Presets — curated settings for two kinds of models

Two profiles encode hard-won tuning for the two worlds agents live in: frontier models (big, capable, expensive) and constrained models (small, local, cheap — and easily distracted). Source: src/presets.rs.

---

ConstrainedProfile — the small-model kit

Built on four assumptions about small models: they drift off goal, repeat tool calls, ship broken edits, and emit malformed arguments. Each assumption gets a counter-measure:

Assumption	Counter-measure
drifts off goal	a GoalReminder contributor re-sends the original task every 5 turns
repeats read-only calls	a memoize layer caching Read/Glob/Grep/LS (TTL 5 turns)
ships broken edits	a verify layer after Write/Edit/MultiEdit
emits fat/loose payloads	output cap (16,384 chars), strict tool schemas, tight context window

Applying it:

use loopctl::presets::ConstrainedProfile;

let mut agent = BareLoop::new(
    client,
    tools,
    ConstrainedProfile::session_config(),      // context window 32,768
);
agent.set_request_options(ConstrainedProfile::request_options());  // ToolConstraint::Strict
ConstrainedProfile::apply(&mut agent)?;        // compaction manager + middleware + reminder
let result = agent.run(task, &ConstrainedProfile::run_config()).await?;  // max 100 turns

Three pieces, three ways they reach the loop — config at construction, options via setter, and apply() for everything that needs the live loop. apply() itself runs exactly three steps, in order: (1) build a ContextManager around the default truncating compactor, synced to the session’s window and threshold (whatever the constructor seeded is replaced by the values your SessionConfig actually carries); (2) install the middleware pipeline; (3) register the goal-reminder contributor. It deliberately does not touch the loop’s config or request options — those are the other two pieces, and keeping them separate means the profile composes with your own construction code instead of fighting it.

The pipeline it builds, in order (first = outermost):

OutputLimitMiddleware(16_384)     ← caps everything below
VerifyMiddleware(NoopVerifier)    ← appends [verify] blocks after writes
MemoizingMiddleware(TTL 5)        ← caches raw results innermost
[core: your registry]

The order is the design: memoize innermost caches the raw result (cache never holds verify text; every write is verified anew); the cap outermost truncates the combined output (verify diagnostics can’t escape the cap).

The upgrade path is deliberate: NoopVerifier verifies nothing until you swap in a real one (cargo check, tsc, your test suite) — then every agent write gets proven or disproven on the spot. Same for the NoopPathExtractor: replace it with a real path extractor and cache invalidation becomes file-aware.

Gotcha: middleware you chain onto pipeline_builder() lands inside the memoize layer — the wrong slot for anything that must see the original call (e.g. redaction). Add such layers when installing the pipeline, after the preset’s layers.

FrontierProfile — the explicit baseline

Everything default: 200k window, 200 turns, no pipeline, no constraints. It exists so “no profile” can be an explicit, self-documenting choice rather than an omission.

GoalReminder — useful on its own

The shipped contributor behind the profile’s anti-drift measure:

agent.add_contributor(Box::new(GoalReminder::new(5)));

Re-emits the first user message verbatim as a system reminder every N turns (skipping turn 0; n: 0 disables). Sensible values 3–10. See contributors.

---

Which profile when

Situation	Profile
7B–30B local model, coding agent	Constrained
Frontier API model (GPT-4o / Claude / Gemini class)	Frontier — add layers à la carte as problems appear
Frontier model doing lots of file reading	Frontier + your own memoize layer
Any model whose writes you must trust	Either + a real Verifier

---

Related pages

• Middleware — the layers the profile stacks.
• Contributors — GoalReminder’s home turf.
• Session config — what the 32k window means.