repospend 0.0.4 → 0.0.6

package/CHANGELOG.md

@@ -2,6 +2,26 @@
 
 All notable changes to RepoSpend will be documented in this file.
 
+## 0.0.6
+
+- Improve Codex session metadata recovery from rollout logs, including standalone Codex Desktop sessions and imported Claude model metadata.
+- Add an optional source-scoped app display mode so shared surfaces can be shown as source-specific entries such as "Codex on VS Code".
+- Improve sparse timeline charts and RTK unavailable-state guidance in the dashboard.
+- Document why RepoSpend token totals can differ from `ccusage` while API-equivalent costs still match.
+- Show a git-branch icon next to repos with a confirmed git root and a folder icon for path-inferred repos, giving a quick visual signal of data confidence.
+- Fix Windows extended-length path prefix (\\?\) handling in git root detection, which caused repos on Windows to incorrectly appear as unverified.
+
+
+## 0.0.5
+
+- Fix Codex token under-count on long sessions by summing per-turn `last_token_usage` instead of taking the max of `total_token_usage`, which Codex resets on each context compaction.
+- Track Codex context compactions per session (auto vs. manual `/compact`) and surface them in the session detail panel.
+- Filter Codex startup-context blobs (AGENTS.md/INSTRUCTIONS/environment_context) out of the prompt timeline and prompt counts.
+- Dedupe Claude Code token usage across resumed `.jsonl` files using `requestId` + message `id`, so replayed assistant turns are not double-billed.
+- Label Codex sub-agent sessions as "Codex subagent" instead of folding them into "Codex app".
+- Show the running RepoSpend version in the dashboard footer.
+- Publish a GitHub release with the matching `CHANGELOG.md` section whenever a `v*.*.*` tag is pushed.
+
 ## 0.0.4
 
 - Automatically retry Codex SQLite access after attempting to rebuild `better-sqlite3` when an `npx` or local install has a Node native module ABI mismatch.

package/README.md

@@ -156,6 +156,31 @@ Claude Code cost may show as unknown when local files do not include token count
 or a model name. RepoSpend shows those sessions with unknown tokens/cost rather
 than guessing.
 
+### Why RepoSpend's token totals can look lower than `ccusage`
+
+If you compare RepoSpend to `ccusage` you may see matching costs but different
+"total tokens", especially for Codex where cache reads dominate. This is
+expected.
+
+RepoSpend treats `cached_input_tokens` as a **subset of** `input_tokens`, which
+is how the OpenAI and Anthropic APIs document the field. The "total tokens"
+number reflects the tokens the model actually processed. `ccusage` adds cached
+tokens as a separate line item in its "Total Tokens" column, which inflates the
+total but does not change billing.
+
+The dollar cost is the source of truth and should agree between the two tools
+within rounding. If costs diverge, that is a real discrepancy worth
+investigating; token-count divergence on its own usually is not.
+
+### Codex Desktop on Windows
+
+RepoSpend captures Codex **CLI** usage on Windows from `~/.codex/`. The Codex
+**Desktop app** (installed as the MSIX package
+`OpenAI.Codex_<id>` under `%LOCALAPPDATA%\Packages\`) does not persist session
+transcripts or per-turn token usage to disk. Only Electron caches and debug
+logs are stored locally. Real session data lives server-side. Neither RepoSpend
+nor `ccusage` can report on Codex Desktop usage from local files alone.
+
 ## What It Reads
 
 RepoSpend only reads local files. It does not edit Codex, Claude Code, or RTK