Watching the Verifier¶

Every enforcement mechanism in the plugin checks the output of an agentic session — commit and PR constraints, mutation tests, convention-drift garbage collection, the advocatus diaboli. None of them checks the state of the human who approves that output. A green checkmark at 09:00 and a green checkmark at 21:00 are the same colour in the merge log and carry the same authority, but they were not necessarily produced by the same quality of verification. The faculty that says yes, this belongs in our system runs on a finite human and has, until now, been unobserved by the harness that trusts it.

The cognitive-reservoir skill, the reservoir-warden agent, the /reservoir command, and the reservoir-check Stop hook together close that gap. They are the framework's honest-confidence principle applied reflexively: the framework finally observes the one actor it has always trusted blindly, and does so without pretending to a precision it lacks.

Why "stop when you feel tired" fails¶

The intuitive control fails because the metacognition that would notice the fatigue draws on the same capacity being spent. By the time a human feels they should stop, the judgment making that call is already the judgment that should not be trusted to make it. The corrective therefore has to be external and count/time-based, not judgment-based — it counts what it can see rather than asking the human how they feel.

Observable proxies, never a diagnosis¶

The mechanism counts only four observable proxies over a recent git window:

continuous session span — the vigilance decrement says sustained time-on-task reliably degrades performance;
decision volume — approval-like events (commits/merges) as the proxy for "times the human said yes";
context switches — distinct work streams touched; task-switching cost / attention residue says the cost scales with the switching, not the agent count, which is exactly what multi-agent orchestration accumulates invisibly;
wall-clock hour — mapped to a circadian band only when a chronotype is declared.

Every statement carries one confidence flag. Counts are observed. Risk read off the counts is inferred and always defeasible. Anything about the human — chronotype, whether a break was real rest — is asked and never assumed. The mechanism never combines the proxies into a single fatigue score: that would manufacture a precision the inputs cannot support.

Honest about contested science¶

The popular "decision fatigue" narrative rests on two contested pillars that this mechanism deliberately does not assert as fact:

Ego depletion — the strong "willpower is a finite resource that drains with use" model. A 2016 multi-lab Registered Replication Report (23 labs, N ≈ 2,141) found an effect of d = 0.04 with a confidence interval crossing zero. "The reservoir empties" is a useful metaphor — and the name of the skill — but not a measured mechanism.
The hungry-judges study — the headline "65% → near 0%" figure depends on case ordering that later work challenged. It must not be quoted as established.

What the design stands on instead is robust: task-switching cost, vigilance decrement, chronotype-dependent circadian variation, and (as suggestive context) Ericsson's 4–5 h/day deliberate-practice ceiling. Every output is framed as a precaution under uncertainty, never a diagnosis. This honesty is a hard requirement of the design, not editorial flavour.

Advisory only, never a gate¶

The mechanism is opt-in per project via a Cognitive reservoir block in HARNESS.md, which doubles as the marker the Stop hook greps for. It is deliberately not modelled as a Constraint. Constraints are gates with a scope that can fail CI; wiring a human-state advisory into a blocking gate would both defeat its purpose and overclaim a measurement the proxies cannot support. So it never blocks a commit, merge, or session, never fails CI, and never persists a record of the human's state to disk — the human edits the HARNESS.md block themselves.

When a threshold is crossed it surfaces the proxies and the inferred risk with caveats, then offers the one firm principle:

Decide your stop before the next session begins, while the judgment making the decision is still the kind you would trust. Do not negotiate the boundary with your tired self.

It pairs that with one concrete, time-boxed option — "re-review today's last two approvals tomorrow morning on a full reservoir" — and then goes silent. No lecture, no second nudge, no score.

The prohairesis stays with the engineer¶

The lineage is the Twentieth Watt essay and Epictetus: the governance of one's own cognitive state is the one thing the framework cannot do for the engineer. The prohairesis — the capacity for reasoned choice — remains theirs. The warden watches; it never chooses. A cluster of advisories a human routinely ignores is a signal to tune the HARNESS.md thresholds, not to weaken the honesty rule — and certainly not to take the decision out of human hands.