feat(agentex): support SYNC agents in the Slack gateway

[michael-chou359]

What

Let the Slack gateway invoke ACPType.SYNC agents, not just async/agentic ones.

Why

_dispatch hard-coded task/create + event/send. event/send is only in the allowed method set for ASYNC/AGENTIC agents, so routing @agent <a-sync-agent> from Slack failed the turn with:

ClientError: Unsupported method: AgentRPCMethod.EVENT_SEND for ACP type: ACPType.SYNC

and the user just saw "Something went wrong."

How

Branch _dispatch on agent.acp_type:

• SYNC → a single message/send (SendMessageRequestEntity), which get-or-creates the thread task and returns the reply messages synchronously — no task/create, no event/send, no polling.
• ASYNC / AGENTIC (e.g. golden-agent) → unchanged: task/create on the first turn (with the DuplicateItemError create-race fallback) + event/send + poll for the settled reply.

Reply text is extracted the same way for both (_agent_text), and first-turn params (config_id / default MCPs) are built once and passed to whichever create path runs. task_metadata.channel="slack" (origin gating for golden-agent) is still set on the async path only, which is where it applies.

Testing

• New unit test: a SYNC agent triggers exactly one message/send (task_name = the thread key) and the reply is returned directly; the async/agentic tests are unchanged (the fake agent now carries acp_type).
• Full gateway unit suite green (52), lint clean.

🤖 Generated with Claude Code

Greptile Summary

The PR adds Slack gateway support for synchronous ACP agents and introduces bounded recovery for concurrent task-creation races.

• Routes SYNC agents through a single synchronous message/send request.
• Preserves task/create, event/send, and reply polling for ASYNC and AGENTIC agents.
• Adds retry helpers and tests for create races and temporary replica lag.

Confidence Score: 4/5

The PR is not yet safe to merge because concurrent first turns can still be dropped when read-replica lag outlasts the fixed retry window.

Both previously reported create-race paths now retry, but they still depend on the same read-only lookup becoming consistent within a default 0.75-second backoff period; after that period, the SYNC path propagates DuplicateItemError and the async path propagates ItemDoesNotExist, producing the generic Slack failure response.

Files Needing Attention: agentex/src/domain/use_cases/slack_gateway_use_case.py

Important Files Changed

Filename	Overview
agentex/src/domain/use_cases/slack_gateway_use_case.py	Adds ACP-type-specific dispatch and race retries, but bounded read-replica recovery still drops turns when replication lag exceeds the configured window.
agentex/tests/unit/use_cases/test_slack_gateway_use_case.py	Covers SYNC dispatch and short-lived replica lag, but does not establish safe recovery when lag outlasts the retry window.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Slack turn] --> B{Agent ACP type}
    B -->|SYNC| C[message/send]
    C --> D{DuplicateItemError?}
    D -->|No| E[Return synchronous reply]
    D -->|Yes| F[Backoff and retry]
    F --> C
    B -->|ASYNC or AGENTIC| G[Look up thread task]
    G --> H{Task exists?}
    H -->|No| I[task/create]
    I --> J{DuplicateItemError?}
    J -->|Yes| K[Retry read-only task lookup]
    J -->|No| L[event/send]
    K -->|Task visible| L
    K -->|Retries exhausted| M[Generic Slack failure]
    H -->|Yes| L
    L --> N[Poll and return reply]

Loading

Comments Outside Diff (1)

1. agentex/src/domain/use_cases/slack_gateway_use_case.py, line 497 (link)

   Replica retry still drops turns

   If the configured read replica remains behind the primary longer than the fixed retry window, every get_task attempt misses the winning task and the helper raises ItemDoesNotExist, causing _run_turn to drop the Slack turn and return the generic failure response.

   Prompt To Fix With AI

   This is a comment left during a code review.
   Path: agentex/src/domain/use_cases/slack_gateway_use_case.py
   Line: 497
   
   Comment:
   **Replica retry still drops turns**
   
   If the configured read replica remains behind the primary longer than the fixed retry window, every `get_task` attempt misses the winning task and the helper raises `ItemDoesNotExist`, causing `_run_turn` to drop the Slack turn and return the generic failure response.
   
   ---
   
   For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

   

Prompt To Fix All With AI

### Issue 1
agentex/src/domain/use_cases/slack_gateway_use_case.py:497
**Replica retry still drops turns**

If the configured read replica remains behind the primary longer than the fixed retry window, every `get_task` attempt misses the winning task and the helper raises `ItemDoesNotExist`, causing `_run_turn` to drop the Slack turn and return the generic failure response.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

_{Reviews (3): Last reviewed commit: "fix(agentex): make the Slack create-race..." | Re-trigger Greptile}