> ## 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.

# 2026-11-02 — Forge updates

> Six admin lists that stopped at a bounded read now page at the database with a sealed cursor. Seven admin lists page at the database, at most 100 rows a page. Approval-age alerts include the full pending backlog, while daily-cost alerts exclude interface-specific totals before reading them. Three budget lists that answered an empty list past 500 rows now page at the database with a sealed cursor. Six chat admin lists page at the database, at most 100 rows a page. Client health now ranks clients by attention level, explains each column, shows ages in weeks, months and years, and adds sorting, search and filters. Seven lists read from the forge-apps and forge-capabilities components now page at the database with a sealed cursor. A tool's and a procedure's credential lists and an org unit's Slack workspace list now page at the database with a sealed cursor instead of answering their whole range at once. Two lists that read a bounded range and then dropped rows now page at the database with a sealed cursor. The approval policy catalog, a run's approval trail, your credential's own grants and the policy effectiveness list now page at the database with a sealed cursor instead of answering one capped read. A list that holds more rows than an operation reads now ends with hasMore false and truncated true, instead of hasMore true with no cursor. Forge MCP now confirms every operation whose effect cannot be taken back, such as a message sent or a role given, and the OpenAPI document marks them with x-forge-irreversible. Forge MCP's tools/list carries every tool's output schema, and a run Markdown page over 16 KiB links a resource instead of carrying the body. Every Forge MCP refusal is the error and meta envelope with an error.continuation next call, reads are named through forge_call_read_operation, and the runs list answers an unreadable cursor with 400. A first backfill too large to finish now moves History from forward to the earliest day that fits, with an audit event. Admin operations that read a record through a nested call answer their declared refusal instead of 500. Org unit memberships, structure types, spending limits, scheduled org changes and expertise tags now page at the database with a sealed cursor instead of answering their whole range. Two lists that answered one capped read now page at the database with a sealed cursor, in the order they had. listAgentTaskRuns, listAgentConsultations, listWorkflowDebugAttemptHistory and listObjectiveOutcomeContributions page at the database with a sealed cursor. listProjectObjectives pages at the database with a sealed cursor and no longer fails past 1,000 links. Check every committed evidence section and download a portable verification pack. The admin run audit events list now pages at the database with a sealed cursor instead of answering at most 1,000 events at once. The run file references list now pages at the database with a sealed cursor instead of answering one capped read. When the runtime stops a request whose database calls passed its time limit, Forge answers 503 internal_error with Retry-After instead of 500. The external tool configurations, a worker's bandwidth feedback and a credential's verification history are bounded lists, answered whole with no cursor. An API call hit by a transient runtime fault now answers 503 with Retry-After where it used to answer 500. A path under /api/ that no route serves answers 404 not_found in the JSON error envelope, with Allow when another method serves it. The four vendor registry admin lists now page at the database with a sealed cursor instead of answering their whole range at once.

**2026-11-02**

Six admin lists page at the database.

Six admin lists now page through every row at the database. Before, five filtered a bounded read
and said nothing when the bound cut it short, and the org chart snapshots answered `[]` past 500
snapshots. Each reads at most 100 rows a page, and a `limit`
above 100 reads 100.

* `GET /api/v2/identity/admin/unmapped-attempts`: `limit` 1–100, default 50, attempts still
  awaiting a mapping first.
* `GET /api/v2/environments/admin/connections/list`: `limit` 1–100, default 100, oldest first,
  within your key's org units.
* `GET /api/v2/org-units/admin/snapshots`: `limit` 1–100, default 100, newest published first.
* `GET /api/v2/trust-fabric/admin/command-decisions`: `limit` 1–100, default 50, newest first.
* `GET /api/v2/connections/admin/credentials`: `limit` 1–100, default 100, oldest first, only the
  credentials your key may read.
* `GET /api/v2/chat/admin/channel-attachments`: `limit` 1–100, default 50, newest first.

Continue while `meta.hasMore` is `true` by passing `meta.nextCursor` as `cursor` with the same
other parameters. The cursor is sealed (it starts `an2_`) and bound to the operation, its other
parameters and your credential. Any other cursor answers `400 Invalid cursor`. A page may hold
fewer rows than `limit` while more remain, and none of these lists answers `meta.partial`.

`GET /api/v2/projects/admin/tasks` is unchanged.

**2026-11-02**

Seven admin lists page at the database.

Seven admin lists now page through every row at the database. Before, each answered its whole
list at once or stopped at a read bound. Each reads at most 100 rows a page, and a `limit` above
100 reads 100.

* `GET /api/v2/email/admin/routes`: `limit` 1–100, default 100, oldest first. One request used to
  read up to 200 routes and stop there.
* `GET /api/v2/notifications/admin/slack-sends`: `limit` 1–100, default 50, newest first, optionally
  for one `surface`. One request used to read up to 200 sends.
* `GET /api/v2/ops/admin/notes`: `limit` 1–100, default 100, newest first. A resource's notes used
  to be answered whole.
* `GET /api/v2/projects/admin/requests/plans`: `limit` 1–100, default 100, by status (descending).
  A request's plans used to be answered whole.
* `GET /api/v2/projects/admin/requests`: `limit` 1–100, default 100. One request used to read up
  to 1,000 intake requests.
* `GET /api/v2/runs/admin/checkpoints`: `limit` 1–100, default 100, oldest first. A run's
  checkpoints used to be answered whole.
* `GET /api/v2/webhooks/admin/dead-letters`: `limit` 1–100, default 50, newest first, optionally
  by `status` and `eventType`. One request used to read up to 1,000 entries.

Continue while `meta.hasMore` is `true` by passing `meta.nextCursor` as `cursor` with the same
other parameters. The cursor is sealed (it starts `an2_`) and bound to the operation, its other
parameters and your credential. Any other cursor answers `400` with "Invalid cursor: restart paging
from the first page, without a cursor". None of the seven leaves a row out after it reads it.

**2026-11-02**

Approval-age and daily-cost alerts use narrower database reads.

Approval-age alerts find the oldest pending approval across the full backlog, including backdated
approvals beyond the former 5,000-record scan limit. Stored pending approvals with missing or empty
legacy identifiers also contribute to the age metric.

Daily-cost alerts read fewer records by excluding interface-specific copies of the same spend
before loading the totals. The cost total, configured thresholds, hourly debounce, and combined
delivery when several rules fire remain unchanged.

**2026-11-02**

A budget's alerts and transactions page at the database.

A budget's alerts, pending alerts and transactions now page through every row at the database.
Before, each answered `[]` once the budget held more than 500 alerts or transactions. Each answers
at most 100 rows a page, and a `limit` above 100 answers 100.

* `GET /api/v2/costs/admin/budgets/alerts`: `limit` 1–100, default 50, untriggered alerts first,
  then triggered ones, each in creation order.
* `GET /api/v2/costs/admin/budgets/alerts/pending`: `limit` 1–100, default 50, triggered and
  unacknowledged alerts in creation order.
* `GET /api/v2/costs/admin/budgets/transactions`: `limit` 1–100, default 50, newest first. It
  read up to 200 a request before; continue with the cursor instead.

Continue while `meta.hasMore` is `true` by passing `meta.nextCursor` as `cursor` with the same
`budgetId`. The cursor is sealed (it starts `an2_`) and bound to the operation, its other
parameters and your credential. Any other cursor answers `400 Invalid cursor`. A pending-alerts
page may hold fewer rows than `limit`, or none, while more remain: an acknowledged alert is left
out of the page it was read in.

**2026-11-02**

Six chat admin lists page at the database.

Six chat admin lists now page through every row at the database. Before, each answered its whole
list at once or stopped at its `limit`. Each reads at most 100 rows a page, and a `limit` above 100
reads 100.

* `GET /api/v2/chat/admin/channel-participants`: `limit` 1–100, default 100. A channel's roster
  used to be answered whole.
* `GET /api/v2/chat/admin/my-pending-drafts`: `limit` 1–100, default 100, oldest first. The held
  drafts used to be answered whole.
* `GET /api/v2/chat/admin/channel-sessions` and `.../channel-session-cards`: `limit` 1–100,
  default 50, most recently active first, or oldest first with `status`. Without `status`, a
  session that becomes active while you page moves ahead of your cursor, so that walk does not
  return it; start again from the first page to see it.
* `GET /api/v2/chat/admin/my-channels`: `limit` 1–100, default 50.
* `GET /api/v2/chat/admin/session-messages`: `limit` 1–100, default 100, oldest first. One request
  used to read up to 200 messages and stop there; follow `meta.nextCursor` for the rest.

Continue while `meta.hasMore` is `true` by passing `meta.nextCursor` as `cursor` with the same
other parameters. The cursor is sealed (it starts `an2_`) and bound to the operation, its other
parameters and your credential. Any other cursor answers `400` with "Invalid cursor: restart paging
from the first page, without a cursor". A page may hold fewer rows than `limit` while more remain:
a participant that is not a person or agent, a draft or session your key may not use, and a
channel it may not read are left out of the page they were read in.

**2026-11-02**

Client health attention levels, readable ages, sorting and filters.

On **Planning → Client health**, each client linked to a customer now has an attention level
(**Needs attention**, **At risk**, **Watch** or **On track**) and the board lists the most urgent
first. Column names are plain (**Past due**, **Late first reply**, **Quietest**), and an info icon
beside each one explains what it counts. Ages read as "3 wk", "7 mo" or "4 yr", and **Age mix**
splits open tickets into under a week, 1–4 weeks, 1–3 months, 3–12 months and over a year.

Click a column name to sort, search clients by name, or tick **Customers only**. The **Past due**,
**Late first reply** and **Needs attention (clients)** cards filter the table. Departments not yet
linked to a customer are listed last on gray rows. See
[Client health](https://docs.gloo.com/forge/console/planning/client-health.md#attention-levels).

**2026-11-02**

Seven component-backed lists page at the database.

Seven lists now page through every row at the database. Before, each answered its whole list at
once, or stopped at a cap.

* `GET /api/v2/webhooks`: `limit` 1–200, default 200, by status and then creation order. It no
  longer stops at 200 subscriptions.
* `GET /api/v2/externalApps/list`: `limit` 1–100, default 100, in creation order.
* `GET /api/v2/webhooks/admin/list`: `limit` 1–100, default 100, newest first.
* `GET /api/v2/webhooks/admin/deliveries`: `limit` 1–100, default 25, newest first. One request
  used to read up to 1,000 deliveries; follow `meta.nextCursor` for more than 100.
* `GET /api/v2/models/admin/slots`, `.../approaching-deprecation` and `.../routing/configs`:
  `limit` 1–100, default 100, in creation order.

Continue while `meta.hasMore` is `true` by passing `meta.nextCursor` as `cursor` with the same
other parameters. The two platform lists' cursors start `d1.`, and the admin lists' cursors start
`an2_`. Each cursor is bound to the operation, its other parameters and your credential, and any
other cursor answers `400`. A page may hold fewer rows than `limit`, or none, while more remain:
an app you may not manage, a slot in another `environment`, a model outside the deprecation window
and a routing config whose agent your key may not read are left out of the page they were read in.

The Forge MCP `forge://external-apps` resource follows `meta.nextCursor` for at most five pages
or until it holds at least 100 apps (at most 199, since a page holds up to 100), and says `truncated: true` when more exist, as `forge://agents` does.

**2026-11-02**

The connections admin credential and Slack workspace lists page at the database.

`GET /api/v2/connections/admin/tools/credentials`, `/procedures/credentials` and
`/slack/workspaces` now page at the database. Before, each read its whole range and answered it in
one list. Each takes `limit` 1–100 (default 100; a `limit` above 100 answers 100) and `cursor`, and
keeps its order: a tool's or a procedure's credentials in creation order, and an org unit's active
Slack workspaces in the order they were installed.

A credential your key may not read, or a workspace outside your key's reach, is still left out, so
a page may hold fewer rows than `limit` while more follow. Continue while `meta.hasMore` is `true`
by passing `meta.nextCursor` as `cursor` with the same other parameters. The cursor is sealed (it
starts `an2_`) and bound to the operation, its other parameters and your credential. Any other
cursor answers `400 Invalid cursor`. An unknown `toolId`, `procedureId` or `orgUnitId` still
answers an empty list.

**2026-11-02**

Two bounded lists page at the database.

Two lists now page at the database. Before, each read a bounded range, then dropped rows, with
nothing to say when the bound cut.

* `GET /api/v2/connections/admin/credentials/by-state` lists the credentials you may read in one
  approval state, in creation order, with no 5,000-credential registry bound.
* `GET /api/v2/trust-fabric/admin/tier3-invocations` keeps its order: by disposition, newest first
  within each. It now reads only decisions inside `sinceMs`, so older decisions no longer crowd out
  recent ones. Its default `limit` is now 100, not 200; continue with the cursor for the rest.

Both take `limit` 1–100 (default 100; a `limit` above 100 answers 100) and `cursor`. Continue
while `meta.hasMore` is `true` by passing `meta.nextCursor` as `cursor` with the same other
parameters. A row you may not read is left out after the read and still advances the cursor, so a
page may hold fewer rows than `limit`, or none, while more follow. The cursor is sealed and bound
to the operation, its other parameters and your credential. Any other cursor answers
`400 Invalid cursor`.

`GET /api/v2/workflows/listDebugFiles` and `GET /api/v2/capabilities/listAuditEvents` do not change.
`listDebugFiles` still lists at most 551 files, because the file download looks up the same 551; an
uncapped download lookup would have to come first. `listAuditEvents` merges the events of
`synchronized_knowledge` with those of its old name, `knowledge_source`, by time, which one index
walk cannot reproduce.

**2026-11-02**

Four governance and grants lists page at the database.

`GET /api/v2/governance/listPolicies`, `/listRunApprovals` and `/listMyGrants`, and
`GET /api/v2/trust-fabric/admin/policy-effectiveness`, now page at the database. Before, each read
its whole list, or up to 500 rows, and answered it in one list. Each takes `limit` 1–100 (default
100; a `limit` above 100 answers 100) and `cursor`, and keeps its order: policies in creation
order, a run's approval decisions oldest first, and your grants in the order they were created.

A run's trail and your grants are no longer cut at 500, so `meta.partial` on those two lists is
always `false`. A compatibility role's grant is still left out of `listMyGrants`, so a page may hold
fewer grants than `limit` while more follow. Continue while `meta.hasMore` is `true` by passing
`meta.nextCursor` as `cursor` with the same other parameters. The cursor is sealed and bound to the
operation, its other parameters and your credential. Any other cursor answers `400 Invalid cursor`.

The Forge MCP tools `forge_list_run_approvals` and `forge_list_my_grants` take the same optional
`cursor` and `limit` and return `meta.hasMore` and `meta.nextCursor`.

`listTrustPolicyRevisions` and `GET /api/v2/roles/admin/grants/holder` still answer their whole
list, a page at a time.

**2026-11-02**

A list page offers more rows only with a cursor that reaches them.

A list page could answer `meta.hasMore: true` with `meta.nextCursor: null`: more rows existed, but
no cursor reached them. A client that pages until `hasMore` is `false` could not finish.

`meta.hasMore: true` now always comes with a `meta.nextCursor` string. When a list holds more rows
than its operation reads (a bounded read reached its cap, or the operation reported a cut read),
its last page answers `meta.hasMore: false`, `meta.nextCursor: null` and `meta.truncated: true`.
Narrow the request with the operation's filters to reach the rest. `meta.partial` and an
operation's own fields are unchanged. `ListPageMeta` in the OpenAPI spec gains `truncated` and
requires a `nextCursor` string whenever `hasMore` is `true`.

**2026-11-02**

MCP confirms irreversible operations.

Each operation whose effect cannot be taken back now carries `x-forge-irreversible` in the
OpenAPI document: `data-lost`, `delivered`, `authority` or `live`. Before, Forge MCP confirmed an
operation only when it was a `DELETE` or its name led with a destructive verb, so
`giveRole`, `transferProjectOwnership`, `publishWorkflowDraft`,
`approveSynchronizedSourceShrink` and the `send*` operations ran through `forge_call_operation`
without asking.

`forge_call_operation` now asks the user before it runs any of the 57 declared operations, and
the question says what cannot be taken back. `forge_reassign_human_task` asks too, since the new
assignee is told at once. A host without elicitation gets `confirmation_required` with a
`confirmationToken`, as for any destructive call. See
[Destructive calls are confirmed](https://docs.gloo.com/forge/reference/mcp-server.md#destructive-calls-are-confirmed).

**2026-11-02**

MCP lists output schemas; a large run page links its resource.

`tools/list` now lists each Forge MCP tool's output schema as
`{ type: "object", anyOf: [<success>, <refusal>] }`. A client that checks results against it
accepts a refusal too: the refusal branch requires `error.code`, `error.message` and
`meta.requestId`. Models are not sent output schemas, so they cost no model-visible bytes.

`forge_get_run_markdown` answers a page over 16 KiB with a summary instead of the Markdown, in its
text and its structured content alike. `data.markdown` is `null`, and `omitted` gives the body's
size and the resource that reads it, also linked as a `resource_link`. A later page links the new
`forge://runs/{runId}/report/{cursor}.md` resource. Read the resource for the Markdown. See
[MCP server](https://docs.gloo.com/forge/reference/mcp-server.md).

**2026-11-02**

MCP refusals are one envelope and name their next step.

* Every Forge MCP refusal is the `{ error, meta }` envelope. A refusal from `forge_call_operation` or `forge_call_read_operation` now carries it too, with Forge's own `status` and `body` beside it, unchanged. See [Error Handling](https://docs.gloo.com/forge/reference/mcp-server.md#error-handling).
* A refusal names its next call in `error.continuation`: `forge_explain_denial` with the refused permission for a `403`, and, for a `404` from a tool that takes an id, the read that lists those ids, such as `listRuns` for `forge_get_run`.
* The server names reads through `forge_call_read_operation`, as its instructions route them: in `forge_find`, `forge_explain_denial` and `forge_triage_run` continuations, prompts, resources, tool descriptions and an unlisted tool's refusal.
* [List workflow runs](https://docs.gloo.com/forge/reference/http-api/runs/list.md) answers a cursor it did not issue with `400` "Invalid cursor" naming the `cursor` field, not `500`. `forge_list_runs` returns that refusal.

**2026-11-02**

A first backfill now moves History from forward when its window is too large to finish.

Before a Freshservice mirror's first backfill reads its first page, it now counts the tickets in
its window. If there are more than 40,000, it moves **History from** forward to the earliest day
whose tickets fit, and records the change in the audit log with the old and new dates. Before,
such a backfill stopped at the page ceiling, started over and needed an operator to narrow the
window. If the count fails, the window stays as it is and the backfill runs as before. The count
uses Freshservice's Filter Tickets API and has not yet been tested against a live account. See
[Client health](https://docs.gloo.com/forge/console/planning/client-health.md).

**2026-11-02**

Nested not-found refusals.

Some admin operations read their record through a nested call: an action that runs another action, for example `updateExternalToolCredentialsAdmin`. When that record was missing, they answered `500 internal_error` instead of the refusal the operation declares, usually `404 not_found`. They now answer the declared refusal. See [Errors](https://docs.gloo.com/forge/reference/errors.md).

The `503` Forge answers when it cannot record a refusal now carries `Retry-After`, like its other transient `503`. See [internal\_error](https://docs.gloo.com/forge/reference/errors.md#internal_error).

**2026-11-02**

The org and people admin lists page at the database.

`GET /api/v2/org-units/admin/memberships/by-member`, `/org-units/admin/structure-types`,
`/org-units/admin/scheduled-changes`, `/costs/admin/limits` and `/people/admin/expertise-tags` now
page at the database. Before, each
read its whole range and answered it in one list. Each takes `limit` 1–100 (default 100; a `limit`
above 100 answers 100) and `cursor`.

Each keeps its order. A member's memberships, the structure types and a scope id's spending limits
are still listed by `createdAt`, then id. The scheduled changes and the expertise tags no longer
stop at 200 and 500 rows.

A membership in customer territory, or a structure type that is not active, is still left out, so
a page may hold fewer rows than `limit` while more follow. Continue while `meta.hasMore` is `true`
by passing `meta.nextCursor` as `cursor` with the same other parameters. The cursor is sealed (it
starts `an2_`) and bound to the operation, its other parameters and your credential. Any other
cursor answers `400 Invalid cursor`.

`GET /api/v2/org-units/admin/structure-types/tiers` and `/people/admin/workers/expertise-tags` are
unchanged. The tiers still answer in one list, in `tierOrder` order, where a tier with no
`tierOrder` counts as 0. The worker's tags still answer in one `{ tags, truncated }` object.

**2026-11-02**

Two person identity lists page at the database.

`GET /api/v2/workforce/listPersonIdentities` and `GET /api/v2/workforce/listPersonEmailClaims` now
page at the database. Each takes `limit` 1–100 (default 100; a `limit` above 100 answers 100) and
`cursor`, and keeps its order and its item shape: a person's links oldest first with the `workos`
sign-in row last, and a person's email claims oldest first.

Before, a person with more than 500 identity links answered `413`, and one with more than 100
email claims answered `500`; both lists now answer every row, a page at a time.

Continue while `meta.hasMore` is `true` by passing `meta.nextCursor` as `cursor` with the same
other parameters. The cursor is sealed and bound to the operation, its other parameters and your
credential. Any other cursor answers `400 Invalid cursor`.

`GET /api/v2/knowledge/goldenSets/list` and `GET /api/v2/synthesis/listTemplates` still answer
their whole list, a page at a time.

**2026-11-02**

Four platform lists that took no page parameters page at the database.

Four platform lists took no page parameters, because they filtered rows after a bounded read and
could not say when the bound cut. They now page through every row at the database:

* `GET /api/v2/agents/listTaskRuns` and `GET /api/v2/agents/listConsultations`: `limit` 1–100,
  default 50, newest first; with `status`, foreground rows first. They no longer stop at one page
  of 100 rows. A row that names a Personal Agent is still left out. A row whose parent run cannot
  be decided within the request's limits now withholds its content, as a run you cannot see does,
  instead of answering `meta.partial`.
* `GET /api/v2/workflows/listDebugAttemptHistory`: `limit` 1–25, default 25. The first page starts
  with the attempt itself, then your earlier attempts of the same node, newest first. It no longer
  stops at 25 attempts, and `data.truncated` is gone. `data.retentionExpiresAt` is unchanged.
* `GET /api/v2/outcomes/listObjectiveContributions`: `limit` 1–50, default 50, newest first in
  `data.items`. It no longer stops at the newest 50 contributions.

Each list leaves out rows the credential cannot read after the read, so a page may hold fewer rows
than `limit` while more remain. Continue while `meta.hasMore` is `true` by passing
`meta.nextCursor` as `cursor` with the same other parameters. None of them answers
`meta.partial`. The cursor starts `d1.`, is bound to the operation, its other parameters and the
credential, and any other cursor answers `400 Invalid cursor`. A `limit` above a list's maximum
now reads the maximum instead of answering `400`, and a malformed `limit` answers `400` with
"limit must be a positive integer".

**2026-11-02**

A project's objective links page at the database.

`GET /api/v2/projects/listObjectives` now pages through every link at the database: `limit`
1–200, default 200. The project's `projectObjectives` links come first, oldest first, then its
primary link, unless one of those links names the same objective. A project with more than
1,000 links no longer fails; page through it instead. When the last page of links is full, the
next page holds only the primary link.

Continue while `meta.hasMore` is `true` by passing `meta.nextCursor` as `cursor` with the same
`id`. The list never answers `meta.partial`. The cursor starts `d1.`, is bound to the operation,
`id` and the credential, and any other cursor answers `400 Invalid cursor`. A `limit` above 200
reads 200, and a malformed `limit` answers `400` with "limit must be a positive integer". The
links are in `data` as before.

The link writes (`linkObjective`, `unlinkObjective`) still answer all of a project's links in
the same order.

**2026-11-02**

Complete receipt verification and portable evidence.

Trust receipts offer a single action to verify the signature and every committed evidence page.
Download an evidence pack with the immutable receipt, its public key, full evidence pages and an
offline verifier. Offline verification requires a separately confirmed Forge key fingerprint.
Successful verification preserves any reported evidence gaps and the run's execution result.
Historical receipts without page commitments are explicitly verified as signature-only.

**2026-11-02**

A run's audit events page at the database.

`GET /api/v2/ops/admin/runs/audit-events` now pages through every audit event of the run at the
database, newest first. Before, it read at most 1,000 events and answered them in one list. It
takes `limit` 1–100 (default 100; a `limit` above 100 answers 100) and `cursor`.

Continue while `meta.hasMore` is `true` by passing `meta.nextCursor` as `cursor` with the same
`runId`. The cursor is sealed (it starts `an2_`) and bound to the operation, its other parameters
and your credential. Any other cursor answers `400 Invalid cursor`. A page may hold fewer events
than `limit`, or none, while more remain: an event about a person, which needs `workforce:read`,
is left out of the page it was read in.

**2026-11-02**

A run's file references page at the database.

`GET /api/v2/runs/listArtifacts` now pages at the database. Before, it read up to 500 file
references and answered one list. It keeps its order: a run's files oldest first.

`listArtifacts` takes `limit` 1–100 (default 100; a `limit` above 100 answers 100) and `cursor`.
A file you may not read is still left out, so a page may hold fewer references than `limit` while
more follow.

The list is not cut at a read cap any more, so `meta.partial` is always `false`. Continue while
`meta.hasMore` is `true` by passing `meta.nextCursor` as `cursor` with the same `runId`. The cursor
is sealed and bound to the operation, its `runId` and your credential. Any other cursor answers
`400 Invalid cursor`.

The Forge MCP tool `forge_list_run_artifacts` takes the same optional `cursor` and `limit` and
returns `meta.hasMore` and `meta.nextCursor`.

`GET /api/v2/tools/listNeedingAttention`, `/models/list`, `/sandboxes/listActive` and
`/customers/listScopes` still answer their whole list, a page at a time.

**2026-11-02**

A request out of database time answers a retryable 503.

When the runtime stops a request because its database calls passed the runtime's time limit, an
`/api/v2` or `/api/v3` route now answers `503 internal_error` with `Retry-After: 1` and the message
`Forge is briefly unavailable. Retry after the Retry-After delay.` It used to answer
`500 internal_error`. It usually means the host was overloaded, and the same request then succeeds
when sent again. Resend a read after the delay; resend a write with the same `Idempotency-Key`. If
it still fails after a few resends, it is not transient: file a support ticket with
`meta.requestId`.

**2026-11-02**

Three lists are bounded by construction.

Three lists are now `x-forge-paging: bounded`. Each answers its whole list in one response, with
`meta.hasMore: false` and `meta.nextCursor: null`. Any `cursor` answers `400 Invalid cursor`. The
spec states each bound as `x-forge-bound`.

* `GET /api/v2/connections/admin/external-tools/configurations` holds at most one configuration per
  external tool provider. `limit` does not cut it.
* `GET /api/v2/workforce/listBandwidthFeedback` answers the worker's newest `limit` reports,
  1–50 (default 20).
* `GET /api/v2/connections/admin/credentials/verifications` answers the credential's newest `limit`
  verification results, 1–100 (default 20).

For the last two, `limit` is how many of the newest items to answer. It is no longer a page size,
and no cursor follows it. The items and their order are unchanged.

**2026-11-02**

Transient runtime faults answer 503 with Retry-After.

When a `/api/v2` or `/api/v3` call is hit by a transient runtime fault, such as a query past its
time limit, it now answers `503` with the `internal_error` code and a `Retry-After` header. Before, a route that handled the failure itself answered `500 internal_error`, so
your client could not tell the fault was temporary. `meta.requestId` still matches the
`X-Request-Id` header. Resend a read after the delay; check a write's result before repeating it.

When the route handled the fault itself, a call sent with an `Idempotency-Key` keeps its `500`,
so it is never told to retry with a key whose answer is already saved; check the result, then
retry with a new key. Other server
errors still answer `500`. See [`internal_error`](https://docs.gloo.com/forge/reference/errors#internal_error).

**2026-11-02**

Unknown API paths answer 404 not\_found in the error envelope.

A path under `/api/` that no route serves now answers `404 not_found` in the standard
[error envelope](https://docs.gloo.com/forge/reference/errors.md#not_found), with `docUrl`, `meta.requestId` and the
`X-Request-Id` header. Before, these requests got a plain-text `404` with no request id. When
another method serves the path, the `404` carries an `Allow` header naming those methods, as
`/api/v3` collections already did. Every path a route serves answers as before.

**2026-11-02**

The vendor registry admin lists page at the database.

`GET /api/v2/vendors/admin/list`, `/specs`, `/lessons` and `/connections` now page at the
database. Before, each read its whole range and answered it in one list. Each takes `limit` 1–100
(default 100; a `limit` above 100 answers 100) and `cursor`, and keeps its order: the registry
newest first, a vendor's API specs and active lessons in the order they were added, and its
connections by name.

Continue while `meta.hasMore` is `true` by passing `meta.nextCursor` as `cursor` with the same
other parameters. The cursor is sealed (it starts `an2_`) and bound to the operation, its other
parameters and your credential. Any other cursor answers `400 Invalid cursor`. An unknown
`vendorKey` still answers an empty list.


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