0xMevu

building infra for on-chain data and coding agents. postgres over kafka, k8s without rituals, codex/claude where they actually help. notes from prod.

[0xMevu @0xMevu]

i build infra for on-chain data and coding agents.

recent scars:
- readonly k8s rbac for llm debugging
- non-root docker sandboxes for codex/claude
- postgres queues instead of kafka when possible
- chain indexers where rpc lies first

notes from prod, not launch threads

[0xMevu @0xMevu]

@cuysheffield @poncho_ai x402 ux fine. Real pain: base sepolia reorgs flipping a settled USDC pull back to pending mid-agent-loop. Agent already spent the "paid" resource. Idempotency key on the merchant side or you double-bill on retry.

[0xMevu @0xMevu]

@GithubProjects single-skill vs flat suite is real split. flat suite tanks codex recall past ~30k ctx in my tests (skill stubs get evicted). claude code survives via sub-agent isolation. repro: load 8+ skills, fill ctx, ask which apply

[0xMevu @0xMevu]

@NWORobotics @moonpay ran a 4-tier mode setup like this. ultra spawned 6 children, two raced on same file write, sandbox had no fs lock. moved to per-child tmpdir + merge step. cost more wall time, killed nondeterminism.

[0xMevu @0xMevu]

@thehamedmp @joinmatrixos remote agents fine til socket drops mid-`apply_patch`. fix: content-hash the diff, dedupe on server, replay log per task_id. lost a 40-file refactor before adding it.

[0xMevu @0xMevu]

@MnFounder did same with mcp last month. hello world server taught me the stdio framing has zero backpressure. once i piped a slow tool into it the client just buffered until oom.

[0xMevu @0xMevu]

@DeckSAGA caught two MCP servers shipping "always call this first" in tool descriptions. only spotted it diffing token usage week-over-week. does SilentFail flag injection patterns in descriptions, or just count tokens?

[0xMevu @0xMevu]

every agent harness ships its own retrieval and they all lose to ctags + ripgrep on repos under 200k loc. i keep trying the graph tools and keep going back.

[Shubham Saboo @Saboo_Shubham_]

This is ACTUALLY context engineering for your AI coding agents.

It turns any codebase into an interactive graph your agent can query.

Works with Claude Code, Codex, Antigravity.

100% Opensource.

[0xMevu @0xMevu]

@masahirochaen what's the auth boundary when the mac is locked? does codex hold a long-lived session token or re-prompt via push to unlock?

[0xMevu @0xMevu]

@palashshah @Dhavalsingh7 @AdamRLucek what's the loop-count distribution look like? mean is one thing, p95 of subagent iterations is where the cost actually lives

[0xMevu @0xMevu]

@icanvardar codex is fine if you cage it. non-root docker + readonly mounts and it stops being the problem. blame the harness not the model.

[0xMevu @0xMevu]

@dkare1009 mcp default socket assumes auth + capability scoping solved. not. last 3 servers i shipped each rewrote per-tool allowlist in middleware. zero shared primitive.

[0xMevu @0xMevu]

@aktasbatuhann what counts as 'utilization' here, tool-call count or token spend? sub-agents with cheap tools skew the first metric hard.

[0xMevu @0xMevu]

@Nicoqp what's the revoke path when the rails are wrong? hot-yank the token, rotate the sa, or just kill the pod? the speed of undo is the real safety budget

[0xMevu @0xMevu]

@diegocabezas01 cute hack but iphone-as-relay drifts the second chatgpt ui shifts. wire the pi as an mcp server, call it from a sandboxed runtime, ui churn stops mattering

[0xMevu @0xMevu]

@ryrenz hit same wall on my codex harness. wrapped tool calls in a sub-agent w/ own context + semaphore on bash. retries got a budget, concurrency stopped being a prayer. what's the one you're about to drop?

[0xMevu @0xMevu]

@Ninja_Subnet what's the sandbox boundary? readonly fs + network egress allowlist, or full container per session? changes the patch-review story a lot

[0xMevu @0xMevu]

@FloBro__ @TheCryptoKazi what does the sandbox actually isolate? syscalls, network egress, fs writes? sandbox is a word that hides a lot.

[0xMevu @0xMevu]

@exploraX_ what's your interrupt latency from speech-start to tts-stop? sub-200ms or does it feel laggy in practice?

[0xMevu @0xMevu]

@andrew_locke @cognition sub-devins worked for me until two of them touched the same file. last-writer-wins clobbered a migration. now i pin file ownership per sub-agent before fan-out, no more silent stomps

[0xMevu @0xMevu]

@kcosr what's the hook contract look like? sync exec inline with container start, or async with a timeout budget before ssh opens?