[sergo] Sergo Report: No-Tool-Change Precision Audit of Two CI-Enforced Linters - 2026-06-17

[github-actions[bot]]

Executive Summary

Run R39 found no tool change (Serena LSP stable at 23 tools R4-39; custom-linter registry stable at 29, no 30th linter). With no new linter to anchor on, this run executed a 50/50 cached + new strategy targeting two fresh, never-before-audited linters that are both CI-enforced — so any false positive is a build-blocker. Two distinct, single-file precision issues were filed.

Serena LSP tools	23 (stable R4-39)
Custom linter registry	29 (stable; cmd/linters/main.go:52-80)
Strategy	No-tool-change precision audit (50 cached pattern / 50 new target)
Findings examined	6
Issues created	2 (both CI-enforced linters)
Reconcile	R38 #39493 / #39494 confirmed open + unfixed-in-code
Success score	9 / 10

Tool / Registry Update

• Serena LSP: unchanged — 23 tools. activate_project --project warm.
• Linter registry: unchanged — 29 analyzers (multichecker.Main, cmd/linters/main.go:52-80). No 30th linter appeared.
• CI-enforcement correction: the R38 cache note dropped regexpcompileinfunction from the enforced set. The actual 14 enforced flags at .github/workflows/cgo.yml:1123 are: errstringmatch panicinlibrarycode manualmutexunlock osexitinlibrary rawloginlib regexpcompileinfunction fprintlnsprintf strconvparseignorederror jsonmarshalignoredeerror uncheckedtypeassertion fmterrorfnoverbs tolowerequalfold httpnoctx timeafterleak. Both of this run's targets are in that list.

Strategy: 50/50 Split

Cached half — proven syntactic_stdlib_match pattern. The strongest recurring land-arg in this project: a linter that identifies a stdlib call by comparing the package identifier's text (ident.Name == "pkg") instead of resolving package identity via the type checker. Already merged for sortslice (#38029) and open for ctxbackground (#38789). Applied here to a fresh linter.

New half — a never-audited, CI-enforced linter chosen for a parent-shape precision probe. uncheckedtypeassertion was read end-to-end (123 LOC); its comma-ok safe-form detection was found to recognize only one AST parent shape.

Expected outcome: two small, single-file fixes with proven-acceptable shapes (the team lands single-file precision PRs in 1-2 days).

Findings & Generated Tasks

Task 1 — uncheckedtypeassertion: comma-ok safe form only recognized for AssignStmt (false positive, CI-enforced)

• Location: pkg/linters/uncheckedtypeassertion/uncheckedtypeassertion.go:76-82 (parent check) and :99-101 (isSafeTwoValueAssertion).
• Evidence: the safe two-value form is only skipped when parents[typeAssert] is an *ast.AssignStmt. The Go comma-ok form is equally valid in initialization, so two safe constructs have a different parent and are wrongly reported:
  1. var v, ok = x.(T) — parent is *ast.ValueSpec.
  2. v, ok := (x.(T)) — parent is *ast.ParenExpr (the AssignStmt is a grandparent).
• Impact: the linter is CI-enforced, so the first idiomatic var v, ok = x.(T) fails the build on legal, safe code and forces a spurious //nolint. Latent today (no prod trigger).
• Fix: unwrap *ast.ParenExpr parents, then accept either an AssignStmt (len(Lhs)==2 && len(Rhs)==1) or a ValueSpec (len(Names)==2 && len(Values)==1).
• Effort: small, single file + testdata.

Task 2 — regexpcompileinfunction: package matched by identifier name, not type identity (FP + FN, CI-enforced)

• Location: pkg/linters/regexpcompileinfunction/regexpcompileinfunction.go:72-82 (isRegexpCompileCall), enforced at .github/workflows/cgo.yml:1123.
• Evidence: isRegexpCompileCall returns ident.Name == "regexp" && ... with no pass.TypesInfo. This is the same defect already merged for sortslice (sortslice precision: match sort.Slice/SliceStable via package identity (pass.TypesInfo), not the syntactic identifier name "sort [Content truncated due to length] #38029).
  • False negative: import re "regexp"; re.MustCompile("...") → ident.Name == "re" → missed.
  • False positive: a local/field named regexp of an unrelated type with a Compile/MustCompile method → flagged.
• Impact: medium — the FP path can block CI on code that never touches regexp; the FN path silently drops a real performance finding. Both latent today.
• Fix: resolve via pass.TypesInfo.ObjectOf(ident).(*types.PkgName).Imported().Path() == "regexp", mirroring the merged sortslice.go:58-67; pkg/linters/internal/astutil/astutil.go:72 already does the equivalent Imported().Path() check for "fmt".
• Effort: small, single function; thread pass through and add go/types.

Reconciliation (issues since last run)

• R38 ssljson (29th linter) graph-integrity gap: logic-step scene_id is parsed but never validated — orphan steps pass silently #39493 (ssljson scene_id graph-integrity FN) and ssljson precision: duplicate scene/step IDs are silently deduped by map[string]bool — the spec's "unique IDs" Pass-4 rule is uni [Content truncated due to length] #39494 (ssljson unique-IDs gap) — both open and still unfixed in code: ssljson.go:61 still parses SSLStep.SceneID with no validation rule consuming it. No re-file.
• No new sergo issues were opened by anyone between 2026-06-16 and this run; the phantom-issue reconcile (R37) did not recur.
• Long-standing open items left untouched per do-not-refile: execcommandwithoutcontext precision: false positive on nil-guarded exec.Command fallback — autofix injects a nil context that pa [Content truncated due to length] #38281/execcommandwithoutcontext enforce-readiness: propagate context in connectStdioMCPServer (2 sites), add nolint support, then enfo [Content truncated due to length] #38282 (execcommand, 7d+), doc-sync: linters/doc.go still says "All 24 active analyzers" after two new linters (hardcodedfilepath, timesleepnocontext) land [Content truncated due to length] #38787 (doc-sync resolved-in-code), hardcodedfilepath precision: hasFormatVerb misses flag/width/precision verbs (%05d, %-10s, %8.2f) and 9 verb letters (%x %t %c % [Content truncated due to length] #38788/ctxbackground precision: isContextBackgroundCall matches identifier name "context" syntactically while the same file detects the [Content truncated due to length] #38789, timeafterleak (28th linter): 3 prod loop+select time.After sites leak a timer per iteration — convert to time.NewTimer/Stop, the [Content truncated due to length] #39180/timeafterleak precision: a stored time.After channel (timer := time.After(d); <-timer) inside a loop+select escapes detection — [Content truncated due to length] #39181, timesleepnocontext precision: false positive on time.Sleep inside deferred cleanup closures #39323/timesleepnocontext enforce-readiness: zero production violations after the defer-cleanup FP fix — add to CI #39324.

Metrics

Metric	Value
Total runs	39
Cumulative findings	287
Cumulative tasks	78
Avg success score	8.79
This run	6 findings → 2 issues, score 9

Historical Context

The syntactic_stdlib_match pattern continues to be the highest-yield land-arg: sortslice #38029 merged, ctxbackground #38789 open, and now regexpcompileinfunction filed. A new sibling pattern was recorded this run — parent-shape-too-narrow — where a linter gates safe/unsafe on a single direct-parent AST type but the safe construct also appears under other parents (ValueSpec, ParenExpr); this is the FP analogue of the recurring alias_intermediate_var_FN.

Recommendations & Next-Run Focus (R40)

1. Reverify landing of sg39a1 / sg39a2 (single-file → expect 1-2 days).
2. Probe remaining never-audited linters for syntactic_stdlib_match (fprintlnsprintf, strconvparseignorederror, jsonmarshalignoredeerror, fmterrorfnoverbs, errstringmatch) — any ident.Name == <literal pkg> without TypesInfo.
3. Audit a fresh single-file linter (e.g. strconvparseignorederror, 88 LOC) for precision; avoid complex control-flow linters (execcommand-class lingers).
4. Re-check stale opens for closure/no-op: execcommandwithoutcontext precision: false positive on nil-guarded exec.Command fallback — autofix injects a nil context that pa [Content truncated due to length] #38281/execcommandwithoutcontext enforce-readiness: propagate context in connectStdioMCPServer (2 sites), add nolint support, then enfo [Content truncated due to length] #38282 (now 7d+), hardcodedfilepath precision: hasFormatVerb misses flag/width/precision verbs (%05d, %-10s, %8.2f) and 9 verb letters (%x %t %c % [Content truncated due to length] #38788/ctxbackground precision: isContextBackgroundCall matches identifier name "context" syntactically while the same file detects the [Content truncated due to length] #38789, timeafterleak precision: a stored time.After channel (timer := time.After(d); <-timer) inside a loop+select escapes detection — [Content truncated due to length] #39181, timesleepnocontext precision: false positive on time.Sleep inside deferred cleanup closures #39323/timesleepnocontext enforce-readiness: zero production violations after the defer-cleanup FP fix — add to CI #39324, ssljson (29th linter) graph-integrity gap: logic-step scene_id is parsed but never validated — orphan steps pass silently #39493/ssljson precision: duplicate scene/step IDs are silently deduped by map[string]bool — the spec's "unique IDs" Pass-4 rule is uni [Content truncated due to length] #39494.
5. Watch cmd/linters/main.go for a 30th linter.

References:

• §27667564805
• §27525869216

Generated by 🤖 Sergo - Serena Go Expert · ◷

• expires on Jun 17, 2026, 9:30 PM UTC-08:00