feat(agent-builder): author, save, and dry-run custom tools

[dmarticus]

What

Wires the backend's custom-tool authoring loop (posthog#66584) into PostHog Code. Until now the agent builder could only read tool source (tools/<id>/source.ts, read-only) — this adds the client + a minimal GUI to create, save, and dry-run a custom tool end to end, behind the existing agent-platform flag.

This is v0: correctness over polish. No capabilities panel beyond a save-time readout, no monaco intellisense, no per-team quota UX — those are v1.

How

Part A — API client + types (hand-written; agent-platform types are not generated from OpenAPI)

• @posthog/shared/agent-platform-types: ToolCompileError/ToolCompileErrorKind, ToolCapabilities, WriteToolRequest/WriteToolResult, DryRunToolRequest/DryRunToolEnvelope/DryRunToolResult.
• posthog-client.ts:
  • putRevisionTool — 422 returns a typed compile-failed result (rendered inline); other non-2xx (400/409 sealed/…) still throw.
  • deleteRevisionTool — 404 (tool_not_found) treated as success.
  • dryRunRevisionTool — reads ok from the body (a throwing tool is HTTP 200 with ok:false); 200 and 500 both return the envelope; 429 → throttled (with max_concurrent), 503 → unavailable. Never thrown, never retried.
  • parseFailedRequest helper recovers status + body from the shared fetcher's Failed request: [status] body throw.

Part B — hooks

• useSaveRevisionTool (invalidates the bundle(...) query only when the compile persisted), useDeleteRevisionTool (invalidates bundle), useDryRunRevisionTool (no invalidation). All expose isPending.

Part C — UI (agent builder)

• ToolSourceEditor — draft-only editable source with compile-on-save, inline ToolCompileError diagnostics (kind + message + 1-based line/column), Revert, double-submit guard, and a secret_refs / dynamic_secret_refs readout from the save response.
• ToolDryRunPanel — args JSON editor, optional mock_secrets key/value rows, Test button, and a result area handling success (result + duration_ms), tool failure (error.code/message), 429 busy, and 503 unavailable.
• ToolBody gates edit + dry-run behind AGENT_PLATFORM_FLAG; non-draft / non-custom tools stay read-only.

Testing

• pnpm typecheck (23/23 packages) and pnpm biome check on all touched files pass.
• New client tests: 200/422/409 writes, delete 404-as-success, dry-run 200-ok / 200-fail / 500 / 429 / 503 + mock_secrets passthrough. New hook test for the invalidate-only-when-persisted branch. @posthog/api-client 78 pass, @posthog/ui 1325 pass.

Note

End-to-end dry-run requires the deployed backend (janitor + main app) to include posthog#66584 with the sandbox backend configured. Until then dry-run returns 503, which is the expected "not available here" state this UI is built against.

🤖 Generated with Claude Code

[trunk-io]

Merging to main in this repository is managed by Trunk.

• To merge this pull request, check the box to the left or comment /trunk merge below.

After your PR is submitted to the merge queue, this comment will be automatically updated with its status. If the PR fails, failure details will also be posted here

[github-actions]

React Doctor found no issues in the changed files. 🎉

_{Reviewed by React Doctor for commit 89510b2.}

[greptile-apps]

Comments Outside Diff (1)

1. packages/ui/src/features/agent-applications/components/ToolSourceEditor.tsx, line 982 (link)

   Uncleaned timeout on unmount

   setTimeout is started but never cleared, so if the component unmounts within the 2-second window (e.g. the user navigates away right after saving), the captured setJustSaved closure fires on a stale closure. React 18 makes this a silent no-op, but the pattern accumulates stale closures. Store the timer ID in a useRef and clear it in a cleanup callback or on the next onSave call.

   Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

_{Reviews (1): Last reviewed commit: "feat(agent-builder): author, save, and d..." | Re-trigger Greptile}