Skip to content

Release v0.7.1

Choose a tag to compare

@homer6 homer6 released this 26 Oct 03:42
· 269 commits to main since this release
7241d54

Released: 2026-10-25 · Branch: main · Previous: v0.7.0 (2026-10-24) ·
Tracking: #1219, epic #1158.

A patch:

  • The Quickstart decides what to build with you. Its model step is an Ask Orama conversation on the
    dataset you dropped. It trains on any column types, a clustering model trains, and every state after the
    conversation says what happens next. The website code comes as a copy-ready agent prompt.
  • A failed app tells you, and the Application Agent fixes it. A failed rollout or build is a
    notification with Fix with the Application Agent; an app can be changed with instructions, and auto-fix
    is opt-in. The Apps page counts Degraded apps as needing attention.
  • Ask Orama's confirm card is readable, and it never proposes a change that cannot work.
  • Auto-Suggest finds opportunities again, and a scan survives leaving the page.

No migrations and no new environment variables since v0.7.0. Ships with agent-ask-orama v0.2.1 and
agent-application v0.2.1 (Before it rolls).

Features

PR Change
#1199 The Quickstart's model step is an Ask Orama conversation (#1182). Once search works, Decide What to Build opens Ask Orama on the dataset: its name, row count, each column's kind and whether it can be predicted, and the index. It leads with ideas: a model, a target derived from a column (a threshold, bands or a set of values), M-3 clustering, Auto-Suggest opportunities, an analysis, or an app built by the Application Agent (such as a dashboard). Confirming a model trains it, and the steps continue. Nothing to Predict Yet is gone; Choose the columns myself stays as a fallback. New capability start_quickstart_conversation. quickstart_status names the conversation. train_quickstart gains derived_target and methodology: "M-3" (a clustering model is not published for the website: zetta serves predictions only for gradient-boosted models). The conversation lives in the workspace Zetta (../api/quickstart.md)
#1186 The Quickstart trains on any column types, detected automatically, and never fails on transform_required (#1181). The training copy reads up to 1,000,000 rows and must fit upload_dataset's 16 MiB (#1190)
#1187 Website code as a Copy agent prompt tab (the default), carrying the source and everything an agent needs to install it, beside a Source code tab (#1185)
#1218 A failed app is a notification, fixed or changed with the Application Agent, with opt-in auto-fix (#1207). A background pass records a notification in the app's workspace Zetta when a rollout ends Degraded, Failed or crash-looping at the desired generation, or the newest build fails; Ready again resolves it. The Apps list shows a banner per failed app; the app page has Fix or Change This App (open notifications, Fix with the Application Agent, a Change this app box, and Auto-fix: off by default, 1 to 10 attempts, default 3). The agent proposes a diff; once confirmed it commits through commit_repository_files (liquid drips, attributed to the person) and redeploys. Auto-fix makes one attempt per failed generation as the person who turned it on, then stops and says so. Account → Notifications lists every notification of the workspace. New capabilities list_notifications, dismiss_notification, get_auto_fix, set_auto_fix, repair_app, change_app; agent_turn takes kind: repair or change for agent-application. Records: lake_resource_notifications and lake_resource_repairs in the workspace Zetta; nothing in Postgres (../api/notifications.md)

Fixes

PR Change
#1213 A Quickstart clustering model trains, and every state after the conversation has one next step (#1203, #1204). Lake no longer sends a held-out share for M-3, which zetta refuses; the engine chooses the number of segments (2 to 8). After a confirm the page switches to the training at once and says how far along it is; a refused or failed training gives the reason in plain words with Try again (same choices) and Ask Orama to fix it. A confirmed train_quickstart writes what started into the conversation, or ok: false, code: training_failed (new error code) with the plain reason, instead of "Done: train_quickstart.". The conversation page carries quickstart_state, model, model_problem and quickstart_next so Ask Orama can follow it (../api/quickstart.md, ../ui/quickstart.md)
#1211 The Apps page counts Degraded apps as needing attention (#1208). Failed becomes Needs attention (Failed or Degraded), linked to ?filter=attention, with Show all apps back (../ui/apps.md)
#1189 Auto-Suggest suggests opportunities again: plan_features works from a dataset alone, with its snapshot cutoff. With agent-autosuggest v0.2.0 (#24), Scan my workspace keeps its Opportunities (#1153; e2e T6 passes on dev)
#1201 A scan survives leaving the page (#1195). A finished scan's Opportunities are saved when the agent publishes them, not when someone reads them. list_opportunities returns last_scan, so the page shows Scanning… started 2m ago or the results without ?task=. A second scan while one runs joins it (../api/discover.md)
#1198 Ask Orama's proposals:
• the confirm card shows each argument as name, value and description, wrapping across the full width, with long values folded (#1192);
• a proposal naming a project that does not exist is withheld before it is shown, and a new project's pipeline is one confirmable compile_pipeline with create_project: true (#1193);
• every turn carries Lake's record of the conversation's proposals (pending, confirmed, failed, withheld), so after a failure "create it then" gets a new, corrected proposal (#1194) (../api/ask-orama.md)
#1217 A test pins Lake's half of the App Zetta binding contract (#1205). HttpResponseIsolation.AnAppsZettaBindingReadsItsWorkspaceThroughLake builds the binding as the operator renders it after orama-operator#1001 and proves the app's read of its own workspace is served, and another workspace, an action and a wrong token are refused. No Lake code changes (../api/workspaces.md)

Before it rolls

  • Schema: unchanged since v0.7.0 (git diff v0.7.0 v0.7.1 -- db/ is empty). The operator runs lake migrate
    from the image, which applies nothing new.
  • Settings: nothing new is read from the environment (git grep getenv between v0.7.0 and v0.7.1 adds only
    the test variables LAKE_TEST_ZETTA_URL and LAKE_TEST_ZETTA_TOKEN); environment_variables() and
    stripe_webhook_events are unchanged. The new refusals are Lake's own (a repair turn naming another app, auto-fix
    past its attempts, a clustering of fewer than 2 rows) and need no setting.
  • Agents: roll with agent-ask-orama v0.2.1 (prompt v17: it follows the Quickstart's state after a confirm,
    agent-ask-orama#56) and agent-application v0.2.1
    (repair and change turns, agent-application#40; without
    it a repair turn gets the old new-app behaviour). The other agents stay at v0.2.0.
  • Apps with workspace data start only once orama-operator#1001
    (the binding as JSON) is released and rolled (#1205).

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 (commit_repository_files, build_app, set_app_env, deploy_app) are
    capability-level, so it could act on another app or repository of the same workspace the person may write
    (#1216).
  • Those of v0.7.0: the open security issues, #1151 and #1161, the flaky timing
    test #1150, and the follow-ups listed there. #1153 is fixed here.
  • Apps with workspace data crash-loop until orama-operator#1001 rolls (#1205,
    Lake's side is done). An app that queries with workspace.sql() will then fail on its first query: app-runtime
    calls /zetta/v1/sql, which Lake does not serve on that route (#1212).
  • app_logs is empty on nonprod: no Loki-compatible log source is configured there
    (#1206).
  • A failed app is not yet told to the Quickstart or the conversation that built it
    (#1214), and notifications are listed per workspace, not across
    the account (#1215). Custom agents are not yet a repairable
    kind (#1209).
  • The v0.7.0 end-to-end failures on testing (#1183,
    #1184, #1188,
    T7 and T11/T12) are not fixed here. The e2e gate reruns on testing for this release.
  • Follow-ups filed:
    • datasets larger than the 16 MiB training copy (#1190);
    • Ask Orama can't find the last-uploaded dataset (#1191);
    • saved dataset dashboards (#1196);
    • a clustering model's segment per row (#1197);
    • a clustering's K range once zetta's train_model takes one (#1210);
    • the running-scan indicator on the nav (#1200);
    • the Project Corpus Agent's sandbox replay (#1202). Until
      then, agent-project-corpus v0.2.0 refuses Lake turns.

Where it has run

lake-dev ran dev-edb8cf5d (through #1199). The four PRs merged after it (#1213, #1217, #1211, #1218) had not
run on any cluster before this release; Steve asked for the release with them (2026-10-25). The release image goes
to testing first, where the e2e gate runs.

Image: oramasearch/lake:v0.7.1, also main-7241d549, digest
sha256:86483b2f11577823f411ec5fe37d3a8e7b46b0a0f7312737f726a7651d23970c, built on the cluster BuildKit (linux/amd64)
from a detached checkout of the tag (7241d549, 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 6bff8e73 plus the version and these notes, b6f6d2ae), macOS arm64, Release,
against zetta v0.5.0 (e11c02f5) and liquid at liquid-bridge's pin, PostgreSQL 17 (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 (270 s):

Check Result
Build every target builds and links
Full suite 1,643 tests: 1,628 passed, 0 failed, 14 skipped, 1 disabled
Live Zetta 12 tests ran and passed against the real engine: QuickstartZetta.* (3: search and model answers, a dated dataset of every column kind, and a clustering model trained to done), WorkspaceStoreZetta.* (6), WorkspaceRecordsZetta.*, RequestRecorderZetta.* and FeaturePlanZetta.*; none skipped
Skipped the PostgreSQL variants of five HttpResponseIsolation cases whose state is written only on the memory store (×10, by design, among them #1217's binding test), a browser (InstanceOrigin.BrowserToTerminal), the Hugging Face Python client, a stdio MCP runner, and BrowserKeys.AnExpiredKeyIsToldItExpired on the memory store (expiry is the PostgreSQL store's); 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, unit tests, 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 b6f6d2ae: the whole lake-tests binary exits 0; ctest 1,573 of 1,574 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 image on any cluster, the testing e2e gate, and a repair turn against a live Lake, Zetta and
model.