Fix 'ok' short predictions: add explicit instruction to agent_input

[Turtle-Hwan]

Problem

Clone MCP kept returning single-word or bare-affirmation predictions —
"ok", "looks good", "go ahead" — instead of specific, actionable next
prompts. This caused the loop to escalate to humans even on iterations
where the work was clearly proceeding well.

Root cause

formatConversationHistory() (shared by both hooks) dumped the
conversation context and then just ended. Clone MCP had no guidance on
what kind of response to produce, so it defaulted to the shortest
plausible user reply — a terse approval.

The ask-user-question hook already appends:

"Predict the exact natural-language answer... Return only the answer text."
…and gets specific responses. The stop hook had no equivalent.

Fix

Append a 5-rule instruction block to formatConversationHistory():

• write as the user typing their next message
• reference concrete artifacts, file names, or test results
• minimum 1 full sentence — no single words, no "ok", no bare affirmations
• if work is complete, ask for the next logical action

Since formatConversationHistory() is shared, both hooks get the fix
in one change.

Evidence

Live e2e against https://api.clone.is/mcp (demo token):

	predicted_response
Before	"yeah go ahead"
After	"Go ahead and write the supertest integration tests for the validation error cases."

Reasoning also improved — now explicitly cites prior turns.

Test plan

• pnpm test → 23/23 passing, including new assertions that the
  instruction block appears in agent_input.
• node scripts/manual-e2e-multiturn.mjs → specific full-sentence
  prediction confirmed.

🤖 Generated with Claude Code