h-abduct - SKILL.md Agent Skill

name: h-abduct description: | INTERNAL SUBROUTINE — used by h-diagnose for parallel rival-hypothesis generation. Manual invocation possible but the right user-facing entry point is almost always h-diagnose (which uses h-abduct internally with parallel testing). Generates ≥3 typed rival explanations for an observed signal per FPF B.5.2 abductive cycle. Do not auto-select this skill — when failure investigation is needed, select h-diagnose; when problem framing is needed, select h-frame. when_to_use: | Internal subroutine only. Use h-diagnose for failures; use h-frame for problem framing. argument-hint: "[abductive prompt: signal or observation requiring explanation]" disable-model-invocation: true allowed-tools: mcphafthaft_problem mcphafthaft_solution mcphafthaft_query

h-abduct — Pure abductive four-step (subroutine)

You are running the B.5.2 abductive four-step micro-cycle: frame the prompt → generate candidates → apply plausibility filters → select prime hypothesis. Per CC-B.5.2-2 rivals stay visible; per CC-B.5.2-1 every run starts from a declared typed AbductivePrompt.

Explicit-only — disable-model-invocation: true. The skill is invoked manually when the operator wants the discipline without the parallel-test overhead of h-diagnose.

Step 1 — Frame the typed AbductivePrompt

State the initiating prompt precisely. Per B.5.2.0 the typed species are:

AnomalyStatement — " doesn't fit existing explanation"
ProblemCuePrompt — "<sketch of a problem we don't yet know how to solve>"
OpportunityCuePrompt — ""
ProbeCuePrompt — ""

Capture the prompt in writing before generating candidates — otherwise the abduction floats.

Step 2 — Generate ≥3 candidate hypotheses

Per CC-B.5.2-2 at least one rival required. Aim for 3-5 candidates that differ in kind, not degree. Each candidate states:

Claim (one sentence)
Why-plausible (one sentence)
Discriminating probe — what would falsify or confirm

If all candidates trend in one direction (all blame the same component), force a rival from a structurally different direction (data flow, race condition, environmental, configuration, infrastructure, etc.).

Step 3 — Apply plausibility filters (≥2 declared)

For each candidate, evaluate against explicit filters. Typical:

Parsimony — does the candidate introduce only the additional structure the prompt requires?
Explanatory reach — how much of the prompt does the candidate actually account for?
Consistency — does the candidate avoid collision with already-trusted pillars / mechanisms / scope declarations?
Falsifiability / probeability — does the candidate create a path for deduction, testing, contrast, or evidence acquisition?
Scope fit — is the candidate framed for the declared prompt scope rather than for an inflated or shifted target?

Pick at least two filters and apply them. Record the rationale.

Step 4 — Select prime hypothesis + publish

Pick one candidate for downstream work. Per CC-B.5.2-1 publish:

The selected prime hypothesis as a new U.Episteme at AssuranceLevel:L0
Linked back to the original AbductivePrompt
With selection rationale referencing the applied filters

In haft kernel this translates to either:

mcp__haft__haft_solution(
  action="explore",
  problem_ref="<prob-... if existing or framed>",
  variants=[
    // 3-5 variants, prime hypothesis marked in description
  ],
  no_stepping_stone_rationale="<if no stepping stone among candidates>"
)

OR (lighter — just record the abductive cycle without portfolio):

mcp__haft__haft_problem(
  action="frame",
  signal="<prime hypothesis statement>",
  problem_type="diagnosis",
  acceptance="prime hypothesis confirmed or falsified by <discriminating probe>",
  mode="tactical"
)

The lighter path is for exploratory abduction where the operator doesn't yet want full portfolio comparison. The heavier path is when subsequent /h-compare is expected.

Step 5 — Hand off

Surface to operator:

Prime hypothesis
Rivals (per CC-B.5.2-2 stay visible)
Filter rationale
Recommended next step:
- /h-verify if the prime hypothesis is verifiable against existing evidence
- /h-explore if rivals deserve deeper variant exploration before commitment
- /h-decide (manual-only) if the prime is binding and ready to commit

What NOT to do

DO NOT generate only one candidate — rivals are required (CC-B.5.2-2).
DO NOT skip plausibility filters — at least two declared, per CC-B.5.2.
DO NOT collapse to a "best guess" without documenting why other candidates were rejected.
DO NOT publish prime as L1+ — abductive output is L0 (unsubstantiated until tested).
DO NOT auto-invoke from anywhere except explicit operator command or h-diagnose subagent spawn.

FPF spec references

B.5.2 — Abductive Loop (parent procedure)
B.5.2.0 — U.AbductivePrompt (typed entry surfaces)
B.5.2.1 — Creative Abduction with NQD
CC-B.5.2-1, CC-B.5.2-2 — Conformance: typed prompt + visible rivals
EXP-01 (abductive loop), EXP-04 (WLNK per variant)

Look up via mcp__haft__haft_query(action="fpf", query="B.5.2").