Discover Debug Log Schema

Discover Debug Log Schema Skill

Reads VS Code Copilot Chat's debug logs on this machine and reports what is actually in them.

The extension reads copilotUsageNanoAiu out of these logs to bill VS Code Chat sessions exactly rather than estimating — see docs/logFilesSchema/vscode-chat-debug-log-format.md. That format is undocumented by Microsoft, we parse exactly one of its event types, and validate-session-schemas does not cover it. This skill closes that gap.

What it answers

1. Is the billing contract intact? Are the five llm_request attrs that _eatdlProcessLlmRequest() in src/tokenEstimation.ts depends on (inputTokens, outputTokens, cachedTokens, model, copilotUsageNanoAiu) still present? This matters because a rename fails silently: the parser still returns a result, copilotNanoAiu becomes 0, and every affected session quietly falls back to estimated pricing. No test fails, no error surfaces.
2. What are we leaving on the table? Every event type and field path in the log, marked PARSED/IGNORED per type and * per field for what the repo reads today. We consume llm_request only; request_start and request_end are known to exist and are parsed by nothing. If those carry timings, per-call latency would be available with no user configuration — nobody has checked.

Usage

Run from the repository root. Dependency-free Node, no build step:

# Default: last 30 days, 10 most-recent logs
node .github/skills/discover-debug-log-schema/discover-debug-log-schema.js

# Every log regardless of age, more files
node .github/skills/discover-debug-log-schema/discover-debug-log-schema.js --days 0 --max 50

# Machine-readable
node .github/skills/discover-debug-log-schema/discover-debug-log-schema.js --json

Options

Exit codes

• 0 — billing contract intact, or no debug logs found on this machine
• 1 — DRIFT: a field the parser depends on is missing from observed records
• 2 — bad arguments

NOT_OBSERVED (no llm_request events in the sampled files) is not drift — it means the window caught no LLM calls. Widen it with --days 0 before concluding anything.

Where it looks

<vscode-user>/workspaceStorage/<hash>/<ext-folder>/debug-logs/<sessionId>/main.jsonl

across every VS Code variant the extension supports (Code, Insiders, Exploration, VSCodium, Cursor, and the .vscode-server* remote roots) and all four extension-folder spellings (GitHub.copilot-chat, github.copilot-chat, GitHub.copilot, github.copilot) — mirroring readTokensFromDebugLog() in vscode-extension/src/extension.ts.

No logs found means Copilot Chat has not written any on this machine, not that the format changed. The script exits 0 in that case so it is safe in CI, where a runner has no local sessions.

Privacy

Field values are never printed by default — only paths, JSON types and counts. These logs contain prompts, responses and absolute file paths. Use --include-examples only on your own machine, and never paste that output into an issue or PR without reading it first.

Acting on results

• DRIFT → exact billing is broken right now. Open a real log, find what the field was renamed to, and fix _eatdlProcessLlmRequest() in src/tokenEstimation.ts plus the CONSUMED map in this skill's script. Update the format doc.
• New event types or fields → check them against the format doc. Anything useful (timings, tool durations, edit outcomes) is a candidate for a real feature, and needs no user configuration because this log is always written.
• Either way → fold what you learn into docs/logFilesSchema/vscode-chat-debug-log-format.md. That page exists because this path was undocumented once already and caused a wrong conclusion.

Keeping it honest

The CONSUMED map at the top of the script is the list of fields the repo actually reads. If you change what _eatdlProcessLlmRequest() parses, change that map in the same PR — otherwise this skill reports a contract that no longer matches the code.