feat: Add A2A parser plugin for inbound request inspection by huang195 · Pull Request #363 · rossoctl/cortex

huang195 · 2026-05-04T16:05:57Z

Summary

Adds a2a-parser pipeline plugin that parses inbound A2A JSON-RPC 2.0 requests (message/send, message/stream) and populates pctx.Extensions.A2A for downstream policy plugins (e.g., guardrails)
Expands A2AExtension type with RPCID, SessionID, MessageID, Role fields to match the actual A2A protocol payload from Kagenti UI
Renames A2APart.Type to Kind to align with A2A protocol field naming
Registers a2a-parser in the plugin registry (opt-in via config, not added to defaults)

Design

Follows the exact same pattern as mcp-parser (PR #361):

Declares BodyAccess: true and Writes: ["a2a"]
Always returns Continue (never rejects -- parsing only)
Gracefully handles nil/empty body, invalid JSON, unknown methods
Reuses jsonRPCRequest helper from the same package

Pipeline placement

pipeline:
  inbound:
    plugins:
      - jwt-validation
      - a2a-parser        # this PR
      # future: - guardrails (reads "a2a" extension slot)

Debug logging

Comprehensive slog.Debug calls at every parsing step for early debugging.
Enable with LOG_LEVEL=debug in authbridge-config ConfigMap.

Test plan

11 unit tests covering: capabilities, message/send, message/stream, multiple parts, file parts, unknown methods, nil/empty/invalid body, missing message, OnResponse no-op
go vet ./... passes
go test ./... passes (all packages including existing MCP/JWT/token-exchange tests)
go build ./... passes for cmd/authbridge
End-to-end: add a2a-parser to inbound pipeline config in Kind, send chat from UI, verify a2a-parser: parsed message/stream in logs

Assisted-By: Claude (Anthropic AI) noreply@anthropic.com

Parses A2A JSON-RPC 2.0 request bodies (message/send, message/stream) and populates pctx.Extensions.A2A with session, message, role, and parts for downstream policy plugins (e.g., guardrails). - Expand A2AExtension with RPCID, SessionID, MessageID, Role fields - Rename A2APart.Type to Kind to match A2A protocol field name - Register a2a-parser in the plugin registry - Add comprehensive debug logging for LOG_LEVEL=debug troubleshooting Assisted-By: Claude (Anthropic AI) <noreply@anthropic.com> Signed-off-by: Hai Huang <huang195@gmail.com>

- Make message extraction generic: extract from params.message regardless of method name (forward-compatible with future A2A methods) - Add uri fallback for file parts (A2A spec allows uri as alternative to data) - Consolidate debug logging: one summary log instead of per-field calls - Rename test to AnyMethod, add FutureMethodWithMessage test - Add DataPart test (exercises json.Marshal path) - Add FilePartURI test (exercises uri fallback) Assisted-By: Claude (Anthropic AI) <noreply@anthropic.com> Signed-off-by: Hai Huang <huang195@gmail.com>

evaline-ju · 2026-05-04T17:01:04Z

+		case "file":
+			content, _ = partMap["data"].(string)
+			if content == "" {
+				content, _ = partMap["uri"].(string)
+			}


Ref https://a2a-protocol.org/latest/definitions/#protobuf I'm curious how this case maps. Does another part of Kagenti form file data?

Good question. Looking at the A2A spec, the current spec (June 2025 draft) uses a unified Part with oneof content: text, raw (base64 bytes), url (file pointer), or data (structured JSON), plus optional mediaType/filename metadata.

The older SDK samples (which Kagenti UI currently uses) still send {"kind": "file", "data": "..."} or {"kind": "file", "uri": "..."} — that is what this code handles today. No part of Kagenti currently creates file parts; this branch just ensures we do not lose content if a future A2A client sends one.

As the spec stabilizes, we should update this to match the canonical Part structure (no kind discriminator, use mediaType + content field presence instead). I will add a TODO comment noting this.

Dead code in this PR -- no callers remain after removing the method switch. The A2A parser (PR rossoctl#363) will re-add them when it merges. Assisted-By: Claude (Anthropic AI) <noreply@anthropic.com> Signed-off-by: Hai Huang <huang195@gmail.com>

Malformed parts (non-string kind, empty kind) are now silently skipped instead of producing empty A2APart entries. Hardens the parser against non-compliant or malicious payloads. Assisted-By: Claude (Anthropic AI) <noreply@anthropic.com> Signed-off-by: Hai Huang <huang195@gmail.com>

Add test for type-mismatched part content (text: false, data: 0, uri: {}). Skip nil data values in the data-part branch to avoid serializing JSON null as content. Add TODO noting spec evolution. Signed-off-by: Hai Huang <hai@us.ibm.com> Signed-off-by: Hai Huang <huang195@gmail.com>

mrsabath

Clean A2A parser plugin — mirrors MCP parser pattern, comprehensive tests (14 cases), all prior review feedback addressed. LGTM.

mrsabath · 2026-05-05T00:37:34Z

+				content, _ = partMap["uri"].(string)
+			}
+		case "data":
+			if dataVal, ok := partMap["data"]; ok && dataVal != nil {


nit: This logs at Info level on every successfully parsed request. Since the parser runs on all inbound traffic (not just A2A), this could be noisy in production. Consider slog.Debug here and keeping only the existing Debug log below for detailed fields. The MCP parser uses slog.Info too, so this is consistent — but both might want to be Debug in the long run.

Dead code in this PR -- no callers remain after removing the method switch. The A2A parser (PR rossoctl#363) will re-add them when it merges. Assisted-By: Claude (Anthropic AI) <noreply@anthropic.com> Signed-off-by: Hai Huang <huang195@gmail.com>

- Make message extraction generic: extract from params.message regardless of method name (forward-compatible with future A2A methods) - Add uri fallback for file parts (A2A spec allows uri as alternative to data) - Consolidate debug logging: one summary log instead of per-field calls - Rename test to AnyMethod, add FutureMethodWithMessage test - Add DataPart test (exercises json.Marshal path) - Add FilePartURI test (exercises uri fallback) Assisted-By: Claude (Anthropic AI) <noreply@anthropic.com> Signed-off-by: Hai Huang <huang195@gmail.com>

huang195 added 2 commits May 4, 2026 12:05

evaline-ju reviewed May 4, 2026

View reviewed changes

esnible requested changes May 4, 2026

View reviewed changes

Comment thread authbridge/authlib/plugins/a2aparser.go Outdated

huang195 mentioned this pull request May 4, 2026

refactor(mcp-parser): Generic method handling and improved logging #364

Merged

3 tasks

esnible reviewed May 4, 2026

View reviewed changes

Comment thread authbridge/authlib/plugins/a2aparser.go

mrsabath approved these changes May 5, 2026

View reviewed changes

huang195 merged commit 4215fa6 into rossoctl:main May 5, 2026
17 checks passed

huang195 deleted the feat/a2a-parser-plugin branch May 5, 2026 00:48

mrsabath mentioned this pull request Jun 2, 2026

epic: Plugin Pipeline: Extending AuthBridge Beyond Token Exchange rossoctl/rossoctl#1458

Closed

9 tasks

huang195 mentioned this pull request Jun 24, 2026

track(merged): feat: Add A2A parser plugin for inbound request inspection (ext#363) #551

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat: Add A2A parser plugin for inbound request inspection#363

feat: Add A2A parser plugin for inbound request inspection#363
huang195 merged 4 commits into
rossoctl:mainfrom
huang195:feat/a2a-parser-plugin

huang195 commented May 4, 2026

Uh oh!

evaline-ju May 4, 2026 •

edited

Loading

Uh oh!

huang195 May 4, 2026

Uh oh!

Uh oh!

Uh oh!

mrsabath left a comment

Uh oh!

mrsabath May 5, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

huang195 commented May 4, 2026

Summary

Design

Pipeline placement

Debug logging

Test plan

Uh oh!

evaline-ju May 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

huang195 May 4, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

mrsabath left a comment

Choose a reason for hiding this comment

Uh oh!

mrsabath May 5, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

evaline-ju May 4, 2026 •

edited

Loading