the skill-body flag control's skill set is a hand list: a newly added skill body is not covered #67

The finding

#66 landed the control that holds this repo's skill bodies to their own Kong grammar (candy/plugin-box/skill_doc_flags_test.go). The org's validator accepted it and recorded the fragility on the thread:

Non-blocking fragility worth noting for a future change: the skill set is a hand-maintained slice, so a newly added skill body is not auto-covered; and the list case assumes that body documents no flag.

This issue carries that finding to its root fix.

The RCA

• Mechanism — the control walks skillDocCases, a hand-written slice of ten {skill, command, grammar} entries (candy/plugin-box/skill_doc_flags_test.go:81-91), and reads each body from ../../.agents/skills/<skill>/SKILL.md. A skill body added to .agents/skills/ LATER is not in the slice, so it is never read: the control is green over a body it has never seen. The set of bodies in the repository and the set the control reads are two lists, and only one of them is in the repository.
• Missed control — the control's subject is the set of skill BODIES; the set it enumerates is a literal in the test file. Nothing compares the two, so the drift is invisible until a body documents a flag its plugin does not accept — the exact defect the build skill documents --stage-parallelism / CHARLY_STAGE_PARALLELISM — spellings this plugin's grammar does not accept #65 and refactor(cli)!: the build concurrency flag is --stage-parallelism, and a control holds every skill body to the grammar #66 were about, re-entering silently. This is the same shape as the umbrella's hand-maintained retired-claim key set (the umbrella's own gate does not assert #647's post-condition: a clean tree with an absent nested level passes charly task verify opencharly#710), which is why that one is COMMITTED and scanned.
• Blast radius — this repo's ten skill bodies, and every repo that adopts the control by copying it.
• Root fix — the set of BODIES must come from the repository (readdir .agents/skills), and the code fact that stays hand-written is the mapping from a skill to the command and the Kong struct it documents. The control then asserts BOTH directions: every body in the tree has a case (an uncovered body is RED), and every case names a body that exists (a stale case is RED).

MEASURED: the repo's .agents/skills/ holds exactly the ten directories the slice names — build generate inspect list load merge new pull reconcile validate — so the control is complete TODAY. That is exactly why the gap is invisible: the slice and the tree agree by hand, and the next skill body breaks the agreement without a word.

The list case

charly box list dispatches its subcommands by hand (dispatchList), so it has no Kong grammar and its body may document no flag. The control already asserts that generically — a nil grammar yields an empty flag surface, so any flag the body offers is reported. Under the root fix it stops being an assertion about one body and becomes the property of any body whose command has no flag surface.

How it will be proven (R7)

The control must be able to fail, and the new arm must be the one that was missing:

• a fixture tree with an EXTRA skill directory that has no case → RED ("a skill body is not covered");
• a case naming a directory that does not exist → RED ("a stale case");
• the ten real bodies → GREEN.

The repo's own Go gate (gofmt, go vet, go test -count=1), charly box validate and the R10 bed check-box-local run on the committed tree, and the arms are pasted as the output, not described.

Change classification

• Class: Go source work in this repo (a test), plus the fixture-comment correction recorded on the refactor(cli)!: the build concurrency flag is --stage-parallelism, and a control holds every skill body to the grammar #66 thread (charly.yml's "did NOT look beside the candy" wording is over-stated: the lookup resolves the project root, so the fixture's move to the project root is functionally safe — the comment says more than the lookup does).
• Attribution tier: analysed on a live system — the finding rests on the LIVE repository: the slice's ten entries (skill_doc_flags_test.go:81-91) against the tree's ten skill directories, and the validator's own report quoted above. No code is changed by this issue; the tier describes the basis of the analysis.

Agent: charly · session ses_8efb271d-f2c0-47d3-85e3-2a1e00b11ae0

Assisted-by: Claude Code minimax/minimax-m3.1-flash:cloud (analysed on a live system)