init-greptile-review（初始化爬虫审查）

Init Greptile Review

Generate repo-specific Greptile configuration that makes Greptile catch meaningful bugs, security issues, architectural violations, and contract drift without wasting semantic review budget on lint-like trivia.

What good Greptile setup looks like

Greptile is a semantic PR reviewer, not a linter.

• Use Greptile for architecture boundaries, security invariants, API contracts, performance pitfalls, dependency misuse, compliance, and cross-file reasoning.
• Leave semicolons, formatting, import sorting, and other linter/formatter work to ESLint, Prettier, Ruff, RuboCop, golangci-lint, and similar tools.
• Greptile reads repo-local config from the source branch of the PR.
• Changes apply on the next PR or re-review, not retroactively.
• Prefer .greptile/ over legacy greptile.json.

Decision tree

What are you doing?
│
├── First-time setup, permissions, or indexing ───────────────► references/setup.md
├── Authoring or tuning .greptile/config.json / files.json ──► references/config-spec.md
├── Designing repo-specific review rules ─────────────────────► references/rules-engineering.md
├── Matching a common stack or monorepo layout ───────────────► references/patterns.md
├── Needing full end-to-end examples ─────────────────────────► references/scenarios.md
├── Wiring status checks, API triggers, or CI/CD ─────────────► references/integration.md
└── Debugging ignored config, noisy reviews, or rollout checks ► references/troubleshooting.md + references/anti-patterns.md

Quick start

By default, generate only the files the repo actually needs:

.greptile/
├── config.json      # always
├── rules.md         # only when rules need examples, migration notes, or nuanced narrative
└── files.json       # only when docs/specs/schemas materially improve review

packages/payments/.greptile/config.json  # only when a subtree needs different strictness or extra rules

Start lean:

• Default to strictness: 2 and commentTypes: ["logic", "syntax"]
• Add "style" only when the repo lacks strong formatting/linting
• Start with 5-10 high-signal rules, not 20 generic ones
• If migrating from greptile.json, move to .greptile/ and delete the legacy file

Workflow

Follow these six phases in order.

Phase 1: Classify the request and current Greptile state

Read first: references/setup.md for platform and placement questions; references/config-spec.md for parameter formats.

Before writing anything, change into the target repo root. If you are working from a shared fixture or sample repository, copy it to a scratch repo/worktree before writing .greptile/ files so the source fixture stays unchanged.

Then determine what kind of job this is:

• New setup — no existing .greptile/ or greptile.json
• Tune existing config — root config exists and should be extended, not replaced
• Monorepo split — different packages/services need different strictness or rules
• Troubleshooting — existing config is ignored, noisy, or missing real issues
• Integration — status checks, API triggers, CI/CD, or merge gating

Search the repository for these items (e.g. find. -name '.greptile' -type d, find. -name 'greptile.json'):

• Existing .greptile/ folders at root and child directories
• Legacy greptile.json
• Docs, specs, and schemas that could feed files.json — look for: Prisma/Drizzle schemas, OpenAPI/Swagger specs, ADR documents, architecture docs (docs/, architecture.md), shared type definition files. Exclude READMEs and changelogs. Note these for Phase 2 refinement.
• Existing linters/formatters — check root config files (.eslintrc*, prettier.config*, biome.*, .stylelintrc*, pyproject.toml, .rubocop.yml) and package.json / pyproject.toml devDependencies for linter/formatter packages

Shell-safe detection recipe:

find . -type d -name '.greptile' -not -path '*/node_modules/*' 2>/dev/null
find . -type f -name 'greptile.json' -not -path '*/node_modules/*' 2>/dev/null

find . -maxdepth 2 -type f \( \
  -name '.eslintrc*' -o -name 'eslint.config.*' -o \
  -name 'prettier.config*' -o -name '.prettierrc*' -o \
  -name 'biome.json' -o -name 'biome.jsonc' -o \
  -name '.stylelintrc*' -o -name 'stylelint.config*' -o \
  -name '.rubocop.yml' -o -name '.rubocop_todo.yml' -o \
  -name '.golangci.yml' -o -name '.golangci.yaml' -o \
  -name 'pyproject.toml' -o -name 'setup.cfg' -o -name '.flake8' -o -name '.pylintrc' \
\) 2>/dev/null

python3 - <<'PY'
from pathlib import Path
import json

package_json = Path("package.json")
if not package_json.exists():
    print("package.json not found")
else:
    data = json.loads(package_json.read_text())
    dev = data.get("devDependencies", {})
    linters = [name for name in dev if any(tag in name.lower() for tag in ("eslint", "prettier", "biome", "stylelint", "lint", "oxlint"))]
    print("Linters/formatters:", ", ".join(linters) if linters else "none found in devDependencies")
PY

If none of these checks finds lint/format tooling, treat style coverage as absent only after confirming the repo is not delegating linting to another manifest or CI wrapper.

⚠ Steering: When searching for docs/specs/schemas, look for file extensions and paths, not file contents. Use shell-safe commands such as find. -type f \(-name '*.prisma' -o -name 'openapi*' -o -name '*.graphql' -o -name '*.proto' \) -not -path '*/node_modules/*' and find. -maxdepth 3 -type f \(-path './docs/*' -o -name 'architecture.md' -o -name 'ARCHITECTURE*' -o -name 'ADR*' \)! -name 'README*'! -name 'CHANGELOG*' | head -20. Exclude READMEs and changelogs — they rarely contain information a reviewer needs.

If config already exists, preserve working pieces and tighten only what the repo evidence supports.

Phase 2: Map the review surface

Read first: references/patterns.md for stack-specific patterns and ignore lists.

Map the repository by reading the workspace config (e.g. pnpm-workspace.yaml, lerna.json, Cargo.toml), listing top-level directories, reading root and app-level dependency manifests, and searching for risk-sensitive directories. Cover these seven areas:

1. Structure — monorepo vs single-service, top-level packages/apps/services, critical directories
2. Stack — languages, frameworks, ORM, transport layers, testing stack
3. Risk zones — auth, payments, PII, file system, shell execution, IPC, migrations, infra
4. Generated/noisy files — build output, generated code, lockfiles, snapshots, vendored code
5. Useful context files — refine the list from Phase 1; add any newly discovered architecture docs, ADRs, OpenAPI specs, schemas, shared types, or contribution guides
6. Existing quality controls — linters, formatters, CI, custom rules, security scanners
7. Recurring failure patterns — user-stated pain points, incident history, review gaps, repeated bugs. If no user input is available, infer from recent commit history (e.g. git log --oneline --grep='fix' --grep='bug' --grep='revert' | head -20) or skip this item.

⚠ Steering: "Map the repository" means reading config files and directory structure, not reading every source file. For item 7 (recurring failure patterns), if the user hasn't stated pain points, run git log --oneline --grep='fix' --grep='bug' --grep='revert' | head -20 to infer from history, or skip this item entirely.

Depth guidance: for a small repo (<50 files), a directory listing and dependency manifest may suffice. For a large monorepo, check each top-level package's purpose, key dependencies, and risk-sensitive directories. Stop when you can confidently answer all 7 items above.

Exploration checklist — confirm you can answer these before proceeding:

• What is the repo structure? (monorepo/single-service, key directories)
• What languages and frameworks are used?
• Where are the risk-sensitive paths? (auth, payments, PII, shell, IPC)
• What linters/formatters already exist?
• Are there context-worthy docs, schemas, or specs?
• What should be ignored? (generated code, build output, lockfiles)

Treat file targeting as part of the design: the same repo map should drive rules, files.json, ignorePatterns, child configs, and any patternRepositories.

Phase 3: Choose file topology and strictness

Read first: references/config-spec.md for parameter formats and cascading behavior.

Always generate: root .greptile/config.json

Generate only when needed:

• .greptile/rules.md — when a rule cannot be expressed as a single sentence under 200 characters, or when good/bad code examples are essential for the reviewer to apply it correctly
• .greptile/files.json — when a context file provides information a reviewer would need to validate correctness (e.g. a database schema to validate model field usage, an API spec to validate endpoint contracts, shared types to check interface conformance)
• Child .greptile/config.json — a subtree needs different strictness, extra rules, or disabledRules

⚠ Steering: The decision thresholds are: (a) use rules.md when a rule needs >200 characters or good/bad code examples to be understood; (b) use files.json when a context file provides information the reviewer *needs* to validate correctness — not just "nice to have" background reading; (c) use child configs only when a subtree's rules, strictness, or comment types genuinely differ.

Monorepo rule: prefer root defaults plus child overrides only where review needs truly diverge.

Inheritance that matters:

• Child rules extend parent rules
• Child disabledRules turns off parent rules by id
• Child strictness, commentTypes, and triggerOnUpdates override
• Child ignorePatterns extend

Strictness guidance:

• 2 for most production code
• 1 for payments, auth, security-critical, or compliance-heavy paths
• 3 for internal tools, prototypes, or low-risk directories where noise must stay low

Only tune PR filters (labels, authors, branches, fileChangeLimit) or summary behavior when the user asks for it or the repo workflow clearly needs it.

Phase 4: Engineer semantic rules from architecture signals

Read first: references/rules-engineering.md for the rule quality bar and category catalog.

Every rule must be repo-specific, specific, measurable, actionable, scoped, semantic, and identifiable.

Use the repo signals to decide which rule categories matter:

Rule-writing heuristics:

• Start with 5-10 rules max
• Tie each rule to an actual repo path, library, contract, or failure mode
• Tell Greptile what the correct alternative is, not just what to avoid
• Give every parent-level rule an id if a child config might disable it
• If a rule applies to the majority of files in the repository rather than a specific subsystem, it is probably too broad — narrow the scope or reconsider whether it's linter work

⚠ Steering: A rule that applies to the majority of files in the repository is almost certainly too broad. Ask: "Does this rule fire on a specific subsystem, or on everything?" If it fires on everything, it's probably linter work (e.g. "no console.log") or needs aggressive scoping. Also, to verify no rule duplicates a linter: for each rule, ask "Could ESLint/Pylint/RuboCop catch this with a regex or AST rule?" If yes, delete it.

Phase 5: Scope rules, context, and file targeting precisely

Read first: references/patterns.md for universal ignore patterns and stack-specific targeting.

Use Greptile's semantic budget where it matters:

• scope must be an array of globs, never a comma-separated string
• ignorePatterns must be a newline-separated string, not an array
• Scope context files to the code they inform; do not attach a Prisma schema to frontend components
• Use patternRepositories only for real cross-repo contracts or design systems, in org/repo format
• Start with the universal ignore patterns from references/patterns.md and add project-specific patterns for any generated or high-noise directories discovered in Phase 2
• If strong linting/formatting already exists, do not add style-focused Greptile rules
• If the repo uses status checks, merge gating, or CI/CD triggers, see references/integration.md for wiring Greptile into the delivery workflow

When deciding between JSON rules and prose rules:

• Use config.json for crisp pass/fail rules
• Use rules.md only when examples or narrative make the rule materially easier for Greptile to apply

Phase 6: Validate, output, and recover if needed

Read first: references/troubleshooting.md for diagnostic flow and canary verification; references/anti-patterns.md for preflight checks.

Before finalizing any output:

• JSON validates locally (for example: python3 -m json.tool.greptile/config.json)
• scope values are arrays of strings
• ignorePatterns is a newline-separated string
• strictness is 1, 2, or 3
• commentTypes only uses "logic", "syntax", "style", "info"
• patternRepositories uses org/repo, not full URLs
• files.json only points to files that actually exist
• No rule duplicates lint/format tooling — review each rule against the litmus test: if ESLint, Pylint, RuboCop, or a regex grep could catch it, remove it
• No .greptile/ + greptile.json coexistence remains after migration

⚠ Steering: Output format matters. Agents commonly forget to include (a) the file tree, (b) reasoning annotations tied to repo evidence, or (c) the canary test. Use the output template: file tree → complete file contents → reasoning annotations (markdown list with rule-id: reason format) → canary rule → migration notes. See references/scenarios.md for complete examples of properly formatted output.

Output every generated configuration in your response to the user with:

1. A markdown code block showing the .greptile/ directory structure (file tree)
2. If filesystem access exists, write/update the .greptile/ files in the repo first unless the user explicitly asked for draft-only output
3. Complete file contents for each generated file
4. Reasoning annotations as a markdown list tied to specific repo evidence, e.g.:

- auth-wrapper-required: Added because all API routes use withWorkspace/withSession wrappers (see lib/auth/workspace.ts:58) - stripe-webhook-sig: Stripe webhook handlers process financial events at app/api/stripe/webhook/; unverified payloads allow forged billing

1. A canary test: a temporary rule to verify Greptile reads the config. Example: {"id": "canary", "rule": "Comment 'canary active' on any PR modifying a README.", "scope": ["**/README.md"], "severity": "low"}. Open a test PR, confirm the canary fires, then remove the rule. See references/troubleshooting.md for the full canary protocol.
2. Migration notes if replacing greptile.json

See references/scenarios.md for complete end-to-end output examples to model your output on.

If Greptile still behaves incorrectly, recover in this order:

1. Repo indexed?
2. JSON valid?
3. .greptile/ conflicting with greptile.json?
4. Config on the PR's source branch?
5. scope actually matches changed files?
6. strictness suppressing the severity you used?
7. Author/branch/file-limit filters excluding the PR?
8. Force a re-review, then remove any canary rule after verification

Do this, not that

Minimal reading sets

"I need to set up Greptile from scratch"

• references/setup.md
• references/config-spec.md
• references/troubleshooting.md

"I need repo-specific rules that catch real issues"

• references/rules-engineering.md
• references/patterns.md
• references/anti-patterns.md

"I need a monorepo design"

• references/config-spec.md
• references/patterns.md
• references/scenarios.md

"I need CI/CD or status-check integration"

• references/integration.md
• references/setup.md
• references/config-spec.md

"The config is noisy, ignored, or not firing"

• references/troubleshooting.md
• references/anti-patterns.md
• references/rules-engineering.md

Reference files

Key gotchas

• Greptile reads config from the source branch, not the target branch
• Config changes affect the next PR or re-review, not past reviews
• ignorePatterns skips review, not indexing
• includeAuthors: [] means all authors
• fileChangeLimit: 0 skips all PRs
• statusCheck: true suppresses "X files reviewed" comment spam better than statusCommentsEnabled: false
• .greptile/ silently overrides greptile.json
• scope is an array; ignorePatterns is a newline-separated string
• Dashboard or org rules may still influence behavior outside repo-local files

Final reminder

Start with the smallest relevant reading set, then expand only when the task requires it. The goal is not to write the most rules — it is to generate the smallest .greptile/ configuration that helps Greptile catch the issues this specific repository actually misses.