Repository navigation
Releases
agent-runtime 1.8.6
Faster codex tasks. Each codex container keeps one codex instance warm while it has a free task slot. A new task with the same settings uses it: median 2.26 s against 4.17 s in 1.8.5, about 46% faster, with the same token usage.
- Warm instance switch: new "Warm instance" switch in the console (codex only, on by default), API field hot_spare.
- New event: a log event with op: "codex_spare" on each task says whether the warm instance was used and why not.
- app-server: codex now runs on codex app-server. Event shapes in the API and console are unchanged.
- Fixes: cancel and timeout now work while codex starts, and failed or cancelled starts no longer leave processes behind.
agent-runtime 1.8.5
- Adds an optional Codex setting, "Progress updates": the agent posts one short sentence before each step, saying what it's doing.
- Each update is written in the user's language and arrives as a new progress event, so apps can show it to users as a status update.
- It works for both structured (JSON) and plain text tasks, and the final answer keeps exactly the format the task asked for.
- It can be turned on per container or per message. It is off by default; when off, the agent behaves the same as 1.8.4.
- The console shows progress updates in the task timeline and event feed, and has a "Progress updates" switch next to "Reasoning summaries".
agent-runtime 1.8.4
- Adds an optional Codex setting that sends short summaries of the model's reasoning as task events.
- These events can be shown to users as status updates.
- It can be turned on per container or per message.
- It is off by default. When it is off, the agent behaves the same as 1.8.3.
agent-runtime 1.8.3
New: codex tool toggles. Choose which codex tools an agent can use: web_search (on by default), image_generation, view_image, multi_agent and goals. Set them per agent or template, or override them for a single task.
Upgrade notes
- Existing codex agents keep working as before (web search on). Migration 0029 handles this.
- Toggles only take effect on containers running agent image 1.8.3.
- API change: for codex, "tools": [] now turns all tools off. Leave the field out to get the default.
Rollback: move codex containers back to the previous image before running alembic downgrade 0028_env_vars.
Fixes: the model picker lists all OpenAI models in one section. Switching an agent's or template's driver now resets its tools.
agent-runtime 1.8.2
Agent image 1.8.2
- New model: gpt-6-astra on the codex and opencode drivers. gpt-6-sol and gpt-6-luna are also available on codex.
- Updated tools: Codex 0.152.0 → 0.157.1, opencode 1.17.18 → 1.18.32. Claude Code is unchanged at 2.1.212.
- Removed models:
- Old OpenAI models: gpt-4, gpt-3.5, o1, o3-mini, o4-mini, gpt-5.1/5.2-codex.
- From the codex driver: gpt-5.4, gpt-5.4-mini, gpt-5.3-codex-spark.
- Retired free Zen models.
- Smaller prompts on codex: about 600 fewer prompt tokens per turn.
- Upgrade: set AGENT_IMAGE_TAG=1.8.2; there are no migrations. Move any agent still on a removed model to a current one.
agent-runtime 1.8.1
Fixed
- Codex agents on GPT-5.6 had no shell tool. The 5.6 models (
gpt-5.6-sol,-terra,-luna) aretool_mode: code_mode_only, which needs a second binary,codex-code-mode-host, that the image never installed. Code mode fails closed, so those agents ran with no tools at all. Both binaries now ship from the sameCODEX_VERSION, in full and slim (+63 MB). - Event durations were measuring the database, not the agent. Events were stamped at DB-write time and tasks ended with
datetime.now(); both now use the agent's own timestamp off the SSE frame.
Added
- Per-step durations in the event feed and chat timeline, task duration in the dashboard activity feed.
Upgrading
make image / make image-slim, then recreate the control plane to pick up AGENT_IMAGE_TAG=1.8.1. Running containers stay on their old image — move one with scripts/recycle_container.py <id>. Not pushed to any registry.
agent-runtime 1.8.0
What's new
- codex 0.152.0. The agent image ships the current Codex CLI (was 0.144.3).
- Barebones codex runs. Every codex exec now switches off the CLI's ChatGPT-account
plugin sync, apps (connectors), analytics, OpenTelemetry export, image generation,
view_image, personality, and multi-agent tools. None of these serve a headless
sandboxed agent. Effect on a trivial task: the post-turn tail that waited on codex's
shutdown network calls drops from 0.3-7 s to ~0.02 s, prompt size falls from ~11.2k
to ~7.2k input tokens, and startup network calls go from 22 to 1. web_search stays
enabled; it is the agent's built-in web access. - System prompt delivered once. The agent's system prompt now reaches codex as
developer_instructions (a developer-role message injected once per session) instead
of being written to AGENTS.md and prefixed to every task's stdin. That removes the
doubled prompt on single tasks and the per-turn growth in resumed sessions, and it
ranks the prompt above user-level context. Upgraded agents have their old AGENTS.md
removed automatically.
Upgrade notes
- Set AGENT_IMAGE_TAG=1.8.0. Recycle existing agents to move them onto the new image.
agent-runtime 1.7.0
What's new
- Enforcement on CLI drivers. codex, claude-code, and opencode now validate the agent's final answer against your schema. On a mismatch, the platform resumes the agent's session with the validation error and retries (up to 3 attempts total, all within the task's existing timeout). If the output still doesn't conform, the task fails with error code invalid_structured_output — a completed task always means valid output.
- Native enforcement where available. When your schema fits the providers' strict subset (additionalProperties: false, all properties required, no $ref), the schema is also passed natively to the CLI (codex --output-schema, claude --json-schema) for constrained decoding — better first-shot validity, fewer retries. Schemas outside the subset are never rejected; they simply rely on the platform's own validation. Either way, the platform's validator is the guarantee.
- Server-side validation at submission. Invalid requests now fail fast with 400 invalid_output_schema (missing or malformed JSON Schema) or 400 structured_output_unsupported (driver can't honor the contract) instead of running the task and returning unvalidated text. Applies to both POST .../tasks and POST .../tasks/from-prompt.
- Retry visibility. Each failed validation attempt emits a warn-level event in the task stream (structured output invalid (attempt N/3): …), so you can see why a structured task took extra round-trips.
- Console. The structured-output contract in the Submit Task form is now available for all drivers.
Under the hood
- All drivers share one enforcement core (tolerant JSON extraction, jsonschema validation, session-resume retry loop) — no per-driver behavior drift.
- Claude Code CLI in the agent image bumped to 2.1.212 (versions before 2.1.205 silently ignored invalid --json-schema arguments).
- Robustness hardening: JSON extraction is no longer confused by bracketed prose ([INFO], [1]-style citations) around the answer, and near-strict schemas that would be rejected by provider-native endpoints correctly fall back to platform-side validation instead of failing the task.
agent-runtime 1.6.0
New: api driver — high-throughput single-call tasks
- One direct LLM API call per task: no tools, no subprocesses, no workspace writes. Built for pure-inference workloads (classify, extract, rewrite) at high parallelism.
- Structured output enforced against the task's JSON schema with a bounded retry loop (validation errors are fed back to the model; at most 3 calls).
- api-driver containers default to 32 concurrent tasks (new tenant limit api_driver_max_workers); other drivers keep their existing cap of 4.
- Works with every API-routable model (Anthropic + OpenAI-compatible gateways); codex-family Responses-only models are correctly excluded.
- Also fixed for all drivers along the way: rehydrated/recovered containers previously came back with a worker cap of 1 regardless of configuration.
Web tools: reliable search and a new hosted reader
- web_search now uses Exa as its primary provider when a deployment sets EXA_API_KEY, falling back to the local SearXNG and, as a last resort, the keyless Wikipedia search API. Every degradation is labeled in the tool result so the model knows which tier answered.
- New web_read tool: fetches a URL as clean text via Exa's hosted reader (handles JS-heavy pages, works in slim containers with no chromium), degrading to local text extraction when unkeyed.
- SearXNG configuration overhauled: the engine set is now actually restricted (the previous config silently left every default engine active, including permanently CAPTCHA'd ones), engine suspensions recover in 60s instead of minutes, and per-engine timeouts tolerate slow container DNS. Result: searches return usable results instead of guaranteed-empty degraded responses after the first burst.
Configured system prompts now reach CLI drivers
- The container config's system_prompt was silently ignored by the codex, claude-code, and opencode drivers (vanilla was unaffected). Each now delivers it through the CLI's native mechanism:
- claude-code: --append-system-prompt (augment) or --system-prompt (replace) — a true system-prompt equivalent.
- opencode: wired into the config's instructions list, which opencode merges into its system prompt.
- codex: materialized as $CODEX_HOME/AGENTS.md and reinforced as a standing-instructions block above every task prompt (codex has no system-prompt flag; its harness prompt cannot be replaced, so both prompt modes behave as augment).
- Verified live end-to-end on opencode (a French-only configured prompt produced French answers to English questions through the full stack).
Agent loop and prompt efficiency (vanilla)
- Loop breakers: byte-identical duplicate tool calls and consecutive-failure streaks now stop or redirect a stuck task instead of burning the budget.
- Tool input schemas are no longer duplicated into the system prompt (they already travel in the API tools parameter). With MCP servers attached this had inflated the per-call baseline ~10× (18.6k → ~1.8k tokens plus one schema copy), resent on every iteration.
agent-runtime 1.5.0
This release turns the barebones (vanilla) driver from an Anthropic-only tool
loop into a full platform citizen: it now runs OpenAI and opencode Go models,
loads tenant skills with progressive disclosure, attaches remote MCP servers,
and survived a works-as-expected audit of all nine built-in tools.
Multi-provider models on the barebones driver
- New OpenAI Chat Completions adapter and a per-model LLM router. Bare
claude-*ids keep going to Anthropic;gpt-*/o*ids go to OpenAI with
the tenant's OpenAI API key;opencode-go/*ids go to the opencode Zen Go
gateway with the tenant's opencode key (MiniMax/Qwen over the
Anthropic-compatible endpoint, GLM/Kimi/MiMo/DeepSeek over chat completions). - The model catalog now offers barebones for every OpenAI api-key chat model
(Codex-family ids excluded — they are Responses-API-only) and all opencode
Go plan models. Subscriptions still don't apply to barebones; it is
API-key-only by design. - Tool-call translation handles parallel calls, malformed tool-call JSON
(dropped with a visible note instead of crashing the task), empty
completions, and null token usage from quirky gateways.
Skills on the barebones driver
- Attach tenant skills to barebones containers; they materialize into the
container workspace and surface to the model as a name + description list.
Full instructions load lazily through a newskilltool (Claude-Code-style
progressive disclosure), with YAML frontmatter stripped,
${CLAUDE_SKILL_DIR}substituted so portable skills resolve their own
files, an "already loaded" note instead of duplicate copies, and a generous
dedicated size cap so a large skill is never silently truncated. - Bundled skill files (scripts, references) are readable from the model's
bash/python/read_filetools — including from the sandboxed
subprocess side (see Fixes). - Skill failures never fail the task: invalid bundles are skipped with a
warning event.
MCP on the barebones driver
- Attach remote MCP servers (streamable HTTP) to barebones containers. Remote
tools join the loop asmcp__<server>__<tool>alongside built-ins, with
per-server failure isolation (one unreachable server degrades, never
aborts), per-call timeouts, and guaranteed connection teardown on every
task outcome. - Bearer and custom-header auth supported; secrets are never persisted or
logged, matching the platform's existing credential handling.
Hardened built-in tools
- All nine built-in tools audited end to end; eight crash paths on
missing/malformed model inputs fixed, each with a regression test. - Uniform tool-result size cap (
TOOL_RESULT_MAX_CHARS, default 30k) with an
explicit truncation marker, plus a crash guard so no tool exception can
kill a task. - New end-to-end test sweep drives every built-in tool through a live
barebones loop in CI's integration harness.
Console
- Skills and MCP pickers are now driven by driver capability flags from the
API instead of a hardcoded driver list — barebones shows both pickers, and
future drivers inherit this for free.
Fixes
- Agent image was missing the
mcppackage — MCP tasks on any driver
path that needed the SDK would have failed at import. Now bundled. - Sandboxed subprocesses couldn't reach skill files (
Permission denied
frombash/pythonwhile file tools worked): the runtime dir is now
traverse-only (711) with its private children (event logs, task metadata,
transfer spool) explicitly locked to 700. Existing volumes self-correct on
container restart. - MCP connections no longer leak when a session continuation fails fast;
detached skills are cleaned from the workspace on the next run. - OpenAI wire requests stay valid on empty assistant turns and null usage.
- Dev stack
AGENT_IMAGE_TAGwas pinned to an old image; version references
(Makefile, deploy examples, dev env) are now aligned at 1.5.0.