This ticket is a follow-up to #7660. It aims to make the memory protocol more resilient by explicitly addressing two key failure modes: technical failures (un-savable turns) and behavioral failures (skipped turns).
1. Clarify Recovery for "Un-savable Turns":
The current "Session Recovery Protocol" is robust, but it's not explicitly linked to the scenario where a hard API error aborts a turn before the "Consolidate-Then-Save" step can be reached.
- Action: We will make the connection between the problem and the existing solution explicit.
- Rename the section
### Step 3.1: Session Recovery Protocol to ### Step 3.1: Protocol for Recovering from Un-savable Turns.
- Add a preamble to this section that clearly defines an "un-savable turn" and frames the existing recovery procedure as the specific, mandatory remedy for this scenario.
2. Mitigate "Skipped Turns" (The "Focus Window" Problem):
This addresses the behavioral issue where an agent "forgets" to save a turn. The existing recovery protocol cannot fix this, as the agent is not even aware it failed.
- Action: We will introduce the concept of a "Pre-Flight Check" into the main "Consolidate-Then-Save" protocol. This will serve as a cognitive forcing function.
- A new instruction will be added: Before executing any significant file modification (e.g.,
replace, write_file), the agent MUST add a "Pre-Flight Check" to its thought process, explicitly stating its plan to save the consolidated turn before executing the file change.
This ticket is a follow-up to #7660. It aims to make the memory protocol more resilient by explicitly addressing two key failure modes: technical failures (un-savable turns) and behavioral failures (skipped turns).
1. Clarify Recovery for "Un-savable Turns": The current "Session Recovery Protocol" is robust, but it's not explicitly linked to the scenario where a hard API error aborts a turn before the "Consolidate-Then-Save" step can be reached.
### Step 3.1: Session Recovery Protocolto### Step 3.1: Protocol for Recovering from Un-savable Turns.2. Mitigate "Skipped Turns" (The "Focus Window" Problem): This addresses the behavioral issue where an agent "forgets" to save a turn. The existing recovery protocol cannot fix this, as the agent is not even aware it failed.
replace,write_file), the agent MUST add a "Pre-Flight Check" to itsthoughtprocess, explicitly stating its plan to save the consolidated turn before executing the file change.