synthesis-checkpoint

Synthesis Checkpoint — Mid-Session Refresh & Drift Recovery

Refresh-and-report mode

When asked to refresh after an ecosystem upgrade or report session readiness, load this skill from the current verified installed root, then follow references/refresh-and-report.md. This mode uses a deterministic local inspector and optional campaign feedback. Project feedback destinations follow explicit registry successor links; the reporting session keeps its project identity and execution restrictions. It ends after reporting and grants no project implementation or repair authority. Ordinary checkpoints use the protocol below and do not send campaign feedback.

Current skill text and a native runtime reload are distinct evidence. Never equate an old in-context skill body with a failed current startup, or copying new files with proof a running harness reloaded them.

The Problem

LLMs are stateless at the model level. Across a long conversation, an LLM's sense of "what's true now" can drift from what's actually on disk and in git. Drift sources:

• Time drift. The model has no clock. If the conversation started Tuesday and continues Wednesday, the model often still thinks it's Tuesday.
• State drift. Project CONTEXT.md may have been edited (by user, by sub-agent, by another session) since the model last read it.
• Compaction drift. When the context window approaches its limit, the harness may summarize older turns. Details disappear. The model often cannot detect that this happened.
• Cached-fact drift. Facts the model read earlier in the conversation (last commit, last session date, last decision) may no longer be true.
• Coordination drift. Another live root session may have claimed or changed the same source area since the last tool call.

Self-discipline degrades with conversation length. Rules read at session start lose salience as turns accumulate. Even a model that "knows" to verify often skips verification under the weight of accumulated context.