Author a module

A small contribution should involve one directory and a registry entry. methods/allagma-context-active-brief/ is the minimal executable reference: portable instructions, module.yaml, a thin select.py helper and an input fixture. The full-record alternative keeps the same input/output contract.

1. Copy the directory to a new allagma-* name.
2. Set the matching name in SKILL.md and a new stable id in module.yaml.
3. Describe the behavioral difference and activation conditions. Keep host commands, model pins and native tool names out of shared instructions.
4. Retain or explicitly migrate the input/output contract. Declare required capabilities, all resources and dependencies, settings, owner, lifecycle, source lineage, adaptation, example and evaluation scope.
5. Register the ID/path in registry.json, or use an explicit study-local overlay. Start as experimental until qualification supports a stable scope.
6. Run the targeted module check and the executable example. Test the behavior you changed; broad research claims need a suitable comparison.

python3 -m allagma check --module context/active-brief
python3 methods/allagma-context-active-brief/select.py \
  methods/allagma-context-active-brief/example.json /tmp/context-example.json
python3 -m allagma validate-record /tmp/context-example.json

The thin context helper uses its adjacent manifest’s ID in the output, so a derived method retains its own producer identity without copying the selector implementation. To change the algorithm, supply a helper implementing the same ContextRecord contract and update the example. Source checks validate naming, frontmatter, capability/dependency closure and resource paths.

Recipes and adapters

A recipe has portable instructions and recipe.json: roles, steps with method IDs and consumed/produced contracts, branches, stopping conditions and supported settings. recipe/research and recipe/replication share methods; replication requests two reanalysis passes instead of one. The helper implements this small declared control surface, not an arbitrary workflow language. New loop engines must supply their own helper as a resource and qualify its semantics.

Adapters map capabilities and artifact handoffs to a host or local runner. reviewer/checklist and reviewer/trace accept the same input and produce the same verdict/findings/coverage handoff, then the framework records ReviewRecord. The second checks additional dependency digests. Both are explicitly deterministic examples, not real-model reviewers.

For a host adapter, provide capability mapping, installation paths, missing- capability behavior, fake-host fixtures and usage notes. Native provider settings belong in its generated packaging. A host version probe alone is not activation qualification. Required scientific checks cannot disappear as an undocumented fallback.

Improvement and lifecycle

Run python3 -m allagma compare --output work/comparison-1 to inspect a bounded candidate/proposal/measurement/decision example. Keep controller, evaluator, task split and resources fixed apart from the factor under study. Record cost, noise, intervention and complexity. Preserve rejected and unevaluated candidates; an environment failure before evaluation is not evaluated.

Adoption, replacement, retirement and default selection are separate decisions. The local comparison does not change defaults. See governance for support-state requirements and deprecation windows. Adding another module is not itself an improvement metric.

Reverse the comparison with --baseline context/active-brief --candidate context/full-record to produce the measured cost-based rejection example. Each comparison retains the actual candidate/baseline files, common runtime, fixed fixtures, execution trace and decision. A backend failure produces a not evaluated record with its partial trace and does not become a quality score.

---

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