forks context so heavy intake (build sheets, payloads, eval cases) stays out of the parent thread. All launches are emitted as BYOK curl the user runs with their own key; no tool makes API calls. Inspired by anthropics/launch-your-agent (Apache-2.0). Distinct from engineering/agent-harness (generic domain loop) and engineering/write-a-skill (authors Claude Code skills, not CMAs).
agent-memoryUse when a project's CLAUDE.md has grown past what anyone reads and you want the agent to learn durable facts from its own sessions instead — or when asking why the agent keeps re-learning the same correction, why a remembered rule is wrong, or where a memory line came from. Implements a four-tier store (L0 transcripts / L1 candidates / L2 project context / L3 stable persona) where promotion is earned by recurrence across sessions and days, never by one confident statement, and nothing reaches a committed file without a human adopting it.