[kemalelmizan]

The bug

The REPL built its runAgent call field by field, and the list was incomplete:

const result = await runAgent({
  model: options.model,
  tools: options.tools,
  executor: options.executor,
  messages: next,
  maxSteps: options.maxSteps,
  decoding: options.decoding,
  selectModel: options.selectModel,
  decide: options.decide,
  stream: options.stream,
  onEvent: render,
});

steer and stopGuard were never in it. The consequence is invisible and wrong:

• --plan-interval <n> planned nothing in the REPL.
• --mode tasks could end a turn with the shared list still unfinished.

Both work in a one-shot run, so the flags look healthy; only the interactive path was broken. It is the same shape as a mode whose tools are wired but whose runner is not (#69): the capability exists and the host drops it on the way to the loop.

I found it while adding the planning-cost report (#74): the /steps footer asserted in the e2e never appeared, because no planning call ever happened in the REPL.

The fix

Structural, not a third field on the list:

// ReplOptions extends the loop's options, so a new seam is forwarded by default
export interface ReplOptions extends Omit<RunAgentOptions, "messages" | "onEvent" | "executor"> {
  executor: AgentExecutor & Partial<ApprovalHost>;
  // …the REPL-only options
}

const result = await runAgent({ ...options, messages: next, onEvent: render });

A new seam added to the loop now flows to the REPL automatically; dropping one has to be written down. The ReplOptions doc comment records why, and docs/architecture.md states the rule where the opt-in seams are described.

Verified

• Two unit tests pin the seams that were missing: a steer whose text must reach the model's next prompt, and a stopGuard that must earn a second model call. Both fail against the old hand-picked list (checked by temporarily restoring it).
• One e2e test drives the real REPL over stdin and asserts a --plan-interval 1 turn plans, records the plans, and reports planning: plans=… plan_tokens=… plan_ms=… in /steps. The driver is answer-driven — readline.question drops anything typed while no prompt is pending, so a single stdin dump only ever runs the first line.
• pnpm build, pnpm typecheck, pnpm lint, pnpm test:coverage — green (331 passed / 5 skipped; coverage 90.3 / 84.1 / 89.3 / 90.3).

Note

The /steps footer spacing (planning: plans=… instead of planning:plans=…) is fixed here rather than in #74: the footer could not appear before this change, so the cosmetic problem had no way to be seen.