Lab 10 — Compaction and context management

Capability: Codex manages its own context and provides /compact in interactive mode. It does not expose the build-your-own atomic transcript-rewrite contract or guarantee that every fact survives.

Install Codex CLI; create a private temporary CODEX_HOME, copy this lab’s config.toml.example to $CODEX_HOME/config.toml and fill in your endpoint (ending in /openai/v1) and deployment. Set AZURE_OPENAI_API_KEY from your resource key via a secret manager. The Microsoft guide says Entra ID is not supported for Codex; never save the key in files. Workspace: start in lab10-compaction/ and follow the common preflight. Add these synthetic files:

'demo-order uses queue orders-demo' |
  Set-Content (Join-Path $workdir 'order.txt') -Encoding utf8
'demo-makeline consumes queue orders-demo' |
  Set-Content (Join-Path $workdir 'makeline.txt') -Encoding utf8
codex -C "$workdir" --sandbox read-only

printf 'demo-order uses queue orders-demo\n' > "$workdir/order.txt"
printf 'demo-makeline consumes queue orders-demo\n' > "$workdir/makeline.txt"

Start interactive codex -C "$workdir" --sandbox read-only. Ask for both queue names with citations. Ask a few follow-up questions, record the salient facts and context/usage indicators if shown by your version, then enter /compact in the interactive Codex prompt, not in your shell. Ask again which file supports each queue claim and compare answers to the files. Continue by updating a synthetic fact in a fresh workspace; verify it is reread instead of trusting a summary.

Record what was retained, dropped and re-fetched. If the model does not support manual compaction in your version, mark the lab’s manual step unavailable. Do not claim a specific token threshold, saving percentage or cache gain without provider measurements; compaction never removes the model’s hard context-window limit.

Follow reset and cleanup afterward.