AGENTS.md

Guidelines for AI agents working in this codebase.

Project Overview

Sentry CLI is a command-line interface for Sentry, built with Bun and Stricli.

Goals

• Zero-config experience - Auto-detect project context from DSNs in source code and env files
• AI-powered debugging - Integrate Seer AI for root cause analysis and fix plans
• Developer-friendly - Follow gh CLI conventions for intuitive UX
• Agent-friendly - JSON output and predictable behavior for AI coding agents
• Fast - Native binaries via Bun, SQLite caching for API responses

Key Features

• DSN Auto-Detection - Scans .env files and source code (JS, Python, Go, Java, Ruby, PHP) to find Sentry DSNs
• Project Root Detection - Walks up from CWD to find project boundaries using VCS, language, and build markers
• Directory Name Inference - Fallback project matching using bidirectional word boundary matching
• Multi-Region Support - Automatic region detection with fan-out to regional APIs (us.sentry.io, de.sentry.io)
• Monorepo Support - Generates short aliases for multiple projects
• Seer AI Integration - issue explain and issue plan commands for AI analysis
• OAuth Device Flow - Secure authentication without browser redirects

Cursor Rules (Important!)

Before working on this codebase, read the Cursor rules:

• .cursor/rules/bun-cli.mdc - Bun API usage, file I/O, process spawning, testing
• .cursor/rules/ultracite.mdc - Code style, formatting, linting rules

Quick Reference: Commands

Note: Always check package.json for the latest scripts.

# Development
bun install                              # Install dependencies
bun run dev                              # Run CLI in dev mode
bun run --env-file=.env.local src/bin.ts # Dev with env vars

# Build
bun run build                            # Build for current platform
bun run build:all                        # Build for all platforms

# Type Checking
bun run typecheck                        # Check types

# Linting & Formatting
bun run lint                             # Check for issues
bun run lint:fix                         # Auto-fix issues (run before committing)

# Testing
bun test                                 # Run all tests
bun test path/to/file.test.ts            # Run single test file
bun test --watch                         # Watch mode
bun test --filter "test name"            # Run tests matching pattern
bun run test:unit                        # Run unit tests only
bun run test:e2e                         # Run e2e tests only

Rules: No Runtime Dependencies

CRITICAL: All packages must be in devDependencies, never dependencies. Everything is bundled at build time via esbuild. CI enforces this with bun run check:deps.

When adding a package, always use bun add -d <package> (the -d flag).

When the @sentry/api SDK provides types for an API response, import them directly from @sentry/api instead of creating redundant Zod schemas in src/types/sentry.ts.

Rules: Use Bun APIs

CRITICAL: This project uses Bun as runtime. Always prefer Bun-native APIs over Node.js equivalents.

Read the full guidelines in .cursor/rules/bun-cli.mdc.

Bun Documentation: https://bun.sh/docs - Consult these docs when unsure about Bun APIs.

Quick Bun API Reference

Task	Use This	NOT This
Read file	await Bun.file(path).text()	fs.readFileSync()
Write file	await Bun.write(path, content)	fs.writeFileSync()
Check file exists	await Bun.file(path).exists()	fs.existsSync()
Spawn process	Bun.spawn()	child_process.spawn()
Shell commands	Bun.$\command``	child_process.exec()
Find executable	Bun.which("git")	which package
Glob patterns	new Bun.Glob()	glob / fast-glob packages
Sleep	await Bun.sleep(ms)	setTimeout with Promise
Parse JSON file	await Bun.file(path).json()	Read + JSON.parse

Exception: Use node:fs for directory creation with permissions:

import { mkdirSync } from "node:fs";
mkdirSync(dir, { recursive: true, mode: 0o700 });

Exception: Bun.$ (shell tagged template) has no shim in script/node-polyfills.ts and will crash on the npm/node distribution. Until a shim is added, use execSync from node:child_process for shell commands that must work in both runtimes:

import { execSync } from "node:child_process";
const result = execSync("id -u username", { encoding: "utf-8", stdio: ["pipe", "pipe", "ignore"] });

Architecture

The full project-structure tree — including the live command/subcommand list and the
domain API modules — is generated from the route tree and lives in
docs/src/content/docs/contributing.md
(the project-structure block produced by script/generate-docs-sections.ts). It is
kept in sync automatically, so it is not duplicated here to avoid drift. For the
current command list run ls src/commands/ or sentry --help.

Top-level layout:

• src/bin.ts — entry point; src/app.ts — Stricli application setup;
  src/context.ts — dependency-injection context.
• src/commands/ — one directory per command group (auth, cli, dashboard,
  event, issue, log, org, project, release, replay, repo, sourcemap,
  span, team, trace, trial, local, …) plus standalone command files
  (api.ts, explore.ts, help.ts, init.ts, schema.ts).
• src/lib/ — shared utilities. Key subtrees: api/ (domain API modules),
  db/ (SQLite layer), dsn/ (DSN detection, with per-language extractors under
  dsn/languages/), and formatters/ (output formatting). See the file-locations
  table below and the JSDoc in each module for details.
• src/types/ — TypeScript types and Zod schemas.
• test/ — tests mirroring src/ (unit, *.property.test.ts,
  *.model-based.test.ts, e2e/, fixtures/, mocks/).
• docs/ — documentation site (Astro + Starlight); script/ — build/utility
  scripts; .cursor/rules/ — Cursor AI rules; biome.jsonc — lint config.

Key Patterns

CLI Commands (Stricli)

Commands use Stricli wrapped by src/lib/command.ts.

CRITICAL: Import buildCommand from ../../lib/command.js, NEVER from @stricli/core directly — the wrapper adds telemetry, --json/--fields injection, and output rendering.

Pattern:

import { buildCommand } from "../../lib/command.js";
import type { SentryContext } from "../../context.js";
import { CommandOutput } from "../../lib/formatters/output.js";

export const myCommand = buildCommand({
  docs: {
    brief: "Short description",
    fullDescription: "Detailed description",
  },
  output: {
    human: formatMyData,                // (data: T) => string
    jsonTransform: jsonTransformMyData, // optional: (data: T, fields?) => unknown
    jsonExclude: ["humanOnlyField"],    // optional: strip keys from JSON
  },
  parameters: {
    flags: {
      limit: { kind: "parsed", parse: Number, brief: "Max items", default: 10 },
    },
  },
  async *func(this: SentryContext, flags) {
    const data = await fetchData();
    yield new CommandOutput(data);
    return { hint: "Tip: use --json for machine-readable output" };
  },
});

Key rules:

• Functions are async *func() generators — yield new CommandOutput(data), return { hint }.
• output.human receives the same data object that gets serialized to JSON — no divergent-data paths.
• The wrapper auto-injects --json and --fields flags. Do NOT add your own json flag.
• Do NOT use stdout.write() or if (flags.json) branching — the wrapper handles it.

Command File Structure

Command files in src/commands/ should focus on three concerns:

1. Argument parsing — positional args, flags, URL detection
2. API orchestration — fetching data, error handling, enrichment
3. Output dispatch — yield new CommandOutput(data)

Formatting and rendering logic belongs in src/lib/formatters/<domain>.ts. If a command file exceeds ~400 lines, extract formatting helpers into a dedicated formatter module.

Reference: src/lib/formatters/replay.ts (extracted from replay/view.ts), src/lib/formatters/trace.ts, src/lib/formatters/human.ts.

Lint enforcement: stderr.write() is banned in command files (GritQL rule). Use logger for diagnostics and CommandOutput for data output.

Route Maps (Stricli)

Route groups use Stricli's buildRouteMap wrapped by src/lib/route-map.ts.

CRITICAL: Import buildRouteMap from ../../lib/route-map.js, NEVER from @stricli/core directly — the wrapper auto-injects standard subcommand aliases based on which route keys exist:

Manually specified aliases in aliases are merged with (and take precedence over) auto-generated ones. Do NOT manually add aliases that are already in the standard set above.

import { buildRouteMap } from "../../lib/route-map.js";

export const myRoute = buildRouteMap({
  routes: {
    list: listCommand,
    view: viewCommand,
    create: createCommand,
  },
  defaultCommand: "view",
  // No need for aliases — ls, show, and new are auto-injected.
  // Only add aliases for non-standard mappings:
  // aliases: { custom: "list" },
  docs: {
    brief: "Manage my resources",
  },
});

Positional Arguments

Use parseSlashSeparatedArg from src/lib/arg-parsing.ts for the standard [<org>/<project>/]<id> pattern. Required identifiers (trace IDs, span IDs) should be positional args, not flags.

import { parseSlashSeparatedArg, parseOrgProjectArg } from "../../lib/arg-parsing.js";

// "my-org/my-project/abc123" → { id: "abc123", targetArg: "my-org/my-project" }
const { id, targetArg } = parseSlashSeparatedArg(first, "Trace ID", USAGE_HINT);
const parsed = parseOrgProjectArg(targetArg);
// parsed.type: "auto-detect" | "explicit" | "project-search" | "org-all"

Reference: span/list.ts, trace/view.ts, event/view.ts

Markdown Rendering

All non-trivial human output must use the markdown rendering pipeline:

• Build markdown strings with helpers: mdKvTable(), colorTag(), escapeMarkdownCell(), renderMarkdown()
• NEVER use raw muted() / chalk in output strings — use colorTag("muted", text) inside markdown
• Tree-structured output (box-drawing characters) that can't go through renderMarkdown() should use the plainSafeMuted pattern: isPlainOutput() ? text : muted(text)
• isPlainOutput() precedence: SENTRY_PLAIN_OUTPUT > NO_COLOR > FORCE_COLOR (TTY only) > !isTTY
• isPlainOutput() lives in src/lib/formatters/plain-detect.ts (re-exported from markdown.ts for compat)

Reference: formatters/trace.ts (formatAncestorChain), formatters/human.ts (plainSafeMuted)

Create & Delete Command Standards

Mutation (create/delete) commands use shared infrastructure from src/lib/mutate-command.ts,
paralleling list-command.ts for list commands.

Delete commands MUST use buildDeleteCommand() instead of buildCommand(). It:

1. Auto-injects --yes, --force, --dry-run flags with -y, -f, -n aliases
2. Runs a non-interactive safety guard before func() — refuses to proceed if
   stdin is not a TTY and --yes/--force was not passed (dry-run bypasses)
3. Options to skip specific injections (noForceFlag, noDryRunFlag, noNonInteractiveGuard)

import { buildDeleteCommand, confirmByTyping, isConfirmationBypassed, requireExplicitTarget } from "../../lib/mutate-command.js";

export const deleteCommand = buildDeleteCommand({
  // Same args as buildCommand — flags/aliases auto-injected
  async *func(this: SentryContext, flags, target) {
    requireExplicitTarget(parsed, "Entity", "sentry entity delete <target>");
    if (flags["dry-run"]) { yield preview; return; }
    if (!isConfirmationBypassed(flags)) {
      if (!await confirmByTyping(expected, promptMessage)) return;
    }
    await doDelete();
  },
});

Create commands import DRY_RUN_FLAG and DRY_RUN_ALIASES for consistent dry-run support:

import { DRY_RUN_FLAG, DRY_RUN_ALIASES } from "../../lib/mutate-command.js";

// In parameters:
flags: { "dry-run": DRY_RUN_FLAG, team: { ... } },
aliases: { ...DRY_RUN_ALIASES, t: "team" },

Key utilities in mutate-command.ts:

• isConfirmationBypassed(flags) — true if --yes or --force is set
• guardNonInteractive(flags) — throws in non-interactive mode without --yes
• confirmByTyping(expected, message) — type-out confirmation prompt
• requireExplicitTarget(parsed, entityType, usage) — blocks auto-detect for safety
• DESTRUCTIVE_FLAGS / DESTRUCTIVE_ALIASES — spreadable bundles for manual use

List Command Pagination

All list commands with API pagination MUST use the shared cursor-stack
infrastructure for bidirectional pagination (-c next / -c prev):

import { LIST_CURSOR_FLAG } from "../../lib/list-command.js";
import {
  buildPaginationContextKey, resolveCursor,
  advancePaginationState, hasPreviousPage,
} from "../../lib/db/pagination.js";

export const PAGINATION_KEY = "my-entity-list";

// In buildCommand:
flags: { cursor: LIST_CURSOR_FLAG },
aliases: { c: "cursor" },

// In func():
const contextKey = buildPaginationContextKey("entity", `${org}/${project}`, {
  sort: flags.sort, q: flags.query,
});
const { cursor, direction } = resolveCursor(flags.cursor, PAGINATION_KEY, contextKey);
const { data, nextCursor } = await listEntities(org, project, { cursor, ... });
advancePaginationState(PAGINATION_KEY, contextKey, direction, nextCursor);
const hasPrev = hasPreviousPage(PAGINATION_KEY, contextKey);
const hasMore = !!nextCursor;

Cursor stack model: The DB stores a JSON array of page-start cursors
plus a page index. Each entry is an opaque string — plain API cursors,
compound cursors (issue list), or extended cursors with mid-page bookmarks
(dashboard list). -c next increments the index, -c prev decrements it,
-c first resets to 0. The stack truncates on back-then-forward to avoid
stale entries. "last" is a silent alias for "next".

Hint rules: Show -c prev when hasPreviousPage() returns true.
Show -c next when hasMore is true. Include both nextCursor and
hasPrev in the JSON envelope.

Navigation hint generation: Use paginationHint() from
src/lib/list-command.ts to build bidirectional navigation strings.
Pass it pre-built prevHint/nextHint command strings and it returns
the combined "Prev: X | Next: Y" string (or single-direction, or "").
Do NOT assemble navParts arrays manually — the shared helper ensures
consistent formatting across all list commands.

import { paginationHint } from "../../lib/list-command.js";

const nav = paginationHint({
  hasPrev,
  hasMore,
  prevHint: `sentry entity list ${org}/ -c prev`,
  nextHint: `sentry entity list ${org}/ -c next`,
});
if (items.length === 0 && nav) {
  hint = `No entities on this page. ${nav}`;
} else if (hasMore) {
  header = `Showing ${items.length} entities (more available)\n${nav}`;
} else if (nav) {
  header = `Showing ${items.length} entities\n${nav}`;
}

Three abstraction levels for list commands (prefer the highest level
that fits your use case):

1. buildOrgListCommand (team/repo list) — Fully automatic. Pagination
   hints, cursor management, JSON envelope, and human formatting are all
   handled internally. New simple org-scoped list commands should use this.

2. dispatchOrgScopedList with overrides (project/issue list) — Automatic
   for most modes; custom "org-all" override calls resolveCursor +
   advancePaginationState + paginationHint manually.

3. buildListCommand with manual pagination (trace/span/dashboard list) —
   Command manages its own pagination loop. Must call resolveCursor,
   advancePaginationState, hasPreviousPage, and paginationHint directly.

Auto-pagination for large limits:

When --limit exceeds API_MAX_PER_PAGE (100), list commands MUST transparently
fetch multiple pages to fill the requested limit. Cap perPage at
Math.min(flags.limit, API_MAX_PER_PAGE) and loop until results.length >= limit
or pages are exhausted. This matches the listIssuesAllPages pattern.

const perPage = Math.min(flags.limit, API_MAX_PER_PAGE);
for (let page = 0; page < MAX_PAGINATION_PAGES; page++) {
  const { data, nextCursor } = await listPaginated(org, { perPage, cursor });
  results.push(...data);
  if (results.length >= flags.limit || !nextCursor) break;
  cursor = nextCursor;
}

Never pass a per_page value larger than API_MAX_PER_PAGE to the API — the
server silently caps it, causing the command to return fewer items than requested.

Reference template: trace/list.ts, span/list.ts, dashboard/list.ts

ID Validation

Use shared validators from src/lib/hex-id.ts:

• validateHexId(value, label) — 32-char hex IDs (trace IDs, log IDs). Auto-strips UUID dashes.
• validateSpanId(value) — 16-char hex span IDs. Auto-strips dashes.
• validateTraceId(value) — thin wrapper around validateHexId in src/lib/trace-id.ts.

All normalize to lowercase. Throw ValidationError on invalid input.

Sort Convention

Use "date" for timestamp-based sort (not "time"). Export sort types from the API layer (e.g., SpanSortValue from api/traces.ts), import in commands. This matches issue list, trace list, and span list.

Generated Docs & Skills

All command docs and skill files are generated via bun run generate:docs (which runs generate:command-docs then generate:skill). This runs automatically as part of dev, build, typecheck, and test scripts.

• Command docs (docs/src/content/docs/commands/*.md) are gitignored and generated from CLI metadata + hand-written fragments in docs/src/fragments/commands/.
• Skill files (plugins/sentry-cli/skills/sentry-cli/) are committed (consumed by external plugin systems) and auto-committed by CI when stale.
• Edit fragments in docs/src/fragments/commands/ for custom examples and guides.
• bun run check:fragments validates fragment route consistency.
• Positional placeholder values must be descriptive: "org/project/trace-id" not "args".

Zod Schemas for Validation

All config and API types use Zod schemas:

import { z } from "zod";

export const MySchema = z.object({
  field: z.string(),
  optional: z.number().optional(),
});

export type MyType = z.infer<typeof MySchema>;

// Validate data
const result = MySchema.safeParse(data);
if (result.success) {
  // result.data is typed
}

Type Organization

• Define Zod schemas alongside types in src/types/*.ts
• Key type files: sentry.ts (API types), config.ts (configuration), oauth.ts (auth flow), seer.ts (Seer AI)
• Re-export from src/types/index.ts
• Use type imports: import type { MyType } from "../types/index.js"

SQL Utilities

Use the upsert() helper from src/lib/db/utils.ts to reduce SQL boilerplate:

import { upsert, runUpsert } from "../db/utils.js";

// Generate UPSERT statement
const { sql, values } = upsert("table", { id: 1, name: "foo" }, ["id"]);
db.query(sql).run(...values);

// Or use convenience wrapper
runUpsert(db, "table", { id: 1, name: "foo" }, ["id"]);

// Exclude columns from update
const { sql, values } = upsert(
  "users",
  { id: 1, name: "Bob", created_at: now },
  ["id"],
  { excludeFromUpdate: ["created_at"] }
);

Error Handling

All CLI errors extend the CliError base class from src/lib/errors.ts:

// Error hierarchy in src/lib/errors.ts
// Exit codes are defined in the EXIT constant object — use EXIT.* constants
// when constructing errors, never hardcode numeric exit codes outside errors.ts.
CliError (base, exitCode=1)
├── HostScopeError (exitCode=13)
├── ApiError (exitCode=30 — HTTP/API failures)
├── AuthError (exitCode=10–12 by reason — 'not_authenticated' | 'expired' | 'invalid')
├── ConfigError (exitCode=20 — configuration/DSN)
├── OutputError (exitCode=60 — data rendered, but operation failed)
├── ContextError (exitCode=22 — missing context)
├── ResolutionError (exitCode=23 — value provided but not found)
├── ValidationError (exitCode=21 — input validation)
├── DeviceFlowError (exitCode=51 — OAuth flow)
├── SeerError (exitCode=40–42 by reason — 'not_enabled' | 'no_budget' | 'ai_disabled')
├── TimeoutError (exitCode=31 — operation timed out)
├── UpgradeError (exitCode=50 — upgrade failures)
└── WizardError (exitCode=61–64 by workflow step — init wizard error)

Exit code ranges: 1x=auth, 2x=input/config, 3x=API/network, 4x=feature/billing,
5x=operations, 6x=command-specific. See EXIT in src/lib/errors.ts and
https://cli.sentry.dev/exit-codes/ for the full reference.

Choosing between ContextError, ResolutionError, and ValidationError:

ContextError rules:

• command must be a single-line CLI usage example (e.g., "sentry org view <slug>")
• Constructor throws if command contains \n (catches misuse in tests)
• Pass alternatives: [] when defaults are irrelevant (e.g., for missing Trace ID, Event ID)
• Use " and " in resource for plural grammar: "Trace ID and span ID" → "are required"

CI enforcement: bun run check:errors scans for ContextError with multiline commands, CliError with ad-hoc "Try:" strings, and silent catch blocks (advisory).

// Usage examples
throw new ContextError("Organization", "sentry org view <org-slug>");
throw new ContextError("Trace ID", "sentry trace view <trace-id>", []); // no alternatives
throw new ResolutionError("Project 'cli'", "not found", "sentry issue list <org>/cli", [
  "No project with this slug found in any accessible organization",
]);
throw new ValidationError("Invalid trace ID format", "traceId");

Fuzzy suggestions in resolution errors:

When a user-provided name/title doesn't match any entity, use fuzzyMatch() from
src/lib/fuzzy.ts to suggest similar candidates instead of listing all entities
(which can be overwhelming). Show at most 5 fuzzy matches.

Reference: resolveDashboardId() in src/commands/dashboard/resolve.ts.

Catch Block Logging

Silent catch blocks are prohibited in src/ production code. Biome's noEmptyBlockStatements catches syntactically empty catch {} blocks, but blocks with only a return statement and no logging are equally problematic — errors vanish silently, making debugging impossible.

Every catch block must either:

1. Re-throw the error
2. Log with log.debug() or log.warn() for diagnostic visibility
3. Return a fallback value with a log.debug() call explaining the suppression

// WRONG — error vanishes silently
try { data = await fetchOptionalData(); }
catch { return []; }

// RIGHT — error is visible in debug logs
try { data = await fetchOptionalData(); }
catch (error) {
  log.debug("Failed to fetch optional data", error);
  return [];
}

Use logger.withTag("command-name") for tagged logging in command files.

CI enforcement: bun run check:errors includes a silent-catch scan that flags
catch blocks which are empty, comment-only, or return-only without surfacing the
error. It is currently advisory (warns, does not fail CI) because of a pre-existing
backlog; run with SENTRY_STRICT_SILENT_CATCH=1 to enforce. Do not add new silent
catches — they will appear in the scan output during review.

Auto-Recovery for Wrong Entity Types

When a user provides the wrong type of identifier (e.g., an issue short ID
where a trace ID is expected), commands should auto-recover when the
user's intent is unambiguous:

1. Detect the actual entity type using helpers like looksLikeIssueShortId(),
   SPAN_ID_RE, HEX_ID_RE, or non-hex character checks.
2. Resolve the input to the correct type (e.g., issue → latest event → trace ID).
3. Warn via log.warn() explaining what happened.
4. Show the result with a return hint nudging toward the correct command.

When recovery is ambiguous or impossible, keep the existing error but add
entity-aware suggestions (e.g., "This looks like a span ID").

Detection helpers:

• looksLikeIssueShortId(value) — uppercase dash-separated (e.g., CLI-G5)
• SPAN_ID_RE.test(value) — 16-char hex (span ID)
• HEX_ID_RE.test(value) — 32-char hex (trace/event/log ID)
• /[^0-9a-f]/.test(normalized) — non-hex characters → likely a slug/name

Reference implementations:

• event/view.ts — issue short ID → latest event redirect
• span/view.ts — traceId/spanId slash format → auto-split
• trace/view.ts — issue short ID → issue's trace redirect
• hex-id.ts — entity-aware error hints in validateHexId/validateSpanId

Async Config Functions

All config operations are async. Always await:

const token = await getAuthToken();
const isAuth = await isAuthenticated();
await setAuthToken(token, expiresIn);

Adding New Utility Files

Before creating a new src/lib/*.ts utility file, check whether existing shared modules already cover your use case:

If an existing module covers ≥80% of what you need, extend it with new exported functions rather than creating a new file. New files are appropriate when the domain is genuinely new (e.g., replay-search.ts for replay-specific field resolution).

Every new src/lib/**/*.ts file must start with a module-level JSDoc comment describing the module's purpose.

Imports

• Use .js extension for local imports (ESM requirement)
• Group: external packages first, then local imports
• Use type keyword for type-only imports

import { z } from "zod";
import { buildCommand } from "../../lib/command.js";
import type { SentryContext } from "../../context.js";
import { getAuthToken } from "../../lib/config.js";

List Command Infrastructure

Two abstraction levels exist for list commands:

1. src/lib/list-command.ts — buildOrgListCommand factory + shared Stricli parameter constants (LIST_TARGET_POSITIONAL, LIST_JSON_FLAG, LIST_CURSOR_FLAG, buildListLimitFlag). Use this for simple entity lists like team list and repo list.

2. src/lib/org-list.ts — dispatchOrgScopedList with OrgListConfig and a 4-mode handler map: auto-detect, explicit, org-all, project-search. Complex commands (project list, issue list) call dispatchOrgScopedList with an overrides map directly instead of using buildOrgListCommand.

Key rules when writing overrides:

• Each mode handler receives a HandlerContext<T> with the narrowed parsed plus shared I/O (stdout, cwd, flags). Access parsed fields via ctx.parsed.org, ctx.parsed.projectSlug, etc. — no manual Extract<> casts needed.
• Commands with extra fields (e.g., stderr, setContext) spread the context and add them: (ctx) => handle({ ...ctx, flags, stderr, setContext }). Override ctx.flags with the command-specific flags type when needed.
• resolveCursor() must be called inside the org-all override closure, not before dispatchOrgScopedList, so that --cursor validation errors fire correctly for non-org-all modes.
• handleProjectSearch errors must use "Project" as the ContextError resource, not config.entityName.
• Always set orgSlugMatchBehavior on dispatchOrgScopedList to declare how bare-slug org matches are handled. Use "redirect" for commands where listing all entities in the org makes sense (e.g., project list, team list, issue list). Use "error" for commands where org-all redirect is inappropriate. The pre-check uses cached orgs to avoid N API calls — when the cache is cold, the handler's own org-slug check serves as a safety net (throws ResolutionError with a hint).

3. Standalone list commands (e.g., span list, trace list) that don't use org-scoped dispatch wire pagination directly in func(). See the "List Command Pagination" section above for the pattern.

Project Filtering in API Calls

Different Sentry API endpoints use different project filtering mechanisms. Never apply both simultaneously:

When adding a new dataset to explore, verify which filtering mechanism the underlying API expects and handle it in resolveDatasetConfig. The explore command centralizes dataset-specific behavior (sort, query, fetch, field validation) in resolveDatasetConfig — add new datasets there rather than scattering if (dataset === ...) checks through the func body.

Commenting & Documentation (JSDoc-first)

Default Rule

• Prefer JSDoc over inline comments.
• Code should be readable without narrating what it already says.

Required: JSDoc

Add JSDoc comments on:

• Every exported function, class, and type (and important internal ones).
• Types/interfaces: document each field/property (what it represents, units, allowed values, meaning of null, defaults).

Include in JSDoc:

• What it does
• Key business rules / constraints
• Assumptions and edge cases
• Side effects
• Why it exists (when non-obvious)

Inline Comments (rare)

Inline comments are allowed only when they add information the code cannot express:

• "Why" - business reason, constraint, historical context
• Non-obvious behavior - surprising edge cases
• Workarounds - bugs in dependencies, platform quirks
• Hardcoded values - why hardcoded, what would break if changed

Inline comments are NOT allowed if they just restate the code:

// Bad:
if (!person) // if no person  
i++          // increment i   
return result // return result 

// Good:
// Required by GDPR Article 17 - user requested deletion
await deleteUserData(userId)

Prohibited Comment Styles

• ASCII art section dividers - Do not use decorative box-drawing characters like ───────── to create section headers. Use standard JSDoc comments or simple // Section Name comments instead.

Goal

Minimal comments, maximum clarity. Comments explain intent and reasoning, not syntax.

Testing (bun:test + fast-check)

Prefer property-based and model-based testing over traditional unit tests. These approaches find edge cases automatically and provide better coverage with less code.

fast-check Documentation: https://fast-check.dev/docs/core-blocks/arbitraries/

Testing Hierarchy (in order of preference)

1. Model-Based Tests - For stateful systems (database, caches, state machines)
2. Property-Based Tests - For pure functions, parsing, validation, transformations
3. Unit Tests - Only for trivial cases or when properties are hard to express

Test File Naming

Test Environment Isolation (CRITICAL)

Tests that need a database or config directory must use useTestConfigDir() from test/helpers.ts. This helper:

• Creates a unique temp directory in beforeEach
• Sets SENTRY_CONFIG_DIR to point at it
• Restores (never deletes) the env var in afterEach
• Closes the database and cleans up temp files

NEVER do any of these in test files:

• delete process.env.SENTRY_CONFIG_DIR — This pollutes other test files that load after yours
• const baseDir = process.env[CONFIG_DIR_ENV_VAR]! at module scope — This captures a value that may be stale
• Manual beforeEach/afterEach that sets/deletes SENTRY_CONFIG_DIR

Why: Bun's test runner uses --isolate --parallel (see test:unit in package.json), so each test file runs in a fresh global environment within a worker process. That bounds most cross-file leaks to a single worker, but process.env is still shared within a file's lifecycle — if your afterEach deletes the env var, the next describe/test's module-level code (or a beforeEach that re-reads env) gets undefined, causing TypeError: The "paths[0]" property must be of type string. Also, TEST_TMP_DIR is namespaced by BUN_TEST_WORKER_ID in test/constants.ts so parallel workers don't wipe each other's temp state during preload.

// CORRECT: Use the helper
import { useTestConfigDir } from "../helpers.js";

const getConfigDir = useTestConfigDir("my-test-prefix-");

// If you need the directory path in a test:
test("example", () => {
  const dir = getConfigDir();
});

// WRONG: Manual env var management
beforeEach(() => { process.env.SENTRY_CONFIG_DIR = tmpDir; });
afterEach(() => { delete process.env.SENTRY_CONFIG_DIR; }); // BUG!

Property-Based Testing

Use property-based tests when verifying invariants that should hold for any valid input.

import { describe, expect, test } from "bun:test";
import { constantFrom, assert as fcAssert, property, tuple } from "fast-check";
import { DEFAULT_NUM_RUNS } from "../model-based/helpers.js";

// Define arbitraries (random data generators)
const slugArb = array(constantFrom(..."abcdefghijklmnopqrstuvwxyz0123456789".split("")), {
  minLength: 1,
  maxLength: 15,
}).map((chars) => chars.join(""));

describe("property: myFunction", () => {
  test("is symmetric", () => {
    fcAssert(
      property(slugArb, slugArb, (a, b) => {
        // Properties should always hold regardless of input
        expect(myFunction(a, b)).toBe(myFunction(b, a));
      }),
      { numRuns: DEFAULT_NUM_RUNS }
    );
  });

  test("round-trip: encode then decode returns original", () => {
    fcAssert(
      property(validInputArb, (input) => {
        const encoded = encode(input);
        const decoded = decode(encoded);
        expect(decoded).toEqual(input);
      }),
      { numRuns: DEFAULT_NUM_RUNS }
    );
  });
});

Good candidates for property-based testing:

• Parsing functions (DSN, issue IDs, aliases)
• Encoding/decoding (round-trip invariant)
• Symmetric operations (a op b = b op a)
• Idempotent operations (f(f(x)) = f(x))
• Validation functions (valid inputs accepted, invalid rejected)

See examples: test/lib/dsn.property.test.ts, test/lib/alias.property.test.ts, test/lib/issue-id.property.test.ts

Model-Based Testing

Use model-based tests for stateful systems where sequences of operations should maintain invariants.

import { describe, expect, test } from "bun:test";
import {
  type AsyncCommand,
  asyncModelRun,
  asyncProperty,
  commands,
  assert as fcAssert,
} from "fast-check";
import { createIsolatedDbContext, DEFAULT_NUM_RUNS } from "../../model-based/helpers.js";

// Define a simplified model of expected state
type DbModel = {
  entries: Map<string, string>;
};

// Define commands that operate on both model and real system
class SetCommand implements AsyncCommand<DbModel, RealDb> {
  constructor(readonly key: string, readonly value: string) {}
  
  check = () => true;
  
  async run(model: DbModel, real: RealDb): Promise<void> {
    // Apply to real system
    await realSet(this.key, this.value);
    
    // Update model
    model.entries.set(this.key, this.value);
  }
  
  toString = () => `set("${this.key}", "${this.value}")`;
}

class GetCommand implements AsyncCommand<DbModel, RealDb> {
  constructor(readonly key: string) {}
  
  check = () => true;
  
  async run(model: DbModel, real: RealDb): Promise<void> {
    const realValue = await realGet(this.key);
    const expectedValue = model.entries.get(this.key);
    
    // Verify real system matches model
    expect(realValue).toBe(expectedValue);
  }
  
  toString = () => `get("${this.key}")`;
}

describe("model-based: database", () => {
  test("random sequences maintain consistency", () => {
    fcAssert(
      asyncProperty(commands(allCommandArbs), async (cmds) => {
        const cleanup = createIsolatedDbContext();
        try {
          await asyncModelRun(
            () => ({ model: { entries: new Map() }, real: {} }),
            cmds
          );
        } finally {
          cleanup();
        }
      }),
      { numRuns: DEFAULT_NUM_RUNS }
    );
  });
});

Good candidates for model-based testing:

• Database operations (auth, caches, regions)
• Stateful caches with invalidation
• Systems with cross-cutting invariants (e.g., clearAuth also clears regions)

See examples: test/lib/db/model-based.test.ts, test/lib/db/dsn-cache.model-based.test.ts

Test Helpers

Use test/model-based/helpers.ts for shared utilities:

import { createIsolatedDbContext, DEFAULT_NUM_RUNS } from "../model-based/helpers.js";

// Create isolated DB for each test run (prevents interference)
const cleanup = createIsolatedDbContext();
try {
  // ... test code
} finally {
  cleanup();
}

// Use consistent number of runs across tests
fcAssert(property(...), { numRuns: DEFAULT_NUM_RUNS }); // 50 runs

When to Use Unit Tests

Use traditional unit tests only when:

• Testing trivial logic with obvious expected values
• Properties are difficult to express or would be tautological
• Testing error messages or specific output formatting
• Integration with external systems (E2E tests)

Avoiding Unit/Property Test Duplication

When a *.property.test.ts file exists for a module, do not add unit tests that re-check the same invariants with hardcoded examples. Before adding a unit test, check whether the companion property file already generates random inputs for that invariant.

Unit tests that belong alongside property tests:

• Edge cases outside the property generator's range (e.g., self-hosted DSNs when the arbitrary only produces SaaS ones)
• Specific output format documentation (exact strings, column layouts, rendered vs plain mode)
• Concurrency/timing behavior that property tests cannot express
• Integration tests exercising multiple functions together (e.g., writeJsonList envelope shape)

Unit tests to avoid when property tests exist:

• "returns true for valid input" / "returns false for invalid input" — the property test already covers this with random inputs
• Basic round-trip assertions — property tests check decode(encode(x)) === x for all x
• Hardcoded examples of invariants like idempotency, symmetry, or subset relationships

When adding property tests for a function that already has unit tests, remove the unit tests that become redundant. Add a header comment to the unit test file noting which invariants live in the property file:

/**
 * Note: Core invariants (round-trips, validation, ordering) are tested via
 * property-based tests in foo.property.test.ts. These tests focus on edge
 * cases and specific output formatting not covered by property generators.
 */

import { describe, expect, test, mock } from "bun:test";

describe("feature", () => {
  test("should return specific value", async () => {
    expect(await someFunction("input")).toBe("expected output");
  });
});

// Mock modules when needed
mock.module("./some-module", () => ({
  default: () => "mocked",
}));

File Locations

Automated Fix PRs (BugBot / agents)

Automated bug-fix PRs (e.g. Cursor BugBot) must follow these rules to avoid the
duplication and staleness that caused five overlapping PRs to pile up:

1. Check for existing work first. Before opening a PR, search open PRs and
   recently-closed PRs/issues for the same file + symbol:

   gh pr list --state open --search "in:title <file-or-symbol>"
   gh issue list --state all --search "<symbol>"
   

   If an open PR already touches the target function, comment on it or extend
   it instead of opening a duplicate. Multiple BugBot PRs independently re-fixed
   the same JSON.parse guard, withTTY helper, and pagination code.

2. Rebase before review. A PR that is many commits behind main may fail CI
   on unrelated drift (e.g. a lint error in a file the PR never touched) and its
   fix may already be superseded. Rebase onto main and re-verify the bug still
   exists before requesting review. Verify against current main, not the
   snapshot the PR was generated from.

3. Separate correctness fixes from opinion. A real bug (wrong output, crash,
   skipped data) is in scope. A subjective UX change (different hint wording,
   different default) is not a bug — main's current behavior is often
   deliberate. Do not bundle UX opinions into bug-fix PRs; they waste review
   cycles and are usually dropped.

4. Prefer shared helpers over re-deriving fixes. If a correct implementation
   already exists (e.g. autoPaginate() for pagination, safeParseJson() for
   cached JSON), use it rather than hand-rolling a one-off fix. The recurring
   pagination-overshoot and parse-crash bugs were classes solved once centrally.

Long-term Knowledge

For long-term knowledge entries managed by lore (gotchas, patterns, decisions, architecture), see .lore.md in the project root.