analyze-chat-export（分析聊天导出）

Analyze Chat Export

Purpose

Extract structured metrics from VS Code Copilot chat exports to support retrospective analysis. Provides data on model usage, tool invocations, manual approvals, rejection patterns, file edit statistics, and session timing.

Hard Rules

Must

• Use the extract-metrics.sh script for analysis (consolidates all queries).
• Redact sensitive information before committing chat logs.
• Save analysis results alongside the chat export in the feature folder.

Must Not

• Commit unredacted chat logs containing passwords, tokens, API keys, secrets, or PII.
• Load the entire JSON file into memory (use streaming jq queries).

Pre-requisites

• jq command-line JSON processor installed.
• Chat export files (.json) saved via workbench.action.chat.export command. Multiple files may exist per feature, one per agent chat session (e.g., developer.chat.json, architect.chat.json).

Known Limitations

Custom agent names are NOT recorded in the export.

The chat export only contains the VS Code infrastructure agent (github.copilot.editsAgent), not the custom agent definition file (e.g., developer.agent.md, @Developer).

Impact:

• Cannot analyze metrics per custom agent
• Cannot determine which agent definitions performed best
• Cross-feature analysis loses agent context

Note: A single feature chat typically includes work from multiple agents, so per-agent analysis would require VS Code to record this information in the export format.

Quick Start

Recommended: Use the extraction script

# Generate analysis files (both markdown and JSON)
.github/skills/analyze-chat-export/extract-metrics.sh docs/features/NNN-<feature-slug>/chat.json docs/features/NNN-<feature-slug>/chat-metrics

This creates:

• chat-metrics.md - Human-readable report for review
• chat-metrics.json - Raw data for cross-feature analysis (commit this file)

Export Structure Reference

See these reference documents:

• Chat Export Structure - Empirical analysis of exported data
• Chat Export Format Specification - VS Code source-based type definitions

Quick Reference: Top-Level Keys

{
  "initialLocation": "panel",
  "requests": [...],
  "responderAvatarIconUri": { "id": "copilot" },
  "responderUsername": "Copilot"
}

Quick Reference: Request Fields

Quick Reference: Confirmation Types (isConfirmed.type)

Quick Reference: Response State (modelState.value)

Actions

1. Export Chat (Prerequisite)

Ask the Maintainer to export each relevant agent session chat:

1. Focus the chat panel for the agent session to export.
2. Run command: workbench.action.chat.export
3. Save to: docs/features/NNN-<feature-slug>/<agent-name>.chat.json (e.g., developer.chat.json, architect.chat.json)

2. Run Extraction Script (Recommended)

# Generate analysis report (creates both .md and .json files)
.github/skills/analyze-chat-export/extract-metrics.sh docs/features/NNN-<feature-slug>/chat.json docs/features/NNN-<feature-slug>/chat-metrics

This creates two files:

• chat-metrics.md - Human-readable markdown report
• chat-metrics.json - Raw metrics data for cross-feature analysis (commit this file)

The script outputs a markdown report with:

• Session overview (duration, requests, time breakdown)
• Model usage statistics
• Tool usage breakdown (top 15)
• Automation effectiveness (auto vs manual approvals)
• Model success rates
• Response times by model
• Error summary
• User feedback votes

3. Individual jq Queries (Advanced)

For custom analysis or debugging, use individual jq queries.

Session Metrics

CHAT_FILE="docs/features/NNN-<feature-slug>/chat.json"

# Total requests/turns
jq '.requests | length' "$CHAT_FILE"

# Session duration in minutes
jq '((.requests | last.timestamp) - (.requests | first.timestamp)) / 1000 / 60 | floor' "$CHAT_FILE"

# First and last timestamps (for start/end times)
jq '.requests | first.timestamp, last.timestamp' "$CHAT_FILE"

# Time breakdown (all in seconds)
jq '
{
  session_duration_sec: (((.requests | last.timestamp) - (.requests | first.timestamp)) / 1000 | floor),
  user_wait_time_sec: (([.requests[].timeSpentWaiting // 0] | add) / 1000 | floor),
  agent_work_time_sec: (([.requests[].result.timings.totalElapsed // 0] | add) / 1000 | floor)
}
| . + {
  user_wait_pct: (if .session_duration_sec > 0 then (.user_wait_time_sec / .session_duration_sec * 100 | floor) else 0 end),
  agent_work_pct: (if .session_duration_sec > 0 then (.agent_work_time_sec / .session_duration_sec * 100 | floor) else 0 end)
}
' "$CHAT_FILE"

# Format time breakdown as human-readable
jq '
  def format_time(s): "\(s / 3600 | floor)h \((s % 3600) / 60 | floor)m";
  {
    session: ((.requests | last.timestamp) - (.requests | first.timestamp)) / 1000,
    user_wait: ([.requests[].timeSpentWaiting // 0] | add) / 1000,
    agent_work: ([.requests[].result.timings.totalElapsed // 0] | add) / 1000
  }
  | {
    session_duration: format_time(.session),
    user_wait_time: format_time(.user_wait),
    agent_work_time: format_time(.agent_work)
  }
' "$CHAT_FILE"

3. Extract Model Usage

# Models used with counts
jq '[.requests[].modelId] | group_by(.) | map({model: .[0], count: length}) | sort_by(-.count)' "$CHAT_FILE"

4. Extract Tool Usage

# Total tool invocations
jq '[.requests[].response[] | select(.kind == "toolInvocationSerialized")] | length' "$CHAT_FILE"

# Tool usage breakdown
jq '[.requests[].response[] | select(.kind == "toolInvocationSerialized") | .toolId] | group_by(.) | map({tool: .[0], count: length}) | sort_by(-.count)' "$CHAT_FILE"

5. Extract Approval Patterns

# Approval type distribution
jq '[.requests[].response[] | select(.kind == "toolInvocationSerialized") | .isConfirmed.type // "unknown"] | group_by(.) | map({type: .[0], count: length})' "$CHAT_FILE"

# Count manual approvals (type 0 = pending/cancelled, type 4 = manual)
jq '[.requests[].response[] | select(.kind == "toolInvocationSerialized") | select(.isConfirmed.type == 0 or .isConfirmed.type == 4)] | length' "$CHAT_FILE"

6. Calculate Premium Request Estimate

# Model multipliers (update as needed based on docs/ai-model-reference.md)
jq '
  def multiplier:
    if . == "copilot/gpt-5.1-codex-max" then 50
    elif . == "copilot/claude-opus-4.5" then 50
    elif . == "copilot/gpt-5.2" then 10
    elif . == "copilot/gemini-3-pro-preview" then 1
    elif . == "copilot/claude-sonnet-4.5" then 1
    elif . == "copilot/gemini-3-flash-preview" then 0.33
    elif . == "copilot/gpt-5-mini" then 0.25
    elif . == "copilot/claude-haiku-4.5" then 0.05
    else 1
    end;
  [.requests[].modelId | multiplier] | add
' "$CHAT_FILE"

7. Redact Sensitive Data

# Create redacted copy
jq '
  .requests |= map(
    .message.text |= (
      gsub("(?i)(password|token|secret|key|bearer)[=: ]+[^\\s\"]+"; "[REDACTED]") |
      gsub("[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\\.[a-zA-Z]{2,}"; "[EMAIL_REDACTED]")
    )
  )
' "$CHAT_FILE" > "${CHAT_FILE%.json}-redacted.json"

8. Extract Response Timings

# Average response time (totalElapsed) in seconds
jq '[.requests[].result.timings.totalElapsed // 0] | add / length / 1000' "$CHAT_FILE"

# Average time to first progress in milliseconds
jq '[.requests[].result.timings.firstProgress // 0] | add / length' "$CHAT_FILE"

# Response state distribution (1=Complete, 2=Cancelled, 3=Failed)
jq '[.requests[].modelState.value] | group_by(.) | map({state: .[0], count: length})' "$CHAT_FILE"

9. Extract User Feedback

# Vote distribution (0=down, 1=up)
jq '[.requests[] | select(.vote != null) | .vote] | group_by(.) | map({vote: (if .[0] == 1 then "up" else "down" end), count: length})' "$CHAT_FILE"

# Vote down reasons
jq '[.requests[] | select(.voteDownReason != null) | .voteDownReason] | group_by(.) | map({reason: .[0], count: length})' "$CHAT_FILE"

10. Extract File Edit Statistics

# Files edited with accept/reject status (1=Keep, 2=Undo, 3=UserModification)
jq '[.requests[].editedFileEvents[]? | {uri: .uri.path, status: (if .eventKind == 1 then "kept" elif .eventKind == 2 then "undone" else "modified" end)}]' "$CHAT_FILE"

# Count of edits by status
jq '[.requests[].editedFileEvents[]?.eventKind] | group_by(.) | map({status: (if .[0] == 1 then "kept" elif .[0] == 2 then "undone" else "modified" end), count: length})' "$CHAT_FILE"

11. Detect Errors and Cancellations

# Failed requests (modelState.value == 3)
jq '[.requests[] | select(.modelState.value == 3) | {id: .requestId, error: .result.errorDetails.message}]' "$CHAT_FILE"

# Cancelled requests (modelState.value == 2)
jq '[.requests[] | select(.modelState.value == 2)] | length' "$CHAT_FILE"

# Error codes
jq '[.requests[] | select(.result.errorDetails != null) | .result.errorDetails.code] | group_by(.) | map({code: .[0], count: length})' "$CHAT_FILE"

11b. Rejection Analysis

Rejections include cancelled requests, failed requests, and cancelled/rejected tool invocations.

# Rejections grouped by model
jq '
  [.requests[] | {
    model: .modelId,
    state: .modelState.value,
    error_code: .result.errorDetails.code,
    cancelled_tools: ([.response[] | select(.kind == "toolInvocationSerialized" and .isConfirmed.type == 0)] | length)
  }]
  | group_by(.model)
  | map({
      model: .[0].model,
      total_requests: length,
      cancelled: ([.[] | select(.state == 2)] | length),
      failed: ([.[] | select(.state == 3)] | length),
      tool_rejections: ([.[].cancelled_tools] | add),
      error_codes: ([.[] | select(.error_code != null) | .error_code] | group_by(.) | map({code: .[0], count: length}))
    })
  | map(. + {rejection_rate: (if .total_requests > 0 then (((.cancelled + .failed + .tool_rejections) / .total_requests) * 100 | floor) else 0 end)})
  | sort_by(-.total_requests)
' "$CHAT_FILE"

# Common rejection reasons (error codes across all requests)
jq '
  [.requests[] | select(.result.errorDetails != null) | {
    code: .result.errorDetails.code,
    message: .result.errorDetails.message
  }]
  | group_by(.code)
  | map({code: .[0].code, count: length, sample_message: .[0].message})
  | sort_by(-.count)
' "$CHAT_FILE"

# User vote-down reasons (explicit rejection feedback)
jq '
  [.requests[] | select(.voteDownReason != null) | .voteDownReason]
  | group_by(.)
  | map({reason: .[0], count: length})
  | sort_by(-.count)
' "$CHAT_FILE"

12. Terminal Commands Analysis (Automation Opportunities)

# Identify repeated command patterns (candidates for scripts)
jq '
  [.requests[].response[]
    | select(.kind == "toolInvocationSerialized" and .toolId == "run_in_terminal")
    | (.invocationMessage // "" | tostring | gsub("^[^`]*`"; "") | gsub("`[^`]*$"; "") | split("\n")[0] | split(" ")[0:2] | join(" "))
  ]
  | group_by(.)
  | map({pattern: .[0], count: length})
  | sort_by(-.count)
  | .[0:10]
' "$CHAT_FILE"

13. Model Performance

# Response time statistics grouped by model
jq '
  [.requests[] | select(.result.timings.totalElapsed != null) | {
    model: .modelId,
    elapsed: .result.timings.totalElapsed,
    first_progress: (.result.timings.firstProgress // 0)
  }]
  | group_by(.model)
  | map({
      model: .[0].model,
      count: length,
      avg_elapsed_sec: (([.[].elapsed] | add) / length / 1000 | . * 100 | floor / 100),
      avg_first_progress_ms: (([.[].first_progress] | add) / length | floor),
      total_elapsed_sec: (([.[].elapsed] | add) / 1000 | floor)
    })
  | sort_by(-.count)
' "$CHAT_FILE"

# Model effectiveness: cancelled/failed rate by model
jq '
  [.requests[] | {model: .modelId, state: .modelState.value}]
  | group_by(.model)
  | map({
      model: .[0].model,
      total: length,
      complete: ([.[] | select(.state == 1)] | length),
      cancelled: ([.[] | select(.state == 2)] | length),
      failed: ([.[] | select(.state == 3)] | length),
      success_rate: (
        ([.[] | select(.state == 1)] | length) as $ok |
        (length) as $total |
        if $total > 0 then (($ok / $total) * 100 | floor) else 0 end
      )
    })
  | sort_by(-.total)
' "$CHAT_FILE"

Metrics Available

✅ Reliably Extractable

• Total requests/turns
• Models used (with counts)
• Session start/end timestamps
• Response timings (totalElapsed, firstProgress)
• Tool usage breakdown
• Manual vs auto-approval counts
• Terminal command exit codes
• Response states (complete, cancelled, failed)
• User feedback votes and reasons
• File edit acceptance/rejection status

⚠️ Partially Available

• Extended thinking content (may be encrypted)
• timeSpentWaiting - appears to be time waiting for user confirmation, not agent processing time

❌ Not Available

• Custom agent names - export only shows github.copilot.editsAgent, not custom agent files (see Known Limitations)
• Token counts
• Actual cost in dollars
• User reaction/thinking time between responses
• Agent handoff events as distinct records

Output

Metrics extracted from chat export for inclusion in retrospective.md.