case.
- Write all the letters of an initialism in one case. `revive` checks
common initialisms.
-- Put entry points first, then each struct's methods as a group, then
- standalone utility functions. Do not put a utility function between
- the methods of one type.
-- In each group, order functions by first use in the file.
+- Order declarations as imports, constants, variables, types,
+ constructors, methods, then standalone helpers. Types include structs,
+ interfaces, and aliases. Put all types before any constructor or
+ method, also when a file holds more than one type. The Tests section
+ gives the exception for files with tests.
+- Keep each type's methods in one group. Do not put a helper between
+ them. Order functions by first use within their declaration group, not
+ across groups. Entry points do not go before types or constructors
+ merely because they are entry points.
+- Group related constants in one `const (...)` block and related
+ variables in one `var (...)` block. Put an empty line between entries
+ that have comments.
## Entry point and context
Call `defer cancel()` immediately after you create the context.
- Reuse `ctx` when a child context replaces its parent. Use a different
name only when both contexts stay in use. Return `ctx.Err()` directly.
+- Derive each context from the context of the caller. When work must
+ finish even after the caller is cancelled, such as a final write, use
+ `context.WithoutCancel` of the caller's context. Use
+ `context.Background()` only where no caller context exists: in `run`,
+ and in methods that follow `io.Closer`.
+- Name a duration that limits one operation after that operation, and
+ end the name with `Timeout`, so that the name says what it limits
+ without its comment. Define it in the file that applies it, not in the
+ file of another type.
## Errors
-- Handle every returned error.
+- Handle every returned error, including an error assigned to `_`. Never
+ use a blank assignment to hide an unchecked error from a linter.
+ Handle meaningful errors. When the user approves an intentional
+ discard, call the operation with an explicit `// nolint: errcheck`
+ directive and a comment that gives the reason, not a blank assignment.
+ The `check` skill gives the approval rule.
- Reuse an existing `err` for later operations in the same scope. Use a
different variable only to keep or join errors.
- Use `errors.Is` and `errors.As` to examine errors. Do not compare an
error directly with a sentinel, because errors can be wrapped.
`errorlint` checks comparisons and type assertions on errors.
+- Do not match the text of an error from another package. When an
+ external fault leaves no other way, test it through the real code
+ path, so that a change of the text fails a test.
+- Handle a documented sentinel error explicitly when it does not show a
+ failure.
- When a function has more than one error path, wrap each returned error
with the context of its operation. A function with one error path can
return the error directly. `wrapcheck` checks package and interface
- When deferred cleanup can fail, use a named `err` result and join the
cleanup error with `errors.Join`.
+## Logs and cancellation
+
+- Log with the `*slog.Logger` of the part (`daemonLogger`). The handler
+ formats the line; do not format values into the message.
+- Write the message as a short event in plain words, such as "receive
+ failed" or "subagent done". Do not explain what it means or what the
+ user must do.
+- Put names, IDs, counts, and durations in attributes, with short keys
+ (`name`, `id`, `after`, `retry`). Put the cause in `err`; the handler
+ writes it last.
+- Use info for normal events, warn for a failure that is retried or
+ loses nothing, and error for a failure that loses data or stops a
+ part.
+- Cancellation of a service or request context is an expected stop, not
+ a failure. Do not log it as an error, and do not hide all errors only
+ because the context is canceled. A joined error with an independent
+ failure stays a failure.
+- Use the worker or request context for this check. An operation that
+ reaches its own timeout while its worker is active stays a failure.
+- Never log secrets, tokens, message text, or phone numbers. Redact IDs
+ as the existing code does.
+
## Style
- Keep function signatures on one line. Write function bodies on more
table-driven test. A small type of two fields that acts as one value,
such as a key or a coordinate pair, can use a positional literal on
one line.
-- Put an empty line between logical blocks, also after multiline
- declarations, assignments, literals, and calls. Keep an operation with
- its immediate error or result check. Keep a short declaration with its
- immediate use. `wsl_v5` and `nlreturn` check part of this rule. Review
- the remaining logical blocks and immediate-use groups by hand.
+- Put an empty line between logical blocks, such as two independent `if`
+ statements. Put an empty line before and after loops, switches,
+ selects, standalone blocks, and multiline declarations, assignments,
+ literals, and calls, when there is an adjacent statement. These
+ boundaries also apply when a loop immediately uses the declaration
+ before it.
+- The only exception after a multiline construct is its immediate error
+ check. Keep a single-line declaration with its immediate use, and an
+ operation with its immediate error check. A result assertion does not
+ waive the separator after a multiline construct. `wsl_v5` and
+ `nlreturn` check only part of these rules. Review the other boundaries
+ by hand.
+- Group imports in this order: standard library, external dependencies,
+ project imports. Review these groups by hand when the formatter does
+ not check them.
- Use GNU-style command flags: two dashes for long flags, such as
`--cwd`, and one dash only for one-letter flags, such as `-h`.
- Prefer the Go standard library. Prefer Go over cgo.
## Tests
-- Add tests for meaningful behavior. Do not add tests that only repeat
- implementation details.
+- Add tests for behavior changes. Do not add tests that only repeat
+ implementation details, and do not add a test file without behavior
+ worth a test.
- Use `testify/require` for assertions. `testifylint` checks the
assertion forms.
- Name a table-driven case slice `testCases` and its range variable
`tc`. A subtest name describes the behavior, not a generic outcome or
input.
-- In a test file, put constants immediately after the imports, then the
- tests, then supporting types, methods, and helpers.
+- Each file with tests must have the name of its owning production file
+ with `_test` before `.go`: `memory_test.go` tests `memory.go`. Move
+ scenario and integration tests into their owner's test file. Do not
+ add an empty production file to make a test filename valid. Only
+ `_helper_test.go` files are exempt from this pairing rule.
+- In an ordinary `_test.go` file, order declarations as imports,
+ constants, variables, tests, supporting types, constructors, methods,
+ then helpers. This tests-first order is the exception to the
+ production order. Put all supporting types before their constructors
+ and methods.
+- A `_helper_test.go` file has no tests. Order its declarations as
+ imports, constants, variables, types, constructors, methods, then
+ helpers, as in a production file.
- Keep a helper that only one test file uses in that file. Put a helper
that more than one test file uses in a file without tests. Name that
file for what it holds, and end the name with `_helper_test.go`.
- Supply valid required dependencies in tests. Do not make a dependency
- optional only to make a test simpler.
+ optional or add a nil fallback only to make a test simpler.
projects / onechat / blobdiff