[FEATURE]: Debugging loop detection and cross-session memory

[openchat-ai]

DescriptionTwo problems that compound each other during multi-session debugging sessions.## Problem A — Hypothesis loop detectionWhen debugging a test failure, the model often proposes a hypothesis, changes code, runs tests, sees the same failure, and proposes a new hypothesis — all at the source-code layer. It never descends to examine evidence (logs, core dumps, network traces) because nobody tells it to.opencode sits between the model and the tools. It can detect when the model has made N attempts at the same layer without success and intervene with a suggestion to go deeper.## Problem B — Cross-session memoryEvery new session re-reads files already read in the previous session. Same files, re-read multiple times across sessions debugging the same CI failure. This wastes context and time.## Proposed SolutionProblem A: Add a hypothesis tracker to the conversation loop:- Track (hypothesis text hash, layer indicator) pairs per conversation- After 2 falsified hypotheses at the same layer, inject a system message: "You have tested 2 hypotheses at the source-code layer without success. Consider examining the actual evidence: logs, core dumps, strace output."- This is purely an opencode loop-level feature — the model doesn't need changesProblem B: Add a .opencode/session.json cache:- On session end, write: branch name, goal description, list of files read with their mtime- On session resume with the same branch, skip re-reading files whose mtime hasn't changed- Also track the last hypothesis so the model can pick up where it left off- This is an opencode session management feature## Verification- [x] I have checked that this feature has not been suggested before

[github-actions]

Thanks for updating the issue! It now follows the feature request template — the title, description, and verification checkbox all look good. We'll take it from here.

[hjqcan]

Problem B maps closely to a boundary we've been working on in GoodMemory (I'm the maintainer): durable, inspectable cross-session memory for coding agents through MCP.

OpenCode already works with it through a local MCP block; the OpenCode-specific setup is here: https://github.com/hjqcan/GoodMemory/blob/main/docs/GoodMemory-OpenCode-Setup-Guide.md

Quick start:

npm install -g goodmemory@0.7.1

Then add the documented mcp.goodmemory block to opencode.json and verify with goodmemory_stats. Writes are opt-in; the default surface is read-only.

This would not solve Problem A's native hypothesis-loop detection, nor the proposed file-mtime cache. Those still belong in OpenCode. It can, however, preserve goals, decisions, corrections, and debugging hypotheses across sessions so the next session does not start from zero.