Skip to content

Release v0.7.2

Choose a tag to compare

@homer6 homer6 released this 27 Oct 01:27
· 253 commits to main since this release
ef4a519

Released: 2026-10-26 · Branch: main · Previous: v0.7.1 (2026-10-25) ·
Tracking: #1234.

A patch:

  • An App's workspace.sql(...) works, at the access its binding grants: read-only always, and at most
    5,000 rows a query.
  • Auto-fix sees an App whose code keeps failing while its pod looks healthy, and Restart records what it
    restarted so the operator can replace the pods.
  • People choose the model their workspace's agents use. The gateway serves and prices the gpt-6.x models
    (default gpt-6.1-sol), and a prompt over 272K input tokens is billed at what it costs.

Two migrations (0104_app_restarts, 0105_agent_model_prices), applied by lake migrate from the image. No new
environment variables (Before it rolls).

Fixes

PR Change
#1231 sql on the App gateway runs as workspace_sql, at the binding's access level (#1229, also #1212). sql on /workspaces/<customer>/zetta/v1/ (and under /w/<workspace>/) is no longer unknown_capability: GET takes the input from the query string (what app-runtime sends), POST a JSON body. The App's read binding, the read-only check, safe mode and the audit apply as for people and agents. Writes, DDL, CREATE USER and two statements answer 400 invalid_input and never reach the workspace. A statement without a LIMIT gets LIMIT 5001; more than 5,000 rows, or a LIMIT over 5,000, answers the new error code too_many_rows (422). An answer is never cut short without saying so. A direct workspace_sql call is uncapped. workspace_sql accepts the engine's intent and passes it on (../api/workspaces.md, ../api/web-apps.md, ../api/errors.md)
#1232 Auto-fix sees a healthy App whose script keeps failing, and Restart is recorded (#1228). For a running, Ready App with auto-fix on, the repair pass reads its newest 500 log lines from the last five minutes through app_logs; two or more Streamlit or FastAPI tracebacks record an app_errors notification (the exception line as its reason, the traceback as detail) and start auto-fix's one attempt for that generation. Apps with auto-fix off are never read. Metric lake_repair_log_reads_total. Restart now sets apps.restart_generation (migration 0104_app_restarts), and get_app answers it. The pods are replaced only once orama-operator#1033 is released and rolled (it writes the number on the pod template); until then Restart is recorded but no pod is replaced (../api/notifications.md, ../control-plane/web-apps.md)

Features

PR Change
#1233 People choose the agents' model per workspace; the gateway serves and prices gpt-6.x and bills long prompts at cost (#1230). The gateway serves gpt-6.1-sol, gpt-6.1-terra, gpt-6-astra and gpt-6-luna beside the existing models. A workspace setting agent_model (get_workspace_settings for members, set_workspace_settings for owners and admins), kept as a sealed lake_workspace_settings record in the workspace Zetta (nothing in Postgres), names the model every agent turn's chat call is sent with; unset means gpt-6.1-sol. A chosen model that is no longer served is refused, never swapped. The console's Workspaces → Settings has a picker with prices. Migration 0105_agent_model_prices seeds the rates at OpenAI's standard price plus 30% (read 2026-10-26), with long-prompt classes <model>@long/input and <model>@long/output (also for gpt-5.5): more than 272,000 input tokens is recorded under them. service_tier is pinned to default; any other value is refused. With billing on, a model is served only while its rates, long ones included, are published (../api/model-gateway.md, ../api/billing.md, ../api/workspaces.md)

Docs

PR Change
#1222 The v0.7.1 image and its digest on its release page
#1223 A runbook for the human steps that give app hosts an edge certificate (../deployment/edge-certificates.md, #798); a human action, open

Before it rolls

  • Schema: two migrations since v0.7.1 (git diff v0.7.1 v0.7.2 -- db/): 0104_app_restarts (a column
    apps.restart_generation, guard app_restarts_guard) and 0105_agent_model_prices (rates only, no table
    changes, guard agent_model_prices_guard). The operator runs lake migrate from the image before the new
    binary serves (orama-operator#936), so nothing is vendored. Without 0105 a Lake with billing on refuses the
    gpt-6.x models and gpt-5.5.
  • Settings: nothing new is read from the environment (git grep getenv between v0.7.1 and v0.7.2 adds
    nothing); environment_variables() and stripe_webhook_events are unchanged. The new refusals are Lake's own
    and need no setting: SQL writes and over-cap reads on the App gateway, a service_tier other than default,
    a model whose rates are not published while billing is on, and a workspace model that is no longer served.
  • Agents follow the workspace's choice through the gateway; no agent image changes. The sets name
    agents.model: gpt-6.1-sol, which only says what agents ask for.
  • Restart replaces pods only with orama-operator#1033,
    an operator patch released separately. Either may roll first: the operator reads the column tolerantly.

Known issues

  • Security: a confirmed or auto-fix repair or change turn is checked against its app when it starts, but its
    task credential's write scopes are capability-level, so it could act on another app or repository of the same
    workspace the person may write (#1216). Auto-fix now also
    starts on app_errors, so this applies to more turns.
  • Security: Lake's read-only SQL guard and Zetta's statement parser disagree on comments and statements
    (#926); the App gateway's sql goes through that guard.
  • The other open security issues listed in v0.7.0 remain open.
  • app_logs is empty on nonprod: no Loki-compatible log source is configured there
    (#1206), so the app_errors detector reads nothing there.
  • The gateway was not exercised against OpenAI by the pull requests' suites (they use a fake provider); the dev
    environment's gate checked the workspace model setting, not a paid long prompt.
  • The end-to-end failures from v0.7.1's gate on testing are not fixed here: Ask Orama ranks the wrong at-risk table
    (#1225), walkthrough.mjs fails after the Quickstart's failure
    path (#1226), and the Experiments Agent's turn crashes
    (agent-experiments#26). Also open: a file dropped
    before a new workspace is ready is never loaded (#1224), and
    the v0.7.0 and v0.7.1 follow-ups listed on their pages.

Where it has run

lake-dev ran dev-11153627, the commit this release is cut from, and passed the dev gate: health, an App's
gateway SQL with its read binding, the 5,000-row cap, and the workspace model setting. The release image goes to
testing first, where the e2e gate runs.

Image: oramasearch/lake:v0.7.2, also main-ef4a519f, digest
sha256:b05e5fe553fe16e520903f2496d5e063fb9c8c94623092fb8edd52f2b8df6e63, built on the cluster BuildKit (linux/amd64)
from a detached checkout of the tag (ef4a519f, LAKE_BUILD_REVISION) against zetta v0.5.0 and liquid-bridge's
liquid pin (27d56ba). The Docker Hub repository stays private.

Verification

On the release branch (dev at 11153627 plus the version and these notes, defd7172), macOS arm64, Release,
against zetta v0.5.0 (e11c02f5) and liquid at liquid-bridge's pin (27d56ba), PostgreSQL (a database of its own),
LAKE_TEST_SKIP_STALE_SWEEP=1, a live Zetta v0.5.0 node (LAKE_TEST_ZETTA_URL, LAKE_TEST_ZETTA_TOKEN,
LAKE_TEST_ZETTA_BINARY), ctest -j6 (248 s):

Check Result
Build every target builds and links
Full suite 1,658 tests: 1,642 passed, 0 failed, 16 skipped, 1 disabled
Live Zetta 12 tests ran and passed against the real engine: QuickstartZetta.* (3), WorkspaceStoreZetta.* (6), WorkspaceRecordsZetta.*, RequestRecorderZetta.* and FeaturePlanZetta.*; none skipped
Skipped the PostgreSQL variants of six HttpResponseIsolation cases whose state is written only on the memory store (×12, by design, among them #1231's AnAppsZettaBindingRunsReadOnlySqlThroughLake, which passes on the memory store), a browser (InstanceOrigin.BrowserToTerminal), the Hugging Face Python client, a stdio MCP runner, and BrowserKeys.AnExpiredKeyIsToldItExpired on the memory store; SearchBenchmark is disabled by design
Version ReleaseVersion.* (3) pass: the three places agree and this page is indexed
Web app npm ci && npm run build (typecheck, 449 of 449 unit tests, vite build, effect and list-error checks), lake-ui-token-gates and lake-ui-contrast pass
Linux cluster BuildKit lane (linux/amd64, clang-19, kernel 6.8 azure) at defd7172: the whole lake-tests binary exits 0; ctest 1,587 of 1,588 passed, the one failure the lane's lake-postgres-tests_NOT_BUILT placeholder (that lane builds no PostgreSQL tests and runs no Zetta), 1 disabled

No test crashed, so no AddressSanitizer run was needed.

Not yet run: the release image on any cluster, the testing e2e gate, the gateway against OpenAI with a long
prompt, and Restart replacing pods (needs orama-operator#1033).