docs(bugfixes): PreCompact deadlock writeup for #863#915
docs(bugfixes): PreCompact deadlock writeup for #863#915Robins163 wants to merge 2 commits intoMemPalace:developfrom
Conversation
Standalone writeup of the symptom, root cause, and fix published in MemPalace#863 (mvalentsev). Includes the session-log repro, the "decision: allow is invalid" correction from MemPalace#872, the escape hatch for anyone stuck on an unpatched version, and the rationale for removing the block entirely vs keeping it with a stateful guard. Net-new file — docs/bugfixes/ does not exist on develop yet.
|
Been thinking about this. The realistic repeat scenario here is an LLM agent making the same mistake, not a human dev. The |
|
Fully agree — that's the right framing. The "allow" mistake in #872 was an agent writing a return value without checking the spec, and the fix for that class of mistake has to live where agents actually load context, not in I checked Proposing to add a short Hook return contract section to
The Does that scope extension work for you on this PR, or would you prefer I split it into a separate PR so #915 stays docs-only? |
Document the hook return value rules that caused the PreCompact deadlock (MemPalace#863, MemPalace#856, MemPalace#858) and the undocumented "allow" value (MemPalace#872). AGENTS.md picks this up automatically via symlink. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Follow-up to #867 / #863 as discussed in #867 (comment).
Standalone writeup of the PreCompact deadlock — symptom, root cause, and the fix that shipped in #863 (@mvalentsev). Net-new file; `docs/bugfixes/` does not exist on develop yet.
Contents
Branch rebased on `upstream/develop`. Tests aren't included — #863's `test_precompact_allows` already covers the shape.