A
agentic harnesscreate context

Create Context

Create the baseline session context from the in-folder of raw project documents, classifying each file by priority, writing a manifest, summarising oversized files and producing a token-budgeted baseline. Use when bootstrapping project context or setting up the ctx snapshot once per project. Triggers include "create context", "bootstrap context", "setup context", "init context".

npx skills add github:pantheon-org/tekhne --skill agentic-harness--create-context

Turn the raw documents in the in-folder into a prioritised, token-budgeted snapshot that later sessions load in one read.

Layout

.context/session/in/                      immutable input: users drop documents here, no command edits it
.context/session/ctx/                     generated snapshot: manifest, copies, summaries
.context/session/CONTEXT-baseline-llm.md  single load point, 2000 tokens at most

Mindset

  • Use the in-folder as the source of truth. The manifest mirrors it faithfully.
  • Keep the snapshot scoped to the session’s open questions. Include only what shapes decisions.
  • Keep one manifest per session. Two manifests in one folder leave load-context unable to choose.
  • State the origin of every constraint, assumption and decision that enters the baseline.

When to Use

  • Use it once per project when raw documents sit in the in-folder and no ctx snapshot exists.
  • Use it with --force after the source documents changed materially.
  • Use it when load-context reports a missing manifest.

When Not to Use

  • Do not use it when the ctx snapshot is current. A rerun without --force stops, and --force discards existing summaries.
  • Do not use it when the in-folder is empty. There is nothing to scan.
  • Do not use it to refresh one document. Update that summary and its manifest entry by hand.
  • Use save-context instead to store live session state in a named stream.

Steps

  1. Guard the prerequisites. Stop on a missing in-folder, and stop on an existing ctx folder unless --force was passed.
[ -d .context/session/in ] || { echo "No in-folder found. Add source files first." >&2; exit 1; }
[ -d .context/session/ctx ] && [ "$ARGUMENTS" != "--force" ] && { echo "ctx exists. Use --force." >&2; exit 1; }
  1. Scan the in-folder with the bundled script. It lists md, txt, csv, yaml, yml and json files with a token estimate, and skips sensitive paths.
./scripts/scan-in-folder.sh .context/session/in
  1. Classify every file as HIGH, MEDIUM or LOW with a one-line description. Batch several files into each AskUserQuestion call, and check for blank answers after every call (see the guard below).
  2. Size each file with wc -w, where tokens are words divided by 0.75, then apply the thresholds.
Priority At or under threshold Over threshold, up to 25K Over 25K
HIGH (1500) inline summarise directly summarise via sub-agent
MEDIUM (2500) inline summarise directly summarise via sub-agent
LOW reference only, no copy not applicable not applicable
  1. Copy HIGH and MEDIUM files into the ctx folder and keep their subdirectories. Write a summary of about 500 tokens for each oversized file, and delegate files above 25K tokens to the summarize-for-context sub-agent.
  2. Write manifest.yaml in the ctx folder after sizing, because each action value depends on it. Then validate it.
./scripts/validate-manifest.sh .context/session/ctx/manifest.yaml
  1. Write the baseline within 2000 tokens: inline content, summary pointers and LOW references. Report the file counts and the RISEN INPUT table, then suggest /save-context baseline.

AskUserQuestion Guard

Outside Plan Mode the tool can return blank answers without showing its UI. After every call, check the answers. If they are blank, do not assume a priority. Print the options as a numbered list, ask for the number, and wait for the reply.

Anti-Patterns

  • NEVER build the manifest from memory. WHY: the in-folder defines the scope, so a manifest without a scan misleads every downstream agent.
# BAD: fabricated, no scan, wrong shape
high_priority: []

# GOOD: sources.high/medium/low from a real scan
sources:
  high:
    - {file: design-doc.md, desc: "Architecture decisions", action: inline}
  medium: []
  low: []
  • NEVER omit high, medium or low, even when empty. WHY: validate-manifest.sh fails and consumers cannot iterate predictably.
  • NEVER copy files matching .env*, *credentials*, *secrets*, *token*, *.key, *.pem, *.crt, *.p12 or *.pfx. WHY: the ctx folder gets committed and shared, so a copied secret leaks.
# BAD: copies everything, secrets included
cp -r .context/session/in/* .context/session/ctx/

# GOOD: select with the scanner, which also skips secrets inside a secrets/ directory
./scripts/scan-in-folder.sh .context/session/in
  • NEVER inline a file above its threshold. WHY: one oversized inline breaks the 2000-token baseline budget on every later load.
  • NEVER run the skill concurrently with save-context or load-context. WHY: nothing locks the ctx folder, so the last writer wins and leaves a partial manifest.
  • Pitfall: a filename pattern alone misses secrets/prod.yaml. Check directory names too, as the scanner does.

References

  • Reference - manifest schema, sizing rules, baseline template, validation rules and error messages
  • Worked examples - bootstrap, forced rebuild and large-file runs with expected output
Edit this page on GitHub