[Feature Request] Chat Turn Management (Rewind/Delete/Edit) & Response Sanitization Middleware

[Master-Antonio]

Summary

Currently, dsh treats the conversation log as strictly append-only. When an upstream provider returns malformed responses, fails a tool call, or causes a loop, the context window gets polluted with unrecoverable noise (e.g., repeated error recovery messages or broken
HTML artifacts).

I would like to propose two key UX and backend enhancements:

1. Interactive turn management (rewind, edit, or delete turns).
2. Response sanitization / cleanup filter for third-party provider artifacts.

---

1. Interactive Turn Management (Rewind / Delete / Branch)

Problem:

• If a model fails to format tool parameters properly or gets stuck in a loop, the user is forced to manually tell the agent to resume or restart.
• Repeating instructions like "I closed the file, try again" inflates the context token count and degrades future reasoning.
• There is currently no UI option (or slash command) to delete, edit, or roll back to a previous message.

Proposed Solution:

• Rewind / Fork from Turn: Add an action button on past messages to rollback/branch the session state to that specific checkpoint (discarding subsequent turns and failed tool calls).
• Delete / Prune Turn: Allow removing specific failed assistant turns or user messages from session.jsonl.
• Edit User Message: Allow editing a previous prompt and re-running the trajectory from that point.

---

2. Provider Output Sanitization & Compatibility Handling

Problem:

• Some reverse proxies and aggregator endpoints inject telemetry/metadata directly into the content field (e.g., <details><summary></summary>```Response ID / Request ID```</details>).
• This breaks the chat markdown layout (spawning empty code boxes) and frequently corrupts dsh agentic tool-calling JSON parsers (e.g., missing required argument errors on Read or Ask question).
• Additionally, varying reasoning_effort requirements (e.g., TokenRouter expecting xhigh instead of high/max) trigger 400 Bad Request errors unless manually mapped.

Proposed Solution:

• Output Sanitization Regex / Middleware: Introduce an optional provider-level config (e.g. in settings.yaml under compat.stripHtmlTags: true or compat.cleanResponsePatterns) to strip injected <details> metadata blocks before markdown rendering and tool parsing.
• Flexible Reasoning Effort Fallback: Automatically omit or remap reasoning_effort if unsupported by the target provider.

---

Why this benefits dsh:

• Greatly improves agent resilience when working across diverse OpenAI-compatible endpoints and local models.
• Saves tokens and prevents degraded context without forcing users to abandon entire sessions with + New Session.