[Feature]: make read_session tool able to read at offset and/or an agents last message

[NamedIdentity]

Prerequisites

• I have searched existing issues and discussions to avoid duplicates
• This feature request is specific to oh-my-opencode (not OpenCode core)
• I have read the documentation

Problem Description

The read_session tool can read the entire session, or return a number of messages, but cannot read offsets or the last message in the session. This prevents using the read_session tool to let one agent gain context from subagents responses without reading the entire subagent session.

This becomes a problem in OpenCode builds where subagent to subagent tasking and persistent sessions are used. I introduced a PR to OpenCode with that feature (anomalyco/opencode#7756 (comment)).

Proposed Solution

Update the read_session tool to allow offset reads or reading the last message in a session.

Alternatives Considered

Instruct the Primary agent to use it's in-context memory of subagent responses to write files which it then instructs subagents to read (relay-files). Presumably, writing files from existing text-in-context is less straining on the context window then generating text to send in prompts to subagents via A2A.

Doctor Output (Optional)

Additional Context

The read_session tool having offset reads or last message reads would be generally beneficial. However, it is especially beneficial when subagents can task subagents and maintain persistent sessions.

An example of a subagent-subagent persistent session workflow.

A Primary agent tasks a substantive-work subagent to extract and consolidate data from a corpus in stages. The substantive-work subagent produces multiple consolidated documents. The Primary agent sends one of the documents to be reviewed by a group of review-agents. The review-agents return responses to the Primary. The Primary reports to the user a summary of review-agents findings.

The substantive-work subagent can then be tasked to receive the review-agents responses, but the Primary would have to waste context generating text to relay messages from the review-agents to the substantive-work subagent. Being able to instruct the substantive-work subagent to read the last message in review-agents sessions would efficiently relay responses/data. The substantive-work subagent could then use that data, and it's prior context (it's a persistent session) to determine if the review-agents findings are true, and if so, issue corrections to the document.

Alternative workflow with stateless sessions:

Primary agent tasks to review a corpus of data by tasking multiple subagents. All subagents return response. Primary then instructs on subagent to read_session for all subagent sessions and use that distilled data to perform a task.

This feature is critical to agentic workflows performing data extraction, transformation, and consolidation; to knowledge-work.

Feature Type

New Tool

Contribution

• I'm willing to submit a PR for this feature
• I can help with testing
• I can help with documentation