m0n0x41d/haft v5.3.0

on GitHub

Quint Code v5.3.0

FPF-native engineering decision tool. 6 MCP tools + interactive terminal dashboard. Works with Claude Code, Cursor, Gemini CLI, Codex CLI.

Install

curl -fsSL https://raw.githubusercontent.com/m0n0x41d/quint-code/main/install.sh | bash

See CHANGELOG for full details.

Interactive Terminal Dashboard

The headline feature: quint-code board — a Bubbletea v2 TUI that gives you a real-time view of your project's decision health without starting an agent session.

• Overview — health summary, activity timeline, depth distribution, module coverage, context map, evidence freshness
• Problems — backlog with drill-in to full signal/constraints/acceptance
• Decisions — list with R_eff scores and drift status, drill-in with glamour markdown rendering
• Modules — coverage tree with decision counts per module
• Live refresh — auto-updates every 3 seconds, no restart needed
• --check mode — exit code 1 if critical issues exist, usable as pre-push hook

Navigation: tab/1-4 switch views, j/k navigate, enter drill in, esc back, q quit.

Other Changes

• Decision mode computed from artifact chain — mode now reflects the actual reasoning path taken (problem with characterization + portfolio with comparison = standard minimum), not what the agent declared. Agent can escalate but not downgrade.
• FTS5 search keyword enrichment — search_keywords column on artifacts for agent-generated synonyms at write time. Accepted on quint_note and quint_decision.
• C/C++ header-only module detection — -I include directories registered as modules so dependency edges aren't dropped.
• /q-refresh scan rescans modules — keeps dependency graph fresh alongside drift and stale checks.
• Notes excluded from drift detection — observations don't need baselines.
• Module scanner excludes .claude and .context — worktrees and reference repos no longer inflate module count.
• q-reason context-aware entry — skill no longer always falls through into full FPF cycle. Three paths based on user intent: think-and-respond, prepare-and-wait (human drives), full autonomous. Default is prepare-and-wait.
• C/C++ symlink-safe include resolution — fixes silent edge loss on macOS symlinked checkouts.

Acknowledgments

Thanks to @karabelaselias for contributing C/C++ language support, reporting the variant parsing bug (#33), and raising the question about q-reason's interaction with anti-deskilling workflows that led to the context-aware entry fix.

What's Changed

• 5.3.0 by @m0n0x41d in #37