dsh-session-tools
English | 中文
A DSH Web plugin for session management across conversations:
session_archive: archive the calling session, or another session by exact ID.session_read: read DSH's bounded, untrusted text snapshot of another session by exact ID.session_send: send a message to another ordinary session and receive a durable request ID.session_wait: wait for, or poll, the exact reply associated with a session_send request.Copy session ID in every populated session row's three-dot menu.ID header button: copy the exact ID of the currently open session.
session_send places one durable user message in the target's next ordinary turn. DSH's native agent resolver can resume an inactive ordinary session with its recorded context before delivery. session_wait correlates the request ID with that exact target turn and returns only its visible assistant text; it does not expose reasoning or tool traces.
Archiving is persistent but non-destructive: the session disappears from normal lists while its log remains stored. Cross-session reads reuse DSH's session-reference projection, so each reference keeps the host's configured byte limit.
Requirements
- DSH
0.1.1-rc.2 / 0.1.2-rc.1
- A Web profile (
web in the examples below)
Version 0.2.2 reads live event snapshots on DSH 0.1.2-rc.1 and retains the legacy event-array path for DSH 0.1.1-rc.2. Polling sees newly committed replies and terminal status without reusing an older snapshot.
Install
dsh plugin --profile web add github:Unintendedz/dsh-session-tools#v0.2.2
Restart the running DSH Web service. Plugins are loaded when the service starts.
Usage
Use Copy session ID in a row's three-dot menu, or open a session and click ID in its header.
To ask another ordinary session to do work, call session_send:
{
"session_id": "session-target-id",
"message": "Check the test results and summarize any failures."
}
The tool returns immediately with a request_id. Pass both IDs to session_wait:
{
"session_id": "session-target-id",
"request_id": "request-id-from-session-send",
"timeout_seconds": 60
}
session_wait returns completed, running, or failed. Its wait defaults to 60 seconds and accepts 0–600 seconds; use 0 to poll. A running result is safe to wait on again with the same IDs because waiting never resends the message or cancels the target.
You can also paste a copied ID into another conversation and ask it to call session_read, or ask it to call session_archive with that ID. Calling session_archive without session_id archives the current conversation.
The sidebar menu is repositioned after the copy action is injected, so all four rows remain inside the viewport even when the source conversation is near the bottom of the screen.
Update
Install the desired tag, then restart DSH Web:
dsh plugin --profile web add github:Unintendedz/dsh-session-tools#v0.2.2
Uninstall
dsh plugin --profile web remove dsh-session-tools
Safety
session_read uses DSH's native session-reference resolver. Returned text is bounded by the host and explicitly treated as untrusted context.session_send and session_wait require two different ordinary sessions. Self-targeting is rejected; subagent sessions must use DSH's built-in send_message tool.- Only the session that created a request may wait for its result.
session_wait returns only visible assistant text from the request's exact turn. Reasoning and tool traces are excluded, and returned UTF-8 text is bounded to 65,536 bytes.- A wait timeout returns
running; it does not resend the request or cancel the target. Cross-session wait cycles are rejected.
session_archive preserves the session log; it does not delete conversation data.- Tool arguments require exact, non-empty IDs and reject unknown fields.
Development
npm install
npm test
Set DSH_NATIVE_ROOT to the installed DSH package root containing its node_modules, and use a fresh temporary DSH_HOME, to include the native Session polling regression. It uses synthetic data and does not boot a profile.
The host entry point is lib/index.js; browser source lives in src/client.jsx. npm test rebuilds lib/client.js and runs the host and browser-bundle tests.
Request lifecycle and native output
session_wait associates an inbox claim with its active turn. If that turn is aborted, errors, or is blocked before the user message is recorded, waiting returns failed with that reason. A claim whose turn is still active remains running; ownership and queued-request cancellation checks still apply. Native tool output always includes the request status and any terminal reason, followed by the available reply text bounded to 64 KiB of UTF-8. A partial reply therefore remains visibly failed or running.
