fix(ui): rotate model credentials in a dedicated modal so a normal save can't overwrite secrets

[ryan-crabbe-berri]

Relevant issues

Linear ticket

Resolves LIT-3169

Summary

Adds a dedicated "Update API Key" modal so a model's key can be rotated without hand-editing raw litellm_params, and closes a silent credential-corruption bug in the model edit form

The bug this fixes

/v1/model/info redacts secrets by masking them (e.g. azur****BBCC), not by removing them. The edit form seeded the read-only LiteLLM Params textarea with the entire litellm_params blob, masked secrets included, and re-sent the whole blob on every save. Saving any unrelated field (TPM, tags, cost) therefore re-sent the masked placeholder, which the backend encrypted straight over the real secret. The masked-but-kept fields are azure_ad_token, aws_session_token, watsonx token/zen_api_key and the OCI key fields; api_key, client_secret, vertex_credentials and the AWS access/secret keys were safe only because the backend removes those from the snapshot entirely. An earlier revision of this PR added an inline auth section that sat on top of that broken blob path, which made the footgun easier to trigger

What changed

API-key rotation lives in UpdateModelCredentialsModal, opened from an "Update API Key" button on the model. It is a single field that PATCHes only { api_key }; the backend already merges partial litellm_params, so everything else on the deployment is untouched. The modal and its trigger button are antd, so the feature introduces no new tremor

The general edit form no longer touches secrets at all: masked values are stripped from both the textarea seed and the outbound payload, so a normal save can never carry a redacted secret. Two stray console.logs that echoed credentials (the provider upload handler and the model-update response) are removed, and a react-hooks/use-memo error that was failing the frontend-lint CI job is fixed

Screenshots / Proof of Fix

Driven through the Admin UI at http://localhost:4000/ui/models-and-endpoints, capturing the outbound PATCH from the browser and reading the stored secret back from the DB by decrypting the row (the API only ever shows it masked)

Before the fix, a plain save destroyed a masked-but-kept secret:

# azure model with azure_ad_token set to a known value
azure_ad_token (DB, decrypted) = azureADtoken_ORIGINAL...
# open the model, Edit Settings, Save (touch nothing)
# the browser PATCH carried "azure_ad_token": "azur****...BBCC"
azure_ad_token (DB, decrypted) = azur****************************BBCC   <- real secret gone

After the fix, the same plain save leaves it alone, and the modal rotates only the key:

# Edit Settings, Save (touch nothing)
# PATCH litellm_params no longer contains the masked secret at all  -> secret preserved
# Update API Key -> enter a new key -> submit
# PATCH body: {"litellm_params":{"api_key":"sk-rotated..."},"model_info":{"id":...}}
api_key (DB, decrypted)        = sk-rotated...     # rotated
azure_ad_token (DB, decrypted) = azureADtoken_ORIGINAL...   # untouched by the api-key-only PATCH

To re-capture screenshots: go to http://localhost:4000/ui/models-and-endpoints, open a DB model, click "Update API Key", enter a new key, submit; then Edit Settings and Save without touching anything and confirm the key still works. Browsers cache old JS chunks across rebuilds, so hard-refresh if the UI looks stale

Test plan

• model_info_view.test.tsx: regression test asserting a masked secret in litellm_params is never present in the save PATCH (verified it fails when the strip guard is disabled)
• update_model_credentials_modal.test.tsx: the modal sends a minimal { api_key } PATCH and is a no-op when the field is blank
• provider_specific_fields.test.tsx plus the Add Model and credential-modal suites still pass (the shared component is back to its pre-PR render path)
• End to end against a live proxy: corruption reproduced before the fix and gone after, and api-key rotation confirmed, captured from the browser PATCH and the decrypted DB row

Type

🐛 Bug Fix

🆕 New Feature

Changes

UI only. New UpdateModelCredentialsModal (single api-key field, antd); model_info_view wires the antd trigger button, strips masked secrets on seed and on submit, and drops the inline auth section; provider_specific_fields loses the unused excludeKeys/disableRequired/useProviderAuthFieldKeys additions; networking drops a credential-echoing log

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

This PR fixes a silent credential-corruption bug where the model edit form re-sent masked secret placeholders (e.g. azur****BBCC) on every save, causing the backend to encrypt the placeholder over the real secret. A dedicated UpdateModelCredentialsModal now handles API-key rotation as a minimal PATCH containing only the new key, while the general edit form strips any masked value before sending.

• Corruption fix: isMaskedSecret (regex \\*{2,}) and stripMaskedSecrets are applied at both the textarea seed and the outbound PATCH payload; a regression test confirms masked values never reach the wire.
• New modal: UpdateModelCredentialsModal sends exactly { litellm_params: { api_key }, model_info: { id } }, leaving all other deployment config untouched; antd Form validation blocks empty submission.
• Cleanup: removes two console.log calls that echoed freshly-entered provider secrets from the modelPatchUpdateCall path and the vertex credential upload handler.

Confidence Score: 5/5

Safe to merge — the change is UI-only, tightly scoped to the model credential edit path, and introduces no new backend calls or auth boundary changes.

The corruption fix is correctly layered: masked secrets are stripped at both the form-seed stage and the final PATCH assembly, so neither a user typing nor a code-path regression can re-introduce the bug. The query-key invalidation (["models", "list"] prefix) correctly matches the useModelsInfo query. destroyOnHidden is valid in the project's antd 5.29.3. The regression test actively catches the masked-secret-on-wire scenario. No new network handlers or auth changes are introduced.

No files require special attention.

Important Files Changed

Filename	Overview
ui/litellm-dashboard/src/components/model_info_view.tsx	Core fix: adds isMaskedSecret/stripMaskedSecrets guards that strip masked credential placeholders from both the textarea seed and the PATCH payload, plus toolbar migration from Tremor to antd and wiring of the new credentials modal.
ui/litellm-dashboard/src/components/update_model_credentials_modal.tsx	New modal that sends a minimal { litellm_params: { api_key }, model_info: { id } } PATCH for key rotation; uses antd Form with required validation, destroyOnHidden (valid in antd 5.29.3), and proper error handling.
ui/litellm-dashboard/src/components/update_model_credentials_modal.test.tsx	New test suite covering the minimal-PATCH contract and the no-op-on-blank guard for the credentials modal; correctly mocks networking and notification dependencies.
ui/litellm-dashboard/src/components/model_info_view.test.tsx	Adds a regression test asserting that a masked secret in litellm_params never appears in the save PATCH; the test is well-targeted and would catch re-introduction of the corruption bug.
ui/litellm-dashboard/src/components/networking.tsx	Removes two console.log calls that echoed the full PATCH payload (including freshly-entered provider secrets) before and after the model update request.
ui/litellm-dashboard/src/components/add_model/provider_specific_fields.tsx	Removes debug console.log statements (vertex credential upload handler and upload onChange) that were added in an earlier revision; the component interface is simplified back to its pre-PR state.

_{Reviews (4): Last reviewed commit: "refactor(ui): convert the model detail t..." | Re-trigger Greptile}

[veria-ai]

PR overview

Provider auth editing in model edit view

This PR adds provider-specific authentication fields to the model edit form, omits blank secret inputs from update payloads, and avoids copying newly entered auth fields into the local read-only params view after save. I checked the form merge order, credential-name branch, provider field key resolution, and PATCH helper for secret exposure or authorization issues.

Security review

• No new security issues were flagged in the latest review.
• 1 previously flagged issue(s) appear fixed in the latest changes.
• No review issues remain open on this pull request.

Risk: 2/10

[ryan-crabbe-berri]

@greptile re review

[ryan-crabbe-berri]

@greptileai re review

[codspeed-hq]

Merging this PR will not alter performance

✅ 30 untouched benchmarks

---

_{Comparing litellm_model_auth_fields (f9cfb9b) with litellm_internal_staging (d4c33b2)}

[ryan-crabbe-berri]

@greptileai review

Reworked credential editing into a dedicated api-key rotation modal that PATCHes only the new key, stopped the model edit form from re-sending masked secrets (the corruption fix), and converted the model detail toolbar to antd