Reasoning & FPF
Haft is built on the First Principles Framework (FPF) by Anatoly Levenchuk — a rigorous, transdisciplinary architecture for thinking. The skill set gives your coding agent an FPF-native operating system for engineering decisions: framing before solutions, characterization before comparison, parity enforcement, evidence with congruence penalties, weakest-link assurance, and a cycle that reopens itself when evidence ages or a measurement fails.
Seven MCP tools
| Tool | What it does |
|---|---|
haft_note | Micro-decisions — atomic facts with typed anchors, validation, auto-expiry |
haft_problem | Frame problems, declare comparison dimensions with indicator roles |
haft_solution | Explore variants with diversity check, compare under parity |
haft_decision | Decision contracts: invariants, claims, evidence, baseline lifecycle |
haft_commission | WorkCommission lifecycle for optional execution harnesses |
haft_refresh | Lifecycle management for every artifact kind |
haft_query | Search, status dashboard, code graph (callers / callees / impact / explore, each reached symbol fused with the decisions governing it), FPF spec search |
Fifteen skills installed by haft init
| Skill | Mode | What it does |
|---|---|---|
| h-reason | auto (umbrella) | Full FPF reasoning palette — framing, exploration, comparison, verification, notes, plus slideument patterns (Goldilocks, NQD, BLP, Scaling-Law Lens). Manual /h-reason always works. |
| h-frame | auto | Frame a problem with stabilize + problem typing + umbrella-word repair |
| h-diagnose | auto | Diagnose a failure with parallel hypothesis testing (one subagent per hypothesis to prevent anchoring) |
| h-explore | auto | Generate distinct candidate variants with NQD diversity discipline |
| h-compare | auto | Fair comparison with dim-wise parallel scoring + Pareto front (not a scalar winner) |
| h-decide | manual | Record a binding DecisionRecord with full DRR — Transformer Mandate |
| h-verify | auto | Baseline → measure → evidence loop with drift detection |
| h-status | auto | Read-only project FPF state dashboard |
| h-onboard | auto | First-frame ceremony for projects new to haft |
| h-spec-cover | auto | Spec-coverage check with blind/stale module triage |
| h-note | auto | Lightweight micro-decision recording |
| h-commission | manual | WorkCommission lifecycle — manual per Transformer Mandate |
| h-abduct | subroutine | Pure B.5.2 abductive four-step (frame → ≥3 rivals → filters → prime) |
| h-boundary-unpack | subroutine | A.6.B decomposition of boundary statements |
| h-semio-review | subroutine | Concept-rename / spec-consistency audit (X-FANOUT-AUDIT) |
Auto-triggering skills fire when their description matches operator context.
Manual-only skills (h-decide, h-commission) require
explicit invocation per the Transformer Mandate — binding artifacts come from the
human principal, not the agent. Subroutines are called from other skills or invoked
explicitly when working a specific FPF sub-discipline.
Routing reliability is testable: haft check routing runs 40 golden
prompts (current pass rate 82.5%).
Semantic FPF spec search
haft fpf search (and haft_query(action="fpf") from MCP)
searches the indexed FPF specification. Retrieval is hybrid: exact pattern id first,
then keyword (FTS5) fused with semantic recall over baked section vectors. A reworded
"how do I think about X" finds the pattern that answers it, even when it shares no
keywords with the pattern's own heading.
The vectors are baked into the binary at build time, so end users never embed the
spec themselves. Query embedding runs through the optional local
haft-embed sidecar; when the sidecar is absent, semantic recall
degrades cleanly to keyword search — it never hard-fails.
haft fpf search "how do I find the part of a design that breaks first"
# → X-WLNK (Weakest-Link analysis), by meaning, not keywordEvidence workflow
Attach evidence with haft_decision(action="evidence", ...). Evidence
carries formality levels (F0–F3), congruence levels (CL0–CL3), and expiry dates.
Trust scores (R_eff) degrade as evidence ages; stale evidence triggers refresh. Use
haft_decision(action="measure", ...) for post-implementation
verification, or the /h-verify skill from the host agent.
See also Enforcement check for how drift, staleness, and evidence decay roll into one CI gate.