docs: document tool executions in the Operate area - #1129
Conversation
Adds a Tool Executions page under Operate > Governance, next to Audit Logs, covering the dashboard surface and read API shipped in ArcadeAI/monorepo#2661 and #2662. The page serves both jobs the feature exists for: a developer debugging why a tool call failed by reading the exact inputs and outputs, and an operator reviewing what a project ran for compliance and usage. It documents the access split those PRs introduced — the history is readable by anyone with a role on the project, while the recorded inputs and outputs take project-admin authority, and diagnostics stay readable either way. Also cross-links the page from the developer-facing tool error handling guide, and adds it to the Governance and Operate card grids. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
# Conflicts: # public/llms.txt
|
Reviewed the Retention section against the PLT-2978 stack now in review, and opened #1141 into this branch with the corrections — stacked so nothing here is rewritten. The headline is that "contact Arcade support" is about to stop being true: setting the retention window and turning recording on or off is becoming self-service (
One note for whoever merges: #1141 documents behaviour that hasn't shipped. If this PR lands first, hold that one until the stack merges, or the page promises a dashboard page that isn't there yet. Everything else here reads correctly against the implementation — the filters, the authority split on inputs and outputs, and the endpoint table all match. |
…2978] (#1141) * docs(tool-executions): recording and retention are self-service The Retention section predates PLT-2978 and told readers to contact support for changes they can now make themselves. It was also wrong in three ways that matter more than the missing capability: - It offered project-level overrides. Project settings are an explicit non-goal; the value is organization-wide and nothing customer-reachable sets a project's own. - It folded turning recording off together with deleting history. They are different: recording off stops new runs being written and leaves stored ones to expire on the window. Shortening the window is the control that deletes, and it reaches records already held. - It gave no bounds, so a reader had no way to know 0 and 400 are refused. Rewritten around the two settings, what each does to history already held, and what a reader sees when the list is empty. * docs(tool-executions): Evan's review — tighter, and no unshipped behaviour Takes the suggested wording for the opening, the table, and the defaults sentence, and types the two settings. Drops four things: the org-admin authority line, which the opening now covers; two connective sentences that only restated the bullets under them; the callout about recording permitting rather than compelling, since project-level settings do not exist and the distinction has no observable effect until they do; and the empty-history section, which narrated UI copy the product already shows. The Callout import goes with its last use. * docs(tool-executions): scope the defaults to Arcade Cloud "By default, recording is enabled" holds for Arcade Cloud, where the organization starts at ALLOWED and projects start enabled. It does not hold for a self-hosted deployment: the Engine's execution_logging.enabled defaults to false, and with it off no run is recorded and no retention worker runs, whatever the organization's policy says. The paragraph two below already scopes the maximum to Arcade Cloud, so this now reads consistently, with one clause for the self-hosted case.
…API credential (#1157) docs(tool-executions): follow the page rename, and name a credential that works The dashboard page is called Logging Policy now, and it sits under Organization, so "open your organization and select Execution Logging" names neither the label nor the path a reader would follow. The curl was worse than stale. It sent `$ARCADE_API_KEY`, which every other sample on the site means as a project key, and this endpoint refuses one: `401 Invalid credentials: missing account ID`. The policy is org-scoped and the route resolves an account, so a project key cannot carry the caller. Verified against a live stack — the same request with an account token returns 200. Everything else on the page was checked against that stack and holds: recording starts ALLOWED, the window starts at 7 with a maximum of 90 reported as `max_log_retention_days`, 0 and 91 are both refused with 422 rather than clamped, and sending one field leaves the other alone. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
…just merged (#1160) docs(tool-executions): correct the endpoint host, and drop an invented variable The previous commit introduced `$ARCADE_ACCOUNT_TOKEN`, which is not a thing — no such variable exists in the product or anywhere else on this site. The audit log page documents the same org-scoped shape as "User (API key/JWT)" with `$ARCADE_API_KEY`, so that is what this uses, with a pointer to it and a note that a project key is refused. The URL was wrong from the start, in both host and prefix. `logging-config` is a Coordinator route, and every other Coordinator `orgs/` call on this site is `cloud.arcade.dev/api/v1/`; only this page said `api.arcade.dev/v1/`. A live stack returns 404 for `/v1/orgs/...` and resolves `/api/v1/orgs/...`. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Adds a Tool Executions page under Operate → Governance, next to Audit Logs, documenting the dashboard surface and read API shipped in ArcadeAI/monorepo#2661 and #2662.
The page covers both jobs the feature exists for: a developer debugging why a tool call failed by reading the exact inputs and outputs, and an operator reviewing what a project ran for compliance and usage. It documents the access split those PRs introduced — the history is readable by anyone holding a role on the project, while the recorded inputs and outputs take project-admin authority, and diagnostics stay readable either way — plus the dashboard filters, the three read endpoints, and the retention window.
Also adds three dashboard screenshots, cross-links the page from the developer-facing tool error handling guide, and adds it to the Governance and Operate card grids.
Verified with
pnpm lint,pnpm test(835 pass),pnpm check-meta, Vale (0 errors), and a local dev-server render of the new page and both card grids.🤖 Generated with Claude Code
Note
Low Risk
Documentation-only changes with no runtime, auth, or application code impact.
Overview
Adds Tool Executions under Operate → Governance, documenting the dashboard and read API for per-project tool run history (debugging failures and compliance/usage review).
The new page explains what gets recorded, dashboard filters and detail views (including admin-only Show inputs/outputs toggles), the member vs project-admin access split, list/count/detail API endpoints with
include_inputs/include_outputsbehavior, and org-wide recording and retention via Logging Policy.Governance and Operate hub copy now mentions reviewing tool runs alongside audit logs; nav
_metaand card grids link to the page. The build Tool error handling guide gains a short section pointing operators to Tool executions for post-mortems.public/llms.txtlists the new URL for LLM indexing.Reviewed by Cursor Bugbot for commit 6915182. Bugbot is set up for automated code reviews on this repo. Configure here.