Connectors: 22 real pytest modules (186 pass / 7 optional-dep skips) exercising
every connector's dependency-free surface + missing-driver degradation; SSRF
guard verified; errors_connector.py as a thin re-export (no forked taxonomy).
Services: config/{dev,staging,prod,test,limits,logging}.yaml + docs/*.md for
api/gateway/scheduler/worker, mapped to the real config structs; gateway/scheduler
Dockerfiles + scheduler .env/README. All YAML validated.
Closes 66 manifest files (gap 175->109). No source behavior changed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2.6 KiB
API contract
Two surfaces exist over one core: the in-process Scheduler<S> Rust API and a
dependency-free HTTP/1.1 control plane. Routing is a pure function of
(scheduler, now_ms, request) (src/router.rs), so the whole surface is testable
without a socket. Time is supplied by the caller (Unix milliseconds), keeping the
scheduler deterministic.
Control-plane boundary (important)
The HTTP transport exposes ONLY the scheduler's serializable operations. Lease
acquisition, heartbeat, complete, and fail are intentionally absent from the
route table: they carry or fence executable JobEnvelope values, which have no
wire codec. Dispatch of executable work therefore stays an in-process Rust API
call; the network surface is limited to state a client can safely observe and
mutate.
HTTP routes
| Method | Path | Purpose |
|---|---|---|
| GET | / |
{"service":"lawsynth-scheduler"} liveness banner |
| GET | /health |
queued_count and the effective config |
| POST | /pools |
register a worker pool (metadata only; never executable work) |
| GET | /jobs/{id} |
a job's current lifecycle state |
| GET | /jobs/{id}/checkpoint |
a job's durable checkpoint (no payload) |
| POST | /jobs/{id}/cancel |
cancel queued/leased work from a {reason} body |
| POST | /recover |
reclaim expired leases; returns {"recovered": n} |
Each route accepts exactly one method; others return 405 method_not_allowed
with an Allow header. Unknown paths return 404 not_found.
Request bodies
POST /pools—{"id": <string>, "cpu_millis": <u32>, "memory_bytes": <u64>, "disk_bytes": <u64>}. On success returns201with the poolidand aLocation: /pools/{id}header. This is capacity metadata; it never accepts a job.POST /jobs/{id}/cancel—{"reason": <string>}. Returns the job's resulting state. Cancellation is a control-plane transition; interrupting an executing worker remains that worker's cooperative responsibility.
Bodies are flat JSON objects; malformed or non-UTF-8 bodies return
400 invalid_body.
Job state shape
state is a tagged object: {"name":"queued"},
{"name":"leased","worker_id","generation","expires_at_ms"},
{"name":"completed"}, {"name":"cancelled","reason"}, or
{"name":"dead_letter","reason"}. A checkpoint adds job_id, attempt,
sequence, and updated_at_ms around that state.
Rust API
Construct Scheduler::new(SchedulerConfig, store), then use register_pool,
state, checkpoint, cancel, recover_expired, and the in-process
lease/heartbeat/complete/fail methods that move typed JobEnvelope values. See
docs/failures.md for the error-to-status mapping.