Progress & Recovery

Progress & Recovery

Techniques for transparency, state management, and error recovery during long research tasks.

Inspired by: Gemini's thinking panel, Manus's context engineering, DeepSeek's extended thinking

Contents

• Progress Markers
• Interim Save
• Error Recovery
• Context Management

Progress Markers

Purpose: Show research progress during long tasks.

When to use:

• Deep/Exhaustive tier (always)
• Research exceeds 5 minutes
• User requests visibility

Format:

> 🔍 **Phase 3:** Retrieved 15 sources, identified 3 gaps, refining queries...
> ✅ **Phase 4:** Verified 8 C1 claims, 2 contradictions flagged for Red Team
> 🔄 **Phase 3.5:** Gap analysis complete, executing 4 follow-up queries
> ⚠️ **Backtrack:** Dead-end on [topic], pivoting to alternative angle

Rules:

• Keep updates to single line
• Include phase number
• Show concrete numbers (sources, claims, gaps)

Interim Save

Purpose: Save progress to prevent context loss.

When to save:

• After Phase 4 (TRIANGULATE) completion
• After 20+ sources collected
• Before Red Team phase
• When context approaches limit

Save Format:

## Interim Research State

**Timestamp:** [ISO timestamp]
**Phase completed:** [Current phase]
**Sources collected:** [N]
**C1 claims verified:** [N]

### Key Findings So Far
1. [Finding 1] — Confidence: [LEVEL]
2. [Finding 2] — Confidence: [LEVEL]

### Gaps Remaining
- [Gap 1]
- [Gap 2]

### Failed Paths (avoid repeating)
- ❌ [Query 1] → [Reason]
- ❌ [Query 2] → [Reason]

### Next Steps
1. [Action 1]
2. [Action 2]

Save Location:

• For file output: Append to research document
• For conversation: Include in context for continuation

Error Recovery

If research is interrupted:

1. Check interim save for last known state
2. Resume from last completed phase
3. Avoid re-searching failed paths (check error trace)
4. Validate existing claims still hold (for time-sensitive topics)

Error Trace Format:

### Failed Paths (avoid repeating)
- ❌ "[query]" → Paywall, no accessible content
- ❌ "[query]" → Results outdated (>6 months)
- ❌ "[query]" → Tangential, not relevant to core question

Context Management

From Manus's context engineering learnings:

Attention Anchoring

Problem: "Lost-in-the-middle" — model forgets objectives in long context.

Solutions:

• Keep active todo/task list updated throughout research
• Recite key objectives periodically
• Place critical info at context start AND end

Memory Extension

Use file system for:

• Large interim data (>20 sources)
• Detailed source notes
• Failed path logs

Compression strategies:

• Compress observations while preserving key pointers (URLs, dates)
• Keep summaries in context, details in files
• Reference files by path, don't inline large content

Failed Path Tracking

Keep failed searches in context to avoid:

• Repeating same unsuccessful queries
• Hitting same paywalls
• Searching outdated terms

### Failed Paths
- ❌ "GPT-4 features 2024" → Outdated, use "GPT-5 features 2025"
- ❌ "OpenAI pricing page" → 403, use press releases instead
- ❌ "[product] API docs" → Requires auth, use r.jina.ai fallback

Tier-Specific Recommendations