context-degradation-review

name	context-degradation-review
description	Review agent work for context poisoning, lost-in-the-middle failures, distraction, context clash, and stale carryover. Use before long-session handoffs, SET runner handoffs, skill changes, memory writes, or any task where too much or conflicting context may degrade behavior.
license	MIT
metadata	{"abvx_status":"experimental","abvx_origin":"adapted"}

Context Degradation Review

Use this skill to find when the context itself is making the agent worse.

Use When

• a long session, handoff, or compaction summary may contain stale or conflicting instructions;
• a SET bundle, AGENTS.md, SKILL.md, or memory write mixes useful facts with outdated task history;
• the agent keeps optimizing for irrelevant prior work;
• tool output, logs, or retrieved docs may bury the decisive evidence;
• multiple sources disagree and the answer treats them as one coherent contract.

Failure Modes

• lost-in-middle: key constraints are present but buried between low-signal material.
• context-poisoning: incorrect, stale, or speculative text is treated as authority.
• distraction: large but irrelevant context pulls work away from the current task.
• context-clash: two instructions, docs, memories, or artifacts conflict.
• stale-carryover: a prior goal continues after the user changed or stopped the task.

Workflow

1. Identify the active task, current authority order, and expected output.
2. List the context sources used: user message, repo files, memory, docs, tool output, web sources, generated summaries.
3. Mark each source as authoritative, supporting, stale-risk, or irrelevant.
4. Check for the five failure modes above.
5. Propose the smallest context repair:
   • remove or ignore a stale source;
   • promote the decisive constraint near the active plan;
   • split task history from durable guidance;
   • ask for clarification only when the conflict blocks action;
   • rewrite a handoff or memory proposal with source boundaries.
6. Continue only after the active context is narrow enough to act safely.

Output Shape

Use:

• verdict: CONTEXT_OK, CONTEXT_DEGRADED, or CONTEXT_BLOCKED;
• active task;
• context sources reviewed;
• degradation findings by mode;
• smallest repair;
• remaining risk.

Guardrails

• Do not use this as a generic writing critique.
• Do not remove context just because it is long; remove or demote it only when it is low-signal, stale, or conflicting.
• Do not treat memory as higher authority than the current user message.
• Do not let stale rollout context override an explicit stop, pause, or changed task.

Provenance

Adapted from context degradation patterns in muratcankoylan/Agent-Skills-for-Context-Engineering, narrowed into an ABVX review lens.