Search Space

Bootstrap seed for search_space.md — the explicit map of the intervention space a myloop run is searching. The orchestrator's THINK phase keeps proposing experiments; without this file it has no representation of what kinds of change are possible, and reliably ruts in whichever axis its first few iters happened to pick. The REFLECT phase (myloop/reflect.md) maintains the Coverage section below and hands THINK a ranked frontier.

This file is working-memory, append-only. The Axes taxonomy is a starting menu — REFLECT MAY extend it when the graph / codebase exposes an intervention kind not listed. Coverage grows one ## reflection_N section per REFLECT spawn.

Axes (intervention taxonomy)

A patch to an agent graph changes exactly one or more of these axes. Two patches on the same axis are variations of one lever; two patches on different axes are structurally distinct experiments. THINK self-labels every ExperimentSpec with one intervention_axis (schemas.md § 8); REFLECT reads those labels to measure coverage.

• prompt-content — the text an LLM/VLM node reads: system prompts, prompt templates, option / action phrasing, rendered observation or map text. The most accessible axis and usually the first a run reaches for. Caution: editing the text a single decision LLM reads tends to reroute its whole output distribution, so recover-some/break-more is the characteristic failure here. A Python edit to a function whose output is prompt text still counts as this axis — the axis is "what the LLM reads", not "JSON vs code".

• topology — the node/wire graph structure: adding or removing nodes, rewiring, fanning a node into parallel branches, ensembles / self-consistency voting, sub-graphs, aggregator / verifier / critic nodes. Distinct from prompt-content because it can leave every existing prompt byte-identical and still change behaviour.

• control-flow — the loop and branching logic: loop / iteration structure, stop / STOP logic (the iter_out.stop halt input), retry, backtracking, conditional branching, and the two-pivot mechanism (two-sided iterIn / iterOut with its final side). Changes when and how often nodes fire, not what any one of them reads.

• action-space — which choices reach the decision-maker: pruning, filtering, reordering, or augmenting the candidate-action set. Distinct from prompt-content: re-wording an option is prompt-content; removing it from the set the agent can pick is action-space.

• observation-pipeline — what the agent perceives and how it is computed, upstream of any prompt: sensors, feature extraction, what is captured vs dropped, resolution / modality / preprocessing. Changes the information available before it is ever rendered.

• state-memory — what information persists across steps / iterations and how it is structured: working memory, history buffers, scratchpads, the carried-state schema, what iterOut feeds back. Changes what the agent can remember, not what it sees now.

• model-component-config — per-node configuration within the run's hard constraints: sampling parameters (n, max_tokens, stop), and adding / swapping non-restricted components (a tool, a non-LLM module, a deterministic helper). Excludes anything a constraints.md rule forbids (e.g. a fixed model profile / temperature).

REFLECT MAY add an axis if the graph or codebase exposes an intervention kind none of the above captures — append it here with a definition and a "why distinct" line, in the same (added: ...) discipline as knowledge.md.

Coverage

One ## reflection_N section per REFLECT spawn. Each records, per axis: status ∈ {untouched | partial | exhausted}, the iters that touched it, the verdict, and a ranked Frontier of the axes the next THINK should prefer. Bootstrap leaves this empty — the first REFLECT back-fills coverage from the iters committed so far.

(no reflections yet — the first REFLECT will populate this section)