feat(codex): add tool-excluded active Tok/s metrics

CHANGELOG.md

@@ -2,6 +2,11 @@
 
 ## Unreleased
 
+### Added (CLI)
+- **Codex throughput tracking**: per-model Tok/s in the dashboard and report, active time excludes tool wait. (#805, thanks @ihearttokyo)
+
+### Fixed (CLI)
+- **Codex parser corrections**: fork-replay no longer double-counts `patch_apply_end` and `mcp_tool_call_end`; `exec` is normalized to Bash; `custom_tool_call` events are handled; token_count lines larger than 32 KiB now parse exact token counts instead of estimating. Codex session cache bumps from v7 to v8 for a one-time re-parse. Only tool attribution changes for ordinary sessions, leaving their cost identical; sessions that logged an oversized token_count line are repriced from exact counts instead of an estimate. (#805)
 ### Fixed
 - Claude Desktop and Cowork sessions are discovered for Windows Microsoft Store (MSIX) installs. (#611)
 

src/codex-cache.ts

@@ -11,9 +11,10 @@ import type { ParsedProviderCall } from './providers/types.js'
 // v5: also attribute CLI-wrapped MCP calls (`mcp-cli call server tool`) that
 // Codex logs as a plain exec_command (issue #478 follow-up). Force a re-parse
 // so sessions cached under v4 pick up the CLI-MCP attribution.
-// v6: rich-session-capture — per-call locAdded/locRemoved/editFailed from
+// v6/v7: rich-session-capture — per-call locAdded/locRemoved/editFailed from
 // patch_apply_end. Sessions cached under v5 lack these fields; re-parse to add.
-const CODEX_CACHE_VERSION = 7
+// v8: persist native MCP timing and compact invocation attribution.
+const CODEX_CACHE_VERSION = 8
 const CACHE_FILE = 'codex-results.json'
 
 type FileFingerprint = { mtimeMs: number; sizeBytes: number }