Loop Mode

Loop mode is Inferoa's loop-engineering surface for recursive long-horizon work. Run /loop to define the outcome once; Inferoa keeps inspecting, changing, testing, verifying, deciding, and continuing until the work is proven.

Use it when a task may span multiple turns, context compaction, tool failures, verification passes, or a later resumed session.

When To Use It

Use loop mode when the desired outcome is clear but the work is long:

• the agent needs to keep working until an objective is complete;
• progress needs an internal checklist, evidence, and status;
• completion should not depend on a single assistant turn;
• you want the session to preserve the objective across interruptions.

Do not use loop mode as a substitute for planning ambiguous scope. If the task needs approval before edits begin, start with Plan mode.

Intelligent Model Selection

With INFEROA_MODE=auto and vLLM Semantic Router, each loop turn can route to a different model. The TUI shows selected: <model> / <decision>.

See Model endpoints for setup.

Basic Commands

/loop Improve the docs site and verify the Docusaurus build.
/loop run deliver Improve the docs site and verify the Docusaurus build.
/loop run deliver --at-least 24h Improve this package and handle related high-value issues.
/loop run discover Reduce benchmark latency without hurting accuracy.
/loop run replay --count 100 say hi to me
/loop status
/loop pause
/loop resume
/loop drop

/loop status displays the active loop, current loop task, attempts, verification, skills, pending review state, and latest loop decisions.

Preference And Runtime

Bare /loop <objective> starts the creation flow. Inferoa asks for the objective, a preference, runtime, and optional human-in-the-loop review.

Preference:

• Deliver closes an end-to-end objective with planning, execution, verification, and decision passes.
• Discover runs autonomous research and lets the agent choose benchmarks, metrics, harnesses, controls, and comparison shape.
• Replay repeats the original visible prompt for a fixed attempt count.

Runtime:

• Auto lets the agent stop only after the loop's coverage, frontier, evidence, and residual-risk state has been reconciled.
• At least keeps the loop running until the minimum duration is satisfied, then the normal decision and verification gates still apply.

How It Works

The active loop stores:

• the original objective;
• preference (deliver, discover, or replay);
• runtime policy (auto or at_least);
• an internal loop task plan and step status;
• the current loop task, starting with a Deliver or Discover bootstrap;
• attempts, which are runs interpreted as work on the loop task;
• verification records from commands, research metrics, checker runs, human review, or structured model evidence;
• a candidate ledger of open, completed, and rejected work;
• notes, resources, tool traces, skill snapshots, token usage, tool usage, and time usage;
• the latest loop decision.

The agent should keep step status and evidence current while working. An empty checklist is not enough to finish the loop.

For Deliver loops, the delivery contract starts with a coverage inventory derived from the objective's risk surfaces. Mapping a repository, reading a topology document, or naming a module does not make that surface covered. A surface must be covered with evidence, or explicitly rejected with a rationale or accepted residual risk.

Decisions And Completion

When the current loop task appears exhausted, Inferoa runs an internal decision pass. The decision pass steps back from the current plan and asks whether more work is needed to satisfy the original objective.

The decision has three useful outcomes:

• expand: open a new loop task with concrete steps that materially affect the original objective;
• done: record evidence-backed semantic completion;
• blocked: pause because user input or an external state change is required.

For broad loops, completion is also gated by the candidate ledger. If a decision says done while high-value candidates remain open, Inferoa expands the next loop task instead of silently finishing.

Completion is also gated by coverage debt. If the current loop task closes its seeded frontier but coverage surfaces remain pending or in_progress, Inferoa opens a coverage-continuation loop task instead of completing. If a reflection packet names residual risk, that risk must be persisted in loop state before completion can pass.

Loop completion is gated by verification and loop decisions. A loop is not done because the checklist is empty; it is done after verification records evidence-backed semantic completion.

Discover Loops

Discover loops reuse the same loop supervisor, but each loop task is optimized for research. The agent chooses the benchmark shape, metric, harness, controls, and comparison path from the workspace and task evidence. A loop task can create, continue, complete, or reject multiple experiments. Each experiment represents one hypothesis or solution line, and each run records benchmark output and parsed METRIC name=value evidence.

Research completion requires logged metric evidence. The loop cannot complete while a benchmark run is pending, and a done decision should cite run history, the best observed metric, and guardrail or regression evidence.

Self-Improve

/self-improve and inferoa self-improve turn verified loop evidence into a reviewable workspace skill proposal. The first implementation uses structured replay/gating over recorded evidence rather than a live model rerun:

/self-improve help
/self-improve status
/self-improve propose
/self-improve run --replay
/self-improve report
/self-improve adopt

Self-improve artifacts are stored under .inferoa/self-improve/, and adopted skills are written under .inferoa/skills/.

Relationship To Other Modes

Use Plan mode before loop mode when the scope needs approval.