Skip to content
Allagma

Resource supervision

allagma resource runs study-owned commands under a finite, frozen resource profile. It contains no scientific model, metric, training loop or result criterion. Existing v0.2 campaign budgets and historical bundle helpers keep their original behavior. A study may wrap environment setup, campaigns and independent analysis with this additional supervisor.

Create a profile using the development example, then initialize a new ledger and launch a command:

Terminal window
python3 -m allagma resource init --ledger /path/to/ledger \
  --profile /path/to/profile.json --workdir /path/to/study
python3 -m allagma resource run --ledger /path/to/ledger \
  --category compute --label pilot-01 --timeout 120 --attempt -- \
  .venv/bin/python domain/runner.py inputs.json outputs.json
python3 -m allagma resource status --ledger /path/to/ledger

Categories have separate cumulative wall-time ceilings and per-command timeouts. For resource run, --workdir may select a subdirectory of the budgeted study; storage accounting still covers the entire frozen work directory. A path that escapes that directory is rejected. Use a separate setup category for dependency installation. Charge scientific checks, training, sampling, analysis and recomputation to compute. Record Codex tokens and account usage separately; an agent’s entire deliberation time is not scientific computation. --attempt counts a bounded computational job, not a statistical replicate. The study must also enforce any scientific run/seed cap.

Before launch, the supervisor reserves the timeout, termination grace, polling interval and four seconds of control overhead. Completed outcomes charge actual elapsed time; a missing outcome charges the whole reservation. Jobs use distinct directories with immutable reservation.json, process.json and result.json files and retained stdout/stderr. Nonzero commands and interrupted attempts are preserved. A concurrent operation or unresolved reservation prevents new work.

After a supervisor crash, inspect the recorded process and outputs. Do not restart because a wait timed out. resource recover --ledger /path/to/ledger refuses recovery when the recorded process group is still live. Once the work is confirmed stopped, it retains an abandoned outcome charged at the full reservation. It never fabricates scientific completion or restores budget for unmeasured work. To adopt a changed ceiling, obtain the required authorization and create a new, explicitly linked policy; do not edit an existing ledger.

The helper requires POSIX, ps, Python 3.11+ and a main-thread invocation. It monitors sampled process-tree RSS and logical regular-file bytes inside the work directory without following symlinks. An OS per-file size limit bounds individual child writes. It terminates the process group and observed descendants, including identity-checked individual signals where macOS rejects group signals. Time, RSS and storage watchdogs have polling and termination overshoot. RSS omits some GPU allocations; studies using MPS need a separate device-allocation limit. Unobserved detached descendants and writes outside the work directory require host isolation. This helper is not an adversarial sandbox, a filesystem quota, or a guarantee of instantaneous hard limits.

After termination, rc2 measures final persistent storage again. A nominally successful command above the cap becomes storage_exceeded; an existing failure retains its reason and also records final_storage_limit_exceeded. The final measurement contributes to the reported peak. This closes a fast-exit sampling race, while retaining the limits of polling and logical-byte accounting.

The evaluation controller must keep the policy, ledger and scorer outside candidate-writable paths. A study-owned ledger alone cannot prove resistance to candidate tampering. Deterministic process tests cover failure/retry history, timeout, RSS and storage termination, per-file limits, missing-outcome charging, live-job recovery refusal, budget reservation, profile validation and CLI argv.

Canonical source · Generated from the maintained repository document at this build’s commit.