Autonomous Multi-Phase Codebase Exploration
When to use
- Need to understand how a feature/system works across the codebase
- Starting work on unfamiliar codebase or component
- Planning changes that span multiple layers/components
- User asks "how does X work?" for non-trivial X
- Need architecture understanding before implementation
Skip when
- Pure reference lookup (function signature, type definition)
- Checking if specific file exists (yes/no question)
- Reading error message from known file location
Sequence
Runs before: ring:writing-plans
Related
Similar: dispatching-parallel-agents
Multi-phase approach: Phase 0 scopes the target, Phase 1 discovers the natural structure of the codebase, Phase 2 deep-dives into each discovered area in parallel, Phase 3 collects results, and Phase 4 synthesizes findings.
Announce at start: "Using ring:exploring-codebases for multi-phase autonomous exploration."
How It Works
Phase 1: Discovery (3-4 parallel agents) → Architecture, Components, Layers, Organization Phase 2: Deep Dive (N adaptive agents, one per discovered perspective) → Target implementation in each area Phase 3: Synthesis → Actionable guidance with file:line evidence
Phase 0: Scope Definition
Extract from user request: core subject, context/intent, depth needed. Set exploration boundaries (include/exclude directories).
Phase 1: Discovery Pass
⛔ STOP-CHECK BEFORE DISPATCH
Before emitting any Task call, count the discovery agents you intend to launch in this turn.
- Count MUST equal the number of discovery perspectives you scoped (typically 3-4: Architecture, Components, Layers, Organization).
- If your dispatch count diverges from your scoped count → STOP and reconcile.
- No substitutions, no omissions.
⛔ MUST NOT trickle-dispatch
All discovery agents leave in the SAME TURN, before reading any agent output.
Forbidden sequences:
- Dispatch agent 1 → read result → dispatch agent 2
- Dispatch a subset → wait → dispatch the rest
- Dispatch follow-up agents conditioned on partial output
- Loop sequentially over the discovery angles
If you find yourself about to dispatch a discovery agent in a turn AFTER any agent has already returned a result → STOP. You violated parallel dispatch. Report the violation and mark the phase INCOMPLETE rather than completing the trickle.
Self-verify after dispatch
After the dispatch turn, verify all scoped Task calls were emitted in that single turn. If fewer went out than scoped, the phase did NOT execute correctly. Mark INCOMPLETE and surface the dispatch failure — do NOT silently continue with a partial pool.
Parallel dispatch — atomic batch
Emit all scoped Task calls (the count established in the STOP-CHECK above) in a SINGLE TURN, as one atomic batch.
If your runtime exposes a multi_tool_use.parallel wrapper, use it to dispatch the complete pool in one wrapped invocation. This is the canonical fan-out mechanism on OpenAI-style tool envelopes and on certain Anthropic SDK consumers — naming it explicitly activates parallel emission on runtimes where trickle-dispatch is the default behavior.
If your runtime emits parallel tool_use blocks natively (Claude Code with Claude models), multi_tool_use.parallel may not be needed — but naming it is harmless and serves as an enforcement anchor.
The STOP-CHECK, anti-trickle, and self-verify guards above remain binding regardless of which mechanism your runtime uses.
Dispatch 3-4 discovery agents in a SINGLE turn (parallel):
Architecture Discovery: Find pattern (Hexagonal, Layered, Microservices, Monolith, etc.). Evidence: top-level directory structure, layer separation, file paths. Output: pattern name + confidence + ASCII diagram.
Component Discovery: Identify all major components/modules. For each: name, location, responsibility, tech stack, size. Map dependencies between components.
Layer Discovery: Within each component, identify layers (HTTP/API, Business Logic, Data Access, Infrastructure). Document how layers are separated and how they communicate.
Organization Discovery: Find organizing principle (by layer vs by feature vs by domain). Document file naming conventions, test organization, config locations.
After Phase 1: Validate quality — all areas have file:line evidence, no major "unknowns" remain. Determine how many deep-dive agents to launch (one per discovered perspective).
Phase 2: Deep Dive Pass
⛔ STOP-CHECK BEFORE DISPATCH
Before emitting any Task call, count the deep-dive agents you intend to launch in this turn.
- Count MUST equal the number of perspectives discovered in Phase 1.
- If your dispatch count diverges from your scoped count → STOP and reconcile.
- One agent per discovered perspective. No substitutions, no omissions.
⛔ MUST NOT trickle-dispatch
All deep-dive agents leave in the SAME TURN, before reading any agent output.
Forbidden sequences:
- Dispatch agent 1 → read result → dispatch agent 2
- Dispatch a subset → wait → dispatch the rest
- Dispatch follow-up agents conditioned on partial output
- Loop sequentially over the perspective list
If you find yourself about to dispatch a deep-dive agent in a turn AFTER any agent has already returned a result → STOP. You violated parallel dispatch. Report the violation and mark the phase INCOMPLETE rather than completing the trickle.
Self-verify after dispatch
After the dispatch turn, verify all scoped Task calls were emitted in that single turn. If fewer went out than scoped, the phase did NOT execute correctly. Mark INCOMPLETE and surface the dispatch failure — do NOT silently continue with a partial pool.
Parallel dispatch — atomic batch
Emit all scoped Task calls (the count established in the STOP-CHECK above) in a SINGLE TURN, as one atomic batch.
If your runtime exposes a multi_tool_use.parallel wrapper, use it to dispatch the complete pool in one wrapped invocation. This is the canonical fan-out mechanism on OpenAI-style tool envelopes and on certain Anthropic SDK consumers — naming it explicitly activates parallel emission on runtimes where trickle-dispatch is the default behavior.
If your runtime emits parallel tool_use blocks natively (Claude Code with Claude models), multi_tool_use.parallel may not be needed — but naming it is harmless and serves as an enforcement anchor.
The STOP-CHECK, anti-trickle, and self-verify guards above remain binding regardless of which mechanism your runtime uses.
Dispatch N adaptive agents in a SINGLE turn (parallel). One agent per discovered perspective.
Each agent prompt includes:
- Discovered context (architecture, component responsibility, location)
- Specific target to find
- Scope boundaries (directory, depth)
Each agent traces the target through: entry points → execution flow (with file:line) → data transformations → integration points.
Phase 3: Result Collection
Organize results into Discovery (structure) and Deep Dive (target implementation per area). Cross-reference: do deep dive findings align with Phase 1 structure? Note gaps explicitly.
Phase 4: Synthesis
Integrate findings into a unified document:
markdown# Codebase Exploration: [Target] ## Executive Summary [Architecture + how target works — 2-3 sentences] ## Phase 1: Discovery Findings ### Architecture Pattern | Component Structure | Layer Organization | Tech Stack | Diagram ## Phase 2: Deep Dive Findings ### [Discovered Area 1] - Entry point: file:line - Flow: step-by-step with file:line - Patterns: [observed] - Integration: [connections] ### [Discovered Area 2] [same structure] ## Cross-Cutting Insights ### Pattern Consistency | Variations | Integration Points | Data Flow | Key Decisions ## Implementation Guidance ### Adding new functionality: where, patterns to follow, integration requirements ### Modifying existing: files to change, ripple effects ### Debugging: starting points, data inspection points, common failures
Phase 5: Action Recommendations
Context-aware next steps based on user's goal:
- Implementation → suggest ring:writing-plans with discovered entry points
- Debugging → investigation starting points with file:line
- Learning → reading path through key files
Discovery Agent Templates
Architecture Agent
Goal: Discover architecture pattern. Examine top-level structure. Identify: Hexagonal/Layered/Microservices/Monolith/MVC/Other. Provide: pattern name, directory evidence, key paths, confidence level, ASCII diagram.
Component Agent
Goal: Identify all major components/modules. For each component: name, path, responsibility, stack, size. Map dependencies. Identify shared libraries.
Layer Agent
Goal: Discover layers within components. For each layer: name, location, responsibility, dependencies. Check for layer violations. Document communication patterns (DI, interfaces, direct).
Organization Agent
Goal: Find organizing principle. Identify: by-layer vs by-feature vs by-domain. Document: file naming conventions, test co-location, config locations.
Deep Dive Agent Template
Goal: Explore [TARGET] in [DISCOVERED_PERSPECTIVE]. Context: architecture=[X], component=[Y], location=[Z], responsibility=[R]. Task: 1. Find [TARGET] — entry points with file:line 2. Trace execution flow — each step with file:line 3. Document patterns — error handling, validation, testing 4. Identify integration points — inbound/outbound with file:line Scope: Stay within [directory]. Maximum depth: [based on layer]. Output: entry points, execution flow, patterns, integration points, key files.
Verification Checklist
Phase 1 complete:
- Architecture pattern identified with evidence
- All major components enumerated
- Layers/boundaries documented
- File:line references for structural elements
Phase 2 complete:
- All discovered perspectives explored
- Target found and documented per area
- Execution flows traced with file:line
- Integration points identified
Synthesis quality:
- Discovery and deep dive integrated
- Cross-cutting insights identified
- Implementation guidance is specific and actionable
