[sergo] Sergo Report: CI Config Sweep and Fix Verification - 2026-10-05

[github-actions[bot]]

Executive summary

Registry and Serena tool counts are both stable (76 analyzers, 24 Serena tools). The open sergo issue backlog dropped to zero this run: a maintainer bulk-closed all 14 issues from the R74-R86 reconcile sweeps (13 confirmed landed-fixed, plus the timenowsub refile from the previous run) and cleaned up one junk test-probe issue. With nothing left to reconcile, this run spent its budget verifying the recent fixes hold up and sweeping a part of the codebase never swept before: the CI workflow configuration itself. That surfaced one confirmed, still-unfixed reverse-phantom bug, filed as a single new issue.

Tool and registry status

• Serena: 24 tools, unchanged, reconfirmed live via serena --help.
• Linter registry: 76 analyzers (grep -c Analyzer,$ pkg/linters/registry.go), unchanged since R85 (fprintferrorunchecked, PR [linter-miner] Add fprintf-error-unchecked linter #65073).
• No new linter to audit fresh this run.

Strategy: 50/50 cached vs new split

Cached component (verification, not refiling): with zero opens to reconcile, the usual refile work from the bounded 3-run memory window did not apply. Instead, re-read pkg/linters/internal/resourcetracker/resourcetracker.go and pkg/linters/bufioscannererunchecked/bufioscannererunchecked.go end to end to stress-test the fixes landed since R85/R86 (the per-acquisition-position state map in resourcetracker, and the rest-threading in bufioscannererunchecked). Both held up under harder scenarios (branch-local reassignment, nested blocks) - audited clean, no regression found.

New-exploration component: a first-ever full diff of .github/workflows/cgo.yml native vs wasm LINTER_FLAGS strings (previously only spot-checked per newly-added linter, never swept end-to-end). This is new ground: no prior run had compared the two lists directly against each other.

Finding: contextcancelnotdeferred missing from wasm CI enforcement

The native custom-linter CI step and the GOOS=js GOARCH=wasm step run the same linter suite with two separately-maintained LINTER_FLAGS strings that are supposed to be identical (modulo the wasm-only LINTER_PACKAGES scoping). They are, except for one flag: -contextcancelnotdeferred appears in the native string only.

This turned out to be a rediscovery, not a new finding: issue #55932 (temporary_id sg61a1) filed this exact gap on 2026-08-26 and auto-expired closed not_planned on 2026-09-02, without ever being fixed. It was never refiled in the 25 runs since, because the topic (CI workflow config) never happened to fall inside any single runs bounded 3-entry strategy-memory window - a blind spot in how this workflow samples its own history. Re-reading the current file confirms the gap is still present today, more than five weeks and dozens of edits later.

The root cause traces to the introducing commit (PR #49783, 2026-08-02), which added both contextcancelnotdeferred and wgdonenotdeferred to the native list but only added wgdonenotdeferred to the wasm list - a copy-paste omission at the moment the linter was introduced. It also was not caught by the dedicated drift-prevention test (TestCIEnforcedLintersMatchRegistry in pkg/linters/doc_sync_test.go, added by issue #55636 specifically to guard registry-to-CI drift): that test unions both LINTER_FLAGS matches into a single set before checking registry membership, so an analyzer enforced on only one of the two build targets still satisfies the check. The guard cannot detect this class of asymmetry by construction.

Impact is currently latent - no *_wasm.go file under the 5 wasm-linted packages uses context.WithCancel/WithTimeout/WithDeadline today - but any future wasm-only code using a cancellable context and a manual (non-deferred) cancel would pass CI silently on both targets.

Filed as a single issue (sg87a1) recommending both the one-line cgo.yml fix and hardening the drift-guard test to assert set equality between the two LINTER_FLAGS lists, not just their union.

Generated tasks

1. Add -contextcancelnotdeferred to the wasm LINTER_FLAGS string in cgo.yml, and extend TestCIEnforcedLintersMatchRegistry to assert the native and wasm flag sets are identical (modulo -test=false and LINTER_PACKAGES) so this class of drift cannot silently reappear.

Metrics this run

• Findings: 4 (zero-opens reconcile, resourcetracker audited clean, bufioscannererunchecked audited clean, wasm LINTER_FLAGS asymmetry).
• Issues created: 1 (sg87a1).
• Success score: 8/10.

Historical context

87 runs total, 537 cumulative findings, 154 cumulative tasks, running average success score 8.73. This is the first run in recent history to reconcile against a fully empty open-issue backlog - the previous standing veins (paren_unwrap_gap, branch_state_merge, presnapshot_position_blind, idiom_reflag_no_allowlist, block_scope_state_reset, ident_only_equality_gap) all closed out in the R86 sweep. A new pattern class, ci_target_asymmetry, has been added to the shared pattern catalog.

Recommendations for maintainers

• Apply the sg87a1 fix (one-line workflow change plus a test hardening) when convenient; it is low effort and closes a real, if currently latent, blind spot.
• Consider a periodic non-linter-logic sweep (doc.go/README/cgo.yml cross-file consistency) as a standing addition to this workflows rotation, since that is exactly how sg87a1 resurfaced after 25 runs of silence.

Next-run focus (R88)

Verify sg87a1 landed. Expect exactly one open sergo issue at reconcile start - the cleanest slate in this workflows history. If the linter registry grows past 76, audit the new linter fresh; otherwise continue rotating cgo.yml/doc.go consistency sweeps alongside the usual per-linter bug hunts.

References

• §37261626096

Generated by 🤖 Sergo - Serena Go Expert · claude · agent · 271 AIC · ⌖ 7.23 AIC · ⊞ 5.2K · ◷

• expires on Oct 5, 2026, 8:11 PM UTC-08:00

[github-actions[bot]]

This discussion has been marked as outdated by Sergo - Serena Go Expert.

A newer discussion is available at Discussion #66010.