Ghostspend Fix

GhostSpend Fix Skill

Purpose

ghostspend-audit diagnoses; this skill guides remediation without ever taking a destructive or spend-affecting action on its own. Every proposed fix is shown to the user verbatim — the exact command or file diff — before it runs. This is a permanent boundary (see SECURITY.md and docs/gs-fix-dev-plan.md Section 6), not a v0.2 limitation to be relaxed later.

This phase (Phase 2 of the dev plan) supports only two options per finding: Safest and Skip. Balanced and Other (free-text) are Phase 3 and are not implemented here — do not improvise them.

Freshness Gate

Before acting on any finding, confirm it came from a recent ghostspend-audit run. If ~/.ghostspend/config.json's setup_completed_at is absent, or you have no evidence an audit ran this session, re-run ghostspend-audit rather than acting on findings you cannot confirm are current. Offer the user an override ("use the findings I already have") if they'd rather not re-audit.

Option Model — Finding-Type-Aware

Each finding type declares which options are actually meaningful for it. Render only those. Never present two options that would execute the identical action — that erodes trust in the prompt (see dev plan §9, resolved 2026-09-13). If a finding type isn't in the table below, it isn't supported yet in this phase — say so and stop (see "Everything Else" below).

Supported Finding Types This Phase

Findings arrive from scripts/ghostspend.sh encoded as severity|message (see skills/ghostspend-audit/SKILL.md's Severity Classification table). Match on the message content to identify the finding type below.

Everything Else In This Phase

For any finding type not in the table above (MCP pending-approval, project- level .mcp.json drift, or anything else ghostspend-audit produces), report it as "not yet supported by /gs-fix — manual follow-up needed" and move to the next finding. Do not invent a plausible-looking fix for an unsupported type; that is exactly the kind of unreviewed action this skill exists to prevent.

Hard Boundaries

These mirror SECURITY.md and apply without exception:

• Never disable, delete, uninstall, or reconfigure another vendor's CLI tool (Codex, Gemini, Copilot, OpenCode) directly. The only file this skill writes to is ~/.ghostspend/config.json.
• Never execute a command or write a file without first showing the user the exact command or diff and receiving an explicit yes.
• If a user's request (even a "Safest" or "Skip" follow-up question) implies an action outside the two tables above, decline and explain why, rather than attempting a plausible-sounding substitute.
• No network call (e.g. npm install) proceeds without the same explicit confirmation step already required by CLAUDE.md's non-negotiables.

Procedure

1. Confirm a recent audit exists (Freshness Gate above). If not, re-run ghostspend-audit or get the user's explicit override.
2. Sort findings critical → high → medium → low (already the order scripts/ghostspend.sh prints them in).
3. For each finding, in order:
   • If its type is in the Supported Finding Types table, present only Safest and Skip (never more, never less).
   • If its type is not in that table, report it as unsupported per "Everything Else" and move on — do not ask the user to choose between options that don't exist yet.
4. On Safest, show the exact command or diff, wait for an explicit yes, then execute and confirm the result (e.g. re-check claude mcp list).
5. On Skip, take no action and note it as skipped in the summary.
6. Close with a summary: what was fixed, what was skipped, what needs manual follow-up (unsupported types), grouped the same way the audit summary is.