hpc-agent-core 0.4.6

Shared runtime for the HPC MCP agent family: SSH middleware, PSI/J-style job models, scheduler backends, docs RAG, doctor, and serving, reused across per-machine plugins (Rikyu, HOKUSAI, Octopus, ...).

hpc-agent-core

Shared runtime for the HPC MCP agent family — the Claude Code / Codex plugins that let an agent submit and monitor Slurm/Grid Engine jobs, manage files, and search documentation on a supercomputer (Rikyu, HOKUSAI BigWaterfall2, Octopus, R-CCS Cloud, TSUBAME4, and others).

Each of those plugins used to be a full copy-pasted fork of the same template. This package extracts the parts that were already generic in practice — SSH middleware, PSI/J-style job models, scheduler backends, the docs RAG pipeline, health checks (doctor), and the MCP serving glue — so a machine repo becomes a thin "skin": its own <machine>_config.json, a hand-written guide, skills, and packaging, depending on this package for everything else.

See PLAN.md in the merge_computers workspace for the full design rationale and migration plan. This repo implements that plan's Phase 1 (extract the generic modules) — it is not yet a drop-in replacement for any machine repo's server code; nothing currently depends on it yet.

What's here (Phase 1)

• config.py — generic env/file/default settings resolution. A machine's own config.py calls hpc_agent_core.config.configure(...) once to register its SSH default, embedding endpoint, docs source, and (optionally) any remotemanager.Computer option it needs to differ from the shared defaults (computer_defaults={...} — see COMPUTER_OPTION_NAMES for the full supported set: shell, timeout, keyfile, landing_dir, transport, ...). A machine sets these once in its own repo; the end user never has to.
• middleware.py — the SSH execution layer (base64 payloads, login shell, clean error surfacing). Never touches config or the network above module scope — the MCP server must never fail to start just because config is missing (see PORTING.md's invariant in every machine repo).
• models.py — PSI/J-style JobSpec/ResourceSpec/JobAttributes/Job/ JobState, with no per-machine defaults baked in. Also Scheduler (an optional hint field for a machine that composes more than one SchedulerBackend) and map_ge_state for Grid Engine.
• compute/base.py — the SchedulerBackend ABC and scheduler-neutral script-body rendering (env vars, container wrapping, launcher prefix).
• compute/slurm.py — a config-driven Slurm backend: has_accounting (sacct vs. squeue+scontrol), gpu_request_style ("gpus_total" vs. "gres") and the independent nodes_always_explicit (whether Slurm derives node count from the GPU count, or --nodes is always emitted — these two don't always move together, see PHASE4_AUDIT.md §1.1), no_gpu_flag_prefixes (partitions needing no GPU flag at all, e.g. a unified CPU+GPU superchip), and gpu_vendor_map (container GPU flag by partition prefix). Verified against Rikyu's, Octopus's, and RCCS-Cloud's actual rendered scripts. The has_accounting=False path (Banyan/Dgx1-style) is implemented from the porting knowledge-transfer reports and passes a mocked end-to-end test, but is not yet verified against a real no-accounting cluster — see the module docstring before trusting it on a live machine. Also provides get_live_resources()/ get_drained_nodes() (live per-partition occupancy via sinfo) — a machine's get_resources tool should call this rather than returning static config data, a real gap the first clean-room port using this guide found in practice.
• compute/gridengine.py — a Grid Engine backend (qsub/qstat/qacct/qdel), promoted from shinobulab-cell-cluster-mcp (the only GE machine so far) and verified to reproduce its exact rendered scripts. host_pins/ queue_aliases/default_queue generalize its host-pinning quirk into config, since it's a plausible shape for other small GE clusters, not a one-off. Live-tested as part of shinobulab-cell-cluster-mcp; not yet re-verified from this promoted copy against a real cluster.
• rag/ — embedding client, BM25 + vector docs index, and an ingest pipeline that only ever chunks a bundled local guide (never clones a remote docs site — see the module docstring for why).
• docs_server.py, doctor.py, serving.py — generic MCP docs server, health checks, and CLI entry point.
• mcp_server.py — the only file that imports the MCP SDK's server class. Machine repos import MCPServer from here, never from mcp directly, and don't declare "mcp" as their own dependency — so a future SDK rename (like 2.0.0's FastMCP → MCPServer) is a one-file fix here instead of a fix in every machine repo.

What's deliberately not here yet

• hpc_server.py (the IRI-grouped Slurm/filesystem tool surface) and a machine_profile.py that turns <machine>_config.json into config-driven scheduler dispatch — a machine currently wires a SlurmBackend/ GridEngineBackend by hand in its own compute.py; that's a deliberate choice (PLAN.md §2a/§2b), not a placeholder for a missing feature.
• Repointing any machine repo at this package as a dependency.

Development

python3 -m venv .venv && .venv/bin/pip install -e .

No machine repo depends on this yet, so there isn't a meaningful smoke test to run standalone — validate changes against a machine repo once one is repointed here.