crystalium

Load at the start of the APIVR-Δ Analyze phase. Techniques for building a repo map, running progressive disclosure, and maximising context relevance for coding tasks. Use whenever you need to understand unfamiliar code before editing it, or when an existing Discovery Report feels under-grounded.

Phase A (Abstract) — AgentFold compression of the Locate trajectory into a ≤2000-token working-memory summary plus a Memex excerpt store. Use after findings.md is complete and before synthesis. Runs in a clean-context subagent to prevent trajectory contamination. Mechanically validates that every FINDING-XXX, GAP-XXX, and ESCALATION_TRIGGER is preserved.

Phase L (Locate) — bounded probes and scatter subagents to answer DECISION_TARGET sub-questions. Use after Phase T (map.md) is complete. Descends a symbol → graph → lexical → windowed-read → test-dry-run ladder; emits FINDING-XXX records with evidence anchors and confidence tiers. Hot phase (≤60% of mission budget).

Phase S (Synthesize) — emit the scout-report.md artifact that downstream agents (SPECTRA, APIVR-Δ, human) consume. Use after Phase A (Abstract) fold is validated. Hard cap 3000 tokens; every factual clause cited with FINDING-XXX; every recommended action carries a handoff label.

Phase T (Traverse) — deterministic structural mapping of a codebase. Use when a mission.md brief exists and before any meaning-based search. Zero LLM calls during retrieval; only symbol index, AST, ripgrep, and git log. Emits map.md with entrypoints, modules, call graph, and heatmap.

Decision-ready specifications via the SPECTRA cycle. Use for complex features (complexity ≥7/12), multi-component/service changes, ambiguous requirements needing structured decomposition, or any task where 'just start coding' would require significant rework. Produces a dual-format spec (Markdown + YAML/JSON), never code.

Phase G (Graph) — build the Information Dependency Graph from candidate fault nodes. Distinguish propagated symptoms from root candidates using information flow, not temporal order. Rank root candidates by descendant count.

Phase I (Intervene) — falsify hypotheses via counterfactual replay. Design minimal interventions for each root candidate; run in sandbox; root cause is the one whose intervention flips failure→success. Hard cap 5 interventions.

Phase I (Isolate) — narrow the fault surface to ≤8 candidate nodes via delta-debugging-style reduction. Preserves failure throughout reduction; records ruled-out evidence.

Phase L (Learn) — emit verified root-cause finding with evidence chain, walk the cause back to its originating decision, persist failure signature to semantic memory for future de-duplication, route handoff to the correct downstream Eidolon.

Phase V (Verify) — establish a reproducible failure. Deterministic-first; on two failed deterministic attempts, switch to statistical replay with confidence bands. No attribution begins without this phase passing.