runtime-development

Files (2)

• SKILL.md
• runtime-architecture.md · references/

SKILL.md (first lines)

# Runtime development

Use this skill for functional changes under `rlm/core/`, `rlm/domains/`,
`rlm/backends/`, or `rlm/api.py`.

## Why it matters

Runtime design separates cheap, deterministic inspection from model judgment,
and keeps large source material out of the parent context. That separation is
the product's answer to context rot, escalating inference cost, and unsupported
claims about an unfamiliar codebase.

Inspect the closest implementation and focused test first. Read
[`references/runtime-architecture.md`](references/runtime-architecture.md)
only when a change crosses those module boundaries or changes recursion,
budgets, or data binding. Load `../testing/SKILL.md` for validation details.

Preserve data locality: source bytes stay in the REPL/container, while model
calls receive only metadata or a narrow slice. Prefer Python analysis over a
model call when it can determine the answer.

## Execution architectures

Repository and corpus execution is selected by `Config.architecture`, not by
ad-hoc conditionals in the API or runtime. Keep each architecture in
`rlm.core.architecture` behind the `ResearchArchitecture` protocol and return
only a `PreparedArchitecture` (scope, validated plan, and deterministic
fallback targets). This keeps experimental architectures independently
selectable and comparable.

`direct` is the normal REPL-led path. `planned` builds a source-free scope and
executes only a validated plan; `planner_enabled` remains its compatibility
alias. A new architecture must preserve the prompt ceilings, source-free final
rendering, target enforcement, budget inheritance, and trajectory visibility.
Do not make a new architecture an unbounded fan-out mechanism.

For planner-enabled work, treat the scope manifest as the admission boundary:
the planner may select only IDs, while runtime code resolves targets and
executes their fixed leaf/child route. Keep final rendering source-free (only
compact findings), and make failed planning fall back to a manifest-restricted
view rather than an unrestricted domain.

From williamntlam/recursive-language-model, MIT.

More from williamntlam/recursive-language-model