Four context moments
Illustrative reconstruction. These boxes show categories, not token sizes. Check your own visible reads and native display.
1. Bootstrap
- Instructions, prompt and history or summary
- Root instructions and setup prompt
- File excerpts and returned tools
- Any read manifest or instruction excerpt
- Unknown
- Exact hidden assembly and unread dataset contents
Mark each item observed, reconstructed or unknown in your own report. Conversation may be compacted into a summary; exact hidden contents are unavailable.
2. Inventory tool result
- Instructions, prompt and history or summary
- Current prompt and visible conversation
- File excerpts and returned tools
- Returned filenames, sizes and any sampled excerpt
- Unknown
- Unopened file contents; exact occupancy
Mark each item observed, reconstructed or unknown in your own report. Conversation may be compacted into a summary; exact hidden contents are unavailable.
3. Nested instructions and skill
- Instructions, prompt and history or summary
- Explicit nested AGENTS.md read; skill description and body if read
- File excerpts and returned tools
- Source-row excerpt and check result
- Unknown
- Automatic harness loading unless observed
Mark each item observed, reconstructed or unknown in your own report. Conversation may be compacted into a summary; exact hidden contents are unavailable.
4. Fresh final-review thread
- Instructions, prompt and history or summary
- New prompt; explicitly reloaded root, skill, progress and stage JSON
- File excerpts and returned tools
- Saved results and safe numeric summary
- Unknown
- Prior chat history unless explicitly supplied
Mark each item observed, reconstructed or unknown in your own report. Conversation may be compacted into a summary; exact hidden contents are unavailable.
A skill description and its loaded body are separate items. Saved instructions need loading. Capacity, cumulative tokens and current occupancy are different measurements. Leave numeric size unknown when no native field is available.
Resume in a fresh thread by reading AGENTS.md, the reproducibility skill, records/progress.md and the latest stage JSON. Retain that thread’s actual result.