Session Analysis Report
Generate a diagnostic report for an agent session by running the Analyze-Session.ps1 script included with this skill. The script auto-detects whether the current session was produced by GitHub Copilot CLI or Claude Code from environment variables and on-disk file format, and dispatches to the appropriate parser. If neither harness can be detected, the script exits with a clear error.
Privacy and sensitivity — surface this guidance to the user
Analyze-Session.ps1 always:
- Embeds a "Privacy and sensitivity" section at the top of the generated
session-report.md (right above the Overview table), and
- Prints a yellow PRIVACY NOTICE banner to the console when it finishes writing the file.
You (the agent) must surface this guidance to the user in your response — do not let it stay buried in script output the user might not have read. When you finish running the script and reporting the findings, include a short privacy reminder in your reply to the user, in plain second-person language. Use this template, adapting wording as needed:
⚠️ Heads-up before you share session-report.md — this file contains your unredacted session transcript: file contents and paths the agent read or edited, your prompts verbatim (including any secrets you may have pasted), tool output, environment values, and local paths under C:\Users\<you>\…. You're responsible for what you share — please open the file in your editor and read it end-to-end before attaching it to a public issue, posting it in chat, or sending it outside your organization. Redact anything sensitive. If you only need to share the high-level metrics, ask me to summarize the file instead of attaching it.
If the user only wants the high-level metrics (turn counts, skill usage, build success rate) without the per-turn detail, summarize the report and share the summary instead of the file — and tell the user that's what you're doing so they don't have to read it themselves to confirm.
Steps
- Run the analysis script to generate the report:
# Analyze the most recent session (auto-detects harness) and save report
.\Analyze-Session.ps1 -OutputFile session-report.md
# Or analyze a specific session by ID (searched in both harness locations)
.\Analyze-Session.ps1 -SessionId "<session-id>" -OutputFile session-report.md
# Or analyze a transcript file directly (format sniffed from content)
.\Analyze-Session.ps1 -EventsFile <path-to-transcript.jsonl> -OutputFile session-report.md
# Force a specific format if auto-detection picks the wrong harness
.\Analyze-Session.ps1 -Format ClaudeCode -OutputFile session-report.md
# Skip subagent transcripts (Claude Code only) for a parent-only view
.\Analyze-Session.ps1 -SkipSubagents -OutputFile session-report.md
Detection rules:
- The current session is preferred when an explicit ID is available:
COPILOT_AGENT_SESSION_ID (Copilot CLI) or CLAUDE_SESSION_ID (Claude Code) take priority over "most recently modified" so a parallel session in another terminal can't shadow the one the skill was invoked from.
- Environment first:
CLAUDECODE=1 or CLAUDE_CODE_ENTRYPOINT -> Claude Code; COPILOT_* env vars -> Copilot.
- For Claude Code, the most-recent JSONL whose
cwd matches the current working directory is preferred.
- For an explicit
-EventsFile, the format is sniffed from the first events.
- If neither harness is detected, the script exits with a non-zero status and a message naming both supported locations.
Review the generated report — read session-report.md and summarize key findings for the user:
- How many turns, how long, token usage
- What skills were loaded and when
- Build success/failure pattern
- Any stuck patterns or tooling issues detected
Add your own observations — append a section to the report with any additional context:
- Was the final app working? What's missing?
- Quality assessment of the generated code
- Suggestions specific to what went wrong
Include any tooling improvements or recommendations based on the analysis.
- Are there rules that need to be added to the Roslyn analyzer to prevent common mistakes detected during the session?
- Were there bugs or issues with winapp run or the BuildAndRun.ps1 script?
- Are there features that could be added to lower the number of turns required to complete a task?
What the Report Covers
| Section |
Details |
| Overview |
Harness, session ID, model, duration, turns, tokens (incl. cache tokens for Claude Code) |
| Prompt |
The original user request |
| Turn Breakdown |
Turns and tokens by category (building, coding, exploring, subagent dispatch, etc.) |
| Skills |
Which were invoked and when, including from inside subagent transcripts |
| Subagents |
(Claude Code only) Per-agent breakdown of dispatched subagents and their work |
| Build Analysis |
Build attempts, failures, errors, whether BuildAndRun.ps1 was used |
| Stuck Patterns |
Build loops, repeated file reads, obj/ clean cycles |
| Tooling Issues |
Auto-detected improvement opportunities |
| Turn Detail |
Every turn with tools used and errors flagged, parent and subagent transcripts shown separately |
When to Use
- When the user asks for a session report to understand what happened during an agent session.
1---2name: winui-session-report3description: Analyze the current or a recent agent session (GitHub Copilot CLI or Claude Code) and generate a diagnostic report. Use when asking for session feedback, debugging agent behavior, or reviewing what happened during a build session.4---56### Session Analysis Report78Generate a diagnostic report for an agent session by running the `Analyze-Session.ps1` script included with this skill. The script auto-detects whether the current session was produced by GitHub Copilot CLI or Claude Code from environment variables and on-disk file format, and dispatches to the appropriate parser. If neither harness can be detected, the script exits with a clear error.910### Privacy and sensitivity — surface this guidance to the user1112`Analyze-Session.ps1` always:13141. **Embeds a "Privacy and sensitivity" section at the top of the generated `session-report.md`** (right above the Overview table), and152. **Prints a yellow PRIVACY NOTICE banner to the console** when it finishes writing the file.1617**You (the agent) must surface this guidance to the user in your response — do not let it stay buried in script output the user might not have read.** When you finish running the script and reporting the findings, include a short privacy reminder in your reply to the user, in plain second-person language. Use this template, adapting wording as needed:1819> ⚠️ **Heads-up before you share `session-report.md`** — this file contains your unredacted session transcript: file contents and paths the agent read or edited, your prompts verbatim (including any secrets you may have pasted), tool output, environment values, and local paths under `C:\Users\<you>\…`. You're responsible for what you share — please open the file in your editor and read it end-to-end before attaching it to a public issue, posting it in chat, or sending it outside your organization. Redact anything sensitive. If you only need to share the high-level metrics, ask me to summarize the file instead of attaching it.2021If the user only wants the high-level metrics (turn counts, skill usage, build success rate) without the per-turn detail, summarize the report and share the summary instead of the file — and tell the user that's what you're doing so they don't have to read it themselves to confirm.2223### Steps24251. **Run the analysis script** to generate the report:2627```powershell28# Analyze the most recent session (auto-detects harness) and save report29.\Analyze-Session.ps1 -OutputFile session-report.md3031# Or analyze a specific session by ID (searched in both harness locations)32.\Analyze-Session.ps1 -SessionId "<session-id>" -OutputFile session-report.md3334# Or analyze a transcript file directly (format sniffed from content)35.\Analyze-Session.ps1 -EventsFile <path-to-transcript.jsonl> -OutputFile session-report.md3637# Force a specific format if auto-detection picks the wrong harness38.\Analyze-Session.ps1 -Format ClaudeCode -OutputFile session-report.md3940# Skip subagent transcripts (Claude Code only) for a parent-only view41.\Analyze-Session.ps1 -SkipSubagents -OutputFile session-report.md42```4344Detection rules:45- The current session is preferred when an explicit ID is available: `COPILOT_AGENT_SESSION_ID` (Copilot CLI) or `CLAUDE_SESSION_ID` (Claude Code) take priority over "most recently modified" so a parallel session in another terminal can't shadow the one the skill was invoked from.46- Environment first: `CLAUDECODE=1` or `CLAUDE_CODE_ENTRYPOINT` -> Claude Code; `COPILOT_*` env vars -> Copilot.47- For Claude Code, the most-recent JSONL whose `cwd` matches the current working directory is preferred.48- For an explicit `-EventsFile`, the format is sniffed from the first events.49- If neither harness is detected, the script exits with a non-zero status and a message naming both supported locations.50512. **Review the generated report** — read `session-report.md` and summarize key findings for the user:52 - How many turns, how long, token usage53 - What skills were loaded and when54 - Build success/failure pattern55 - Any stuck patterns or tooling issues detected56573. **Add your own observations** — append a section to the report with any additional context:58 - Was the final app working? What's missing?59 - Quality assessment of the generated code60 - Suggestions specific to what went wrong61624. Include any tooling improvements or recommendations based on the analysis.63 - Are there rules that need to be added to the Roslyn analyzer to prevent common mistakes detected during the session?64 - Were there bugs or issues with winapp run or the BuildAndRun.ps1 script?65 - Are there features that could be added to lower the number of turns required to complete a task?6667### What the Report Covers6869| Section | Details |70|---------|---------|71| Overview | Harness, session ID, model, duration, turns, tokens (incl. cache tokens for Claude Code) |72| Prompt | The original user request |73| Turn Breakdown | Turns and tokens by category (building, coding, exploring, subagent dispatch, etc.) |74| Skills | Which were invoked and when, including from inside subagent transcripts |75| Subagents | (Claude Code only) Per-agent breakdown of dispatched subagents and their work |76| Build Analysis | Build attempts, failures, errors, whether BuildAndRun.ps1 was used |77| Stuck Patterns | Build loops, repeated file reads, obj/ clean cycles |78| Tooling Issues | Auto-detected improvement opportunities |79| Turn Detail | Every turn with tools used and errors flagged, parent and subagent transcripts shown separately |8081### When to Use8283- When the user asks for a session report to understand what happened during an agent session.
unliftedq