[ghost]

Relevant issues

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have Added testing in the tests/test_litellm/ directory, Adding at least 1 test is a hard requirement - see details
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

CI (LiteLLM team)

CI status guideline:

• 50-55 passing tests: main is stable with minor issues.
• 45-49 passing tests: acceptable but needs attention
• <= 40 passing tests: unstable; be careful with your merges and assess the risk.

• Branch creation CI run
  Link:

• CI run for the last commit
  Link:

• Merge / cherry-pick CI run
  Links:

Type

🆕 New Feature
🐛 Bug Fix
🧹 Refactoring
📖 Documentation
🚄 Infrastructure
✅ Test

Changes

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Apr 3, 2026 9:54pm

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

ishaan-berri seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codspeed]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing litellm_ishaan_april1_try2 (cc8ed6e) with main (fc885af)}

[ryan-crabbe-berri]

lgtm

[greptile-apps]

Greptile Summary

This PR is a multi-feature release covering: (1) prompt environment/versioning support — new environment and created_by fields on LiteLLM_PromptTable with a DB migration, CRUD endpoint updates scoped by environment, and a revamped UI for browsing environments and version history; (2) user_custom_key_update hook — a new config-loadable async hook that can approve/deny key updates (mirroring user_custom_key_generate); (3) _enforce_upperbound_key_params refactor — extracted from _common_key_generation_helper and now also applied to key updates; (4) Amazon Nova Canvas image-edit on Bedrock — a new transformation class supporting INPAINTING, OUTPAINTING, IMAGE_VARIATION, and BACKGROUND_REMOVAL; (5) Bedrock Anthropic cache-token fix — promotes cache_creation_input_tokens/cache_read_input_tokens from message_stop onto message_delta; (6) Firecrawl self-hosted support; (7) failed-request timing fix in proxy_track_cost_callback.

Key findings:

• P1 — Edit button sends null template for non-selected versions (prompt_info.tsx): Clicking "Edit" directly on a version row that hasn't been explicitly loaded first calls onEdit with raw_prompt_template: null — the editor opens with no prompt content for that version.
• P2 — Double fetchPromptInfo on initial load (prompt_info.tsx): The isInitialMount ref is consumed on the first selectedEnv = null render; when selectedEnv is first populated, the effect triggers a second API call on every prompt open.
• P2 — Hardcoded environment filter options (prompts.tsx): Only development/staging/production are offered; custom environment names are invisible.
• P2 — get_prompt_versions returns empty list when prisma_client is None: The old in-memory fallback was removed.
• DB migration is safe for existing data; _enforce_upperbound_key_params refactor is correct; good unit test coverage for all server-side features.

Confidence Score: 4/5

Safe to merge after fixing the Edit button null-template bug; all server-side changes are solid but the UI has a functional issue that could silently open the editor with no content.

Score of 4 reflects one P1 UI bug (Edit on a non-selected version passes null template, so the editor opens empty) and two P2 issues (double fetch on initial load, hardcoded environment filter). Server-side logic — migration, endpoint changes, key management refactor, Bedrock transforms, and Firecrawl support — is well-tested and correct.

ui/litellm-dashboard/src/components/prompts/prompt_info.tsx (P1 Edit button bug and double-fetch); ui/litellm-dashboard/src/components/prompts.tsx (hardcoded env filter options)

Important Files Changed

Filename	Overview
ui/litellm-dashboard/src/components/prompts/prompt_info.tsx	Major refactor adding environment tabs, version history table, and per-version edit; has a P1 bug where clicking Edit on a non-selected version passes null for raw_prompt_template, and a P2 double-fetch issue with the isInitialMount pattern.
ui/litellm-dashboard/src/components/prompts.tsx	Adds environment filter Select with hardcoded options (development/staging/production); dynamic population from loaded prompts would be more robust.
litellm/proxy/prompts/prompt_endpoints.py	Adds environment and created_by scoping to all CRUD endpoints; get_prompt_versions switches from in-memory to DB queries which silently returns empty list when prisma_client is None.
litellm/proxy/management_endpoints/key_management_endpoints.py	Refactors upperbound key param enforcement into _enforce_upperbound_key_params helper (correctly preserving behavior), and adds user_custom_key_update hook support for single and bulk key updates.
litellm-proxy-extras/litellm_proxy_extras/migrations/20260331000000_add_prompt_environment_and_created_by/migration.sql	Adds environment (default 'development') and created_by columns, replaces unique(prompt_id, version) with unique(prompt_id, version, environment); safe for existing data since all rows get the same default environment.
litellm/llms/bedrock/image_edit/amazon_nova_canvas_image_edit_transformation.py	New Nova Canvas image edit transformation supporting INPAINTING, OUTPAINTING, IMAGE_VARIATION, and BACKGROUND_REMOVAL task types; model support detected via model_cost JSON flag following existing patterns.
litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py	Adds _promote_message_stop_usage to merge Bedrock's cache usage fields from message_stop onto message_delta so SDK clients that only read message_delta see the full token breakdown.
litellm/proxy/proxy_server.py	Adds user_custom_key_update global variable, loaded from general_settings.custom_key_update config, following the same pattern as user_custom_key_generate.
litellm/proxy/hooks/proxy_track_cost_callback.py	Fixes failed request timing by using actual start_time from the logging object instead of datetime.now(), ensuring correct duration is recorded for failed requests.

_{Reviews (1): Last reviewed commit: "fix(test): update model armor streaming ..." | Re-trigger Greptile}

[greptile-apps]

Edit button sends null template for non-selected versions

When a user clicks "Edit" directly on a version row that isn't the currently loaded version, raw_prompt_template will be null because isSelected is false (vNum !== selectedVersion). The edit button uses e.stopPropagation(), so the row click handler (handleVersionClick) is not called — the template is never fetched for that version before onEdit is invoked.

This means clicking "Edit" on any version except the one currently displayed will open the editor with an empty/null template, silently discarding the existing prompt content.

The fix is to always fetch the version's info before calling onEdit:

onClick={async (e) => {
  e.stopPropagation();
  let template = isSelected ? promptTemplate : null;
  if (!isSelected) {
    try {
      const versionedId = `${promptId}.v${vNum}`;
      const resp = await getPromptInfo(accessToken, versionedId, selectedEnv || undefined);
      template = resp.raw_prompt_template;
    } catch {
      NotificationsManager.fromBackend(`Failed to load version v${vNum} for editing`);
      return;
    }
  }
  const editData = {
    prompt_spec: { ...v, prompt_id: basePromptId, environment: selectedEnv },
    raw_prompt_template: template,
  };
  onEdit?.(editData);
}}

[greptile-apps]

Double fetchPromptInfo on every initial load

On initial component mount the flow is:

1. promptId/accessToken effect runs → fetchPromptInfo() (no env filter) → response sets selectedEnv via setSelectedEnv(...).
2. The selectedEnv effect fires with selectedEnv = null → isInitialMount.current is set to false, then returns.
3. React re-renders with the new selectedEnv → selectedEnv effect fires again. Because isInitialMount.current is already false it goes to the else branch and calls fetchPromptInfo(selectedEnv) again.

This causes two sequential API calls for every prompt the user opens. Consider tracking whether selectedEnv was set by the initial load vs. a user tab click using a separate ref flag:

const envSetByInitialLoad = useRef(false);
// Inside fetchPromptInfo, after setSelectedEnv:
envSetByInitialLoad.current = true;
setSelectedEnv(...);
// In the selectedEnv effect:
if (envSetByInitialLoad.current) {
  envSetByInitialLoad.current = false;
  if (selectedEnv && accessToken) fetchVersionHistory(selectedEnv);
  return;
}

[greptile-apps]

Hardcoded environment list in filter dropdown

The environment filter only shows development, staging, and production. Users who define custom environments won't see them in the dropdown and can't filter to them from the UI.

Consider populating the options dynamically from the loaded promptsList:

const environmentOptions = [
  ...new Set(promptsList.map(p => p.environment).filter(Boolean) as string[])
].sort().map(env => ({ label: env.charAt(0).toUpperCase() + env.slice(1), value: env }));

[greptile-apps]

get_prompt_versions silently returns empty list when prisma_client is None

The new implementation only queries the database. If prisma_client is None (no DB configured), the function returns an empty list rather than falling back to the in-memory registry as the old code did. Similarly, get_prompt_info only falls back to in-memory when environment is None — specifying an environment without a DB yields a 400.

Consider adding an explicit error message explaining that environment-scoped version history requires a database connection, or add a fallback path.