Awaitless

Adaptive durable execution for coding agents.

Run commands through one execution layer. Quick work returns inline; longer or queued work becomes durable across local, SSH, and Slurm. Your workload stays on infrastructure you already own.

Agents submit work. Awaitless owns execution.

Awaitless is the adaptive durable execution layer between coding agents and the compute they use. It gives agents one stable job contract while reusing your local machine, SSH hosts, and Slurm clusters underneath.

简体中文 · Documentation · Benchmarks · PyPI

One job lifecycle across your existing compute

Coding agent → submit work → Awaitless owns the job lifecycle → Local / SSH / Slurm

Awaitless owns the job lifecycle, not the hardware. It does not discover resources, understand GPU topology, allocate multiple resources, or replace a cluster scheduler. Operators name queues and set fixed concurrency; Slurm continues to handle requests such as --gpus 2 --mem 64G and all physical cluster scheduling.

Awaitless SSH submit, disconnect, resume, and Artifact demo

Your coding agent should write code, not babysit jobs

An agent can write its own run → sleep → check loop. The harder problem is making job identity, disconnect recovery, queue admission, cancellation, and result delivery reliable across long workloads and changing sessions. Without that execution layer, the agent repeatedly pulls the same growing log back into its context:

ssh gpu 'run_benchmark > job.log 2>&1 &'
ssh gpu 'tail -n 200 job.log'  # again...
ssh gpu 'tail -n 200 job.log'  # and again...

Awaitless turns that lifecycle into one adaptive execution call and one result boundary:

awaitless run --json --host gpu --artifact results.json -- ./run_benchmark
# quick: {"state":"succeeded","delivery":"inline","exit_code":0,...}
# longer: {"job_id":"job_019F...","state":"running","delivery":"detached",...}

awaitless wait job_019F... --json
# {"state":"succeeded","exit_code":0,"parsed_results":{...}}

Detached JSON also includes job_state, wait_state, delivery_state, and a ready-to-copy next_command. A client-side wait timeout is not a workload failure: use awaitless wait --last --json for the most recently detached Job, or use the returned command with its stable Job ID. To inspect benchmark lines without reading a large tail, use awaitless logs <job-id> --grep 'PASS|FAIL|median|CV'.

Every run is durable before launch. Finishing within the inline window looks like an ordinary command result; crossing it only detaches the waiter. Interrupt the waiter, close the MCP client, or start a fresh agent session: the Job keeps running and its stable ID recovers the result.

Queue work before a named resource is free

Create a durable FIFO queue once, then submit every command immediately:

awaitless queue create gpu0 --concurrency 1

awaitless submit --queue gpu0 -- python train_a.py
awaitless submit --queue gpu0 -- python train_b.py
awaitless submit --queue gpu0 -- python train_c.py

The first command runs and the others report queued. Each starts automatically when capacity becomes available. This is durable admission control for a named scarce resource: fixed concurrency and FIFO ordering, with no priority or preemption. Awaitless never kills running work to make room for a later job.

Operators can also bind adaptive runs to a queue globally or per host:

[hosts.gpu]
hostname = "gpu.example.com"
queue = "gpu0"

The Agent can then call run without choosing a queue or probing the GPU first.

This queue does not discover resources, understand GPU topology, dynamically allocate devices, issue leases, or combine requests such as two GPUs plus 64 GB of memory. Use Slurm or another scheduler for those responsibilities; Awaitless provides the Agent-facing job lifecycle around that scheduler.

Consume whichever job finishes next

v0.7 adds completions ... --drain --json for consuming a small parallel set in one call without client-side cursor bookkeeping. Long jobs can emit structured heartbeat updates with wait --progress-interval 30s. Use --capture-log PATH for command-owned logs and --resource gpu=0 or --device 0 for explicit exclusive admission; terminal results freeze bounded logs, diagnostics, timing, environment, and a SHA-256 identified snapshot.

Submit independent work up front, keep every Job ID, then wait at one durable completion boundary:

awaitless completions job_A job_B job_C --json
# {"completions":[...],"next_cursor":"cmp_...","active_job_ids":[...]}

awaitless completions job_A job_B job_C --after cmp_... --json

The first call returns already-finished work immediately or blocks until at least one selected Job completes. Process the batch before advancing to next_cursor; reusing an older cursor safely replays the same completion IDs. If the client disappears, a new session can continue from the saved cursor. Awaitless makes continuation results durably available—it does not run the agent's next reasoning step or require a resident notification service.

The v0.8 evidence suite replaces historical call-count demos with four questions: does an Agent choose the protocol correctly, does a Job survive faults without duplicate launch, does Awaitless keep execution-management state out of the reasoning loop, and does adaptive run preserve low friction for short commands? See the v0.8 evidence plan.

v0.8 evidence status

Release evidence is model- and commit-specific. The checked-in suite does not carry numbers from earlier versions or from a different model. Run the v0.8 benchmarks, inspect every raw record, then publish a dated report with model, config hash, git commit, skipped workloads, and all failures in the denominator. The reviewed v0.8 evidence report includes the complete raw records and analysis summaries rather than a selected score.

Try the recovery story in 30 seconds

Linux, Python 3.10+, and Bash are required. Run the built-in demo without a persistent install:

uvx --from awaitless-runner awaitless demo --json

The demo submits two local jobs, terminates their first completion waiter, then uses new clients to consume both bounded results and JSON Artifacts by cursor.

For regular CLI use:

uv tool install awaitless-runner
awaitless doctor --json

pip install awaitless-runner works too.

Give it to your coding agent

Add one stdio MCP server to your client's configuration (adapt the outer key to your client):

{
  "mcpServers": {
    "awaitless": {
      "command": "uvx",
      "args": ["awaitless-runner"]
    }
  }
}

The preferred run tool returns quick commands inline and automatically gives longer or queued work a durable handle. Tasks-aware clients can still use run_job, while low-level clients retain submit_job and wait_for_job. Retrying an expensive submission with the same client_request_id cannot launch a duplicate job. For parallel work, every client can use wait_for_completions regardless of MCP Tasks support. The normative identity, lifecycle, continuation, completion, Artifact, and compatibility contract is Awaitless Agent Job Protocol.

Codex plugin

This repository is also a Codex plugin. Its manifest bundles the Awaitless agent skill with the stdio MCP server, which Codex launches through uvx. Install the repository from a local Codex marketplace, then start a new Codex task so the skill and MCP tools are loaded together.

The plugin requires uvx on PATH; the first MCP launch downloads awaitless-runner from PyPI if it is not already cached.

For direct CLI use, the whole loop is:

awaitless run --json --name tests -- python -m pytest -q
# If delivery is detached, save the returned job_id, then:
awaitless wait <job-id> --json
# Or recover the most recent detached job:
awaitless wait --last --json

One interface, three places to run

Use --backend, --host, or configuration defaults to switch targets without changing how the agent submits and collects work.

Why not just use a shell or tmux?

Awaitless does not replace interactive terminals or Slurm. It gives coding agents durable fixed-concurrency queues on local/SSH machines and delegates cluster resource scheduling to Slurm.

How it works

flowchart LR
    A["Coding agent"] -->|"run"| B["Awaitless MCP / CLI"]
    B --> C[("SQLite job record")]
    C --> Q["Optional queue admission"]
    Q -->|"capacity available"| D{"Backend"}
    D --> L["Local process"]
    D --> S["SSH host"]
    D --> H["Slurm allocation"]
    L --> I{"Finished inline?"}
    S --> I
    H --> I
    I -->|"yes: result"| A
    I -->|"no: durable handle"| A
    A -. "reconnect with stable ID" .-> C
    C --> E["Durable completion cursor"]
    E -->|"state + exit code + bounded logs + Artifacts"| A

There is no Awaitless daemon, HTTP service, or hosted sandbox. Each invocation opens the same SQLite store; submitted runners and scheduler jobs outlive the stdio server that created them. Full logs remain on disk while only bounded tails enter the agent context.

Documentation

• Documentation index
• Product positioning and evolution principles
• v0.6 adaptive run
• v0.7 immutable completion snapshots
• v0.5 durable completion feed
• CLI, configuration, SSH, Slurm, persistence, Artifacts, and troubleshooting
• MCP Tasks protocol and compatibility
• Benchmark definitions, methodology, and interpretation
• Real MCP → Slurm disconnect evidence
• Architecture and design rationale

License

MIT