← Back to labs index

Lab 10 — Context and compaction

Every question resends the whole session, so a long investigation fills the context window (Lab 7’s /context showed how). This standalone snapshot lets you act on that in harness ask: compact the history into a model-written summary, optionally steered by your instructions, and check which facts survive.

The exercise matches the Claude Code Lab 10. make_log.py writes a 6,000-line synthetic app.log with a few important facts buried in noise.

What changes

/compact [instructions]:
- sends the history to the model with a request to summarize it (your instructions are appended), then replaces the history with that summary. The CLI prints the summary and the token estimates before and after.
- the session file keeps the full record: a reset entry marks where the compacted history starts, so -c/-r resume from the summary.
/clear empties the history; files, todos and session approvals stay.
harness ask --compact-at TOKENS compacts automatically before a question once the previous request reached that many input tokens. It doesn’t compact in the middle of a run.
/context adds the current history size, so you can see the drop right after compacting.
read_file takes optional start_line and max_lines (up to 1000 lines per call), so the model can read files over the 100 KB whole-file limit in parts. A whole-file read cut at 20,000 characters now says so.
With --trace, each compaction is a compact event.

The summary prompt and history rewrite are in harness/compaction.py, next to the earlier in-process models of a compaction policy, summary handoff and repository map.

Learner steps

Start from: labs/app/ at tag lab09-done (see the track guide).

Install this lab and configure Foundry as before:

python -m venv .venv
. .venv/bin/activate          # Windows: .\.venv\Scripts\Activate.ps1
pip install -e '.[dev]'
cp .env.example .env
az login
cd ../app

Generate the incident log:
```
python ../lab10-compaction/make_log.py
```
app.log is ignored by the app’s *.log rule. Don’t open it yet.

Investigate:

harness ask -n incident

Last night checkout failures spiked. Investigate app.log: find the root cause, the deploy involved, the affected order IDs and region, and when it was fixed. Read the whole log and show your evidence.

Then type /context.

Observe: the read_file calls with start_line/max_lines, and how much of the request the log takes (tool results).

Compact with instructions:
```
/compact Keep: root cause, config key and values, deploy id, affected order IDs and region, rollback time, and the fix plan. Drop raw log lines.
```
Read the printed summary, then type /context again. Ask, without letting it re-read the log:
```
Without reading any files, answer: what config changed, from what to what, in which deploy, which orders failed, in which region, and when was it rolled back?
```
Check against the answer key: payment.timeout_ms changed from 5000 to 500 in deploy 7f3a2c; orders ORD-2047, ORD-2113, ORD-2190 timed out in eu-west; then a rollback followed. The “out of stock sku=P3” lines are noise, not the cause.
Compare with an unguided compact. /exit, start a new session (harness ask -n incident-plain), repeat step 3, then use plain /compact and ask the same question. Optionally try /clear to see what losing everything looks like.

To see automatic compaction, start one more session with a threshold below the size /context reported, for example harness ask -n incident-auto --compact-at 30000, investigate, then ask the follow-up question.

Record

	Facts kept (of 6)	`/context` before	after
`/compact` with instructions
plain `/compact`

Note any hallucinated “facts” after compaction.

Checkpoint. Nothing to commit (the log is ignored). In app/:
```
git tag lab10-done
```
Back in the lab folder, run the offline checks and inspect the declared capabilities:
```
pytest checks/
harness lab-info
```

harness ask keeps subagents and background agents (Lab 9), skills and MCP (Lab 8), tracing (Lab 7), permissions (Lab 6), file memory (Lab 5), plan mode and todos (Lab 4), sessions (Lab 3) and the Lab 2B tools, built-in policy, project hooks and rules. This is a teaching harness, not a sandbox: only use it on the practice app.