insider — Claude Code plugin

Deep-dive industry and company analysis using parallel specialized agents with checkpointing. VC / equity-research-grade breakdowns with no fixed time limits.

Architecture

Research layer (markdown)         Data layer (jsonl)         Reading layer (HTML)
└─ machines/agents                └─ queries/analytics       └─ humans
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

/industry <industry>
  ↓
Phase 1: macro              (blocking)
  ↓
Phase 2: economics          (parallel)  ─┐
         competition        (parallel)  ─┘
  ↓
Phase 3: synthesis          (blocking)
  ↓
Phase 3.5: Committee Vote         (3 agents in parallel)
  ↓
Phase 3.7: extractor  (blocking)   ← reads raw md → writes data layer jsonl
  ↓
Phase 4: assembler            (blocking)   ← single HTML file, inline SVG, EN/中文 toggle
  ↓
Phase 4.5: verify-numerics + verifier  (blocking, two-stage)
  ↓
research/ + data/ + reading/

No fixed time limits. Agents cannot reliably perceive wall-clock time or tool-call counts. Stopping is driven by marginal insight per source — when the next 3 fetches return nothing new, the agent writes its output and stops.

Checkpointing. Every phase writes a checkpoint. Crash at phase 3? Resume from phase 3 on next invocation.

Parallel agents. Economics and competitive analysis run simultaneously after macro foundation completes. Committee members (investor / expert / skeptic) also vote in parallel.

Hooks auto-loaded. The plugin ships hooks/hooks.json with a PostToolUse hook on WebSearch | WebFetch | mcp__gemini-search__web_search that appends every search to .checkpoint/search-backup.jsonl in the user's CWD. When the plugin is enabled via /plugin install, the hook loads automatically — no setup script required. Permissions are handled by Claude Code's runtime prompts; granted tools land in .claude/settings.local.json automatically.

Commands

Command	Use when
/industry <industry>	Decode a sector
/company <name-or-ticker>	Diligence a specific company
/consume <slug>	Regenerate the reading from existing research
/lens <slug> <stakeholder>	Stakeholder-filtered prep doc
/audit <slug>	Freshness check using data layer

Examples

/industry semiconductor capital equipment
/industry B2B SaaS
/company ASML
/company Snowflake
/lens semiconductor-capital-equipment c-suite
/audit ai-coding-assistants

Pipeline

Industry research (7 phases)

Phase	Agent	Output	Parallel?
1	macro	macro.md	No
2	economics	economics.md	Yes
2	competition	players.md	Yes
3	synthesis	thesis.md, scenarios.md, analogs.md, gaps.md, open-secrets.md, sources.md	No
3.5	investor + expert + skeptic	committee vote (no file)	Yes
3.7	extractor	claims.jsonl, sources.jsonl, entities.json, metrics.jsonl (append/merge)	No
4	assembler	reading//index.html	No
4.5	verify-numerics.sh + verifier	phase-4.5-logic-review.json	No

Company research (6 phases)

Phase	Agent	Output	Parallel?
1	filings	narrative.md	No
2	financials	financials.md	Yes
2	competition	competitive.md	Yes
3	synthesis	thesis.md, scenarios.md, gaps.md, sources.md	No
3.7	extractor	data layer jsonl	No
4	assembler	reading//index.html	No
4.5	verify-numerics.sh + verifier	phase-4.5-logic-review.json	No

Dependencies

Required

• Claude Code (the plugin is a Claude Code plugin)

Optional but recommended

• Gemini CLI (npm install -g @anthropic-ai/gemini-cli) — underlying dependency for the gemini search MCP server (mcp__gemini-search__web_search). Not bundled (avoids MCP namespace collision). After installing, manually edit .insider/search-priority.json to include it in your search tool order.

Why Gemini CLI is not bundled: The plugin intentionally does not ship a .claude/mcp.json to avoid namespace collision with a Gemini MCP server the user may already have configured globally. See references/known-open-problems.md § Open problem 4 for the full collision analysis.

Permissions

.claude-plugin/required-permissions.json documents the tools sub-agents need: Read, Write, Edit, WebFetch, WebSearch, mcp__gemini-search__web_search, Bash, Agent.

The plugin does not auto-merge these into .claude/settings.json. Instead, Claude Code prompts the user the first time each tool is invoked, and granted tools accumulate in .claude/settings.local.json. To restrict a specific tool, add it to the deny array in settings.local.json — sub-agents will fail loudly and the orchestrator will report which tool needs to be re-allowed.

Why this matters: Sub-agents inherit the parent's permission context. When a permission is missing, Claude Code's prompt is the natural place to grant it; baking permissions into a JSON file pre-emptively just creates drift.

Setup

Installed as a Claude Code plugin (recommended) — nothing to do. hooks/hooks.json auto-loads when the plugin is enabled. The search-backup hook creates .checkpoint/ in your project on first fire.

Standalone clone (no plugin install) — run the fallback script from your project root to copy the same hook into your project's .claude/settings.json:

/path/to/insider/tools/setup.sh

Optional flags: --check-only (verify only) and --project-root <dir> (target a directory other than $(pwd)).

Recovery

Failure	Recovery
Agent crashes	Recovery-agent reconstructs from backup, marks [RECOVERED]; if no backup, mark partial
Agent returns thin output	Review gate detects; spawn Round 2 with gap list
Orchestrator crashes	Resume from last checkpoint on next invocation
Web search API error	Wait 5s, retry with reformulated query
All parallel agents fail	Report to user; offer abort or continue with partial
Sub-agent permission denied	Claude Code prompts the user; if denied, the orchestrator reports which tool needs re-allow

Output structure

research/                          # Research layer (markdown) — for machines/agents
├── industries/<slug>/
│   ├── thesis.md
│   ├── macro.md
│   ├── economics.md
│   ├── players.md
│   ├── scenarios.md
│   ├── analogs.md
│   ├── gaps.md
│   ├── open-secrets.md
│   ├── sources.md
│   └── meta.json
└── companies/<slug>/
    ├── thesis.md
    ├── financials.md
    ├── narrative.md
    ├── competitive.md
    ├── scenarios.md
    ├── gaps.md
    ├── sources.md
    └── meta.json

data/                              # Data layer (jsonl) — for queries/analytics, cross-project append/merge
├── claims.jsonl
├── entities.json
├── sources.jsonl
└── metrics.jsonl

reading/                           # Reading layer (HTML) — for humans, single file per slug
└── <slug>/
    └── index.html

.checkpoint/                       # Resume state
├── industries/<slug>/
│   ├── phase-1-macro.json
│   ├── phase-2-deep-dive.json
│   ├── phase-3-synthesis.json
│   ├── phase-3-committee.json
│   ├── phase-3.7-data.json
│   └── phase-4-complete.json
└── companies/<slug>/
    └── ...

Trust-signal discipline

Every claim carries source class ([reported], [community], [inference]) and confidence ([high], [medium-high], [medium], [low]). Inference capped at 20% of items per file. Top 3 claims require direct quotes. Self-check pass verifies citations and deletes stale items.

Known limitations

The verifier stack catches a defined set of error classes (fabrication, broken arithmetic, format mismatches, coverage gaps, declared-constituent overlaps). It structurally cannot catch three classes that require domain modeling or expert review — see references/known-open-problems.md:

1. Conceptual-model fit — chart primitives (flow / Sankey) applied to underlying structures that are actually hierarchies / overlaps
2. Constituent plausibility — declared values that are syntactically correct but factually wrong
3. Novel-class semantic errors — error patterns we haven't seen yet

Mitigations are layered (anti-amorphous-label rule, transition constituent decomposition, skeptic structural-overlap check), but residual risk is real and acknowledged. Domain-expert review of high-stakes claims remains necessary.

Query tools

./tools/query.sh claims --project <slug> --confidence high
./tools/query.sh sources --status dead
./tools/query.sh entities --type company
./tools/query.sh metrics --entity <name>
./tools/query.sh contradicts --entity <name> --metric <metric>
./tools/query.sh stale --project <slug>
./tools/query.sh stats
./tools/query.sh webfetch-failures --stats

Cleanup

./tools/clean.sh           # dry-run: list cruft candidates, no deletion
./tools/clean.sh --apply   # actually delete

Removes anything outside the per-directory allowlist:

Directory	Allowlist	Anything else flagged
reading/<slug>/	index.html, numerics.json	screenshots, drafts, .backup/, etc.
research/<slug>/	*.md, meta.json	tmp / draft files
data/	*.jsonl, *.json	tmp / backups
.checkpoint/	*.json, *.jsonl, README.md	tmp / backups

Plus always-cruft anywhere: *.tmp, *.bak, *.old, *~, .DS_Store, __pycache__/. Empty reading/<slug>/ (no index.html) is also flagged.

Consumption

• Primary: Open reading/<slug>/index.html in any browser (single file, EN/中文 toggle, zero deps)
• Deep dive: Open research/ folder as Obsidian vault
• Query: Use ./tools/query.sh for cross-research search

License

MIT License — see LICENSE. Author: Chuang.