vigil-graph

SKILL: Graph — build the Information Dependency Graph

When to use

Load when fault-surface.md has ≤8 candidates. Unload when idg.md identifies ≥1 ROOT_CANDIDATE or escalates with [GAP].

---

Contract

Field	Value
LLM calls permitted	For edge inference when AST/graph queries are insufficient — with explicit evidence requirement
Tool budget	≤15% of mission budget
Output	idg.md — schema-valid per schemas/idg.v1.json
Critical rule	Every edge must cite trace evidence or AST-derived fact. Speculative edges are rejected.

---

Why Dependency Graph, Not Temporal Order

The most common attribution error in agentic debugging is treating the earliest observed symptom as the root cause. Research is unambiguous here: GraphTracer (arXiv:2510.10581) demonstrated +18.18% attribution accuracy over state-of-the-art temporal methods by using information dependency graphs; CHIEF (NeurIPS 2024) showed that hierarchical causal graphs outperform flat log analysis; the Lifecycle of Failures study (arXiv:2509.23735) quantified the lift at 46.3% → 65.8%.

The principle: A failure propagates through a dependency chain. The first node to show the failure is usually downstream of where it originated. Temporal order tells you where to start looking; dependency order tells you where it came from.

---

Node Construction

Each candidate from fault-surface.md becomes a node. Each node records:

id: N-001
candidate_ref: C-001           # from fault-surface.md
path: "app/flows/record_vote.rb"
lines: "42-78"
kind: code                     # code | test | config | schema | data | env | tool_output
observed_state: |
  At the point of failure: ballot.token = nil
expected_state: |
  At the point of failure: ballot.token = valid UUID
observation_source: |
  trace:stack_frame_1 + assertion_message + view_file(record_vote.rb:56)

Hard rule: observed_state must come from the reproduction trace, stderr, or direct file inspection. expected_state comes from the test assertion, spec, or type signature. If either cannot be anchored, the node is not admissible — return to Isolate.

---

Edge Construction

Edges are directed, A → B, meaning "B's state depends on A's state". An edge requires one of:

Edge type	Evidence required
Data flow	AST/call-graph shows A's return value flows to B's input, OR trace shows A writes state that B reads
Control flow	A's execution precedes and conditions B's execution within the failing path
Shared state	A and B both read/write a named resource (DB table, global, file); trace confirms timing of access
Contract boundary	A produces output that B consumes across a boundary (API call, message queue, RPC)

Prohibited edges:

• Edges based on "probably calls" inference without AST or trace evidence
• Edges from proximity alone (same file, same class, similar name)
• Edges derived from unit test structure (what a test happens to set up is not a dependency)
• Bidirectional edges (cycles handled separately — see below)

Edge-building procedure

For each node B:

1. Use dep_graph_query(readers_of: <B's inputs>) to find what writes B's state
2. Use dep_graph_query(callers_of: <B's entry points>) to find what triggers B
3. Check trace for any node A whose output logically precedes B's state
4. Record only edges with at least one anchor from steps 1-3

---

Identifying Root Candidates

A root candidate is a node in the graph with zero incoming edges among other candidate nodes.

Formal procedure:

1. Construct the graph G from NODES and EDGES.
2. For each node n: if in_degree_within_candidate_set(n) == 0, mark as ROOT_CANDIDATE.
3. Compute descendant_count(n) = number of candidate nodes reachable from n.
4. Rank ROOT_CANDIDATES by descendant_count descending.

Interpretation:

• Top-ranked root candidate = the node whose failure could explain the most symptoms.
• Low-ranked root candidate = may be real but explains fewer downstream failures; likely a compound-failure co-root.
• Nodes with incoming edges = SYMPTOM_NODES. Mark with [SYMPTOM]. These are not root causes — do NOT generate hypotheses for them unless all root candidates are falsified.

---

Handling Special Graph Shapes

Cycles

If the graph contains a cycle, one of two things is happening:

1. False edge — one of the edges in the cycle is not evidence-backed. Re-audit; likely an inferred edge that should be removed.
2. Genuine feedback loop — recursion, mutual callbacks, or circular state dependency. In this case: halt, emit [DISPUTED], escalate to human review with the cycle documented. VIGIL does not auto-resolve feedback loops.

Disconnected subgraphs

If fault-surface.md candidates form 2+ disconnected components, this is a signal for COMPOUND failure. Each component has its own root candidate. Classify the mission as compound; the Intervene phase will address each component independently.

Single-node graph

If only one candidate survived Isolate, the graph has one node with no edges. It's the root candidate by default. Proceed to Intervene, but note in idg.md that no dependency analysis was performed (it was vacuous).

No root candidates

Every node has at least one incoming edge → the true root is outside the candidate set. Escalate to Isolate with expanded scope, or emit [GAP] and escalate upward.

---

Writing idg.md

mission_id: VIGIL-YYYYMMDD-NNN
upstream_fault_surface: <ref>
nodes:
  - id: N-001
    candidate_ref: C-001
    path: "app/flows/record_vote.rb"
    lines: "42-78"
    kind: code
    observed_state: "ballot.token = nil at line 56"
    expected_state: "ballot.token = valid UUID"
    observation_source: "trace:stack_frame_1 + view_file"
    role: symptom       # symptom | root_candidate | both
  - id: N-002
    ...
edges:
  - from: N-003
    to: N-001
    kind: data_flow
    evidence: "dep_graph_query(writers_to: ballot.token) → N-003 at line 24"
  - from: N-002
    to: N-001
    kind: control_flow
    evidence: "trace shows N-002 executes immediately before N-001 within same call"
root_candidates:
  - node: N-003
    descendant_count: 2
    rank: 1
  - node: N-004
    descendant_count: 0
    rank: 2
symptom_nodes: [N-001, N-002]
graph_shape: normal       # normal | cyclic | disconnected | single_node | no_roots
disputed:
  - null                  # populated if cycles or evidence conflicts
compound_components:      # only if disconnected
  - null

---

Pitfalls

• Treating the failing test's location as the root cause. The test is the observer of failure, rarely its source.
• Asserting an edge because "it makes sense." Makes-sense is not evidence. Cite the AST query or the trace line.
• Ranking by suspicion instead of descendant count. Suspicion is Isolate's job; Graph ranks by structural reach.
• Skipping [SYMPTOM] marking. Downstream phases must not waste interventions on symptom nodes; mark them explicitly.

---

VIGIL Phase G — structure beats sequence for attribution