PIMP MY IDEBack to dispatches
October 10, 2026 / context control / tool output

Shrink the stream. Keep the source.

A smaller model context can make an agent faster and cheaper. It can also hide the exact output you need to debug the next wrong turn.

A context-savings meter needs three labels: what entered the model, where the raw result lives, and how to retrieve or delete it.

Tool output is part of the context bill

Anthropic's context documentation says tool results, messages, system instructions, images, documents, and tool definitions all count toward the context window. It also says prompt caching changes the price of reused tokens, not the amount of context they occupy.[2]

That makes raw logs, browser snapshots, issue lists, and repository scans obvious targets for reduction. The useful move is to compute against the large result and send the model the small answer. The dangerous move is to discard the large result before anyone defines its custody.

Compression answers how much the model sees. Custody answers whether anyone can check it.

A trending project makes the split concrete

Context Mode is an open repository and npm package that routes tool work through an MCP server. Its README says a 315 kilobyte input became 5.4 kilobytes in one example. The project stores session events in SQLite, indexes content with FTS5, retrieves selected passages, and exposes a purge command.[1]

Those are project claims, not results we reproduced. We inspected release 1.0.169 at commit 8de5f73. The package declares Node.js 22.5 or newer, an Elastic License 2.0 license, a post-install script, local databases, hooks, and eleven MCP tools. We did not install or run it because setup changes agent clients, hooks, and local history paths.

Compaction and retrieval solve different failures

OpenAI's compaction guide says its server can emit an encrypted compaction item after a configured token threshold. The item carries prior state into the next window. The guide also says it is opaque and not meant for human interpretation.[3]

Anthropic documents server-side compaction, old tool-result clearing, and thinking-block clearing as separate controls.[2] A compacted conversation can continue. That does not mean a reviewer can recover one exact command result, compare it with the current repository, or prove that retained data was deleted.

The protocol already has room for an address

The MCP tool specification allows text, structured content, embedded resources, and resource links in tool results. It also allows output schemas. Clients should validate structured results when a schema exists.[4]

A compact result can therefore carry a count, status, source address, digest, retention rule, and retrieval hint. The raw payload does not need to flood every turn. It does need a named location and a deletion route that the operator can inspect.

Use a two-record contract

Keep the model record small. Include the answer, query, source address, timestamp, and digest. Keep the custody record outside the model context. Include the raw artifact, access scope, retention deadline, deletion action, and retrieval log.

The inspection rack below is a teaching fixture. Its byte reduction is illustrative. It does not measure Context Mode or predict savings for another agent. Use it to draft the contract before routing real output through a filter.

Interactive teaching fixture

Context output sump

Traditional purpose: trim tool output before the next model turn. This version combines a reduction preview, custody checks, retrieval scope, and a copyable contract on one panel.

CaptureFilterAddressRetrieve

Routing controls

The byte values are an invented fixture. Select a route and the contract fields you want a real system to supply.

180 KB
20 KB320 KB
Context route
Contract sections to request
CONTRACT INCOMPLETE0 of 4 sections selected

Flow preview

The fixture sends a small answer into context and keeps an address for exact retrieval. Real storage and search behavior remain unverified.

Raw fixture180 KB
Model context7.2 KB
RETRIEVAL ROUTE SELECTEDThe fixture includes a retrieval address. Real evidence still needs a tested location, digest, query, scope, and purge result.
Sources read / October 10, 2026 UTC

Source log

Open the evidence boundary
  1. Context Mode repository at commit 8de5f73, committed and read October 10, 2026. We inspected the README, package manifest, source layout, tests, declared version 1.0.169, license, engine requirement, hooks, tool list, storage paths, retrieval description, purge route, and install behavior. The 98 percent figure and 315 kilobyte to 5.4 kilobyte example are project claims. We did not reproduce them.
  2. Anthropic, "Context windows", read October 10, 2026. This supports the context accounting for tool results and other request content. It also supports the distinction among compaction, tool-result clearing, thinking-block clearing, and prompt caching.
  3. OpenAI, "Compaction", read October 10, 2026. This supports the configured threshold, encrypted compaction item, continuation behavior, and the statement that the item is opaque rather than human-readable.
  4. Model Context Protocol specification, "Tools," version 2025-06-18, read October 10, 2026. This supports tool-result content types, resource links, embedded resources, structured content, output schemas, and client validation guidance.
  5. GitHub Trending, read October 10, 2026. It surfaced Context Mode during this editorial pass. Trending placement is a discovery signal, not proof of quality, safety, or adoption.

Evidence boundary: Pimp My IDE inspected the repository at the named revision and confirmed the npm registry listed version 1.0.169. We did not install the package, modify an agent client, run its hooks, open its hosted dashboard, or benchmark its storage and retrieval behavior. The interactive numbers are teaching fixtures.