blindspot 0.1.0

The bus factor for the AI era. Map team knowledge blindspots before someone leaves.

blindspot

Engineering resilience for AI-accelerated teams. Detect hidden organizational fragility — concentrated ownership, shallow reviews, knowledge decay, single-engineer dependencies — before they become delivery risks.

pip install blindspot
blindspot scan /path/to/repo

One self-contained HTML report. No server. No signup. No telemetry.

🚧 Pre-alpha. APIs and metrics will change.

---

What you get

One report, grouped for both audiences.

blindspot scan /repo produces a single HTML file organised under four headers — TL;DR (executive summary, resilience score with A–F grades, top actions, trend), then People & Ownership, Knowledge State, Process Quality for drill-down. A board reads the top; an engineering manager scrolls.

A specific question, answered.

blindspot scan /repo --simulate-departures "[e-mail],[e-mail],[e-mail]"

"If these three engineers leave together after a re-org, how many files become orphans, which services take the hit, and what should we do first?"

---

Why this exists

AI coding tools made engineering teams faster. But faster does not mean understood. Codebases now carry a new kind of risk: services that shipped quickly, owned by one person, reviewed by no one in depth — and that one person can leave tomorrow.

Existing tools measure velocity. blindspot measures resilience.

The product is built around three rules:

• Service-first, not person-first. Default views show service-level risk; person views require service context.
• Evidence over inference. Signals come from observable git/PR/filesystem state — not from guessing how someone codes.
• Reports, not surveillance. blindspot answers "is this service fragile?", not "is this person slacking?".

---

What it measures

Signal	What it tells you
Ownership concentration	Who actually understands each part of the codebase, weighted by recency and review depth
Bus factor per service / folder	How many people would need to leave before knowledge is critically lost
Departure simulation	What happens if a specific person — or a specific group — leaves: orphan files, affected services, coverage loss
Knowledge decay	Code volatility and contributor drift, projected 30 / 60 / 90 days forward
Review lineage	Where approvals land without substantive comments (rubber-stamp ratio), and where one reviewer carries everything (diversity HHI)
Correction load	Share of recent commits to each file that are follow-up fixes or reverts — observable evidence of stability debt, not a verdict on people
AI-native operational context	Per-service coverage of agent rules, specs, prompts, architecture decisions and skills. How much organizational memory exists for new humans and AI agents to load. Not an AI-generated-code detector
Engineering Resilience Score	Composite 0–100 number + letter grade, with band (Strong / Moderate / Fragile / Critical)

Each signal that's actionable generates a concrete recommendation in the Top actions table, tagged with the relevant fragility pattern (single-owner-concentration, review-without-scrutiny, fragile-velocity).

---

Quick start

# Install
pip install blindspot

# Full report (grouped: TL;DR + People · Knowledge · Process)
blindspot scan /path/to/repo --output report.html

# Multi-person departure scenario (combined card at the top)
blindspot scan /path/to/repo \
    --simulate-departures "[e-mail],[e-mail]"

# Single-person deep-dive
blindspot simulate --person [e-mail] /path/to/repo

A rule-based narrator ships in the package — the executive summary is deterministic, in-process, no network, no key required. Configure a cloud LLM key (Anthropic or OpenAI) for richer prose; the report itself shows you how.

---

Documentation

Full end-to-end documentation lives in docs/ — the algorithms (with formulas and parameters), the architecture, the CLI reference, configuration, and how to read every section of the report.

---

Roadmap

Phase	Surface	Status
1	CLI + static HTML report	Shipping
1	Multi-person departure scenarios	Shipped (0.0.5b)
2	Knowledge graph visualization	Planned
2	Hidden silo detection + Change fear index	Planned
3	GitHub Action + Checks API output	Planned
3	Self-hosted dashboard	Planned
3	Slack / Jira / incident integration	Planned

---

License

MIT. See LICENSE.