Roadmap

The through-line: loop-generator's differentiator is verifiable outcomes — the checks are the spec, backed by layered false-positive defenses. The roadmap protects and extends that differentiator first, then improves the loop's feedback quality, usability, and ecosystem.

Status legend: planned (agreed, not started) · exploring (needs design) · deferred (revisit later). Items already shipped move off this page and into the docs.

Tier 1 — near-term, high leverage, low cost

Nothing open right now. The previous Tier 1 item — filesystem-walk change detection for non-git workspaces — shipped as the content-hash fallback (an independent hash walk of the working tree, with driver-reported files demoted to a secondary signal) and is documented in the trust model.

Tier 2 — medium-term, needs design

Human-in-the-loop controls exploring

The onIteration hook can't currently pause or abort; the only cancellation path is an external AbortSignal that stops after the current iteration. Planned: a review/step mode (show the diff + evaluations, wait for continue/abort/edit) and an iteration hook that can signal continue vs. abort. The biggest usability gap for higher-stakes or expensive runs.

Resolve the structured-feedback dead affordance shipped

FeedbackSummary.evaluations is passed to driver.run via invocation.feedback. Built-in drivers now consume it: CLI and SDK drivers append a machine-readable JSON block to the agent prompt; the mock driver can apply details.files fixes when useStructuredFeedback is set. Custom drivers can reuse augmentPromptWithStructuredFeedback from the public API.

A native coverage evaluator exploring

Only two evaluators ship today (command, experiment). Ship a native coverage evaluator plus a documented custom-evaluator recipe — the recipe matters more than any single built-in primitive. (The other half of this item — first-class capture of what happens inside a run — shipped as the observer plug-in point: --trace, the jsonl/otlp built-ins, and the AgentEvent driver seam.)

Deferred / not planned

• AST-query and log/event-assertion evaluators — speculative and large; the command evaluator plus the custom-evaluator plugin path already cover most cases. Let the documented recipe prove demand first.
• "Replay last iteration with an edited prompt" — niche; the pause/step control delivers most of the same value.
• Full third-party driver ecosystem (external-package driver loading, plugin discovery, a "publish your driver" guide) — the conformance harness is already exported (loop-generator/testing) and documented for extenders. A growth play, not a correctness one; revisit after Tiers 1–2 land.

What we are explicitly not trying to solve

These are fundamental to the agentic-loop approach, not implementation gaps. The roadmap makes their failure modes more visible; it does not eliminate them.

• Mis-specified or insufficient checks (the checks are the contract — bad checks, bad contract).
• Agents that game metrics or ship minimal patches that pass surface checks.
• Context degradation over many iterations (feedback diffs and human-in-the-loop controls mitigate; they don't cure).
• Tasks whose success is inherently subjective or requires human judgment.
• The underlying unreliability of LLM agents on large, novel refactors.