Repository navigation
Releases
v0.1.11
What's Changed
- Add GLM 5.3 and GLM 5.3 Flash to SEED_MODELS via OpenRouter by @willjohnson in #48
- Add Gemini Decisions (gemini-decision-1) by @willjohnson in #49
- Add Claude Sonnet 5.5 and Haiku 5.5, remove Haiku 4.5, sort SEED_MODELS by @willjohnson in #50
- Add gemini-4-neon seed model (v0.1.11) by @willjohnson in #51
Full Changelog: v0.1.10...v0.1.11
v0.1.10
What's Changed
- Bump the uv group across 1 directory with 2 updates by @dependabot[bot] in #45
- Bump the uv group across 1 directory with 2 updates by @dependabot[bot] in #46
- Add Fable, GPT-6 Terra, and Gemini 3/4 seed models (v0.1.10) by @willjohnson in #47
Full Changelog: v0.1.9...v0.1.10
v0.1.9
Versioned team-level system prompts, plus a restyled admin UI.
- Adds a team-level system prompt between the tenant and agent prompts, so the prompt sandwich is now tenant → team → agent → memory. Team owners edit it with
PATCH /v1/teams/{id}or on the team page. - Versions tenant and team prompts in a new
prompt_versionstable. Each job records the versions it ran with (tenant_prompt_version_id,team_prompt_version_id). History is atGET /v1/tenants/{id}/prompt-versionsandGET /v1/teams/{id}/prompt-versions, and on the settings and team pages. - Restyles the admin UI with Tailwind CSS and keeps the selected tenant on pages outside a tenant.
- Demo: each poller run gets its own event ids; screenshots are refreshed; adds a note on fixing MinIO volume ownership after the image switch.
- Dependency updates: opentelemetry-api, certifi, and two uv group bumps.
Migration required (alembic upgrade head). It moves each tenant's system prompt in as version 1 and drops tenants.system_prompt. 0.1.8 processes can't read tenants after it runs, so deploy at a quiet time and check for jobs stuck in queued or running afterwards.
Validation includes the full test suite (435 tests, no warnings), lint/format checks, lockfile consistency, and source/wheel builds. Live vendor API calls were not tested.
Full changelog: v0.1.8...v0.1.9
v0.1.8
Dependency security updates and a warning-free test suite.
- Upgrades pydantic-ai to 2.53.0, mako to 1.4.2, and multidict to 6.9.1, closing 20 Dependabot alerts (4 high), including GHSA-6fqq-452j-qhrp in pydantic-ai's
ConcurrencyLimitedModel. - Uses pydantic-ai 2.53's profile key names (
supports_forced_tool_choice,thinking_enabled_by_default) for the GPT-6 and Opus 5.5 models; behavior is unchanged. - Uses Starlette's current status-code names and requires
starlette>=0.48, the first release that has them. API responses are unchanged (still 422 and 413).
No database migration is required. Rebuild images or re-sync dependencies to pick up the upgraded packages.
Validation includes the full test suite (417 tests, no warnings), lint/format checks, lockfile consistency, and source/wheel builds. Live vendor API calls were not tested.
Full changelog: v0.1.7...v0.1.8
v0.1.7 — OpenAI Decisions support
Adds GPT-6 Luna Decisions to the starter model registry, running decision agents over the OpenAI Decisions API.
- Select
openai:decisions/gpt-6-luna(registry namegpt-6-luna-decisions) and put OpenAI's native question list (predicate, choice, and score questions) in the version's Params; leave Output schema blank. - Jobs accept
context.prompt, optional text files/links, and up to 128 image attachments.output.answersis a list in question order with names, probabilities, typed choices, scores, and refusals; a refusal is returned as an answer so the others remain available. - Tool grants, delegation, memory, learning, human escalation, custom output schemas, and chat parameters are rejected for decision agents. Injection screening, PII redaction, timeouts, retries, and budget pre-flight checks still apply.
- Without a price for the Decisions entry, jobs record a
cost_unpricedevent rather than advancing monthly spending limits. - Uses an
openaiprovider credential at agent, team, or tenant scope, orOPENAI_API_KEY. - Dependency updates: oauthlib 4.0.0, pyjwt 2.15.0, urllib3 2.8.0.
After updating, register the model with:
docker compose exec api sleeper seed-modelsFor non-container installations, run sleeper seed-models. No database migration is required. See the README's "OpenAI decision agents" section for a full Params example.
Validation includes the full test suite (417 tests), lint/format checks, and source/wheel builds.
Full changelog: v0.1.6...v0.1.7
v0.1.6
Adds GPT-6 Sol, GPT-6 Astra, GPT-6 Luna, and Claude Opus 5.5 to the starter model registry.
- Routes GPT-6 through the Responses API with reasoning support for scoped and environment credentials.
- Uses native structured output for Opus 5.5, which rejects forced tool use.
- Preserves existing model registrations and agent selections.
After updating, register the models with:
docker compose exec api sleeper seed-modelsFor non-container installations, run sleeper seed-models. No database migration or dependency upgrade is required.
Validation includes regression tests, the full test suite, lint/format checks, lockfile consistency, and source/wheel builds. Live vendor API calls were not tested.
Full changelog: v0.1.5...v0.1.6
v0.1.5
What's Changed
- Add Jev decision support and bump version to 0.1.5 by @willjohnson in #30
Full Changelog: v0.1.4...v0.1.5
v0.1.4 — Render deploys and provider reliability
What’s new
Deployment
- A documented Render deployment path:
render.yamlfor api, worker, Postgres and Key Value, with a walkthrough covering bootstrap andPUBLIC_BASE_URL. Object storage stays external (Cloudflare R2 or any S3 endpoint), since Render has no S3-compatible service. - The OIDC redirect URI is built from
public_base_url, so per-tenant login works behind a real hostname.
Provider failures are retried, not blamed on the agent
- An aggregator that loses its upstream mid-generation reports it inside an HTTP 200 (
finish_reason: "error"). That now retries like any 429/5xx, instead of failing the job. A model that genuinely misbehaved still fails, since retrying that only spends the budget again. - A model that refuses forced tool choice falls back to asking for the schema in the prompt, so an agent with an output schema is no longer unusable on that model.
- A job whose worker was shut down mid-run is requeued rather than marked failed.
- An eval run the provider never let finish ends
erroredwith no pass rate, so a hiccup cannot become the baseline a real regression is measured against, or make a pending memory version look like it failed its gate.
Cost accounting says what it doesn’t know
- OpenRouter's real per-generation charge is taken from OpenRouter, since no static table can price an aggregator.
- A model with no price records the gap instead of a silent zero, which would leave spending limits inert.
- A cost missing a generation's record is flagged (
cost_incomplete) rather than reported as though it were exact.
Observability and fixes
- Tool calls are recorded in the job trail, so a finished job can answer whether the model actually read its source.
- Store listings come back in the coordinate system the tools accept.
- A model row's provider must match its model string, caught at registration instead of at job time.
- The create-agent form publishes a complete first version.
Storage images
- MinIO archived
minio/minioand removed it from Docker Hub. CI and compose now usecgr.dev/chainguard/minio, pinned by digest: the same server, built from a fork that carries the CVE fixes, rebuilt daily. Langfuse moved to the same image (langfuse/langfuse#10488).
Upgrade
No migrations in this release.
Compose users get a new storage image; docker compose pull before up. The image is pinned by digest, so updating it is deliberate. Deployments that already use external S3 (R2 on Render) are unaffected.
Verification
314 tests passed.
v0.1.3 — Human work inbox
What’s new
- A unified human work inbox for memory approvals and agent escalations.
- Opt-in agent-to-human escalation with notifications through the team’s existing notification channels.
- Memory approval actions stay synchronized with inbox work items.
- New work-item API endpoints and admin inbox UI.
- Sleeper Service is now documented as part of the CI Everywhere paradigm.
Upgrade
Apply the included Alembic migration before starting the updated application.
Verification
291 tests passed, including focused coverage for the human work-item workflow.
v0.1.2 — the admin console
Everything the API can do, the pages can now do. The gap this closes was
81 API routes against 36 UI routes: an operator could build and ship an
agent from the browser, but not create the user who would run it, not set
the credential it bills to, not give it a webhook to answer, and not put a
file where a job could reference it.
Where each surface landed follows the gates rather than one settings page,
because a page whose panels answer to different roles cannot state one
honest rule at the top:
/ui/t/{id}/settings tenant admin — system prompt and settings blob,
tenant provider credentials, OIDC config,
tenant-wide invoke keys
/ui/t/{id}/files any tenant member — upload, list, download
team page owner — Apprise alert channels, team credentials,
team-scoped invoke keys
agent page event sources, agent credentials, memory history,
and delete while nothing has ever run on it
/ui/users superuser — create users, issue and revoke keys
/ui/account anyone — rotate your own key
/ui/tenants superuser — create tenants
Tenant- and team-scoped invoke keys had been parked on the API "until
there is a settings section with somewhere to put them". This release
builds that section, so the reason expired and they came with it.
Validation reuses the API's own rules — validate_hooks_settings,
sniff_content_type, the outbound URL checks — so the pages cannot accept
what the API would reject, or become a way around a gate. Two places
diverge deliberately: file downloads are forced to attachment with
nosniff, because the UI serves uploaded bytes from the origin holding the
session cookie and the API has no cookie to steal; and an emptied settings
field clears the blob, because a form posts whole state where a PATCH does
not.
283 tests green.