refactor(monorepo): reorganización lógica + renames + SDDs + split CHANGELOG

Reorganización física de crates/:
- core/ (mezclaba 6 propósitos) se divide en protocol/, init/, runtime/, compat/
- shared/ (3 crates) se redistribuye en protocol/ e init/
- lapaloma (sub-módulo de ui_engine) se promueve a modules/pineal/

Renames de proyectos:
- shipote → shuma (runtime de sandboxes)
- nouser → akasha (explorador de Mónadas)
- yahweh → nahual (motor GPUI, antes ui_engine/)
- lapaloma → pineal (data-viz agnóstica)

Fraccionamiento UI → core agnóstico:
- vista-core (DeckState + snap, 175 LOC, 5 tests verdes)
- barra-core (Task + render_html + sanitize, 90 LOC, 5 tests verdes)
- vista-web y barra-web ahora son thin DOM bindings

Documentación nueva:
- 16 SDDs por subdirectorio (≤80 LOC c/u): protocol/init/runtime/compat
  + 10 módulos + apps/
- docs/STATUS.md con cifras reales por proyecto
- docs/ROADMAP.md con plan a finalización (6 hitos, ~6-8 semanas)
- CHANGELOG.md particionado en docs/changelog/<proyecto>.md (7 buckets)

Automatización:
- scripts/reorg.py — script idempotente que: git mv directorios, renombra
  package names, recomputa path = refs, reescribe imports rust, actualiza
  workspace Cargo.toml. Soporta --dry-run.
- scripts/split-changelog.py — particiona CHANGELOG por componente.

Validación:
- cargo check --workspace pasa (124 crates + 2 nuevos cores).
- 10 tests adicionales (5 en vista-core + 5 en barra-core) verdes.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
sergio
2026-05-19 14:48:34 +00:00
parent 86fb6ae20b
commit 550c98f275
375 changed files with 8512 additions and 7155 deletions
+73
View File
@@ -0,0 +1,73 @@
# shipote
**Runtime de espacios aislados con flujo tipado** dentro del fractal brahman.
shipote no es un shell POSIX. Es un daemon long-running que:
- Aísla procesos en **Workspaces** (cada uno con su `SomaSpec`: namespaces, cgroup, rlimits).
- Encadena comandos en **Pipelines** (DAG con fan-out, fan-in, taps, replay buffer).
- **Discierne contenido** sobre cada flow (detecta `application/json`, `text/markdown`, magic bytes, hasta `brahman:card`).
- Expone **data plane** real: subscribers externos se conectan a Unix sockets y reciben los bytes del stream.
- Se integra al ecosistema vía `brahman-sidecar` (anuncia Cards al broker).
```
┌─────────────┐
│ shipote │ daemon long-running
│ -daemon │ $XDG_RUNTIME_DIR/shipote.sock
└──────┬──────┘
│ admin protocol (postcard)
┌─────────────────┼─────────────────┐
│ │ │
┌─────▼─────┐ ┌─────▼─────┐ ┌─────▼─────┐
│ shipote │ │ shipote- │ │ HTTP / │
│ (CLI) │ │ shell │ │ custom │
│ │ │ (GUI) │ │ (futuro) │
└───────────┘ └───────────┘ └───────────┘
```
## Quick start
```sh
# 1. Compilar
cargo build -p shipote-daemon -p shipote-cli
# 2. Arrancar daemon (en una terminal):
SHIPOTE_LOG=info ./target/debug/shipote-daemon
# 3. En otra terminal:
./target/debug/shipote health
./target/debug/shipote workspace create demo.toml
./target/debug/shipote run -w <ws-id> /bin/echo -- hola
```
## Estado actual
- **11 crates** en el monorepo. Ver [ARCHITECTURE.md](docs/ARCHITECTURE.md).
- **85 tests** pasando.
- Compatible con cualquier supervisor (`ente-incarnate` es reusable — no requiere ser PID 1).
## Documentación
- [**ARCHITECTURE.md**](docs/ARCHITECTURE.md) — qué crates lo componen, cómo se conectan, decisiones clave.
- [**CLI.md**](docs/CLI.md) — referencia de comandos del binario `shipote`.
- [**RECIPES.md**](docs/RECIPES.md) — specs TOML de workspaces y pipelines con casos prácticos.
- [**DEVELOPMENT.md**](docs/DEVELOPMENT.md) — cómo correr daemon, GUI, tests, snapshot.
## Acoplamiento con el resto del monorepo
| Reusa | Propósito |
|---|---|
| `ente-incarnate` (extraído de `ente-soma` en Fase 0) | clone(2) + namespaces + cgroup + rlimits |
| `brahman-card` | tipos `Card`, `SomaSpec`, `Flow`, `TypeRef` |
| `brahman-sidecar` + `brahman-handshake` | anuncio al broker |
| `ente-snapshot` filosofía | persistencia (shipote tiene su propio `ShipoteSnapshot` v4) |
| `yahweh-launcher` + widgets | GUI `shipote-shell` |
| `shipote-discern` ← usado por | `yahweh-provider-fs` (mime per archivo) y `nouser-core::cluster::pick_lens` |
## Limitaciones declaradas
- **`mount/pid/net/uts/ipc/cgroup` namespaces** requieren `CAP_SYS_ADMIN` o combo con `CLONE_NEWUSER`.
- **`user` namespace** puede estar bloqueado por sysctl o LSM (apparmor/selinux).
- **`cgroup` v2 limits** (memory.max, pids.max) requieren delegation. Sin delegation, accounting funciona pero el kernel no enforce.
- **Pipeline restart** preserva spec pero **no ULIDs** del pipeline_id (cada relaunch genera uno nuevo).
- **Replay buffer broadcast**: subscribers tarde reciben el snapshot del replay (cap configurable por bytes o chunks), pero pueden perder chunks intermedios si el ring rota antes de conectarse.
+36
View File
@@ -0,0 +1,36 @@
# modules/shuma/ — Runtime de espacios aislados (era shipote)
**Propósito.** Cada Workspace = un proceso aislado (namespaces +
cgroups + capabilities filtradas) que expone un wire protocol tipado.
El daemon es dueño de los workspaces; los clientes (cli/shell/gateway)
hablan postcard sobre Unix socket.
## Crates
| crate | tipo | rol |
| ----------------- | ---- | ------------------------------------------------------- |
| `shuma-card` | lib | Card del daemon + spec del Workspace |
| `shuma-protocol` | lib | Wire types: requests/responses + framing |
| `shuma-discern` | lib | Lookup de daemon vía broker brahman |
| `shuma-core` | lib | Pipeline: parse spec → encarnar → supervisar → persist |
## Dependencias
- `shuma-core``init/ente-incarnate` (encarnación real).
- `shuma-protocol``protocol/brahman-card`.
- Apps: `shuma-daemon` (dueño), `shuma-cli`, `shuma-shell` (GUI),
`shuma-gateway` (HTTP/JSON ↔ postcard).
## Docs
Documentación completa en `crates/modules/shuma/docs/`:
- `ARCHITECTURE.md` — diseño + flow
- `CLI.md` — referencia del cli
- `DEVELOPMENT.md` — guía de contribución
- `RECIPES.md` — specs ejemplo de Workspace
## Estado
LOC 6,907. Backend completo (daemon + cli funcionales). Tests sobre
discern + protocol. 14 TODOs en core (supervisión avanzada). Ver
`docs/changelog/shuma.md`.
+159
View File
@@ -0,0 +1,159 @@
# Arquitectura de shipote
## Diagrama de capas
```
┌──────────────────────────────────────────────────────────────────┐
│ Apps (binarios) │
│ │
│ shipote-daemon shipote (CLI) shipote-shell │
│ crates/apps/... crates/apps/... crates/apps/... │
└─────────┬─────────────────────┬──────────────────────┬───────────┘
│ │ │
│ postcard frames (length-prefixed)
│ │ │
└─────────────────────┴──────────────────────┘
shipote-protocol — wire types
crates/modules/shipote/shipote-protocol
┌──────────────────────────────────────────────────────────────────┐
│ shipote-core — runtime │
│ crates/modules/shipote/shipote-core │
│ │
│ WorkspaceManager (Arc<Mutex<Inner>>) │
│ ├ workspaces: HashMap<WorkspaceId, WorkspaceState> │
│ ├ saved_pipelines, pipeline_flows, pipeline_supervisors │
│ ├ restart_specs, pending_pipeline_restarts │
│ └ dirty: AtomicBool (snapshot incremental) │
│ │
│ Modules: │
│ - pipeline.rs (run_pipeline, splitter, merger, TokenBucket) │
│ - flow_channel.rs (Unix socket + tokio::broadcast + replay) │
│ - logbuf.rs (ring buffer 64 KiB para stdout/stderr) │
│ - stats.rs (WorkspaceStats: RSS/CPU/peak/cores) │
│ - persist.rs (ShipoteSnapshot v4 con JSON atomic write) │
└──────────────────────────────────────────────────────────────────┘
┌──────────────────────────────────────────────────────────────────┐
│ Soporte compartido │
│ │
│ shipote-card shipote-discern │
│ - WorkspaceSpec - DiscernPipeline │
│ - PipelineSpec - MagicBytes, JsonProbe, etc. │
│ - CommandRef - Detecta brahman:card │
│ - DiscernPolicy │
│ │
│ ente-incarnate (crates/shared/) — extraído de ente-soma │
│ - Incarnator { caps, env, stdio, pre_exec } │
│ - CapabilitySet::detect (user_ns, cgroup, kernel version) │
│ - ChildStdio, ChildSetup (NoNewPrivs, Chdir, ...) │
│ - clone(2) + namespaces + cgroup move (post-clone setup) │
└──────────────────────────────────────────────────────────────────┘
brahman-card · ente-incarnate · nix · libc
```
## Crates (11 nuevos)
### Core del runtime
- **`crates/shared/ente-incarnate`** — rutina extraída del Init para encarnar Cards en procesos aislados. Reusable por cualquier supervisor. NO requiere ser PID 1.
- **`crates/modules/shipote/shipote-card`** — tipos: `WorkspaceSpec`, `PipelineSpec`, `CommandRef`, `FlowEdge`, `DiscernPolicy`, `QuotaEnforcement`. Compilan a `brahman_card::Card`.
- **`crates/modules/shipote/shipote-protocol`** — wire postcard length-prefixed sobre Unix socket en `$XDG_RUNTIME_DIR/shipote.sock`.
- **`crates/modules/shipote/shipote-discern`** — `DiscernPipeline` + discerners default. Reusado por `yahweh-provider-fs` y `nouser-core::cluster`.
- **`crates/modules/shipote/shipote-core`** — `WorkspaceManager` + módulos `pipeline`, `flow_channel`, `logbuf`, `persist`, `stats`.
### Binarios
- **`crates/apps/shipote-daemon`** — long-running. Escucha admin socket, ejecuta el dispatch, reaper periódico.
- **`crates/apps/shipote-cli`** — `shipote` (clap). Para administración + scripting.
- **`crates/apps/shipote-shell`** — GUI con `yahweh_launcher`. Dashboard: estado, workspaces, comandos, flows, quotas, sparkline.
- **`crates/apps/shipote-gateway`** — HTTP/JSON adapter. `POST /rpc` con body JSON `Request` → traduce a postcard, round-trip al daemon, retorna JSON. Default `127.0.0.1:7378`.
### Consumidores del discerner (externos a shipote pero usándolo)
- **`crates/modules/ui_engine/libs/providers/fs`** — `FileDataProvider` ahora puebla `mime_type` con `DiscernPipeline`.
- **`crates/modules/nouser/core/src/cluster.rs`** — `pick_lens` usa el discerner como fallback cuando la extensión cae a `Lens::Grid`.
## Modelo conceptual
### Workspace
Espacio aislado raíz. Una **Card con `kind: Ente`, `payload: Virtual`** que aloja N comandos hijos. Su `SomaSpec` define el sandbox compartido.
```toml
label = "demo"
on_exit = "reap"
ttl = 60000 # opcional, ms
[soma.rlimits]
mem_bytes = 10485760 # 10 MiB
nproc = 4
[soma.cgroup]
path = "shipote/demo"
[quota_enforce]
mem = "kill" # None | Log | Kill
```
### CommandRef
Un comando dentro del workspace. Su `SomaSpec` se **intersecta** con el del padre (OR de namespaces, MIN de rlimits).
### PipelineSpec
DAG de `CommandRef` con edges tipados. Soporta:
- **1→1**: pipe directo (sin task intermedio).
- **1→N (fan-out)**: splitter task replica.
- **N→1 (fan-in)**: merger task con mpsc (una sub-task por reader → channel → escribir al consumer.stdin).
### Flow channel (data plane)
Unix socket en `$XDG_RUNTIME_DIR/shipote-flow-<pipeline_id>-<edge_idx>.sock`. Cada subscriber recibe primero el replay (cap por chunks y/o bytes), después el broadcast live. Tokio `broadcast::Sender<Arc<Vec<u8>>>` — zero-copy entre subscribers.
### Discerner
Pipeline ordenado: `MagicBytes``CardProbe``JsonProbe``TomlProbe``Utf8Probe`. Cada uno devuelve `Discernment { ty: TypeRef, confidence, mime, lens }`.
## Decisiones clave de arquitectura
### Por qué `ente-incarnate` es separado
La rutina de aislamiento (clone+ns+cgroup+rlimits) vivía dentro del Init (`ente-soma`). Era global y atada a PID 1. La separamos en Fase 0 para que shipote (y futuros supervisores) puedan reusarla sin implicar privilegio.
Hoy `ente-soma` es un wrapper de ~30 líneas sobre `ente-incarnate` que preserva la semántica histórica (`set_bus_sock` global + `strict_caps=false`). `ente-zero` sigue funcionando intocado.
### Por qué pipeline-level restart relanza el pipeline ENTERO
Los pipes intermedios de un comando que muere ya están cerrados. Restart parcial no funciona. La identidad ULID del pipeline cambia entre intentos; `restart_count` y `current_backoff_ms` se preservan en el supervisor.
### Por qué replay buffer guarda ANTES del broadcast
Si guardas después, un subscriber que conecta entre `broadcast::send` y `replay.push_back` puede perder ese chunk. El orden importa: replay → broadcast.
### Por qué `pipe2(O_CLOEXEC)` siempre
Sin `O_CLOEXEC`, el siguiente comando del pipeline hereda copias del write-end del pipe anterior. El read-end nunca ve EOF y el consumidor cuelga. Bug encontrado y arreglado en Fase F.
### Por qué AsyncFd + non-blocking en taps/mergers
`tokio::fs::File::from_std` con un FD blocking bloquea el reactor entero. La solución: `fcntl(F_SETFL, O_NONBLOCK)` antes de envolver en `AsyncFd<OwnedFd>`. Patrón usado en splitter, merger y log drainer.
### Por qué token-bucket real reemplazó el sleep simple
El sleep `chunk_size / rate` era uniforme y no permitía burst legítimo. `TokenBucket` con refill por wall time + `capacity=rate_bps` (1s burst) permite burst real y throttle suave.
### Por qué `dirty: AtomicBool` para snapshot incremental
Cada SIGTERM hacía full re-serialize aunque no hubiera cambios. `Ordering::Relaxed` está bien porque el peor caso es un save innecesario. `restore_snapshot` resetea dirty=false (acabamos de hidratar, no es mutation).
### Por qué auth con SO_PEERCRED y `target: "audit"`
SO_PEERCRED es automático en Unix sockets — leer `ucred.uid` del peer y comparar con el propio uid. Sin grupos, sin caps. `SHIPOTE_TRUST_ANYONE=1` es el escape hatch.
`target: "audit"` separa el log de mutaciones del log normal. Filtrable con `EnvFilter`: `SHIPOTE_LOG=warn,audit=info`.
## Versión de snapshot
| Versión | Campos | Forward-compat |
|---------|--------|----------------|
| v1 | `workspaces` | ✓ |
| v2 | + `saved_pipelines` | ✓ (default vacío) |
| v3 | + `live_pipelines` | ✓ (default vacío) |
| v4 | + `stats_history` per workspace (cap 16) | ✓ (default vacío) |
`SNAPSHOT_VERSION = 4`. Reader rechaza versiones > 4 (forward-incompat). Todos los campos nuevos usan `#[serde(default)]`.
## Wire protocol
Todos los mensajes son enums tagged externamente, serializados con `postcard`. Framing: u32 BE length-prefix + payload. Max frame: 1 MiB.
### Requests
- Reads: `Ping`, `Health`, `Capabilities`, `WorkspaceList`, `WorkspaceStats`, `WorkspaceStatsHistory`, `WorkspaceFullSummary`, `WorkspaceQuota`, `CommandList`, `CommandLogs`, `PipelineSavedList`, `FlowList`, `FlowThroughput`, `Discern`.
- Mutaciones (loguean audit): `WorkspaceCreate`, `WorkspaceStop`, `Run`, `PipelineRun`, `PipelineRunSaved`, `PipelineStop`, `PipelineSave`, `PipelineDrop`, `FlowDrop`.
+134
View File
@@ -0,0 +1,134 @@
# Referencia CLI — `shipote`
El binario `shipote` (compilado por `cargo build -p shipote-cli`) se conecta al daemon vía Unix socket. Por default usa `$XDG_RUNTIME_DIR/shipote.sock`; override con `--socket <path>`.
```sh
shipote [--socket PATH] <comando> [...]
```
## Comandos globales
### `shipote ping`
Health-check rápido. Imprime `pong` si el daemon responde.
### `shipote health`
Health endpoint estructurado.
```
version: 0.1.0
uptime: 11411 ms
alive_workspaces: 1
alive_commands: 1
alive_pipelines: 0
active_flows: 0
dirty: true
```
### `shipote caps`
Capacidades runtime detectadas: kernel version, `user_ns` status, `cgroup_v2`, `cgroup_delegated`, `cap_sys_admin`.
### `shipote discern <path>`
Discierne ad-hoc un archivo (no requiere workspace). Imprime `ty`, `confidence`, `mime`, `lens`.
## Workspaces
### `shipote workspace create <spec.toml|.json>`
Crea workspace desde un spec. Imprime el ULID asignado. Si el cgroup está delegated y el spec declara rlimits, se aplica `memory.max`/`pids.max`.
### `shipote workspace list`
Lista workspaces vivos.
### `shipote workspace stats <id>`
Resource accounting:
```
commands: 1 alive / 1 total
rss: 2.05 MiB
rss_peak: 58.89 MiB
cpu: 0.300 s
cpu_pct: 98.7 % (24.7% total / 4 cores)
source: proc
uptime: 1002 ms
```
### `shipote workspace quota <id>`
Reporta breaches del `soma.rlimits`. Si `quota_enforce.{mem,nproc}=Kill`, el daemon mata automáticamente al detectar.
### `shipote workspace stop <id> [--grace-ms N]`
Stop graceful. Default `grace_ms=1000`. `0` = SIGKILL inmediato.
## Comandos directos
### `shipote run -w <ws-id> [--restart-on-failure] <exec> -- <argv>...`
One-shot dentro del workspace. Si `--restart-on-failure`, el reaper relanza con backoff exponencial cuando exit != 0.
**Nota**: usar `--` antes de los argv del comando para que clap no los confunda con flags suyos:
```sh
shipote run -w 01... /bin/sh -- -c "echo hola"
```
### `shipote commands <ws-id>`
Lista comandos del workspace (vivos + exited) con pid, status, bytes_log.
### `shipote logs <ws-id> <cmd-id> [--stream {stdout|stderr|both}] [--tail N] [--follow]`
Tail del ring buffer de logs.
- `--stream` default `stdout`. `both` concatena stderr-tras-stdout.
- `--tail N`: últimos N bytes (0 = todo).
- `--follow` (`-f`): poll cada 200ms, imprime suffix nuevo hasta que el comando termine.
## Pipelines
### `shipote pipeline run <spec.toml> [--tap] [--tail] [--var KEY=VAL ...]`
Lanza pipeline.
- `--tap`: crea flow channels (data plane) por cada edge.
- `--tail`: auto-implica `--tap`, suscribe al primer flow socket y vuelca a stdout.
- `--var KEY=VAL` (repetible): sustituye `${KEY}` en el spec antes de lanzar.
### `shipote pipeline stop <pipeline-id> [--grace-ms N]`
Stop selectivo del pipeline (no afecta otros comandos del workspace).
### `shipote pipeline save <name> <spec.toml>`
Persiste el spec bajo `name` (sobrevive restart vía snapshot).
### `shipote pipeline saved-list`
Lista pipelines guardados.
### `shipote pipeline drop <name>`
Elimina un saved pipeline.
### `shipote pipeline run-saved <name> [--tap] [--tail] [--var KEY=VAL ...]`
Ejecuta un pipeline guardado con vars opcionales.
## Flows
### `shipote flow list`
Lista pipelines vivos con sockets activos (uno por edge con `--tap`).
### `shipote flow throughput`
Bytes acumulados + rate por socket.
```
shipote-flow-<ULID>-0.sock 0.1 KiB total 0.07 KiB/s
```
### `shipote flow tail <socket-path>`
Conecta directo al Unix socket y vuelca hasta EOF. Si el splitter tiene replay buffer, lo recibís primero.
### `shipote flow drop <pipeline-id>`
Cierra el data plane de un pipeline (drop de FlowChannels).
## Variables de entorno
- `SHIPOTE_LOG` — filtro `tracing-subscriber` (`info`, `warn`, `audit=info,warn`, ...).
- `SHIPOTE_TRUST_ANYONE=1` — daemon acepta peers con cualquier uid (testing).
- `XDG_RUNTIME_DIR` — dónde se crea el socket admin y los flow sockets.
- `XDG_STATE_HOME` — dónde se guarda el snapshot. Fallback: `$HOME/.local/state/shipote/state.json`.
## Trucos del shell zsh
`shipote run` toma flags propios antes del `--`. Si el comando que ejecutás tiene un argumento que arranca con `-`, usá `--`:
```sh
# Mal (clap intenta parsear `-c` como flag de shipote):
shipote run -w <ws> /bin/sh -c "echo hi"
# Bien:
shipote run -w <ws> /bin/sh -- -c "echo hi"
```
+244
View File
@@ -0,0 +1,244 @@
# Development guide — shipote
Cómo compilar, correr, testear y debuggear shipote desde la raíz del monorepo (`/home/sergio/brahman`).
## Compilar
Todos los crates de shipote y su soporte:
```sh
cargo build -p shipote-daemon -p shipote-cli -p shipote-shell
```
Sólo lo esencial (daemon + cli, sin GUI):
```sh
cargo build -p shipote-daemon -p shipote-cli
```
Workspace completo:
```sh
cargo check --workspace
```
## Correr el daemon
```sh
SHIPOTE_LOG=info ./target/debug/shipote-daemon
```
Filtros útiles:
```sh
SHIPOTE_LOG=warn,shipote_core=info ./target/debug/shipote-daemon # menos ruido
SHIPOTE_LOG=warn,audit=info ./target/debug/shipote-daemon # sólo audit
SHIPOTE_LOG=info,shipote_core::pipeline=debug ./target/debug/shipote-daemon # debug del pipeline
```
El daemon:
- Escucha `$XDG_RUNTIME_DIR/shipote.sock` (fallback `/run/user/$UID/shipote.sock`).
- Restaura snapshot de `$XDG_STATE_HOME/shipote/state.json` (fallback `$HOME/.local/state/shipote/state.json`).
- Relanza pipelines `live` (con `restart_on_failure=true`) automáticamente.
- Reaper cada 500 ms (cosecha zombies + drena pending restarts + chequea quota breaches).
- SIGTERM/SIGINT → snapshot + `stop_with_grace(1s)` de todos los workspaces + exit.
## Correr el shell (GUI)
```sh
cargo run -p shipote-shell
```
Polling cada 2 s al daemon. Si el daemon no está corriendo, muestra "Daemon DOWN" banner rojo. Sparkline de RSS por workspace. Cards: Estado, Capabilities, Workspaces, Comandos, Saved pipelines, Flow channels, Quota breaches, Live tail.
Requiere DISPLAY (X11/Wayland) — no corre en sandbox sin gráfico.
## Correr el CLI
```sh
./target/debug/shipote ping
./target/debug/shipote health
./target/debug/shipote workspace create demo.toml
```
Ver [CLI.md](CLI.md) para referencia completa.
## Tests
Suite completa (85 tests al cierre de Fase R):
```sh
cargo test -p ente-incarnate \
-p shipote-card \
-p shipote-protocol \
-p shipote-discern \
-p shipote-core \
-p yahweh-provider-fs \
-p nouser-core \
--no-fail-fast
```
Breakdown:
| Crate | Tests |
|---|---|
| `ente-incarnate` | 16 (clone+ns, stdio, pre_exec, caps detect, cgroup paths) |
| `nouser-core` | 27 (pre-existentes — discerner integrado sin regresiones) |
| `shipote-card` | 8 (roundtrip TOML/JSON, compile-to-card, intersect_soma) |
| `shipote-core` | 26 (workspaces, pipeline fan-in/out, flow_channel replay, stats history, restart, quota enforce, snapshot dirty-skip, ...) |
| `shipote-discern` | 5 (MagicBytes, JsonProbe, CardProbe, TomlProbe, Utf8Probe) |
| `yahweh-provider-fs` | 3 (discern_head + integración) |
Tests específicos útiles:
```sh
# Fan-in (merger):
cargo test -p shipote-core --lib pipeline_fanin -- --nocapture
# Replay buffer:
cargo test -p shipote-core --lib replay -- --nocapture
# Quota enforcement:
cargo test -p shipote-core --lib quota_enforce -- --nocapture
# Snapshot incremental:
cargo test -p shipote-core --lib save_snapshot_skips
```
> Algunos tests del pipeline (fan-in/fan-out) son tokio-async y necesitan `--test-threads=1` cuando todos los tests corren juntos. La suite por crate ya funciona bien.
## Smoke E2E manual
Demo end-to-end de los features principales:
```sh
# 1. Limpiar estado previo
rm -f $HOME/.local/state/shipote/state.json
pkill -f shipote-daemon
# 2. Arrancar daemon (en background o terminal aparte)
SHIPOTE_LOG=info ./target/debug/shipote-daemon &
sleep 0.3
# 3. Crear workspace
cat > /tmp/ws.toml <<'EOF'
label = "demo"
on_exit = "reap"
EOF
WS=$(./target/debug/shipote workspace create /tmp/ws.toml)
echo "workspace: $WS"
# 4. Health check
./target/debug/shipote health
# 5. Run comando con log capture
./target/debug/shipote run -w $WS /bin/echo -- "hello shipote"
sleep 0.3
CMD=$(./target/debug/shipote commands $WS | head -1 | awk '{print $1}')
./target/debug/shipote logs $WS $CMD
# 6. Pipeline con tap (data plane real)
cat > /tmp/pipe.toml <<EOF
label = "demo-pipe"
workspace = "$WS"
discern = { sample_bytes = 4096, enrich_producer = true }
[[nodes]]
label = "p1"
payload.Native = { exec = "/bin/echo", argv = ['{"hello": 1}'], envp = [] }
[[nodes]]
label = "p2"
payload.Native = { exec = "/bin/cat", argv = [], envp = [] }
[[edges]]
from = 0
from_output = "stdout"
to = 1
to_input = "stdin"
EOF
./target/debug/shipote pipeline run /tmp/pipe.toml --tap
# 7. Flow throughput
./target/debug/shipote flow throughput
# 8. Stats con CPU%
./target/debug/shipote workspace stats $WS
# 9. SIGTERM (drain + snapshot)
pkill -TERM -f shipote-daemon
```
## Debug del daemon
### Logs del daemon en archivo
```sh
SHIPOTE_LOG=info ./target/debug/shipote-daemon 2>&1 | tee /tmp/shipote-daemon.log
```
### Filtrar audit log
```sh
SHIPOTE_LOG=warn,audit=info ./target/debug/shipote-daemon 2>&1 | grep audit
```
### Verificar caps runtime
```sh
shipote caps
```
Te dice si `user_ns` está bloqueado por sysctl/LSM y si tu cgroup está delegado.
### Inspeccionar el snapshot
```sh
cat $HOME/.local/state/shipote/state.json | jq .
```
JSON con: version, timestamp_ms, workspaces, saved_pipelines, live_pipelines, stats_history persistida.
### Debugging FDs
```sh
ls -la /proc/$(pgrep -x shipote-daemon)/fd | head -20
```
Si ves muchos `pipe:[N]` o `socket:[N]` huérfanos, hay leak en algún spawn.
## Arquitectura del repositorio
```
crates/
├── apps/
│ ├── shipote-daemon/ ← binario long-running
│ ├── shipote-cli/ ← binario `shipote`
│ └── shipote-shell/ ← GUI GPUI
├── modules/shipote/
│ ├── shipote-card/ ← tipos WorkspaceSpec/PipelineSpec/...
│ ├── shipote-protocol/ ← wire postcard
│ ├── shipote-discern/ ← MagicBytes/Json/Toml/Card/Utf8
│ ├── shipote-core/ ← WorkspaceManager + pipeline + flow_channel + ...
│ ├── README.md ← entry point (este archivo es vecino)
│ └── docs/
│ ├── ARCHITECTURE.md
│ ├── CLI.md
│ ├── RECIPES.md
│ └── DEVELOPMENT.md ← estás acá
└── shared/
└── ente-incarnate/ ← extraído de ente-soma; reusable por shipote y otros
```
## Memoria del proyecto
Toda la historia de fases (F a R) está documentada en:
```
/home/sergio/.claude/projects/-home-sergio-brahman/memory/project_shipote.md
```
Esa memoria persiste entre conversaciones con Claude Code. Si arrancás una sesión nueva, Claude la consulta automáticamente.
## Issues conocidos
### El remote `origin` está mal configurado
`https:/sergio:...` con un solo `/`. Para subir los 10 commits locales:
```sh
git remote set-url origin https://sergio:<password>@gitea.gioser.net/sergio/brahman
git push origin main
```
### El daemon no se conecta al broker
Si el Init (`brahman-init.sock`) no está corriendo, el sidecar loguea `no conectado` y el daemon sigue standalone. Esto es **diseñado** (graceful degradation), no un bug. Los announcements de edge-card al broker tampoco llegan en este modo.
### Pipeline restart pierde ULIDs
Cada relaunch genera un pipeline_id nuevo. Trackers externos que dependen del ULID se rompen. Workaround: usar el `label` del pipeline (estable entre restarts).
### `tokio::test` + `waitpid` síncrono
Algunos tests del pipeline necesitan `--test-threads=1` porque el waitpid síncrono dentro de un `current_thread` runtime bloquea el reactor antes de que las tareas spawneadas corran.
+279
View File
@@ -0,0 +1,279 @@
# Recetas de specs
Specs TOML para casos comunes. Todas asumen que `<WS>` ya existe (creado con `shipote workspace create <ws-spec>`).
## Workspaces
### Workspace mínimo
```toml
label = "demo"
on_exit = "reap"
```
### Workspace con TTL (auto-stop tras N ms)
```toml
label = "ephemeral"
on_exit = "reap"
ttl = 30000 # 30 s
```
### Workspace con rlimits (sólo accounting)
```toml
label = "bounded"
on_exit = "reap"
[soma.rlimits]
mem_bytes = 10485760 # 10 MiB — visible en `shipote workspace quota`
nproc = 4
```
### Workspace con enforcement automático
```toml
label = "strict"
on_exit = "reap"
[soma.rlimits]
mem_bytes = 5242880
nproc = 2
[quota_enforce]
mem = "kill" # None | Log | Kill
nproc = "kill"
```
### Workspace con cgroup delegado (kernel enforces)
```toml
label = "cgroup-enforced"
on_exit = "reap"
[soma.rlimits]
mem_bytes = 10485760
nproc = 4
[soma.cgroup]
path = "shipote/bounded" # bajo $cgroup_actual/shipote/bounded
```
> Requiere `cgroup_delegated: true` en `shipote caps`. Sino el accounting funciona pero el kernel no enforces.
### Workspace con namespacing real
```toml
label = "isolated"
on_exit = "reap"
[soma.namespaces]
user = true
pid = true
mount = true
net = false
uts = false
ipc = false
cgroup = false
[soma.rlimits]
mem_bytes = 0
nproc = 0
nofile = 0
[soma.cgroup]
path = ""
```
> Requiere `user_ns: Allowed` (o `cap_sys_admin: true`).
## Pipelines
### Pipeline lineal con tap (data plane)
```toml
label = "echo-cat"
workspace = "<WS>"
discern = { sample_bytes = 4096, enrich_producer = true }
[[nodes]]
label = "producer"
payload.Native = { exec = "/bin/echo", argv = ['{"hello": 1}'], envp = [] }
[[nodes]]
label = "consumer"
payload.Native = { exec = "/bin/cat", argv = [], envp = [] }
[[edges]]
from = 0
from_output = "stdout"
to = 1
to_input = "stdin"
```
Run:
```sh
shipote pipeline run echo-cat.toml --tap
# imprime: edge ty=json mime=application/json conf=0.95
```
### Pipeline con fan-out (1 → N)
```toml
label = "broadcast"
workspace = "<WS>"
discern = { sample_bytes = 4096, enrich_producer = true }
[[nodes]]
label = "src"
payload.Native = { exec = "/bin/echo", argv = ["mensaje compartido"], envp = [] }
[[nodes]]
label = "wc-c"
payload.Native = { exec = "/usr/bin/wc", argv = ["-c"], envp = [] }
[[nodes]]
label = "wc-l"
payload.Native = { exec = "/usr/bin/wc", argv = ["-l"], envp = [] }
[[edges]]
from = 0
from_output = "stdout"
to = 1
to_input = "stdin"
[[edges]]
from = 0
from_output = "stdout"
to = 2
to_input = "stdin"
```
### Pipeline con fan-in (N → 1)
```toml
label = "merge"
workspace = "<WS>"
[[nodes]]
label = "p1"
payload.Native = { exec = "/bin/echo", argv = ["from-p1"], envp = [] }
[[nodes]]
label = "p2"
payload.Native = { exec = "/bin/echo", argv = ["from-p2"], envp = [] }
[[nodes]]
label = "merge-sink"
payload.Native = { exec = "/bin/cat", argv = [], envp = [] }
[[edges]]
from = 0
from_output = "stdout"
to = 2
to_input = "stdin"
[[edges]]
from = 1
from_output = "stdout"
to = 2
to_input = "stdin"
```
### Pipeline con replay y rate-limit
```toml
label = "throttled"
workspace = "<WS>"
[discern]
sample_bytes = 4096
enrich_producer = true
replay_chunks = 32 # default
replay_bytes = 65536 # cap adicional por bytes (0 = sólo chunks)
max_bytes_per_sec = 1024 # token-bucket con burst de 1s
[[nodes]]
label = "fast"
payload.Native = { exec = "/bin/sh", argv = ["-c", "for i in 1 2 3 4 5; do echo line-$i; done"], envp = [] }
[[nodes]]
label = "sink"
payload.Native = { exec = "/bin/cat", argv = [], envp = [] }
[[edges]]
from = 0
from_output = "stdout"
to = 1
to_input = "stdin"
```
### Pipeline supervisado (restart on failure con backoff)
```toml
label = "supervised"
workspace = "<WS>"
restart_on_failure = true
restart_backoff_ms = 200 # inicial; escala x2
restart_max_backoff_ms = 30000 # cap
restart_max = 5 # 0 = infinito
[[nodes]]
label = "flaky"
payload.Native = { exec = "/bin/false", argv = [], envp = [] }
```
Después de 5 restarts (`/bin/false` siempre exit=1), el daemon loguea `restart_max reached — giving up` y el supervisor se descarta.
### Pipeline con templating
Spec con placeholders:
```toml
label = "tmpl-${VARIANT}"
workspace = "<WS>"
discern = { sample_bytes = 4096, enrich_producer = true }
[[nodes]]
label = "gen-${VARIANT}"
payload.Native = { exec = "/bin/echo", argv = ["greeting from ${VARIANT}"], envp = [] }
[[nodes]]
label = "sink"
payload.Native = { exec = "/bin/cat", argv = [], envp = [] }
[[edges]]
from = 0
from_output = "stdout"
to = 1
to_input = "stdin"
```
Run con vars:
```sh
shipote pipeline run tmpl.toml --var VARIANT=alpha
shipote pipeline run tmpl.toml --var VARIANT=beta
```
Variables sin match quedan intactas (útil para detectar olvidos).
## Subscribers externos
### Tail directo a un flow socket
```sh
shipote pipeline run mypipe.toml --tap &
sleep 0.3
SOCK=$(shipote flow list | grep shipote-flow | xargs)
shipote flow tail "$SOCK"
```
Si conectás tarde, el replay buffer te entrega los últimos N chunks (según `replay_chunks` y `replay_bytes` del spec).
### Modo live-tail integrado
```sh
shipote pipeline run mypipe.toml --tail
# vuelca el primer flow_socket a stdout hasta que el productor termine.
```
## Combinatorias útiles
### Workspace con cleanup automático
```toml
label = "burst-and-die"
on_exit = "reap"
ttl = 10000 # auto-stop a los 10s
[soma.rlimits]
mem_bytes = 5242880
```
### Pipeline JSON-aware con discern enriched
- Producer escribe JSON.
- Discern detecta `application/json` con confidence 0.95.
- Card efímera anunciada al broker (si está corriendo): `shipote.flow.<id>.<from>.<output>.json`.
- Subscribers downstream pueden filtrar por TypeRef en el broker.
@@ -0,0 +1,17 @@
[package]
name = "shuma-card"
version.workspace = true
edition.workspace = true
rust-version.workspace = true
license.workspace = true
authors.workspace = true
publish.workspace = true
description = "Tipos de shipote: WorkspaceSpec, PipelineSpec, CommandRef, FlowEdge. Compilan a Cards de brahman-card."
[dependencies]
brahman-card = { path = "../../../protocol/brahman-card" }
serde = { workspace = true }
serde_json = { workspace = true }
toml = { workspace = true }
thiserror = { workspace = true }
ulid = { workspace = true }
+655
View File
@@ -0,0 +1,655 @@
//! `shuma-card` — tipos del runtime shuma.
//!
//! Tres entidades nuevas encima del `brahman-card::Card`:
//!
//! - [`WorkspaceSpec`] — espacio aislado raíz con su propio `SomaSpec`.
//! - [`CommandRef`] — un comando dentro de un workspace.
//! - [`PipelineSpec`] — DAG de `CommandRef` conectados por `FlowEdge`.
//!
//! Cada `WorkspaceSpec`/`CommandRef` se **compila** a una o varias
//! [`brahman_card::Card`] que el daemon entrega al [`Incarnator`] de
//! `ente-incarnate`. Esto preserva el contrato canónico del fractal.
#![forbid(unsafe_code)]
use brahman_card::{Card, Payload, Permissions, SomaSpec, Supervision};
use serde::{Deserialize, Serialize};
use std::time::Duration;
use thiserror::Error;
use ulid::Ulid;
// =====================================================================
// Identidades
// =====================================================================
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize)]
pub struct WorkspaceId(pub Ulid);
impl WorkspaceId {
pub fn new() -> Self {
Self(Ulid::new())
}
}
impl Default for WorkspaceId {
fn default() -> Self {
Self::new()
}
}
impl std::fmt::Display for WorkspaceId {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
self.0.fmt(f)
}
}
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize)]
pub struct PipelineId(pub Ulid);
impl PipelineId {
pub fn new() -> Self {
Self(Ulid::new())
}
}
impl Default for PipelineId {
fn default() -> Self {
Self::new()
}
}
impl std::fmt::Display for PipelineId {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
self.0.fmt(f)
}
}
// =====================================================================
// Workspace
// =====================================================================
/// Espacio aislado de shuma. Es la raíz de aislamiento — cualquier comando
/// que corre dentro hereda restricciones y no puede aflojarlas.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct WorkspaceSpec {
pub label: String,
/// Aislamiento del workspace mismo (cuando se materializa como Card raíz).
#[serde(default)]
pub soma: SomaSpec,
/// Permisos máximos para hijas. Hijas pueden bajar pero no subir.
#[serde(default)]
pub permissions: Permissions,
/// `None` = vive hasta `stop`. `Some(d)` = el daemon lo termina tras d.
#[serde(default, with = "opt_duration_millis")]
pub ttl: Option<Duration>,
/// Slots de flow pre-declarados. Limitan qué consumidores externos al
/// workspace pueden empatar contra los productores internos.
#[serde(default)]
pub flow_dirs: Vec<FlowSlot>,
/// Política al terminar el workspace.
#[serde(default)]
pub on_exit: ExitPolicy,
/// Política de enforcement automático cuando un recurso excede su
/// rlimit declarado en `soma.rlimits`. Default = sólo accounting
/// (None) — el quota report sigue funcionando, pero no hay kill.
#[serde(default)]
pub quota_enforce: QuotaEnforcement,
}
/// Acción cuando un recurso excede su límite. Aplica por recurso (mem,
/// nproc, ...).
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize, Deserialize)]
#[serde(rename_all = "lowercase")]
pub enum QuotaAction {
/// Sólo accounting: la breach aparece en `workspace_quota`.
#[default]
None,
/// Loguear la breach (info-level del daemon).
Log,
/// Matar todos los comandos vivos del workspace (SIGKILL, sin grace).
Kill,
}
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct QuotaEnforcement {
#[serde(default)]
pub mem: QuotaAction,
#[serde(default)]
pub nproc: QuotaAction,
}
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct FlowSlot {
pub name: String,
pub direction: FlowDirection,
/// Si `Workspace`, sólo otros nodos del mismo workspace pueden empatar.
/// Si `Public`, el broker global puede emparejar.
#[serde(default)]
pub scope: FlowScope,
}
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "lowercase")]
pub enum FlowDirection {
Input,
Output,
}
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize, Deserialize)]
#[serde(rename_all = "lowercase")]
pub enum FlowScope {
#[default]
Workspace,
Public,
}
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize, Deserialize)]
#[serde(rename_all = "kebab-case")]
pub enum ExitPolicy {
/// Reapear procesos hijos y descartar estado.
#[default]
Reap,
/// Mantener el workspace en `stopped` para inspección.
Keep,
/// Tomar snapshot del estado (para restart posterior).
Snapshot,
}
mod opt_duration_millis {
use serde::{Deserialize, Deserializer, Serialize, Serializer};
use std::time::Duration;
pub fn serialize<S: Serializer>(d: &Option<Duration>, s: S) -> Result<S::Ok, S::Error> {
d.map(|x| x.as_millis() as u64).serialize(s)
}
pub fn deserialize<'de, D: Deserializer<'de>>(d: D) -> Result<Option<Duration>, D::Error> {
let v: Option<u64> = Option::deserialize(d)?;
Ok(v.map(Duration::from_millis))
}
}
// =====================================================================
// CommandRef
// =====================================================================
/// Un comando que vive dentro de un workspace. Se compila a una `Card` con
/// `pin_to` apuntando al workspace padre (label) y su `SomaSpec`
/// intersectado con el del workspace.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct CommandRef {
pub label: String,
pub payload: Payload,
/// SomaSpec del comando. El compilador lo intersecta con el del workspace.
#[serde(default)]
pub soma: SomaSpec,
/// Inputs/outputs tipados (mismos `Flow` de brahman-card).
#[serde(default)]
pub flows: brahman_card::Flows,
/// Política de supervisión. Default `OneShot` (un comando se ejecuta y muere).
#[serde(default = "default_oneshot")]
pub supervision: Supervision,
}
fn default_oneshot() -> Supervision {
Supervision::OneShot
}
// =====================================================================
// Pipeline
// =====================================================================
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct PipelineSpec {
pub label: String,
pub workspace: WorkspaceId,
pub nodes: Vec<CommandRef>,
#[serde(default)]
pub edges: Vec<FlowEdge>,
#[serde(default)]
pub discern: DiscernPolicy,
/// Si `true` y cualquier comando del pipeline termina con exit!=0,
/// el daemon relaunch el pipeline ENTERO (stop + nuevo run_pipeline).
/// Útil para pipelines de procesamiento continuo.
#[serde(default)]
pub restart_on_failure: bool,
/// Backoff inicial entre restarts (ms). Crece exponencialmente
/// hasta `restart_max_backoff_ms`. Default 200ms = ~5 restarts/s
/// inicial, escalando rápido.
#[serde(default = "default_restart_backoff")]
pub restart_backoff_ms: u64,
/// Backoff máximo (ms). Default 30s. El backoff no crece más allá.
#[serde(default = "default_restart_max_backoff")]
pub restart_max_backoff_ms: u64,
/// Máximo de restarts antes de dar up. `0` = infinito. Default 0.
/// Útil para fail-loud cuando un pipeline siempre falla.
#[serde(default)]
pub restart_max: u32,
}
fn default_restart_backoff() -> u64 {
200
}
fn default_restart_max_backoff() -> u64 {
30_000
}
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct FlowEdge {
/// Índice en `PipelineSpec.nodes` del productor.
pub from: usize,
/// Nombre del Flow output del productor.
pub from_output: String,
/// Índice en `PipelineSpec.nodes` del consumidor.
pub to: usize,
/// Nombre del Flow input del consumidor.
pub to_input: String,
}
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct DiscernPolicy {
/// Bytes a samplear por flow para el discernidor. Default 4 KiB.
#[serde(default = "default_sample_bytes")]
pub sample_bytes: usize,
/// Si `true`, enriquece la Card del producer con el TypeRef detectado.
#[serde(default = "default_true")]
pub enrich_producer: bool,
/// Chunks que el FlowChannel guarda en replay buffer para subscribers
/// tarde. Default 32. Subir si los productores escriben en ráfagas y
/// querés que los consumidores tardíos vean toda la salida.
#[serde(default = "default_replay_chunks")]
pub replay_chunks: usize,
/// Tope adicional por **bytes** acumulados en el replay buffer. Lo
/// que se exceda primero (chunks o bytes) drop-ea el chunk más viejo.
/// `0` = sin tope por bytes (sólo aplica `replay_chunks`). Útil para
/// productores con chunks de tamaño variable.
#[serde(default)]
pub replay_bytes: usize,
/// Rate-limit del flow channel (bytes/s). `0` = sin límite. Si está
/// definido, el splitter sleeps proporcional al tamaño del chunk
/// antes de re-broadcastear. Protege subscribers lentos.
#[serde(default)]
pub max_bytes_per_sec: u64,
}
impl Default for DiscernPolicy {
fn default() -> Self {
Self {
sample_bytes: default_sample_bytes(),
enrich_producer: default_true(),
replay_chunks: default_replay_chunks(),
replay_bytes: 0,
max_bytes_per_sec: 0,
}
}
}
fn default_sample_bytes() -> usize {
4096
}
fn default_true() -> bool {
true
}
fn default_replay_chunks() -> usize {
32
}
// =====================================================================
// Compilación a Card
// =====================================================================
#[derive(Debug, Error)]
pub enum CompileError {
#[error("workspace label vacío")]
EmptyWorkspaceLabel,
#[error("comando con label vacío en posición {0}")]
EmptyCommandLabel(usize),
#[error("edge fuera de rango: from={from}, to={to}, nodes={nodes}")]
EdgeOutOfBounds { from: usize, to: usize, nodes: usize },
}
impl WorkspaceSpec {
/// Compila el WorkspaceSpec a una Card raíz que el Incarnator puede
/// encarnar. Usa `Payload::Virtual` (el workspace no es un proceso por
/// sí solo; sólo aloja hijos).
pub fn to_card(&self, id: WorkspaceId) -> Result<Card, CompileError> {
if self.label.trim().is_empty() {
return Err(CompileError::EmptyWorkspaceLabel);
}
let mut c = Card::new(format!("shuma.workspace.{}", self.label));
c.id = id.0;
c.soma = self.soma.clone();
c.permissions = self.permissions.clone();
c.payload = Payload::Virtual;
c.supervision = Supervision::OneShot;
Ok(c)
}
}
impl CommandRef {
/// Compila un CommandRef a Card hija de un workspace. La Card resultante
/// referencia al workspace por label en `pin_to` de cada Flow.
pub fn to_card(&self, idx: usize, workspace_label: &str) -> Result<Card, CompileError> {
if self.label.trim().is_empty() {
return Err(CompileError::EmptyCommandLabel(idx));
}
let mut c = Card::new(format!("shuma.cmd.{}.{}", workspace_label, self.label));
c.payload = self.payload.clone();
c.soma = intersect_soma(&self.soma, /*workspace*/ &SomaSpec::default());
c.supervision = self.supervision.clone();
c.flow = self.flows.clone();
// pin_to del workspace en cada Flow input/output → el broker prefiere
// resolver dentro del mismo workspace cuando hay candidatos múltiples.
let pin = format!("shuma.workspace.{}", workspace_label);
for f in c.flow.input.iter_mut().chain(c.flow.output.iter_mut()) {
if f.pin_to.is_none() {
f.pin_to = Some(pin.clone());
}
}
Ok(c)
}
}
/// Intersección conservadora: si el workspace pidió aislamiento, la hija
/// también lo tiene (no puede aflojar). Si la hija pidió aislamiento extra,
/// se respeta.
fn intersect_soma(child: &SomaSpec, ws: &SomaSpec) -> SomaSpec {
let mut out = child.clone();
out.namespaces.mount |= ws.namespaces.mount;
out.namespaces.pid |= ws.namespaces.pid;
out.namespaces.net |= ws.namespaces.net;
out.namespaces.uts |= ws.namespaces.uts;
out.namespaces.ipc |= ws.namespaces.ipc;
out.namespaces.user |= ws.namespaces.user;
out.namespaces.cgroup |= ws.namespaces.cgroup;
// rlimits: el menor (más restrictivo) gana.
out.rlimits.mem_bytes = min_opt(out.rlimits.mem_bytes, ws.rlimits.mem_bytes);
out.rlimits.nproc = min_opt(out.rlimits.nproc, ws.rlimits.nproc);
out.rlimits.nofile = min_opt(out.rlimits.nofile, ws.rlimits.nofile);
out
}
fn min_opt<T: Ord + Copy>(a: Option<T>, b: Option<T>) -> Option<T> {
match (a, b) {
(Some(x), Some(y)) => Some(x.min(y)),
(Some(x), None) | (None, Some(x)) => Some(x),
(None, None) => None,
}
}
impl PipelineSpec {
pub fn validate(&self) -> Result<(), CompileError> {
let n = self.nodes.len();
for (i, c) in self.nodes.iter().enumerate() {
if c.label.trim().is_empty() {
return Err(CompileError::EmptyCommandLabel(i));
}
}
for e in &self.edges {
if e.from >= n || e.to >= n {
return Err(CompileError::EdgeOutOfBounds {
from: e.from,
to: e.to,
nodes: n,
});
}
}
Ok(())
}
}
// =====================================================================
// I/O conveniencia (TOML + JSON)
// =====================================================================
#[derive(Debug, Error)]
pub enum LoadError {
#[error("io: {0}")]
Io(#[from] std::io::Error),
#[error("toml: {0}")]
Toml(#[from] toml::de::Error),
#[error("json: {0}")]
Json(#[from] serde_json::Error),
#[error("formato desconocido (esperado .toml o .json)")]
UnknownFormat,
}
pub fn load_workspace_spec(path: &std::path::Path) -> Result<WorkspaceSpec, LoadError> {
let raw = std::fs::read_to_string(path)?;
match path.extension().and_then(|s| s.to_str()) {
Some("toml") => Ok(toml::from_str(&raw)?),
Some("json") => Ok(serde_json::from_str(&raw)?),
_ => Err(LoadError::UnknownFormat),
}
}
pub fn load_pipeline_spec(path: &std::path::Path) -> Result<PipelineSpec, LoadError> {
let raw = std::fs::read_to_string(path)?;
match path.extension().and_then(|s| s.to_str()) {
Some("toml") => Ok(toml::from_str(&raw)?),
Some("json") => Ok(serde_json::from_str(&raw)?),
_ => Err(LoadError::UnknownFormat),
}
}
/// Sustituye `${KEY}` en todos los strings del spec por el valor de
/// `vars["KEY"]`. Variables sin match quedan intactas (no se borra el
/// placeholder — útil para detectar olvidos).
///
/// Walk recursivo sobre la representación JSON intermedia para cubrir
/// labels, argv, envp, paths y cualquier String del schema.
pub fn substitute_vars(
spec: &PipelineSpec,
vars: &std::collections::HashMap<String, String>,
) -> Result<PipelineSpec, serde_json::Error> {
if vars.is_empty() {
return Ok(spec.clone());
}
let mut v = serde_json::to_value(spec)?;
walk_subst(&mut v, vars);
serde_json::from_value(v)
}
fn walk_subst(v: &mut serde_json::Value, vars: &std::collections::HashMap<String, String>) {
match v {
serde_json::Value::String(s) => {
*s = subst_str(s, vars);
}
serde_json::Value::Array(arr) => {
for item in arr {
walk_subst(item, vars);
}
}
serde_json::Value::Object(obj) => {
for (_, val) in obj.iter_mut() {
walk_subst(val, vars);
}
}
_ => {}
}
}
fn subst_str(s: &str, vars: &std::collections::HashMap<String, String>) -> String {
let mut out = String::with_capacity(s.len());
let bytes = s.as_bytes();
let mut i = 0;
while i < bytes.len() {
if i + 1 < bytes.len() && bytes[i] == b'$' && bytes[i + 1] == b'{' {
// Buscar el cierre `}`.
if let Some(close) = bytes[i + 2..].iter().position(|&b| b == b'}') {
let key = std::str::from_utf8(&bytes[i + 2..i + 2 + close]).unwrap_or("");
if let Some(val) = vars.get(key) {
out.push_str(val);
i += 2 + close + 1;
continue;
}
}
}
out.push(bytes[i] as char);
i += 1;
}
out
}
#[cfg(test)]
mod subst_tests {
use super::*;
use std::collections::HashMap;
#[test]
fn substitute_in_argv_and_label() {
let mut vars = HashMap::new();
vars.insert("MSG".into(), "hola-mundo".into());
vars.insert("LABEL".into(), "renamed".into());
let spec = PipelineSpec {
label: "p-${LABEL}".into(),
workspace: WorkspaceId::new(),
nodes: vec![CommandRef {
label: "node-${LABEL}".into(),
payload: Payload::Native {
exec: "/bin/echo".into(),
argv: vec!["${MSG}".into()],
envp: vec![],
},
soma: Default::default(),
flows: Default::default(),
supervision: Supervision::OneShot,
}],
edges: vec![],
discern: DiscernPolicy::default(),
restart_on_failure: false,
restart_backoff_ms: 200,
restart_max_backoff_ms: 30_000,
restart_max: 0,
};
let out = substitute_vars(&spec, &vars).unwrap();
assert_eq!(out.label, "p-renamed");
assert_eq!(out.nodes[0].label, "node-renamed");
match &out.nodes[0].payload {
Payload::Native { argv, .. } => assert_eq!(argv[0], "hola-mundo"),
_ => panic!("wrong payload"),
}
}
#[test]
fn unknown_var_left_intact() {
let vars = HashMap::new();
let spec = PipelineSpec {
label: "p-${UNDEFINED}".into(),
workspace: WorkspaceId::new(),
nodes: vec![],
edges: vec![],
discern: DiscernPolicy::default(),
restart_on_failure: false,
restart_backoff_ms: 200,
restart_max_backoff_ms: 30_000,
restart_max: 0,
};
let out = substitute_vars(&spec, &vars).unwrap();
assert_eq!(out.label, "p-${UNDEFINED}");
}
}
#[cfg(test)]
mod tests {
use super::*;
fn sample_workspace() -> WorkspaceSpec {
WorkspaceSpec {
label: "demo".into(),
soma: SomaSpec::default(),
permissions: Permissions::default(),
ttl: Some(Duration::from_secs(60)),
flow_dirs: vec![FlowSlot {
name: "out".into(),
direction: FlowDirection::Output,
scope: FlowScope::Public,
}],
on_exit: ExitPolicy::Reap,
quota_enforce: Default::default(),
}
}
#[test]
fn workspace_toml_roundtrip() {
let ws = sample_workspace();
let s = toml::to_string(&ws).unwrap();
let back: WorkspaceSpec = toml::from_str(&s).unwrap();
assert_eq!(back.label, ws.label);
assert_eq!(back.ttl, ws.ttl);
assert_eq!(back.flow_dirs.len(), 1);
}
#[test]
fn workspace_json_roundtrip() {
let ws = sample_workspace();
let s = serde_json::to_string(&ws).unwrap();
let back: WorkspaceSpec = serde_json::from_str(&s).unwrap();
assert_eq!(back.label, ws.label);
}
#[test]
fn workspace_compiles_to_card() {
let ws = sample_workspace();
let id = WorkspaceId::new();
let c = ws.to_card(id).unwrap();
assert_eq!(c.id, id.0);
assert!(c.label.starts_with("shuma.workspace."));
assert!(matches!(c.payload, Payload::Virtual));
}
#[test]
fn empty_label_rejected() {
let mut ws = sample_workspace();
ws.label = String::new();
assert!(ws.to_card(WorkspaceId::new()).is_err());
}
#[test]
fn pipeline_validates_edges() {
let p = PipelineSpec {
label: "p".into(),
workspace: WorkspaceId::new(),
nodes: vec![CommandRef {
label: "a".into(),
payload: Payload::Virtual,
soma: SomaSpec::default(),
flows: brahman_card::Flows::default(),
supervision: Supervision::OneShot,
}],
edges: vec![FlowEdge {
from: 0,
from_output: "x".into(),
to: 5,
to_input: "y".into(),
}],
discern: DiscernPolicy::default(),
restart_on_failure: false,
restart_backoff_ms: 200,
restart_max_backoff_ms: 30_000,
restart_max: 0,
};
assert!(p.validate().is_err());
}
#[test]
fn intersect_soma_takes_more_restrictive() {
let mut child = SomaSpec::default();
child.rlimits.mem_bytes = Some(1_000_000);
let mut ws = SomaSpec::default();
ws.rlimits.mem_bytes = Some(500_000);
ws.namespaces.user = true;
let r = intersect_soma(&child, &ws);
assert_eq!(r.rlimits.mem_bytes, Some(500_000));
assert!(r.namespaces.user);
}
}
@@ -0,0 +1,27 @@
[package]
name = "shuma-core"
version.workspace = true
edition.workspace = true
rust-version.workspace = true
license.workspace = true
authors.workspace = true
publish.workspace = true
description = "Runtime de shipote: WorkspaceManager sobre ente-incarnate. Estado in-memory, lifecycle, reaping."
[dependencies]
shuma-card = { path = "../shuma-card" }
shuma-discern = { path = "../shuma-discern" }
brahman-card = { path = "../../../protocol/brahman-card" }
ente-incarnate = { path = "../../../init/ente-incarnate" }
nix = { workspace = true }
libc = { workspace = true }
anyhow = { workspace = true }
thiserror = { workspace = true }
tokio = { workspace = true }
tracing = { workspace = true }
ulid = { workspace = true }
serde = { workspace = true }
serde_json = { workspace = true }
[dev-dependencies]
tempfile = { workspace = true }
@@ -0,0 +1,478 @@
//! Flow channels: data plane sobre Unix socket por edge enriquecido.
//!
//! Cuando un splitter detecta el TypeRef de un edge, además de replicar a
//! los consumers internos del pipeline, se levanta un FlowChannel que
//! expone los bytes a subscribers externos (otros módulos del fractal).
//!
//! ## Diseño
//!
//! - `tokio::sync::broadcast::channel` para fan-out lock-less entre el
//! splitter (sender) y los N subscribers conectados.
//! - `UnixListener` accept-loop: por cada cliente nuevo, spawn una task
//! que drena el receiver y escribe al socket.
//! - Subscribers lentos pueden perder mensajes (broadcast::Receiver::Lagged)
//! — se loguea warn y se sigue. Esto es deliberado para no bloquear el
//! splitter en consumers lentos.
//!
//! ## Lifetime
//!
//! `FlowChannel` se construye con `new(path)`. Cuando se drop:
//! - El `accept_task` se cancela (vía drop del `tokio::task::JoinHandle`
//! que tenemos abort-on-drop).
//! - El socket file se borra del FS (`Drop` impl).
//!
//! Sender clones son baratos; los subscribers conectados se enteran del
//! cierre cuando todos los senders se dropean.
use std::collections::VecDeque;
use std::path::{Path, PathBuf};
use std::sync::{Arc, Mutex};
use tokio::io::AsyncWriteExt;
use tokio::net::UnixListener;
use tokio::sync::broadcast;
use tokio::task::AbortHandle;
use tracing::{debug, warn};
/// Capacidad del broadcast channel. Si un subscriber está más de N chunks
/// atrasado, queda `Lagged` y empieza a perder mensajes.
const BROADCAST_CAP: usize = 64;
/// Chunks default del replay buffer. Cuando un cliente nuevo se conecta,
/// recibe hasta estos N chunks antes de iniciar el broadcast live.
/// Override via `FlowChannel::with_replay_cap`.
pub const DEFAULT_REPLAY_CHUNKS: usize = 32;
pub struct FlowChannel {
sender: broadcast::Sender<Arc<Vec<u8>>>,
replay: Arc<Mutex<VecDeque<Arc<Vec<u8>>>>>,
replay_caps: ReplayCaps,
socket_path: PathBuf,
meter: Arc<FlowMeter>,
_accept_handle: AbortOnDrop,
}
/// Contador de bytes y rate (bytes/s ventana 1s).
#[derive(Debug)]
pub struct FlowMeter {
/// Bytes acumulados desde la creación del FlowChannel.
total_bytes: std::sync::atomic::AtomicU64,
/// Ring buffer de (timestamp_ms, bytes_acumulados) para calcular
/// el rate sobre los últimos N samples.
rate_window: Mutex<VecDeque<(u64, u64)>>,
}
const RATE_WINDOW_SAMPLES: usize = 32;
impl FlowMeter {
fn new() -> Self {
Self {
total_bytes: std::sync::atomic::AtomicU64::new(0),
rate_window: Mutex::new(VecDeque::with_capacity(RATE_WINDOW_SAMPLES)),
}
}
fn record(&self, delta: u64) {
let now = self.total_bytes
.fetch_add(delta, std::sync::atomic::Ordering::Relaxed)
+ delta;
let ts = std::time::SystemTime::now()
.duration_since(std::time::UNIX_EPOCH)
.map(|d| d.as_millis() as u64)
.unwrap_or(0);
if let Ok(mut w) = self.rate_window.lock() {
if w.len() >= RATE_WINDOW_SAMPLES {
w.pop_front();
}
w.push_back((ts, now));
}
}
/// Bytes totales acumulados desde la creación.
pub fn total_bytes(&self) -> u64 {
self.total_bytes.load(std::sync::atomic::Ordering::Relaxed)
}
/// Bytes por segundo (rolling sobre la ventana). 0 si no hay
/// historia suficiente o si el último sample es muy viejo (>5s).
pub fn bytes_per_sec(&self) -> f64 {
let now_ms = std::time::SystemTime::now()
.duration_since(std::time::UNIX_EPOCH)
.map(|d| d.as_millis() as u64)
.unwrap_or(0);
let w = match self.rate_window.lock() {
Ok(w) => w,
Err(_) => return 0.0,
};
if w.len() < 2 {
return 0.0;
}
let last = w.back().copied().unwrap();
// Si el último sample tiene >5s, asumimos idle.
if now_ms.saturating_sub(last.0) > 5000 {
return 0.0;
}
let first = w.front().copied().unwrap();
let dt_ms = last.0.saturating_sub(first.0).max(1);
let d_bytes = last.1.saturating_sub(first.1);
(d_bytes as f64 * 1000.0) / dt_ms as f64
}
}
#[derive(Debug, Clone, Copy)]
pub struct ReplayCaps {
/// Máximo de chunks retenidos.
pub chunks: usize,
/// Máximo de bytes (sumando len de chunks). `0` = sin tope.
pub bytes: usize,
}
impl ReplayCaps {
pub fn chunks_only(chunks: usize) -> Self {
Self { chunks: chunks.max(1), bytes: 0 }
}
pub fn new(chunks: usize, bytes: usize) -> Self {
Self { chunks: chunks.max(1), bytes }
}
}
#[derive(Clone)]
pub struct FlowSender {
sender: broadcast::Sender<Arc<Vec<u8>>>,
replay: Arc<Mutex<VecDeque<Arc<Vec<u8>>>>>,
replay_caps: ReplayCaps,
meter: Arc<FlowMeter>,
}
impl FlowSender {
/// Pushea al broadcast y al replay buffer. Si no hay subscribers,
/// el broadcast::send retorna Err pero igual guardamos en replay
/// (subscribers tarde verán los chunks pasados).
pub fn send(&self, data: Arc<Vec<u8>>) {
let incoming = data.len();
let caps = self.replay_caps;
if let Ok(mut g) = self.replay.lock() {
evict_for_incoming(&mut g, caps, incoming);
g.push_back(data.clone());
}
self.meter.record(incoming as u64);
let _ = self.sender.send(data);
}
}
/// Evict los chunks más viejos para hacer espacio a un chunk entrante de
/// `incoming` bytes — el buffer post-push queda dentro de los caps.
fn evict_for_incoming(buf: &mut VecDeque<Arc<Vec<u8>>>, caps: ReplayCaps, incoming: usize) {
// 1) chunks: dejar lugar para 1 más.
while buf.len() + 1 > caps.chunks {
if buf.pop_front().is_none() {
break;
}
}
// 2) bytes (si está activado).
if caps.bytes > 0 {
let mut current: usize = buf.iter().map(|a| a.len()).sum();
while current + incoming > caps.bytes {
match buf.pop_front() {
Some(c) => current = current.saturating_sub(c.len()),
None => break,
}
}
}
}
impl std::fmt::Debug for FlowChannel {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
f.debug_struct("FlowChannel")
.field("socket_path", &self.socket_path)
.field("subscribers", &self.sender.receiver_count())
.finish()
}
}
impl FlowChannel {
/// Crea un FlowChannel atado al path `socket_path`. Si el path ya
/// existe, lo borra antes de bind (asume restart limpio).
pub fn new(socket_path: PathBuf) -> std::io::Result<Self> {
Self::with_replay_caps(socket_path, ReplayCaps::chunks_only(DEFAULT_REPLAY_CHUNKS))
}
pub fn with_replay_cap(socket_path: PathBuf, chunks: usize) -> std::io::Result<Self> {
Self::with_replay_caps(socket_path, ReplayCaps::chunks_only(chunks))
}
pub fn with_replay_caps(socket_path: PathBuf, caps: ReplayCaps) -> std::io::Result<Self> {
if socket_path.exists() {
let _ = std::fs::remove_file(&socket_path);
}
if let Some(parent) = socket_path.parent() {
let _ = std::fs::create_dir_all(parent);
}
let listener = UnixListener::bind(&socket_path)?;
let (tx, _rx_unused) = broadcast::channel::<Arc<Vec<u8>>>(BROADCAST_CAP);
let replay: Arc<Mutex<VecDeque<Arc<Vec<u8>>>>> =
Arc::new(Mutex::new(VecDeque::with_capacity(caps.chunks)));
let tx_for_accept = tx.clone();
let replay_for_accept = replay.clone();
let path_for_log = socket_path.clone();
let join = tokio::spawn(async move {
debug!(path = %path_for_log.display(), "flow channel listening");
loop {
let (mut stream, _addr) = match listener.accept().await {
Ok(p) => p,
Err(e) => {
warn!(?e, "flow channel accept failed");
return;
}
};
// Snapshot del replay buffer Y subscribe al broadcast.
// El orden es crítico: subscribe ANTES de drenar el replay
// para no perder chunks que llegan justo en el medio.
let mut rx = tx_for_accept.subscribe();
let snapshot: Vec<Arc<Vec<u8>>> = {
let g = replay_for_accept.lock().expect("replay lock");
g.iter().cloned().collect()
};
tokio::spawn(async move {
// Fase 1: drenar replay snapshot al subscriber.
for chunk in &snapshot {
if let Err(e) = stream.write_all(chunk).await {
debug!(?e, "flow subscriber dropped during replay");
return;
}
}
// Fase 2: live broadcast.
loop {
match rx.recv().await {
Ok(chunk) => {
if let Err(e) = stream.write_all(&chunk).await {
debug!(?e, "flow subscriber dropped");
return;
}
}
Err(broadcast::error::RecvError::Closed) => return,
Err(broadcast::error::RecvError::Lagged(n)) => {
warn!(skipped = n, "flow subscriber lagged");
}
}
}
});
}
});
Ok(Self {
sender: tx,
replay,
replay_caps: caps,
socket_path,
meter: Arc::new(FlowMeter::new()),
_accept_handle: AbortOnDrop(join.abort_handle()),
})
}
pub fn meter(&self) -> &FlowMeter {
&self.meter
}
/// Push un chunk al channel. Si no hay subscribers, drop silencioso.
/// Siempre se guarda en el replay buffer (con cap rotation por chunks
/// y opcionalmente por bytes).
pub fn send(&self, data: Vec<u8>) {
let incoming = data.len();
let arc = Arc::new(data);
let caps = self.replay_caps;
if let Ok(mut g) = self.replay.lock() {
evict_for_incoming(&mut g, caps, incoming);
g.push_back(arc.clone());
}
self.meter.record(incoming as u64);
let _ = self.sender.send(arc);
}
pub fn socket_path(&self) -> &Path {
&self.socket_path
}
/// Handle clone-able para que tasks externas (splitter) pushen al
/// channel sin tener ownership del FlowChannel. Cada push se guarda
/// también en el replay buffer y se contabiliza en el meter.
pub fn sender_handle(&self) -> FlowSender {
FlowSender {
sender: self.sender.clone(),
replay: self.replay.clone(),
replay_caps: self.replay_caps,
meter: self.meter.clone(),
}
}
pub fn subscriber_count(&self) -> usize {
self.sender.receiver_count()
}
}
impl Drop for FlowChannel {
fn drop(&mut self) {
// El AbortOnDrop cancela el accept loop; sólo nos queda limpiar el
// socket file.
let _ = std::fs::remove_file(&self.socket_path);
}
}
struct AbortOnDrop(AbortHandle);
impl Drop for AbortOnDrop {
fn drop(&mut self) {
self.0.abort();
}
}
/// Path canónico para un flow channel: `$XDG_RUNTIME_DIR/shuma-flow-<id>.sock`.
pub fn default_flow_socket_path(id: &str) -> PathBuf {
let base = std::env::var("XDG_RUNTIME_DIR").unwrap_or_else(|_| {
let uid = nix::unistd::getuid().as_raw();
let p = format!("/run/user/{uid}");
if std::path::Path::new(&p).exists() {
p
} else {
"/tmp".into()
}
});
PathBuf::from(base).join(format!("shuma-flow-{id}.sock"))
}
#[cfg(test)]
mod tests {
use super::*;
use tokio::io::AsyncReadExt;
use tokio::net::UnixStream;
#[tokio::test]
async fn channel_delivers_to_subscriber() {
let tmp = tempfile::tempdir().unwrap();
let path = tmp.path().join("flow.sock");
let ch = FlowChannel::new(path.clone()).unwrap();
// Subscriber se conecta.
let path_clone = path.clone();
let task = tokio::spawn(async move {
let mut stream = UnixStream::connect(&path_clone).await.unwrap();
let mut buf = vec![0u8; 64];
let n = stream.read(&mut buf).await.unwrap();
buf.truncate(n);
buf
});
// Damos tiempo al accept.
tokio::time::sleep(std::time::Duration::from_millis(50)).await;
// Hasta que haya 1 receiver_count, el send no llega.
for _ in 0..50 {
if ch.subscriber_count() >= 1 {
break;
}
tokio::time::sleep(std::time::Duration::from_millis(20)).await;
}
ch.send(b"hello-flow".to_vec());
let received = tokio::time::timeout(std::time::Duration::from_secs(2), task)
.await
.expect("timeout")
.unwrap();
assert_eq!(received, b"hello-flow");
}
#[tokio::test]
async fn replay_buffer_serves_late_subscriber() {
let tmp = tempfile::tempdir().unwrap();
let path = tmp.path().join("flow.sock");
let ch = FlowChannel::new(path.clone()).unwrap();
// Pushes ANTES de cualquier subscriber: van solo al replay.
ch.send(b"chunk-1".to_vec());
ch.send(b"chunk-2".to_vec());
ch.send(b"chunk-3".to_vec());
// Subscriber LATE — debe recibir los 3 chunks del replay.
let path_clone = path.clone();
let task = tokio::spawn(async move {
let mut stream = UnixStream::connect(&path_clone).await.unwrap();
let mut buf = vec![0u8; 256];
// Leemos hasta recibir los 3 chunks (21 bytes esperados).
let mut total = Vec::new();
for _ in 0..20 {
let n = stream.read(&mut buf).await.unwrap();
if n == 0 {
break;
}
total.extend_from_slice(&buf[..n]);
if total.len() >= 21 {
break;
}
}
total
});
let received = tokio::time::timeout(std::time::Duration::from_secs(2), task)
.await
.expect("timeout")
.unwrap();
let s = String::from_utf8_lossy(&received);
assert!(s.contains("chunk-1"), "got: {s:?}");
assert!(s.contains("chunk-2"), "got: {s:?}");
assert!(s.contains("chunk-3"), "got: {s:?}");
}
#[tokio::test]
async fn replay_evicts_by_bytes_cap() {
let tmp = tempfile::tempdir().unwrap();
let path = tmp.path().join("flow.sock");
// chunks=100 (no limita), bytes=20: deberíamos retener sólo los
// últimos chunks cuyos bytes sumen ≤ 20.
let ch = FlowChannel::with_replay_caps(path.clone(), ReplayCaps::new(100, 20)).unwrap();
ch.send(b"AAAAAAAA".to_vec()); // 8 bytes
ch.send(b"BBBBBBBB".to_vec()); // 8 → total 16
ch.send(b"CCCCCCCC".to_vec()); // 8 → total 24 > 20, evict A → 16
ch.send(b"DDDDDDDD".to_vec()); // 8 → total 24 > 20, evict B → 16
let path_clone = path.clone();
let task = tokio::spawn(async move {
let mut stream = UnixStream::connect(&path_clone).await.unwrap();
let mut buf = vec![0u8; 64];
let mut total = Vec::new();
for _ in 0..20 {
let n = stream.read(&mut buf).await.unwrap();
if n == 0 {
break;
}
total.extend_from_slice(&buf[..n]);
if total.len() >= 16 {
break;
}
}
total
});
let got = tokio::time::timeout(std::time::Duration::from_secs(2), task)
.await
.expect("timeout")
.unwrap();
let s = String::from_utf8_lossy(&got);
// Sólo C y D (los más viejos A y B fueron evicted).
assert!(!s.contains("AAAA"), "should have evicted A: {s:?}");
assert!(!s.contains("BBBB"), "should have evicted B: {s:?}");
assert!(s.contains("CCCC"), "should keep C: {s:?}");
assert!(s.contains("DDDD"), "should keep D: {s:?}");
}
#[tokio::test]
async fn drop_removes_socket() {
let tmp = tempfile::tempdir().unwrap();
let path = tmp.path().join("flow.sock");
{
let _ch = FlowChannel::new(path.clone()).unwrap();
assert!(path.exists());
}
// Después del drop, el socket file no debe quedar.
// Damos un pelín de tiempo al runtime para que el drop corra
// mientras estamos en task.
tokio::time::sleep(std::time::Duration::from_millis(10)).await;
assert!(!path.exists());
}
}
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,122 @@
//! Ring buffer en memoria para capturar stdout/stderr de comandos.
//!
//! Tamaño fijo por comando (config: `MAX_LOG_BYTES`). Cuando se llena,
//! descarta los bytes más viejos. Pensado para diagnostico rápido, no
//! para retención histórica — eso es trabajo de un journald-like aparte.
use std::sync::{Arc, Mutex};
/// Bytes máximos retenidos por comando. 64 KiB cubre logs típicos sin
/// abusar de memoria si el daemon tiene cientos de comandos vivos.
pub const MAX_LOG_BYTES: usize = 64 * 1024;
#[derive(Debug, Clone)]
pub struct LogBuf {
inner: Arc<Mutex<Inner>>,
}
#[derive(Debug)]
struct Inner {
/// Bytes raw. Cuando se acerca al cap, descartamos head para mantener
/// el tail.
buf: Vec<u8>,
cap: usize,
/// Total escrito alguna vez (no decrementado al recortar).
written_total: u64,
}
impl LogBuf {
pub fn new() -> Self {
Self::with_cap(MAX_LOG_BYTES)
}
pub fn with_cap(cap: usize) -> Self {
Self {
inner: Arc::new(Mutex::new(Inner {
buf: Vec::with_capacity(cap.min(4096)),
cap,
written_total: 0,
})),
}
}
pub fn append(&self, data: &[u8]) {
let Ok(mut g) = self.inner.lock() else { return };
g.written_total += data.len() as u64;
g.buf.extend_from_slice(data);
// Recorte cuando excede cap (con un pequeño slack para evitar
// shift en cada append). El usuario ve sólo el tail.
if g.buf.len() > g.cap + 1024 {
let drop = g.buf.len() - g.cap;
g.buf.drain(..drop);
}
}
/// Devuelve el tail de hasta `n` bytes (o todo si `n=0`).
pub fn tail(&self, n: usize) -> Vec<u8> {
let g = match self.inner.lock() {
Ok(g) => g,
Err(_) => return Vec::new(),
};
if n == 0 || n >= g.buf.len() {
return g.buf.clone();
}
g.buf[g.buf.len() - n..].to_vec()
}
/// Cuántos bytes hay actualmente en el buffer.
pub fn len(&self) -> usize {
self.inner.lock().map(|g| g.buf.len()).unwrap_or(0)
}
pub fn is_empty(&self) -> bool {
self.len() == 0
}
pub fn written_total(&self) -> u64 {
self.inner.lock().map(|g| g.written_total).unwrap_or(0)
}
}
impl Default for LogBuf {
fn default() -> Self {
Self::new()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn append_and_tail_basic() {
let lb = LogBuf::with_cap(100);
lb.append(b"hello ");
lb.append(b"world\n");
let t = lb.tail(0);
assert_eq!(t, b"hello world\n");
}
#[test]
fn cap_drops_oldest() {
let lb = LogBuf::with_cap(10);
lb.append(&[b'a'; 8]);
lb.append(&[b'b'; 8]);
// Después del recorte, debe quedar ~10 bytes pero el slack
// permite hasta 10+1024. Como pasamos slack, no se recorta aún
// en este caso (16 bytes < 10+1024). Forzamos un append grande.
lb.append(&[b'c'; 2048]);
assert!(lb.len() <= 10 + 1024);
let t = lb.tail(0);
// El tail debe contener 'c's (los más recientes).
assert!(t.iter().filter(|&&b| b == b'c').count() > 0);
}
#[test]
fn written_total_tracks_all() {
let lb = LogBuf::with_cap(10);
lb.append(b"abcdef");
lb.append(b"ghijkl");
assert_eq!(lb.written_total(), 12);
}
}
@@ -0,0 +1,383 @@
//! Persistencia del estado del WorkspaceManager.
//!
//! v1: sólo `WorkspaceSpec`s vivos. Los comandos (PIDs) NO se persisten —
//! el kernel los mata al cerrar el daemon. Sólo la *intención declarada*
//! (Workspaces creados con su spec) sobrevive a un reboot del daemon.
use crate::WorkspaceManager;
use serde::{Deserialize, Serialize};
use shuma_card::{PipelineSpec, WorkspaceId, WorkspaceSpec};
use std::path::{Path, PathBuf};
use tracing::{info, warn};
/// v2 agregó `saved_pipelines`. v3 agrega `live_pipelines`. v4 agrega
/// `stats_history` por workspace (sparkline survives daemon restart).
/// Versiones inferiores leen campos ausentes como vacío.
pub const SNAPSHOT_VERSION: u16 = 4;
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ShipoteSnapshot {
pub version: u16,
pub timestamp_ms: u64,
pub workspaces: Vec<WorkspaceEntry>,
#[serde(default)]
pub saved_pipelines: Vec<PipelineEntry>,
/// Pipelines vivos con supervisor (`restart_on_failure=true`) al
/// momento del snapshot. El daemon los relanza al restore.
#[serde(default)]
pub live_pipelines: Vec<LivePipelineEntry>,
}
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct WorkspaceEntry {
pub id: WorkspaceId,
pub spec: WorkspaceSpec,
/// Stats history persistida — cap reasonable para no inflar el JSON.
/// Sólo se guardan campos serializables (no Instant).
#[serde(default)]
pub stats_history: Vec<PersistedStats>,
}
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct PersistedStats {
pub commands_alive: u32,
pub commands_total: u32,
pub rss_bytes: Option<u64>,
pub rss_peak_bytes: Option<u64>,
pub cpu_usec: Option<u64>,
pub cpu_percent: Option<f32>,
pub cpu_cores: u32,
pub uptime_ms: u64,
}
impl From<&crate::stats::WorkspaceStats> for PersistedStats {
fn from(s: &crate::stats::WorkspaceStats) -> Self {
Self {
commands_alive: s.commands_alive,
commands_total: s.commands_total,
rss_bytes: s.rss_bytes,
rss_peak_bytes: s.rss_peak_bytes,
cpu_usec: s.cpu_usec,
cpu_percent: s.cpu_percent,
cpu_cores: s.cpu_cores,
uptime_ms: s.uptime_ms,
}
}
}
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct PipelineEntry {
pub name: String,
pub spec: PipelineSpec,
}
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct LivePipelineEntry {
pub workspace: WorkspaceId,
pub spec: PipelineSpec,
pub tap: bool,
}
impl ShipoteSnapshot {
pub fn write(&self, path: &Path) -> anyhow::Result<()> {
let bytes = serde_json::to_vec_pretty(self)?;
if let Some(parent) = path.parent() {
std::fs::create_dir_all(parent).ok();
}
let tmp = path.with_extension("tmp");
std::fs::write(&tmp, &bytes)?;
std::fs::rename(&tmp, path)?;
Ok(())
}
pub fn read(path: &Path) -> anyhow::Result<Self> {
let bytes = std::fs::read(path)?;
let snap: ShipoteSnapshot = serde_json::from_slice(&bytes)?;
// v1 y v2 son compatibles forward (v1 sin saved_pipelines lee como vec vacío).
if snap.version > SNAPSHOT_VERSION {
anyhow::bail!(
"snapshot version {} no soportada (esperada ≤ {})",
snap.version,
SNAPSHOT_VERSION
);
}
Ok(snap)
}
}
/// Path canónico del snapshot: `$XDG_STATE_HOME/shuma/state.json`,
/// fallback `$HOME/.local/state/shuma/state.json`,
/// fallback `/tmp/shuma-state-$UID.json`.
pub fn default_snapshot_path() -> PathBuf {
if let Ok(state) = std::env::var("XDG_STATE_HOME") {
return PathBuf::from(state).join("shuma/state.json");
}
if let Ok(home) = std::env::var("HOME") {
return PathBuf::from(home).join(".local/state/shuma/state.json");
}
let uid = nix::unistd::getuid().as_raw();
PathBuf::from(format!("/tmp/shuma-state-{uid}.json"))
}
fn now_ms() -> u64 {
std::time::SystemTime::now()
.duration_since(std::time::UNIX_EPOCH)
.map(|d| d.as_millis() as u64)
.unwrap_or(0)
}
impl WorkspaceManager {
/// Toma snapshot del estado actual.
pub async fn snapshot(&self) -> ShipoteSnapshot {
const PERSIST_STATS_CAP: usize = 16;
let g = self.inner.lock().await;
let workspaces = g
.workspaces
.iter()
.map(|(id, ws)| {
// Persist sólo los últimos N samples — el resto crece
// y el JSON se infla.
let take = ws.stats_history.len().min(PERSIST_STATS_CAP);
let skip = ws.stats_history.len() - take;
let stats_history: Vec<PersistedStats> = ws
.stats_history
.iter()
.skip(skip)
.map(PersistedStats::from)
.collect();
WorkspaceEntry {
id: *id,
spec: ws.spec.clone(),
stats_history,
}
})
.collect();
let saved_pipelines = g
.saved_pipelines
.iter()
.map(|(name, spec)| PipelineEntry {
name: name.clone(),
spec: spec.clone(),
})
.collect();
// Pipelines vivos con supervisor — preserva la intención. Los
// pids/sockets/discernments son ephemeral y se regeneran al
// restore (relaunch desde cero).
let live_pipelines = g
.pipeline_supervisors
.values()
.map(|sup| LivePipelineEntry {
workspace: sup.workspace,
spec: sup.spec.clone(),
tap: sup.tap,
})
.collect();
ShipoteSnapshot {
version: SNAPSHOT_VERSION,
timestamp_ms: now_ms(),
workspaces,
saved_pipelines,
live_pipelines,
}
}
/// Escribe snapshot a disco. Si `is_dirty()` es false **y** el path
/// existe (snapshot previo válido), skip la escritura.
pub async fn save_snapshot(&self, path: &Path) -> anyhow::Result<()> {
if !self.is_dirty() && path.exists() {
info!(path = %path.display(), "snapshot SKIPPED (clean)");
return Ok(());
}
let snap = self.snapshot().await;
snap.write(path)?;
// Clear dirty: lo que está en disco es el current state.
self.dirty
.store(false, std::sync::atomic::Ordering::Relaxed);
info!(path = %path.display(), workspaces = snap.workspaces.len(), "snapshot saved");
Ok(())
}
/// Carga snapshot desde disco y restaura los Workspaces + saved
/// pipelines. Devuelve los `live_pipelines` para que el caller
/// (daemon) los relance — no podemos relanzarlos desde acá porque
/// `run_pipeline` necesita `Incarnator` + `DiscernPipeline`.
/// Errores no-fatales (workspaces inválidos) se loguean y se saltan.
pub async fn restore_snapshot(
self: &std::sync::Arc<Self>,
path: &Path,
) -> anyhow::Result<RestoreOutcome> {
let snap = match ShipoteSnapshot::read(path) {
Ok(s) => s,
Err(e) => {
warn!(?e, path = %path.display(), "no snapshot — start fresh");
return Ok(RestoreOutcome::default());
}
};
let mut out = RestoreOutcome::default();
for entry in snap.workspaces {
// v2+: reusamos el id original así clients que tracking
// workspace_id no se rompen al restart.
let label = entry.spec.label.clone();
let id = entry.id;
let history = entry.stats_history;
match self.create_with_id(id, entry.spec).await {
Ok(_) => {
out.workspaces_restored += 1;
// Hidratar history persistida. Convertimos
// PersistedStats → WorkspaceStats (perdemos
// los campos no serializables como `source`).
if !history.is_empty() {
let mut g = self.inner.lock().await;
if let Some(ws) = g.workspaces.get_mut(&id) {
for ps in history {
ws.stats_history.push_back(crate::stats::WorkspaceStats {
commands_alive: ps.commands_alive,
commands_total: ps.commands_total,
rss_bytes: ps.rss_bytes,
rss_peak_bytes: ps.rss_peak_bytes,
cpu_usec: ps.cpu_usec,
cpu_percent: ps.cpu_percent,
cpu_cores: ps.cpu_cores,
source: "persisted".into(),
uptime_ms: ps.uptime_ms,
});
}
}
}
}
Err(e) => warn!(?e, %label, "skipped workspace en restore"),
}
}
for entry in snap.saved_pipelines {
self.save_pipeline(entry.name, entry.spec).await;
out.saved_pipelines_restored += 1;
}
out.live_pipelines = snap.live_pipelines;
// Restore no cuenta como mutación — lo que está en disco es lo
// que acabamos de cargar. Sin esto, el próximo SIGTERM siempre
// re-escribiría aunque no hubiese cambios reales.
self.dirty
.store(false, std::sync::atomic::Ordering::Relaxed);
info!(
workspaces = out.workspaces_restored,
saved_pipelines = out.saved_pipelines_restored,
live_pipelines = out.live_pipelines.len(),
"snapshot restored"
);
Ok(out)
}
}
/// Lo que el caller del restore obtiene. Las `live_pipelines` requieren
/// `Incarnator + DiscernPipeline` para relanzarlas → el caller las
/// procesa (típicamente el daemon).
#[derive(Debug, Default)]
pub struct RestoreOutcome {
pub workspaces_restored: usize,
pub saved_pipelines_restored: usize,
pub live_pipelines: Vec<LivePipelineEntry>,
}
#[cfg(test)]
mod tests {
use super::*;
use crate::WorkspaceManager;
use ente_incarnate::IncarnatorConfig;
use shuma_card::{ExitPolicy, WorkspaceSpec};
use std::sync::Arc;
fn sample_ws(label: &str) -> WorkspaceSpec {
WorkspaceSpec {
label: label.into(),
soma: Default::default(),
permissions: Default::default(),
ttl: None,
flow_dirs: vec![],
on_exit: ExitPolicy::Reap,
quota_enforce: Default::default(),
}
}
#[tokio::test]
async fn roundtrip_snapshot_preserves_ulids() {
let tmp = tempfile::tempdir().unwrap();
let path = tmp.path().join("state.json");
let mgr1 = Arc::new(WorkspaceManager::new(IncarnatorConfig::default()));
let (id1, _) = mgr1.create(sample_ws("a")).await.unwrap();
let (id2, _) = mgr1.create(sample_ws("b")).await.unwrap();
mgr1.save_snapshot(&path).await.unwrap();
let mgr2 = Arc::new(WorkspaceManager::new(IncarnatorConfig::default()));
let out = mgr2.restore_snapshot(&path).await.unwrap();
assert_eq!(out.workspaces_restored, 2);
let listed = mgr2.list().await;
let restored_ids: std::collections::HashSet<_> = listed.iter().map(|s| s.id).collect();
assert!(restored_ids.contains(&id1));
assert!(restored_ids.contains(&id2));
}
#[tokio::test]
async fn save_snapshot_skips_when_clean() {
let tmp = tempfile::tempdir().unwrap();
let path = tmp.path().join("state.json");
let mgr = Arc::new(WorkspaceManager::new(IncarnatorConfig::default()));
let _ = mgr.create(sample_ws("dirty-test")).await.unwrap();
assert!(mgr.is_dirty(), "create debería marcar dirty");
mgr.save_snapshot(&path).await.unwrap();
assert!(!mgr.is_dirty(), "save_snapshot debería limpiar dirty");
let mtime1 = std::fs::metadata(&path).unwrap().modified().unwrap();
// Esperamos un pelín para que mtime cambie si fuera re-escrito.
tokio::time::sleep(std::time::Duration::from_millis(20)).await;
// Segundo save sin mutación → skip.
mgr.save_snapshot(&path).await.unwrap();
let mtime2 = std::fs::metadata(&path).unwrap().modified().unwrap();
assert_eq!(mtime1, mtime2, "skip cuando clean — mtime no cambia");
}
#[tokio::test]
async fn snapshot_includes_saved_pipelines() {
use shuma_card::{CommandRef, DiscernPolicy, PipelineSpec};
let tmp = tempfile::tempdir().unwrap();
let path = tmp.path().join("state.json");
let mgr1 = Arc::new(WorkspaceManager::new(IncarnatorConfig::default()));
let (ws_id, _) = mgr1.create(sample_ws("ws")).await.unwrap();
let spec = PipelineSpec {
label: "echo-cat".into(),
workspace: ws_id,
nodes: vec![CommandRef {
label: "n1".into(),
payload: brahman_card::Payload::Native {
exec: "/bin/echo".into(),
argv: vec!["hi".into()],
envp: vec![],
},
soma: Default::default(),
flows: Default::default(),
supervision: brahman_card::Supervision::OneShot,
}],
edges: vec![],
discern: DiscernPolicy::default(),
restart_on_failure: false,
restart_backoff_ms: 200,
restart_max_backoff_ms: 30_000,
restart_max: 0,
};
mgr1.save_pipeline("daily".into(), spec).await;
mgr1.save_snapshot(&path).await.unwrap();
let mgr2 = Arc::new(WorkspaceManager::new(IncarnatorConfig::default()));
mgr2.restore_snapshot(&path).await.unwrap();
let saved = mgr2.list_saved_pipelines().await;
assert_eq!(saved, vec!["daily".to_string()]);
let got = mgr2.get_saved_pipeline("daily").await.expect("saved");
assert_eq!(got.label, "echo-cat");
}
#[test]
fn default_path_ends_with_state_json() {
let p = default_snapshot_path();
assert!(p.to_string_lossy().ends_with("state.json"));
}
}
@@ -0,0 +1,808 @@
//! Pipeline runtime: encadena nodos con pipes y opcionalmente intercepta
//! cada flow para discernir su contenido.
//!
//! Cada nodo se encarna via [`ente_incarnate::Incarnator`] — eso significa
//! que **cada comando puede tener su propio SomaSpec** (namespaces, cgroup,
//! rlimits) heredado del workspace. La conexión stdin↔stdout se hace con
//! `pipe2(2)` + `ChildStdio` declarativo: el callback de clone(2) hace los
//! `dup2` pre-execve sin romper la regla async-signal-safe.
use crate::CoreError;
use brahman_card::Payload;
use ente_incarnate::{ChildStdio, Incarnator};
use nix::fcntl::OFlag;
use nix::unistd::pipe2;
use shuma_card::PipelineSpec;
use shuma_discern::{DiscernPipeline, Discernment, Hint};
use std::os::fd::{AsRawFd, IntoRawFd, RawFd};
use std::sync::Arc;
use tokio::io::unix::AsyncFd;
use tokio::io::Interest;
use tracing::{debug, info, warn};
use ulid::Ulid;
/// Resultado de lanzar un pipeline.
#[derive(Debug)]
pub struct PipelineLaunch {
pub pipeline: Ulid,
pub command_pids: Vec<(String, i32)>,
/// Discernments por edge, en el mismo orden que `spec.edges`.
pub edge_discernments: Vec<EdgeDiscernment>,
}
#[derive(Debug, Clone)]
pub struct EdgeDiscernment {
pub from_label: String,
pub from_output: String,
pub to_label: String,
pub to_input: String,
pub discernment: Option<Discernment>,
/// Path del Unix socket donde otros módulos pueden suscribirse al
/// stream replicado por este edge. `None` cuando tap=false (no hay
/// data plane porque no hay sampling).
pub flow_socket: Option<std::path::PathBuf>,
}
/// Lanza un pipeline conectando nodos por stdin/stdout. Cada nodo se
/// encarna via `Incarnator` (con o sin namespacing según su SomaSpec).
///
/// Soporta:
/// - Pipeline lineal (1 producer → 1 consumer).
/// - **Fan-out** (1 producer → N consumers): shuma interpone un
/// splitter que duplica bytes a cada destino. Cuando `tap=true`, el
/// splitter además samplea para discernir.
/// - Múltiples predecessors por nodo NO se soporta aún (fan-in): sólo se
/// honra el primer edge entrante.
pub async fn run_pipeline(
spec: &PipelineSpec,
workspace_label: &str,
tap: bool,
discerner: Arc<DiscernPipeline>,
incarnator: Arc<Incarnator>,
manager: Option<Arc<crate::WorkspaceManager>>,
) -> Result<PipelineLaunch, CoreError> {
spec.validate()?;
let n = spec.nodes.len();
info!(
nodes = n,
edges = spec.edges.len(),
tap,
"launching pipeline (incarnated)"
);
// Pre-compute grafo:
// - `consumers[i]` = índices de edges salientes de `i`.
// - `predecessors[j]` = índices de edges entrantes a `j`.
let mut consumers: Vec<Vec<usize>> = vec![Vec::new(); n];
let mut predecessors: Vec<Vec<usize>> = vec![Vec::new(); n];
for (idx, e) in spec.edges.iter().enumerate() {
consumers[e.from].push(idx);
predecessors[e.to].push(idx);
}
// Por cada edge: par (r_to_consumer, w_from_producer_side).
// El consumer recibe r_to_consumer; el producer escribe a w_from_producer_side
// (directa o vía splitter).
let mut edge_r: Vec<RawFd> = vec![-1; spec.edges.len()];
let mut edge_w: Vec<RawFd> = vec![-1; spec.edges.len()];
for i in 0..spec.edges.len() {
let (r, w) = pipe2(OFlag::O_CLOEXEC).map_err(|e| {
CoreError::Incarnate(ente_incarnate::IncarnateError::Pipe(e))
})?;
edge_r[i] = r.into_raw_fd();
edge_w[i] = w.into_raw_fd();
}
let mut consumer_stdin_fd: Vec<Option<RawFd>> = vec![None; n];
let mut producer_stdout_fd: Vec<Option<RawFd>> = vec![None; n];
let mut splitter_specs: Vec<SplitterSpec> = Vec::new();
let mut merger_specs: Vec<MergerSpec> = Vec::new();
// Stdout del producer: directo a edge_w[único] si tiene 1 consumer y NO tap;
// sino, pipe propio que va al splitter task.
for i in 0..n {
if consumers[i].is_empty() {
continue;
}
if consumers[i].len() == 1 && !tap {
producer_stdout_fd[i] = Some(edge_w[consumers[i][0]]);
continue;
}
// Splitter: pipe propio para el productor → splitter lee y replica a edge_w[*].
let (prod_r, prod_w) = pipe2(OFlag::O_CLOEXEC).map_err(|e| {
CoreError::Incarnate(ente_incarnate::IncarnateError::Pipe(e))
})?;
producer_stdout_fd[i] = Some(prod_w.into_raw_fd());
let prod_r_fd = prod_r.into_raw_fd();
let mut consumer_writes: Vec<RawFd> = Vec::with_capacity(consumers[i].len());
let mut edge_meta: Vec<EdgeMeta> = Vec::with_capacity(consumers[i].len());
for edge_idx in &consumers[i] {
let edge = &spec.edges[*edge_idx];
consumer_writes.push(edge_w[*edge_idx]);
edge_meta.push(EdgeMeta {
from_label: spec.nodes[edge.from].label.clone(),
from_output: edge.from_output.clone(),
to_label: spec.nodes[edge.to].label.clone(),
to_input: edge.to_input.clone(),
});
}
splitter_specs.push(SplitterSpec {
producer_r_fd: prod_r_fd,
consumer_w_fds: consumer_writes,
edges: edge_meta,
tap,
sample_bytes: spec.discern.sample_bytes,
max_bytes_per_sec: spec.discern.max_bytes_per_sec,
});
}
// Stdin del consumer: edge_r[único] si tiene 1 predecessor; sino, merger.
for j in 0..n {
match predecessors[j].len() {
0 => {}
1 => {
consumer_stdin_fd[j] = Some(edge_r[predecessors[j][0]]);
}
_ => {
// Merger: lee de N edge_r y escribe a un nuevo pipe cuyo
// read end es el stdin del consumer.
let (cons_r, cons_w) = pipe2(OFlag::O_CLOEXEC).map_err(|e| {
CoreError::Incarnate(ente_incarnate::IncarnateError::Pipe(e))
})?;
consumer_stdin_fd[j] = Some(cons_r.into_raw_fd());
let inputs: Vec<RawFd> = predecessors[j]
.iter()
.map(|eidx| edge_r[*eidx])
.collect();
merger_specs.push(MergerSpec {
producer_r_fds: inputs,
consumer_w_fd: cons_w.into_raw_fd(),
});
}
}
}
// Encarnamos cada nodo con su stdin/stdout fd asignado.
let mut pids = Vec::with_capacity(n);
for (i, node) in spec.nodes.iter().enumerate() {
match &node.payload {
Payload::Native { .. } | Payload::Legacy { .. } => {}
_ => {
return Err(CoreError::Incarnate(
ente_incarnate::IncarnateError::NonExecutablePayload,
))
}
}
let card = node.to_card(i, workspace_label)?;
let stdio = ChildStdio {
stdin_fd: consumer_stdin_fd[i],
stdout_fd: producer_stdout_fd[i],
stderr_fd: None,
};
let outcome = incarnator
.incarnate_with(&card, stdio)
.map_err(CoreError::Incarnate)?;
let pid = outcome.pid;
pids.push((node.label.clone(), pid.as_raw()));
debug!(label = %node.label, pid = pid.as_raw(), "node incarnated");
}
let pipeline_id_for_flows = Ulid::new();
// Si tap=true, creamos un FlowChannel por edge para el data plane.
// Cada splitter pushea al sender del channel correspondiente.
let pipeline_id = pipeline_id_for_flows;
let mut flow_channels: Vec<crate::flow_channel::FlowChannel> = Vec::new();
let mut splitter_channels: Vec<Vec<Option<crate::flow_channel::FlowSender>>> =
Vec::with_capacity(splitter_specs.len());
let mut edge_socket_for_splitter: Vec<Vec<Option<std::path::PathBuf>>> = Vec::new();
for s in &splitter_specs {
let mut senders_per_edge = Vec::with_capacity(s.edges.len());
let mut paths_per_edge = Vec::with_capacity(s.edges.len());
for (i, _em) in s.edges.iter().enumerate() {
if !s.tap {
senders_per_edge.push(None);
paths_per_edge.push(None);
continue;
}
// Socket name = pipeline_id full (26 chars ULID) + edge_idx.
// ULID es único globalmente → cero colisiones entre runs.
// Edge_idx desambigua múltiples sockets del mismo pipeline.
// No incluimos from_label en el name (puede tener chars que
// no van en paths Unix — los hints van en `EdgeDiscernment`).
let id = format!("{}-{}", pipeline_id, i);
let mut socket = crate::flow_channel::default_flow_socket_path(&id);
// Fallback: si el path existe (raro — daemon crashed sin
// cleanup), agregar suffix numérico hasta encontrar libre.
let mut suffix = 1u32;
while socket.exists() {
let alt = format!("{id}-{suffix}");
socket = crate::flow_channel::default_flow_socket_path(&alt);
suffix += 1;
if suffix > 1000 {
warn!(orig = id, "flow socket collision: 1000 retries — using as-is");
break;
}
}
match crate::flow_channel::FlowChannel::with_replay_caps(
socket.clone(),
crate::flow_channel::ReplayCaps::new(spec.discern.replay_chunks, spec.discern.replay_bytes),
) {
Ok(fc) => {
senders_per_edge.push(Some(fc.sender_handle()));
paths_per_edge.push(Some(socket));
flow_channels.push(fc);
}
Err(e) => {
warn!(?e, "flow channel new failed");
senders_per_edge.push(None);
paths_per_edge.push(None);
}
}
}
splitter_channels.push(senders_per_edge);
edge_socket_for_splitter.push(paths_per_edge);
}
// Registramos los flow_channels en el manager AHORA, antes de await
// las tasks. Esto permite que clientes externos hagan `flow list` y
// se suscriban mientras el pipeline aún produce data.
if let Some(mgr) = &manager {
if !flow_channels.is_empty() {
let drained: Vec<crate::flow_channel::FlowChannel> = flow_channels.drain(..).collect();
mgr.retain_pipeline_flows(pipeline_id, drained).await;
}
}
// Spawn mergers + splitters después del incarnate. Cada task posee
// sus fds y los cierra al terminar (via Drop de OwnedFd).
let mut merger_handles: Vec<tokio::task::JoinHandle<()>> = Vec::new();
for m in merger_specs {
merger_handles.push(spawn_merger(m));
}
let mut tap_handles: Vec<SplitterHandle> = Vec::new();
for (s, senders) in splitter_specs.into_iter().zip(splitter_channels.into_iter()) {
tap_handles.push(spawn_splitter(s, discerner.clone(), senders));
}
let mut edge_discernments = Vec::new();
for (h, paths) in tap_handles.into_iter().zip(edge_socket_for_splitter.into_iter()) {
match h.handle.await {
Ok(eds) => {
for (mut ed, path) in eds.into_iter().zip(paths.into_iter()) {
ed.flow_socket = path;
edge_discernments.push(ed);
}
}
Err(e) => warn!(?e, "splitter handle joined with error"),
}
}
for h in merger_handles {
if let Err(e) = h.await {
warn!(?e, "merger handle joined with error");
}
}
Ok(PipelineLaunch {
pipeline: pipeline_id,
command_pids: pids,
edge_discernments,
})
}
#[allow(dead_code)]
fn short_ulid(u: &Ulid) -> String {
let s = u.to_string();
s[s.len() - 6..].to_string()
}
#[derive(Debug, Clone)]
struct EdgeMeta {
from_label: String,
from_output: String,
to_label: String,
to_input: String,
}
struct SplitterSpec {
producer_r_fd: RawFd,
consumer_w_fds: Vec<RawFd>,
edges: Vec<EdgeMeta>,
tap: bool,
sample_bytes: usize,
/// Rate-limit en bytes/s (0 = sin limit). Tras cada chunk de `n`
/// bytes, splitter sleeps `n / max_bytes_per_sec` segundos.
max_bytes_per_sec: u64,
}
struct SplitterHandle {
handle: tokio::task::JoinHandle<Vec<EdgeDiscernment>>,
}
struct MergerSpec {
producer_r_fds: Vec<RawFd>,
consumer_w_fd: RawFd,
}
fn spawn_merger(spec: MergerSpec) -> tokio::task::JoinHandle<()> {
for fd in &spec.producer_r_fds {
set_nonblocking(*fd);
}
set_nonblocking(spec.consumer_w_fd);
// Patrón: una task lectora por cada producer reenvía bytes a un mpsc.
// El merger principal consume del mpsc y escribe al consumer.
// Esto evita el "block en reader idle" del enfoque round-robin sobre
// AsyncFd::ready() (los readers idle nunca dejan turno).
tokio::spawn(async move {
let (tx, mut rx) = tokio::sync::mpsc::channel::<Vec<u8>>(32);
let nr = spec.producer_r_fds.len();
for fd in spec.producer_r_fds {
let tx = tx.clone();
tokio::spawn(async move {
// SAFETY: ownership transferida.
let owned = unsafe { std::os::fd::OwnedFd::from_raw_fd_compat(fd) };
let r = match AsyncFd::with_interest(owned, Interest::READABLE) {
Ok(a) => a,
Err(e) => {
warn!(?e, "merger reader AsyncFd");
return;
}
};
let mut buf = [0u8; 4096];
loop {
match async_read(&r, &mut buf).await {
Ok(0) => break,
Ok(n) => {
if tx.send(buf[..n].to_vec()).await.is_err() {
break;
}
}
Err(_) => break,
}
}
// Drop de tx → cuando todos los readers cerraron, el rx
// recibe None y el merger termina.
});
}
drop(tx); // sólo los reader tasks tienen sus clones ahora.
// SAFETY: ownership transferida al task.
let w_owned = unsafe { std::os::fd::OwnedFd::from_raw_fd_compat(spec.consumer_w_fd) };
let w = match AsyncFd::with_interest(w_owned, Interest::WRITABLE) {
Ok(a) => a,
Err(e) => {
warn!(?e, "merger AsyncFd w");
return;
}
};
let mut total: u64 = 0;
while let Some(chunk) = rx.recv().await {
if async_write_all(&w, &chunk).await.is_err() {
return;
}
total += chunk.len() as u64;
}
debug!(bytes = total, readers = nr, "merger finished");
})
}
fn spawn_splitter(
spec: SplitterSpec,
discerner: Arc<DiscernPipeline>,
edge_senders: Vec<Option<crate::flow_channel::FlowSender>>,
) -> SplitterHandle {
set_nonblocking(spec.producer_r_fd);
for fd in &spec.consumer_w_fds {
set_nonblocking(*fd);
}
let handle = tokio::spawn(async move {
// SAFETY: ownership transferida al task.
let r_owned = unsafe { std::os::fd::OwnedFd::from_raw_fd_compat(spec.producer_r_fd) };
let r = match AsyncFd::with_interest(r_owned, Interest::READABLE) {
Ok(a) => a,
Err(e) => {
warn!(?e, "splitter AsyncFd r");
return Vec::new();
}
};
let mut writers: Vec<AsyncFd<std::os::fd::OwnedFd>> = Vec::with_capacity(spec.consumer_w_fds.len());
for fd in spec.consumer_w_fds {
let owned = unsafe { std::os::fd::OwnedFd::from_raw_fd_compat(fd) };
match AsyncFd::with_interest(owned, Interest::WRITABLE) {
Ok(a) => writers.push(a),
Err(e) => warn!(?e, "splitter AsyncFd w"),
}
}
let mut sample: Vec<u8> = Vec::with_capacity(spec.sample_bytes);
let mut buf = [0u8; 4096];
let mut total: u64 = 0;
let mut eof = false;
let mut bucket = if spec.max_bytes_per_sec > 0 {
Some(TokenBucket::new(spec.max_bytes_per_sec))
} else {
None
};
// Fase 1: sampling (sólo si tap=true) + replicación.
while !eof && (spec.tap && sample.len() < spec.sample_bytes) {
let n = match async_read(&r, &mut buf).await {
Ok(0) => { eof = true; 0 }
Ok(n) => n,
Err(e) => { warn!(?e, "splitter read"); break; }
};
if n == 0 { break; }
if spec.tap {
let take = n.min(spec.sample_bytes - sample.len());
sample.extend_from_slice(&buf[..take]);
}
// Token bucket: reserva ANTES de broadcast — si hay debt,
// sleep antes de mandar al subscriber.
if let Some(b) = bucket.as_mut() {
let wait = b.reserve(n as u64);
if !wait.is_zero() {
tokio::time::sleep(wait).await;
}
}
broadcast_chunk(&writers, &edge_senders, &buf[..n]).await;
total += n as u64;
}
let d = if spec.tap {
discerner.discern(&sample, &Hint { path: None, size_total: None })
} else {
None
};
// Fase 2: replicación pura.
while !eof {
let n = match async_read(&r, &mut buf).await {
Ok(0) => { eof = true; 0 }
Ok(n) => n,
Err(_) => break,
};
if n == 0 { break; }
if let Some(b) = bucket.as_mut() {
let wait = b.reserve(n as u64);
if !wait.is_zero() {
tokio::time::sleep(wait).await;
}
}
broadcast_chunk(&writers, &edge_senders, &buf[..n]).await;
total += n as u64;
}
debug!(bytes = total, consumers = writers.len(), "splitter finished");
// Mismo discernment para todos los edges del splitter (es el mismo
// stream replicado). Devolvemos N entries (una por edge) para que
// la UI/CLI los liste todos. flow_socket lo rellena el caller.
spec.edges
.into_iter()
.map(|em| EdgeDiscernment {
from_label: em.from_label,
from_output: em.from_output,
to_label: em.to_label,
to_input: em.to_input,
discernment: d.clone(),
flow_socket: None,
})
.collect()
});
SplitterHandle { handle }
}
/// Token-bucket real con capacidad de burst.
/// - `rate_bps`: tokens (bytes) por segundo de refill.
/// - `capacity`: máx tokens acumulables. Default = 1 segundo de rate.
/// - `tokens`: tokens disponibles (puede negativos para "debt").
/// - `last_refill`: para calcular cuántos refill desde la última call.
struct TokenBucket {
rate_bps: u64,
capacity: u64,
tokens: f64,
last_refill: std::time::Instant,
}
impl TokenBucket {
fn new(rate_bps: u64) -> Self {
Self {
rate_bps,
capacity: rate_bps, // 1 second worth of burst.
tokens: rate_bps as f64,
last_refill: std::time::Instant::now(),
}
}
/// Refill desde la última call según wall time. Reserva `cost`
/// tokens; si no alcanza, retorna el sleep necesario.
fn reserve(&mut self, cost: u64) -> std::time::Duration {
let now = std::time::Instant::now();
let elapsed_secs = now.duration_since(self.last_refill).as_secs_f64();
self.tokens = (self.tokens + elapsed_secs * self.rate_bps as f64)
.min(self.capacity as f64);
self.last_refill = now;
self.tokens -= cost as f64;
if self.tokens >= 0.0 {
std::time::Duration::ZERO
} else {
// Debt: tiempo para recuperar a 0 tokens.
let secs_needed = -self.tokens / self.rate_bps as f64;
std::time::Duration::from_secs_f64(secs_needed)
}
}
}
async fn broadcast_chunk(
writers: &[AsyncFd<std::os::fd::OwnedFd>],
edge_senders: &[Option<crate::flow_channel::FlowSender>],
data: &[u8],
) {
// Internal pipes a los consumers del pipeline.
for w in writers {
let _ = async_write_all(w, data).await;
}
// Externos: broadcast a subscribers vía FlowChannel.
// Cada edge tiene su propio sender (mismo data — el sample/discernment
// viaja por broadcast separados para que un subscriber por edge vea su
// stream específico).
if edge_senders.iter().any(|s| s.is_some()) {
let shared = std::sync::Arc::new(data.to_vec());
for s in edge_senders {
if let Some(s) = s {
let _ = s.send(shared.clone());
}
}
}
}
async fn async_read(
afd: &AsyncFd<std::os::fd::OwnedFd>,
buf: &mut [u8],
) -> std::io::Result<usize> {
loop {
let mut guard = afd.readable().await?;
let fd = afd.as_raw_fd();
// SAFETY: lectura sobre fd válido propiedad del AsyncFd.
let r = unsafe { libc::read(fd, buf.as_mut_ptr() as *mut _, buf.len()) };
if r >= 0 {
return Ok(r as usize);
}
let err = std::io::Error::last_os_error();
if err.kind() == std::io::ErrorKind::WouldBlock {
guard.clear_ready();
continue;
}
return Err(err);
}
}
async fn async_write_all(
afd: &AsyncFd<std::os::fd::OwnedFd>,
mut buf: &[u8],
) -> std::io::Result<()> {
while !buf.is_empty() {
let mut guard = afd.writable().await?;
let fd = afd.as_raw_fd();
// SAFETY: escritura sobre fd válido propiedad del AsyncFd.
let r = unsafe { libc::write(fd, buf.as_ptr() as *const _, buf.len()) };
if r > 0 {
buf = &buf[r as usize..];
continue;
}
if r == 0 {
return Err(std::io::Error::new(
std::io::ErrorKind::WriteZero,
"write 0",
));
}
let err = std::io::Error::last_os_error();
if err.kind() == std::io::ErrorKind::WouldBlock {
guard.clear_ready();
continue;
}
return Err(err);
}
Ok(())
}
fn set_nonblocking(fd: RawFd) {
// SAFETY: fcntl con F_SETFL es seguro para fds válidos.
unsafe {
let flags = libc::fcntl(fd, libc::F_GETFL, 0);
if flags >= 0 {
libc::fcntl(fd, libc::F_SETFL, flags | libc::O_NONBLOCK);
}
}
}
// Extension trait para abstraer la API de OwnedFd entre versiones (compat).
trait OwnedFdFromRawCompat: Sized {
unsafe fn from_raw_fd_compat(fd: RawFd) -> Self;
}
impl OwnedFdFromRawCompat for std::os::fd::OwnedFd {
unsafe fn from_raw_fd_compat(fd: RawFd) -> Self {
use std::os::fd::FromRawFd;
// SAFETY: el caller transfiere ownership de `fd` a la `OwnedFd`.
unsafe { std::os::fd::OwnedFd::from_raw_fd(fd) }
}
}
// Re-export para que el unused warning del AsRawFd se calle si no se usa.
#[allow(dead_code)]
fn _keep_raw(_: &dyn AsRawFd) {}
#[cfg(test)]
mod tests {
use super::*;
use brahman_card::Payload;
use ente_incarnate::IncarnatorConfig;
use shuma_card::{CommandRef, DiscernPolicy, FlowEdge, PipelineSpec, WorkspaceId};
fn cmd(label: &str, exec: &str, argv: &[&str]) -> CommandRef {
CommandRef {
label: label.into(),
payload: Payload::Native {
exec: exec.into(),
argv: argv.iter().map(|s| s.to_string()).collect(),
envp: vec![],
},
soma: Default::default(),
flows: Default::default(),
supervision: brahman_card::Supervision::OneShot,
}
}
#[tokio::test]
async fn pipeline_isolated_echo_to_cat_runs() {
let spec = PipelineSpec {
label: "echo-cat".into(),
workspace: WorkspaceId::new(),
nodes: vec![
cmd("p1", "/bin/echo", &["hola pipeline aislado"]),
cmd("p2", "/bin/cat", &[]),
],
edges: vec![FlowEdge {
from: 0,
from_output: "stdout".into(),
to: 1,
to_input: "stdin".into(),
}],
discern: DiscernPolicy::default(),
restart_on_failure: false,
restart_backoff_ms: 200,
restart_max_backoff_ms: 30_000,
restart_max: 0,
};
let disc = Arc::new(DiscernPipeline::default_pipeline());
let inc = Arc::new(Incarnator::new(IncarnatorConfig::default()));
let launch = run_pipeline(&spec, "ws", false, disc, inc, None).await.unwrap();
assert_eq!(launch.command_pids.len(), 2);
// Cosecha.
for (_, pid) in &launch.command_pids {
let _ = nix::sys::wait::waitpid(nix::unistd::Pid::from_raw(*pid), None);
}
}
#[tokio::test]
async fn pipeline_fanin_two_to_one() {
// 2 productores → 1 consumer (cat). El merger multiplexa.
let spec = PipelineSpec {
label: "fanin".into(),
workspace: WorkspaceId::new(),
nodes: vec![
cmd("p1", "/bin/echo", &["from-p1"]),
cmd("p2", "/bin/echo", &["from-p2"]),
cmd("c", "/bin/cat", &[]),
],
edges: vec![
FlowEdge {
from: 0,
from_output: "stdout".into(),
to: 2,
to_input: "stdin".into(),
},
FlowEdge {
from: 1,
from_output: "stdout".into(),
to: 2,
to_input: "stdin".into(),
},
],
discern: DiscernPolicy::default(),
restart_on_failure: false,
restart_backoff_ms: 200,
restart_max_backoff_ms: 30_000,
restart_max: 0,
};
let disc = Arc::new(DiscernPipeline::default_pipeline());
let inc = Arc::new(Incarnator::new(IncarnatorConfig::default()));
let launch = run_pipeline(&spec, "ws", false, disc, inc, None).await.unwrap();
assert_eq!(launch.command_pids.len(), 3);
for (_, pid) in &launch.command_pids {
let _ = nix::sys::wait::waitpid(nix::unistd::Pid::from_raw(*pid), None);
}
}
#[tokio::test]
async fn pipeline_fanout_one_to_two() {
// 1 productor (echo) → 2 consumers (wc -c). Splitter replica.
let spec = PipelineSpec {
label: "fanout".into(),
workspace: WorkspaceId::new(),
nodes: vec![
cmd("p", "/bin/echo", &["fanout-test"]),
cmd("c1", "/bin/cat", &[]),
cmd("c2", "/bin/cat", &[]),
],
edges: vec![
FlowEdge {
from: 0,
from_output: "stdout".into(),
to: 1,
to_input: "stdin".into(),
},
FlowEdge {
from: 0,
from_output: "stdout".into(),
to: 2,
to_input: "stdin".into(),
},
],
discern: DiscernPolicy::default(),
restart_on_failure: false,
restart_backoff_ms: 200,
restart_max_backoff_ms: 30_000,
restart_max: 0,
};
let disc = Arc::new(DiscernPipeline::default_pipeline());
let inc = Arc::new(Incarnator::new(IncarnatorConfig::default()));
let launch = run_pipeline(&spec, "ws", false, disc, inc, None).await.unwrap();
assert_eq!(launch.command_pids.len(), 3);
for (_, pid) in &launch.command_pids {
let _ = nix::sys::wait::waitpid(nix::unistd::Pid::from_raw(*pid), None);
}
}
#[tokio::test]
async fn pipeline_isolated_with_tap_captures_discernment() {
let spec = PipelineSpec {
label: "json-cat".into(),
workspace: WorkspaceId::new(),
nodes: vec![
cmd("p1", "/bin/echo", &["{\"hello\": 1}"]),
cmd("p2", "/bin/cat", &[]),
],
edges: vec![FlowEdge {
from: 0,
from_output: "stdout".into(),
to: 1,
to_input: "stdin".into(),
}],
discern: DiscernPolicy {
sample_bytes: 4096,
enrich_producer: true,
replay_chunks: 32,
replay_bytes: 0,
max_bytes_per_sec: 0,
},
restart_on_failure: false,
restart_backoff_ms: 200,
restart_max_backoff_ms: 30_000,
restart_max: 0,
};
let disc = Arc::new(DiscernPipeline::default_pipeline());
let inc = Arc::new(Incarnator::new(IncarnatorConfig::default()));
let launch = run_pipeline(&spec, "ws", true, disc, inc, None).await.unwrap();
assert_eq!(launch.edge_discernments.len(), 1);
let d = &launch.edge_discernments[0];
let dis = d.discernment.as_ref().expect("discernment present");
assert_eq!(dis.mime.as_deref(), Some("application/json"));
// Cosecha.
for (_, pid) in &launch.command_pids {
let _ = nix::sys::wait::waitpid(nix::unistd::Pid::from_raw(*pid), None);
}
}
}
@@ -0,0 +1,210 @@
//! Resource accounting por workspace.
//!
//! Dos fuentes:
//! - **Per-proc** (`/proc/<pid>/status` + `stat`): suma RSS y CPU ticks de
//! los comandos vivos del workspace. Siempre disponible. Costo: O(N pids).
//! - **Cgroup v2** (`memory.current`, `cpu.stat`): un read por workspace si
//! `SomaSpec.cgroup.path` está y es leíble. Más preciso (incluye descendants).
//!
//! Si ambos están disponibles, devolvemos el cgroup (más preciso) y dejamos
//! el per-proc como `sample_via_proc`.
use std::path::Path;
use std::time::Instant;
#[derive(Debug, Clone, Default)]
pub struct WorkspaceStats {
pub commands_alive: u32,
pub commands_total: u32,
/// RSS sumado en bytes. `None` si no se pudo medir.
pub rss_bytes: Option<u64>,
/// High-water mark de RSS (peak alguna vez observado). Cgroup v2:
/// `memory.peak` (≥6.5). Per-proc: suma de `VmHWM` de cada pid.
pub rss_peak_bytes: Option<u64>,
/// Tiempo CPU acumulado en microsegundos. `None` si no se pudo medir.
pub cpu_usec: Option<u64>,
/// %CPU instantáneo derivado entre dos samples consecutivos. `None`
/// en el primer sample (no hay baseline). `100.0` = 1 core saturado.
/// `400.0` con 4 cores activos = la máquina al 100%.
pub cpu_percent: Option<f32>,
/// Cores online detectados (sysconf `_SC_NPROCESSORS_ONLN`). Útil
/// para normalizar `cpu_percent / cpu_cores` → 0..100 absoluto.
pub cpu_cores: u32,
/// Fuente del dato: "proc" | "cgroup" | "mixed".
pub source: String,
/// Wall-clock uptime del workspace en milisegundos.
pub uptime_ms: u64,
}
impl WorkspaceStats {
/// CPU% normalizado al 100% total de la máquina (no por core).
/// Útil para comparar workspaces independiente del paralelismo.
pub fn cpu_percent_total(&self) -> Option<f32> {
self.cpu_percent
.map(|p| if self.cpu_cores == 0 { p } else { p / self.cpu_cores as f32 })
}
}
/// Reporte de quotas: comparación entre el accounting real y los
/// `rlimits` declarados en `SomaSpec`. NO hace enforcement automático
/// en v1 — sólo accounting + reporting. El caller decide qué hacer.
#[derive(Debug, Clone, Default)]
pub struct QuotaReport {
/// Límite de memoria declarado (bytes). None = sin límite.
pub mem_limit: Option<u64>,
/// Límite de procesos declarado.
pub nproc_limit: Option<u32>,
/// Lista de violaciones detectadas (strings humano-legibles).
/// Empty = todo dentro de quota.
pub breaches: Vec<String>,
}
/// Detecta cores online runtime. Cacheado vía OnceLock — el valor no
/// cambia salvo hotplug, que es raro y aceptamos sample stale.
fn online_cores() -> u32 {
static CACHED: std::sync::OnceLock<u32> = std::sync::OnceLock::new();
*CACHED.get_or_init(|| {
let n = unsafe { libc::sysconf(libc::_SC_NPROCESSORS_ONLN) };
if n > 0 { n as u32 } else { 1 }
})
}
/// Mide stats para un set de PIDs vivos + un path de cgroup opcional.
pub fn measure(
alive_pids: &[i32],
cgroup_path: Option<&Path>,
workspace_started: Instant,
) -> WorkspaceStats {
let mut rss_proc: u64 = 0;
let mut rss_peak_proc: u64 = 0;
let mut cpu_proc: u64 = 0;
let mut proc_ok = false;
for &pid in alive_pids {
if let Some((rss, peak, cpu)) = read_proc_pid(pid) {
rss_proc += rss;
rss_peak_proc += peak;
cpu_proc += cpu;
proc_ok = true;
}
}
let cgroup = cgroup_path.and_then(read_cgroup_stats);
let (rss, rss_peak, cpu, source) = match (cgroup, proc_ok) {
(Some(cg), _) => (Some(cg.rss), cg.rss_peak, Some(cg.cpu_usec), "cgroup".to_string()),
(None, true) => (
Some(rss_proc),
Some(rss_peak_proc),
Some(cpu_proc),
"proc".to_string(),
),
(None, false) => (None, None, None, "none".to_string()),
};
WorkspaceStats {
commands_alive: alive_pids.len() as u32,
commands_total: 0,
rss_bytes: rss,
rss_peak_bytes: rss_peak,
cpu_usec: cpu,
cpu_percent: None, // El caller lo rellena con el diff vs prev sample.
cpu_cores: online_cores(),
source,
uptime_ms: workspace_started.elapsed().as_millis() as u64,
}
}
struct CgroupStats {
rss: u64,
rss_peak: Option<u64>,
cpu_usec: u64,
}
/// Lee `(rss_bytes, rss_peak_bytes, cpu_usec)` de `/proc/<pid>/`. None si el proc desapareció.
fn read_proc_pid(pid: i32) -> Option<(u64, u64, u64)> {
let (rss_kb, hwm_kb) = {
let status = std::fs::read_to_string(format!("/proc/{pid}/status")).ok()?;
let mut rss = 0u64;
let mut hwm = 0u64;
for l in status.lines() {
if let Some(rest) = l.strip_prefix("VmRSS:") {
rss = rest
.trim()
.split_whitespace()
.next()
.and_then(|s| s.parse().ok())
.unwrap_or(0);
} else if let Some(rest) = l.strip_prefix("VmHWM:") {
hwm = rest
.trim()
.split_whitespace()
.next()
.and_then(|s| s.parse().ok())
.unwrap_or(0);
}
}
(rss, hwm)
};
let cpu_usec = {
let stat = std::fs::read_to_string(format!("/proc/{pid}/stat")).ok()?;
// formato: pid (comm) state ppid pgrp ... utime stime cutime cstime
// Cuidado: comm puede tener espacios y paréntesis. Buscamos la última `)`.
let end_comm = stat.rfind(')')?;
let after = &stat[end_comm + 1..];
let fields: Vec<&str> = after.split_whitespace().collect();
// Tras `)`, índice 0 = state, índice 11 = utime, 12 = stime.
let utime = fields.get(11).and_then(|s| s.parse::<u64>().ok()).unwrap_or(0);
let stime = fields.get(12).and_then(|s| s.parse::<u64>().ok()).unwrap_or(0);
let ticks = utime + stime;
// Convertimos ticks → microsegundos. SC_CLK_TCK típicamente 100.
let clk_tck = unsafe { libc::sysconf(libc::_SC_CLK_TCK) }.max(1) as u64;
ticks * 1_000_000 / clk_tck
};
Some((rss_kb * 1024, hwm_kb * 1024, cpu_usec))
}
/// Lee `CgroupStats` del cgroup. None si no existe o no es leíble.
/// `memory.peak` requiere kernel ≥6.5; si falta, `rss_peak` queda None.
fn read_cgroup_stats(cgroup_path: &Path) -> Option<CgroupStats> {
let mem = std::fs::read_to_string(cgroup_path.join("memory.current"))
.ok()
.and_then(|s| s.trim().parse::<u64>().ok())?;
let cpu_stat = std::fs::read_to_string(cgroup_path.join("cpu.stat")).ok()?;
let cpu_usec = cpu_stat
.lines()
.find_map(|l| l.strip_prefix("usage_usec"))
.and_then(|s| s.split_whitespace().next())
.and_then(|s| s.parse::<u64>().ok())
.unwrap_or(0);
let peak = std::fs::read_to_string(cgroup_path.join("memory.peak"))
.ok()
.and_then(|s| s.trim().parse::<u64>().ok());
Some(CgroupStats {
rss: mem,
rss_peak: peak,
cpu_usec,
})
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn measure_with_no_pids_returns_zero() {
let stats = measure(&[], None, Instant::now());
assert_eq!(stats.commands_alive, 0);
assert_eq!(stats.rss_bytes, None);
assert_eq!(stats.source, "none");
}
#[test]
fn measure_self_pid_returns_data() {
let me = std::process::id() as i32;
let stats = measure(&[me], None, Instant::now());
assert_eq!(stats.commands_alive, 1);
// Nuestro propio RSS debería ser > 0.
assert!(stats.rss_bytes.unwrap_or(0) > 0);
assert_eq!(stats.source, "proc");
}
}
@@ -0,0 +1,15 @@
[package]
name = "shuma-discern"
version.workspace = true
edition.workspace = true
rust-version.workspace = true
license.workspace = true
authors.workspace = true
publish.workspace = true
description = "Discernidor de contenido sobre buffers: MIME, codificación, parser hints. Compartible con file_explorer y nouser."
[dependencies]
brahman-card = { path = "../../../protocol/brahman-card" }
serde = { workspace = true }
serde_json = { workspace = true }
toml = { workspace = true }
@@ -0,0 +1,307 @@
//! `shuma-discern` — detección de tipo de contenido sobre buffers.
//!
//! Trait + pipeline + discerners default. Devuelve un [`Discernment`] con
//! `TypeRef` consistente con el broker, confidence, MIME y un `lens` hint
//! para UIs (reusa el espíritu del `dominant_lens` de akasha).
#![forbid(unsafe_code)]
use brahman_card::TypeRef;
#[derive(Debug, Clone)]
pub struct Hint<'a> {
pub path: Option<&'a str>,
pub size_total: Option<u64>,
}
#[derive(Debug, Clone)]
pub struct Discernment {
pub ty: TypeRef,
pub confidence: f32,
pub mime: Option<String>,
pub lens: Option<String>,
}
pub trait Discerner: Send + Sync {
fn name(&self) -> &str;
fn discern(&self, sample: &[u8], hint: &Hint<'_>) -> Option<Discernment>;
}
pub struct DiscernPipeline {
discerners: Vec<Box<dyn Discerner>>,
}
impl DiscernPipeline {
pub fn new() -> Self {
Self { discerners: Vec::new() }
}
/// Pipeline con los discerners default. Orden importa: el primer match
/// con confidence ≥ `accept_threshold` corta.
pub fn default_pipeline() -> Self {
let mut p = Self::new();
p.push(Box::new(MagicBytes));
// CardProbe antes que JsonProbe: una Card es JSON, pero queremos el
// TypeRef más específico cuando aplique.
p.push(Box::new(CardProbe));
p.push(Box::new(JsonProbe));
p.push(Box::new(TomlProbe));
p.push(Box::new(Utf8Probe));
p
}
pub fn push(&mut self, d: Box<dyn Discerner>) {
self.discerners.push(d);
}
/// Recorre los discerners y devuelve el primer Discernment con
/// confidence ≥ 0.5, o el más confidente si ninguno alcanza el umbral.
pub fn discern(&self, sample: &[u8], hint: &Hint<'_>) -> Option<Discernment> {
let mut best: Option<Discernment> = None;
for d in &self.discerners {
if let Some(r) = d.discern(sample, hint) {
if r.confidence >= 0.9 {
return Some(r);
}
best = match best {
Some(prev) if prev.confidence >= r.confidence => Some(prev),
_ => Some(r),
};
}
}
best
}
}
impl Default for DiscernPipeline {
fn default() -> Self {
Self::default_pipeline()
}
}
// =====================================================================
// Discerners
// =====================================================================
/// Magic-bytes para formatos comunes. Confidence alta cuando hay match.
pub struct MagicBytes;
impl Discerner for MagicBytes {
fn name(&self) -> &str { "magic-bytes" }
fn discern(&self, s: &[u8], _h: &Hint<'_>) -> Option<Discernment> {
let d = |ty: &str, mime: &str, lens: Option<&str>| Discernment {
ty: TypeRef::Primitive { name: ty.into() },
confidence: 0.99,
mime: Some(mime.into()),
lens: lens.map(String::from),
};
match s {
x if x.starts_with(&[0x89, b'P', b'N', b'G']) => Some(d("png", "image/png", Some("gallery"))),
x if x.starts_with(&[0xFF, 0xD8, 0xFF]) => Some(d("jpeg", "image/jpeg", Some("gallery"))),
x if x.starts_with(b"%PDF-") => Some(d("pdf", "application/pdf", Some("reader"))),
x if x.starts_with(&[0x7F, b'E', b'L', b'F']) => Some(d("elf", "application/x-executable", None)),
x if x.starts_with(&[0x00, 0x61, 0x73, 0x6D]) => Some(d("wasm", "application/wasm", None)),
x if x.starts_with(&[0x1F, 0x8B]) => Some(d("gzip", "application/gzip", None)),
x if x.starts_with(b"PK\x03\x04") || x.starts_with(b"PK\x05\x06") => {
Some(d("zip", "application/zip", None))
}
x if x.starts_with(b"GIF87a") || x.starts_with(b"GIF89a") => {
Some(d("gif", "image/gif", Some("gallery")))
}
_ => None,
}
}
}
/// JSON: parsea el inicio. No requiere parsearlo entero; con que arranque
/// con `{`/`[` y haga progreso cuenta.
pub struct JsonProbe;
impl Discerner for JsonProbe {
fn name(&self) -> &str { "json" }
fn discern(&self, s: &[u8], _h: &Hint<'_>) -> Option<Discernment> {
let trimmed = trim_left(s);
let first = *trimmed.first()?;
if first != b'{' && first != b'[' {
return None;
}
// Intento parsear tal cual; si falla por truncated, igualmente confidence media.
let txt = std::str::from_utf8(trimmed).ok()?;
match serde_json::from_str::<serde_json::Value>(txt) {
Ok(_) => Some(Discernment {
ty: TypeRef::Primitive { name: "json".into() },
confidence: 0.95,
mime: Some("application/json".into()),
lens: Some("tree".into()),
}),
Err(_) => Some(Discernment {
ty: TypeRef::Primitive { name: "json".into() },
confidence: 0.6, // sample truncado
mime: Some("application/json".into()),
lens: Some("tree".into()),
}),
}
}
}
pub struct TomlProbe;
impl Discerner for TomlProbe {
fn name(&self) -> &str { "toml" }
fn discern(&self, s: &[u8], h: &Hint<'_>) -> Option<Discernment> {
let txt = std::str::from_utf8(s).ok()?;
// Heurística: presencia de `[seccion]` y/o `clave = valor` y extensión.
let looks_like = txt.lines().any(|l| {
let l = l.trim();
l.starts_with('[') && l.ends_with(']')
}) || txt.lines().any(|l| {
let l = l.trim();
!l.starts_with('#') && l.contains(" = ")
});
if !looks_like {
return None;
}
let confidence = if h.path.map_or(false, |p| p.ends_with(".toml")) {
0.95
} else {
0.55
};
// Si parsea, sube confidence.
let parsed = toml::from_str::<toml::Value>(txt).is_ok();
Some(Discernment {
ty: TypeRef::Primitive { name: "toml".into() },
confidence: if parsed { 0.93 } else { confidence },
mime: Some("application/toml".into()),
lens: Some("tree".into()),
})
}
}
/// Si el JSON parsea como Card, lo emite como Wit { brahman:card }.
pub struct CardProbe;
impl Discerner for CardProbe {
fn name(&self) -> &str { "card" }
fn discern(&self, s: &[u8], _h: &Hint<'_>) -> Option<Discernment> {
let trimmed = trim_left(s);
if trimmed.first()? != &b'{' {
return None;
}
let txt = std::str::from_utf8(trimmed).ok()?;
let v: serde_json::Value = serde_json::from_str(txt).ok()?;
let obj = v.as_object()?;
if obj.contains_key("schema_version") && obj.contains_key("id") && obj.contains_key("payload") {
Some(Discernment {
ty: TypeRef::Wit {
package: "brahman:card".into(),
interface: None,
name: "card".into(),
},
confidence: 0.97,
mime: Some("application/json".into()),
lens: Some("card".into()),
})
} else {
None
}
}
}
/// Texto UTF-8 plano. Fallback de baja confidence.
pub struct Utf8Probe;
impl Discerner for Utf8Probe {
fn name(&self) -> &str { "utf8" }
fn discern(&self, s: &[u8], h: &Hint<'_>) -> Option<Discernment> {
if s.is_empty() {
return None;
}
let valid = std::str::from_utf8(s).is_ok();
if !valid {
return None;
}
// Detectar binario disfrazado: bytes de control fuera de \t\n\r.
let suspicious = s.iter().filter(|&&b| b < 0x09 || (b > 0x0D && b < 0x20)).count();
if suspicious * 100 / s.len().max(1) > 5 {
return None;
}
let lens = h.path.and_then(|p| {
if p.ends_with(".md") { Some("markdown") }
else if p.ends_with(".rs") || p.ends_with(".py") || p.ends_with(".go") || p.ends_with(".js") || p.ends_with(".ts") {
Some("code")
} else { None }
}).map(String::from);
Some(Discernment {
ty: TypeRef::Primitive { name: "text".into() },
confidence: 0.5,
mime: Some("text/plain; charset=utf-8".into()),
lens,
})
}
}
fn trim_left(s: &[u8]) -> &[u8] {
let mut i = 0;
while i < s.len() && (s[i] == b' ' || s[i] == b'\t' || s[i] == b'\n' || s[i] == b'\r') {
i += 1;
}
&s[i..]
}
#[cfg(test)]
mod tests {
use super::*;
fn discern(sample: &[u8]) -> Option<Discernment> {
DiscernPipeline::default_pipeline().discern(sample, &Hint { path: None, size_total: None })
}
#[test]
fn png_detected() {
let r = discern(&[0x89, b'P', b'N', b'G', 0x0D, 0x0A, 0x1A, 0x0A, 0, 0]).unwrap();
assert_eq!(r.mime.as_deref(), Some("image/png"));
assert!(r.confidence > 0.9);
}
#[test]
fn json_detected() {
let r = discern(b"{\"hello\": 1}").unwrap();
assert_eq!(r.mime.as_deref(), Some("application/json"));
}
#[test]
fn card_wins_over_plain_json() {
let payload = br#"{"schema_version":1,"id":"01ARZ3NDEKTSV4RRFFQ69G5FAV","label":"x","payload":{"Virtual":null},"supervision":"OneShot"}"#;
let r = discern(payload).unwrap();
match r.ty {
TypeRef::Wit { ref package, .. } => assert_eq!(package, "brahman:card"),
_ => panic!("expected card"),
}
}
#[test]
fn utf8_text_fallback() {
let r = discern(b"hello world\nthis is text").unwrap();
// Puede ser detected as toml (= heurística) o text. Ambos son aceptables, sólo aseguro algo razonable.
assert!(r.mime.is_some());
}
#[test]
fn binary_rejected_by_utf8() {
let mut bytes = vec![0u8; 100];
bytes[0] = 0x00;
bytes[1] = 0x01;
bytes[2] = 0x02;
let r = DiscernPipeline::default_pipeline().discern(&bytes, &Hint { path: None, size_total: None });
// Tras Utf8Probe rechazar, no hay match → None.
// Si por casualidad otro discerner mata antes, también es OK.
if let Some(r) = r {
assert_ne!(r.mime.as_deref(), Some("text/plain; charset=utf-8"));
}
}
}
@@ -0,0 +1,19 @@
[package]
name = "shuma-protocol"
version.workspace = true
edition.workspace = true
rust-version.workspace = true
license.workspace = true
authors.workspace = true
publish.workspace = true
description = "Wire protocol entre shipote-daemon y clientes (cli/gui). Postcard length-prefixed sobre Unix socket."
[dependencies]
shuma-card = { path = "../shuma-card" }
brahman-card = { path = "../../../protocol/brahman-card" }
serde = { workspace = true }
postcard = { workspace = true }
thiserror = { workspace = true }
tokio = { workspace = true }
ulid = { workspace = true }
nix = { workspace = true }
@@ -0,0 +1,444 @@
//! `shuma-protocol` — wire daemon ↔ cliente (cli/gui).
//!
//! Framing: u32 BE length-prefix + payload postcard. Mismo patrón que
//! `ente-bus`/`brahman-handshake` para que clientes existentes compartan
//! reader/writer helpers si quieren.
#![forbid(unsafe_code)]
use serde::{Deserialize, Serialize};
use shuma_card::{PipelineSpec, WorkspaceId, WorkspaceSpec};
use std::path::PathBuf;
use thiserror::Error;
use tokio::io::{AsyncReadExt, AsyncWriteExt};
use tokio::net::UnixStream;
use ulid::Ulid;
pub const DEFAULT_SOCK_NAME: &str = "shuma.sock";
pub const MAX_FRAME: usize = 1 << 20;
fn default_grace_ms() -> u64 {
1000
}
// =====================================================================
// Mensajes
// =====================================================================
#[derive(Debug, Clone, Serialize, Deserialize)]
pub enum Request {
/// Health-check.
Ping,
/// Health endpoint estructurado: versión + uptime + counts.
Health,
/// Crear un workspace nuevo.
WorkspaceCreate { spec: WorkspaceSpec },
/// Listar todos los workspaces vivos.
WorkspaceList,
/// Detener un workspace y reapear sus comandos. `grace_ms`: tiempo
/// que se espera tras SIGTERM antes de SIGKILL. 0 = SIGKILL inmediato.
WorkspaceStop {
id: WorkspaceId,
#[serde(default = "default_grace_ms")]
grace_ms: u64,
},
/// Ejecutar un comando one-shot dentro de un workspace existente.
Run {
workspace: WorkspaceId,
exec: String,
argv: Vec<String>,
envp: Vec<(String, String)>,
/// Si `true` y el comando muere con exit_status != 0, el reaper
/// lo relaunch con backoff exponencial.
#[serde(default)]
restart_on_failure: bool,
},
/// Lanzar un Pipeline completo dentro de un workspace.
PipelineRun {
spec: PipelineSpec,
/// Si `true`, el daemon interpone un tap entre productor y
/// consumidor de cada FlowEdge, sampleando los primeros bytes
/// y discerniendo el TypeRef.
tap: bool,
/// Variables para sustitución `${KEY}` en strings del spec
/// antes de spawn (templating).
#[serde(default)]
vars: std::collections::BTreeMap<String, String>,
},
/// Discernir un buffer ad-hoc (sin workspace). Útil para `shuma discern <file>`.
Discern { sample: Vec<u8>, hint_path: Option<PathBuf> },
/// Capacidades runtime del kernel/proceso del daemon.
Capabilities,
/// Listar comandos vivos+pasados de un workspace.
CommandList { workspace: shuma_card::WorkspaceId },
/// Tail del log capturado para un comando.
CommandLogs {
workspace: shuma_card::WorkspaceId,
command: Ulid,
tail_bytes: usize,
/// "stdout" | "stderr" | "both" (default "both" si vacío).
stream: String,
},
/// Guardar (o reemplazar) un PipelineSpec bajo un nombre.
PipelineSave { name: String, spec: PipelineSpec },
/// Listar nombres de pipelines guardados.
PipelineSavedList,
/// Eliminar un pipeline guardado.
PipelineDrop { name: String },
/// Ejecutar un pipeline guardado.
PipelineRunSaved {
name: String,
tap: bool,
#[serde(default)]
vars: std::collections::BTreeMap<String, String>,
},
/// Resource accounting de un workspace.
WorkspaceStats { workspace: shuma_card::WorkspaceId },
/// Reporte de quotas (rlimits declarados vs uso actual).
WorkspaceQuota { workspace: shuma_card::WorkspaceId },
/// History de samples del workspace (server-side). Sobrevive
/// restart del shell. `tail`: cantidad de samples desde el final
/// (0 = todo).
WorkspaceStatsHistory {
workspace: shuma_card::WorkspaceId,
tail: usize,
},
/// Resumen completo de un workspace: stats + quota + commands +
/// flow sockets en una sola roundtrip. Reduce N×4 requests del
/// shell a N×1.
WorkspaceFullSummary { workspace: shuma_card::WorkspaceId },
/// Detener selectivamente los comandos de un pipeline (no el workspace
/// entero). `grace_ms`: SIGTERM → wait → SIGKILL.
PipelineStop {
pipeline: Ulid,
#[serde(default = "default_grace_ms")]
grace_ms: u64,
},
/// Listar pipelines activos con sus flow channels (data plane).
FlowList,
/// Throughput por flow socket: bytes_total + bytes_per_sec.
FlowThroughput,
/// Cerrar el data plane de un pipeline (drop sockets + canales).
FlowDrop { pipeline: Ulid },
}
#[derive(Debug, Clone, Serialize, Deserialize)]
pub enum Response {
Pong,
Health {
version: String,
uptime_ms: u64,
alive_workspaces: u32,
alive_commands: u32,
alive_pipelines: u32,
active_flows: u32,
dirty: bool,
},
WorkspaceCreated {
id: WorkspaceId,
warnings: Vec<String>,
},
WorkspaceList {
items: Vec<WorkspaceSummary>,
},
WorkspaceStopped {
id: WorkspaceId,
reaped: u32,
},
RunStarted {
workspace: WorkspaceId,
command_id: Ulid,
pid: i32,
},
PipelineStarted {
pipeline: Ulid,
command_pids: Vec<(String, i32)>,
/// Discernments por edge cuando tap=true. Vacío sin tap.
edges: Vec<EdgeDiscernmentInfo>,
},
Discernment {
ty: String,
confidence: f32,
mime: Option<String>,
lens: Option<String>,
},
Capabilities {
kernel_version: (u32, u32, u32),
user_ns: String,
cgroup_v2: String,
cgroup_delegated: bool,
has_cap_sys_admin: bool,
},
CommandList {
items: Vec<CommandInfo>,
},
CommandLogs {
bytes: Vec<u8>,
},
PipelineSaved {
name: String,
},
PipelineSavedList {
names: Vec<String>,
},
PipelineDropped {
name: String,
existed: bool,
},
PipelineStopped {
pipeline: Ulid,
reaped: u32,
},
WorkspaceStats {
info: WorkspaceStatsInfo,
},
WorkspaceQuota {
info: QuotaReportInfo,
},
WorkspaceStatsHistory {
samples: Vec<WorkspaceStatsInfo>,
},
WorkspaceFullSummary {
stats: WorkspaceStatsInfo,
quota: QuotaReportInfo,
commands: Vec<CommandInfo>,
flow_sockets: Vec<PathBuf>,
},
FlowList {
items: Vec<FlowInfo>,
},
FlowThroughput {
items: Vec<FlowThroughputInfo>,
},
FlowDropped {
pipeline: Ulid,
existed: bool,
},
Error {
message: String,
},
}
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct QuotaReportInfo {
pub mem_limit: Option<u64>,
pub nproc_limit: Option<u32>,
pub breaches: Vec<String>,
}
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct WorkspaceStatsInfo {
pub commands_alive: u32,
pub commands_total: u32,
pub rss_bytes: Option<u64>,
#[serde(default)]
pub rss_peak_bytes: Option<u64>,
pub cpu_usec: Option<u64>,
#[serde(default)]
pub cpu_percent: Option<f32>,
#[serde(default = "default_cpu_cores")]
pub cpu_cores: u32,
pub source: String,
pub uptime_ms: u64,
}
fn default_cpu_cores() -> u32 {
1
}
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct FlowThroughputInfo {
pub socket: PathBuf,
pub bytes_total: u64,
pub bytes_per_sec: f64,
}
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct FlowInfo {
pub pipeline: Ulid,
pub sockets: Vec<PathBuf>,
}
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct CommandInfo {
pub id: Ulid,
pub label: String,
pub pid: i32,
pub alive: bool,
pub exit_status: Option<i32>,
pub log_bytes: u64,
}
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct EdgeDiscernmentInfo {
pub from_label: String,
pub from_output: String,
pub to_label: String,
pub to_input: String,
/// `Some(ty)` si el discerner detectó algo. `None` si no hubo data
/// suficiente o no matcheó ningún discerner.
pub ty: Option<String>,
pub mime: Option<String>,
pub lens: Option<String>,
pub confidence: f32,
/// Path del Unix socket donde otros módulos pueden suscribirse a los
/// bytes replicados de este edge (data plane). `None` si tap=false.
pub flow_socket: Option<PathBuf>,
}
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct WorkspaceSummary {
pub id: WorkspaceId,
pub label: String,
pub commands: u32,
pub uptime_ms: u64,
}
// =====================================================================
// Errores
// =====================================================================
#[derive(Debug, Error)]
pub enum ProtocolError {
#[error("frame oversize: {0} bytes (max {MAX_FRAME})")]
FrameOversize(usize),
#[error("io: {0}")]
Io(#[from] std::io::Error),
#[error("postcard: {0}")]
Postcard(#[from] postcard::Error),
#[error("connection closed")]
Closed,
}
// =====================================================================
// Framing helpers
// =====================================================================
pub async fn write_frame<T: Serialize>(stream: &mut UnixStream, msg: &T) -> Result<(), ProtocolError> {
let bytes = postcard::to_allocvec(msg)?;
if bytes.len() > MAX_FRAME {
return Err(ProtocolError::FrameOversize(bytes.len()));
}
let len = (bytes.len() as u32).to_be_bytes();
stream.write_all(&len).await?;
stream.write_all(&bytes).await?;
stream.flush().await?;
Ok(())
}
pub async fn read_frame<T: for<'de> Deserialize<'de>>(
stream: &mut UnixStream,
) -> Result<T, ProtocolError> {
let mut len_buf = [0u8; 4];
stream.read_exact(&mut len_buf).await.map_err(|e| {
if e.kind() == std::io::ErrorKind::UnexpectedEof {
ProtocolError::Closed
} else {
ProtocolError::Io(e)
}
})?;
let len = u32::from_be_bytes(len_buf) as usize;
if len > MAX_FRAME {
return Err(ProtocolError::FrameOversize(len));
}
let mut buf = vec![0u8; len];
stream.read_exact(&mut buf).await?;
Ok(postcard::from_bytes(&buf)?)
}
/// Path canónico del socket del daemon: `$XDG_RUNTIME_DIR/shuma.sock`,
/// fallback `/run/user/$UID/shuma.sock`, fallback `/tmp/shuma-$UID.sock`.
pub fn default_socket_path() -> PathBuf {
if let Ok(xdg) = std::env::var("XDG_RUNTIME_DIR") {
return PathBuf::from(xdg).join(DEFAULT_SOCK_NAME);
}
let uid = nix::unistd::getuid().as_raw();
let p = PathBuf::from(format!("/run/user/{uid}"));
if p.exists() {
return p.join(DEFAULT_SOCK_NAME);
}
PathBuf::from(format!("/tmp/shuma-{uid}.sock"))
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn ping_roundtrip() {
let bytes = postcard::to_allocvec(&Request::Ping).unwrap();
let back: Request = postcard::from_bytes(&bytes).unwrap();
assert!(matches!(back, Request::Ping));
}
#[test]
fn workspace_create_roundtrip() {
let req = Request::WorkspaceCreate {
spec: WorkspaceSpec {
label: "demo".into(),
soma: Default::default(),
permissions: Default::default(),
ttl: None,
flow_dirs: vec![],
on_exit: shuma_card::ExitPolicy::Reap,
quota_enforce: Default::default(),
},
};
let bytes = postcard::to_allocvec(&req).unwrap();
let back: Request = postcard::from_bytes(&bytes).unwrap();
match back {
Request::WorkspaceCreate { spec } => assert_eq!(spec.label, "demo"),
_ => panic!("wrong variant"),
}
}
#[test]
fn default_socket_path_uses_runtime_dir() {
let p = default_socket_path();
assert!(p.to_string_lossy().ends_with("shuma.sock"));
}
}