convez.ai
Agente
mtschoen

agent-walker

mtschoen/agent-walker

Ver no GitHub
InicialPluginScore · de 0 a 100

Native pace-walker for Claude Code session JSONLs. Side-by-side Rust/C++/Zig/Go implementations behind a shared conformance corpus.

Instalação

$npx convez-ai mtschoen/agent-walker

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 2 semanas
Versão
—
Issues abertas
0
Contribuidores
6

README

A tradução para o português deste README ainda está na fila.

claude-walker

Native pace-walker for Claude Code session JSONLs, implemented side-by-side in Rust, C++, Zig, and Go as a language comparison.

What it does

Walks ~/.claude/projects/**/*.jsonl (parent transcripts + their subagents/agent-*.jsonl siblings), filters to a time range, dedupes assistant turns by message.id, and prices each turn against the canonical Anthropic per-MTok rate table. Emits two dollar totals on stdout.

The walker reads a configurable set of project roots (default ~/.claude/projects plus any extras listed in ~/.claude/walker-roots.json), so the same subscription used from multiple machines can be aggregated by the host that has filesystem access to all of them. See "Multi-host setup" below and the ## Roots section in SPEC.md.

Drop-in optional speedup for schoen-claude-status (the Python statusline does this in ~250ms with an orjson + ProcessPoolExecutor walk; the goal here is 30-80ms).

Multi-host setup

The sliding-window pace projection only works correctly if the walker sees every transcript billed to the subscription. When the same plan is used from two machines, neither machine's ~/.claude/projects is a complete picture on its own.

Make the other machine's transcripts reachable via the filesystem (SMB/NFS mount, sync, whatever), then point the walker at the mounted path via ~/.claude/walker-roots.json:

{
  "extra_roots": [
    "/mnt/other-host/Users/mtsch/.claude/projects"
  ]
}

The file is per-host and lives outside any synced memory — chonkers' config points at llamabox's mount, llamabox's config points at chonkers'.

Latency caveat. Walking JSONLs over a remote mount is slower than the local case. On a representative Windows-→-Linux SMB mount the cost-mode walk goes from ~200ms (local-only) to ~3s (with the SMB extra root). The statusline pace cache (15s TTL) absorbs the cost so the user sees one slow render per cache window, but if the remote host is asleep or the mount times out, the walker still exits 0 with primary-only totals — it never blocks the caller indefinitely.

Why four languages

The work is small, well-defined, CPU-bound, parallelizable, and performance-sensitive — a tight benchmark for "what does it actually feel like to ship a fast CLI in $LANG." Comparison axes are tracked in RESULTS.md: runtime, binary size, lines of code, build time, distribution friction.

Layout

shared/
  corpus/        JSONL fixtures with expected outputs (the ground truth)
  conformance.py runs each binary, asserts ±$0.01 agreement
  bench.py       times each binary on the live ~/.claude/projects fleet
rust/            cargo, simd-json or sonic-rs
go/              go modules, encoding/json or sonic-go
cpp/             cmake, simdjson
zig/             build.zig, std.json or hand-rolled

See SPEC.md for the CLI contract every implementation must honor and the correctness rules they all share.

Status

Early — see PLAN.md. Spec + conformance harness first; Rust is the first reference implementation; the others land behind it.

Ver o original no GitHub

Bom para

Relacionados