Context Mode: shrink tool output without losing the clue

What the documentation establishes

The README documents a Claude Code plugin, output indexing/search tools and a ctx-doctor installation check. Source ↗

The walkthrough below is an original worked exercise. We have not independently tested this product’s performance.

1. Install and confirm the route

The documented Claude Code plugin route is shown below. Run the commands inside Claude Code, then restart the session or use its reload command. The doctor command checks the installation; a successful check does not demonstrate that a particular task used the expected tool.

Before any real task, inspect which tools are available and ask the assistant to report the tool names used for this fixture. If it reads the entire file through another tool, you have tested a different path. Record that deviation rather than reporting a context-saving result.

/plugin marketplace add mksglu/context-mode
/plugin install context-mode@context-mode
/reload-plugins
/context-mode:ctx-doctor

2. Create a small failure that a summary could hide

Save this original Python script as make_fixture.py in an empty directory and run python3 make_fixture.py. It writes only incident.txt in that directory. The file deliberately has normal activity, two different faults and a final line that sounds reassuring but does not prove either fault was repaired.

Do not give the answer key to the assistant. For the first question, the expected findings are a permission-denied invoice export at event 271 and a mail delivery timeout at event 843. A queue-empty message at event 1000 is insufficient to conclude either operation succeeded. Nothing in the fixture establishes a database outage.

from pathlib import Path

lines = []
for n in range(1, 1001):
    text = "heartbeat ok"
    if n == 271:
        text = "job=invoice-export result=failed reason=permission-denied retry=none"
    if n == 843:
        text = "job=delivery-email result=unknown reason=timeout receipt=missing"
    if n == 1000:
        text = "scheduler queue=empty"
    lines.append(f"event={n:04d} {text}")
Path("incident.txt").write_text("\n".join(lines) + "\n", encoding="utf-8")
print("Wrote incident.txt; fictional data only.")

3. Ask the same question through two paths

In a baseline session, provide the fixture directly. In a separate session with the plugin enabled, ask the assistant to index or search the fixture through the available Context Mode tools. Keep the model, effort and question the same where possible. Do not add the failure event numbers to the second question: that would make retrieval artificially easy.

If a search returns no matches, require the assistant to say what it searched and whether it inspected the surrounding records. One empty query is not evidence that no incident happened. The desired answer should distinguish a known failure, an unknown delivery outcome and an absence of evidence for a third incident.

Inspect incident.txt and prepare an incident handoff.
List each failed or uncertain job, its event ID, the exact evidence,
and what remains unknown. Explain whether the final scheduler line
proves the jobs succeeded. Is there evidence of a database outage?
Do not invent a retry, root cause, or completion receipt.
Report the tools and search terms you actually used.

4. Grade evidence before comparing size

Record these five checks separately for every answer. Keep incorrect answers in the comparison instead of retrying until one looks good. If your interface exposes token use, record the same metric for both paths; otherwise write not measured. Do not estimate a token-saving percentage from the number of visible words.

Repeat after swapping the positions of the two faults and changing their names. If the answer only works on the first fixture, you may have tested memorized positions or unusually convenient search terms. A small test cannot establish production reliability, but it can disprove an overconfident recommendation quickly.

5. Diagnose failures and decide whether to keep it

If the doctor reports a missing dependency, resolve that specific runtime or registration error using the current README before continuing. If the file indexes but the permission failure disappears, inspect the retrieval query and chunk boundaries. Ask for surrounding records, then record the extra work as a correction rather than silently replacing the first result.

If the assistant claims a repair happened, compare the cited event with the original file. That is an evidence error even if the prose sounds reasonable. If the indexed path is slower or requires more corrections on your normal task size, keep the direct-reading baseline for that task. The useful outcome is a routing rule: which tasks benefit, which need raw evidence, and when to escalate.

Keep the fixture, prompts, tool trace, answers and scorecard together with the installed plugin version and date. Re-run the same cases after an upgrade or a change in routing hooks. When working with actual logs, decide what may be sent to the selected model; a local index does not by itself mean every downstream model call is local.

Guide history

Revisions preserve the original instructions. A source recheck alone does not create a new revision.

Revision 1 · New documented setup plus original 1,000-line incident fixture, expected answers, paired comparison and troubleshooting. Tool runtime not executed.

Continue learning