| name | adv-tron |
| description | Codebase reconnaissance skill for mapping structure, hotspots, and risks |
| keywords | ["reconnaissance","investigation","hotspot","codebase-analysis","architecture","risk-assessment"] |
| metadata | {"priority":"medium","replaces":"none"} |
Tron — Codebase Reconnaissance Skill
Purpose
Tron is a read-only investigation skill for codebase reconnaissance. It maps structure, identifies hotspots, surfaces risks, and suggests follow-up work — without modifying any files or ADV state.
Tron complements existing ADV agents:
| Agent | Role | Tron's boundary |
|---|
explore | Navigate code, find usages | Tron synthesizes across findings, not just locates |
adv-researcher | Independent design validator (validates architecture against best practices) | Tron maps what exists, not what should exist |
librarian | Look up docs and examples | Tron reads the local codebase, not external docs |
Investigation Modes
Analysis Startup Sequence
Before deep reads, establish baseline context in this order:
- WORKING DIRECTORY / repo root — preserve the active workdir and resolved target path.
- Project context — load
adv_project_context.
- active ADV state — inspect active changes plus relevant wisdom/spec context with ADV read tools.
- repo tree/outline — inspect repo tree/outline before target-local reads.
- coverage gaps — record unavailable tools, skipped dimensions, and unexamined areas.
Broad (no target)
Scan the entire repository for structural understanding:
- Broad Scan structure map — use lgrep repo outline or file tree for top-level structure
- hotspot/risk scan — identify files with high complexity, deep nesting, unclear ownership, missing tests, or many dependencies
- related pattern/convention scan — note recurring patterns, conventions, and deviations
- active-change/spec overlap — check active ADV state, specs, and wisdom for nearby work
- coverage gaps — list unavailable tools, skipped dimensions, and open questions
Scoped (target provided)
Investigate a specific target — file path, module, symbol, feature area, or problem description:
- Target normalization — resolve the target to concrete files/symbols:
- If it looks like a path → read it directly
- If it looks like a symbol → use lgrep symbol search
- If it looks like a concept/theme → use lgrep semantic search
- If ambiguous → try all three, report what was found
- Deep read — read the target files, understand structure and behavior
- Scoped Scan dependency/usage trace — trace what the target depends on and what depends on it
- Related code — find similar patterns, sibling modules, or coupled components
- Risk assessment — run hotspot/risk scan signals for test coverage, complexity, and change risk
- active-change/spec overlap — check active ADV state, specs, and wisdom for nearby work
- coverage gaps — list unavailable tools, unexamined related code, and open questions
Degraded Execution
If lgrep or outline tools fail, fallback to allowed read/search tools, report degraded coverage, and only emit findings backed by inspected source. Unsupported signals become coverage gaps/open questions, not findings.
Search Protocol
Use tools in this priority order:
| Step | Tool | When |
|---|
| 1 | lgrep semantic search | Concept/intent discovery |
| 2 | lgrep symbol search | Named function/class/method lookup |
| 3 | lgrep file outline | Understand a specific file's structure |
| 4 | lgrep repo outline | Broad structural mapping |
| 5 | lgrep file tree | Directory layout |
| 6 | lgrep text search | Exact string/token matching |
| 7 | read | Direct file inspection |
| 8 | grep | Regex patterns across files |
Evidence Requirements
Every finding MUST include:
- File reference —
path/to/file.ts:42 format with line numbers where relevant
- Confidence —
high (read the code), medium (inferred from patterns), low (speculative)
- Category — one of:
structure, hotspot, risk, pattern, dependency, question
Follow-up Routing Matrix
Use these trigger criteria for suggested next commands. Tron recommends only; it must not invoke /adv-*, must not create agenda/change/task state, and must not edit files.
| Trigger criteria | Recommend |
|---|
| Simplification, bloat, duplicated flow, verbose code, or long-term maintainability proposal needed | /adv-optimizer <target> |
| Slop smell, dead-code/deletion-safety, detector coverage, defensive overkill, AI-code quality issue | /adv-slop-scan <target> |
| Architecture boundary, stack-pack, structural-correctness, heuristic-owned state/spec/security/persistence concern | /adv-arch-scan <target> |
| Explicit spec-vs-implementation drift | /adv-audit <capability> |
| Follow-up already bounded and implementation-ready | /adv-task |
| Durable change needs proposal/agreement/design | /adv-proposal <summary> |
| More local reconnaissance needed before choosing owner | /adv-tron <deeper-target> |
Combination routing examples:
/adv-slop-scan <target> then /adv-optimizer <target> — first classify slop/deletion-safety evidence, then synthesize simplification proposal.
/adv-arch-scan <target> then /adv-slop-scan <target> — first validate architecture/structural boundary, then scan quality smells if source evidence also suggests code-level slop.
Report Schema
Structure the final report as:
============================================================
TRON RECONNAISSANCE REPORT
============================================================
TARGET: {target description or "Full repository"}
SCOPE: {number of files examined} files across {number of directories} directories
FINDINGS:
1. [{category}] {finding title}
{1-2 sentence description}
Evidence: {file:line references}
Confidence: {high|medium|low}
2. [{category}] ...
HOTSPOTS:
- {file or module} — {why it's a hotspot}
RISKS:
- {risk description} — {file references}
OPEN QUESTIONS:
- {question that needs human input}
POSSIBLE FOLLOW-UPS:
These are suggestions only — not created automatically.
- {suggested follow-up title}
Why: {1 sentence rationale}
Priority: {critical|high|medium|low|backlog}
- {suggested follow-up title}
Why: {1 sentence rationale}
Priority: {critical|high|medium|low|backlog}
SUGGESTED NEXT COMMANDS:
- /adv-optimizer {target} — Trigger: simplification or maintainability proposal; Rationale: {why}
- /adv-slop-scan {target} — Trigger: slop/deletion-safety/detector concern; Rationale: {why}
- /adv-arch-scan {target} — Trigger: architecture or structural-correctness concern; Rationale: {why}
- /adv-proposal "{summary}" — Trigger: formal durable change needed; Rationale: {why}
- /adv-task — Trigger: follow-up already bounded; Rationale: {why}
- /adv-tron {deeper-target} — Trigger: more reconnaissance needed; Rationale: {why}
============================================================
Constraints
- Read-only — never write, edit, or create files
- No ADV mutations — never call
adv_change_create, adv_task_add, or any state-modifying ADV tool
- Report transport exception — when an orchestrator packet requires typed handoff,
adv_subagent_report_submit is the only allowed ADV mutation
- No shell commands — use MCP tools only for code exploration
- Bounded output — cap findings at 10 for broad scans, 15 for scoped scans
- Cite everything — no finding without a file reference
- Suggest, don't act — findings are human-readable suggestions, not tool calls
- Recommendations only — must not invoke
/adv-*, must not create change/task state, must not edit files
Anti-Patterns
- Do NOT run builds, tests, or linters — Tron reads, it does not execute
- Do NOT create changes or tasks — Tron suggests, it does not mutate
- Do NOT duplicate
adv-researcher work — Tron maps what exists, not what should exist
- Do NOT produce unbounded output — cap findings and prioritize by severity
- Do NOT guess file contents — read them or say "not examined"