Skip to content

Release v0.2.0

Choose a tag to compare

@homer6 homer6 released this 25 Oct 21:08
· 28 commits to main since this release
db8a750

agent-creator v0.2.0

Released 2026-10-25. A minor release (Steve named the version on 2026-10-25). One private image,
docker.io/oramasearch/agent-creator:v0.2.0, built on
agent-runtime v0.2.0. Release
tracking: agent-creator#16.

Since v0.1.0, the Agent Creator Agent can read the methodology registry again, and its
tools are limited to exactly the reads a turn makes. Underneath, the agent-runtime it runs on carries
the 2026-10-21 security-review fixes. Everything v0.1.0 lists is otherwise unchanged: prompt v5, the
rule checks in defects, the turn answer, and how Lake reaches it.

What it includes

Every pull request merged into dev since the tag v0.1.0 (git log --first-parent v0.1.0..origin/dev):

PR What it does
#15 The methodology registry is readable (agent-creator#11). definition/record.json grants search, summarize and pipelines; in agent-runtime only train and pipelines grant list_methodologies, and train would also give direct check_pipeline and start_run. definition/agent.toml now has a tools allow-list of exactly the twelve reads a turn makes (list_solutions, get_solution, list_apps, get_app, receive_app_events, list_builds, get_build, receive_build_events, list_methodologies, list_datasets, list_projects, get_project), so none of pipelines' proposals, and no other tool, is one of its tools. New tests/test_grants.py runs the turn's reads against a workspace that answers only what agent-runtime would make a tool for this record (from a vendored copy of agent-runtime's skills.json)
#14 Records the v0.1.0 image digest in the v0.1.0 notes
#13 The v0.1.0 release pull request (dev promoted to main), carried back into dev

What this changes for a person: methodology-bound proposals and declines pass their checks again,
and approved Pipelines that name a runnable methodology are offered for grants instead of being
listed as not runnable.

From agent-runtime v0.2.0, underneath

The image is FROM oramasearch/agent-runtime:${RUNTIME_TAG} built with RUNTIME_TAG=v0.2.0, so it
carries every change in agent-runtime since its v0.1.0:

  • agent-runtime#19: POST /v1/turn compares
    the service token in constant time, and refuses every turn with 503 when
    AGENT_SERVICE_TOKEN is unset or empty
    (it used to be open). A wrong or missing bearer is 401.
  • agent-runtime#21: a turn's credential
    goes only to Lake at this service's LAKE_URL.
    The turn's own lake_url is no longer read; with
    LAKE_URL unset or not an absolute http(s) address without user info, the turn is refused before
    any request goes out.
  • agent-runtime#26: dataset-confined agents
    are limited to qualified dataset reads (LSR-181). AgentCreator's work area is workspace, so its
    own turns are unaffected; an agent it proposes with work area dataset gets these limits (see
    Known issues).

Settings it needs

No new setting of its own. agent-runtime v0.2.0 now requires AGENT_SERVICE_TOKEN and
LAKE_URL (and reads AGENT_MODEL when the record names no model; the record names gpt-5.5). The
Orama operator already renders AGENT_SERVICE_TOKEN, LAKE_URL and AGENT_MODEL for every library
agent, so no operator change has to roll first. A hand-run service without AGENT_SERVICE_TOKEN now
refuses every turn.

What it does not include

Unchanged from v0.1.0:

  • It creates nothing. Creating the approved agent is lake's (lake#233).
  • It does not change existing agents (lake#271).
  • It proposes only what agent-runtime's twelve skills can do.
  • It checks that Python compiles, not that it does the right thing.
  • No live eval of the current prompt. The only recorded eval is prompt v1 (15/16).
  • Conversation memory is per pod and in memory; a restart loses it.

Known issues

Security-relevant first.

  1. Workspace identity is not passed explicitly on every call
    (agent-creator#9, with
    agent-runtime#6): App and build reads
    carry only customer_id, and the idle bus poll runs every 10 seconds. Related, in lake:
    lake#918 (LSR-009, agent launch and schedule
    credentials escaping their workspace and inner action scopes).
  2. The runtime underneath keeps unbounded conversation histories and claims bus tasks before a
    worker is free
    (agent-runtime#25,
    LSR-076; agent-runtime#24, LSR-075).
  3. Lake's shared agent service token is accepted by every library agent, and a user's service
    launch holds it (agent-runtime#14, on
    hold).
  4. The runtime still writes a trace per turn under TRACE_DIR in the pod.
  5. Dataset-confined agents lose some reads on agent-runtime v0.2.0
    (agent-runtime#26): index search, App and
    build reads, Pipeline and MCP tools are refused or omitted in dataset work areas. An agent
    AgentCreator proposes with work area dataset and a skill that needs those reads will not get them.
  6. The structured answer is shipped, the issue is still open
    (agent-creator#1), for lake's side to read
    result.
  7. The skills fixture is vendored (tests/fixtures/agent-runtime-skills.json, from agent-runtime
    cf964f4). When agent-runtime's skills change, refresh it as docs/methodologies/README.md says.

Where it has run

The v0.2.0 tree has not run on the dev environment. No dev-<sha> build of agent-creator ran on
the dev member of lake-nonprod; every Lake still takes the operator's default
oramasearch/agent-creator:v0.1.0 (orama-operator internal/controller/lake_defaults.go on dev).
Moving it to :v0.2.0 is a separate step, through testing (the e2e gate), staging and prod.

The image

Built from the tag v0.2.0 (db8a750) on the cluster BuildKit with build-agent.sh (linux/amd64, the repository's Dockerfile, built on oramasearch/agent-runtime:v0.2.0 with RUNTIME_TAG=v0.2.0), then tagged :v0.2.0 on that digest without rebuilding:

docker.io/oramasearch/agent-creator:db8a750-0@sha256:441eec21b709c06aa5c4d0f6d3d6fc125f54d42609275408be502ed3e544f644, also tagged :v0.2.0 (same digest). The Docker Hub repository answers 404 anonymously (private).

It is not rolled anywhere yet: pinning it on an environment is a separate change to the operator's set files, through dev, testing, staging and prod in order.

Verification

Proof Result
Test suite (uv run --frozen --group dev pytest -q -ra, macOS arm64, Python 3.14 and again on 3.12) 29 passed, 0 failed, 0 skipped on each: tests/test_creator.py (25), tests/test_grants.py (3), tests/test_release.py (1, reads the version from pyproject.toml and requires these notes and their index line)
Offline evals None: every eval case runs against a live workspace
Live evals (evals/run.py) Not run (needs a live model and a live workspace)
Linux image build (cluster BuildKit, from the tag) Built from v0.2.0 with build-agent.sh: oramasearch/agent-creator:db8a750-0, also :v0.2.0, sha256:441eec21b709c06aa5c4d0f6d3d6fc125f54d42609275408be502ed3e544f644; private (404 anonymously)