Iter 50's parseMcpScanText reverse-engineered `harness mcp-scan` text
output to extract structured findings. If upstream changes the format
(drops `[]` markers, switches to JSON, changes the Result: regex), the
parser silently returns empty findings and iter-49's introduced/cleared
diff regresses to the original dead-code state.
This iter adds a compat tripwire mirroring scripts/check-metaharness-
compat.mjs (which protects the @metaharness/router API surface).
WHAT SHIPPED
1. scripts/check-mcp-scan-format.mjs — runs `harness mcp-scan` against
ruflo and asserts 4 invariants:
✓ At least one ` [SEV] message` line in expected regex shape
✓ Result line matches `Result: SEV (N finding, M high)`
✓ parseMcpScanText loadable (sanity)
✓ parseMcpScanText extracts ≥ 1 finding from real output
Exit 0 — invariants intact OR metaharness not installed (graceful skip)
Exit 1 — invariant violated, parser will silently break
Exit 2 — unexpected error
The "extract ≥1 finding from real output" assertion is the
load-bearing check. It catches drift the regex above might miss
(e.g., if upstream switched to identical-looking but slightly
different whitespace).
2. .github/workflows/metaharness-ci.yml — new step in the
`metaharness-real-data` job runs the tripwire BEFORE the roundtrip.
If upstream format drifted, CI fails with a clear "format invariant
violated" message before chasing downstream symptoms.
3. plugins/ruflo-metaharness/scripts/smoke.sh — step 17z43 anchors:
- tripwire script exists + parses
- All 3 invariant text markers (finding/result/parser-roundtrip)
- CI workflow references the script
- Runtime: tripwire passes against the locally-installed metaharness
WHY TWO TRIPWIRES NOW EXIST
scripts/check-metaharness-compat.mjs (iter 12)
→ Protects @metaharness/router public API (constructor, route(),
fromExamples(), TrainedRouter.fromJSON())
scripts/check-mcp-scan-format.mjs (THIS ITER)
→ Protects `harness mcp-scan` text format (regex + parser-roundtrip)
These cover different upstream surfaces. The router compat-check ran
at the API level; the new check is at the CLI-output level. Both are
necessary because ruflo depends on both surfaces in different ways.
VERIFIED LOCALLY
$ node scripts/check-mcp-scan-format.mjs
✓ All 4 invariants intact.
GATES THIS ITER
smoke: 82 → 83 (step 17z43 green)
ci: new pre-roundtrip step in metaharness-real-data job
runtime: 4 format invariants pass against metaharness@0.1.14
PROTECTION RATIONALE
Iter 50 was a downstream workaround for issue ruvnet/agent-harness-
generator#16 (`harness mcp-scan` emits text only with --json). When
upstream eventually fixes that issue and emits structured JSON, the
parser's regex-based extraction will STILL work (the parser falls back
gracefully). But the tripwire will fail on the FIRST run after upstream
adds the JSON output AND drops the text format — which is a real risk
because upstream is in v0.1.x.
When the tripwire eventually fires, the fix is upstream-aware:
- If text format dropped → switch parseMcpScanText to read r.json directly
- If JSON added alongside text → prefer r.json, keep parser as fallback
The tripwire makes that decision visible at the PR that introduces it,
not 4 weeks later when someone notices drift detection is silent.
Co-Authored-By: RuFlo <ruv@ruv.net>
rUv