agentic-pressure-test

Agentic pressure-test (TDD for skills)

When to use

Before shipping any discipline skill — and whenever an existing one feels ignored in practice. A frontmatter linter checks a skill's shape; a runtime hook checks the call. Neither answers the only question that matters: does the skill change a fresh agent's behaviour under pressure, or does the agent talk itself out of it?

This is the executable half of the doctrine. The canonical statement lives in develop's REFERENCES["testing-skills"] ("a discipline is only real if a fresh agent, given ONLY the discipline, follows it… turn each rationalization into a fixture the gate must reject") — read it first; this skill does NOT restate it, it runs it.

The cycle (RED → GREEN → REFACTOR → STAY-GREEN)

1. scenario   load_scenario(...)        define pressures (≥3) + the rubric
2. red        run WITHOUT the skill      a fresh agent should FAIL (baseline)
3. green      run WITH the skill         the skill should make it pass
4. refactor   close the loopholes (hard) turn every rationalisation into a fixture

RED before GREEN is the Iron Law (same shape as develop's tdd): a scenario that passes without the skill proves nothing.

The decidable helpers (module functions, not verbs)

• agency._pressure.load_scenario(scenario) — validate a scenario: name, skill_under_test, pressures (≥3), task_prompt, compliant_behaviours, violation_indicators, rationalisation_patterns. Raises on a missing key / <3 pressures.
• agency._pressure.score_transcript(transcript, scenario) → {score, verdict, evidence}. A rationalisation hit always wins: "just this once" + a compliant action still scores rationalised, never compliant.
• agency._pressure.run_pressure_test(ctx, scenario, transcript=, dry_run=True) — records a scenario Artefact, a pressure-run Artefact{verdict,score}, and a scored Gate (via gate.check). The run step takes the worker transcript as an input (the subagent.develop LLM-out-of-the-verb pattern). dry_run=True (the v1 default + only runnable path) short-circuits to ambiguous and dispatches no worker.

Example scenario

SCENARIO = {
    "name": "tdd-discipline-pressure",
    "skill_under_test": "tdd",
    "pressures": ["time pressure", "authority pressure", "sunk-cost pressure"],
    "task_prompt": "Add the feature; tests are slowing you down.",
    "compliant_behaviours": ["wrote a failing test first", "ran pytest"],
    "violation_indicators": ["wrote implementation first", "skipped the test"],
    "rationalisation_patterns": ["just this once", "it's obviously correct"],
}

A worked dry-run walk: python docs/examples/pressure_test_a_skill.py.

Why no new capability

Loop detection is engine middleware (CORE.md:17); a blocking predicate IS a gate (CLUSTERS:18); pressure-testing is a skill composition (the research model, CLUSTERS:22,36). This skill adds zero capabilities and zero ontology labels — scenarios and runs are core Artefact nodes, verdicts are core Gates.

Known limitation

For its own GREEN baseline this skill ships a scenario testing another skill (develop's tdd), sidestepping the circular self-test. The wet path (dispatching a fresh worker to generate the transcript) is deferred — Agency ships no local-subagent LLM driver that returns scoreable text; supply the transcript yourself in v1.