Repository navigation
Releases
Release v0.3.1
Released 2026-11-16. A patch release. The Dockerfile has one build stage again, pinned by digest, since v0.3.0. It fixes v0.3.0's known issue 3 (#78). Nothing the agent does has changed. Image: docker.io/oramasearch/agent-application:v0.3.1.
This release is the tree of dev after #82 (7a85b94), with these notes and the version.
What changed
Every pull request merged into dev since v0.3.0 (git log --first-parent v0.3.0..origin/dev):
| PR | What it does |
|---|---|
| #81 | Records the v0.3.0 image digest in the v0.3.0 notes; documentation only |
| #82 | One digest-pinned build stage: the unpinned stage a merge had left beside it is removed, so the venv is built on the pinned bases. Fixes #78 |
One build stage, pinned by digest — #82
In v0.3.0 the Dockerfile had two stages named build. The second, which builds the venv, named python:3.12-slim and ghcr.io/astral-sh/uv:0.11.7 by tag only, so the digest pins did not hold for it. Now there is one build stage, and every FROM and COPY --from names the multi-platform index digest of its tag. The bases are the same digests v0.3.0 pinned. The comments in the Dockerfile header no longer reference internal issues.
No source file under src/ changed, and no setting is new or removed (git diff v0.3.0..v0.3.1 -- src is empty).
Not included
- Claiming tasks by reference and naming the task on every call (#74): deferred for review, not in this release.
Known issues
Security-relevant first.
- Library agents hold live task credentials for every tenant they serve: one shared process claims tasks for every person, each with a credential that acts as that person. A design issue on Lake's side.
- The image bundles GPL and LGPL code (#77): it is built on
python:3.12-slim(Debian: glibc, bash, coreutils, dpkg, readline). The move toscratchwith static binaries is deferred; it is noted here and does not block this release. - Confirmation gates do not establish a separate human confirmer (on Lake's side, low).
ensure_repositorydoes not match and create repositories in the turn's workspace (#18): a repository with the same name in another workspace makes the deploy fail with a conflict.- No explicit workspace IDs on data-plane calls (#3), and no idle backoff on the bus.
- Generated Apps may still select whole tables (#51) where they should aggregate in SQL and page row-level reads.
- No evals for repair and change turns (#41).
v0.3.0's known issue 3, the two build stages (#78), is fixed by #82.
Where it has run
The v0.3.1 tree had not run on a development environment when it was released. Moving a Lake to :v0.3.1 is a separate change to the operator's defaults or the Lake's resource.
The image
Built from the tag v0.3.1 (a256372) with BuildKit (linux/amd64, the repository's Dockerfile), then tagged :v0.3.1 on that digest without rebuilding:
docker.io/oramasearch/agent-application:a256372@sha256:e87d1ec1df245e0b2f8bb23e6dd164fb0b18a46f0c24ef95a41303f03a1de3a4, also tagged :v0.3.1 (same digest). The image is published on Docker Hub; its GPL/LGPL base is a known issue (#77).
Pinning it on a Lake is a separate change to the operator's defaults or the Lake's resource.
Verification
| Proof | Result |
|---|---|
Test suite (uv run --frozen --group dev pytest -q -ra, macOS arm64, Python 3.12) |
102 passed, 0 failed, 0 skipped: tests/test_agent.py 23, tests/test_service_auth.py 13, tests/test_repair.py 10, tests/test_turn_destinations.py 10, tests/test_conversations.py 9, tests/test_generated_apps.py 9, tests/test_service.py 8, tests/test_bus_backpressure.py 6, tests/test_evals.py 4, tests/test_repository_file_bytes.py 4, tests/test_release.py 3, tests/test_lake_bus.py 2, tests/test_repository_paths.py 1 |
| Version | pyproject.toml and uv.lock 0.3.1; importlib.metadata.version("agent-application") reports it (tests/test_release.py), and GET /healthz returns that value |
SPDX headers (python3 tools/spdx-headers.py --check) |
Passed |
SBOM (python3 tools/sbom.py --version v0.3.1 --check) |
Passed, regenerated for v0.3.1 |
| Live evals (a model through Lake's gateway) | Not run for this release (needs a live model) |
| Dev environment | Not run (above) |
| Linux image build (BuildKit, from the tag) | Built from v0.3.1: oramasearch/agent-application:a256372, also :v0.3.1, sha256:e87d1ec1df245e0b2f8bb23e6dd164fb0b18a46f0c24ef95a41303f03a1de3a4 |
Release v0.3.0
Released 2026-11-16. A minor release. The agent's licence and open-source preparation, the bus consumer and conversation memory bounded, and the image's bases pinned by digest, since v0.2.7. The bounded bus and memory change behaviour a caller can see and add settings, which is why this is a minor release and not a patch. Image: docker.io/oramasearch/agent-application:v0.3.0.
This release is the tree of dev after #75 (c9cc363), with these notes and the version.
What changed
Every pull request merged into dev since v0.2.7 (git log --first-parent v0.2.7..origin/dev):
| PR | What it does |
|---|---|
| #72 | Adds a security policy, SECURITY.md: how to report a vulnerability privately, and what we commit to |
| #73 | Pins the Dockerfile's base images (python:3.12-slim, ghcr.io/astral-sh/uv:0.11.7) by digest as well as by tag |
| #70 | The bus consumer claims a task only when a worker is free, and one customer can no longer take every worker |
| #76 | The MIT licence: LICENSE, COPYRIGHT, an SPDX header in every source file, THIRD_PARTY_LICENSES, CONTRIBUTING.md, CLA.md, CODE_OF_CONDUCT.md, and the SBOM (sbom.spdx.json, sbom.cdx.json, tools/sbom.py) |
| #71 | Bounds the conversation memory, and keys a bus turn's conversation by the account and workspace Lake names |
| #75 | Prepares the tree for a public export: internal links, names and wording removed from the README, the docs and the release notes |
The licence — #76
agent-application is released under the MIT licence, copyright OramaSearch, Inc. (LICENSE, COPYRIGHT). Every source file carries SPDX-License-Identifier: MIT, added and checked by tools/spdx-headers.py. THIRD_PARTY_LICENSES lists the licence of every dependency. Contributions are made under CLA.md (contributors assign copyright to OramaSearch, Inc.), following CONTRIBUTING.md and CODE_OF_CONDUCT.md. tools/sbom.py writes the software bill of materials in SPDX 2.3 and CycloneDX, regenerated for this release.
A security policy — #72
SECURITY.md says how to report a vulnerability privately and what we commit to in reply.
The bus consumer claims only what it can run — #70
A task is claimed only when a worker is free to run it. Nothing is received while every worker is busy, so claimed tasks (and their credentials) no longer wait in memory, and a task left unclaimed stays on the bus for another replica. AGENT_BUS_WORKERS (default 4) run at once; one customer holds at most AGENT_BUS_TENANT_WORKERS (default one fewer), so one tenant's burst cannot take every worker.
The conversation memory is bounded, and keyed by Lake's claim — #71
- At most
AGENT_CONVERSATIONSconversations (default 256), each at mostAGENT_CONVERSATION_TURNSturns (20) andAGENT_CONVERSATION_BYTES(1 MiB), all of them at mostAGENT_CONVERSATIONS_BYTES(64 MiB); one idle forAGENT_CONVERSATION_IDLE_SECONDS(3600) is dropped. - Every bus turn carries Lake's claim (task, account, workspace). A bus turn's conversation is keyed by the account and workspace of the claim plus its
conversation_id, so a conversation id named in another tenant's turn reaches nothing. A directPOST /v1/turnkeeps its own key in a separate namespace. Seeding from thehistoryLake sends works as before. A turn that raises is rolled back.
The image's bases are pinned by digest — #73
Each FROM and COPY --from names the multi-platform index digest its tag pointed to. See known issue 3: a merge left an unpinned build stage beside the pinned one.
Prepared for a public export — #75
Links to issues and pull requests, internal names and wording are removed from the README, the docs and the earlier release notes.
Behaviour changes to know before it rolls
- A bus turn's conversation is scoped to the account and workspace of Lake's claim. The same
conversation_idin another account or workspace is another conversation. - Conversations are forgotten past the limits above, oldest or idle first.
- The bus consumer runs at most
AGENT_BUS_WORKERSturns at once and leaves the rest on the bus.
Settings it needs
No setting is required. Seven optional settings are new, each with the default above: AGENT_BUS_WORKERS, AGENT_BUS_TENANT_WORKERS, AGENT_CONVERSATIONS, AGENT_CONVERSATION_TURNS, AGENT_CONVERSATION_BYTES, AGENT_CONVERSATIONS_BYTES and AGENT_CONVERSATION_IDLE_SECONDS. git diff v0.2.7..v0.3.0 -- src adds no other environment read.
Not included
- Claiming tasks by reference and naming the task on every call (#74, for lake#871): deferred for review, not in this release.
Known issues
Security-relevant first.
- Library agents hold live task credentials for every tenant they serve: one shared process claims tasks for every person, each with a credential that acts as that person. A design issue on Lake's side.
- The image bundles GPL and LGPL code (#77): it is built on
python:3.12-slim(Debian: glibc, bash, coreutils, dpkg, readline), so the image is not published. It moves toscratchwith static binaries before it is. - The Dockerfile has two stages named
build(#78): a merge kept the pre-pin stage, which uses the bases by tag only and is the one that builds the venv, beside the pinned one. The digest pins of #73 do not yet hold for the build stage. - Confirmation gates do not establish a separate human confirmer (lake#989, low).
ensure_repositorydoes not match and create repositories in the turn's workspace (#18): a repository with the same name in another workspace makes the deploy fail with a conflict.- No explicit workspace IDs on data-plane calls (#3), and no idle backoff on the bus.
- Generated Apps may still select whole tables (#51) where they should aggregate in SQL and page row-level reads.
- No evals for repair and change turns (#41).
Where it has run
The v0.3.0 tree had not run on a development environment when it was released. Moving a Lake to :v0.3.0 is a separate change to the operator's defaults or the Lake's resource.
The image
Built from the tag v0.3.0 (c8d1645) with BuildKit (linux/amd64, the repository's Dockerfile), then tagged :v0.3.0 on that digest without rebuilding:
docker.io/oramasearch/agent-application:c8d1645@sha256:91fb836f22e6c0a9185984d4dd4b20e4e0ea16073e3847410a7cdcbf3061369c, also tagged :v0.3.0 (same digest).
Pinning it on a Lake is a separate change to the operator's defaults or the Lake's resource.
Verification
| Proof | Result |
|---|---|
Test suite (uv run --frozen --group dev pytest -q -ra, macOS arm64, Python 3.12) |
102 passed, 0 failed, 0 skipped: tests/test_agent.py 23, tests/test_service_auth.py 13, tests/test_repair.py 10, tests/test_turn_destinations.py 10, tests/test_conversations.py 9 (new), tests/test_generated_apps.py 9, tests/test_service.py 8, tests/test_bus_backpressure.py 6 (new), tests/test_evals.py 4, tests/test_repository_file_bytes.py 4, tests/test_release.py 3, tests/test_lake_bus.py 2, tests/test_repository_paths.py 1 |
| Version | pyproject.toml and uv.lock 0.3.0; importlib.metadata.version("agent-application") reports it (tests/test_release.py), and GET /healthz returns that value |
SPDX headers (python3 tools/spdx-headers.py --check) |
Passed |
SBOM (python3 tools/sbom.py --version v0.3.0 --check) |
Passed, regenerated for v0.3.0 |
| Live evals (a model through Lake's gateway) | Not run for this release (needs a live model) |
| Dev environment | Not run (above) |
| Linux image build (BuildKit, from the tag) | Built from v0.3.0: oramasearch/agent-application:c8d1645, also :v0.3.0, sha256:91fb836f22e6c0a9185984d4dd4b20e4e0ea16073e3847410a7cdcbf3061369c |
Release v0.2.7
agent-application v0.2.7
Released 2026-10-27. A patch release, with lake v0.8.1 (lake#1260):
files read exactly; filters applied everywhere. Since v0.2.6. Private repository and image, docker.io/oramasearch/agent-application:v0.2.7.
What it includes
| PR | Change |
|---|---|
| #65 | An App's files are read from their exact bytes (content_base64), so Change App and Fix no longer commit \xNN escapes into them (#63, lake#1238). |
| #66 | Generated Apps apply every filter to every metric, chart and table and keep tables and charts small; a proposal with a widget whose value is never used is refused; a filtered-dashboard eval (#64). |
Known issues
As in v0.2.6.
Verification
- The repository's test suite on each pull request above, each with a regression test that fails on the old code; not
re-run on the release branch. - Image built from the tag on the cluster BuildKit; checked by lake v0.8.1's end-to-end gate on
lake-testing.
Release v0.2.6
agent-application v0.2.6
Released 2026-10-26. A patch release. Since v0.2.5: repository paths are written without '.' components, so a confirmed App at path '.' commits. Replaces v0.2.5, whose uv.lock did not parse (a release-script error), so it has no image.
Private repository and image, docker.io/oramasearch/agent-application:v0.2.6.
What it includes
| PR | Change |
|---|---|
| #58 | repository paths are written without '.' components, so a confirmed App at path '.' commits |
Found by lake v0.8.0's end-to-end gate on lake-testing.
Known issues
As in v0.2.5.
Verification
- The repository's test suite on #58, with a regression test; not re-run on the release branch.
- Image built from the tag on the cluster BuildKit; re-checked by the e2e gate on
lake-testing.
Release v0.2.5
Superseded by v0.2.6, never deployed. This tag's uv.lock does not parse (an error in the release script), so no image was built. Use v0.2.6, which carries the same change.
Release v0.2.4
agent-application v0.2.4
Released 2026-10-26. A patch release. Since v0.2.3: a change conversation keeps its earlier turns.
Private repository and image, docker.io/oramasearch/agent-application:v0.2.4.
What it includes
| PR | Change |
|---|---|
| #55 | seeded_history() seeds a conversation from the history Lake v0.8.0 sends with a turn (lake#1240), so a thread continues after the agent restarts |
Not included
- Nothing else changed. Apps already generated are not changed.
Known issues
As in v0.2.1.
Verification
- Test suite on #55: 75 passed. Not re-run on the release branch (Steve: ship without waiting).
- Image built from the tag on the cluster BuildKit. Runs with Lake v0.8.0, whose gateway sends this agent's
gpt-6.1-sol tool calls through the Responses API (lake#1245).
Release v0.2.3
agent-application v0.2.3
Released 2026-10-26. A patch release. Since v0.2.2: generated Apps read only what the page shows, and the
agent's fallback model is gpt-6.1-sol. Private repository and image, docker.io/oramasearch/agent-application:v0.2.3.
This release is the tree of dev at ba264e5 with these notes, the version and the README's release line.
What it includes
| PR | Change |
|---|---|
| #52 | The prompt says to aggregate in SQL and page row-level reads with LIMIT/OFFSET, never SELECT * without LIMIT, and never to join whole tables in the App. unbounded_selects() flags a literal sql(...) statement with neither a LIMIT nor an aggregate, in validation and repair, naming the line (#51). The fallback model is gpt-6.1-sol |
| #50 | The v0.2.2 release merge |
Why:
- A production App copied three whole tables into memory every ten minutes.
- Lake v0.7.2 answers
too_many_rowsto an App query past 5,000 rows. - In Lake v0.7.2 people choose the agents' model per workspace (default
gpt-6.1-sol), and the gateway applies that
choice to this agent's calls. The fallback matters only when the agent is run without a Lake model.
Not included
- Apps already generated are not changed. One that reads whole tables is fixed by a
repairorchangeturn.
Known issues
As in v0.2.1.
Verification
| Proof | Result |
|---|---|
Full suite (uv run pytest) |
74 passed, 0 failed, 0 skipped |
| New test | the production loader is flagged; GROUP BY, an aggregate and LIMIT/OFFSET paging pass |
| Dev environment | agent-application:ba264e5 (this tree's code) ran on dev with lake dev-11153627, Ready |
| Image | built from the tag on the cluster BuildKit (linux/amd64) |
Release v0.2.2
agent-application v0.2.2
Released 2026-10-26. A patch release. One change since v0.2.1: generated Apps call orama_app's
zetta() before reading from it, and the agent refuses zetta.sql(...) before it commits an App. Private repository
and image, docker.io/oramasearch/agent-application:v0.2.2. Tracking issue:
#47.
This release is the tree of dev at d0db208 with these notes, the version and the README's release line.
What it includes
| PR | Change |
|---|---|
| #48 | The prompt shows zetta() with an example. proposal_defects and repair refuse attribute reads on the bare zetta/service functions, naming the line and the right form |
Why: a production App written by the agent (customer-insights-dashboard.us.lake.live, 2026-10-26) read
zetta.sql(...). That compiles, then fails on every page with AttributeError: 'function' object has no attribute 'sql'. The prompt's wording ("use its sql") invited it, and validation only checked that the code parsed.
Not included
- Apps already generated are not changed. One that reads
zetta.sqlis fixed by arepairturn, or by editing its
code.
Known issues
As in v0.2.1.
Verification
| Proof | Result |
|---|---|
Full suite (uv run pytest) |
73 passed, 0 failed, 0 skipped |
| New test | the production line is refused with the right form; zetta(), with zetta() as w, service("db")[...] pass; a local zetta that is not orama_app's is left alone |
| Image | built from the tag on the cluster BuildKit (linux/amd64) |
Release v0.2.1
agent-application v0.2.1
Released 2026-10-25. A patch release. One change since v0.2.0: the Application Agent
can repair an App that failed and change an existing App from more instructions, always as a
diff the person confirms, plus Lake's opt-in auto-fix (lake#1207). Private repository and image,
docker.io/oramasearch/agent-application:v0.2.1. Tracking issue:
#42.
It is a patch: the existing turns (design, confirm, deploy) and every setting are unchanged; the two
new turn kinds are only sent by a Lake that carries lake#1218.
This release is the tree of dev at 73efc06 with these notes, the version and the README's release
line.
How it works: repairing and changing an App,
building, confirming and deploying Apps, service and bus operation.
What it includes
Every pull request merged into dev since v0.2.0:
| PR | Change |
|---|---|
| #40 | repair and change turns on an existing App, with commits, a diff the person confirms, and opt-in auto-fix (lake#1207) |
| #39 | Records the v0.2.0 image digest in the v0.2.0 notes |
| #38 | Merges main back into dev after v0.2.0 |
Repair and change turns (#40)
- Reads the App's facts:
receive_app_events,app_logs(when logs are unavailable it says so,
lake#1206, and never gives missing logs as the
reason),list_buildsandbuild_logs, and every file under the App's folder, from the source Lake
recorded at the last deploy when the App runs a built image. - Proposes, writes nothing: the proposal is
{kind: proposal, confirm: true, turn, diagnosis, diff, files, delete, app}, holding only the changed files and a unified diff computed by the agent, not
the model. It is checked before it is shown: safe paths, the code compiles, no secret literal or
turn credential,app.tomlkeeps its name, framework and port, and at least one file changes. A
failed check goes back to the model, twice at most. - On confirm (
context.confirmed_proposal): refuses a proposal for another App, commits the files
onmainin onecommit_repository_files(each write a liquid drip attributed to the person), sets
APP_DEPSand builds when a package outside app-runtime's baked set is now needed, deploys the new
revision or image, follows the rollout and reports. - Auto-fix (
context.auto_fix: {attempt, max_attempts}, granted by Lake only when the person opted
in): applies its own proposal, names the attempt, and writes nothing when it finds no fix. Turn.kindcomes from the body'skind;list_buildsandget_auto_fixjoin the read set. New
promptprompts/repair-v1.md; the design prompt is stillv1.
Settings it needs
No new setting. The settings of v0.2.0 still apply
(AGENT_SERVICE_TOKEN and LAKE_URL required, AGENT_MODEL read, all rendered by the operator).
The new turns need the Lake that sends them, grants their tools and holds the auto-fix gate:
lake#1218, merged into lake's dev on 2026-10-25 and
not in a lake release yet (lake v0.7.0 does not have it). An older Lake never sends repair or
change.
What it does not include
Unchanged from v0.2.0:
- no conversation memory that survives the process;
prior_turnis not read; - no explicit
workspace_idon its calls; - no creating secrets;
- no frameworks beyond FastAPI, Streamlit and static sites;
- no idle backoff on the bus;
- no live eval result, and no eval cases yet for repair and change turns
(#41).
Known issues
Security-relevant first. All open at this tag.
- Library agents hold live task credentials for every tenant they serve
(lake#871). - The bus consumer claims tasks into an unbounded queue before a worker is free
(#33, LSR-075). - Conversation histories are unbounded
(#34, LSR-076). - A confirmed turn cannot create an App with a Lake secret reference
(lake#891). - Confirmation gates do not establish a separate human confirmer
(lake#989, LSR-183, low). Auto-fix applies its own
proposal without a confirm, by design, only when the person opted in on Lake. - Repair and change turns are proven with scripted model answers only
(#41), and have not run against a live
Lake carrying lake#1218. - App logs may be unavailable (lake#1206); a
repair then works from events, build logs and source. - A repository with the same name in another workspace makes the deploy fail with a conflict
(#18). - No explicit workspace IDs, and no idle backoff on the bus
(#3).
Where it has run
The v0.2.1 tree has not run on the dev environment. At this tag the dev member of
lake-nonprod pins oramasearch/agent-application:32caa4c (the v0.2.0 tree); no build of #40 has
been pinned there. Steve asked for the release on 2026-10-25 ("merge it all, cut releases, push all
to nonprod"), together with lake#1218.
The image
Built from the tag v0.2.1 (9301750) on the cluster BuildKit with build-agent.sh (linux/amd64, the
repository's Dockerfile), then tagged :v0.2.1 on that digest without rebuilding:
docker.io/oramasearch/agent-application:9301750@sha256:ff5740591a5b2e056b7fa8b2519f7901a8ad01baf5839a351db4f8f2939e5e4c, also tagged :v0.2.1 (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 pytest -q -ra, macOS arm64, Python 3.12.11, release branch) |
72 tests: 72 passed, 0 failed, 0 skipped. test_agent.py 21, test_service_auth.py 13, test_repair.py 10 (new, #40), test_turn_destinations.py 10, test_service.py 7 (one new: the service passes the turn kind), test_release.py 3 (the installed package is 0.2.1, these notes exist and are indexed), test_evals.py 3, test_generated_apps.py 3, test_lake_bus.py 2 |
| Release test | tests/test_release.py reads the version from pyproject.toml and the installed package, and fails when docs/releases/v<version>.md or its index line is missing |
Live model evals (application-agent-evals) |
Not run (needs a live model; no repair or change cases yet, #41) |
| Linux image build (cluster BuildKit, from the tag) | Built from v0.2.1 with build-agent.sh: oramasearch/agent-application:9301750, also :v0.2.1, sha256:ff5740591a5b2e056b7fa8b2519f7901a8ad01baf5839a351db4f8f2939e5e4c; private (404 anonymously) |
Release v0.2.0
agent-application v0.2.0
Released 2026-10-25. A minor release. Six pull requests since v0.1.0, most of them
fixes from the 2026-10-21 security review: the service token is compared in constant time and turns
are refused when it is unset, a turn's credential goes only to the service's own LAKE_URL, and no
turn trace is kept on the pod's disk. It also proposes Lake's secret references and takes the Lake's
AGENT_MODEL. Private repository and image, docker.io/oramasearch/agent-application:v0.2.0.
This release is the tree of dev at 32caa4c with these notes and the version. Tracking issue:
#35.
How it works: service and bus operation,
building, confirming and deploying Apps, evaluation scenarios.
What it includes
Every pull request merged into dev since v0.1.0:
| PR | Change |
|---|---|
| #29 | POST /v1/turn compares the bearer with hmac.compare_digest, and fails closed: with AGENT_SERVICE_TOKEN unset or empty every turn is refused with 503 (it used to be open). A wrong or missing bearer is still 401 (#23) |
| #31 | A turn's credential goes only to Lake at this service's own LAKE_URL, for capabilities and the model gateway alike. The turn's lake_url is ignored. LAKE_URL must be an absolute http(s) URL with a host and no user info, query or fragment; otherwise, or when unset, the turn is refused before any request goes out (#30) |
| #27 | A turn keeps nothing on the pod's disk: no <TRACE_DIR>/<trace_id>.json. TRACE_DIR is no longer read; the answer still carries trace_id. Lake keeps the answer in the task's sealed record in the workspace Zetta (#21) |
| #28 | A secret environment value must be one of Lake's two reference forms, {"secret": {"mount_id", "name", "field"}} or {"secret": "lake-secret://<mount id>/<name>#<field>"}. A Kubernetes Secret name is refused. The prompt (still v1) proposes a reference only when the person named an existing secret, and otherwise leaves the variable out and points to the App's Environment settings (#19) |
| #32 | The model is APPLICATION_AGENT_MODEL when set, else the Lake's AGENT_MODEL, else gpt-5.5, so the Lake's spec.agents.model now reaches this agent (#22) |
| #26 | Records the v0.1.0 image digest in the v0.1.0 notes |
Settings it needs
| Variable | Why | Who sets it |
|---|---|---|
AGENT_SERVICE_TOKEN |
Now required: without it every POST /v1/turn is 503 (#29). Bus turns do not use it |
The operator, on every library agent pod |
LAKE_URL |
Now required for every turn, and must be a clean absolute http(s) URL (#31) | The operator |
AGENT_MODEL |
Now read (#32); APPLICATION_AGENT_MODEL still overrides it |
The operator, from spec.agents.model |
No new setting. The operator still sets TRACE_DIR=/tmp/traces; the agent ignores it. A hand-run
service must now set AGENT_SERVICE_TOKEN and LAKE_URL.
What it does not include
Unchanged from v0.1.0, except that turn traces are gone:
- no conversation memory that survives the process;
prior_turnis not read; - no explicit
workspace_idon its calls; - no creating secrets;
- no frameworks beyond FastAPI, Streamlit and static sites;
- no idle backoff on the bus;
- no live eval result: the live evals need a live model and were not run.
Known issues
Security-relevant first. All open at this tag.
- Library agents hold live task credentials for every tenant they serve
(lake#871): one shared process claims tasks for
every person, each with a credential that acts as that person. A design issue on Lake's side. - The bus consumer claims tasks into an unbounded queue before a worker is free
(#33, LSR-075): claimed tasks wait in
the executor holding their credentials. - Conversation histories are unbounded
(#34, LSR-076): they are never evicted
from the process's memory. - A confirmed turn cannot create an App with a Lake secret reference
(lake#891): Lake checks a reference with the
caller'sget_secret_mountandread_secret, which a confirmed turn's task credential lacks. The
agent now proposes the right form (#28); creating the App with it waits on Lake. - Confirmation gates do not establish a separate human confirmer
(lake#989, LSR-183, low). - A repository with the same name in another workspace makes the deploy fail with a conflict
(#18). - No explicit workspace IDs, and no idle backoff on the bus
(#3). - Lake's e2e T11 and T12 confirm without
proposal_task_idandproposal_sha256, so Lake refuses
the confirming turn (lake#1183). The script is
out of date, not this agent; until it is fixed the e2e gate does not exercise the confirm step. - The model is offered reads a design turn cannot make:
build_logs,get_buildand
receive_build_eventscome back as errors on a design turn.
Where it has run
The v0.2.0 tree has not run on the dev environment. No dev-<sha> build of these changes was
pinned on the dev member of lake-nonprod; every Lake runs the operator default,
oramasearch/agent-application:v0.1.0 (orama-operator internal/controller/lake_defaults.go on dev).
Moving the default to :v0.2.0 is a separate operator change, rolled through testing, staging and prod.
The image
Built from the tag v0.2.0 (8a10773) on the cluster BuildKit with build-agent.sh (linux/amd64, the repository's Dockerfile), then tagged :v0.2.0 on that digest without rebuilding:
docker.io/oramasearch/agent-application:8a10773@sha256:120a208d38e4655e31e54afcb1dcef06faf6f999cae0e36e5ac4a44e5d3f6982, 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 pytest -q -ra, macOS arm64, Python 3.12) |
62 tests: 62 passed, 0 failed, 0 skipped. test_agent.py 21, test_service_auth.py 13 (new, #29), test_turn_destinations.py 10 (new, #31), test_service.py 7, test_release.py 3 (the installed package is 0.2.0, these notes exist and are indexed), test_evals.py 3, test_generated_apps.py 3 (generated Apps loaded by app-runtime at 692faac), test_lake_bus.py 2 |
| Release test | tests/test_release.py reads the version from pyproject.toml and the installed package, and fails when docs/releases/v<version>.md or its index line is missing |
/healthz version |
test_healthz_reports_the_release_version: version is the installed package's version |
Live model evals (application-agent-evals) |
Not run (needs a live model) |
| Linux image build (cluster BuildKit, from the tag) | Built from v0.2.0 with build-agent.sh: oramasearch/agent-application:8a10773, also :v0.2.0, sha256:120a208d38e4655e31e54afcb1dcef06faf6f999cae0e36e5ac4a44e5d3f6982; private (404 anonymously) |