Verifies recorded decisions against measured evidence with drift detection and decay tracking.
npx skills add https://github.com/m0n0x41d/haft --skill h-verify
Please help me install this Agent Skill. Skill: h-verify Source: https://github.com/m0n0x41d/haft/tree/main/internal/cli/skill/h-verify Command: npx skills add https://github.com/m0n0x41d/haft --skill h-verify
Engineering decisions are often made and then forgotten, leaving teams unsure whether a past choice still holds after code changes, time passes, or evidence expires. This Skill closes the loop by re-checking recorded DecisionRecords against real measurements so stale or invalidated decisions are surfaced before further work relies on them.
Ask the agent to verify whether decision dec-20260525-abc still holds, or run it with no argument to check what is stale across the project.
High-intent search queries and answers about installing and using this skill.
Invoke the skill with a decision reference such as dec-20260525-abc, or with no argument to surface all stale decisions. It reads the decision's claims and predictions, runs each observable against its threshold, attaches the measurements as evidence, and records an accepted, partial, or failed verdict.
Run the skill without a decision reference. It calls haft_query status to surface refresh-due decisions and haft_refresh drain in dry-run mode to preview machine-safe closures, then applies safe closures and reports anything needing operator review.
The baseline action snapshots content hashes of a decision's affected_files. Later verification compares current files against that baseline, classifying changes as cosmetic, incidental, or material so you can decide whether to re-baseline or reopen the decision.
Skip it for one-off sanity checks where you can just run the test directly, and for re-framing the underlying problem, which belongs to the h-frame skill. It is meant for post-implementation reality checks of recorded DecisionRecords.
The kernel rejects measure calls made without prior evidence, since the protocol requires evidence before verdict. Attach at least one evidence item per material prediction using the evidence action, then record the measure verdict.
Congruence level rates how closely the evidence context matches yours: 3 for same-context (own tests, own production), 2 for similar contexts, 1 for different contexts like vendor benchmarks, and 0 for opposed contexts. It feeds the FPF trust calculus, so misclassifying it inflates the reliability signal.
samjhechtRun verification commands and capture outputs before any completion claim.
Satcomx00-x00Supersede an existing Architecture Decision Record with a new one.
ChiroHDAudit pod vaults to repair drift and refresh indexes.
wizelineGenerate formal test plan documents in Markdown or .docx from test suites and requirements.
