Add Hermes Agent session transcript support to AgentRolloutSeedSource

[eric-tramel]

Priority Level

Medium (Nice to have)

Is your feature request related to a problem? Please describe.

The existing AgentRolloutSeedSource supports two vendor-specific formats (Claude Code and Codex). Hermes Agent stores session transcripts under ~/.hermes/sessions, and those session artifacts preserve the full agent loop with structured message roles, tool_calls, tool_call_id, and reasoning fields. That makes Hermes session transcripts a natural ingestion target for trace distillation, analysis, and training-data preparation without requiring a custom seed reader.

Describe the solution you'd like

Add a built-in Hermes Agent rollout format focused on session transcript ingestion rather than Hermes' ShareGPT trajectory export format.

The handler should support the session artifacts Hermes writes under ~/.hermes/sessions, including:

• Gateway session transcripts stored as per-session .jsonl files
• CLI session logs stored as session_*.json files containing top-level session metadata plus a messages array

The format should normalize both shapes into Data Designer's standard agent rollout schema, with Hermes-specific metadata stored in source_meta.

High level changes

1. Add HERMES_AGENT to AgentRolloutFormat enum
2. Implement HermesAgentRolloutFormatHandler
3. Register the handler with BUILTIN_AGENT_ROLLOUT_FORMAT_HANDLERS
4. Default the format path to ~/.hermes/sessions
5. Parse Hermes gateway session_meta / tool-definition metadata into source_meta
6. Normalize Hermes session messages into the standard messages payload used by AgentRolloutSeedSource

Example

from data_designer import DataDesigner, AgentRolloutSeedSource, AgentRolloutFormat

dd = DataDesigner()
config = dd.config_builder()
config.with_seed_dataset(
    AgentRolloutSeedSource(
        format=AgentRolloutFormat.HERMES_AGENT,
        path="~/.hermes/sessions",
    )
)

Describe alternatives you've considered

We could target Hermes' ShareGPT trajectory export format instead, but the session transcript format is a better fit for AgentRolloutSeedSource because it is closer to the existing Claude Code / Codex rollout handlers: one session per file, structured tool-call fields, and richer trace metadata.

Additional context

Hermes session storage currently has two related file conventions under ~/.hermes/sessions:

• Gateway transcripts: *.jsonl
• CLI session logs: session_*.json

A single Hermes handler can likely support both via file-shape auto-detection.

References:

• https://hermes-agent.nousresearch.com/docs/user-guide/sessions
• https://github.com/NousResearch/hermes-agent/blob/main/run_agent.py
• https://github.com/NousResearch/hermes-agent/blob/main/gateway/session.py

[eric-tramel]

This was added in #500 , closing.