Multiple open PRs had unresolved review comments blocking merge. This consolidates the two substantive PRs (#26: runtime bug fixes, #6: reply targeting) with all reviewer feedback applied.

PR #26 — Runtime bug fixes + review feedback

• stdout protection: Redirect console.log → stderr to avoid corrupting MCP StdioServerTransport protocol messages. Uses unknown[] per reviewer.
• since_id pagination: Wire up lastMentionId with URLSearchParams (reviewer suggestion) to avoid refetching processed mentions each poll cycle.
• fetchThread hardening: Guard with Array.isArray(response.data) before calling parseThread — prevents crash on unexpected API shapes.
• Sort immutability: [...tweets].sort() instead of in-place mutation. parseThread signature narrowed from any → unknown.
• Bounded memory: Cap processedMentions Set at 10k, prune with safe { value, done } iterator (not type assertion). Process mentions oldest-first so Set insertion order is chronological and pruning evicts correctly.

PR #6 — Reply targeting fix + review feedback

• Add mentionPostId param to analyzeAndDecide() — replies now target the specific mention post instead of always targeting thread.root_post.id.
• Replace any with typed interface for Grok API response parsing.

// Before: always replied to conversation root
const analysis = await grok.analyzeAndDecide(mention.post.text, thread);
// analysis.action.target_post_id === thread.root_post.id ❌

// After: replies to the actual mention
const analysis = await grok.analyzeAndDecide(mention.post.text, thread, mention.post.id);
// analysis.action.target_post_id === mention.post.id ✅

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.