Development Tasks

此内容尚不支持你的语言。

Conventions and checks that apply to every change. The guide for each task is listed in the Contributing overview.

Checks

bun run check           # TypeScript strict, including tests/
bun run lint            # Biome; applies fixes in place
bun run check-locales   # when locale keys or command metadata changed
bun run find-stale-translations --reason=unfollowed --base=origin/main   # advisory translation follow-up
bun run db:lifecycle    # when schema.sql changed

bun run lint rewrites files, so commit what it changes. CI runs bun run lint:ci, which fails on formatting instead of fixing it.
bun run db:lifecycle needs a local PostgreSQL user that can create and drop databases. It tests fresh initialization, backup and restore, and the maintenance scripts in a temporary database.
bun run test <path> runs selected files through the same disposable-database setup.
NovelAI fitting and streaming tests require the pinned local tokenizer definitions. Provision novelai_nerdstash_v2, novelai_llama3, and novelai_glm with bun run setup:tokenizers --family <family> before running them; the existing CI test job does this.

bun run vl runs every gate and prints one verdict per gate. Its last line is for scripts: vl-status: PASS exit=0 pass=<n> warn=<n> fail=<n> skip=<n>. Passing gates are silent and failing gates print everything; --verbose prints all output. To keep the output, redirect it to a file instead of piping through grep or tail, which replaces the gate’s exit code with the filter’s:

bun run vl > /tmp/vl.log 2>&1; echo "VL_EXIT=$?" >> /tmp/vl.log

Conventions

Two-space indentation and double quotes (Biome enforces both).
Strict TypeScript without any. Shared types go in src/types/. Validate untrusted input with Zod.
camelCase file names, @/* imports for src/*, and node: imports for Node built-ins.
Comments follow the comment policy. JSDoc on exported functions only where behavior is not obvious from the name and types.
Workspace: in code and contributor docs, a workspace is a Discord server or a DM scope. User-facing text says “server”, or “your DMs” where the DM case matters.
Log with log from src/utils/misc/logger.ts, with context such as errorType and IDs. Handle startup-critical failures differently from recoverable ones. Pick the level by reader:
- log.error: a failure the hoster must act on. It reaches production and error_logs, so keep it deliberate. Pass { persist: false } when the failure is the database path itself.
- log.warn: developer trace (causes, free text, stack details). Hidden in production unless the hoster sets LOG_LEVEL=warn, except provider diagnostics passed with { retain: true }, which reach JSON sinks at every level without creating an error_logs row.
- log.metric: a sample to count or graph. Numbers and string fields from a fixed set (codes, reason keys) only; free text cannot be grouped, so it goes to log.warn. An expected refusal is a metric plus a warn, never an error. The development console hides metric lines unless LOG_LEVEL=debug, so do not remove a metric to quiet it. A sample Grafana should graph also goes to metricSampleRepository.recordSample, which records in every environment.
- Where a provider failure becomes a result object, a reply, or a tool error, call reportProviderFailure (src/utils/provider/providerErrorMetrics.ts). It picks the error or the metric plus warn from expected, and writes the provider detail into the message unchanged so the record matches what Show Full Error displays. Expected failures retain their warning details in JSON logs even at the default production level.
- Streaming error owners pass the resolved request key through the logger’s redactValues option so messages, attached errors and metadata remove that exact value before truncation.
Slash commands only. Every user-facing string goes through localizer(). Follow the interaction timing rules in Command System.
Query with Bun SQL template literals, and write idempotent migrations (IF NOT EXISTS, guarded blocks). Schema reference: Database Schema.
After a write that affects cached reads, invalidate the cache keys, and only after the write succeeds. Never edit a cached object in place. See Caching.
Constants and environment variables: name every magic number as a constant in the module that owns it, with a comment when the value was measured or comes from an external limit. Make a value an environment variable only when it is a deployment choice: host resources, network behavior, external credentials or quotas, or an operator decision that differs between installations. Probabilities, lookbacks, UI geometry, and algorithm tuning stay constants. Use one variable per setting, never an engine-specific name plus a shared fallback. Settings that servers or users change at runtime go in the database. Placement and naming: Adding an Environment Variable.

Proportionality

Every test, variable, CI job, abstraction, and comment costs maintenance. Add one only when you can name the failure it prevents.

Reuse and abstraction

Search for an existing helper, registry, or pattern before writing a new one, and extend it.
Extract a shared helper once a second real caller exists. An expected future caller does not count.
Fix a bug in the shared function after checking every caller.
Do not add compatibility shims, fallback paths, or options for states the code cannot reach. Validate at trust boundaries and trust typed internal values.
File size alone does not justify a split. Name the maintenance or correctness problem.

Tests

Test behavior, regression risk, or an interface other code depends on (an exported API, a stored data shape, a Discord limit). A bug fix gets the regression test that would have caught it.
Do not test that wiring or a helper merely exists.
Loop locales or panels inside one test that collects every failure, instead of generating one test per locale. expectForEveryLocale and collectCaseFailures in tests/helpers/localeCases.ts do the collecting.
Assert localized copy by key, never by quoting its English. Use localizedCopy(locale, key), or localizedProse when panel formatting may wrap the line. Both fail on an unknown key, so a copy edit passes while a wrong or deleted key does not. Quote a literal only for an identifier, such as a command name, or for text the test itself supplied.
Assert an exact count or a complete member list only when that exact count or list is what must not change. Otherwise assert stable IDs, uniqueness, or that the required members are present.
Optional local servers (servers/), installers, and devtools do not get dedicated test suites or CI jobs by default.
A function whose only remaining caller is a test is dead code; delete both.
Add CI coverage only when unattended enforcement is valuable and the environment is supportable.
Before a broad audit, write the candidate filter and the decision rubric.

Review findings

Give every finding, from an agent or a human, one disposition before implementation:

Disposition	Meaning
Accept now	Fix it in this change.
Defer	Record the concrete trigger that will make it worth doing.
Not worth it	Real, but the fix costs more to carry than the risk it removes.
Reject	The premise is false or already handled.

Severity describes impact if the finding is real. It does not make a fix mandatory. A recommendation states its expected value, its implementation and maintenance cost, and the source evidence for its premise. Implement only accepted findings.