Host support and qualification
All hosts route to the same bundle catalog, canonical instructions and evidence
contracts. Generic delivery uses ALLAGMA.md. Codex uses study-local
.agents/skills/allagma-*/SKILL.md; Claude Code uses
.claude/skills/allagma-*/SKILL.md. Native wrappers select the campaign before
loading method text. Existing root instructions and user skills are preserved.
These paths follow OpenAI’s Build skills documentation and Claude Code’s skills documentation. Shared frontmatter follows the Agent Skills specification. The sources were rechecked during the self-audit on 7 October 2026. No provider-specific frontmatter, hooks, subagents, tool allowlists or model overrides are required by the default recipe. Optional metadata can be added in a host adapter when actually needed.
| Path | Qualification in v0.2 | What is not established |
|---|---|---|
| Generic/Python local | Actual complete toy execution, restart, recomputation and artifact audit | General scientific or model research quality |
| Codex CLI 0.160.1 | Real explicit skill activation, campaign-lock routing, interrupted/fresh-session continuation and complete MPS diffusion study | Other models/versions, implicit triggering, desktop UI interaction and general research quality |
| Claude Code | Format and contract checked; registration/conflict/campaign-routing fixtures | Real native activation/tool-use smoke; model quality |
The Linux/Python 3.13 onboarding receipt records the literal sparse-checkout tutorial and offline conformance on Debian 12/aarch64, CPython 3.13.16, in a local Docker Linux VM. This measures one container/interpreter combination; it does not extend native-host qualification or establish every Linux distribution, architecture or Python version.
The subsequent v0.3 comparison retains twelve distinct fresh CLI 0.160.1 sessions under the same inherited model/settings. All twelve complete required scientific execution; ten complete all evidence requirements. All six Allagma sessions read the exact locked recipe and seven methods. The corrected rc2 scientific broker separately passes real MPS training/checkpoint and access-denial checks. The frozen cohort includes the earlier allocator defect and cannot establish corrected GPU throughput. Other models, hosts, implicit skill triggering and broad research superiority remain unqualified.
The acceptance report records OS, Python, installed host versions when found,
date, capabilities and exact tested scope. Local inspection found codex-cli 0.151.0 and Claude Code 2.1.217 on 6 October 2026; these are version probes,
not assertions that either host ran the research workflow. The no-account kit
does not launch paid or authenticated model sessions.
On 8 October 2026 (Asia/Seoul), the separate
weight EMA diffusion study executed the
native path using the desktop app’s bundled CLI 0.160.1 and inherited
gpt-6-astra / max settings. No project model override was added. The older
0.151.0 CLI rejected that model, and its failure is retained. The native study
used actual MPS training, four pilots, ten new confirmation trajectories, a
deliberate interruption and fresh-session recovery, then analysis, audit,
independent recomputation and a reporting correction following native critique.
The host report and qualification receipt link exact prompts, JSONL traces, thread IDs, helper paths, model observations, locks and scientific outputs. This qualifies the recorded explicit workflow on that environment; it does not infer implicit skill-trigger reliability or general research quality. The default no-account conformance kit still does not launch native model sessions. Other hosts need their own execution evidence.
Without native skills, read or inject the exact generic files. Without native subagents, use sequential artifact handoffs. Without local execution or an explicitly equivalent runner, report the missing capability and continue only the applicable design work. Do not bypass qualification or fabricate results.
Canonical source · Generated from the maintained repository document at this build’s commit.