Validate

Validate - the combined Arcgram self-check entry

One call. Two questions answered between the siblings: is the flow well-formed and coherent, AND do its @spec values still match the code?

Validate is a thin wrapper, not a new check. It imports [Checkpoint](../checkpoint/SKILL.md) and [Reconcile](../reconcile/SKILL.md) untouched and composes their pure functions:

• Checkpoint (one input: the diagram) checks structure - well-formed, coherent.
• Reconcile (two inputs: diagram + code) checks values - @spec literals == code constants.

Run both -> the diagram is a faithful, current spec. Every finding still comes from one sibling; Validate only merges, source-tags, and unions the attestation.

Code is optional

| input | what runs | clean possible? | |---|---|---| | validate(diagram, { code }) | Checkpoint and Reconcile | yes | | validate(diagram) (no code) | Checkpoint only | no - values unchecked |

Without code there is nothing to reconcile, so Reconcile is not run (running it would just flag every @spec pin as missing). Instead the value half is reported as one uncovered value-check-not-run finding, so a structure-only run is never falsely "clean" on values - the same anti-silent-pass rule both siblings enforce.

Locked invariant - uncovered is a third state, never a pass. A diagram-only run (no code) can never produce an overall green clean, even when the structure is perfect. The attestation explicitly appends | code consistency: NOT CHECKED (no code supplied), summary.codeConsistency reads "NOT_CHECKED", and the CLI verdict is INCOMPLETE, never the word CLEAN. This guards the exact silent-pass the family exists to prevent: claiming "checked & clean" while the code half was never checked.

Output (same contract as both siblings, plus source)

{ findings:[{id,type,severity,note,source}], summary, attestation, clean }
self-check ran HH:MM:SS | N findings | M unchecked

• source is "checkpoint" | "reconcile" | "validate" - so you can route a finding back

to the check that raised it. Everything else (id, type, severity, note) is verbatim from the sibling.

• attestation is the union: N and M are the combined defect / uncovered totals. When the

value half did not run it appends an explicit | code consistency: NOT CHECKED (no code supplied).

• clean === checkpoint.clean && reconcile.clean (identically N=0 && M=0); uncovered is a

third state and is never folded into clean, so a skipped value-check is not clean.

• summary nests each sibling's own summary under .checkpoint / .reconcile, plus ran and

codeConsistency ("checked" | "NOT_CHECKED").

• The full sub-results are returned verbatim on .checkpoint / .reconcile for a caller that wants

the per-check breakdown.

Human reads the marker; agent reads the type (and source).

Use

node validate.mjs  [code-file ...] [--tier-b | --no-tier-b] [--json]

import { validate } from './validate.mjs';
validate({ nodes, edges, meta }, { code });   // code: string | string[]; omit -> structure only

--tier-b / --no-tier-b forward to Checkpoint's justification tier (else auto-detected by a decision diamond). Exit 1 on any defect (advisory - never blocks; an uncovered value-check is not a defect).

Verify: node validate-selftest.mjs (36 checks - source tags, union attestation, clean rule, the code-omitted path, and byte parity with running the two siblings separately). Example: examples/example-workflow.html vs example-workflow.config.cs (clean).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jovesun-lab
• Source: jovesun-lab/arcgram
• License: Apache-2.0
• Homepage: https://arcgram.io

Install and usage instructions live in the source repository linked above.