> ## Documentation Index
> Fetch the complete documentation index at: https://docs.gloo.com/llms.txt
> Use this file to discover all available pages before exploring further.

# 2026-10-18 — Forge updates

> Software asks named people for a decision or input through a human task, hears the outcome by webhook and reads the answer. An Architect turn refused before dispatch can reuse its idempotency key after recovery. An agent node placed in a sandbox VM pauses on its question and resumes on the same conversation, as an inline node does. Router models, detached tools and sandbox commands require recorded admission. A voided approval cancels all of its in-flight implement runs, and the launch check reads the inputs a run stores. A human node's answer passed through a procedure node or a failure input reaches an agent only in a fenced notice, and an LLM router that has run many times sees agent inputs whole again. Create immutable collection schemas and manage home-authorized Knowledge records. Ranked knowledge search accepts exact synchronized record-kind keys. Keep Forge-owned notes on a synchronized record through Console, HTTP and MCP. Attachment OCR waits for governance, and unverifiable agent definitions cannot run. Read a provider's recent changes and an organization's capability binding readiness over HTTP and MCP. Ingress decisions and successful Console launches are recorded before starting a run. Workflow publication checks stops and approval policies before delivery. Expiring file links keep email and Slack downloads behind current Forge access. Integrations now use the canonical synchronized source HTTP and MCP names. Open Synchronized sources in Console and use the canonical synchronized source HTTP and MCP names.

**2026-10-18**

External Apps and platform keys can ask a person.

* An External App or a platform key can ask 1 to 25 named people for a decision or input with [Ask a person](https://docs.gloo.com/forge/reference/http-api/governance/requestHumanTask.md). It declares its own form and a due date from 15 minutes to 14 days. It reads the outcome with [Read one of your asks](https://docs.gloo.com/forge/reference/http-api/governance/getRequestedHumanTask.md) and cancels an open ask with [Withdraw one of your asks](https://docs.gloo.com/forge/reference/http-api/governance/withdrawRequestedHumanTask.md). People answer in the Forge Console, Slack or email. Software never answers a human task.
* The new permission `human_tasks:use` (Run family, not elevated) lets software ask. No preset or default role includes it. An operator gives it through a role, and an app's token must also name it in its `scope`. A grant reaches only the people homed under the unit it covers.
* A person may take one of the app's actions only while holding the app's non-elevated eligibility permission at the task's home. Anyone asked may decline. Asks are capped at 200 open per requester, 10 open per person asked, and 200 created per 24 hours (10 per person asked); a breach of a 24-hour cap answers `429` with `Retry-After`.
* Every surface names who asked, for example `Requested by the External App “Vendor desk”`, and says "Forge never asks for passwords or codes in a request." Email for these tasks shows links in the app's text inert (`hxxps://example[.]com`).
* Through a Personal key, an app's task is summary-only and decline-only: the reads show fixed copy and no form, and every declared action is refused with the task's Console link.
* An External App hears its own asks through the [`human_task.completed` and `human_task.canceled` webhooks](https://docs.gloo.com/forge/reference/http-api/webhooks#human-task-events), which carry ids only. Admin subscriptions see these ids for every app. A platform key, or an app whose grant covers only a sub-unit, polls instead. No webhook is sent when an ask closes because its requester can no longer ask.

**2026-10-18**

Architect turn retries after refusal recording recovers.

If an Architect turn hits its rate limit and Forge cannot record the refusal, the API returns
503 without starting the turn. After recording and rate capacity recover, retry the same request
with the same `Idempotency-Key`. A completed turn or a failure after dispatch keeps the key spent.

**2026-10-18**

Detached agent nodes pause for a person and resume.

* An agent node placed detached (a declared timeout above 8 minutes, so it runs in a sandbox VM) no longer waits up to 2 minutes for an answer. Its `human_ask` returns `status: "pending"` at once, as an inline node's does. When its turn ends, the VM stops and the node pauses: it holds no agent slot, no VM and no timer reaps it.
* When every question the node asked has closed, it resumes on the same conversation in a new VM turn with a fresh timeout. Each answer arrives as a Forge notice. The resumed turn replays earlier messages' final text, not earlier tool calls or their results, and the agent is told to call its tools again for anything it needs.
* For a detached node, `human_check` and `human_withdraw` report a question's status without its answer.
* A node can pause at most 5 times per attempt. Each paused turn's spend counts toward the run's cost cap as it happens.
* A value a person supplied to a step, such as a human node's answer passed downstream, is no longer pasted into an agent node's opening prompt. The prompt refers to the input by name. The value arrives in a fenced Forge notice, after a Forge line that the Console shows as Forge, not as the person. This applies to inline and detached nodes and to debug attempts.
* An agent node already running in a VM when this change shipped keeps the 2-minute wait until it ends.

**2026-10-18**

Exact-call admission across router and sandbox execution.

New workflow instances record governance before starting an LLM router. A held router waits for
approval; an older instance that has not yet executed its LLM router asks for a restart.

Detached file and sandbox callbacks use their frozen tool permissions and current execution
identity. Legacy debug attempts without that frozen context require a fresh attempt.

Command approvals cover one exact reissued call for 15 minutes. Changed commands or sandboxes need
new review. Approval does not automatically replay a command, and a consumed approval remains spent
when recording or execution fails.

**2026-10-18**

Voided Foreman approvals cancel every run they covered.

* When a Foreman plan's or request's approval is voided, every running or paused `sdlc-implement-task` run launched for it is cancelled, however many runs the workflow has. A run started after the void is cancelled too, unless a person has approved the plan or request again.
* The approval check on an `sdlc-implement-task` launch reads the inputs the run stores, including input defaults. A default that names a plan or request needs that plan or request to be approved.
* A launch without the launch permission, or with inputs the workflow's contract rejects, now reports that error before `FOREMAN_APPROVAL_REQUIRED`.

**2026-10-18**

Human answers stay out of more agent prompts.

* A human node's answer that reaches an agent node or an LLM router through a procedure node, whether the procedure is connected to the human node or reads its answer or failure context as one of the procedure's inputs, or through an input wired to a failed node's failure context, is no longer pasted into the prompt. The prompt refers to the input by name, and the value arrives in a fenced Forge notice.
* An LLM router that has run 50 or more times again receives inputs that came from no person whole, in its prompt, with no notice.
* The value in a fenced Forge notice is one line: line breaks become spaces, and a value over 4,000 characters is cut and ends in `…`. Other inputs reach the prompt whole.
* Correction to the 2026-10-18 note on detached agent nodes: only inputs wired from a step a person completed, such as a human node's answer, move to the notice. Values a person supplies when launching a run, such as workflow inputs and `{{workflow.x}}`, are still placed in the prompt.
* When a Foreman plan's or request's approval is voided, a run is cancelled even when it stored the plan or request id in a different accepted form.

**2026-10-18**

Author structured Knowledge records in Console.

Forge Console now provides a **Record collections** page under Knowledge for structured records
authored in Forge. Collection schemas and scope are immutable, while record edits and removals use
current-version checks to prevent silent overwrites.

**2026-10-18**

Search knowledge by synchronized record kind.

Knowledge search can now be narrowed to one to eight exact synchronized record-kind keys through
HTTP and `forge_search_knowledge`. Kind selection preserves existing scope and source access checks
and returns the existing empty result when a known kind has no searchable matches.

**2026-10-18**

Forge-owned record annotations.

People and authenticated machines can read and replace a synchronized record's annotations.
Sync preserves the Forge-owned map while continuing to own provider attributes. Editing requires
current knowledge read and manage authority at the source home, namespace and source access.
Saving replaces the whole map; an empty map clears it and null remains a value. Read more about
[external knowledge sources](https://docs.gloo.com/forge/knowledge/external-sources.md).

**2026-10-18**

Recorded admission before attachment OCR.

External inference checks governance and budgets before reading attached files or extracting image
text. A refused or held request leaves attachments unread. Once a read or OCR attempt starts,
its approval remains consumed even if later preparation fails.

Agent dispatch now stops when its active definition version cannot be verified or has no stored
hash. Restore a verified definition before trying again. Historical versions and their hash
contracts remain intact.

**2026-10-18**

Provider audit feed and capability binding readiness in the API.

`GET /api/v2/providers/listAuditEvents` (`forge_list_provider_audit_events`) lists a provider's
recent changes with `audit:read` at the platform scope: event, category and time, never the actor or
the option diffs. `GET /api/v2/capabilities/getBindingReadiness`
(`forge_get_capability_binding_readiness`) reads a bindable capability's provider binding for one
organization with `credentials:read` there, never a credential id, field or secret. Binding
readiness, in the API and the Console, now finds an organization's own binding however many
organizations the capability binds.

**2026-10-18**

Record ingress launch admission before creating a run.

Ingress workflow launches record a permission Pass before creating the run. Global and agents-door
stops, and denied launch permissions, return the ID of their recorded Block. If Forge cannot record
the required decision, it starts no run. Existing throttle, signature and replay behavior is preserved.

Successful Console production and draft launches also record a Pass before creating a run.
Console permission refusals keep their existing throwing guards; their separate client-reported
References are explicitly untrusted reports, rather than verified server Block decisions.

Pre-launch decisions remain explicitly unresolved in Evidence because a run does not yet exist.

**2026-10-18**

Record publication decisions before delivery.

Workflow publish steps check the global stop, agents door and applicable policies before sending.
A matching approval policy parks the step behind a standard human task. Approval releases one
concrete publication; a changed target or payload needs fresh approval. If Forge cannot record
the decision, it sends nothing. Confirmed-delivery retries continue to reconcile their original receipts.

**2026-10-18**

Published file downloads require sign-in and current access.

New email and Slack file links expire after 24 hours and can be revoked by an operator. Opening
one requires sign-in and your current access to that file in Forge; receiving a message grants no
additional access. Downloads and outbound attachment reads are recorded before bytes are served. Large downloads
check current access and record reads throughout the transfer.
Previously sent permanent storage links and bytes already delivered cannot be recalled.

**2026-10-18**

Retire synchronized source compatibility aliases.

The legacy HTTP and MCP aliases for synchronized sources are retired. Integrations now
call `/api/v2/knowledge/synchronized-sources/*` and the corresponding
`forge_*_synchronized_source*` tools. SDK callers replace `listKnowledgeSources` with
`listSynchronizedSources`, `getKnowledgeSource` with `getSynchronizedSource`,
`listKnowledgeSourceRuns` with `listSynchronizedSourceRuns`, `updateKnowledgeSource` with
`updateSynchronizedSource`, `listKnowledgeSourceProviders` with
`listSynchronizedSourceProviders`, `createKnowledgeSource` with `createSynchronizedSource`,
`removeKnowledgeSource` with `removeSynchronizedSource`, `relistKnowledgeSource` with
`relistSynchronizedSource`, and `requestKnowledgeSourceSync` with
`requestSynchronizedSourceSync`. Retries through the canonical HTTP routes keep the persisted
idempotency identity created during the compatibility interval.

**2026-10-18**

Canonical synchronized source names.

Console now opens **Synchronized sources** at `/knowledge/synchronized-sources`; bookmarks to the
old Console route redirect there. Integrations can use the canonical
`/api/v2/knowledge/synchronized-sources/*` routes and `forge_*_synchronized_source*` MCP tools.
The previous HTTP and MCP names remain compatibility aliases until their containing release has
given callers one release to migrate. See [synchronized knowledge sources](https://docs.gloo.com/forge/knowledge/external-sources.md).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.