create-context Reference

Manifest Schema (.context/session/ctx/manifest.yaml)

project: string                    # Auto-inferred from directory
created: ISO8601                   # Timestamp of creation
source_folder: .context/session/in/  # Reference to immutable bootstrap folder

sources:
  high:                            # ≤1500 tokens inline, >1500 summarized
    - file: string                 # Path relative to .context/session/in/
      desc: string                 # Human-readable description
      action: inline|summarized    # Auto-set based on size
  medium:                          # ≤2500 tokens inline, >2500 summarized
    - file: string
      desc: string
      action: inline|summarized
  low:                             # Reference only, no content copied
    - file: string
      desc: string

tags:                              # Optional categorization
  category_name: [file1, file2]

Context Sizing Behavior

Token estimation: tokens ≈ words / 0.75 (using wc -w)

25K limit: Files exceeding this use summarize-for-context sub-agent with chunked reads.

Baseline Template (.context/session/CONTEXT-baseline-llm.md)

# Baseline Context

saved: {ISO 8601 timestamp}
stream: baseline
manifest: .context/session/ctx/manifest.yaml
source: .context/session/in/

## Project

ref: .context/session/ctx/manifest.yaml
source_folder: .context/session/in/
files: {total_count}

## Sources (inline)

### {filename}
{Full content from .context/session/ctx/{file}}

## Sources (summarized)

### {filename}
See: `.context/session/ctx/{nn}-{basename}-summary-llm.md`

## Sources (reference only)

- `.context/session/in/{path}`: {description}

## Refs

manifest: .context/session/ctx/manifest.yaml
summaries: [.context/session/ctx/01-*.md, .context/session/ctx/02-*.md, ...]
source: .context/session/in/

Token budget: ≤2000 tokens total.

Validation Rules

1. Required: project, created, source_folder, sources (high/medium/low)
2. File paths: Relative to .context/session/in/, must exist at generation time
3. Descriptions: Required, non-empty, concise (1-2 sentences)
4. Action field: Auto-set, user can override
5. Tags: Optional, keys alphanumeric/hyphens/underscores
6. All priority sections: Must be present (can be empty arrays)

Security

Auto-skipped: .env*, *credentials*, *secrets*, *token*, *.key, *.pem, *.crt, *.p12, *.pfx.

The scanner matches every path component, so a file inside a directory named secrets/ is skipped too. Skipped paths are printed on stderr as SKIPPED (security). Set CONTEXT_DIR to scan a context root other than .context.

RISEN INPUT Table (output)

| Priority | File | Description |
|----------|------|-------------|
| 1 (HIGH) | @.context/session/ctx/proposal.md | Main objectives |
| 2 (MED) | @.context/session/ctx/style-guide.md | Style reference |
| 3 (LOW) | @.context/session/in/old-notes.md | Background (ref only) |

Scripts

• scripts/scan-in-folder.sh [IN_DIR] — scan the in-folder, print JSON with a token estimate per file, skip sensitive paths
• scripts/validate-manifest.sh [MANIFEST] — check required fields, the three priority sections, that listed files exist and that every file has a description; exits 1 on a schema error and 0 with warnings when files are missing

Error Messages

Related

• /save-context [stream] - Save session to named stream
• /load-context [stream] - Load saved stream
• scripts/validate-manifest.sh - Validate manifest schema
• summarize-for-context sub-agent - summarises files above 25K tokens in chunks
• Worked examples - end-to-end runs