pre-flight

# pre-flight ## When this skill is the preferred path The user (or the agent itself) is about to send a prompt that might be expensive, redundant, or context-window-exhausting. Instead of sending blindly and finding out only when the bill arrives or the conversation hits 200k tokens, the agent calls `gc_pre_flight` first to get a structured answer. This skill is the gotcontext SKU. **One MCP call replaces context warnings, compression decisions, and cost previews.** ## How to use it Call `gc_pre_flight` with the proposed prompt body. Optionally pass: - `hint` — what the agent is trying to do ("refactor", "debug", "plan"). Guides the compression strategy. - `target_model` — the LLM that will receive this prompt ("gpt-4o", "claude-opus-4-7", etc.). Drives cost preview accuracy. - `current_context_used_pct` — agent's estimate of its own context fill (0-100). Drives the `warn_context_limit` and `clear_first` verdicts. ## What you get back A structured response (JSON) with these fields: - `verdict`: one of `send_as_is`, `send_compressed`, `warn_context_limit`, `clear_first` - `compressed_prompt`: the actual compressed body (when verdict is `send_compressed`); null otherwise - `tokens_in_original` / `tokens_in_compressed`: token counts - `compression_ratio`: 0.0-1.0 - `estimated_cost_original_usd` / `estimated_cost_compressed_usd` / `estimated_savings_usd` - `cache_hit_likelihood`: 0.0-1.0 — probability that the prompt's prefix is in the gotcontext semantic cache - `context_warning`: null OR an object describing the projected context fill if the agent sends as-is - `recommendation`: human-readable action ## How to act on the verdict | Verdict | Action | |---|---| | `send_as_is` | Send the original prompt. Compression wouldn't beat 30% savings. | | `send_compressed` | Use the `compressed_prompt` field as the prompt body. Costs less, fits more. | | `warn_context_limit` | The original prompt would push context fill to 90%+. Send compressed body OR run "gc_session_summary" first. | | `clear_first` | Context is critical. Run "gc_session_summary", then `/clear`, then re-inject the summary, then send. | ## Why this skill exists - Agents have no proactive context-window warnings — they hit the limit mid-task with no heads-up. `gc_pre_flight`'s `warn_context_limit` and `clear_first` verdicts close that gap. - Cursor and Claude Code users routinely complain about 4M-token bursts on trivial tasks. `gc_pre_flight`'s cost preview shows the bill BEFORE you commit to it. - LLMLingua and similar compressors require a separate compression call, then a separate cost-estimation call, then the agent needs context awareness from somewhere else. `gc_pre_flight` is one call that does all three. ## Plan availability Available on every plan including Free. Volume is governed by your existing per-month compression quota — not by allowlist exclusion. This is the conversion driver: Free users see real cost previews and real verdicts; Pro+ users get unlimited calls. ## Before writing code — use the agent-prep tools (Pro+) `gc_pre_flight` answers "should I compress?" before sending a prompt. For the paired question — "what should I actually change?" — two complementary tools are available on Pro+ plans: - `gc_agent_capsule`: returns an actionable context capsule (primary targets, snippets, validation commands, rollback metadata, confidence score) before any non-trivial code change. Use this as the first call whenever an agent is about to touch multiple files or a non-obvious symbol. - `gc_edit_plan`: returns a machine-readable edit plan (which files to modify, what to add/remove, validation_commands to verify the result). Use this instead of free-form "let me think about this" planning loops — the plan comes back structured and executable. Typical pre-code agent sequence: 1. `gc_pre_flight` — is the prompt cost-efficient? 2. `gc_agent_capsule` — what do I need to look at? 3. `gc_edit_plan` — exactly what do I change? 4. Make the edits, run the validation_commands from the plan output.