git:20260812.c7a7642 to git:20260821.28538dc

27 added, 6 removed. Audit A to A.

---

name: dekko-verify

- description: Sanity-check a suspiciously low or zero call-graph result from dekko (get_callers, get_callees, find_usages, impacted_tests, unused) before concluding "no callers" or "dead code." Trigger whenever such a result looks surprising given the symbol's apparent importance, or before deleting/renaming a symbol based on a zero-caller result, or when the repo mixes languages/has any unsupported-language files.

+ description: Sanity-check a suspiciously low or zero call-graph result from dekko (get_callers, get_callees, find_usages, impacted_tests, unused) before concluding "no callers" or "dead code," or a heritage/throws-provenance result (query supertypes, query subtypes, query throws) that labels something (external) when you expect it to be in-repo. Trigger whenever such a result looks surprising given the symbol's apparent importance, before deleting/renaming a symbol based on a zero-caller result, when the repo mixes languages/has any unsupported-language files, or when an (external)/(unresolved) label lands on a name you're confident is first-party code.

---

# Verifying a low-confidence dekko answer

dekko's call-graph resolution is real but conditional — repeated

hands-on evaluation rounds (see this repo's own

`test-repos/reports/`) keep finding the same failure shape: a

**confident wrong answer**, not a visible error. A caller trusts "0

callers" more than it should. This skill exists to catch that before

it leads to deleting live code or missing a real impact.

## When to double-check before trusting a result

Reach for one targeted `grep -rn <name>` (not a full re-read) as a

sanity check — not a full re-verification — when any of these apply:

- **A cross-package/cross-module qualified call is involved.**

`pkg.Func()`-style calls (Go), `namespace::func()` (C++), or any

call where the receiver is a same-repo package/module rather than a

local variable are a known resolver blind spot — confirmed missing

4 real call sites on a live repo as recently as this project's own

round-13 eval. Same caution applies to trait/interface dispatch

(Rust `dyn Trait` calls, Java/Kotlin interface methods) — the

resolver ladder only reliably matches an explicit `Type::method()`

or `Type.method()` form.

- **The result doesn't disclose ambiguity.** A real ambiguous call

should say so (`N call(s) resolved ambiguously`), not just be

silently absent from the count. If a symbol you expect to be widely

used shows a low count with *no* ambiguity disclosure, that's more

suspicious than a low count *with* one.

- **The repo has any unsupported/partially-parsed language files** —

check `dekko stats` or the map-build summary for an "unsupported"

note. Files dekko can't parse are tracked (not silently dropped),

but a symbol only ever called from an unparsed file will still read

as zero-caller.

- **`get_callers`/`get_callees` used their default `--no-tests`

filter.** An empty result may just mean "no *non-test* callers" —

check whether `include_tests`/`--include-tests` was applied before

concluding dead code.

- **A dense-repo common short method name** (`new`, `then`, `map`,

`iter_mut`, or similarly generic names in a 10k+-symbol repo) —

resolver precision degrades under high symbol density; treat a

count from these as directional, not exact.

- **You're about to delete or rename based on `dekko unused`'s

dead-code list.** Same blind spots apply; a callback passed

by reference rather than called directly, or a call from an

unparsed file, can both read as "no inbound calls."

+ - **A heritage or throws-provenance result labels something

+ `(external)`.** `query supertypes`/`subtypes` and `query throws`

+ can mislabel an in-repo type-alias-as-heritage-base or a

+ pattern-bound rethrow as a fake external entry when the extractor

+ doesn't yet model that language's specific syntax shape (confirmed

+ historically on TS `implements <type-alias>` and Java 16+

+ `instanceof`-pattern rethrows — both since fixed, but the general

+ failure shape, a present result that's mislabeled rather than a

+ missing one, can recur in a new syntax shape any of dekko's language

+ extractors hasn't seen yet). An `(external)`/`(unresolved)` label on

+ a name you're confident is first-party code is worth a `query symbol

+ <name>` check before trusting it.

## What "good" looks like

If the grep sanity check agrees with dekko's count (or turns up

nothing dekko missed), trust the structural answer and move on —

this is a spot check, not a mandate to re-verify every query. Most

dekko answers on supported languages with unambiguous calls are

correct; this skill is for the specific conditions above, not a

blanket "always grep after dekko."

## Boundaries

- - This is about **call-graph relation tools** (`get_callers`,

- `get_callees`, `find_usages`, `impacted_tests`, `unused`) —

- `outline`/`query_symbol`/`search_code` describe what's in the repo,

- not what calls what, and don't share this failure mode the same

- way.

+ - This is primarily about **call-graph relation tools**

+ (`get_callers`, `get_callees`, `find_usages`, `impacted_tests`,

+ `unused`) plus **heritage/throws-provenance tools**

+ (`query supertypes`/`subtypes`, `query throws`) — anything that

+ resolves a relationship or provenance rather than just describing a

+ symbol's own shape. `outline`/`query_symbol`/`search_code` describe

+ what's in the repo, not a resolved relationship to something else,

+ and don't share this failure mode the same way.

- One grep, scoped to the symbol name, is enough to sanity-check —

don't fall back to reading whole files or re-deriving the call

graph by hand; that defeats the point of using dekko at all. See

`dekko-orient` for the general "reach for dekko before grep" guidance

this skill is a narrow exception to.

+ - For the call-graph relation tools specifically, `dekko sanity

+ <target>` (`dekko sanity <target> --usages` for a `find_usages`-style

+ check) automates the exact grep sweep this skill describes — same

+ comparison, same blind-spot causes, run deterministically instead of

+ by hand. It doesn't cover heritage/throws-provenance mislabeling yet;

+ the `query symbol <name>` check above is still manual for that case.