| name | mine |
| description | Orchestrates deep exploration of an external technology ecosystem — clones repos, deploys parallel subagents for source analysis, and creates a research document. Use when asked to deeply mine, evaluate an external technology, explore an ecosystem, leave no stone unturned, or compare an external approach against Tau. |
| disable-model-invocation | true |
Technology Mining
Deep-dive into an external technology ecosystem to extract architectural patterns, browser standards, performance techniques, and integration opportunities for Tau.
Workflow
Step 1: Identify Targets
Gather from the user message:
- Repos to clone: GitHub URLs,
owner/repo slugs, or blog post references
- Group tag: Which
repos.yaml group these belong to (or create a new group)
- Mining scope: What to extract — architecture, browser APIs, performance patterns, build systems, API design, POSIX compliance, etc.
- Integration target: Which Tau system the findings relate to (filesystem, runtime, AI, rendering, etc.)
If the user provides blog posts or documentation URLs, fetch them with WebFetch before cloning repos.
Step 2: Clone Repos
Use the repos skill to add and clone all identified repos:
pnpm repos add <owner/repo> -g <group> --clone
If repos are already tracked in repos.yaml, just clone:
pnpm repos clone <name>
Step 3: Deploy Parallel Subagents
Launch one subagent per repo or major aspect using the Task tool with subagent_type="generalPurpose". Each subagent receives a focused mining brief:
Prompt template per subagent:
Deeply explore <repos/name> to extract:
1. Architecture: module structure, dependency graph, key abstractions
2. Browser APIs: which Web APIs are used and how (OPFS, SharedArrayBuffer, WebAssembly, IndexedDB, etc.)
3. Performance patterns: caching, lazy loading, zero-copy, batching, pooling
4. Build system: toolchain, WASM compilation flags, bundle strategy
5. API design: public surface, DX patterns, error handling
6. Testing: test infrastructure, coverage approach
Key files to start with: README, package.json, src/index.*, core/*, lib/*
Return a structured report with numbered findings (F1, F2, ...) and
specific file:line references for each claim.
Adjust the mining dimensions based on the user's scope. Run subagents in parallel (multiple Task calls in one message).
Step 4: Fetch External Context
In parallel with subagent exploration, fetch any blog posts, documentation pages, or GitHub issues the user referenced:
Use WebFetch for each URL, extract:
- Architecture diagrams or descriptions
- Performance claims with numbers
- API examples
- Browser compatibility notes
Step 5: Synthesize Findings
After all subagents return:
- Merge findings across repos into a unified analysis, grouping by theme (architecture, performance, browser APIs, etc.)
- Cross-reference against Tau: For each finding, assess relevance to Tau's architecture — read the relevant Tau source files to compare approaches
- Classify each finding:
| Classification | Meaning |
|---|
| Adopt | Integrate the technology directly into Tau |
| Adapt | Apply the pattern/principle within Tau's existing architecture |
| Reference | Useful knowledge but no immediate action needed |
| Skip | Not applicable to Tau's constraints |
- Fit assessment: Explicitly evaluate whether direct integration is viable. Consider: bundle size impact, architectural alignment, multi-tab/worker compatibility, maintenance burden, and licensing.
Step 6: Create Research Document
Use the create-research skill conventions to produce docs/research/<topic>.md:
- Category:
reference (for ecosystem documentation) or comparison (for fit assessments)
- Include numbered findings (F1-FN) with file:line evidence from the mined repos
- Include a Tau Alignment Analysis section comparing each finding against Tau's current approach
- Include a Recommendations table (R1-RN) with Priority/Effort/Impact columns
- Run
pnpm docs:validate to verify frontmatter
Composing Skills
| Skill | Role in Mining |
|---|
repos | Clone and manage external repos |
create-research | Document findings in standard format |
Anti-Patterns
- Shallow mining: Reading only README files. Subagents must explore source code, not just documentation.
- Ungrounded claims: Every finding must cite a specific file and line from the mined repo.
- Missing fit assessment: Always evaluate adopt vs. adapt — never assume integration is the right answer.
- Skipping Tau comparison: Findings without cross-reference to Tau's current approach are not actionable.
Example Invocation
User: "deeply mine the entire Turso database ecosystem for best practices in databases and filesystems"
Agent actions:
- Fetch blog posts via WebFetch
pnpm repos add tursodatabase/turso -g dev-tools --clone
pnpm repos add tursodatabase/agentfs -g dev-tools --clone
- Deploy 3 subagents: Turso core (WASM, VFS, page cache), AgentFS (POSIX, SQLite schema, CoW overlay), database-wasm (browser runtime, SharedArrayBuffer)
- Synthesize into
docs/research/turso-fs.md