Skip to content

Update trigger - #4

Open
tylerc-govsignals wants to merge 2482 commits into
GovSignals:ConProgramming/two-phase-deployfrom
triggerdotdev:main
Open

tylerc-govsignals wants to merge 2482 commits into
GovSignals:ConProgramming/two-phase-deployfrom
triggerdotdev:main

Conversation

@tylerc-govsignals

Copy link
Copy Markdown
Collaborator

Closes #

✅ Checklist

  • I have followed every step in the contributing guide
  • The PR title follows the convention.
  • I ran and tested the code works

Testing

[Describe the steps you took to test this change]


Changelog

[Short description of what has changed]


Screenshots

[Screenshots]

💯

1stvamp and others added 29 commits November 10, 2026 14:11
…nment

Admin template builds can now select deployments by organisation,
project and
environment type, accept deployment short codes, and wait for slow
builds
to finish, streaming each result as it completes.

Mono-RevId: 3ba72ec11729b3504ea4023252fcc03f20fb0968
Mono-RevId: a8ecd733ed7496919dee9d72f010d949bafe9514
…ne starts

The run engine can detect a second start of an attempt that is already executing (`RUN_ENGINE_DUPLICATE_ATTEMPT_START_CHECK`, logged by default).

Mono-RevId: 39d7ef47a458cd9ade110fe244ccca37a85e520e
Mono-RevId: 2f0eb0ac720ccb1283629d887348feeeeedfe3db
Remove the OIDC private beta and request-access banner now that OIDC is
available to all organizations.

Mono-RevId: 50d79a55977acb8e3279b49e7157280482a121d2
`chat.agent` now validates the messages it accepts from chat clients. Tool approvals and client-side tool outputs keep working, including after a run restarts.

Mono-RevId: 2b23a74ed8d6c0df211dada502b2d70d1ec532ed
Project JWT minting now rejects expiration times longer than 24 hours,
using the same validation as the API-key JWT mint route. Delegated
tokens remain limited by their own expiry.

Mono-RevId: 6ffd715d174118c1262296258c240b92c62cc766
Stops the runs, sessions and webhook-deliveries replication services
from reporting an older replication position after a keepalive. On an
idle database this made `pg_stat_replication` lag appear to grow without
bound, even though the slot itself was current.

The acknowledge loop compared the keepalive LSN (unpadded) with the last
commit LSN (zero-padded) as strings, so they never matched and the older
commit LSN was re-sent. LSNs are now compared numerically, and an ack is
skipped when the last acknowledged LSN is already at or past the latest
commit:

```ts
if (
  runtime.lastAcknowledgedLsn &&
  lsnToUInt64(runtime.lastAcknowledgedLsn) >= lsnToUInt64(runtime.latestCommitEndLsn)
) {
  return;
}
```

Each service has an integration test against real Postgres that fails
before the fix (flush LSN moves backwards) and passes after.

Mono-RevId: 3cbe8f510438e51171aa6577edafb525b30bef1a
Mono-RevId: 269054dce62d10fb948fb4f7f66d44a61d115c6a
Mono-RevId: 4e287271218721fd41a59cad5e99fb08b3dd5032
…min routes

The admin API routes used by the billing service now only accept the
billing service secret. Personal access tokens are no longer accepted on
these routes.

Mono-RevId: 5da5172244268b66f6122b340803cb6178f6a4f2
Adds an `ADMIN_API_ENABLED` setting to the webapp (default `true`). When it is `false`, `/admin/api` routes return 404.

Mono-RevId: 9a88be97cebca6caa780693a6bf9724b6ab42d5d
…it starts, in shadow mode

Adds a check that a run's version finished deploying before the run
starts. A run locked to a version can otherwise start while its
deployment is still in progress, or after the deployment failed.

```ts
await myTask.trigger(payload, { version: "20261102.2" });
```

The check ships in shadow mode, so runs start exactly as before and the
outcome is only logged and counted. Self-hosters can set
`RUN_ENGINE_DEPLOYMENT_STATUS_MODE=enforce` to turn it on: a run locked
to a version whose deployment failed, timed out or was canceled then
fails with a clear error, and a run whose version is still deploying
waits and starts once the deployment finishes. Runs on deployed versions
are unchanged in every mode.

Mono-RevId: 738f996a66f952968216c6b13395ba97aa7dd44a
Mono-RevId: 3d8fd3245ff96259a408de4de1335c95745a731d
…med out

Mono-RevId: d7f666b693185970dc4a3111cc19a56e68bdbd44
Mono-RevId: a33f3c91d528306096a18f7cd823cd2b22ea78e2
Fix the Logs page redirecting to Tasks when Logs is visible in the
sidebar.

Mono-RevId: 139133a72d1e6174523e0c865475f5e0dbd9840b
Invite, email change confirmation and bulk action completion emails now build their links from PUBLIC_APP_ORIGIN when it's set, falling back to LOGIN_ORIGIN.

Mono-RevId: dd89dcdc662faf008586a65bc52b996c69335781
Mono-RevId: 8930af65f06cae345ecdd02653ae4a17da00c7e9
Mono-RevId: 0cfc6fb54ea212ccc24e1a78a0a8377a7533a9a1
Require a dedicated dev worker scope on the dev engine routes and worker
registration, and limit dev worker socket subscriptions to runs in the
authenticated environment.

Mono-RevId: 5ae76c2ceaf05570b8e8455d5a2d114eb403b8f4
Alert charts and evaluations are no longer restricted by the plan's
query period.

Mono-RevId: 3143072e6fbc2ca6a85c514689d188d5497a8ac1
Mono-RevId: 31ac74dd5203497569909231ac7483bc0484f4c3
… at write time

Validate checkpoint locations before they are stored, and report a discarded checkpoint back to the worker instead of acknowledging it as stored.

Mono-RevId: 7f0a5e22aaf1b7051a8681b5ca521e460b93499b
… its restore

Checkpoints can carry the digest of their image manifest, recorded when the checkpoint is taken and handed back with the restore, so the restoring side can verify what it copies.

Mono-RevId: 59bfc7efc9aa64a8eafb24ba0f120ec5c580f78e
Mono-RevId: eab761d4e61f9923c7f3dc0a4dff8bb36423ddb7
Alerts can now be grouped into one shared alert. Select several alerts,
choose whether the group raises when any or all of them fire, and give
the group its own channels and expiry. Member alerts keep evaluating in
the background, but the group owns the shared status and notifications.
Ungroup at any time to return them to standalone alerts.

Mono-RevId: cd0dd4dd0d2ab2c54020ae92cc6c7049a63425a9
Mono-RevId: edd849a674edd3a308c2f8e979ad6ec908395989
…longer needs them

Adds an optional cleaner to the supervisor that deletes microVM suspend
snapshots from a node once their run has finished or has taken a newer
checkpoint. It is off by default.

Mono-RevId: 789f50faae62e43dafbbdc3720b4c5b7932dd35c
pcnc and others added 30 commits November 17, 2026 12:53
Reads of environment variables from the API, the CLI, the build server
and the dashboard are recorded in the audit log. Repeat API reads by the
same caller are recorded once per window. Values are never recorded.

Mono-RevId: 10bd7c124eaa20575596cd76bb6ea8951d251945
Added Watch to widget menus, making it easier to create alerts directly
from queries. Unsupported queries open the Dashboard Agent for help
adapting them.

Mono-RevId: 5bf1797f72dcb0481b485df3957377599061d354
…ogging

Hosted MCP runs now use a distinct `mcp-remote` trigger source. MCP logs
include environment type, denial reasons, and write confirmation
decisions for better visibility.

Mono-RevId: c465970bda74a5c50dd46b26b0be5bfcffdf3aa6
## Summary

Code samples across the docs pointed at models that no longer work or are about to stop working: `gpt-3.5-turbo`, `gpt-4-turbo` and `claude-sonnet-4-20250514` are already shut down, `claude-sonnet-4-5` retires on November 30, 2026, and `gpt-4o` / `gpt-4o-mini` shut down in January 2027. Copying those samples today gives errors.

This swaps model IDs in samples and prose to current models:

| Old | New |
| --- | --- |
| `claude-sonnet-4-5`, `claude-sonnet-4-6`, `claude-sonnet-4-20250514` | `claude-sonnet-5-5` |
| `claude-haiku-4-5` | `claude-haiku-5-5` |
| `claude-opus-4-6` | `claude-opus-5-5` |
| `gpt-4o`, `gpt-4o-mini`, `gpt-4-turbo`, `gpt-3.5-turbo` | `gpt-6-luna` (OpenAI's listed replacement for `gpt-4o`) |

Alongside the swap:

- Prompt `config` examples in `ai/prompts` and `ai-chat/backend` used `temperature: 0.7`, which Claude 5.5 and gpt-6-luna don't support (the AI SDK drops it with a warning), so the examples now use `maxOutputTokens: 1024` instead.
- The prompt caching page now says cache reads cost about 5% of base input on Claude Sonnet 5.5 and Opus 5.5 (it said 10%).

Claude 5.5 models also behave differently from 4.5 in ways a plain ID swap doesn't cover, so a few samples changed further: thinking is on by default and omits its text unless you ask for `display: "summarized"`, a conversation can't end on an assistant turn, and thinking blocks are bound to the earlier system prompt, tools and messages. The compaction summarizers now end with a user instruction, the head-start handler and agent share one system prompt, the Nango sample reads text blocks by type, and the native compaction and per-turn tools pages note where editing history or tools can invalidate thinking blocks.

Every changed sample was run against the live APIs with current packages (`ai` 7, `@ai-sdk/anthropic` 4, `@ai-sdk/openai` 4, `@anthropic-ai/sdk`, `openai`, `@trigger.dev/sdk` 4.8.5), with Anthropic's thinking-block check turned on the way new accounts get it. That surfaced samples that were already broken on current versions, now fixed too:

- The agent pattern pages put `{ role: "system" }` inside `messages`, which `ai` 7 rejects; they now pass `system`.
- Tool samples on the schemaTask and migration pages and the Slack agent had no step limit, so they stopped after the tool call with no reply; they now set `stopWhen`.
- The native compaction example strips reasoning from the messages it keeps after a summary.
- Errors & retrying fell back to a Llama 2 model on Replicate that no longer exists; it now falls back to Claude Haiku 5.5.
- A few missing or wrong imports.

A multi-turn `chat.agent` built from these docs (tools, summarized reasoning, two compactions) also ran end to end in a dev environment without errors.

Left unchanged on purpose: the AI chat changelog (historical entries and a benchmark measured on a specific model), example-project pages that mirror code in other repos, and the DALL·E 3 image example, which needs a code change rather than a model swap and will follow separately.

Sources: [OpenAI deprecations](https://developers.openai.com/api/docs/deprecations), [Anthropic model deprecations](https://platform.claude.com/docs/en/about-claude/model-deprecations).
Mono-RevId: 5ffe77cb272bec674b7edb19de8a3a2ae47a6ff0
…egional CA

Mono-RevId: cec806bfb92fa95d856b3e04823de830f9593143
Sandbox task runners now exit after their run instead of accepting
another run through warm start. Standard task runners keep their
existing warm-start behavior.

Mono-RevId: 50fcda0dc83ba05523c71ac3cd60d311e4d13e15
Restrict sandbox workload tokens to their owning run and prevent sandbox
runners from receiving warm-start work.

Mono-RevId: 221a1f25000ebeca49e9b69b33035441c5a1e271
… proxy runs

Mono-RevId: 58fe5a6866d78745418c1fedac7f09e58e50c078
Mono-RevId: dd9dd39f1b59d7626edbc016ec5a622314adf1ad
New project UI changes

Mono-RevId: 5bd493e069a7dd510860198e15d6ff7a9a4c63fe
The Tasks page now shows the current page and total pages beside the top
controls, with numbered pagination and the filtered task count below the
task list.

Mono-RevId: 06b7d3b42165c514088264f6a0ce6ec04ae66ff7
Scheduled tasks now show their schedules in a tab beside Runs, with
pagination and the existing schedule inspector. The sidebar shows
details for a single schedule or a link to the full list. Next-run times
include an absolute timestamp, and cron timezones are visible.

Mono-RevId: 5542efecdecdfbfe7287c24dc6767d66889384e3
…variables and raw SQL

The run-ops legacy guard now also reports run-graph model calls through
the run-ops client, transaction handles and model accessor variables,
and raw SQL that names a run-graph table.

Mono-RevId: d9e15b1f7be9113ff3dc40675996ca5f87792d76
…t can't show

The session Trace tab no longer scrolls on its own: new spans and runs
below the view show a pill instead. Runs whose spans can't be shown
explain why in the run panel, search shows which runs it covered, and
run reasons only show when notable.

Mono-RevId: 481cc4ee37e9e761ca491c35ab8f1feace0c20ef
…tabase client and the root scripts

The run-ops legacy guard now also checks the run engine, the schedule
and webhook engines, the RBAC, SSO and LLM model catalogue packages,
core and the repository's scripts, and the PR checks run it for changes
to any of them.

Mono-RevId: 8f5ebbb57ab91cb02bae71cd254ec1b2c33860f7
…l on off-target or stale markers

The run-ops legacy guard now requires each allowed direct access to run
tables to carry a marker naming its owning project and reason, and fails
on a marker that covers nothing or whose access has gone.

Mono-RevId: a3e14f188c07bcf9992ae49f4a9048c9f64b3efd
Fix a race where a run released from waiting for a version could be left queued with nothing to pick it up.

Mono-RevId: 85646dc42cdb8101363aca8df74ab7b260acd265
…umbers

The Tasks page now has one pagination bar fixed below the table, with
the task count on the left and page numbers on the right, so the
controls no longer move when a page has fewer rows. Numbered pagination
across the dashboard no longer repeats a page number or shows "…"
between consecutive pages on long lists.

Mono-RevId: d4a956caaf7384f0d1f98ba380343d9cf35f95e2
Mono-RevId: d985f6cadd5210d5109dc0ff201a1725a3235682
Mono-RevId: f5a421984ed7a6476065ef6ba1579def4b13031a
…h run databases

Adds tests that pin how the run store's routing layer reads a run that
must exist when runs are split across two databases.

Mono-RevId: 7621b41586496c09a5cc4df7f98886d00980ca76
Preserve existing secret proxy execution bindings across supervisor
restarts.

Mono-RevId: c2004129a498d1ef0619da091b6a7bc68044aa9c
…nused model support

Removes run store methods and model support that nothing calls;
behaviour does not change.

Mono-RevId: a2d2ccdf3cd66337109834b7c53a5f756d21daf8
…en and additional key events

Mono-RevId: 4cd0d493714f916cb3d759787a4a5e2ad728c3a9
Fix `chat.agent` on AI SDK 7 sending the model the same input on every
step. Once a step's `prepareStep` returned messages (a steering message,
a mid-turn `chat.inject`, context compaction, or your own
`prepareStep`), later steps no longer included the newest tool results,
so the model kept repeating the same tool call until the step limit.

If you call `chat.compact()` from your own `prepareStep`, pass it the
`responseMessages` your `prepareStep` receives, so steps after a
compaction are rebuilt correctly on AI SDK 7.

Mono-RevId: af1564a0fa10d3b1dded2917bf45dfc198e47ad9
…ost-turn hooks in mockChatAgent

`chat.agent` now passes `chatId`, `turn` and `clientData` to the
callbacks an agent can call a model from. That includes the compaction
`summarize` that runs through `chat.MessageAccumulator.prepareStep()`
and a `chat.createSession` turn's `prepareStep()`, `onCompacted`,
`onValidateMessages`, and the `pendingMessages` `onReceived` and
`onInjected` callbacks. So a mock model passed through `clientData` in
`mockChatAgent` tests reaches all of them, and the suite can run
offline. The `mockChatAgent` turn methods (`sendMessage`,
`sendRegenerate`, `sendAction`) now also wait for `onTurnComplete` and
the transcript save to finish, so tests can assert on what those hooks
recorded without a sleep.

Mono-RevId: cb1eee5867c4ab30b1e2684eff97d7492078e93f
…unEngine.getWaitpoint

Removes two internal methods that nothing called: the run store's
`updateManyWaitpoints` and the run engine's `getWaitpoint`.

Mono-RevId: e96257f2662a8c9b7b026bb4f591bf5ddd90ba7a
Removes an internal run store loader for a run relation that nothing
requested.

Mono-RevId: c3cbde4f5130bbd3f5b96f7ad7805058f93965d7
…ng primary through findRunOrThrow

Removes an internal run store read method that nothing outside the store
used.

Mono-RevId: 125b9d8e7163809f00fee48cb24e38bab7c34c1d
## Summary
12 new features, 23 improvements, 23 bug fixes.

## Highlights

- Allow an explicit audience when signing and verifying JWTs and
user-actor tokens.
([`01337977e`](0133797))

## Improvements
- Deploys on the `bun-1.4` runtime now compile the run worker and your
task files into a Bun executable with bytecode, so runs start faster.
`import.meta.url` and `__dirname` still point at your files under
`/app`, so files you read next to your code must be in the image as
before (e.g. with `additionalFiles`). Set `build.bunBytecode: false` in
`trigger.config.ts` to turn compilation off.
([`8130e5e73`](8130e5e))
- `trigger dev` and `trigger deploy` no longer crash with `spawn ...
ENOENT` on bun runtimes when Bun isn't installed at `~/.bun`. The CLI
now falls back to `bun` on your `PATH` (Homebrew, mise and similar
installs), and if Bun still can't be started, `trigger dev` reports a
readable error and `trigger deploy` skips local skill discovery with a
warning and carries on.
([`15797c9f2`](15797c9))
- Bundled task files now keep their project layout (for example
`/app/src/trigger/tasks.mjs`) when the CLI runs from outside the
project, such as `npx trigger.dev deploy`. Before, they could land under
a path containing the project's location, away from the files
`additionalFiles` copies next to them, so reads relative to
`import.meta.url` failed.
([`0c03b59b9`](0c03b59))
- Show CLI command errors in the same error format as other CLI errors,
and keep them visible with `--log-level error`.
([`c515c3f74`](c515c3f))
- Deploys now report a warmup command that loads your config and task
files offline without running any task, so the platform can prepare
faster run starts for your deployment.
([`5479177af`](5479177))
- Generate the machine ID and the managed run controller ID on first use
instead of at import, so importing these modules draws no randomness and
a process restored from a warm start does not reuse IDs drawn before it
was saved.
([`a8ac49cd3`](a8ac49c))
- Deploys on the `node-24` and `node-26` runtimes now build a Node
compile cache for your bundled code during the image build, so runs
spend less time compiling your tasks and their dependencies before they
start. `node-22` deploys are unchanged.
([`cb1e7362f`](cb1e736))
- Report a run whose process the kernel killed for running out of memory
as `TASK_PROCESS_OOM_KILLED`, rather than as a generic exit after
`SIGKILL`, when the out-of-memory kill is counted in the run's cgroup.
([`7a2a7dc80`](7a2a7dc))
- Deployed task processes now receive secret environment variables only
with the run, instead of also in the environment the process starts
with.
([`489a6dff3`](489a6df))
- Deployed task processes now start with the environment variables that
Node, OpenSSL, libuv or the dynamic loader read only at startup (such as
`LD_PRELOAD`, `OPENSSL_CONF`, `NODE_ICU_DATA`, `TZ` and `UV_*`), and the
run's logs warn, by name only, about other startup-only variables (such
as `NODE_OPTIONS`, `MALLOC_*` or `GLIBC_TUNABLES`) that arrive too late
to take effect.
([`60d1774df`](60d1774))
- Builds from `trigger dev` and `trigger deploy` now use esbuild 0.28.
TypeScript parameter properties are now emitted as class fields, and
source map paths are URL-encoded.
([`af2098570`](af20985))
- `webhooks.notion()` endpoints accept Notion's verification request.
While the endpoint has no signing secret, the `verification_token`
Notion sends is stored as its signing secret, so you can paste it back
into Notion and later deliveries verify with it. Redeploy with this
version for existing Notion endpoints to pick it up.
([`4288ea912`](4288ea9))
- Support scoped run tokens without an environment key, including
offloaded task inputs and stream waits. Explain why a run has no usable
credential for an operation.
([`ed0147895`](ed01478))
- Public-token helpers and chat token creation now prefer server minting
for root API keys too, avoiding the legacy claims endpoint on compatible
servers. Trigger, batch, and waitpoint responses missing a token also
try server minting first.
([`90d677b12`](90d677b))
  
Root-key callers retain local-signing compatibility only when the mint
endpoint is missing or an older server specifically rejects an empty
scope list. Named API keys and run tokens never self-sign. Authorization
failures, expiration validation errors, rate limits, and server or
connection failures do not activate the fallback.
  
Server minting with a root key signs with that same authenticated key,
including during its rotation grace period, rather than with its
replacement. Tokens stop authenticating when that key's grace period
ends or it is immediately revoked. Deploy this server update before
upgrading the SDK if you rely on minting during root-key rotation; older
servers can still reject grace-window keys.
  
Server-minted tokens use the API rate-limit budget and follow the
server's expiration rules, including its 30-day maximum. Minting tokens
without scopes requires self-hosted v4.8.4 or later unless the root-key
compatibility fallback is available. The default expiration remains 15
minutes.
- `chat.agent` now passes `chatId`, `turn` and `clientData` to the
callbacks an agent can call a model from, so a mock model injected
through `clientData` reaches them. The compaction `summarize` that runs
through `chat.MessageAccumulator.prepareStep()` and a
`chat.createSession` turn's `prepareStep()`, `onCompacted`,
`onValidateMessages`, and the `pendingMessages` `onReceived` and
`onInjected` callbacks previously lacked it. The `mockChatAgent` turn
methods (`sendMessage`, `sendRegenerate`, `sendAction`) now also wait
for `onTurnComplete` and the transcript save to finish, so tests can
assert on what those hooks recorded without a sleep.
([`579904220`](5799042))
- `concurrencyKey` now also takes an object, so one run can be limited
by two limits keyed by different things, like a cap per session and a
cap per organization. `default` is the run's own key and every other
entry gives a named limit its own key: `concurrencyKey: { default:
sessionId, [orgLimit.name]: orgId }`. A string key works as before.
([`4961fc474`](4961fc4))

## Bug fixes
- Fix `child_process.fork()` and `spawn(process.execPath, ...)` in
compiled `bun-1.4` deploys. They started the run worker again instead of
Bun, so the child never ran and the task hung. `process.execPath` now
points at the image's `bun` binary, as it does in uncompiled deploys.
([`d63b49ef3`](d63b49e))
- Fix webhook verification for `webhooks.brex()`, which now reads the
`webhook-id`, `webhook-timestamp` and `webhook-signature` headers Brex
sends instead of `svix-*` headers, and for `webhooks.composio()`, which
now keys the signature with the secret's raw text as Composio does
instead of base64-decoding it. PagerDuty, Sentry, Vercel, Zendesk,
Attio, DocuSign and Intercom endpoints now default to pasting the
provider-issued signing secret, and Linear and Calendly offer both
generating and pasting one. Redeploy with this version for existing
endpoints to pick up the change.
([`b5b311407`](b5b3114))
- Fix `chat.agent` on AI SDK 7 sending the model the same input on every
step. Once a step's `prepareStep` returned messages (a steering message,
a mid-turn `chat.inject`, context compaction, or your own
`prepareStep`), later steps no longer included the newest tool results,
so the model kept repeating the same tool call until the step limit.
([`1a2c99034`](1a2c990))
  
If you call `chat.compact()` from your own `prepareStep`, pass it the
`responseMessages` your `prepareStep` receives (`chat.compact(messages,
steps, { ...options, responseMessages })`) so steps after a compaction
are rebuilt correctly on AI SDK 7.

## Server changes

These changes affect the self-hosted Docker image and Trigger.dev Cloud:

- Ask the Dashboard Agent to watch something, such as a deploy, a task
or a model, and it proposes alerts you confirm one by one. When the
alerts belong together, it offers to create them as one grouped alert in
a single click.
- Dashboard widget menus now have a Watch action that starts an alert
from the widget's query, and when the query cannot be alerted on as
written or the dashboard has filters applied, the Dashboard agent helps
adapt it. The Watch button in the dashboard header is removed.
- Select up to 100 environment variable values on the Environment
variables page and delete them together. The confirmation lists every
value and its environment, and asks you to type a word before anything
is removed.
- Compare the environment variables of two environments on the
Environment variables page and see only the keys that differ, with
secrets shown as the same or different but never revealed. Copy values
from one environment to the other, one key or up to 100 at a time.
- Click a key on the environment variables page to see its value in
every environment and preview branch in a side panel, with edit and
delete in place. A list showing a single environment can also be sorted
by last updated. The edit dialog now starts in the value field, and
archived preview branches are no longer accepted as targets.
- Environment variables page: a new Needs attention page lists the keys
in one environment that are empty, missing compared with your other
environments, changed in the last day, or replaceable with OIDC, with
copy, delete and ignore actions. The OIDC page is now a grid of
providers with setup in a side panel, it points at environment variables
OIDC could replace, and there are setup guides for Azure, Google Cloud,
HashiCorp Vault, Doppler, Snowflake and Databricks.
- Add opt-in per-environment encryption keys for runtime and build
environment variables, with rotation and migration support.
- Add "Ignored branches" to the Git settings: pull requests from
branches matching these patterns no longer get a preview deployment.
Patterns are globs matched against the whole branch name, like
`dependabot/**`.
- The OIDC page lists Hugging Face and OpenRouter, and the environment
variables page suggests OIDC for their API keys. Each has a setup guide
in the docs.
- Organization owners can now require multi-factor authentication for
every member, and ask for a fresh MFA check before members view
environment variables, create personal access tokens or change the
organization's security settings. Find both under Organization settings
→ Security.
- Add a hosted MCP server so AI clients can work with your projects and
runs. It is opt-in per deployment.
- The allowed IP addresses field on API keys now turns each address or
CIDR range into its own pill as you type, checks it right away with the
same rules used on save, and lets you remove entries one at a time. A
"Test an IP" button checks whether an address would be allowed by the
current list before you save.
- The assistant can now list alerts across environments and projects,
and show which alert channels are set up.
- Environment variables page: after adding a variable, the list switches
to the environment you saved it in and opens the page that holds the new
key, with a toast naming where it went. The `env` filter in the URL also
accepts `development`, `staging` and `production`. The compare view has
a copy button on each key, and each row's edit and delete buttons are
named after the key and environment for screen readers.
- Runs started through the hosted MCP server are now recorded with the
`mcp-remote` trigger source, so they can be told apart from runs started
by the local CLI MCP.
- Personal access tokens created on the dashboard can now expire on a
custom date, picked from a calendar alongside the preset lifetimes. A
token expires at the end of the chosen day (UTC), up to the longest
offered lifetime.
- A scheduled task's schedules now have their own tab next to Runs, with
every column visible. The side panel shows the schedule when the task
has exactly one, and a count with a link when it has more. Next run
shows the local time first, followed by the relative time.
- Runs start faster when the platform is busy.
- The alert group page now shows its header and member list right away,
and member charts fill in as they load.
- The alert wizard now highlights the suggested option on choice steps
instead of the first row.
- The alerts list now shows its rows before the activity sparklines
load, and the sparklines load faster.
- Alerts created from the Dashboard Agent are now limited to 200 active
alerts per project, with bounded names, scopes and notification
channels.
- The dashboard agent now asks you to confirm before resolving an alert.
- The Dashboard Agent can now look up runs, deployments and queries in
non-development environments for members whose role can see them but
can't view API keys.
- Help links (`?aiHelp=`, `/ai-help`) now prefill the chat composer
instead of sending.
- Fix dashboard pages that open as a dialog (such as adding environment
variables or a new alert) re-rendering the whole page when loaded
directly or reloaded.
- Saving an environment variable that someone else deletes at the same
moment now shows a clear message asking you to try again instead of a
database error, and the API answers with a 409. Deleting a value can no
longer remove a value that was being added to the same variable at that
moment.
- Runs that exhaust their dequeue attempts while the run or its
environment can't be read now fail with `TASK_RUN_DEQUEUED_MAX_RETRIES`
instead of staying queued forever.
- Fix waitpoint callback URLs returning authentication errors when
created inside preview branch runs, including previously issued URLs.
- Restore offloaded payload, output, and stream-wait result downloads
for API keys with packet read access.
- Keep waitpoint idempotency keys scoped to a run unless the API key can
both read and write all waitpoints.
- An organization's only Owner can no longer leave or be removed while
other members remain, so the organization always keeps someone who can
manage members, billing and roles.
- Numbered pagination no longer repeats a page number on long lists, and
the Tasks page now keeps its page numbers in one bar fixed below the
table.
- Revoking or expiring a personal access token now also ends the
environment JWTs exchanged from it and the run, batch and session tokens
minted with them.
- Runs triggered in the seconds after a deploy no longer get stuck
waiting for a version until the next deploy.
- A run released from waiting for a version is no longer occasionally
left queued with nothing to pick it up when a worker dequeues it at the
same moment.
- Fix sandbox task retries stalling when retry delays are short.
- A task that returns while its run is resuming from a wait it did not
await (for example a wait raced against a timeout or an abort signal)
can now fail right away with an error that says so. Before, the run
stayed in an executing state until its heartbeat timed out, and could
then be reported as out of memory. This is off by default: set
`RUN_ENGINE_STALE_COMPLETION_MODE=enforce` on the webapp to switch it
on.

<details>
<summary>Raw changeset output</summary>

# Releases
## @trigger.dev/build@4.8.6

### Patch Changes

- Updated dependencies:
  - `@trigger.dev/core@4.8.6`
## trigger.dev@4.8.6

### Patch Changes

- Deploys on the `bun-1.4` runtime now compile the run worker and your
task files into a Bun executable with bytecode, so runs start faster.
`import.meta.url` and `__dirname` still point at your files under
`/app`, so files you read next to your code must be in the image as
before (e.g. with `additionalFiles`). Set `build.bunBytecode: false` in
`trigger.config.ts` to turn compilation off.
([`8130e5e73`](8130e5e))
- Fix `child_process.fork()` and `spawn(process.execPath, ...)` in
compiled `bun-1.4` deploys. They started the run worker again instead of
Bun, so the child never ran and the task hung. `process.execPath` now
points at the image's `bun` binary, as it does in uncompiled deploys.
([`d63b49ef3`](d63b49e))
- `trigger dev` and `trigger deploy` no longer crash with `spawn ...
ENOENT` on bun runtimes when Bun isn't installed at `~/.bun`. The CLI
now falls back to `bun` on your `PATH` (Homebrew, mise and similar
installs), and if Bun still can't be started, `trigger dev` reports a
readable error and `trigger deploy` skips local skill discovery with a
warning and carries on.
([`15797c9f2`](15797c9))
- Bundled task files now keep their project layout (for example
`/app/src/trigger/tasks.mjs`) when the CLI runs from outside the
project, such as `npx trigger.dev deploy`. Before, they could land under
a path containing the project's location, away from the files
`additionalFiles` copies next to them, so reads relative to
`import.meta.url` failed.
([`0c03b59b9`](0c03b59))
- Show CLI command errors in the same error format as other CLI errors,
and keep them visible with `--log-level error`.
([`c515c3f74`](c515c3f))
- Deploys now report a warmup command that loads your config and task
files offline without running any task, so the platform can prepare
faster run starts for your deployment.
([`5479177af`](5479177))
- Generate the machine ID and the managed run controller ID on first use
instead of at import, so importing these modules draws no randomness and
a process restored from a warm start does not reuse IDs drawn before it
was saved.
([`a8ac49cd3`](a8ac49c))
- Deploys on the `node-24` and `node-26` runtimes now build a Node
compile cache for your bundled code during the image build, so runs
spend less time compiling your tasks and their dependencies before they
start. `node-22` deploys are unchanged.
([`cb1e7362f`](cb1e736))
- Report a run whose process the kernel killed for running out of memory
as `TASK_PROCESS_OOM_KILLED`, rather than as a generic exit after
`SIGKILL`, when the out-of-memory kill is counted in the run's cgroup.
([`7a2a7dc80`](7a2a7dc))
- Deployed task processes now receive secret environment variables only
with the run, instead of also in the environment the process starts
with.
([`489a6dff3`](489a6df))
- Deployed task processes now start with the environment variables that
Node, OpenSSL, libuv or the dynamic loader read only at startup (such as
`LD_PRELOAD`, `OPENSSL_CONF`, `NODE_ICU_DATA`, `TZ` and `UV_*`), and the
run's logs warn, by name only, about other startup-only variables (such
as `NODE_OPTIONS`, `MALLOC_*` or `GLIBC_TUNABLES`) that arrive too late
to take effect.
([`60d1774df`](60d1774))
- Builds from `trigger dev` and `trigger deploy` now use esbuild 0.28.
TypeScript parameter properties are now emitted as class fields, and
source map paths are URL-encoded.
([`af2098570`](af20985))
- Updated dependencies:
  - `@trigger.dev/core@4.8.6`
  - `@trigger.dev/build@4.8.6`
  - `@trigger.dev/schema-to-json@4.8.6`
## @trigger.dev/core@4.8.6

### Patch Changes

- Deploys on the `bun-1.4` runtime now compile the run worker and your
task files into a Bun executable with bytecode, so runs start faster.
`import.meta.url` and `__dirname` still point at your files under
`/app`, so files you read next to your code must be in the image as
before (e.g. with `additionalFiles`). Set `build.bunBytecode: false` in
`trigger.config.ts` to turn compilation off.
([`8130e5e73`](8130e5e))
- `trigger dev` and `trigger deploy` no longer crash with `spawn ...
ENOENT` on bun runtimes when Bun isn't installed at `~/.bun`. The CLI
now falls back to `bun` on your `PATH` (Homebrew, mise and similar
installs), and if Bun still can't be started, `trigger dev` reports a
readable error and `trigger deploy` skips local skill discovery with a
warning and carries on.
([`15797c9f2`](15797c9))
- Deploys now report a warmup command that loads your config and task
files offline without running any task, so the platform can prepare
faster run starts for your deployment.
([`5479177af`](5479177))
- Allow an explicit audience when signing and verifying JWTs and
user-actor tokens.
([`01337977e`](0133797))
- Generate the machine ID and the managed run controller ID on first use
instead of at import, so importing these modules draws no randomness and
a process restored from a warm start does not reuse IDs drawn before it
was saved.
([`a8ac49cd3`](a8ac49c))
- Deploys on the `node-24` and `node-26` runtimes now build a Node
compile cache for your bundled code during the image build, so runs
spend less time compiling your tasks and their dependencies before they
start. `node-22` deploys are unchanged.
([`cb1e7362f`](cb1e736))
- `webhooks.notion()` endpoints accept Notion's verification request.
While the endpoint has no signing secret, the `verification_token`
Notion sends is stored as its signing secret, so you can paste it back
into Notion and later deliveries verify with it. Redeploy with this
version for existing Notion endpoints to pick it up.
([`4288ea912`](4288ea9))
- Support scoped run tokens without an environment key, including
offloaded task inputs and stream waits. Explain why a run has no usable
credential for an operation.
([`ed0147895`](ed01478))
- Public-token helpers and chat token creation now prefer server minting
for root API keys too, avoiding the legacy claims endpoint on compatible
servers. Trigger, batch, and waitpoint responses missing a token also
try server minting first.
([`90d677b12`](90d677b))

Root-key callers retain local-signing compatibility only when the mint
endpoint is missing or an older server specifically rejects an empty
scope list. Named API keys and run tokens never self-sign. Authorization
failures, expiration validation errors, rate limits, and server or
connection failures do not activate the fallback.

Server minting with a root key signs with that same authenticated key,
including during its rotation grace period, rather than with its
replacement. Tokens stop authenticating when that key's grace period
ends or it is immediately revoked. Deploy this server update before
upgrading the SDK if you rely on minting during root-key rotation; older
servers can still reject grace-window keys.

Server-minted tokens use the API rate-limit budget and follow the
server's expiration rules, including its 30-day maximum. Minting tokens
without scopes requires self-hosted v4.8.4 or later unless the root-key
compatibility fallback is available. The default expiration remains 15
minutes.

- Fix webhook verification for `webhooks.brex()`, which now reads the
`webhook-id`, `webhook-timestamp` and `webhook-signature` headers Brex
sends instead of `svix-*` headers, and for `webhooks.composio()`, which
now keys the signature with the secret's raw text as Composio does
instead of base64-decoding it. PagerDuty, Sentry, Vercel, Zendesk,
Attio, DocuSign and Intercom endpoints now default to pasting the
provider-issued signing secret, and Linear and Calendly offer both
generating and pasting one. Redeploy with this version for existing
endpoints to pick up the change.
([`b5b311407`](b5b3114))
## @trigger.dev/linear@4.8.6

### Patch Changes

- Updated dependencies:
  - `@trigger.dev/core@4.8.6`
  - `@trigger.dev/sdk@4.8.6`
## @trigger.dev/python@4.8.6

### Patch Changes

- Updated dependencies:
  - `@trigger.dev/core@4.8.6`
  - `@trigger.dev/sdk@4.8.6`
  - `@trigger.dev/build@4.8.6`
## @trigger.dev/react-hooks@4.8.6

### Patch Changes

- Updated dependencies:
  - `@trigger.dev/core@4.8.6`
## @trigger.dev/redis-worker@4.8.6

### Patch Changes

- Updated dependencies:
  - `@trigger.dev/core@4.8.6`
## @trigger.dev/rsc@4.8.6

### Patch Changes

- Updated dependencies:
  - `@trigger.dev/core@4.8.6`
## @trigger.dev/schema-to-json@4.8.6

### Patch Changes

- Updated dependencies:
  - `@trigger.dev/core@4.8.6`
## @trigger.dev/slack@4.8.6

### Patch Changes

- Updated dependencies:
  - `@trigger.dev/core@4.8.6`
  - `@trigger.dev/sdk@4.8.6`
## @trigger.dev/sdk@4.8.6

### Patch Changes

- Fix `chat.agent` on AI SDK 7 sending the model the same input on every
step. Once a step's `prepareStep` returned messages (a steering message,
a mid-turn `chat.inject`, context compaction, or your own
`prepareStep`), later steps no longer included the newest tool results,
so the model kept repeating the same tool call until the step limit.
([`1a2c99034`](1a2c990))

If you call `chat.compact()` from your own `prepareStep`, pass it the
`responseMessages` your `prepareStep` receives (`chat.compact(messages,
steps, { ...options, responseMessages })`) so steps after a compaction
are rebuilt correctly on AI SDK 7.

- `chat.agent` now passes `chatId`, `turn` and `clientData` to the
callbacks an agent can call a model from, so a mock model injected
through `clientData` reaches them. The compaction `summarize` that runs
through `chat.MessageAccumulator.prepareStep()` and a
`chat.createSession` turn's `prepareStep()`, `onCompacted`,
`onValidateMessages`, and the `pendingMessages` `onReceived` and
`onInjected` callbacks previously lacked it. The `mockChatAgent` turn
methods (`sendMessage`, `sendRegenerate`, `sendAction`) now also wait
for `onTurnComplete` and the transcript save to finish, so tests can
assert on what those hooks recorded without a sleep.
([`579904220`](5799042))
- `concurrencyKey` now also takes an object, so one run can be limited
by two limits keyed by different things, like a cap per session and a
cap per organization. `default` is the run's own key and every other
entry gives a named limit its own key: `concurrencyKey: { default:
sessionId, [orgLimit.name]: orgId }`. A string key works as before.
([`4961fc474`](4961fc4))
- `webhooks.notion()` endpoints accept Notion's verification request.
While the endpoint has no signing secret, the `verification_token`
Notion sends is stored as its signing secret, so you can paste it back
into Notion and later deliveries verify with it. Redeploy with this
version for existing Notion endpoints to pick it up.
([`4288ea912`](4288ea9))
- Support scoped run tokens without an environment key, including
offloaded task inputs and stream waits. Explain why a run has no usable
credential for an operation.
([`ed0147895`](ed01478))
- Public-token helpers and chat token creation now prefer server minting
for root API keys too, avoiding the legacy claims endpoint on compatible
servers. Trigger, batch, and waitpoint responses missing a token also
try server minting first.
([`90d677b12`](90d677b))

Root-key callers retain local-signing compatibility only when the mint
endpoint is missing or an older server specifically rejects an empty
scope list. Named API keys and run tokens never self-sign. Authorization
failures, expiration validation errors, rate limits, and server or
connection failures do not activate the fallback.

Server minting with a root key signs with that same authenticated key,
including during its rotation grace period, rather than with its
replacement. Tokens stop authenticating when that key's grace period
ends or it is immediately revoked. Deploy this server update before
upgrading the SDK if you rely on minting during root-key rotation; older
servers can still reject grace-window keys.

Server-minted tokens use the API rate-limit budget and follow the
server's expiration rules, including its 30-day maximum. Minting tokens
without scopes requires self-hosted v4.8.4 or later unless the root-key
compatibility fallback is available. The default expiration remains 15
minutes.

- Fix webhook verification for `webhooks.brex()`, which now reads the
`webhook-id`, `webhook-timestamp` and `webhook-signature` headers Brex
sends instead of `svix-*` headers, and for `webhooks.composio()`, which
now keys the signature with the secret's raw text as Composio does
instead of base64-decoding it. PagerDuty, Sentry, Vercel, Zendesk,
Attio, DocuSign and Intercom endpoints now default to pasting the
provider-issued signing secret, and Linear and Calendly offer both
generating and pasting one. Redeploy with this version for existing
endpoints to pick up the change.
([`b5b311407`](b5b3114))
- Updated dependencies:
  - `@trigger.dev/core@4.8.6`
## @trigger.dev/convex@4.8.6

</details>

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.