v0.8.0

Fixed

• undo apply replays against the target the original write ran on. It dispatched the inverse against whatever target the caller named — in practice the config's first entry — while the write's own target sat unused in the undo record. On a multi-target config the inverse therefore ran against the wrong host; it only looks harmless because the resource usually is not there, but two hosts holding the same name and the inverse succeeds on the wrong one, silently. An explicitly named target still wins. Line-wide: all 24 copies had the identical defect. Caught live in container-host-aiops, where a stop recorded against a Podman target replayed against a Portainer one.

v0.7.0

Changed (BREAKING)

• Requires MCP SDK 2.0 (mcp[cli]>=2.0,<3.0). mcp.server.fastmcp no longer exists in 2.0; the server is now built with MCPServer and reports its package version in the stdio handshake.
• model_deploy no longer takes num_replicas. Ray derives replica counts from the application's own config, and the deployment names needed to set them are only known once the app materialises — use serve scale afterwards.
• drain_replica, restart_replica and update_routing_policy now refuse with a teaching error instead of calling endpoints that do not exist. Ray Serve's REST control plane has no per-replica or routing-policy surface; per-replica draining is a Python-API capability only. Scaling a deployment down lets the controller drain surplus replicas gracefully.

Fixed

• undo apply works from the CLI. Every write tool is imported lazily inside its own CLI command, so a CLI-driven undo ran in a process where the inverse tool was never registered and failed with "inverse tool is not registered" — for every write tool. Only the MCP entry point, which imports the whole server, worked. Found while live-verifying against a real cluster.
• An undetermined outcome is audited unknown, not ok. The harness only classified a result as undetermined when the payload also carried an error key, so a write that looked successful but had not been confirmed was recorded as a success.
• The entire Ray Serve control plane called endpoints Ray has never had. Live-verified against Ray 2.56.1: the REST control plane is declarative — the only mutating endpoint is PUT /api/serve/applications/ with the whole ServeDeploySchema. The per-deployment and per-replica paths this tool used 404 on every real cluster, so scale, scale-to-zero, autoscale, deploy, undeploy, redeploy, drain, restart and routing could never have worked. All rebuilt on the real API: fetch every app config, patch the target, PUT the whole set.
• A single-app change is now refused when another app was deployed imperatively (serve.run) rather than from a Serve config, because the whole-schema PUT would silently delete that bystander.
• get_cluster_resources read fields that no longer exist and returned all-null on every live cluster; it now reads clusterStatus.loadMetricsReport.usageByNode. A resource no node reports yields null, not a fabricated 0.
• get_gpu_utilization used /api/nodes, which 404s; the route is /nodes.

v0.6.0

Release notes — inference-aiops 0.6.0

Previous release: 0.5.0.

BREAKING — the authorization layer is removed

This tool no longer decides whether a write is permitted. Read-only mode
(<PREFIX>_READ_ONLY), the graduated-approval / approver gate, and the
rules.yaml deny engine are all gone. Whether an operation runs is the
agent's judgement, or the permission of the account you connect it with — point
it at a read-only credential and the write fails at the server, the place that
actually owns the permission.

What the tool guarantees instead is that nothing is silent: every operation,
over MCP and the CLI alike, lands a row in the audit log — there is no
unaudited entry point. Destructive writes still capture their before-state and
record an undo token where a clean inverse exists.

• If you set <PREFIX>_READ_ONLY=1, it now has no effect and the MCP server
  logs a warning at startup. Restrict writes via the connecting account instead.
• <PREFIX>_AUDIT_APPROVED_BY / <PREFIX>_AUDIT_RATIONALE still work, but are now
  optional audit annotations — recorded on the row when set, never required.
• The declared risk_level is carried into the audit row as a descriptive tier
  (a label, not a gate).

The governance harness is now: audit (MCP+CLI, unbypassable) · runaway/budget
safety guard · undo recording · output sanitize. policy.py is a small
risk-tier classifier; governance/readonly.py is deleted.

v0.5.0

Release notes — inference-aiops 0.5.0

Previous release: 0.4.0.

In this tool

• BREAKING: hot_swap_model is removed. It POSTed to /v1/hot_swap, an endpoint vLLM has never had. The docstring named vLLM Sleep Mode, so the path appears to have been invented for it — and Sleep Mode suspends the same model rather than swapping base models, so the advertised capability does not exist on the platform at all. Serving a different base model means restarting vLLM with a different --model.
• Replaced by the real thing: model_sleep, model_wake and model_is_sleeping, on vLLM's actual /sleep, /wake_up and /is_sleeping. These require the server to have been started with VLLM_SERVER_DEV_MODE=1; a 404 now says exactly that instead of reading as a generic failure.

Every tool in the line: previews and undetermined outcomes

This release fixes three harness defects that were silently degrading the audit
trail and the undo store.

A write that loses its response is no longer recorded as a failure. The
harness assumed a sanitized error meant nothing had happened. That assumption is
false in exactly the case that matters most: when a write severs its own
connection, the request has already landed, the response cannot come back, and
the operation was recorded as status=error with no undo token created at
all. Transport-level failures are now audited as status=unknown, the result
says plainly that the operation may have taken effect and should be verified
before retrying, and a write that stashed its before-state has its inverse
recorded anyway — flagged effectVerified: false, which undo_list and
undo_apply both surface. Existing undo.db files are migrated in place; their
rows read as verified, which is accurate, since the old code only ever recorded
on the confirmed path.

A dry-run no longer writes an undo token. Previews were recording inverses
built from a before-state they never had: the undo callback's permissive default
filled the gap with a guess, producing a real, applicable token for an operation
that never happened.

A dry-run no longer demands a named approver. Requiring an approval in order
to ask whether something needs approval inverts what a preview is for. The tier
is still computed and still audited, so the preview can tell you an approver
will be needed; it just no longer refuses to answer. The write itself is gated
exactly as before.

The invariant, now stated: a dry_run may read; it must never write. Guards
run on the preview path, which means a preview can and does report that an
operation would be refused.

Also line-wide

• Truncated text now ends in an ellipsis instead of being cut silently. This
  line already treats a silent cut as a defect for lists; it was doing exactly
  that to strings.
• Error messages are capped at 800 characters, not 300. These messages end
  with what to do instead, so the cap was removing the most useful sentence of
  every long refusal.

Verification status

That /v1/hot_swap does not exist is certain — no vLLM has ever served it. The
replacement endpoints are modelled from vLLM's Sleep Mode documentation and
are not live-verified: this repo cannot be exercised on the maintainer's
hardware (the Ray image is amd64-only and dies under emulation). sleep,
wake and is_sleeping are marked accordingly in docs/VERIFICATION.md.

v0.4.0

Release notes — inference-aiops 0.4.0

Previous release: 0.3.0.

Headline: read-only mode

export INFERENCE_READ_ONLY=1

With this set the 15 write tools are never registered — an MCP
client lists 22 tools instead of 37. The writes are not hidden
behind a flag and not merely refused on call: they are absent from the session,
so a model cannot invoke one and cannot be argued into one. For a reviewer this
is checkable rather than promised — connect, list the tools, and the writes are
not there.

Enforcement is two layers deep: the @governed_tool harness refuses every
non-read operation (covering the CLI and in-process callers too), and the MCP
server removes write tools from list_tools(). Changing entry point does not
get around it.

BREAKING — return shapes changed

This release changes payloads that callers may be parsing. Both changes exist
to stop a result from misrepresenting itself:

1. Absent fields are now null, not "". A missing value and an empty value
   were previously indistinguishable, which invited consumers to invent the
   difference. Keys are still always present — only the value may be null.
2. Anything with a limit now returns an envelope —
   {"<items>": [...], "returned": N, "limit": L, "truncated": bool}. Truncation is
   measured (one extra row is fetched), never inferred from the page happening to
   be full. Where a genuine pre-cap total is knowable it is reported as total;
   where it isn't, total is deliberately omitted rather than echoing returned.

Also in this release

• docs/VERIFICATION.md — what the mock suite actually guarantees, a live
  verification checklist, and the criteria for claiming this tool verified.
• skills/inference-aiops/references/agent-guardrails.md — for driving this tool with a
  smaller / local model: which guardrails are now enforced for you, and a
  ready-made system prompt for the rest.
• Expanded operator playbooks in the skill documentation.
• The advertised tool count now matches what an MCP client actually lists
  (it includes undo_list / undo_apply), and a release gate keeps it honest.
• The (preview) label has been dropped. It never meant unreleased; verification
  status now lives in docs/VERIFICATION.md where it can be specific.

v0.3.0

Undo executor (undo list / undo apply) line-wide + this tool's platform/framework expansion + coverage. See CHANGELOG.md.

v0.2.1

Fix release: secrets.enc follows _AIOPS_HOME; sanitized failures audit as status=error (no undo recorded for failed calls); doctor/init fully test-covered. Tool-specific fixes in CHANGELOG.md.

v0.2.0

Security-hardening release from a line-wide code review. Highlights: secure-by-default approver gate for high/critical operations (init seeds a starter rules.yaml), percent-encoded URL path segments, single-sourced version, governance persistence + CLI write-path tests. See CHANGELOG.md for the full list. BREAKING: fresh installs with no rules.yaml now deny high/critical writes until an approver is recorded.

Inference-AIops v0.1.1

Patch: config.yaml now honors the *_AIOPS_HOME env var; CLI writes are now audited + undo-recorded through the governance path (previously MCP-only). See CHANGELOG.

inference-aiops v0.1.0

Governed GPU inference ops over vLLM + Ray Serve/Jobs — 30 MCP tools. Flagship diagnose_latency_spike correlates queue depth + KV pressure + prefix locality. Prometheus-native. Vendored governance harness; scale-down/scale-to-zero/drain/undeploy/hot-swap gated by dry-run + double-confirm. Preview / mock-only.