feat: full Clone MCP session integration (start/record/feedback/stop)

README.md

@@ -6,7 +6,7 @@
   <br/><br/>
   <strong>Keep Agent working, with predicted next prompts.</strong>
   <br/>
-  <sub>A Claude Code plugin for self-driving coding loops powered by Clone MCP.</sub>
+  <sub>An agent plugin for self-driving coding loops powered by Clone MCP.</sub>
 </p>
 
 <p align="center">
@@ -35,9 +35,9 @@
   <a href="#pinning-a-version">Versions</a>
 </p>
 
-Clone is a Claude Code plugin that turns any task into a self-driving loop.
-When Claude tries to stop, Clone predicts your most likely next prompt and
-hands it back to Claude so the work continues — without you having to type
+Clone Loop turns any agent task into a self-driving loop for Claude Code and
+Codex. When the agent tries to stop, Clone predicts your most likely next prompt
+and hands it back so the work continues — without you having to type
 the same nudge ten times.
 
 ```text
@@ -53,9 +53,9 @@ Then walk away.
   you. Threshold-gated so it only auto-continues when it's confident.
 - **Rich context, not a one-line snapshot.** Each prediction sees your
   original task, every prior iteration's user turn, and a full timeline of
-  what Claude did this iteration — text, tool calls, and tool results.
-- **AskUserQuestion popups answered automatically** during an active loop.
-- **One session, one Claude.** No subprocesses, no daemons, no parallel
+  what the agent did this iteration — text, tool calls, and tool results.
+- **AskUserQuestion popups answered automatically in Claude Code** during an active loop.
+- **One session, one agent.** No subprocesses, no daemons, no parallel
   agents to herd.
 
 ## Quick start
@@ -144,22 +144,22 @@ generation: Clone does not generate the interview questions yet.
 
 ## How it works
 
-1. `/clone:loop` writes a state file and Claude starts the task.
-2. When Claude tries to stop, the Stop hook intercepts.
+1. `/clone:loop` or `clone-loop` writes a state file and the agent starts the task.
+2. When the agent tries to stop, the Stop hook intercepts.
 3. The hook asks Clone MCP `predict_next_prompt` for what you'd most likely
    say next.
-4. **Above threshold** → prediction injected, Claude continues.
+4. **Above threshold** → prediction injected, the agent continues.
    **Above threshold + satisfaction signal** → loop exits, asks for input.
    **Below threshold** → loop exits, asks for input.
-5. Mid-loop `AskUserQuestion` popups are auto-answered the same way.
+5. In Claude Code, mid-loop `AskUserQuestion` popups are auto-answered the same way.
 
 ### What Clone actually sees
 
 - **Your original task prompt** — always preserved verbatim.
 - **The conversation so far** — every prior iteration's injected user turn
   plus the assistant timeline that produced it (text + tool calls +
   summarized tool results).
-- **What Claude just did** — this iteration's full timeline.
+- **What the agent just did** — this iteration's full timeline.
 
 Sane caps: 20-turn rolling window on user history, oldest prior-iter
 timelines drop first when the combined size exceeds budget, and long tool
@@ -287,8 +287,8 @@ equivalent AskUserQuestion tool hook event.
 
 ## Requirements
 
-Claude Code with plugin support, Node.js on `PATH`. Windows: PowerShell
-or cmd, no Git Bash needed.
+Claude Code or Codex with plugin support, plus Node.js on `PATH`. Windows:
+PowerShell or cmd, no Git Bash needed.
 
 ## Development
 

commands/help.md

@@ -8,26 +8,27 @@ Please explain the following to the user:
 
 ## What is Clone Loop?
 
-Clone Loop is a Clone-backed automation loop for Claude Code. It runs inside
-the current session and uses Clone MCP to predict the user's likely next prompt
-whenever Claude tries to stop.
+Clone Loop is a Clone-backed automation loop for Claude Code and Codex. It runs
+inside the current agent session and uses Clone MCP to predict the user's likely
+next prompt whenever the agent tries to stop.
 
 **Core concept:**
 
-1. Claude receives the user's initial task.
-2. Claude works on the task, modifying files.
-3. Claude tries to stop.
+1. The agent receives the user's initial task.
+2. The agent works on the task, modifying files.
+3. The agent tries to stop.
 4. The Stop hook intercepts the stop.
 5. The Stop hook calls Clone MCP `predict_next_prompt` directly.
 6. If confidence clears the configured threshold, the hook passes the
-   prediction payload to Claude.
-7. Claude evaluates the prediction in context and continues as if the user had
-   provided the predicted prompt.
+   prediction payload to the agent.
+7. The agent evaluates the prediction in context and continues as if the user
+   had provided the predicted prompt.
 8. The loop continues until max iterations are reached, confidence fails, MCP
    fails, or the user runs `/clone:cancel-loop`.
 
-During an active loop, Clone also watches `AskUserQuestion` and answers with a
-predicted response so the user does not have to handle the popup manually.
+In Claude Code, Clone also watches `AskUserQuestion` during an active loop and
+answers with a predicted response so the user does not have to handle the popup
+manually. Codex v1 does not expose an equivalent question hook.
 
 ## Available Commands
 
@@ -97,7 +98,7 @@ Cancel an active Clone Loop.
 
 Clone Loop resolves API keys in this order:
 
-1. Nonblank `CLONE_API_TOKEN` from the Claude Code process environment.
+1. Nonblank `CLONE_API_TOKEN` from the agent process environment.
 2. Saved plugin config in `${CLAUDE_PLUGIN_DATA}/auth.local.json`.
 3. The public demo fallback token.
 
@@ -106,7 +107,7 @@ Full tokens are never shown in command output.
 ### Human Escalation
 
 If Clone MCP returns low confidence or fails, the hook removes
-`.claude/clone-loop.local.md`. Claude should explain that Clone was not
+`.claude/clone-loop.local.md`. The agent should explain that Clone was not
 confident enough and wait for human input.
 
 ## Learn More

hooks/ask-user-question-hook.mjs

@@ -1,8 +1,7 @@
 #!/usr/bin/env node
 
-import { existsSync, readFileSync } from 'node:fs'
+import { existsSync, readFileSync, writeFileSync } from 'node:fs'
 import { loopHistoryPath, loopStatePath } from '../scripts/clone-paths.mjs'
-import { resolveCloneToken } from '../scripts/clone-auth.mjs'
 import {
   HISTORY_WINDOW_TURNS,
   assistantTextsThisIteration,
@@ -16,7 +15,7 @@ import {
   appendLoopHistory,
   validateActiveLoopState,
 } from '../scripts/loop-state-guard.mjs'
-import { cloneMcpClientInfo, cloneMcpEndpoint, cloneMcpRpc } from '../scripts/clone-mcp-client.mjs'
+import { clonePredictNextPrompt, recordAgentPrompt, submitFeedback } from '../scripts/clone-mcp.mjs'
 import { numeric, parseJson, readStdin } from '../scripts/clone-utils.mjs'
 import { mapPredictionToOption, rankedPredictionCandidates } from '../scripts/interview-answer-utils.mjs'
 
@@ -27,39 +26,6 @@ function appendHistory(record) {
   appendLoopHistory(LOOP_HISTORY_FILE, record)
 }
 
-async function clonePredictNextPrompt({ agent, agentInput, threshold, sessionId }) {
-  const endpoint = cloneMcpEndpoint()
-  const { token } = resolveCloneToken()
-
-  const init = await cloneMcpRpc(endpoint, token, 'initialize', {
-    protocolVersion: '2024-11-05',
-    capabilities: {},
-    clientInfo: cloneMcpClientInfo(),
-  })
-
-  const args = {
-    agent,
-    agent_input: agentInput,
-    k: 1,
-    threshold: Number(threshold || '0.6'),
-  }
-  if (sessionId) args.session_id = sessionId
-
-  const prediction = await cloneMcpRpc(
-    endpoint,
-    token,
-    'tools/call',
-    { name: 'predict_next_prompt', arguments: args },
-    init.sessionId,
-  )
-
-  const content = prediction.payload?.result?.content?.[0]
-  if (!content || content.type !== 'text') {
-    throw new Error('Clone MCP returned no text prediction content.')
-  }
-  return JSON.parse(content.text)
-}
-
 function numericConfidence(value, fallback = 0) {
   return numeric(value, fallback)
 }
@@ -226,6 +192,57 @@ function deferQuestion({ decision, question, threshold, prediction, confidence,
   })
 }
 
+function updateFrontmatter(content, key, value) {
+  const escaped = String(value).replace(/\\/g, '\\\\').replace(/"/g, '\\"')
+  const replacement = `${key}: "${escaped}"`
+  const pattern = new RegExp(`^${key}: .*$`, 'm')
+  if (pattern.test(content)) {
+    return content.replace(pattern, replacement)
+  }
+  return content.replace(/^---\r?\n([\s\S]*?)\r?\n---/, (_, body) => `---\n${body}\n${replacement}\n---`)
+}
+
+async function safeFeedback({ predictionId, status, mcpSessionId }) {
+  if (!predictionId) return
+  try {
+    await submitFeedback({ predictionId, status, mcpSessionId })
+    appendHistory({ event: 'feedback-sent', source: 'ask-user-question', prediction_id: predictionId, status })
+  } catch (error) {
+    appendHistory({ event: 'feedback-sent', source: 'ask-user-question', prediction_id: predictionId, status, error: error?.message || String(error) })
+  }
+}
+
+async function safeReject({ predictionId, mcpSessionId }) {
+  await safeFeedback({ predictionId, status: 'rejected', mcpSessionId })
+}
+
+async function safeRecordAnswer({ cloneSessionId, mcpSessionId, agent, question, answer, iteration }) {
+  if (!cloneSessionId || !answer) return null
+  try {
+    const result = await recordAgentPrompt({
+      cloneSessionId,
+      mcpSessionId,
+      agent,
+      prompt: `Q: ${question}\nA: ${answer}`,
+      source: 'auto-answer',
+      sourceDetail: `clone-loop:ask-user-question:iteration-${iteration || 'unknown'}`,
+    })
+    appendHistory({
+      event: 'record-prompt',
+      source: 'ask-user-question',
+      event_id: result?.eventId || null,
+    })
+    return result
+  } catch (error) {
+    appendHistory({
+      event: 'record-prompt',
+      source: 'ask-user-question',
+      error: error?.message || String(error),
+    })
+    return null
+  }
+}
+
 async function main() {
   const hookInput = parseJson(await readStdin())
   if (hookInput.tool_name && hookInput.tool_name !== 'AskUserQuestion') return
@@ -247,6 +264,7 @@ async function main() {
     clone_agent: cloneAgentRaw,
     clone_session_id: cloneSessionId,
     mcp_session_id: mcpSessionIdInitial,
+    iteration,
   } = state.frontmatter
 
   const toolInput = hookInput.tool_input || {}
@@ -256,7 +274,9 @@ async function main() {
   const cloneThreshold = cloneThresholdRaw || '0.6'
   const cloneAgent = cloneAgentRaw || 'Claude Code Clone Loop'
   const answers = {}
+  const acceptedPredictions = []
   const confidenceValues = []
+  let activeMcpSessionId = mcpSessionIdInitial || ''
 
   for (const [questionIndex, question] of questions.entries()) {
     if (question?.multiSelect) {
@@ -286,7 +306,7 @@ async function main() {
         }),
         threshold: cloneThreshold,
         sessionId: cloneSessionId || undefined,
-        mcpSessionId: mcpSessionIdInitial,
+        mcpSessionId: activeMcpSessionId,
       })
     } catch (error) {
       deferQuestion({
@@ -306,9 +326,16 @@ async function main() {
         prediction,
         confidence: prediction.confidence,
       })
+      await safeReject({ predictionId: prediction.id, mcpSessionId: activeMcpSessionId })
       return
     }
 
+    if (prediction.mcp_session_id && prediction.mcp_session_id !== activeMcpSessionId) {
+      activeMcpSessionId = prediction.mcp_session_id
+      const updated = updateFrontmatter(readFileSync(LOOP_STATE_FILE, 'utf8'), 'mcp_session_id', activeMcpSessionId)
+      writeFileSync(LOOP_STATE_FILE, updated)
+    }
+
     const predictedResponse = String(prediction?.predicted_response || '').trim()
     const selection = predictedResponse
       ? {
@@ -327,6 +354,7 @@ async function main() {
         prediction,
         confidence: prediction?.confidence,
       })
+      await safeReject({ predictionId: prediction?.id, mcpSessionId: activeMcpSessionId })
       return
     }
 
@@ -338,10 +366,16 @@ async function main() {
         prediction,
         confidence: selection.confidence,
       })
+      await safeReject({ predictionId: prediction.id, mcpSessionId: activeMcpSessionId })
       return
     }
 
     answers[questionText] = selection.answer
+    acceptedPredictions.push({
+      predictionId: prediction.id || '',
+      question: questionText,
+      answer: selection.answer,
+    })
     confidenceValues.push(selection.confidence)
   }
 
@@ -353,6 +387,29 @@ async function main() {
     threshold: Number(cloneThreshold),
     answers,
   })
+  for (const accepted of acceptedPredictions) {
+    await safeFeedback({
+      predictionId: accepted.predictionId,
+      status: 'accepted',
+      mcpSessionId: activeMcpSessionId,
+    })
+    const recorded = await safeRecordAnswer({
+      cloneSessionId,
+      mcpSessionId: activeMcpSessionId,
+      agent: cloneAgent,
+      question: accepted.question,
+      answer: accepted.answer,
+      iteration,
+    })
+    if (recorded?.eventId) {
+      const updated = updateFrontmatter(
+        readFileSync(LOOP_STATE_FILE, 'utf8'),
+        'last_prompt_event_id',
+        recorded.eventId,
+      )
+      writeFileSync(LOOP_STATE_FILE, updated)
+    }
+  }
   allowAnswer({ toolInput, answers, confidence, threshold: cloneThreshold })
 }