Slipstream Turns Claude Code Context Management Into a First-Class Runtime Surface

Claude Code’s next productivity fight is not about whether the model can write a better function. It is about whether the runtime can keep the model from wasting half the session reading the wrong 40,000 tokens first.

That is the useful lens for Slipstream, a new Claude Code plugin created on May 31 and pushed through the evening with version 0.2.0. The repository is tiny — one star, no forks, no issues at research time — so this is not a popularity story. It is a pattern story. Slipstream bundles scoped code retrieval, an MCP server, PreCompact memory capture, a local dashboard, statusline budget reporting, output styles, skills, subagents, and lifecycle hooks into one package. That is a lot of moving parts for a small plugin, but the direction is right: context management is becoming a first-class runtime surface.

The README’s token math is the part to read first. Slipstream compares reading all of src/map/retrieve.ts at 4,841 bytes, roughly 1,345 tokens, with calling sp_symbol(retrieve.ts, retrieveSymbol) at 1,381 bytes, roughly 384 tokens — a claimed 71% reduction. The larger example is more pointed: reading the whole src/ tree costs about 146,150 bytes, or 40,597 tokens, while sp_map returns 7,821 bytes, or 2,173 tokens. That is 5.4% of the “just read everything” path.

Those numbers are not a benchmark of agent quality. They are a reminder that many coding-agent failures are self-inflicted. The model is asked to reason after the harness has already filled its working memory with irrelevant implementation detail. Better inference cannot fully compensate for bad retrieval discipline.

The agent that reads less often wins first

Claude Code users already know the failure mode. The agent opens a giant file, spends context on code paths that do not matter, loses the earlier plan, then compacts the transcript into a summary that preserves the vibes while dropping the engineering state. The result is familiar: repeated questions, stale assumptions, and edits that make sense locally but miss the larger invariant.

Slipstream attacks that problem with MCP tools such as sp_map, sp_symbol, sp_lines, sp_search, sp_budget, and sp_mindmap. The shape is boring in the best way. Instead of telling the agent “be careful with context,” give it tools that make careful context use the path of least resistance. A project map is cheaper than a directory dump. A symbol read is cheaper than a file read. A budget tool is better than discovering the context cliff after the session starts hallucinating continuity.

This is especially relevant because Claude Code plugins can now package skills, commands, agents, hooks, statusline commands, output styles, and MCP servers. Slipstream’s manifest uses that whole surface. It runs an MCP server named slipstream via node ${CLAUDE_PLUGIN_ROOT}/dist/mcp/index.js, ships three subagents — sp-shipper, sp-schema, and sp-reviewer — and wires into lifecycle hooks including SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, SubagentStop, Stop, and PreCompact.

That last hook is the most interesting.

PreCompact is where sessions either become durable or become folklore

Compaction is one of the under-discussed boundaries in coding-agent work. A transcript summary is not the same thing as an operational checkpoint. Summaries are optimized for continued conversation; checkpoints are optimized for recovery. A good checkpoint captures the current task, decisions made, files touched, constraints discovered, unresolved risks, and next safe step. If those facts are not written somewhere durable before compaction, the agent may keep talking confidently while the actual engineering state has been sanded down.

Slipstream uses Claude Code’s PreCompact lifecycle surface to capture memory before the transcript is compressed. It also ranks smart memory recall from the git branch, changed files, and the last prompt, then loads only a relevant subset under an approximate 1,200-token ceiling. That is a useful constraint. Durable memory should not mean “dump the junk drawer into every session.” It should mean selective recall tied to the current task.

The dashboard follows the same runtime-first theme. Slipstream appends JSONL events under .claude/slipstream/dashboard/<session>.jsonl, then serves a local dashboard from 127.0.0.1 using server-sent events. The UI shows agents and subagents, discussion/activity, token budget, plan, and a Mermaid mind map. The README says the system is local-only, has no telemetry, and redacts obvious secrets before logging.

That architecture is appropriately plain. Append-only local logs are easy to inspect, replay, and delete. Server-sent events are enough. Not every developer tool needs a cloud dashboard with an enterprise sales funnel attached. But “local-only” is not a magic privacy spell. Event logs can still contain file paths, prompt text, command arguments, tool names, branch names, and snippets that should not live forever. Teams adopting this pattern should decide where logs live, how long they survive, what gets redacted, and whether exports can leave a machine.

Plugins are convenience, but also authority

The supply-chain caveat matters because Slipstream is not merely a prompt file. It is a plugin containing hooks, MCP tools, skills, subagents, output styles, and a statusline integration. Claude Code’s plugin marketplace documentation explicitly supports this kind of bundle. That makes distribution convenient, and it also puts the plugin in the agent’s trust path.

Review it like you would review CI configuration, a GitHub Action, or a repo-local script that developers are expected to run. Inspect the manifest. Read the hooks. Check what the MCP server can access. Verify what gets written under .claude/. Pin versions where possible. Run it first on a disposable repo. A tool designed to save context can still shape tool use, persist sensitive state, and change how the agent interprets a project.

The practical takeaway is not “install Slipstream immediately.” The repo is young, the audience is tiny, and teams should be cautious with any new plugin that touches hooks and MCP. The takeaway is that its design points at the right checklist for serious Claude Code use: scoped retrieval, visible context budgets, structured state before compaction, selective memory, observable subagents, and local logs that can support recovery.

The broader market lesson is sharper. Better models help, but the agent that reads 5% of the codebase instead of 100% often wins before the model race starts. Context discipline is not prompt etiquette. It is runtime engineering. Slipstream is early, but the direction is LGTM: make the harness responsible for the boring constraints humans forget when they are trying to ship.

Sources: Slipstream README, Claude Code hooks reference, Claude Code MCP docs, Claude Code plugin marketplace docs