Context-Reanchor — Prevent Intra-Session Drift

Register: DISCIPLINE Goal: Restore instruction precision before the session drifts away from loaded skills and constraints. Core principle: agents do not only forget between sessions; they drift within long sessions.

Announce: "Using context-reanchor to refresh active obligations before continuing."

Trigger

Run when any of these are true:

• conversation turn > 20;
• multiple large files were loaded;
• before phase transition: design -> implementation -> review -> completion;
• after 2+ failed attempts;
• before a completion claim in a long session;
• active skill was loaded many turns ago;
• evidence anchors are scattered or unclear.

Budget mode reduces verbosity. Context-reanchor restores obligation precision. They are related but not the same.

Reanchor Record

Keep it compact:

RE-ANCHOR
Goal:
Active skill obligations:
Non-negotiable constraints:
Evidence anchors:
Still T4 / ASSUMED:
Next gate:
Need to re-read nano/full skill: yes/no

If any critical obligation is uncertain, re-read the relevant nano/full skill before proceeding.

Stop Conditions

Stop and resolve before continuing if:

• the next step depends on T4 evidence;
• the active goal changed without an ADR/spec update;
• repeated failure suggests systematic-debugging or DPS Spec-Conflict Signal;
• completion would be claimed without verification-before-completion.

Gotchas — Observed Failure Points

<!-- Populated by knowledge-compound after cycles where this skill underperformed --> <!-- New entries should follow shared/gotcha-schema.md -->