Skip to content

quickstart.md

Latest commit

 

History

History
107 lines (73 loc) · 4.26 KB

quickstart.md

File metadata and controls

107 lines (73 loc) · 4.26 KB
Raw

ASIL Quickstart — Wire It Into Any TypeScript Repo

This walks through wiring ASIL into an existing TypeScript codebase. End-to-end in five minutes if you already have a target repo.

Prerequisites

  • Node ≥ 20, pnpm ≥ 9
  • Target repo uses pnpm and has working pnpm install, pnpm test, pnpm typecheck scripts
  • Anthropic API key (primary LLM)
  • OpenAI API key (adversarial gate)

1. Add ASIL to your workspace

Clone or vendor the four packages alongside your code, then add them to your pnpm-workspace.yaml:

packages:
  - "packages/*"
  - "asil/packages/*"   # add this line

pnpm install will link them in. Or, once published, install the library packages from npm:

pnpm add asil-cost-controller asil-thought-multiplier asil-improvement-loop asil-analyzer

(asil-runners is the reference CLI/app, not a published library — vendor it or copy its wiring. The four libraries above are what you compose into your own runner.)

2. Set up environment

Create .env in your repo root:

ANTHROPIC_API_KEY=sk-ant-…
OPENAI_API_KEY=sk-…
REPO_ROOT=/absolute/path/to/your/repo

3. Drop in the runner skills

ASIL's thinkers expect Markdown skill files at <REPO_ROOT>/.asil/skills/. The simplest setup:

mkdir -p .asil/skills
# Copy or symlink any Markdown skills (e.g. from a public skill library) you want to use.
# A minimal install is fine — the thinkers fall back to inline defaults
# if a skill file is missing.

Override the location with ASIL_SKILLS_PATH if you want them somewhere else.

4. Run System B (Thought Multiplier) standalone

This is the lowest-risk way to try ASIL — no code mutations, just generates a build brief.

pnpm --filter asil-runners run:b "add rate limiting to the connect API endpoint"

You'll get a Markdown brief with merged recommendations, conflict resolutions, and acceptance criteria. Use it as a prompt to your editor agent of choice.

5. Run System A (Improvement Loop) in dry-run mode

pnpm --filter asil-runners auto grind --dry-run

This scans, queues tasks, and reports — without touching git or making any LLM calls past the scan. Use it to verify the scanner picks up what you expect.

6. Run a real grind (small budget)

pnpm --filter asil-runners auto grind --max-tasks 3 --skip dependency-update,documentation

The runner will:

  1. Scan and queue tasks
  2. Pause on any unresolved // DOMAIN_QUESTION: markers and offer triage
  3. For each task: isolate, execute, typecheck, test, self-review × 3, adversarial gate, domain guard, open PR
  4. Print a summary with cost spent and PRs opened

7. Inspect the spend

pnpm --filter asil-runners auto report
pnpm --filter asil-runners auto report weekly

Reports come out as Markdown tables with per-task, per-system, and per-day cost rollups.

Going further

  • Customize project rules. Pass projectName and projectDoNotChange via ThoughtMultiplierConfig to inject your house rules into every brief.
  • Plug in a different provider. Replace createAnthropicCaller / createCodexCaller in asil-runners/src/wiring.ts with a wrapper around any LLM that returns {content, inputTokens, outputTokens}.
  • Tighten the budget. Configure BudgetManager with per-category caps and a daily ceiling; the kill switch will halt the loop if either is breached.
  • Mark unresolved domain logic. Sprinkle // DOMAIN_QUESTION: <one-line question> comments in code paths that need human input. The loop will block on them until you triage.

Troubleshooting

Symptom Likely cause Fix
git worktree add fails Filesystem doesn't support worktrees (Google Drive, some FUSE mounts) The runner auto-falls-back to git clone. If both fail, check git status in REPO_ROOT.
pnpm install fails inside worktree Worktree lacks node_modules symlinks The loop runs pnpm install in each worktree by design. Ensure the worktree path is writable.
Adversarial gate keeps blocking Diff is genuinely contradictory or oversized Check the gate's reason string; usually the LLM has a real point. Lower maxAttempts to fail faster.
Cost cap hit immediately Budget too low for selected models Either bump dailyBudgetUSD or downgrade executionModel from opus to sonnet.