Skip to content
AI usage dashboard

Local VS Code built-in Copilot Chat sessions, 2026-10-01

This record preserves the first implementation results. The same-day review and corrections supersede its “10 requests”, “one turn = model_call”, derived input-plus-output total and no-workspace directory conclusions. Current counts are 10 usage turns and 217 observed main-loop rounds; usage and calls are recorded separately.

Background and investigation

  • The user reported existing VS Code Copilot data and requested actual token usage/call counts, beyond credit/premium_interactions quota. This investigation overturned the earlier conclusion that VS Code Copilot Chat saved no per-call token information locally.
  • Read-only inspection of local candidate files/databases used only the permitted fields in the preparation instructions, without printing message text:
    • APPDATA/Code/User/globalStorage/github.copilot-chat/session-store.db: chronicle v3 sessions/turns/checkpoints/search_index; turns.assistant_response is plain text without usage. Related CLI layout remains rejected by the CLI adapter when usage format is unknown.
    • transcripts/*.jsonl: assistant.message/turn_start/turn_end/tool.execution, without token fields or model names.
    • debug-logs/id/main.jsonl: session_start only; chatEditingSessions: editing state.
    • Global/workspace state.vscdb: lmBaseCount/model is the same ordering seed, 204, for every model, rather than actual counts; github-*-usages stores authentication-extension lastUsed.
    • Actual usage files: workspaceStorage/hash/chatSessions/sessionId.jsonl, VS Code native session logs with promptTokens/completionTokens/copilotCredits/elapsedMs/ promptTokenDetails/modelTotals/toolCallRounds per request.
  • microsoft/vscode source inspected on 2026-10-01 established field semantics for local VS Code 1.140.0 / built-in Copilot Chat 0.68.0:
    • chatSessionStore.ts saves workspaceStorageHome/workspaceId/chatSessions, including no-workspace/chatSessions for windows without a workspace.
    • objectMutationLog.ts: kind 0 initializes the object, including after compact rewrites; kind 1 Set(k,v); kind 2 Push(k,v[],i), where i is the length after truncation; kind 3 Delete. More than 1,024 entries triggers a whole-file replacement.
    • chatModel.ts toJSON and chatSessionOperationLog.ts storageSchema v3: promptTokens is the last model call’s input, as IChatUsage documents. completionTokens accumulates output across calls within the turn, through _setUsage; modelTotals replaces that value when present. copilotCredits is turn-level credit converted from nano AIU. Other fields include elapsedMs/outputBuffer/modelTotals, sessionCopilotCredits and modelState with value/completedAt. IChatUsageModelTotal has model/inputTokens/ cachedTokens/outputTokens accumulated across all model calls, for agent-host sessions only. State values: 0 Pending, 1 Complete, 2 Cancelled, 3 Failed, 4 NeedsInput; sealed states are 1/2/3.
    • agentIntent.ts forwards API usage.prompt_tokens/completion_tokens and prompt_tokens_details.cached_tokens directly.
  • Periodic saveState writes sampled streaming counters: for example, 15 updates for 40 model rounds. Summing updates cannot reconstruct per-call input. toolCallRounds contains round records, including modelId/thinking.tokens/timestamp.

Implementation

  • New copilot_chat adapter, independent directory and session_log_v3 version registry; agent=vscode-copilot-chat, matching the OTel adapter’s service.name assignment:
    • Discovery: APPDATA/Code/User/workspaceStorage and Code - Insiders, plus corresponding macOS/Linux paths; bounded two-level enumeration of hash/chatSessions/*.jsonl and no-workspace/chatSessions. Manual roots accept a chatSessions directory, workspaceStorage directory or individual .jsonl file. One instance per chatSessions directory.
    • Detection: first-line kind=0 with v.version=3, sessionId and requests array. Version 3 is KnownVersion; other/missing versions use LatestFallback; empty files remain Pending.
    • Full replay each run: sampled object updates cannot be joined through incremental reads. Compaction bounds the file; native sample 5.25 MB, largest line 837 KB. Event key vscode-chat:sessionId:requestId prevents duplicate insertion. Identical content stays unchanged; growing streaming records use Replace. A compact rewrite changes the source generation and triggers replay, converging on the same event keys.
    • Initial mapping, superseded by the review above: one user turn produced one model_call, matching CLI assistant-message granularity. input_total=promptTokens, the last-call input and a lower bound for the turn, quality=reported; output_total=completionTokens, the whole-turn accumulated output; total was derived. When present, modelTotals replaced these with whole-turn totals per model: cached→cache_read, uncached derived, one event per model for mixed-model turns. occurred_at=completedAt, falling back to request timestamp with source_start; duration=elapsedMs, ttft=result.timings.firstProgress. Model preference: resolvedModel, then modelId without copilot/ prefix, then last-round modelId. State 1/2/3→Final; 0/4→Partial. Requests without token signals are skipped; credits alone do not create token events, and copilotCredits never becomes tokens.
    • No new database tables: existing agent-independent usage_events; quota_history from copilot-user-cache.json remains separate under agent=copilot.
  • Updated outdated CLI hidden_calls capability wording, the copilot_quota.rs header’s claim of being the only locally extractable data, and the matching data-contract.md section.

Validation and remaining limits

  • Synthetic unit tests: cargo test -p llm-usage-core copilot_chat, 17 passed. Covers v3/unknown/non-JSON/empty detection, last Set wins, Partial lifecycle, truncated Push, Delete, compact rewrite reset, no-usage skip, single/multiple modelTotals, malformed rows, missing timestamp and empty-log Pending.
  • Read-only native check: cargo run -p llm-usage-core –example real_verify_copilot_chat with a chatSessions directory and output under build/desktop-usage-validation/copilot-chat-real. files=3, events=10, diagnostics=0; historical count=10, input=3,408,279, output=320,141, one model claude-opus-4-8, partial=0. Repeat count unchanged, added=0, verdict=PASS. Per-request values match independent redacted extraction through Python object-log replay. This historical count does not represent the current model-call count.
  • Application daily_usage contained 2026-09-30 / vscode-copilot-chat / claude-opus-4-8 / historical count 10 / input_known_sum=3408279 / output_known_sum=320141, visible through agent-grouped overview/trend queries.
  • npm run verify exit 0: Markdown/Svelte/fmt/clippy -D warnings, 224+ full Rust tests and web build; npm run test:browser exit 0.
  • Per-call input within a turn is unsaved; default input remains a last-call lower bound, replaced by whole-turn modelTotals when available. Incomplete thinking-token coverage excludes output_reasoning. Inline chat outside chatSessions is excluded. At this stage, simultaneous native/OTel-file collection for the same agent could duplicate counts; choose one. Later partition selection is described in the review. completionTokens may overestimate when upstream isSameUsage fails to deduplicate repeated backend reports, acknowledged by an upstream comment. v4+ uses latest-compatible parsing and still needs native samples before a verified-format claim.