Context Cockpit

context-cockpit — inspect the window, then reclaim it

One pass to answer "why is this session heavy, and what do I run next." A thin orchestrator over Claude Code's native context tools — it reads /context and hands you the exact next move; it does not replace it. Absorbs MCP triage (the former mcp-doctor). Which substrate to reach for belongs to connector-routing; this triages what a session already loaded.

Step 1 — Read the window (/context)

Run /context. It lists every loaded item — system prompt, MCP tool lists, files read, message history — with token counts and % of window. Rank the top 3 token-eaters and mark each keep / trim / drop against what the task at hand (the argument) actually needs.

• Fuel-gauge thresholds: <50%** build freely · **50–70%** getting heavy, wrap or compact · **>85% danger — quality already slipped, don't start real work here.
• Completion criterion: top-3 eaters named, each with a keep/trim/drop verdict.

Step 2 — Trim MCP tools under 50

Every connected MCP loads its FULL tool list at startup, used or not — three servers can eat ~66% of the window before any work begins. Target enabled tool count < 50 and 3–5 servers max (past that, tool-selection accuracy drops, not just tokens).

claude mcp list        # what's connected and taxing you

Also run /mcp for the in-session per-server token cost. Disconnect anything not used THIS session (rule: "every session? no → disconnect until you do"). A heavy-but-rare MCP → wrap it in a Skill Claude reads on demand instead of loading always.

• Completion criterion: enabled tool count < 50, or the over-budget servers named for disconnect.

Step 3 — Clear vs compact

• /clear — full reset. Between UNRELATED tasks (new feature/bug/project). One task = one chat.
• /compact — squeeze, keep the thread. Mid-task when the SAME job's chat gets long.
• Rule: new task → /clear; long task still going → /compact at ~60% (early enough to compress cleanly; past 85% there's nothing healthy left). Never grind a 90%-full window.
• Goal-scoped compact: "Compact to: working code · 3 decisions that mattered · 1 gotcha · next 2 steps. Drop false starts and pasted files."
• Completion criterion: the exact /clear or /compact to run is chosen.

Step 4 — CLI over MCP

A CLI loads NOTHING up front — Claude runs it on demand and only the result returns. "Anytime you can move from an MCP to a CLI, do it." Before adding an MCP, ask: is there a CLI? (A Playwright CLI did an MCP's job for ~90k fewer tokens.) Install the CLI and its matching Skill together so Claude knows how to drive it.

• Completion criterion: each heavy MCP has a CLI-substitution verdict.

Step 5 — Model ladder

Cheapest capable rung wins. Haiku — quick edits, renames, summaries. Sonnet — default workhorse, real building. Opus — hard bugs, architecture, deep planning. Rule: start one rung LOWER than you think; climb only if it stalls. /model switches; /model opus-plan plans on Opus then auto-drops to Sonnet for execution (worth it for multi-file refactors, not one-line fixes). Put folder · model · context % in /statusline so you can see the rung.

• Completion criterion: the session is on the lowest rung that fits the task.

Red-MCP triage (when a server shows red)

Run /mcp, read the error line (right ~80% of the time), then in order, stop when green:

1. Key — missing/wrong API key (#1 cause). Re-paste the full value, no stray spaces/quotes.
2. Installed — server package absent ("command not found") → install per its docs.
3. Restart — config is read at startup; fully quit and reopen Claude Code.
4. Typo — one wrong char / missing comma in the config entry breaks it; validate the file.

• Completion criterion: the server is green, or the failing check is named for the operator.

Related

connector-routing picks WHICH substrate (MCP / CLI / Composio); this skill triages what a session already loaded. Model tiers follow the Anthropic-Max ladder.