feat(storage): add content_filters to DynamoDbChatStorage

[nuthalapativarun]

Issue Link (REQUIRED)

Fixes #131

Summary

Changes

Adds an optional content_filters parameter to DynamoDbChatStorage — a list of (original, placeholder) string pairs. When provided:

• On save (save_chat_message, save_chat_messages): each original value is replaced by its placeholder before the message is written to DynamoDB.
• On fetch (fetch_chat, fetch_chat_with_timestamp, fetch_all_chats): placeholders are restored to their original values before returning to the caller.

This eliminates the need for a custom storage implementation when sensitive content (e.g. company names, tokens) must not be persisted in plain text.

User experience

Before: users had to subclass DynamoDbChatStorage or write a custom storage class to mask sensitive values.

After:

storage = DynamoDbChatStorage(
    table_name="my-table",
    region="us-east-1",
    content_filters=[
        ("acme-corp", "#COMPANY#"),
        ("secret-token-xyz", "#TOKEN#"),
    ]
)

DynamoDB stores #COMPANY#; the agent always sees acme-corp.

Checklist

• I have performed a self-review of this change
• Changes have been tested
• Changes are documented
• I have linked this PR to an existing issue (required)

Acknowledgment

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[cornelcroi]

Hey @nuthalapativarun, a few things need to be fixed before this can merge.

There is a bug in save_chat_message — when a consecutive-role duplicate is detected, the method returns the raw result from _fetch_raw which still has placeholder strings in it instead of the original values. The caller gets masked data back. Also, fetch_all_chats applies _reverse_content_filters differently from the other fetch methods — it has an isinstance guard that silently skips the reversal when content is not a list, so placeholders can leak through. Both of these need tests too. And the docs page for DynamoDB storage needs to be updated to document the new content_filters parameter. CI also hasn't run on the branch.

One thing worth calling out in the PR description: the filter replacement works by doing a json.dumps then naive string replace, so if a filter string contains JSON special characters like quotes or backslashes it can produce invalid JSON. Worth documenting as a known limitation. Also interesting to note — the Swift equivalent (TransformingChatStorage) deliberately does not restore the original values on read, while this PR does. Both approaches are valid but they are different behaviors and worth mentioning.

[nuthalapativarun]

Thanks for the review, @cornelcroi! Pushed a follow-up commit addressing all four points:

1. Unmasked result on consecutive duplicate: save_chat_message's early-return path (when the new message has the same role as the last stored message) now runs the existing conversation through _reverse_content_filters before returning, instead of returning the raw placeholder-masked value straight from storage.
2. fetch_all_chats isinstance guard: removed the isinstance(msg['content'], list) check that gated content-filter restoration — it now always reverses filters regardless of the stored content's shape, so masked placeholder text no longer leaks through for non-list content.
3. Tests: added test_content_filters_consecutive_message_returns_unmasked and test_fetch_all_chats_restores_non_list_content covering both fixes, plus the full suite (14 tests) passes.
4. Docs: added a "Content Filters (Python)" section to docs/src/content/docs/storage/dynamodb.mdx documenting the content_filters parameter and its masking/restoration behavior.

Ready for re-review whenever you get a chance.