Every failure on the authentication, authorization, and configuration
path now surfaces as a typed structured error instead of an ad-hoc
envelope. Users and scripts that consume CLI output get:
- a fixed nine-category taxonomy on the wire, each mapped to a
stable shell exit code (authentication/authorization/config = 3,
network = 4, internal = 5, policy = 6, confirmation = 10)
- identity-aware detail fields (missing_scopes, requested_scopes,
granted_scopes, console_url, log_id, retryable, hint) carried
uniformly on the envelope
- a single canonical policy envelope at exit 6; the legacy
auth_error carve-out is retired
- per-subtype canonical message + hint that preserves Lark's
diagnostic phrasing and routes recovery to the right actor:
app developer (app_scope_not_applied), user (missing_scope,
token_scope_insufficient, user_unauthorized), or tenant admin
(app_unavailable, app_disabled)
- wrong app credentials classify as config/invalid_client whether
surfaced by the Open API endpoint (99991543) or the tenant
access-token mint endpoint (10003 / 10014), instead of
collapsing to a transport error or api/unknown
- local shortcut scope preflight emits the same
authorization/missing_scope envelope (identity + deterministic
missing-scope set) used by the post-call permission path, so AI
consumers read the same structured shape from precheck and from
server-returned permission denial
- streaming download/upload failures keep the same network subtype
split (timeout / TLS / DNS / transport) as the non-stream path
instead of collapsing every cause to a generic transport failure
- console_url is carried only on the bot-perspective
app_scope_not_applied envelope (where the recovery action is
"developer applies the scope at the developer console"); the
user-perspective missing_scope envelope drops the field, since
the only actionable user recovery is `lark-cli auth login --scope`
and pointing an end user at a console they cannot modify is
misleading
- bind workflows (Hermes / OpenClaw / lark-channel) flatten dynamic
Type tags to wire 'config' with the original module name kept
as a metric label
All 10 typed errors are cause-bearing, nil-safe on .Error() and
.Unwrap(), and defensively clone slice setter inputs. Four lint
rules (CheckNilSafeError / CheckBuilderImmutable / CheckUnwrapSymmetry
/ CheckBuildAPIErrorArms) lock these invariants on migrated paths.
29 KiB
lark-cli Error Contract
errs/ defines a typed, RFC 7807–aligned error taxonomy for the CLI. Three
audiences depend on it: AI agents and shell scripts parsing the JSON
envelope on stderr; protocol adapters mapping CLI errors into MCP /
OAuth shapes; and framework + business code producing errors. This file
is the single source of truth for all three.
This document describes the typed authoring target. The refactor lands
in stages; some boundaries (e.g. client.WrapDoAPIError) still operate on
legacy shapes today — see Migration for what is live in each stage.
Migrating an *output.ExitError call site? See Migration. Something off
in production? See Troubleshooting.
Invariants
- Every error belongs to exactly one Category. The set is closed
(
errs/category.go); adding a member requires deliberate review. - Every newly constructed typed error has a Subtype — a stable
lowercase-with-underscores identifier declared in
errs/subtypes*.go. Undeclared subtypes fail CI. The constraint applies only to typed*errs.*literals; stage-1 legacy*core.ConfigErrorflows via the dispatcher'sasExitError→ legacy envelope path (not the typed taxonomy) and is unaffected.errcompat.PromoteConfigErroris a stage-1 passthrough; its stage-2+ typed migration will subject the promoted typed error to this Subtype constraint at that time. Category+Subtypeare wire-stable identifiers consumers may branch on. Renaming either is a breaking change.Codeis the upstream numeric code when known (e.g. Lark API code). It isomitemptyand never carries CLI-internal meaning.- Every typed error embeds
errs.Problem.CheckProblemEmbedrejects exported*Errorstructs that do not. - Wrapping is idempotent: re-wrapping an already-typed error returns it
unchanged across the
errors.As/errors.Unwrapchain. - For the typed-envelope path, exit codes derive from
Categoryonly viaoutput.ExitCodeForCategory— includingSecurityPolicyError, which exits6viaCategoryPolicy. Unmigrated*output.ExitErrorproducers still carry a hand-setCodeuntil they finish migrating.output.ErrBare(code)is the lone exception: a deliberate predicate-command signal that bypasses the envelope (see Predicate commands below).
Wire format
Typed errors render to stderr as one JSON object per process exit:
{
"ok": false,
"identity": "user",
"error": {
"type": "authorization",
"subtype": "missing_scope",
"code": 99991679,
"message": "missing scope `calendar:event:create` for app cli_xxx",
"hint": "run lark-cli auth login --scope calendar:event:create",
"log_id": "20260520-0a1b2c3d",
"missing_scopes": ["calendar:event:create"],
"console_url": "https://open.feishu.cn/app/cli_xxx/auth?q=..."
}
}
| Field | Stability | Notes |
|---|---|---|
ok |
wire-stable | always false for errors |
identity |
wire-stable | user | bot — caller identity; omitted when not resolved |
error.type |
wire-stable | one of the 9 Categories |
error.subtype |
wire-stable | declared Subtype constant |
error.code |
wire-stable | upstream numeric code, omitted when zero |
error.message |
informational | not safe to branch on |
error.hint |
informational | actionable recovery guidance |
error.log_id |
informational | upstream request id (server-side trace) |
error.retryable |
wire-stable | true when present; omitted when false |
| per-Subtype extension fields | per-Subtype-stable | e.g. missing_scopes, console_url, challenge_url |
SecurityPolicyError renders through the same typed envelope as every
other category. error.type is "policy", error.subtype is one of
challenge_required / access_denied, and process exit is 6 via
CategoryPolicy. The legacy auth_error envelope at exit 1 has been
retired.
Categories
| Category | When | Exit | Typed struct |
|---|---|---|---|
validation |
malformed user input | 2 | ValidationError |
authentication |
no valid token / login required | 3 | AuthenticationError |
authorization |
token lacks scope / app permission denied | 3 | PermissionError |
config |
local config missing / unbound | 3 | ConfigError |
network |
DNS, refused, timeout, transport | 4 | NetworkError |
api |
server-side Lark error w/o specific bucket | 1 | APIError |
policy |
content safety / security challenge | 6 | SecurityPolicyError, ContentSafetyError |
internal |
SDK contract violation / decode failure | 5 | InternalError |
confirmation |
high-risk action needs --yes |
10 | ConfirmationRequiredError |
Canonical mapping: internal/output/exitcode.go ExitCodeForCategory.
Note on the
authorization/PermissionErrorasymmetry. The wiretypefield uses the RFC 7807 / taxonomy-formal name"authorization", but the Go type is namedPermissionError. This is deliberate, following the gRPC / Google APIs convention (codes.Unauthenticated+codes.PermissionDenied): each name is chosen to be maximally distinct and readable on its own, not to be perfectly symmetric.AuthenticationErrorandAuthorizationErrordiffer visually only at the 5th character and are easy to confuse in code review;AuthenticationErrorandPermissionErrorcannot be confused. The wire field stays formal because it is the protocol-level taxonomy; the Go type favors call-site readability.
Flow
call site
│ constructs typed error (e.g. *errs.ValidationError)
▼
command runE returns err
│
▼
cmd/root.go handleRootError dispatches:
├─ output.ErrBare(code) → no envelope (stdout already written); exit = code
├─ typed (errs.ProblemOf) → typed JSON envelope; exit = ExitCodeOf(err)
│ (includes *errs.SecurityPolicyError → policy envelope, exit 6)
├─ *core.ConfigError → promoted to typed via errcompat ↑
├─ *output.ExitError → legacy JSON envelope; exit = exitErr.Code
└─ untyped / Cobra error → plain "Error: <msg>" (no envelope); exit 1
Only the typed and *output.ExitError branches emit a JSON envelope on
stderr. Untyped errors (including Cobra's "required flag missing" / unknown
subcommand messages) print plain text and exit 1 — consumers must
tolerate that fallback.
Predicate commands (output.ErrBare)
A small class of commands is predicates: they answer a yes/no
question and signal the answer through the shell exit code so callers
can write if cmd; then ... fi. lark-cli auth check is the canonical
example — its README contract is exit 0 = ok, 1 = missing.
These commands deliberately:
- write a structured JSON answer to stdout themselves, and
- return
output.ErrBare(exitCode)to communicate the exit code to the dispatcher without producing astderrenvelope.
output.ErrBare is not an error in the typed-envelope sense — it
carries no category, subtype, or message. It is a one-bit output-
control signal that lives outside the contract for the same reason
grep -q / diff / systemctl is-active set non-zero exit codes
without printing anything to stderr: pollution of stderr by a
predicate's negative answer would break 2>/dev/null log hygiene in
caller scripts.
New code should not reach for ErrBare unless the command is
genuinely a predicate. Anything carrying recoverable error content
belongs in a typed *errs.XxxError.
Consumers
Go (in-process)
var pe *errs.PermissionError
if errors.As(err, &pe) {
fmt.Println("missing:", pe.MissingScopes)
}
Predicates cover the common categories (errs/predicates.go):
if errs.IsAuthentication(err) { ... }
if errs.IsPermission(err) { ... }
if errs.IsValidation(err) { ... }
Type-agnostic field access:
if p, ok := errs.ProblemOf(err); ok {
log.Printf("cat=%s subtype=%s retryable=%t", p.Category, p.Subtype, p.Retryable)
}
exitCode := output.ExitCodeOf(err) // ExitInternal for non-typed errors
Shell / AI
out=$(lark-cli ... 2>&1)
code=$?
# Untyped / Cobra errors print plain text — guard before jq.
if ! jq -e . >/dev/null 2>&1 <<<"$out"; then
printf '%s\n' "$out" >&2
exit "$code"
fi
case "$(jq -r '.error.type // empty' <<<"$out")" in
authorization) jq -r '.error.missing_scopes[]' <<<"$out" ;;
network) echo "transport failure, safe to retry" ;;
internal) echo "bug — file an issue with log_id $(jq -r '.error.log_id // "n/a"' <<<"$out")" ;;
esac
Unknown fields are forward-compatible additions: ignore, don't fail.
Branch only on type, subtype, code, retryable, and declared
extension fields — message is human-readable prose that may be
reworded without notice.
Producers
Quick reference
The canonical producer surface is the builder API in errs/types.go (per type: struct + NewXxxError + chained WithX setters live in one place):
each NewXxxError(subtype, format, args...) locks Category at the
constructor name, requires Subtype + Message positionally, and exposes
optional fields via chained .WithX(...) setters. Struct literals remain
legal for framework dynamic paths (e.g. classifier fanout) but the lint
CheckTypedErrorCompleteness still requires Category + Subtype +
Message on any literal it sees.
| Situation | Use |
|---|---|
| Bad user input | errs.NewValidationError(subtype, msg).WithParam("--flag") |
| Login required | errs.NewAuthenticationError(errs.SubtypeTokenMissing, msg) |
| Token lacks scope | errclass.BuildAPIError(resp, ctx) |
| Local config missing | errs.NewConfigError(errs.SubtypeNotConfigured, msg) |
| Transport failure | errs.NewNetworkError(errs.SubtypeNetworkTimeout, msg).WithCause(err) (subtype: timeout / tls / dns / server_error / transport) |
| Lark API error | errclass.BuildAPIError(resp, ctx) |
| SDK / decode bug | errs.NewInternalError(errs.SubtypeSDKError, msg).WithCause(err) |
| Policy block | errs.NewSecurityPolicyError(subtype, msg).WithChallengeURL(url) or errs.NewContentSafetyError(subtype, msg).WithRules(...) |
Needs --yes |
errs.NewConfirmationRequiredError(risk, action, msg) |
Authoring discipline
Five rules every producer follows. Some are enforced by lint/errscontract
AST guards (go run -C lint . ..); the rest by code review.
Propagate typed errors unchanged
A function that receives an error already carrying errs.Problem
returns it as-is up the stack. Reclassification at non-boundary frames
(e.g., wrapping a *ValidationError into *InternalError) defeats the
single-source taxonomy and silently downgrades typed signals.
Conforming:
_, err := runtime.DoAPI(req, opts)
if err != nil {
return err // already typed by the framework boundary
}
Non-conforming:
return fmt.Errorf("calling /open-apis: %v", err) // %v strips the typed shape
return &errs.InternalError{Cause: err} // re-decides category
Never return a typed-nil pointer
A typed-nil pointer (var pe *errs.PermissionError; return pe) wraps as
a non-nil interface — errors.As matches and .Error() may panic.
Return interface nil literally.
Non-conforming:
var e *errs.ValidationError // nil pointer
return e // non-nil interface holding nil pointer
Let Category derive the exit code
Do not pick exit codes by hand in new typed producers — ExitCodeForCategory
maps Category to the shell code. A new exit-code requirement means a
new Category, not a one-off override at the call site.
(Legacy *output.ExitError retains hand-set codes until removal;
SecurityPolicyError retains a hand-set code on main until the framework
migration PR retires the carve-out — see Migration.)
Split Message, Hint, and Cause
Each field carries a distinct role:
| Field | Carries | Style |
|---|---|---|
Message |
What is wrong | Direct, lowercase first letter, no trailing period |
Hint |
What to do next | Imperative ("run lark-cli auth login", "use --as user") |
Cause |
The wrapped upstream error, not a stringified copy |
Typed; serialized as json:"-" |
Hint must not be merged into Message. AI agents and humans read them
on separate channels; merging defeats both.
Cause must be a real error. If the upstream returned an error,
place it in Cause so errors.Is and errors.Unwrap walk the chain —
do not inline its .Error() into Message.
Conforming:
return errs.NewNetworkError(errs.SubtypeNetworkTransport,
"request to /open-apis failed after 3 retries").
WithHint("check connectivity and retry; set --log-level debug if it persists").
WithCause(ioErr)
Non-conforming:
Message: fmt.Sprintf("request failed: %v — retry later", ioErr)
// conflates what + what-to-do + cause into one string
ValidationError.Param uses the --flag form
When a *ValidationError originates from a flag value, Param holds the
flag name with leading dashes ("--priority", not "priority"). AI
agents grep this field literally to surface "the bad flag was --X".
For positional arguments, use the canonical name without dashes
("target_user_id").
Constructing typed errors
Prefer the builder API. The constructor pins Category + Subtype +
Message, the chained setters fill optional fields, and the resulting
value retains its concrete *XxxError pointer through the chain so
type-specific setters remain reachable to the end:
return errs.NewValidationError(errs.SubtypeInvalidArgument,
"--data must be a valid JSON object: %v", parseErr).
WithParam("--data")
Why builder over struct literal:
Categoryis locked at the function name — caller cannot mis-specify itSubtypeandMessageare positional arguments —go buildrejects the call site if either is missing- The chain reads top-down: required identity first, optional fields after
- Message is
fmt.Sprintf-formatted from(format, args...), matchingfmt.Errorfmuscle memory and avoiding a separateSprintfline
Struct literals remain legal — CheckTypedErrorCompleteness continues to
enforce Category + Subtype + Message on any literal it sees — and
the framework classifier (internal/errclass/classify.go) still uses
them on the dynamic dispatch path where a Problem value is composed
once and wrapped per Category branch. Outside that pattern, new code
should reach for the builder.
Legacy helpers (output.ErrValidation, output.ErrAuth, output.ErrNetwork)
remain callable during migration but are // Deprecated: — new code goes
through the builder.
Shortcut Execute walkthrough
Adapted from shortcuts/calendar/calendar_suggestion.go:222, whose legacy
form is output.ErrValidation("--duration-minutes must be between 1 and 1440"). The typed migration target (builder form):
Execute: func(ctx context.Context, runtime *common.RuntimeContext) error {
duration := runtime.Int("duration-minutes")
if duration < 1 || duration > 1440 {
return errs.NewValidationError(errs.SubtypeInvalidArgument,
"--duration-minutes must be between 1 and 1440, got %d", duration).
WithHint("pass a value in [1, 1440]").
WithParam("--duration-minutes")
}
_, err := runtime.DoAPI(req, opts)
if err != nil {
return err // already typed by the framework boundary; propagate
}
return nil
}
Two patterns visible: a producer site (the typed *errs.ValidationError
above) and a propagation site (the return err after runtime.DoAPI,
applying Propagate typed errors unchanged).
When the validation logic outgrows a single range check — multiple
flags, format parsing, conditional rules — extract it into a helper that
also returns the typed *errs.ValidationError. The helper, not
Execute, sets Param (a helper bound to one shortcut is normal in
this codebase; see parseTimeRange in
shortcuts/calendar/calendar_agenda.go:144).
Wrapping upstream errors
When a producer receives an error from a function it called, four cases cover the decision:
| Source | Decision | Example |
|---|---|---|
Helper returned a typed *errs.*Error |
Return unchanged | return err |
Helper returned an untyped error tied to user input (strconv.Atoi, json.Unmarshal, …) |
Construct a typed error; put the untyped error in Cause |
return errs.NewValidationError(errs.SubtypeInvalidArgument, "invalid --data: %v", jsonErr).WithCause(jsonErr) |
SDK call via runtime.DoAPI failed |
Return unchanged — the framework boundary already wrapped it | return err |
| Invariant broken (must-not-happen state) | Lift with errs.WrapInternal, set a Message describing the invariant |
return errs.WrapInternal(fmt.Errorf("identity resolver returned nil: %w", err)) |
Prefer the Cause field over fmt.Errorf("ctx: %w", err) when
attaching an upstream error to a typed one. Cause is the chain
errs.UnwrapTypedError walks and the chain consumer code expects;
fmt.Errorf("...: %w", err) only affects .Error() output, which the
wire envelope does not surface.
Boundary helpers (framework-internal)
These helpers are called from framework boundaries, not from domain code:
errs.WrapInternal(err)— lifts an untyped error to*InternalError; already-typed errors pass through unchanged.client.WrapDoAPIError(err)— classifies SDK transport / decode failures into*errs.NetworkError/*errs.InternalErrorat the SDK boundary.client.WrapJSONResponseParseError(body, err)— lifts response-layer JSON parse failures to*errs.InternalError.
If you find yourself reaching for WrapDoAPIError from a shortcuts/**
package, you are probably calling the SDK at the wrong layer — go
through runtime.DoAPI.
Extending the taxonomy
Add a Subtype
- Add a constant in
errs/subtypes.gounder the right Category block. Subtypes are framework-shared — service-specific Subtypes are an anti-pattern (the wirecodefield already identifies the source service; Subtype encodes cross-service semantics likenot_found,quota_exceeded). - If it maps from a Lark code, register the mapping in
internal/errclass/codemeta_<service>.go. - Add a dispatch test in
internal/errclass/classify_test.go. - Reference the constant from a producer.
go run -C lint . ..—CheckDeclaredSubtypefails until the constant is wired through.
ad_hoc_* subtypes are a temporary unblocker that label a value for
follow-up, not a permanent identifier. Resolve any ad_hoc_* to a
declared constant within one week of introduction; CheckAdHocSubtype
emits a warning to keep them visible.
Add a typed Error struct
Rare; the existing structs cover the 9 Categories with room. If you must:
- In
errs/types.go, add a new section with: the struct embeddingerrs.Problem, a nil-receiver-safeUnwrap()if it carriesCause, aNewXxxError(subtype, format, args...)constructor, and one chainedWithXsetter per extension field. - Add an
IsXxxpredicate inerrs/predicates.go. - Add a wire-format pin in
errs/marshal_test.goand a builder-chain pin inerrs/types_builder_test.go.
CheckProblemEmbed enforces the Problem embed at lint time. New
top-level wire fields are forbidden — per-Subtype data goes into the
typed struct as a documented extension field, not into the envelope's
top level.
CI guards
| Check | Enforces | Where |
|---|---|---|
| forbidigo | business path (shortcuts/**, cmd/service/**) must not call legacy output.* error constructors — route through the typed classifier |
.golangci.yml |
CheckProblemEmbed |
every exported *Error embeds errs.Problem |
lint/errscontract/ AST |
CheckNoRegistrar |
no mergeCodeMeta / RegisterServiceMap from service code |
lint/errscontract/ AST |
CheckAdHocSubtype |
ad_hoc_* Subtypes labeled for promotion (warn) |
lint/errscontract/ AST |
CheckDeclaredSubtype |
every Subtype: value is a declared constant or ad_hoc_* |
lint/errscontract/ AST |
CheckTypedErrorCompleteness |
every *errs.<X>Error{Problem: errs.Problem{...}} literal must set Category, Subtype, and Message |
lint/errscontract/ AST |
CI runs lint/ on every PR. Locally: go run -C lint . ... The
lintcheck CLI lives in its own Go module so its golang.org/x/tools
dependency stays out of the shipped lark-cli binary's module graph;
see lint/README.md for how to add a new lint domain.
Stability
| Tier | Surface | Change policy |
|---|---|---|
| Wire-stable | error.type, error.subtype, error.code, error.retryable, declared extension fields, Category enum values |
breaking change ⇒ semver major; deprecation window required |
| Additive | new Category, new declared Subtype, new extension field on an existing struct | minor release; consumers ignore unknown fields by contract |
| Experimental | ad_hoc_* Subtypes; fields documented as such in errs/types.go |
may change or be promoted/removed within one release |
The deprecated *output.ExitError surface is outside these tiers — it
will be removed once business migration completes.
Migration
Strategy shift (2026-05-26). The original plan (docs/design/errors-refactor/spec.md v2.12 §9) was a centrally-driven 4-PR rollout — framework → auth domain → multi-pilot → full-repo + legacy removal. That plan is superseded by a hybrid model: framework owner ships framework-level hardening (including a typed *errs.*Error migration of internal/**) as one focused PR; business-domain typed migration is self-service via docs/errors-guide.md and the builder API, with no central sweep timeline.
Why the shift: 800+ legacy call sites split across 8+ business domains do not all share a single reviewer's bandwidth, and the contract is now expressive enough that each domain owner can migrate their own code from the guide without coordinating with framework owner.
Current state
-
Framework slice — ✅ shipped (PR #984). The
errs/typed taxonomy, classifier (internal/errclass), promotion stub (internal/errcompat, passthrough), dispatcher hook (WriteTypedErrorEnvelope), and thelint/errscontractAST guards. Wire shapes preserved byte-for-byte versus pre-PR, with one intentional semantic fix: config-class errors (*core.ConfigError) now exit3instead of2, aligning withExitCodeForCategory(config errors share the auth exit slot per the taxonomy). The classifier and promote helpers are shipped but unused in production paths — they exist so framework migration can plug in without re-architecting. -
Builder API — ✅ shipped (this branch).
errs/types.goadds the canonical producer surface (errs.NewXxxError(subtype, format, args...).WithX(...)) for all 10 typed types, alongside each struct declaration. Constructor signature pinsCategory(via function name) andSubtype+Message(positional), so the producer cannot mis-specify any of the three identity fields. Optional fields chain through.WithX(...)setters that preserve the concrete pointer type.
Next: framework migration PR (planned)
A single PR consolidates the work the original §9 spec split across PRs 2–4 — restricted to framework code, no business sweep:
- Migrate
internal/**typed construction to the builder API. ~16 call sites ininternal/errclass/classify.go(BuildAPIError fanout),internal/auth/transport.go(SecurityPolicy),internal/auth/uat_client.go,internal/errcompat/promote*.go,internal/client/client.go,internal/client/api_errors.go. - Land the framework-side semantic changes previously scoped to spec §9 PR 2:
SecurityPolicyErrorexit1→6,WrapDoAPIErrortyped (*NetworkErrorwith subtype timeout/tls/dns/server_error/transport,*InternalErrorfor JSON-decode),WrapJSONResponseParseErrortyped,errcompat.PromoteConfigErrorreal Type routing,PromoteAuthErrorhelper + dispatcher wiring, 10 credential Lark codes registered in codeMeta, 99991543 config classification,resolveAccessTokentyped*AuthenticationError,BuildAPIErrorfilling*PermissionError.MissingScopes/Identity/ConsoleURL, deletion ofscopeAwareChecker. - Add
forbidigorule banningoutput.Err*constructors inshortcuts/**andcmd/**(mirrors the contract that new business code must use the builder). - CHANGELOG lists the resulting ~10 shell-exit-code shifts in one release entry (vs the spec §1 spread of 11 — the remaining one site lives in
taskbusiness code).
Business-domain migration (self-service, no central timeline)
Each business package migrates its own output.Err* call sites to the builder when convenient — typically batched within one domain. The guide at docs/errors-guide.md walks owners through the 8 typical error modes (validation / authorization / authentication / config / network / api / internal / policy) with real file:line examples from main. The three-layer extension model (add Subtype / add field / add Category) handles cases the existing taxonomy does not cover.
Helper assertions accept both shapes during migration (see shortcuts/mail/mail_shortcut_validation_test.go assertValidationError) so domain migrations stay green incrementally.
Legacy removal
Deferred until business migration completion approaches the asymptote. Errorf, ErrAPI, ErrAuth, ErrWithHint, ErrBare, ClassifyLarkError, ErrDetail, ExitError, and ErrorEnvelope are // Deprecated: today and stay callable. No fixed removal date.
Before / after at a call site
// before (legacy)
return output.ErrAPI(larkCode, "create event failed", resp.RawBody())
// after (typed) — cc carries Brand / AppID / Identity from the caller's context
return errclass.BuildAPIError(parsedResp, cc)
// before (legacy validation)
return output.ErrValidation("--duration-minutes must be between 1 and 1440")
// after (builder)
return errs.NewValidationError(errs.SubtypeInvalidArgument,
"--duration-minutes must be between 1 and 1440, got %d", duration).
WithParam("--duration-minutes")
Troubleshooting
Envelope shows type=api subtype=unknown for what should be a more
specific category. The Lark code is unknown to LookupCodeMeta and fell
through to the generic bucket (internal/errclass/classify.go). Add the
code to internal/errclass/codemeta_<service>.go with the right Category
and Subtype, plus a dispatch test in classify_test.go.
Envelope shows type=internal subtype=sdk_error. Origin is
client.WrapDoAPIError taking the non-transport branch
(internal/client/api_errors.go). Check: did the SDK fail to decode the
response (look for subtype=invalid_response in the wrapped chain)? Was the
transport detection too narrow for this error (e.g. a *url.Error with an
inner that does not satisfy net.Error)? Either widen the transport
predicate or add an explicit typed wrap upstream.
CheckDeclaredSubtype rejects my Subtype. The constant must be
declared in errs/subtypes*.go and referenced from the dispatch path.
Bare string literals trip CheckDeclaredSubtype unless they match the
ad_hoc_* prefix; ad_hoc_* then trips CheckAdHocSubtype as a
follow-up warning.
errors.As(&typedErr) panics with a nil-pointer receiver. A typed-nil
slipped through. All typed errors define nil-safe Unwrap(), but
returning a typed-nil pointer up the stack still defeats errors.As.
Return interface nil from constructors, never a typed-nil pointer.
Exit code is 5 (internal) when I expected 3 (auth). The error was not
typed before reaching handleRootError. Wrap at the boundary
(client.WrapDoAPIError or a typed constructor) — the bare error.Error()
string cannot be classified retroactively.
Security & privacy
log_idis a server-side trace token. Safe to surface; it does not carry user content.missing_scopesis app configuration, not user data.MessageandHintmust not contain tokens, JWTs, or personally identifying values. CI does not catch this — producer responsibility.- Wrapped
Causeis not serialized to the wire (json:"-"). It is retained for in-processerrors.Is/errors.Unwraptraversal and optional debug logging only.
Pointers (task-driven)
- Which struct to construct? → Producers / Quick reference
- Add a new condition? → Add a Subtype
- Consume from a shell script? → Consumers / Shell / AI
- Understand or fix a CI failure? → CI guards
- Migrate a legacy
ExitErrorcall site? → Migration + the Deprecated note on the symbol being replaced. - Read source. →
errs/doc.go→errs/category.go→errs/types.go→errs/predicates.go→internal/errclass/→cmd/root.gohandleRootError.