context-health

name	context-health
description	Use during long-running sessions or complex multi-step tasks -- detects context window degradation, task drift from original request, and offers recovery strategies including checkpoint restart and re-anchoring

Context Health

Detects when a long conversation or multi-step task is degrading due to context window pressure, task drift, or tool failure spirals. Provides recovery strategies: truncate, checkpoint-restart, re-anchor, or warn. Zero LLM cost, sub-millisecond latency -- pure heuristics and token estimation.

Native Context Awareness (Claude 4.5+ / Opus 4.6)

Claude 4.5 and 4.6 models natively track their remaining context window. This changes how context health works:

What the model does natively:

• Tracks remaining token budget throughout conversation
• Can self-manage context when told about compaction behavior
• Discovers state from the filesystem after context resets (extremely effective on Opus 4.6)
• Naturally wraps up work as context limit approaches (unless told otherwise)

What this skill still provides:

• Task drift detection (models don't self-detect drift reliably)
• Tool failure rate tracking (models don't aggregate failure patterns)
• Structured checkpoint format for cross-session continuity
• Recovery strategy selection (truncate vs restart vs re-anchor)

Tell the model about compaction when using an agent harness that compacts:

Your context window will be automatically compacted as it approaches its limit,
allowing you to continue working indefinitely. Do not stop tasks early due to
token budget concerns. Save progress to files before the window refreshes.

Prefer starting fresh over compacting when possible. Opus 4.6 is extremely effective at rediscovering state from the filesystem -- reading progress files, git logs, and test results. A fresh context with good state files often outperforms a compacted context with degraded history.

When to Check Context Health

• After every N tool calls (every 5-10 is a good interval for complex tasks)
• After producing long outputs (each long output eats significant context)
• After encountering errors -- errors compound; 2+ consecutive failures is a signal
• When your output starts feeling repetitive or off-topic -- you may have drifted
• At the start of each major step in a multi-step task
• When the user says "you already did that" or "that's not what I asked" -- you have drifted

Model-Aware Token Thresholds

Different models degrade at different context utilization levels. These thresholds are based on empirical quality degradation curves.

Full Tier (resilient, less than 5% quality loss across wide range)

Models: opus-4.6, sonnet-4.6, opus-4.5, sonnet-4.5, opus, sonnet, gpt-4o, gpt-4, gpt-5, gemini-pro, gemini-2.5-pro, gemini-3-pro, claude-4, claude-3.5-sonnet, o1, o3

Note: Opus 4.6 has a 1M context window and handles degradation even more gracefully than earlier models. It may not need active health checks until much later thresholds.

Zone	Token Threshold	Action
Sweet spot	0 -- 80,000	Continue normally
Approaching limit	80,000 -- 120,000	Check for drift, warn if detected
Compact zone	120,000 -- 180,000	Truncate old context
Hard limit	180,000+	Must restart with checkpoint

Compact Tier (degrades earlier, ~50% quality drop at upper range)

Models: haiku-4.5, haiku, gpt-4o-mini, gemini-flash, deepseek, codestral, grok, mistral-small, mistral-medium, gemini-2.0-flash, gemini-2.5-flash, gemini-3-flash, perplexity-sonar, groq-llama, claude-3.5-haiku

Zone	Token Threshold	Action
Sweet spot	0 -- 20,000	Continue normally
Approaching limit	20,000 -- 30,000	Check for drift, warn if detected
Compact zone	30,000 -- 50,000	Truncate old context
Hard limit	50,000+	Must restart with checkpoint

Local Tier (small context windows, aggressive thresholds)

Models: llama, mistral-7b, codellama, phi-, qwen-

Zone	Token Threshold	Action
Sweet spot	0 -- 8,000	Continue normally
Approaching limit	8,000 -- 16,000	Check for drift, warn if detected
Compact zone	16,000 -- 32,000	Truncate old context
Hard limit	32,000+	Must restart with checkpoint

Unknown models default to Compact Tier -- conservative is safer than optimistic.

Token Estimation

Quick heuristic: ~4 characters per token. Count the total characters in all messages and divide by 4. This is accurate within ~10% for English text and is sufficient for threshold decisions.

For messages with tool use blocks, also count JSON.stringify(input) content and tool result text.

The Decision Tree

This is the core logic. Evaluate in order -- first match wins.

1. Cooldown active?
   YES --> continue (prevent restart spam after recent restart)

2. Tool failure rate > 40% (with at least 4 tool calls)?
   YES --> restart (agent is confused, context is likely corrupted)

3. Tokens >= hard limit?
   YES --> restart (severe degradation, output quality is collapsing)

4. Tokens >= compact zone AND 2+ prior truncations?
   YES --> restart (truncation has diminishing returns, start fresh)

5. Tokens >= compact zone?
   YES --> truncate (remove old messages, keep recent work)

6. Tokens >= sweet spot AND task drift < 0.15?
   YES --> warn (agent has drifted from original request)

7. Otherwise --> continue

Circuit Breaker: Maximum 3 Restarts

To prevent infinite restart loops, cap restarts at 3 per session. After 3 restarts, downgrade all "restart" actions to "truncate" instead. If truncation also fails to help, warn the user that the task may be too large for a single session.

Task Drift Detection

Drift detection measures how much your current output relates to the original request.

Method: Asymmetric word overlap

Compare the words in the original instruction against the words in the last ~2000 characters of output.

drift_score = |instruction_words INTERSECT recent_output_words| / |instruction_words|

• Score of 1.0 = all instruction keywords present in recent output (on-topic)
• Score of 0.15 or below = significant drift (most instruction keywords absent from recent output)
• Only the last 2000 characters of recent output are checked (not diluted by long history)

This is asymmetric on purpose: a 500-word response to a 4-word instruction scores high as long as it mentions the key words. Unlike Jaccard similarity, long outputs do not dilute the score.

Drift only triggers a warning when tokens exceed the sweet spot threshold. Short conversations do not need drift checks.

Recovery Strategy 1: Truncate

Remove old messages while preserving the ability to continue.

What to keep:

1. System prompt (always)
2. Original user instruction (always)
3. Checkpoint summary (injected after instruction)
4. Last 2 message pairs (most recent context)

Checkpoint summary captures:

• Original task (first 300 characters of instruction)
• Tools used so far (up to 10 tool names)
• Key findings from tool results (last 5, truncated to 200 chars each)
• Last progress (most recent assistant output, first 500 chars)
• Target: 300-500 tokens (~1200-2000 characters)

Checkpoint message format:

[CONTEXT CHECKPOINT -- PREVIOUS PROGRESS]
Original task: {instruction}
Tools used: {tool_names}
Key findings:
  - {finding_1}
  - {finding_2}
Last progress: {last_assistant_output}
[END CHECKPOINT]

Continue from where you left off. Do not repeat completed work.

Track truncation count. After 2 truncations, the next compact-zone trigger escalates to restart (diminishing returns).

Recovery Strategy 2: Checkpoint Restart

A complete context reset with full knowledge carry-forward. Used when truncation is no longer effective.

Fresh context structure:

1. System prompt (original)
2. User message containing: original instruction + checkpoint + continuation directive

Restart message format:

{original_instruction}

[CONTEXT CHECKPOINT -- PREVIOUS PROGRESS]
{checkpoint_summary}
[END CHECKPOINT]

Continue from where the previous execution left off.
Do not repeat completed work. Pick up at the next step.

After restart:

• Activate a 3-iteration cooldown (no health checks for 3 cycles to let the fresh context stabilize)
• Increment restart counter (toward the 3-restart circuit breaker)
• Reset step counter to 0

Recovery Strategy 3: Re-Anchoring

Lighter than truncate or restart. Injects the original request back into context when drift is detected, without removing any messages.

When to re-anchor:

• Health check returns "warn" with task drift detected
• User feedback indicates you have gone off-topic
• You notice your own output diverging from the original request

Re-anchor message format:

[RE-ANCHOR -- Context drift detected]

Your recent output has drifted from the original request.
Reason: Task drift detected (similarity=0.12) at 85,000 tokens

ORIGINAL REQUEST (verbatim):
{original_instruction}

Refocus on satisfying this specific request.
Do not repeat completed work -- continue from where you are,
but ensure your output addresses the original request above.

Recovery Strategy 4: Warn User

When context health is degrading but not yet critical, surface the concern:

• "This task is approaching context limits. Consider breaking remaining work into a follow-up session."
• "Tool failure rate is high (40%+). The remaining context may be unreliable."
• "Output appears to have drifted from the original request."

Tool Failure Tracking

Track tool calls and failures across the session:

Metric	Threshold	Signal
Failure rate > 40%	Minimum 4 tool calls	Agent is confused or context is corrupted
Consecutive failures 3+	Any count	Likely a systematic issue, not transient
Same tool failing repeatedly	2+ failures of same tool	Tool-specific problem, try alternative approach

High tool failure rates often indicate the agent's context has degraded enough that it is generating malformed tool calls, wrong parameters, or calling tools inappropriately.

Checkpoint Format Reference

When you need to capture state before a restart or at a manual checkpoint:

## Checkpoint: {task_name}
**Status:** {percentage or step X of Y}
**Completed:**
- {what was done}
- {what was done}
**Key decisions:**
- {why you chose approach X over Y}
- {important findings that affect remaining work}
**Remaining work:**
- {next step}
- {remaining steps}
**Blockers:**
- {anything that prevented progress}

Integration with Other Skills

• completion-enforcer checks whether the final output is complete. Context health checks whether the environment can still produce quality output.
• constitutional-ai enforces safety rules regardless of context health. Even degraded contexts must respect constitutional boundaries.
• hallucination-detector catches fabricated content, which becomes more likely as context degrades. Run hallucination detection more aggressively when context health is in the compact zone or beyond.

Cost

$0. Zero LLM calls. All checks are character counting, word overlap computation, and threshold comparison. Apply on every iteration of long-running tasks.