Skip to content

CHANGELOG.md

Latest commit

 

History

History
83 lines (61 loc) · 8.11 KB

CHANGELOG.md

File metadata and controls

83 lines (61 loc) · 8.11 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

0.2.0 (2026-07-18)

Features

  • add open source release plan and config (61c9449)
  • optimize loaders with memory-mapped SQLite and parallel workers, add build.sh automation, and integrate release-please (bdf5b2d)

Bug Fixes

  • update author name in readme (19e2c54)

Unreleased

Added

  • build.sh build automation script — versatile build script supporting --release (-ldflags="-s -w" -trimpath), --dev, and --upx modes, cross-compilation (--target and --all), and detailed --help documentation explaining build trade-offs and binary sizes.
  • --verbose / -v diagnostic flag — outputs non-fatal tool loading failures, missing databases, and parsing warnings directly to os.Stderr.
  • thermal upgrade command — self-upgrade by checking GitHub Releases, downloading the matching binary for your OS/arch, and atomically replacing the running executable.
  • thermal version command — prints version, commit hash, and build date (injected via ldflags at build time).
  • internal/version package — build-time version embedding via -ldflags.
  • Token Warriors / Activity Hunters split — the leaderboard now separates tools with real token consumption data from tools that only track activity (messages, steps). Tools are auto-categorized based on whether token data is available.
  • Cost column in the leaderboard for tools that track cost (OpenCode, MiMoCode, codewhale). Sub-cent formatting shows costs like $0.0032.
  • Devin loader — extracts real token metrics from message_nodes.metadata.metrics (input, output, cache, reasoning tokens), session duration from sessions table, and engagement from prompt_history. Devin moved from Activity Hunters to Token Warriors.
  • Codex loader rewrite — reads state_5.sqlite (threads.tokens_used, model, source, reasoning_effort, agent_role, session duration) as the primary source, supplements with rollout JSONL for the input/output/reasoning/cache token breakdown. Codex moved from Activity Hunters to Token Warriors.
  • codewhale loader rewrite — extracts metadata.total_tokens, metadata.cost.session_cost_usd, metadata.model, metadata.mode, and metadata.cumulative_turn_secs from session JSON files. codewhale moved from Activity Hunters to Token Warriors.
  • OpenCode loader rewrite — leverages pre-aggregated session.tokens_* columns, cost, agent, and summary_additions/deletions/files for faster and more comprehensive data extraction.
  • MiMoCode loader — aggregates from message.data.tokens and surfaces session.summary_additions/deletions/files and agent distribution.
  • Agy loader rewrite — reads overview.txt from all brain sessions (was only reading transcript.jsonl from 1 of 14 sessions). Now surfaces 304 steps across 14 sessions instead of 57 steps from 1 session. Extracts model info from transcript.jsonl USER_SETTINGS_CHANGE events.
  • Session duration tracking — LongestSessionMs field added to Summary, surfaced in dashboard for all tools that have session timestamps.
  • Model breakdown — ModelBreakdown field shows which models were used per tool (e.g., Codex: gpt-5.4-mini:6, gpt-5.4:1).
  • Agent/source breakdown — AgentBreakdown field shows source distribution (e.g., Codex: cli:7, vscode:3, exec:1).
  • Lines added/deleted and files touched — LinesAdded, LinesDeleted, FilesTouched fields added to Summary for tools that track code diffs (OpenCode, MiMoCode).
  • Progress bar — new internal/render/progress.go for loading indicator, hardened with atomic.Bool and sync.Once for thread safety.
  • Session cache — internal/loaders/cache.go for caching loader results.
  • GitHub Actions release workflow and GoReleaser config for cross-platform binary builds with version injection.

Performance & Optimization

  • Devin SQLite database optimization — added incremental delta scanning (row_id > c.MaxRowID), memory-mapped I/O (PRAGMA mmap_size), and byte-string pre-filtering (chat_message LIKE '%"assistant"%'), reducing query times on large (7GB+) databases from ~11s down to < 2ms.
  • SQLite memory-mapped I/O across all loaders — LoadOpenCodeData and LoadMiMoCodeData benefit from PRAGMA cache_size = -64000 and PRAGMA mmap_size = 30GB alongside raw string pre-filters.
  • Parallel multi-worker loaders — internal/loaders/agy.go, commandcode.go, and codex.go use bounded parallel worker pools (limit of 8 concurrent workers) to scan multi-file directories and JSONL rollouts simultaneously.
  • Binary size reduction — configured -trimpath and -ldflags="-s -w" in .goreleaser.yml and build.sh to strip DWARF symbol tables and host filepaths, reducing release binaries from ~15MB to ~9.8MB (35% reduction).

Changed

  • Leaderboard format — now split into two sections (Token Warriors, Activity Hunters) with a Cost column. Old format was a single flat list.
  • Tool registry architecture (DataSubdir) — refactored tool registry path resolution (internal/loaders/registry.go) to use clean DataSubdir property mappings instead of hardcoded switch cases.
  • README — updated supported tools table, leaderboard example, build instructions (build.sh), and How It Works section to reflect the new data sources, metrics, and optimizations.
  • .goreleaser.yml — added -trimpath and ldflags (-s -w) to inject version, commit, and date into the binary while minimizing size.

Fixed

  • Daylight Saving Time (DST) streak accuracy — replaced 24-hour time subtraction (24*time.Hour) in streak logic (internal/thermal/streak.go) with calendar date arithmetic (AddDate(0, 0, 1)) to guarantee accuracy during spring-forward (23h) and fall-back (25h) transitions. Ensured strictly ascending quartile thresholds.
  • Self-upgrader network and decompression guards — hardened thermal upgrade (cmd/thermal/upgrade.go) with package-wide 30-second HTTP timeouts, context propagation across requests, exact OS/arch asset checking, and io.LimitReader(tr, 250*1024*1024) decompression limits against disk starvation attacks.
  • Codex token over-counting (260x) — the old parser summed ALL cumulative total_token_usage values from every token_count event in every rollout file, producing 11.6B instead of the correct 43.9M. The new loader uses state_5.sqlite as the source of truth and only reads the LAST token_count event per rollout file for the breakdown.
  • command-code session inflation — the old glob *.jsonl matched 115 files (54 sessions + 54 checkpoints + 7 hooks-audit), inflating counts to 61 sessions and 6,303 messages. Now correctly filters to 54 sessions and 5,614 messages.
  • Agy session under-counting — the old loader only read transcript.jsonl (1 of 14 sessions = 57 steps). Now reads overview.txt from all 14 sessions (304 steps).
  • SQLite NULL handling — Codex state_5.sqlite has NULL values in model, reasoning_effort, and agent_role columns. Switched to sql.NullString to avoid scan errors.

v0.1.0 - 2026-06-12

Added

  • Initial open source release.
  • GitHub-style contribution heatmap for AI coding tools: MiMoCode, OpenCode, Codex, Agy, command-code, codewhale.
  • Leaderboard mode (thermal) showing all installed tools ranked by activity.
  • Per-tool heatmap mode (thermal --tool <name>).
  • JSON output (thermal --json).
  • Configurable weeks range (thermal --weeks 26).
  • Auto-detect mode (thermal --tool auto).
  • MIT license.