Module step 

What a finished step actually did — the deterministic half of step appraisal (docs/GOAL-SYSTEM-DESIGN.md §5.5).

A step is marked done by the agent, and nothing checked it. The symmetry with the tier above is the whole argument: a board task is closed by the owner (TASK-AGENT-DESIGN.md D6), so a person is the check; a todo step is closed by the model, so there is no person and the check has to be structural. D5’s rule — state is derived from the record, never self-reported — reaches one tier further down than it was written for.

Pure, and unit-tested rather than trialled, for compact.rs’s reason: getting it wrong is silent. A finding that fires on honest work is a line the model learns to skip, which is how a check that protects nothing survives; a finding that never fires is indistinguishable from a plan that always lands.

Two readings only, and the omissions are deliberate. §5.5’s table lists five signals; the two here — no calls at all and the last call did not succeed — are facts about the span. The other three (a span far longer than its siblings, a verify-shaped call that passed, the same target read repeatedly) are comparisons, and each needs either a threshold nobody has measured here or a guess about what a tool call meant. A threshold that cries wolf is doctor’s named failure, and the escalation to a model that would settle the ambiguous cases is rung 7’s, not this one’s. The same-target reading is boredom’s (§9.1) and belongs one mechanism over.

What this module does not do is act. The finding is rendered onto the todo result and the plan action — accept, revise the step, revise the plan, escalate — is the model’s, because the plan is the model’s. The harness has no way to author a decomposition and no business having one.

Structs

Span

What happened between a step starting and the model calling it done.

StepEscalation

What the quarantined escalation call is handed. See the module note above on why the step’s own text is safe to include here in a way an appraiser’s evidence (§5.1) could not be.

Work

The run’s work so far, as of one tool call.

Enums

EscalationReason

Which comparison flagged a landed step for a second opinion.

Finding

What the span says about the step.

Outcome

How one executed call ended, as far as the run’s own record knows.

StepVerdict

What the escalation decided: nothing further, or a plan-revision nudge is worth surfacing.

Constants

STEP_ESCALATION_STEM

Marks a folded nudge as the harness’s own words, on boredom::NOTICE_STEM’s exact precedent: agent::is_harness_voice is a closed list the learning miner filters every tool-result message’s text through before deciding whether it is a user’s Steer/Followup intervention. Without an entry here, templated_nudge’s output — folded into the very message escalation_candidate also carries tool results in — would be mined as if a person had typed it, escalation.step and all, and could ride into a future prompt as a Clean-origin learned rule derived from nobody’s words.

Functions

appraise

The deterministic reading. No model, no threshold, no tuned constant.

escalation_candidate

The escalation’s own pre-filter: cheap, deterministic, and it only ever decides whether to ask, never the answer.

escalation_prompt

The prompt the quarantined pass runs. Reasoning first, the typed field last — the front door’s and the appraiser’s own finding: constrained output degrades reasoning when the answer precedes the thinking.

next_run

A fresh run identity. Monotonic within the process, meaningless outside it.

parse_step_verdict

Parse what the escalation returned.

templated_nudge

The nudge folded into the run when the escalation says revise_plan.