Repro: Started a coding task with agency copilot. After auto-compaction fired, the agent re-read the compaction summary and started re-planning instead of executing. This repeated through 217 plan→compact→re-plan cycles before I killed the session.
Root cause (hypothesis): Compaction summarizes "we planned X, Y, Z" into the new context. The agent reads the summary and treats the plan as a fresh prompt to plan against, instead of as decided work to execute. The plan becomes self-perpetuating. With no execution-bias instruction surviving compaction, the loop is structural — not prompt-specific.
Expected behavior: After compaction, agent resumes from the next concrete file edit. Plans already produced are treated as decided, not as a fresh prompt to re-plan against.
Actual behavior: Agent re-summarizes its own plan, validates it, repeats. No file edits across hundreds of turns.
Impact:
- Burned through hours of compute and Copilot quota
- Zero code written despite hundreds of turns
- Severe trust hit for autonomous workflows
- Likely under-reported — users blame themselves before blaming the tool
Proposed fixes (priority order):
- Treat plans in compaction summaries as DECIDED — do not re-plan. Surface "next concrete action" as the primary continuation signal.
- Inject a default post-compact rule: "Resume from the next file edit. Do not re-validate the plan." (Reference: https://github.com/Dicklesworthstone/post_compact_reminder)
- Add an
--mode execute / execution-mode session flag that biases hard against planning preambles.
- Loop detector: N consecutive turns with zero file edits → halt and prompt user (suggested default N=3).
- Improve
/compact defaults: drop planning history, keep only files modified, current error state, next concrete action.
- Add telemetry signal for "planning loop suspected" to quantify real-world frequency.
Current workarounds (should not be required):
- AGENTS.md with explicit "no re-plan after compact" rule
- Post-compact reminder hook (manual third-party install)
- Manual /compact at ~60% with explicit preservation instructions
- HANDOFF.md as ground-truth state file
Happy to share session logs and jump on a triage call to walk through the repro.
— Akhi Shukla, Senior PM, IC3 / UED (Prague)
Manager: Bryan Castro | Skip: Nils Dussart
Repro: Started a coding task with
agency copilot. After auto-compaction fired, the agent re-read the compaction summary and started re-planning instead of executing. This repeated through 217 plan→compact→re-plan cycles before I killed the session.Root cause (hypothesis): Compaction summarizes "we planned X, Y, Z" into the new context. The agent reads the summary and treats the plan as a fresh prompt to plan against, instead of as decided work to execute. The plan becomes self-perpetuating. With no execution-bias instruction surviving compaction, the loop is structural — not prompt-specific.
Expected behavior: After compaction, agent resumes from the next concrete file edit. Plans already produced are treated as decided, not as a fresh prompt to re-plan against.
Actual behavior: Agent re-summarizes its own plan, validates it, repeats. No file edits across hundreds of turns.
Impact:
Proposed fixes (priority order):
--mode execute/ execution-mode session flag that biases hard against planning preambles./compactdefaults: drop planning history, keep only files modified, current error state, next concrete action.Current workarounds (should not be required):
Happy to share session logs and jump on a triage call to walk through the repro.
— Akhi Shukla, Senior PM, IC3 / UED (Prague)
Manager: Bryan Castro | Skip: Nils Dussart