Agent no follow rules after compact session

[0x524c]

I have a Gitops project with Fluxcd used in two Kubernetes clusters. I can't allow the agent to commit and push to GitHub without manual review first, as this could lead to something going wrong and compromising the production infrastructure.
I have an explicit rule prohibiting commit+push, and this works fine until the session is compacted and a new session starts.
It seems that after compaction, the rules are lost. I believe a safer approach would be to compact the conversation and include the user-defined rules in their entirety in the new context after compaction. Otherwise, it poses a significant risk in some use cases.

I have some rules in an AGENTS.md file and also in a custom agent called coder.

This seems like a bug, or at least a shortcoming, in the current workflow.

Is there a way I can explicitly configure it somewhere to ALWAYS include the agent contents and local settings after compaction?

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Session automatically compacted, destroying the entire working context #2945: Session automatically compacted, destroying the entire working context - Reports that compaction causes the agent to forget everything about the session
• Model in BUILD mode does not have enough context to continue after compaction #3031: Model in BUILD mode does not have enough context to continue after compaction - Specifically mentions suspicion that AGENTS.md is not included in context after compaction, which directly relates to your issue with rules being lost
• Soft compaction / AI global workspace metabolism #3032: Soft compaction / AI global workspace metabolism - Discusses how current compaction replaces entire context with brief summary, losing AI assistant "personality" and requiring re-feeding of nuances

Feel free to ignore if none of these address your specific case about GitOps rules and commit/push restrictions being lost after compaction.

[0x524c]

I see some opportunities for improvement in the current compact solution.

1. Introduce a new RulesPart type to preserve AGENTS.md rules across summarization:

When constructing the context to send to the LLM in createUserMessage(), we can introduce a new RulesPart type with a pinned field. If it's the first message, it will be marked as part.type = "rules" and pinned: true (the first message in the context contains critical instructions for LLM's attention). The summary for LLM would always preserve pinned rules + msgs.slice(i).

2. In summarize.txt, include preserve custom rules and maintain agent-specific constraints.
   It would look like this:

You are a helpful AI assistant tasked with summarizing conversations.

When asked to summarize, provide a detailed but concise summary of the conversation.
Focus on information that would be helpful for continuing the conversation, including:
  - What was done
  - What is currently being worked on
  - Which files are being modified
  - What needs to be done next
+ - Preserve custom rules from AGENTS.md
+ - Maintain agent-specific constraints

I'm doing some testing on a local fork.

Something like this:

// New RulesPart type in message-v2.ts:
export const RulesPart = PartBase.extend({
  type: z.literal("rules"),
  rules: z.string(),
  source: z.enum(["AGENTS.md", "custom-agent", "user-instruction"]),
  pinned: z.boolean().default(true),
  hash: z.string(),
}).meta({
  ref: "RulesPart",
})
export type RulesPart = z.infer<typeof RulesPart>

// Enhanced filterSummarized() (message-v2.ts):
export function filterSummarized(msgs: { info: MessageV2.Info; parts: MessageV2.Part[] }[]) {
  const i = msgs.findLastIndex((m) => m.info.role === "assistant" && !!m.info.summary)
  if (i === -1) return msgs.slice()

  // Extract messages before the last summary that contain pinned rules
  const pinnedRulesMessages = msgs
    .slice(0, i)
    .filter((msg) => msg.parts.some((part) => part.type === "rules" && part.pinned))

  // Return: pinned rules messages + all messages from summary onwards
  return [...pinnedRulesMessages, ...msgs.slice(i)]
}

[rekram1-node]

. I can't allow the agent to commit and push to GitHub without manual review first,

ill dig into this to double check AGENTS.md is being persisted, but is this problem not better suited for permissions?

[rekram1-node]

@524c can;t you do this in global opencode config? (this is copied from my config that's why rm is in there )

{
  "$schema": "https://opencode.ai/config.json",
  "permission": {
    "bash": {
      "rm *": "ask",
      "git push *": "ask",
      "git commit *": "ask",
    },
  },
}

[0x524c]

@524c can;t you do this in global opencode config? (this is copied from my config that's why rm is in there )

{
  "$schema": "https://opencode.ai/config.json",
  "permission": {
    "bash": {
      "rm *": "ask",
      "git push *": "ask",
      "git commit *": "ask",
    },
  },
}

Good point, I'll try to work more on permissions. This restriction, specifically for git push, is due to the gitops nature of some projects. In other cases, it's not a problem for agent/llm to bypass a rule and perform a git push.

With the modifications I made to my fork, after a 3 hours session and many compactions I noticed a significant improvement in the agent's behavior, due to forcing the first message with an uncompressed rich context.

[0x524c]

@524c can;t you do this in global opencode config? (this is copied from my config that's why rm is in there )

{
  "$schema": "https://opencode.ai/config.json",
  "permission": {
    "bash": {
      "rm *": "ask",
      "git push *": "ask",
      "git commit *": "ask",
    },
  },
}

But for now, I'm just thinking out loud. It was just an idea I wanted to share. I'll continue doing more local testing.
You can close the issue if you prefer 🙃

[rekram1-node]

ill do a glance over code and some simple test cases just to make sure everything behaves correctly before closing!

[0x524c]

I analyzed the code further and saw that AGENTS.md is injected via the system prompt.

But my empirical experience has shown me that injecting the AGENTS.md into the conversion yields better results. My thoughts on this behavior are:

1. Session starts => AGENTS.md at system prompt
2. User: "Don't commit"
3. Agent: "Understood, I won't commit/push"
4. [Multiple tasks...]
5. SUMMARY => Removes these acknowledgements
6. System prompt still has AGENTS.md BUT conversation has lost its "active contract"

With RulesPart pinned:
• Rules SURVIVE explicit summarization
• Remain visible as part of the conversation's "active contract"

LLMs can treat these two types differently:

• System Prompt: "These are background instructions"
• Conversational Messages: "This is what we are actively discussing and following"

After multiple summarizations, the system prompt can become "invisible infrastructure" while pinned messages remain conversation are treated as "active explicit agreements".

[rekram1-node]

yeah i see what u mean

[rekram1-node]

i think the simple fix for now is just tell the compact prompt to preserve instructions/rules the user set during convo

[0x524c]

i think the simple fix for now is just tell the compact prompt to preserve instructions/rules the user set during convo

Yes, I think it can help a lot.
Something like this in the summary prompt:

including:
 - What was done
 - What is currently being worked on
 - Which files are being modified
 - What needs to be done next
+- Preserve custom rules from AGENTS.md
+- Maintain agent-specific constraints and deletion protection rules

[0x524c]

I'll continue testing these two approaches:

• just improving the summary prompt
• injecting AGENTS.md into the conversation and removing it from the system prompt.

With these two approaches, my Opencode fork now strictly follows the instructions in my AGENTS.md, which tell it to only run git push after explicit authorization.

It always asks for permission now, even after a long session of 8+ hours and dozens of conversation compactions/summaries 🙌

I could even block this through permissions, but that would still be trying to plug a leak.

[rekram1-node]

injecting AGENTS.md into the conversation and removing it from the system prompt.

I don't think we wanna do this one but the other one sounds good

[0x524c]

Thanks, updating the summary prompt will help a lot.