gem-dota 0.4.3

Dota 2 Source 2 replay parser for data science and ML workflows

💻 Parse Source 2 Dota 2 .dem replays in Python or with a command-line interface
⚔️ Access tick-level hero/entity state and combat/event timelines
📈 Calculate match and player statistics (K/D/A, economy curves, lane metrics, teamfights)
🔎 Decode replay internals from bytes to protobuf packets to normalized events
🗺️ Analyze spatial data (wards, movement trails, lane heatmaps, objective timelines)
🎞️ Generate HTML analysis reports with combat, vision, economy, draft, and movement views
🧰 Export analysis-ready outputs as structured models, DataFrames, JSON, and Parquet

[!NOTE] Project status. The core parsing pipeline — binary reader, entity delta system, combat log (S1 + S2), string tables, extractors, and the ParsedMatch output model — is stable and validated against OpenDota outputs (see Known limitations for the remaining parity residuals). Features marked (experimental) — vision estimation, farming-pattern analysis, and Roshan conversion — are best-effort reconstructions, not ground-truth telemetry, and their heuristics may change. Published to PyPI as gem-dota (import name gem).

---

Why Gem?

Named after the Gem of True Sight — it reveals what is normally hidden. Replays are dense binary blobs; gem surfaces that information as structured Python objects you can analyze directly.

Three reasons it exists:

• Python-native. Data/ML/AI work lives in Python. The excellent Go/Java parsers aren't the first language for that audience — gem makes replay parsing approachable from a notebook and easy to plug into pandas and ML pipelines.
• Reach private/high-MMR games. Around 8500+ MMR (Immortal Draft), many matches are effectively invisible to OpenDota/Dotabuff/STRATZ public API flows. Parsing your own replays is the most reliable path for serious self-review.
• Data ownership & transparency. API/GraphQL outputs are already-processed interpretations with information loss and hidden assumptions. gem is open source and inspectable end-to-end, so you can understand the data from first principles.

More context

Skadistats once open-sourced SMOKE (Cython-based rather than pure Python), but it is no longer maintained; gem aims to fill that gap for today's Python/data community. The implementation is an independent Python reimplementation cross-checked against Manta (Go), Clarity (Java), and the OpenDota parser (Java) — see Acknowledgements.

---

Installation

Requires Python 3.10+.

[!IMPORTANT] Parquet export requires either pyarrow or fastparquet in your environment.

Install from PyPI

# pip
pip install gem-dota

# poetry
poetry add gem-dota

# uv (project dependency)
uv add gem-dota

Development / Contributing setup

git clone https://github.com/whanyu1212/gem-dota
cd gem
uv sync --group dev

[!TIP] Most users do not need to download hero/item icon assets. Icon fetching is only required for local report/example rendering that displays portraits or item/rune icons; when you need it, run python -m gem reports assets download --icons.

---

Quick start

[!NOTE] Replace "my_replay.dem" with your own replay path. The parser accepts plain .dem files and returns structured Python objects.

import gem

match = gem.parse("my_replay.dem")

# Draft — who was picked and banned?
for event in match.draft:
    action = "PICK" if event.is_pick else "BAN"
    print(f"{action}: {gem.constants.hero_display(event.hero_name)}")

# Per-player summary
for player in match.players:
    print(
        f"{player.player_name} ({gem.constants.hero_display(player.hero_name)}): "
        f"{player.kills}/{player.deaths}/{player.assists}  "
        f"{player.net_worth_t_min[-1] if player.net_worth_t_min else 0:,} NW  "
        f"{player.stuns_dealt:.1f}s stuns"
    )

# Look up a specific player by hero name (display name, NPC name, or bare suffix)
axe    = gem.find_player(match, "Axe")
am     = gem.find_player(match, "Anti-Mage")
sf     = gem.find_player(match, "npc_dota_hero_nevermore")
if axe:
    print(f"Axe: {axe.kills}/{axe.deaths}/{axe.assists}")

# Parse to DataFrames
dfs = gem.parse_to_dataframe("my_replay.dem")
players   = dfs["players"]     # one row per player per sample tick
positions = dfs["positions"]   # one row per (player, tick) with x/y coords
combat    = dfs["combat_log"]  # all combat log entries
wards     = dfs["wards"]       # ward placements

# Parse to JSON
json_str = gem.parse_to_json("my_replay.dem", indent=2)

# Or convert an already-parsed match
match = gem.parse("my_replay.dem")
json_str = gem.to_json(match)
data     = gem.to_dict(match)  # plain Python dict

# Export all DataFrames to Parquet files (one file per table)
paths = gem.parse_to_parquet("my_replay.dem", output_dir="./parquet_out")

# Or export from an already-parsed match
paths = gem.to_parquet(match, output_dir="./parquet_out")

[!IMPORTANT] If Parquet export fails, install pyarrow (recommended) or fastparquet and retry.

CLI

gem ships with a command-line interface for quick inspection, export, and batch processing:

# Print a match summary
python -m gem match.dem

# Export to JSON (stdout or file)
python -m gem match.dem --format json
python -m gem match.dem --format json --output out.json

# Export all tables to Parquet
python -m gem parse match.dem --format parquet --output ./parquet_out

# Parse a whole folder in parallel (one Parquet subdir per replay)
python -m gem batch replays/ --format parquet --output ./out --workers 4

# Concatenate all replays into one set of DataFrames
python -m gem batch replays/ --format dataframe --output ./out

# Show live progress and a timing breakdown
python -m gem match.dem --progress --timings

# Inspect or populate the HTML report asset cache
python -m gem reports assets status
python -m gem reports assets download --icons
python -m gem reports assets add-map assets/maps/Game_map_7.40.jpg

[!TIP] Start with python -m gem match.dem to validate parsing first, then add --format json or --format parquet for downstream workflows.

---

Showcase — what you can do today

gem can power a full match analysis workflow out of the box, including:

• overview dashboards,
• combat and teamfight breakdowns,
• vision timelines/maps,
• economy progression,
• draft + objectives + chat context,
• movement trails and time-series graphs.

Report screenshots

Overview	Gold / XP	Fight Breakdown & Combat
Damage Breakdown	Interactive Vision Map	Objective Timeline
Purchase Log & Buybacks	Draft Summary	Laning Efficiency

_{Interactive Movement Trail}

Reproduce this analysis

A sample report (TI14 Grand Finals G3 — XG vs Falcons) is available as a download from the docs site: whanyu1212.github.io/gem-dota/reports/

Use the packaged report builder from Python, or run the example wrapper:

import gem
from gem.reports import ReportAssets, write_html_report

match = gem.parse("path/to/your_replay.dem")
assets = ReportAssets.auto(fallback_map="assets/maps/Game_map_7.40.jpg")
write_html_report(match, "report.html", assets=assets)

uv run python examples/match_report.py path/to/your_replay.dem

By default it writes <replay_stem>_report.html to the project root.

[!NOTE] HTML report generation is heavier than plain parse/JSON export because it renders multiple charts, tables, and embedded assets.

---

Expected output of gem.parse(dem_path)

gem.parse(dem_path) returns a ParsedMatch object — a structured, analysis-ready view of the replay.

High-level shape:

• Match metadata: match ID, timing/tick context, and global match-level fields.
• Players (match.players): one ParsedPlayer per player with summary stats (K/D/A, damage, net worth, stuns, logs) plus time-series snapshots.
• Timeline/event collections: draft events, combat log entries, wards/smokes, Roshan/aegis events, objectives, chat, teamfights, and courier snapshots.
• Advantage/time-series arrays: values like radiant gold/XP advantage across game time.

In short: think of ParsedMatch as one container holding both per-player summaries and time-ordered match events, ready for direct Python analysis or conversion via parse_to_dataframe.

---

What you can extract

Data	API
Hero picks and bans with timestamps	ParsedMatch.draft
Per-player K/D/A + core combat summaries	ParsedPlayer.kills / .deaths / .assists / .damage
Gold / XP / net-worth time series	ParsedPlayer.times, .gold_t, .xp_t, .net_worth_t
Minute-aligned economy/XP series	ParsedPlayer.times_min, .total_earned_gold_t_min, .total_earned_xp_t_min
Radiant gold / XP advantage curves	ParsedMatch.radiant_gold_adv / .radiant_xp_adv
Ward placements with exact coordinates	ParsedMatch.wards
Smoke of Deceit activations + grouped heroes	ParsedMatch.smoke_events
Roshan kills + aegis events	ParsedMatch.roshans / .aegis_events
Tower and barracks kills	ParsedMatch.towers / .barracks
Teamfights with per-player breakdown	ParsedMatch.teamfights
Courier state snapshots per team	ParsedMatch.courier_snapshots
Damage by type (physical/magical/pure)	ParsedPlayer.damage_by_type / .damage_taken_by_type
Laning signals (role + lane-phase metrics)	ParsedPlayer.lane_role, .lane_efficiency_pct, .lane_gold_adv, .lane_xp_adv
Lane position heatmaps	ParsedPlayer.lane_pos
Stun seconds dealt per player	ParsedPlayer.stuns_dealt
Per-minute hero damage / healing / deaths / stuns	ParsedPlayer.total_hero_damage_t_min / total_hero_healing_t_min / total_deaths_t_min / total_stuns_t_min
Rune pickups per player	ParsedPlayer.runes_log
Buybacks per player	ParsedPlayer.buyback_log
Chat messages	ParsedMatch.chat
Purchase log per player	ParsedPlayer.purchase_log
Vision modifier events (Slardar, BH Track, Dust, Gem) (experimental)	ParsedMatch.vision_modifiers
Vision source estimation per point (experimental)	gem.estimate_vision(match, team, tick, x, y)
Hero / item / ability display names	gem.catalog (gem.constants remains compatible)
Self-contained HTML reports	gem.reports.build_html_report(), gem.reports.write_html_report()
Look up a player by hero name	gem.find_player(match, "Axe")

---

Releases

Full per-release detail lives in CHANGELOG.md and the GitHub Releases page; the highlights below summarize recent versions.

v0.3.0

A structural + correctness release. The supported top-level API (gem.parse, gem.ParsedMatch, gem.find_player, …) is unchanged.

• Package reorganization — internals grouped into focused subpackages (binary/, schema/, state/, combat/, extractors/, analysis/, catalog/, results/, reports/, replays/), each with its own README.md.
• Source-based combat attribution — damage/healing now attributed to the damage source (matching OpenDota); tower_damage is ~exact offline (up from ~87%), and unmapped combat-log types no longer inflate damage aggregates.
• Correctness fixes (multi-pass adversarial bug hunt) — day/night cycle (10-min, night at 5:00), ward lifespans (observer 6 min / sentry 7 min), teamfight gold/XP attribution, Roshan-conversion window double-count, reincarnation death counting, and coach-index K/D/A remap for HLTV replays.
• New fields — CombatLogEntry.damage_source_name, CombatLogEntry.will_reincarnate, and a backward-compatible CombatLogType enum.
• Removed — root-level compatibility shims (gem.reader, gem.models, …); use the top-level gem.* API or grouped subpackages.

Older releases

v0.2.8

• Neutral item tracking — DOTA_UM_FoundNeutralItem events are parsed into NeutralItemFoundEvent (player, item key, tier, enhancement/trinket fields), with model/DataFrame outputs and constants-audit coverage for newly observed item IDs.
• Neutral camp annotation tooling — scripts/audit_camp_annotations.py groups neutral deaths into camp zones with replay-derived evidence; combat-log parsing now preserves neutral-camp stack metadata and event locations when available.
• 7.41 data refresh — bundled constants updated for 7.41-era items/abilities, camp-zone annotations refreshed for confirmed camp type swaps, and a regenerated 7.40 map fixture with 7.41 camp overlays.
• OpenDota fixture tooling — scripts/fetch_opendota_fixture.py plus DreamLeague Season 29 fixture metadata for patch-7.41 validation.
• Icon-fetch caching — hero/item icon scripts now skip unchanged assets.

v0.2.7

• Experimental farming-pattern analysis — objective-aware camp-path context, larger interactive report view, and detailed documentation for the heuristic model and its limits.
• Roshan conversion analysis — gem.build_rosh_conversions(match) plus a dedicated Roshan Conversion report tab to show how each Roshan translated into fights, objectives, and map pressure.
• Sampling and validation fixes — player time-series sampling now stops at game end, hero movement sampling uses each player's canonical selected hero entity, and the OpenDota validator now supports broader replay-sampling workflows.
• Economy series split — ParsedPlayer.gold_t is now current unspent gold only, with cumulative earned gold exposed separately via ParsedPlayer.total_earned_gold_t.

v0.2.6

• Team identity fields — ParsedMatch.radiant_team_id, radiant_team_name, radiant_team_tag and dire_team_id, dire_team_name, dire_team_tag extracted from CDOTATeam entities. Default to 0/"" for pub games.
• Player Steam/account IDs — ParsedPlayer.steam_id (64-bit) and ParsedPlayer.account_id (32-bit, matches OpenDota/Dotabuff URLs) extracted from CDOTA_PlayerResource.
• HTML report scoreboard — account ID now displayed below each hero name in the scoreboard roster.

v0.2.5

• gem.fetch_replay(match_id, out_dir) — download and decompress a replay from OpenDota in one call. Lower-level helpers fetch_replay_url and download_and_decompress also exposed via the public API.
• gem.resolve_pick_team(event, players) — resolves the team for a draft pick/ban using the post-game player roster (more reliable than m_pGameRules.m_iActiveTeam in HLTV/coach replays).
• gem.net_worth_at(player, tick), gem.ward_vision_impact(ward, match), gem.is_active_teamfight_participant(player_stats), gem.format_npc_name(name) — new analysis helpers.
• Draft fix — DraftExtractor._resolve_name() now correctly halves doubled hero IDs (api_id * 2) before falling back to a direct lookup, fixing wrong hero resolution for bans in modern replays.
• Integration test — uv run pytest skips replay-backed integration tests by default. tests/test_draft_integration.py verifies picks/bans against the OpenDota API with one captains-mode pro replay when run via uv run pytest -m integration; set GEM_DRAFT_INTEGRATION_FULL=1 to run the broader 5-replay sample.
• Sample report — TI14 Grand Finals G3 (XG vs Falcons) report available as a download from the reports gallery.

v0.2.4

• Vision modifier tracking (experimental) — ParsedMatch.vision_modifiers is a new list of VisionModifierEvent records tracking every application of a vision-granting ability or item (Slardar Corrosive Haze, Bounty Hunter Track, Dust of Appearance, Gem of True Sight). Start/end ticks, caster, target, and team are all captured.
• gem.estimate_vision(match, team, tick, x, y) (experimental) — geometry-based vision estimation for a given point at a given tick. Returns a ranked list of VisionSource objects covering allied heroes (day/night radius), observer wards, and active vision modifier reveals. See API reference for limitations.
• Ward killer attribution fix — sentry/observer wards now correctly attribute the killer when the entity lifestate transition and the combat log DEATH event arrive at the same tick (same-tick ordering bug resolved).
• Fights tab — Active Reveals — the match report Fights tab now shows which heroes were under vision modifiers (Corrosive Haze, Dust, Track, Gem) during each teamfight window.
• HTML report size fix — report file size reduced from ~459 MB to ~58 MB by deduplicating base64 map/icon embeds. Map image (9 MB) was embedded 22× (once per teamfight minimap); hero icons were repeated up to 70× each. Both are now hoisted to JS globals and patched on load.
• Ward map heatmap y-flip fix — vision coverage heatmap overlay was rendered upside-down relative to ward dot positions; corrected.
• Sample report gallery — TI14 G3 report available as a download at whanyu1212.github.io/gem-dota/reports/.

v0.2.3

• Per-minute combat totals — total_hero_damage_t_min, total_hero_healing_t_min, total_deaths_t_min, total_stuns_t_min on ParsedPlayer. Monotonically increasing counters; diff any two indices for per-window rates. Targeted at ML feature extraction pipelines.
• gem.find_player(match, hero) — look up a player by display name, NPC name, or bare suffix without manual iteration.
• gem.constants.hero_npc_name(name) — reverse lookup from display name (e.g. "Anti-Mage") to NPC name ("npc_dota_hero_antimage").
• ParsedMatch.duration_minutes / duration_seconds — convenience properties for match length.
• Doc fixes — quickstart guide and match data guide had several references to nonexistent fields; all corrected and verified with a runnable examples/quickstart.py.

v0.2.2

• Batch processing — gem.parse_many(), gem.parse_many_to_dataframe(), gem.parse_many_to_parquet() for parallel multi-replay parsing via ProcessPoolExecutor.
• CLI batch subcommand — python -m gem batch replays/ --format parquet --output ./out; legacy bare-path invocation preserved.
• Docs — home page feature cards, annotated JSON output guide (real TI14 G3 XG vs Falcons replay), CLI reference guide, batch API reference page.

v0.2.1

• JSON export — gem.to_json(), gem.to_dict(), gem.parse_to_json() added to the public API.
• Parquet export — gem.to_parquet(), gem.parse_to_parquet() added (requires pyarrow or fastparquet).
• Rich CLI — live progress bar, timing summary, pixel-art banner in a box, Radiant/Dire colour-coded summary table.
• Docs — architecture page redesigned, diamond icon, laning pages added to nav, export formats documented throughout.
• Bug fixes — two mypy type errors resolved in __main__.py and DataFrame export.

v0.2.0

• Buyback gold cost — HTML report buyback table now shows gold spent per buyback using the exact Dota 2 formula floor(200 + net_worth / 13).
• Removed ParsedMatch.lotus_pickups — healing lotus pickups are not recorded in the .dem combat log under any event type; the field always returned an empty list and has been removed (breaking change).
• Known limitations documented — healing lotus (not parseable from replays) and reliable vs unreliable gold distinction added to README.
• Repo URL fixes — all links corrected from whanyu1212/gem to whanyu1212/gem-dota.
• Test coverage expanded — teamfight helpers, _dedup_purchase_log edge cases, HEAL/gold/XP attribution.
• Screenshot refresh — all report screenshots updated and resized to uniform dimensions.

v0.1.1

• Laning — lane role detection, lane-phase gold/XP efficiency metrics, and positional heatmaps per player (ParsedPlayer.lane_role, .lane_efficiency_pct, .lane_gold_adv, .lane_xp_adv, .lane_pos).
• Damage type breakdown — physical / magical / pure damage split for both dealt and taken (ParsedPlayer.damage_by_type, .damage_taken_by_type).
• Aghanim's Scepter/Shard abilities — extended ability metadata and correct parsing of Aghs-upgraded abilities.
• Teamfight detection — switched to pure temporal windowing; spatial centroid splitting removed for consistency.

v0.1.0

• Initial public release of gem-dota.
• Full Source 2 .dem parser pipeline — stream decoding, send tables, field paths, entity delta system, string tables.
• Game events and combat log ingestion (Source 1 legacy + Source 2 HLTV paths).
• Extractors: players (gold/XP/net worth time series, K/D/A, stuns), objectives (towers, barracks, Roshan, Tormentor), wards (with exact coordinates), courier, draft, teamfights.
• Rune pickups, buybacks, aegis events, smoke groups, chat, purchase log.
• ParsedMatch / ParsedPlayer output models, DataFrame export, and CLI.
• HTML match report example (draft, combat, vision, economy, teamfights, movement).

---

Components

Component	Description
binary/reader.py	BitReader — LSB-first bit reading, varint decoding, all binary primitives
binary/stream.py	DemoStream — outer message loop, Snappy decompression, magic check
schema/sendtable/	Schema layer — serializer + field tree parsed from CDemoSendTables
schema/field_decoder/	Type-dispatch decoders including quantized floats
schema/field_path/	Huffman-coded field path ops for addressing into the serializer tree
schema/field_state.py	Nested mutable field-value tree for entity state storage
schema/field_reader.py	Field decoder dispatch and entity field reading
state/string_table.py	Incremental key-history string tables
state/entities.py	Entity create/update/delete lifecycle and state
state/game_events.py	Game event schema and typed dispatch
combat/log.py	S1 (game event) and S2 (user message) combat log ingestion
parser.py	Top-level orchestrator wiring all subsystems together
api.py	High-level gem.parse(), JSON, DataFrame, and Parquet convenience helpers
cli.py	Command-line implementation behind python -m gem
combat/aggregator.py	Combat-log aggregation for per-player damage/healing/items/economy stats
results/assembly.py	Assembles final ParsedMatch output from extractors/aggregates
results/models.py	ParsedMatch / ParsedPlayer output dataclasses
results/dataframes.py	DataFrame export from ParsedMatch
catalog/	Bundled hero, item, ability, league, XP, and static map-data lookups
constants.py	Compatibility facade for older catalog imports
reports/	Self-contained HTML report generation from ParsedMatch
extractors/	Per-tick polling of entity state — players, lane, objectives, wards, courier, draft, teamfights
analysis/	Post-parse helpers — spatial/combat/vision lookups plus experimental map-context and Roshan-conversion builders
replays/	Bulk parsing (parse_many, parallel workers) and replay download/decompress from OpenDota/Valve CDN

Each package carries its own README.md with a deeper mental model, mechanics, and pitfalls. Module-level architecture is documented in docs/architecture.md.

---

Examples

# Comprehensive HTML analysis report (draft, combat, vision, economy, movement, etc.)
python examples/match_report.py path/to/your.dem

# Full replay summary — combat log + entity snapshots (developer-oriented baseline)
python examples/extraction_demo.py path/to/your.dem

# Match info from Steam API (requires STEAM_API_KEY env var)
python examples/steam_match_info.py <match_id>

examples/ti14_sample.json — annotated real JSON output from TI14 Grand Finals G3 (XG vs Falcons), used as the reference example in the docs.

---

Documentation

VitePress docs (concepts guide, API reference, architecture diagrams, replay parser tab):

cd docs
npm install
npm run docs:dev

Or visit the hosted docs at whanyu1212.github.io/gem-dota.

Topics covered: DEM binary format, Protocol Buffers, varint encoding, the entity delta system, field paths, combat log ingestion, and more.

[!TIP] New to replay internals: start with Getting Started → Bits and Bytes Primer, then continue into the Deep Dive section for stream/parser/entity layers.

---

Contributing

Contributions are welcome — see CONTRIBUTING.md for the full workflow and PR checklist. The repo is managed with uv and enforces lint, format, and type checks via pre-commit.

uv sync --group dev          # install project + dev dependencies

uv run pytest                # fast suite (skips slow/integration markers)
uv run pytest -m integration # replay-backed integration tests (needs fixtures)

uv run ruff check src/ tests/    # lint
uv run ruff format src/ tests/   # format
uv run mypy src/gem/             # type-check

[!TIP] Parser-logic changes should be verified against the reference parsers in refs/ (Manta → Clarity → OpenDota) and ship with a focused regression test. See CONTRIBUTING.md and CLAUDE.md for the parser-safety conventions.

---

AI-Assisted Development

If you use AI coding tools, see CLAUDE.md and AGENTS.md for project context, architecture, and coding conventions.

Use AI as acceleration, not substitution: take ownership of what you submit. Understand the code, run tests, and avoid shipping unreviewed AI slop.

---

Performance

gem is a pure-Python parser. The reference parsers in Go (Manta) and Java (Clarity) are faster in raw throughput — that is the expected trade-off. gem instead optimizes for Python-native ergonomics: parse a replay and work with the result directly in pandas/notebooks/ML pipelines, with an implementation you can read end-to-end. It is built to be fast enough for research and batch analysis (bulk parsing runs across multiple processes via gem.parse_many), not to win raw throughput benchmarks.

Parse cost scales with extraction scope: full per-tick state extraction is much heavier than event-only parsing, and HTML report generation (chart/asset rendering) is heavier still than plain parse/JSON export. For large jobs, prefer parse_many* with multiple workers and export to Parquet.

Published cross-language benchmark numbers are intentionally omitted until they can be produced fairly — same replay set, same extraction scope, same hardware, with warmup and run counts reported. If you run a comparison, please open an issue/PR with hardware specs, the command/config used, the replay sample list, and median/p95 numbers (both replays/sec and time per replay with replay sizes).

---

Known limitations

[!IMPORTANT] Treat advanced outputs like vision estimation and some attribution paths as best-effort reconstructions, not ground-truth telemetry.

• Vision estimation is geometry-only (experimental) — gem.estimate_vision and the vision features in the match report use straight-line distance only. Three things are not modelled: (1) high-ground vision penalties — cliff edges block uphill sight; (2) terrain line-of-sight blocking from trees and cliffs (requires a navmesh/walkability grid from game/dota/pak01_dir.vpk, which is not bundled); (3) per-hero vision range modifiers from abilities/items (e.g. Aghanim upgrades). Vision results should be treated as approximations.
• Roshan drops — Aegis, Cheese, Refresher Shard, and Aghanim's Blessing pickups are not in the combat log. Roshan kills are tracked, but the specific drop items are not.
• Healing Lotus pickups — not recorded in the .dem combat log under any event type, across all tested patches. The CDOTA_BaseNPC_LotusPool entity does not emit per-pickup events either. The lotus count exists in CMsgDOTAFantasyPlayerStats, a Steam Game Coordinator message never written into the replay file. While the Steam Web API (GetMatchDetails) can provide this as a fallback, it has little practical value here: high-MMR and private matches — the primary audience for gem — are often not accessible via the public API, defeating the purpose.
• Reliable vs unreliable gold — the combat log GOLD entries do not distinguish between reliable gold (from kills, objectives) and unreliable gold (from creep bounties). Only total gold earned per reason code is available.
• Smoke empty groups — if a smoke breaks instantly on activation (hero inside sentry truesight), the group list will be empty. This is correct game behaviour, not a parsing gap.
• Truncated/live replays — incomplete replays may return partial parsed output (or stop near the final corrupt block) instead of a perfect full-match result.
• Draft ID quirks — replay pick/ban IDs can differ from static hero API IDs in some patches/formats (commonly transformed IDs). gem normalizes these, but edge cases may still appear.
• Purchase attribution in spectator/HLTV paths — purchase events are not always directly hero-attributed in combat log data; reconstruction relies on entity state and may be incomplete in edge cases.
• Summon ownership edge cases — most summoned-unit attribution is handled, but complex ownership cases can still produce occasional mismatches.
• Report assets — hero/item icons and map images are not bundled in the package. Run python -m gem reports assets status to audit the local cache, python -m gem reports assets download --icons to download missing non-recipe icons, and python -m gem reports assets add-map assets/maps/Game_map_7.40.jpg to copy a map image into the cache. Recipe icons are skipped by default; pass --include-recipes if you need them.

---

Acknowledgements

gem stands on top of years of open work by the Dota replay community.

• Manta (Go, MIT) — primary reference for binary parsing logic and the entity delta system.
• Clarity (Java, BSD 3-Clause) — correctness authority for edge cases and the two-path combat log design.
• OpenDota parser (Java, MIT) — output schema authority for match data structure and conventions.
• dotaconstants (MIT) — hero, item, and ability metadata bundled as static assets.
• Valve for the Dota 2 replay ecosystem and continuously evolving game data surface.

No source code was copied from any of the above projects. They were used as reference implementations to understand protocol behaviour, verify correctness, and inform design decisions. All gem code is an independent Python reimplementation.

See THIRD_PARTY_LICENSES for full license texts.

---

Roadmap

Item	Status
Validation harness against OpenDota-style outputs	Ongoing
Docs expansion (cookbook + parsing-from-scratch walkthroughs)	Done
ML project — using parsed replay data to power supervised/unsupervised models for player and team performance analysis	Planned
Agentic project — LLM-powered replay analysis agent that reasons over structured match data to generate insights and coaching feedback	Planned
Frontend demo application (interactive replay analysis UI showcasing parser capabilities)	Planned
Rust acceleration for selected hot paths (PyO3 + maturin)	Deferred