Summary

When using agy as the active provider, long or tool-heavy turns can pass through an additional AGY checkpoint/compaction layer. After that checkpoint, the model may still complete the immediate tool task, but it can drop post-tool obligations that cli-jaw injected in the system prompt / pre-prompt hook, such as recording durable system events before the final response.

This looks different from ordinary context loss: the current user task is still understood and executed, but the final response is produced without re-checking the obligations that were supposed to apply after tool execution.

Observed behavior

A recent turn looked like this:

1. User reported that an LXC disk resize/restart was complete and asked the agent to verify the state.
2. cli-jaw spawned agy with the current prompt and pre-prompt runtime context.
3. The agent ran verification commands such as df -h / and pm2 list.
4. AGY emitted an internal checkpoint/compaction event in the transcript between tool calls and the final answer.
5. The agent correctly answered that the filesystem and jaw-server were healthy.
6. However, it did not run the diary/recording procedure even though the turn contained durable system-state verification and should have triggered the configured record gate.

In other words, the tool work succeeded, but the post-action obligation was lost after the AGY checkpoint.

Why this matters

For long-running companion/operations agents, post-tool obligations are important:

• record durable system changes or incident recovery
• update state files after verification
• avoid saying a task is complete before persistence checks are done
• emit [SILENT] or a user-visible heartbeat response according to heartbeat policy

With agy, cli-jaw appears to act more like a mediator/toolbox around an external AGY session. AGY also maintains its own transcript/checkpoint state, so cli-jaw's prompt assembly and AGY's internal compaction are not always aligned.

This makes lightweight models particularly vulnerable: they may perform the immediate task but forget the final post-tool checklist.

Expected behavior

After tool execution and before final answer delivery, cli-jaw should have a provider-level way to reassert short post-action obligations, especially for AGY print-mode sessions.

Examples:

• append a small final reminder before response synthesis / delivery
• re-inject a concise post-tool hook after AGY checkpoint/resume boundaries
• expose a provider hook such as beforeFinalResponse / postToolReminder
• preserve and reassert obligations from the pre-prompt hook after compaction-like transcript events

Local mitigation that helped

We added a short pre-prompt runtime reminder, for example:

Before the final response, re-check this turn's tool results. File/package/settings/automation changes, incident recovery, df/pm2/service-state verification, and user operational decisions are recording triggers. If matched, run the diary-recording procedure before answering.

This helps on fresh turns, but it is still only pre-prompt context. If AGY checkpoints between tool execution and final response, a post-tool/final-response hook would be more reliable.

Related notes

This is related to earlier AGY integration edge cases where cli-jaw successfully passed context at spawn time, but AGY's own session/checkpoint layer changed what remained salient later in the turn. It is not necessarily an AGY bug alone; it is an integration boundary issue between cli-jaw prompt assembly and AGY's internal session lifecycle.