fix(hourly): polish cache schema, dispatch, sort defaults, and float guards - #359
Conversation
|
@crhan is attempting to deploy a commit to the Inevitable Team on Vercel. A member of the Team first needs to authorize it. |
![cubic-dev-ai[bot]](https://avatars.githubusercontent.com/in/1082092?s=60&v=4){kind=small}
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
|
Thanks for the thorough review. Addressed in 974b28c: Fixed:
Won't fix / false positives:
|
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
92930f1 to
1bd3758
Compare
|
The latest updates on your projects. Learn more about Vercel for GitHub. 1 Skipped Deployment
|
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
…defaults, and float guards - bump TUI cache schema to 6 and core message_cache schema to 7 so old caches with empty hourly/turn fields don't deserialize as Fresh - restrict is_human_turn's '<' filter to a known set of internal tool tags so user prompts containing HTML/XML are no longer dropped from turn counts - route 'tokscale hourly' through the standard TUI/CLI dispatch so the TUI Hourly tab is reachable interactively and --light is honored - guard format_cost_per_million against NaN/Inf and non-positive denominators - stop re-sorting the entire hourly/daily dataset every render in overview chart; consume already-sorted data - match CLI fallback behavior in TUI hourly bucketing for messages with missing/zero timestamps
junhoyeo
changed the title
1bd3758 to
94af15e
Compare
…nd client filter UX (#467) Pre-release correctness and security fixes across the four PRs landed this cycle (#464, #454, #359, #355). ## Antigravity Trust boundary around the local language-server RPC was too loose: - Bound RPC body sizes at 16 MiB (Content-Length, chunked, and read-to-end paths) - Verify process identity by checking the executable path (`lsof` on macOS, `/proc/<pid>/exe` on Linux); accept paths containing either `antigravity` or `language_server` since some Antigravity-flavored servers launch from generic `language_server` binaries with `--app_data_dir antigravity` - Probe candidate endpoints with a real RPC call and JSON-shape check (probe body capped at 4 KiB) instead of trusting any 200 response; consume HTTP headers before reading body so the cap applies to the JSON body alone - Lock concurrent syncs on a per-cache PID lock file with bounded retry (3 attempts) instead of unbounded recursion. Eviction is gated on PID liveness only — long-running syncs no longer get stomped on by age-based timeouts. - Enforce manifest version on load: future versions abort, older versions start fresh - Recover corrupted manifests by moving them aside as `manifest.json.corrupt-<ts>` instead of failing every subsequent sync ## Codebuff Parser correctness around silent data loss: - Accumulate run-state usage across the full reverse `messageHistory` walk instead of returning on the first signal-bearing entry. Previously a newest assistant entry carrying only a model id would short-circuit the walk and silently drop real token counts on earlier entries. - Include the source-array ordinal in the fallback dedup key so two id-less assistant messages with identical session/timestamp/model/tokens no longer collapse into a single record - Reject non-positive numeric timestamps in the shared `parse_timestamp_value`/`parse_timestamp_str` helpers so messages with `timestamp: 0` or negative epochs fall through to chat-id / file-mtime fallback chain ## TUI + parsing User-visible client-filter UX: - Move `SYNTHETIC_HOTKEY` from `'x'` to `'n'`. `'x'` collided with Mux's hotkey, and the dispatch order made the displayed `[x]` for Synthetic purely cosmetic - Enable `ignore_case` on `--client/-c` so `OPENCODE`, `Codebuff`, and `antigravity` all parse as the same canonical filter ## Test results - `cargo test -p tokscale-cli` — 415 unit + 83 integration, all pass - `cargo test -p tokscale-core` — 566 unit + 10 codebuff + 3 hermes, all pass - `cargo check --workspace` — clean Each fix has a regression test where applicable.
3 participants
Summary
tokscale hourlyCLI subcommand with--json,--light, date filters, and all client flags (including the newly added--kilo)h) to switch between daily and hourly breakdownCache%to clarify it's a multiplier not a percentageDetails
The hourly report aggregates token usage, cost, turn count, and message count by hour. The Turn count uses a 10-second dedup window to collapse rapid consecutive user messages into a single logical turn (matching CC's session behavior).
Cache× is defined as
cache_read_tokens / input_tokens— a value > 1 means more tokens were served from cache than paid for as fresh input, indicating good cache utilization.Test notes
Three pre-existing scanner tests (
test_scan_all_clients_claude,test_scan_all_clients_multiple,test_scan_all_clients_headless_paths) fail on upstreammainas well — they scan the developer's real~/.claude/projectsinstead of the temp dir due to an unrelated env/path issue. Not introduced by this PR.🤖 Generated with Claude Code