vigil-isolate

SKILL: Isolate — narrow the fault surface

When to use

Load when reproduction.md is schema-valid and stable. Unload when fault-surface.md has ≤8 candidates with evidence anchors.

Memory check: if recall fired in Phase V, its results are already in context — consult them when building the initial suspect set. Procedural memories from prior missions on similar error classes can shortcut Source 2 (recent churn) and Source 3 (scope bounds) below. If mcp__crystalium__* tools are unavailable, proceed without memory — never hard-fail.

---

Contract

Field	Value
LLM calls permitted	For classifying candidates and reading source
Tool budget	≤25% of mission budget
Output	fault-surface.md — schema-valid per schemas/fault-surface.v1.json
Hard cap	≤8 candidates. If more survive reduction, narrow scope; if fewer than 1 survives, escalate.

---

The Suspect Set — Where to Look First

Build the initial suspect set from three deterministic sources. Do not enumerate the entire codebase.

Source 1 — Stack trace closure

Every file referenced in the failing stack trace (from reproduction.md) is a suspect. For each frame:

• The file itself → suspect
• Files it imports → secondary suspects if the failure is a RUNTIME_ERROR or INTEGRATION_ERROR
• Call-graph ancestors (who calls this) → tertiary suspects for LOGIC_ERROR

Source 2 — Recent churn

For the commit at which the failure reproduces:

git log --since="30 days ago" --oneline --name-only

Cross-reference churn with Source 1. Files that are both in the stack trace AND recently modified are the highest-priority suspects.

For REGRESSION category: add git bisect to identify the introducing commit. Files touched by that commit become top suspects.

Source 3 — Scope bounds from upstream

If upstream artifact (APIVR-Δ repair-failed-report.md, spec, or scout-report from ATLAS) declared a scope, respect it. Candidates outside the declared scope require explicit justification (e.g., "symptom traces cross the boundary because of [evidence]").

---

Reduction Techniques (Category-Specific)

REGRESSION — use git bisect

Bisect between last known good commit and current. Each iteration:

1. Check out midpoint
2. Run reproduction command
3. Record PASS or FAIL_MATCH
4. Narrow

Terminates at the introducing commit. The diff of that commit is the highest-priority candidate set.

LOGIC_ERROR, TYPE_ERROR, RUNTIME_ERROR — use dependency walk

Starting from the failing assertion's location:

1. Identify all variables/functions read in the failing code path
2. For each, identify its origin (where it's assigned, returned from, passed in)
3. Walk backward until you hit a boundary — user input, IO, constant, or unmodified upstream code
4. Each node in this walk is a candidate

INTEGRATION_ERROR — use contract comparison

The failure is at a component boundary. Candidates are:

• The producer's output schema/contract (actual vs declared)
• The consumer's input expectations (actual vs declared)
• Any serialization layer between them

BUILD_ERROR, LINT_VIOLATION — direct diagnosis

The tooling has already identified the file:line. The "isolation" is trivial. Proceed to Graph with the tooling-reported location as the single candidate.

HEISENBUG — use timing/state differential

With statistical-mode reproduction, run 5× with increased logging. Compare state at key checkpoints across runs:

• What differs between failing and passing runs?
• Are there race windows? Shared state mutations? External clock/random dependencies?

Candidates are the loci where observable state diverges between passing and failing runs.

COMPOUND — accept >1 root candidate

Compound failures have multiple independent roots. Do not force reduction below 2 candidates. Mark them for the Graph phase to confirm independence.

SPEC_DEFECT — validate test intent

Does the test assert the correct thing? Compare test expectation to:

• The spec (if available — from SPECTRA handoff or issue tracker)
• Other tests covering the same behavior
• Recent spec changes

If the test is wrong, the "fix" routes to SPECTRA, not APIVR-Δ.

---

Ruled-Out Trail — First-Class Evidence

For every candidate you remove from the suspect set, record why. This is not noise — it's the most important output of Phase I.

Required ruled-out records:

ruled_out:
  - candidate: "app/flows/auth.rb:120-180"
    reason: "Not in stack trace; not modified in last 90 days; called from unrelated path."
    evidence_source: "git log + stack trace"
  - candidate: "lib/utils/clock.rb:*"
    reason: "Failing test uses frozen time via test helper; clock utility not reached."
    evidence_source: "test_helper.rb:12; dep_graph_query(callers_of: Clock.now)"

Minimum: 3 ruled-out entries per mission. Missions with fewer suggest lazy reduction; harness flags for review.

---

Writing fault-surface.md

mission_id: VIGIL-YYYYMMDD-NNN
upstream_reproduction: <ref>
suspect_set_sources:
  - stack_trace_closure
  - recent_churn
  - scope_bounds
candidates:
  - id: C-001
    path: "app/flows/vote_casting/record_vote.rb"
    lines: "42-78"
    kind: code            # code | test | config | schema | data | env
    suspicion_rank: 1     # 1 = highest
    rationale: "Top of failing stack; modified 3 days ago in commit abc123; writes the asserted-against column"
  - id: C-002
    path: "app/policies/voting_authorizer.rb"
    lines: "12-40"
    kind: code
    suspicion_rank: 2
    rationale: "Delegate target of C-001; recently modified; appears in stack frame 2"
reduction_trail:
  - step: "Stack trace closure produced 17 suspects"
  - step: "Churn filter reduced to 8"
  - step: "Scope bounds (auth module only) reduced to 5"
  - step: "Dependency walk confirmed 2 additional candidates outside initial stack"
ruled_out:
  - <as above, minimum 3 entries>
scope_bounds:
  include: ["app/flows/**", "app/policies/**"]
  exclude: ["spec/**", "lib/legacy/**"]

---

Anti-patterns

• Reading entire files. Windowed reads only, after narrowing by symbol or graph query.
• Inferring candidates. Every candidate needs a path:lines anchor and a traceable reason.
• Cascading reduction too aggressively. If you're down to 0 candidates, you reduced wrong — escalate.
• Chasing a candidate because it "looks suspicious." Anchor the suspicion in stack trace, churn, or dep graph. Vibes are not evidence.
• Skipping ruled-out records. The next VIGIL mission on related code will thank you.

---

VIGIL Phase I — reduction with evidence, ≤8 candidates to Graph