Using codebase analyzer

<SUBAGENT-STOP> If you were dispatched as a subagent to execute a specific task, skip this skill. </SUBAGENT-STOP> <EXTREMELY-IMPORTANT> If you think there is even a 1% chance an analysis skill might apply to what you are doing, you ABSOLUTELY MUST invoke it. This is not negotiable. </EXTREMELY-IMPORTANT>

classify-analysis-target is ALWAYS required. After classification, express lane lets you skip directly to any Track B skill (but Track A context may be missing).

digraph analysis_flow {
    rankdir=TB;
    node [fontname="monospace"];

    "User starts analysis" [shape=doublecircle];
    "classify-analysis-target" [shape=box, style=bold, label="classify-analysis-target\n(ALWAYS FIRST)"];
    "Express lane?" [shape=diamond];
    "Track A: Recon skills" [shape=box];
    "Check trigger signals\n+ security signals" [shape=diamond];
    "HIGH priority" [shape=box, style=filled, fillcolor="#ffcccc"];
    "MEDIUM priority" [shape=box];
    "LOW priority" [shape=box];
    "Track B Phase 1:\nProvenance + Build Pipeline" [shape=box];
    "Track B Phase 2:\nArtifacts + Data Flows + Agent Loop" [shape=box];
    "Track B Phase 3:\nTool Graph + Gates + Simulate + Prompt" [shape=box];
    "Phase 4:\nreconstruct-system-intent" [shape=box, style=bold];

    "User starts analysis" -> "classify-analysis-target";
    "classify-analysis-target" -> "Express lane?";
    "Express lane?" -> "Track B Phase 1:\nProvenance + Build Pipeline" [label=" yes\n(skip Track A)"];
    "Express lane?" -> "Track A: Recon skills" [label=" no"];
    "Track A: Recon skills" -> "Check trigger signals\n+ security signals";
    "Check trigger signals\n+ security signals" -> "HIGH priority" [label=" pause,\noffer deep dive"];
    "Check trigger signals\n+ security signals" -> "MEDIUM priority" [label=" continue,\npresent after"];
    "Check trigger signals\n+ security signals" -> "LOW priority" [label=" note\nfor summary"];
    "HIGH priority" -> "Track B Phase 1:\nProvenance + Build Pipeline";
    "MEDIUM priority" -> "Track B Phase 1:\nProvenance + Build Pipeline";
    "LOW priority" -> "Track B Phase 1:\nProvenance + Build Pipeline";
    "Track B Phase 1:\nProvenance + Build Pipeline" -> "Track B Phase 2:\nArtifacts + Data Flows + Agent Loop";
    "Track B Phase 2:\nArtifacts + Data Flows + Agent Loop" -> "Track B Phase 3:\nTool Graph + Gates + Simulate + Prompt";
    "Track B Phase 3:\nTool Graph + Gates + Simulate + Prompt" -> "Phase 4:\nreconstruct-system-intent";
}

Track A skills: identifying-tech-stack, mapping-architecture, tracing-dependencies, detecting-dead-code, inventorying-api-surface, analyzing-code-quality

Track B phases: Phase 1 (trace-codebase-provenance, analyze-build-pipeline) → Phase 2 (classify-repo-artifacts, trace-data-flows, analyze-agent-loop) → Phase 3 (extract-tool-graph, map-feature-gates, simulate-behavior, analyze-prompt-influence) → Phase 4 (reconstruct-system-intent)

Special: test-hypothesis, detect-hidden-contracts

Orchestration Skills (Development Workflow)

Beyond analysis, this plugin provides development orchestration:

Trigger	Skill	What It Does
"build", "create", "implement", "add feature"	brainstorming	Design-first before implementation
"plan this", "write a plan"	writing-plans	Spec -> bite-sized tasks
"implement plan"	subagent-driven-development	Fresh subagent per task, review between
"execute plan"	executing-plans	Batch execution with checkpoints
"debug", "error", "not working"	systematic-debugging	Root cause investigation
"write tests", "test this"	test-driven-development	RED-GREEN-REFACTOR
"review code", "check my work"	requesting-code-review	Dispatch reviewer subagent
"merge", "finish branch"	finishing-a-development-branch	Merge/PR workflow
"parallel", "concurrent"	dispatching-parallel-agents	Concurrent subagents
"verify", "done"	verification-before-completion	Verify commands pass

Agent Dispatch Protocol:

Skill	Dispatches Agent	When
extract-tool-graph	code-explorer	Tool graph spans 5+ files
simulate-behavior	behavior-simulator	Multiple scenarios to compare
trace-codebase-provenance	code-explorer	Chain-of-custody tracing
test-hypothesis	Either	Targeted investigation

Dispatch only when task exceeds native tool capability (5+ file reads across subsystems).

Platform Capabilities:

Skills reference Claude Code tools and agent dispatch. On other platforms:

Capability	Claude Code	OpenCode	Codex
Full Track A + Track B	Yes	Yes (degraded: no agents)	Yes (degraded: no agents)
Agent dispatch	Yes	No	No
docs/analysis/ output	Yes	Yes	May fall back to inline

When agent dispatch is unavailable: warn user, execute simplified analysis (max 3 trace levels), mark output as Status: partial with degradation note.

See PLATFORM-NOTES.md for tool substitution table and per-platform details.

.state rules: classify-analysis-target creates docs/analysis/.state. Every skill appends its status. Check before Track B (warn-but-continue).

Red Flags — STOP and check yourself:

Thought	Reality
"I can just read the codebase directly"	Without classification you'll miss target-specific patterns
"Track A is overkill for this"	Skipping reconnaissance means missing trigger signals
"I'll skip to Track B directly"	Phase 1 errors cascade into every downstream analysis
"This is just a standard web app"	Similar apps differ. Classify first, assume nothing.
"I'll just run all skills"	Target type determines applicable phases.