AI writes code faster than you can review it. See where you're needed and where AI can handle the rest.

Install · Quickstart · What's inside · Docs

NVIDIA/OpenShell, 1,606 files: red is Harden now, amber Protect, lilac Stabilize, grey Routine. The panel shows what to do before the next merge.

AI writes code faster than people can review it. HighStakes scores every file in a repo on two questions, Will it break? and How bad if it breaks?, and turns the answer into concrete steps: who reviews, which tests come first, and where AI agents may merge on their own.

Install

go install github.com/zanetworker/highstakes/cmd/highstakes@latest
export OPENROUTER_API_KEY="sk-or-..."   # optional: without it, run analyze --no-llm

Needs Go 1.25+ and git.

Quickstart

cd /path/to/repo
highstakes init && highstakes analyze
highstakes get internal/llm/sensitive.go

internal/llm/sensitive.go
PROTECT
Will it break?         16  LOW   repo cutoff 37
How bad if it breaks?  85  HIGH  repo cutoff 61

Do before the next merge
  [ ] Get an owner's review (@zanetworker)
      why: Protect: costly if it breaks
  [ ] Block auto-merge
      why: Protect files never auto-merge
Put on the backlog
  [ ] Add a CODEOWNERS line
      why: not covered by CODEOWNERS
      /internal/llm/sensitive.go @zanetworker

Why
  Break signals: recent fix: "fix: never send credential files to the LLM; fourth review round" (2026-11-01)
  Break cost: security 85 · data 75; "Missed secrets may be sent to the LLM; false positives omit files."

Then highstakes dashboard opens the inbox above, and highstakes starts the terminal UI.

Why HighStakes

• Review where it matters. Each file gets a posture from its two scores, each shown against the repo's own cutoff: Harden now, Protect, Stabilize or Routine.
• Every PR gets a to-do list. pr check and the GitHub Action post a checklist, one-click issues and a merge gate that unblocks once the items are done.
• Know where AI agents may act alone. Each path gets an autonomy level (auto-merge, auto-fix, propose only, assist only), exported as an agent policy and a task queue.
• Learn a codebase in the right order. A reading tour, generated architecture diagrams and runtime units start with the code that matters most.
• Check it against history. check calibration compares the scores with real fixes, reverts and incidents, and a pinned 7-repo benchmark keeps scoring changes honest.

Use plain CODEOWNERS if you already know your critical paths and only need routing. Use a hotspot tool such as CodeScene if you want years of behavioral history and team dashboards rather than a per-PR checklist.

What it doesn't do: it does not review code or find bugs, and it does not run your tests. Scores rank files within one repo, not across repos. How bad if it breaks? is AI-assessed and not yet checked against maintainer ratings.

What's inside

Area	What you get	Docs
Scoring	analyze: likelihood from churn, complexity, centrality and incidents; impact from an LLM blast-radius read (security, data, availability, user); postures with repo cutoffs; --no-llm static mode; results cached by file content	How it works
Look up files	get <file> (verdict, checklists, reasons), list (filter and sort by tier, impact, likelihood), report	CLI reference
See the repo	dashboard (action inbox, treemap, explorer, architecture, quadrant), terminal TUI	Dashboard and TUI
Understand the code	get tour (reading path), get architecture (SeeCode diagram or JSON), list units (binaries and what they link), .heatmap/architecture.yaml for runtime links	Usage guide
Pull requests	pr check (verdict, before-merge checklist, worth doing, backlog, reviewers, change risk), --format markdown PR comment with one-click issues and a quadrant chart, --fail-on, --strict, --shadow	CLI reference
CI	github install (triage workflow), pr signoff, pr state, pr fresh (sign-off label, live state, stale-run checks), GitLab and other CI	CI integration
Ownership and policy	create codeowners (@handles only), create policy (no-auto-approve paths for AI reviewers), owner aliases	Configuration
AI agents	autonomy levels L3 to L0, create policy --agents, list tasks (backlog as a work queue), agent-context (commands as JSON)	Agent autonomy
Is it right?	check calibration (holdout, per-signal metrics, security fixes, maintainer labels), create labels, multi-repo benchmark	How it works, bench/
Incidents	incident create, incident list: recorded incidents raise likelihood	CLI reference
Configuration	.heatmap/config.yaml: excludes, tier thresholds, autonomy.max_level, owners, tour settings	Configuration

How it works

flowchart LR
  src[Source files] --> static[Static analysis<br/>complexity, imports]
  git[Git history] --> hist[Churn, owners,<br/>bug fixes]
  src --> llm[LLM blast radius<br/>security, data, availability, user]
  static --> like[Will it break?]
  hist --> like
  llm --> imp[How bad if it breaks?]
  like --> post[Posture and actions<br/>per file]
  imp --> post
  post --> out1[get, list, tour,<br/>dashboard, TUI]
  post --> out2[pr check and<br/>PR comment, gate]
  post --> out3[CODEOWNERS, AI reviewer<br/>and agent policies]

Loading

Static and git signals give Will it break?; an LLM reads each file for How bad if it breaks?. The two scores place the file in a quadrant, and rules turn the quadrant into actions. Details: How it works.

Models. One OpenRouter key, any model; the default, openai/gpt-6-luna, costs about $0.19 per 500 files, and re-runs only assess changed files. Model comparison: Configuration.

What leaves your machine. Each assessed file's path and first 300 lines go to the model endpoint. Files that look like credentials are never sent, and --no-llm sends nothing. See what is sent.

Contributing

Issues and pull requests are welcome. Build and test with go build ./... && go test ./... -count=1; the project layout is in CLAUDE.md.

License

MIT