Date: 2026-07-10 (unreleased — TENANT_CLOSED + LIMIT_EXCEEDED error-code support. TENANT_CLOSED per runtime spec v0.1.25.13 (cycles-protocol-v0.yaml, runcycles/cycles-protocol#125): ErrorCode.TENANT_CLOSED enum member, TenantClosedError subclass wired into the lifecycle error-code→exception mapping (reservation-creation surfaces), CyclesProtocolError.is_tenant_closed() helper. LIMIT_EXCEEDED per runtime spec v0.1.25.12 (revision 2026-07-04, HTTP 429 rate limiting): enum-only member matching the BUDGET_FROZEN/BUDGET_CLOSED pattern, classified retryable at both the enum and exception layers (429 is transient; previously it fell through to UNKNOWN, which happened to be retryable, so semantics are unchanged — now typed). Enum reordered to mirror spec declaration order. Both purely additive; previously both codes fell through the ErrorCode.from_string forward-compat path to UNKNOWN. See the dated entries at the end of this file. 398 tests pass at 100% coverage.),
2026-07-09 (README + docstring transport-error documentation fix, no version bump — see the dated entry at the end of this file. CyclesTransportError is exported but never raised by the SDK; README and its docstring now describe the actual status == -1 surfacing.),
2026-07-03 (integration-test-only, no version bump — test_health_check now probes the public /actuator/health/readiness endpoint instead of aggregate /actuator/health, which requires X-Admin-API-Key since cycles-server v0.1.25.45 and fails closed with 500 when the server has no admin key configured. The old assertion had failed the org nightly Full-Stack Integration every night since 2026-06-28. No library code change.),
2026-05-22 (v0.4.3 — expires_from/expires_to and finalized_from/finalized_to ISO-8601 window-filter passthrough on list_reservations per cycles-protocol-v0.yaml revision 2026-05-22; closes the Python-client side of runcycles/cycles-server#162. No code change — **query_params already forwards arbitrary kwargs. Added sync + async regression tests; unlike from/to the new param names are plain kwargs (no Python-reserved-word workaround needed). 393 tests pass at 100% coverage.),
2026-05-21 (v0.4.2 — from / to ISO-8601 window-filter passthrough on list_reservations per cycles-protocol-v0.yaml revision 2026-05-21; closes the client side of runcycles/cycles-server#159. No code change — the existing **query_params signature already forwards arbitrary kwargs to the URL query string. Added sync + async regression tests that lock the passthrough in (using the **{"from": ..., "to": ...} dict-unpack form because from is a Python reserved keyword). 391 tests pass at 100% coverage.),
2026-03-14
Spec: cycles-protocol-v0.yaml (OpenAPI 3.1.0, v0.1.25)
Client: runcycles (Python 3.10+ / httpx / Pydantic v2)
Server audit: See cycles-server/AUDIT.md (all passing)
| Category | Pass | Issues |
|---|---|---|
| Endpoints & HTTP Methods | 9/9 | 0 |
| Request Schemas (field names & JSON keys) | 6/6 | 0 |
| Response Schemas (field names & JSON keys) | 10/10 | 0 |
| Enum Values | 5/5 | 0 |
| Nested Object Schemas | 8/8 | 0 |
| Auth Header (X-Cycles-API-Key) | — | 0 |
| Idempotency (header ↔ body sync) | — | 0 |
| Subject Validation | — | 0 |
| Response Header Capture | — | 0 |
| Client-Side Spec Constraint Validation | — | 0 |
| Lifecycle Orchestration | — | 0 |
Overall: Client is protocol-conformant. All endpoints, schemas, field names, JSON keys, and enum values match the OpenAPI spec. No open issues.
Compared the following across spec YAML and client Python source:
- All 9 endpoint paths, HTTP methods, and path/query parameters
- All 6 request body serializations vs spec schemas
- All 10 response model deserializations vs spec schemas
- All 5 enum types and their values
- Nested object schemas (Subject, Action, Amount, SignedAmount, Caps, CyclesMetrics, Balance, ErrorResponse)
- Auth and idempotency header handling
- Subject constraint validation (
anyOf/ at least one standard field) - Pydantic Field constraints vs spec min/max bounds
- Lifecycle orchestration (reserve → execute → commit/release)
| Spec Endpoint | Client Method | HTTP Method | Match |
|---|---|---|---|
/v1/decide |
client.decide() |
POST | PASS |
/v1/reservations (create) |
client.create_reservation() |
POST | PASS |
/v1/reservations (list) |
client.list_reservations() |
GET | PASS |
/v1/reservations/{reservation_id} |
client.get_reservation() |
GET | PASS |
/v1/reservations/{reservation_id}/commit |
client.commit_reservation() |
POST | PASS |
/v1/reservations/{reservation_id}/release |
client.release_reservation() |
POST | PASS |
/v1/reservations/{reservation_id}/extend |
client.extend_reservation() |
POST | PASS |
/v1/balances |
client.get_balances() |
GET | PASS |
/v1/events |
client.create_event() |
POST | PASS |
ReservationCreateRequest — spec required: [idempotency_key, subject, action, estimate]
- Pydantic fields:
idempotency_key,subject,action,estimate,ttl_ms,grace_period_ms,overage_policy,dry_run,metadata— all snake_case, all match spec
CommitRequest — spec required: [idempotency_key, actual]
- Pydantic fields:
idempotency_key,actual,metrics,metadata— all match spec
ReleaseRequest — spec required: [idempotency_key]
- Pydantic fields:
idempotency_key,reason— all match spec
DecisionRequest — spec required: [idempotency_key, subject, action, estimate]
- Pydantic fields:
idempotency_key,subject,action,estimate,metadata— all match spec
EventCreateRequest — spec required: [idempotency_key, subject, action, actual]
- Pydantic fields:
idempotency_key,subject,action,actual,overage_policy,metrics,client_time_ms,metadata— all match spec
ReservationExtendRequest — spec required: [idempotency_key, extend_by_ms]
- Pydantic fields:
idempotency_key,extend_by_ms,metadata— all match spec
| Spec Schema | Client Class | JSON Keys | Match |
|---|---|---|---|
ReservationCreateResponse |
ReservationCreateResponse |
decision, reservation_id, affected_scopes, expires_at_ms, scope_path, reserved, caps, reason_code, retry_after_ms, balances |
PASS |
CommitResponse |
CommitResponse |
status, charged, released, balances |
PASS |
ReleaseResponse |
ReleaseResponse |
status, released, balances |
PASS |
DecisionResponse |
DecisionResponse |
decision, caps, reason_code, retry_after_ms, affected_scopes |
PASS |
EventCreateResponse |
EventCreateResponse |
status, event_id, balances |
PASS |
ReservationExtendResponse |
ReservationExtendResponse |
status, expires_at_ms, balances |
PASS |
BalanceResponse |
BalanceResponse |
balances, has_more, next_cursor |
PASS |
ReservationDetail |
ReservationDetail |
reservation_id, status, idempotency_key, subject, action, reserved, committed, created_at_ms, expires_at_ms, finalized_at_ms, scope_path, affected_scopes, metadata |
PASS |
ReservationSummary |
ReservationSummary |
reservation_id, status, idempotency_key, subject, action, reserved, created_at_ms, expires_at_ms, scope_path, affected_scopes |
PASS |
ReservationListResponse |
ReservationListResponse |
reservations, has_more, next_cursor |
PASS |
| Spec Schema | Client Class | JSON Keys | Match |
|---|---|---|---|
Subject |
Subject |
tenant, workspace, app, workflow, agent, toolset, dimensions |
PASS |
Action |
Action |
kind, name, tags |
PASS |
Amount |
Amount |
unit, amount |
PASS |
SignedAmount |
SignedAmount |
unit, amount |
PASS |
Caps |
Caps |
max_tokens, max_steps_remaining, tool_allowlist, tool_denylist, cooldown_ms |
PASS |
StandardMetrics |
CyclesMetrics |
tokens_input, tokens_output, latency_ms, model_version, custom |
PASS |
Balance |
Balance |
scope, scope_path, remaining, reserved, spent, allocated, debt, overdraft_limit, is_over_limit |
PASS |
ErrorResponse |
ErrorResponse |
error, message, request_id, details |
PASS |
| Spec Enum | Client Enum | Values | Match |
|---|---|---|---|
DecisionEnum |
Decision |
ALLOW, ALLOW_WITH_CAPS, DENY |
PASS |
UnitEnum |
Unit |
USD_MICROCENTS, TOKENS, CREDITS, RISK_POINTS |
PASS |
CommitOveragePolicy |
CommitOveragePolicy |
REJECT, ALLOW_IF_AVAILABLE, ALLOW_WITH_OVERDRAFT |
PASS |
ReservationStatus |
ReservationStatus |
ACTIVE, COMMITTED, RELEASED, EXPIRED |
PASS |
ErrorCode |
ErrorCode |
All 12 spec values + UNKNOWN (client fallback) |
PASS |
Note: Client ErrorCode adds UNKNOWN as a fallback for unrecognized server error codes. This is a client-side convenience and does not violate the spec.
- X-Cycles-API-Key: Set on all requests via
httpx.Clientbase headers inCyclesClient.__init__()(client.py) - X-Idempotency-Key: Extracted from request body
idempotency_keyfield via_extract_idempotency_key()and set as header in_post(). Header and body values always match (copied from body to header), satisfying the spec rule: "If X-Idempotency-Key header is present and body.idempotency_key is present, they MUST match."
validate_subject()in_validation.pycallsSubject.has_at_least_one_standard_field()which checks all 6 standard fields — matches specanyOfconstraint- Pydantic Field constraints enforce
maxLength: 128on all Subject fields andmaxLength: 256on dimension values
_extract_response_headers()inclient.pycapturesx-request-id,x-ratelimit-remaining,x-ratelimit-reset,x-cycles-tenant- Exposed via
CyclesResponseproperties:request_id,rate_limit_remaining,rate_limit_reset,cycles_tenant
All spec constraints are validated both via Pydantic Field validators (on typed request models) and via explicit validation functions (on dict-based lifecycle path):
validate_non_negative():Amount.amount >= 0(specminimum: 0)validate_ttl_ms(): 1000–86400000 (specminimum: 1000, maximum: 86400000)validate_grace_period_ms(): 0–60000 (specminimum: 0, maximum: 60000)validate_extend_by_ms(): 1–86400000 (specminimum: 1, maximum: 86400000)- Pydantic
Field(ge=1, le=86_400_000)onReservationExtendRequest.extend_by_ms - Pydantic
Field(max_length=64)onAction.kind,Field(max_length=256)onAction.name - Pydantic
Field(min_length=1, max_length=256)on allidempotency_keyfields
- Reserve → Execute → Commit flow with proper cleanup (release on failure)
- Heartbeat-based TTL extension at
max(ttl_ms / 2, 1000)ms interval usingextendendpoint - Commit retry engine for transient failures (transport errors, 5xx) with exponential backoff
- Dry-run handling returns
DryRunResultwithout executing guarded function DENYdecision correctly raises typedCyclesProtocolErrorALLOW_WITH_CAPScorrectly propagatesCapsviaCyclesContext- Lifecycle instance cached at decoration time (deferred client resolution on first call)
ContextVar-based context propagation (safe for both sync threads and async tasks)
is_successcorrectly handles 2xx range (200 for most endpoints, 201 for events)- Error responses parsed via
ErrorResponse.model_validate()withErrorCodemapping - Typed exceptions:
BudgetExceededError,OverdraftLimitExceededError,DebtOutstandingError,ReservationExpiredError,ReservationFinalizedError,TenantClosedError(added 2026-07-10, raised at reservation-creation time)
The client is fully protocol-conformant with the Cycles Protocol v0.1.23 OpenAPI spec. All 9 endpoints, 6 request schemas, 10 response schemas, 5 enum types, and all nested object serializations match the spec exactly. JSON field names use correct snake_case throughout. Auth headers, idempotency handling, subject validation, response header capture, and spec constraint validation all follow spec normative rules. No open issues.
Spec version: v0.1.24
Test file: tests/test_contract.py (34 tests, all passing)
Automated contract tests validate sample request/response payloads against the OpenAPI spec schemas using jsonschema.Draft202012Validator with recursive $ref resolution:
- Request schemas validated: DecisionRequest, ReservationCreateRequest, CommitRequest, ReleaseRequest, EventCreateRequest
- Response schemas validated: DecisionResponse, ReservationCreateResponse, CommitResponse, ReleaseResponse, EventCreateResponse, ErrorResponse
- Negative tests: missing required fields, extra fields (additionalProperties), invalid enum values
- Enum value tests: UnitEnum, ErrorCode, DecisionEnum, ReservationStatus, CommitOveragePolicy
- Spec fixture:
tests/fixtures/cycles-protocol-v0.yaml(copy of canonical spec)
Module: runcycles/streaming.py
Test file: tests/test_streaming.py (64 tests, all passing)
Version: 0.3.0
Added StreamReservation and AsyncStreamReservation context managers that automate the reserve → commit/release lifecycle for streaming use cases. This is a DX convenience layer — no protocol changes.
StreamReservation— sync context manager: reserves on__enter__, auto-commits on successful__exit__, auto-releases on exceptionAsyncStreamReservation— async equivalent using__aenter__/__aexit__StreamUsage— mutable accumulator for token counts and cost during streaming- Client convenience methods:
CyclesClient.stream_reservation()andAsyncCyclesClient.stream_reservation()— thin factories that build Subject from config defaults - Cost resolution: explicit
usage.actual_cost>cost_fn(usage)> estimate fallback - Heartbeat: automatic TTL extension, same interval formula as decorator lifecycle (
max(ttl_ms / 2, 1000)ms) - Commit retry: uses existing
CommitRetryEngine/AsyncCommitRetryEngine - Context propagation: sets/clears
CyclesContextviaContextVar, accessible viaget_cycles_context(); respects user-setctx.metricsduring streaming - Spec validation:
validate_ttl_ms()(1000–86400000),validate_grace_period_ms()(0–60000),validate_subject()(at least one standard field) — matches lifecycle.py - Error handling:
RESERVATION_FINALIZED,RESERVATION_EXPIRED, andIDEMPOTENCY_MISMATCHdo not trigger release; other 4xx client errors do trigger release — matches lifecycle.py behavior exactly
Protocol conformance: No new endpoints or protocol changes. All reservation, commit, release, and extend calls use the same client methods and body formats as the decorator path. Verified by 64 unit tests covering success, deny, error, retry, heartbeat, cost resolution, context propagation, spec validation, and all commit error-code branches.
Issue: #45
Files: runcycles/lifecycle.py, runcycles/decorator.py
Test files: tests/test_lifecycle.py, tests/test_decorator.py
Version: 0.4.0
Widened the @cycles decorator to accept callables — in addition to constants — for every field that previously had to be static at decoration time. Mirrors the existing estimate / actual callable contract and re-aligns the Python client with the Java client's @Cycles(workspace = "#workspaceId") SpEL behavior shipped in cycles-spring-boot-starter 0.2.1 (java#50).
- Newly callable fields:
tenant,workspace,app,workflow,agent,toolset,action_kind,action_name,action_tags,dimensions. Each acceptsT | Callable[..., T | None] | None. - Resolution: new
_resolve_value(val, args, kwargs)helper inlifecycle.pyinvokes the callable with the decorated function's*args, **kwargsat reservation time; constants pass through untouched. - Fallback semantics preserved: subject callables returning
Nonefall through todefault_subject_fields(client config);action_kind/action_namereturningNonefall through to"unknown";action_tags/dimensionsreturningNoneare omitted. Constants behave identically to today (regression-tested). - Fail-fast: exceptions raised inside a user callable propagate to the decorator caller without creating a reservation.
- Signature change:
_build_reservation_bodynow takesargsandkwargsparameters; bothCyclesLifecycle.executeandAsyncCyclesLifecycle.executethread them through.
Protocol conformance: No protocol or wire-format changes. The reservation request body shape is unchanged — only the source of each field's value is widened. Verified by new unit tests in TestCallableSubjectFields, TestCallableActionFields, TestCallableDimensions plus an end-to-end decorator test asserting the captured request body.
Files: pyproject.toml
Version: 0.4.1
Metadata-only release retargeting the package for category-search discovery on PyPI. No code, no test, no protocol changes — wire format and API are identical to 0.4.0.
- Description rewritten to lead with the literal category-search phrase:
"Python AI agent budget control — enforce LLM cost limits, tool permissions, and multi-tenant policies before agent actions execute." - Keywords expanded 12 → 21, organized into category-search terms (
ai-agent,agent-budget,budget-control,cost-control,cost-enforcement,spending-limit,llm-cost,runtime-authority,action-control,multi-tenant), framework targeting (langchain,langgraph,crewai,autogen,openai-agents,mcp,openai,anthropic), and brand (cycles,runcycles). - Classifier added:
Topic :: Scientific/Engineering :: Artificial Intelligence.
Driven by Python-side adoption diagnostic finding the biggest sub-gap was discovery, not SDK feature parity. Companion changes: GitHub topics on this repo (governance dropped, mcp added) and Python framework integration guide retitling on runcycles/cycles-docs (PR #568).
Protocol conformance: No protocol or wire-format changes. Existing test suite at 100% coverage; no test additions.
Files: README.md, examples/langchain_integration.py
Version: unreleased (next 0.4.x — docs/examples only)
Documentation-only update pointing users at the new sibling package langchain-runcycles for LangChain agent middleware integration. No SDK code changes; no protocol changes.
- README.md: Added a new
## Integrationssection listinglangchain-runcycles(PyPI:langchain-runcycles) as the canonical path forlangchain.agents.create_agentworkflows. The existingexamples/langchain_integration.pyrow is reframed as the right fit for non-agent LangChain runnables (bareChatOpenAI, chains, RAG); middleware requirescreate_agentso the two patterns serve different surfaces and both remain supported. - examples/langchain_integration.py: Updated the file-level docstring to point at
langchain-runcyclesfor agent workflows while preserving the callback-handler example as-is. No code changes.
Background: LangChain 1.x introduced an AgentMiddleware API with wrap_tool_call, before_model, and wrap_model_call hooks. The new package wraps that API on top of this SDK's existing decide / create_reservation / commit_reservation / release_reservation surface — no new SDK methods needed. Splitting into a sibling repo follows LangChain's publishing guidance ("New integrations should be published as standalone PyPI packages") and the langchain-<service> naming convention used by langchain-anthropic, langchain-openai, etc.
Protocol conformance: No protocol or wire-format changes. The new sibling package consumes this SDK as a normal dependency.
Files: .claude/session-start-global-deny.sh, .github/workflows/python-publish.yml
Version: unreleased (CI/Claude-config only — no package version change)
Cross-cutting hardening landed in response to org-wide tracking issues filed in runcycles/.github. Two distinct changes; both are infra-only.
-
.claude/session-start-global-deny.shsynced from the new canonical atruncycles/.github/shared-config/. The script now (a) carries a top-of-file callout explaining that Part 2 mutates theoriginremote of every sibling repo under/home/user/*, not just the current checkout, and (b) honors aCYCLES_CLAUDE_SKIP_REMOTE_REWRITE=1opt-out env var. Part 1 (MCP deny rules) is unchanged. Tracksruncycles/.github#63. -
.github/workflows/python-publish.ymlgained aVerify pyproject version matches tagstep that runs on tag-triggered builds (refs/tags/v*). The step parsespyproject.tomlviatomlliband fails the workflow before the build phase if the declared version doesn't match the tag (e.g., tagv0.5.0againstpyproject.tomlstill on0.4.1or adev0pre-release). PyPI already rejects duplicate versions server-side, but this surfaces operator error earlier in the pipeline. Python analog of the Java SNAPSHOT-guard tracked inruncycles/.github#61.
Not included in this change: bumping the reusable-workflow ref runcycles/.github/.github/workflows/ci-python.yml@main to @v1 (runcycles/.github#60). That bump is intentionally split into a separate follow-up PR — it depends on the v1 tag existing in runcycles/.github, which is being cut after the canonical-script PR (runcycles/.github#64) merges.
Protocol conformance: No protocol or wire-format changes. No SDK source touched. Test suite unaffected.
Files: README.md
Version: unreleased (docs-only, no version bump, no CHANGELOG entry per repo convention)
Documentation-only correction. The README's exception-hierarchy table described CyclesTransportError as "Network-level failure (connection, DNS, timeout)", implying the SDK raises it — but nothing in the package ever raises it. Actual behavior:
@cyclesdecorator / HOF paths: a transport failure at reserve time raisesCyclesProtocolErrorwithstatus == -1anderror_code=None(via_build_protocol_exceptioninlifecycle.py); commit-time transport failures are retried in the background by the commit retry engine, not raised.- Programmatic client: never raises for transport failures — returns
CyclesResponsewithis_transport_error == Trueandstatus == -1(CyclesResponse.transport_errorconstructor inresponse.py).
The table row now states the class is exported for user code but never raised by the SDK, and a new "Transport errors" subsection documents the status == -1 behavior for both API surfaces with a detection example. Wording matches the docs site (cycles-docs/how-to/error-handling-patterns-in-python.md). CyclesTransportError remains exported from runcycles/__init__.py for use in user code — no API change.
Protocol conformance: No protocol or wire-format changes. No SDK source touched. Test suite unaffected.
Files: runcycles/models.py, runcycles/exceptions.py, runcycles/lifecycle.py, runcycles/__init__.py, README.md, CHANGELOG.md, tests
Version: unreleased
Additive support for the TENANT_CLOSED protocol error code, per runtime spec v0.1.25.13 of cycles-protocol-v0.yaml (runcycles/cycles-protocol#125). Servers return HTTP 409 error=TENANT_CLOSED on reservation create/commit/release/extend when the owning tenant's status is CLOSED — the runtime-surface mirror of governance spec Rule 2 (mutating operations on objects owned by a CLOSED tenant are rejected with 409 TENANT_CLOSED).
ErrorCode.TENANT_CLOSEDadded to the enum inmodels.py, in spec declaration order:… MAX_EXTENSIONS_EXCEEDED, LIMIT_EXCEEDED, TENANT_CLOSED, INTERNAL_ERROR(initially placed afterBUDGET_CLOSED; relocated whenLIMIT_EXCEEDEDlanded so the enum mirrors the spec exactly).TenantClosedError(CyclesProtocolError)added toexceptions.pyfollowing the existing per-code subclass pattern; exported fromruncycles/__init__.py._build_protocol_exceptioninlifecycle.pymapserror == "TENANT_CLOSED"toTenantClosedError. This mapping is invoked on the reservation-creation paths of the@cyclesdecorator,CyclesLifecycle, and streaming surfaces; commit/release-time error codes are handled by the existing commit-failure policy (logged + released) and are not raised as typed exceptions.CyclesProtocolError.is_tenant_closed()helper added alongside the existingis_*per-code helpers.- README exception table gained a
TenantClosedErrorrow.
Forward-compat behavior before this change (verified): an unrecognized TENANT_CLOSED string was mapped to ErrorCode.UNKNOWN by ErrorCode.from_string, so lifecycle surfaces raised plain CyclesProtocolError with error_code == "UNKNOWN" — which is_retryable() reports as retryable. Deserialization never failed. With this change the code is recognized, typed, and correctly non-retryable.
The vendored spec fixture (tests/fixtures/cycles-protocol-v0.yaml, pinned at v0.1.24) is intentionally untouched; it will be refreshed when the v0.1.25.13 spec PR merges. The contract suite validates the fixture, not the client enum, so the client enum being a superset is by design.
Tests: new coverage for the enum member (test_models.py), the exception subclass + helper (test_exceptions.py), and the lifecycle mapping (test_lifecycle.py). 396 tests pass at 100% coverage (gate ≥95%); ruff + mypy --strict clean.
Files: runcycles/models.py, runcycles/exceptions.py, CHANGELOG.md, tests
Version: unreleased (same PR as the TENANT_CLOSED entry above)
Additive support for the LIMIT_EXCEEDED protocol error code, added to the runtime ErrorCode enum in spec revision v0.1.25.12 (2026-07-04). Per the spec's ERROR SEMANTICS, HTTP 429 is reserved for server-side throttling/rate limiting (declared on the public evidence/JWKS endpoints); 429 responses carry error=LIMIT_EXCEEDED plus the Retry-After and X-RateLimit-Reset headers, and clients retry after the indicated delay.
ErrorCode.LIMIT_EXCEEDEDadded in spec declaration order (afterMAX_EXTENSIONS_EXCEEDED);TENANT_CLOSEDrelocated after it so the client enum mirrors the spec order exactly.- Retryability decision: retryable at both layers — added to
ErrorCode.is_retryableand to the code tuple inCyclesProtocolError.is_retryable(). Rationale: a 429 rate limit is transient by definition and the spec instructs retry; this also preserves prior behavior, where the unrecognized string fell back toErrorCode.UNKNOWN(retryable). The repo's status-based rule (status >= 500) does not cover 429, so the code-based classification carries it. - Shape decision: enum-only — no
LimitExceededErrorsubclass, no lifecycle mapping. This matches the sibling pattern (BUDGET_FROZEN/BUDGET_CLOSEDare enum-only): LIMIT_EXCEEDED is not a reservation-lifecycle denial, so a typed exception class is not warranted.
Forward-compat behavior before this change (verified): ErrorCode.from_string("LIMIT_EXCEEDED") returned ErrorCode.UNKNOWN — retryable by accident. Now typed and retryable by design; no semantic change.
Retry-After exposure (codex round-3): the client previously dropped the HTTP Retry-After header, so the spec's "retry after the indicated delay" was not SDK-visible for header-carried 429s. retry-after is now captured in _RESPONSE_HEADERS (client.py), exposed as CyclesResponse.retry_after_ms_header (seconds → ms, non-integer forms ignored), and _build_protocol_exception falls back to it for retry_after_ms when the body field is absent (body wins when both are present). No auto-retry behavior change — no internal path consumes code-level retryability; the delay is surfaced only.
Tests: enum member + retryable (test_models.py), exception-layer retryable with retry_after_ms (test_exceptions.py), Retry-After header conversion + precedence (test_response.py, test_lifecycle.py), and an end-to-end 429 LIMIT_EXCEEDED with a real Retry-After header through the client (test_client.py). Full suite green at 100% coverage.