Every release, in order.

Changed

• plot_zone()’s frame crops to the data instead of a fixed 5.5-foot span: the ceiling now sits a fixed pad above the strike zone, extended only as far as the 97th percentile of the pitches actually drawn needs, rounded up to a clean half foot. A typical sample no longer carries a full foot of empty canvas above its tallest pitch, and the figure itself shrinks along with a tightened frame rather than leaving the blank space behind, since every panel keeps the same feet-to-inch scale on both axes (_INCHES_PER_FOOT). The 97th percentile, not the true max, decides the ceiling, so one stray backstop-bound fastball among 150 real pitches doesn’t stretch the frame for the other 98% of them; that pitch still lands wherever it lands, possibly off the top of the frame, the trade a boxplot makes with its own whiskers. A panel whose in-plot legend key floats in the same top-left corner gets a fixed half-foot of extra headroom, so the key and a chart’s own highest pitches stop competing for the same crop.
• kind="heatmap" bins into hexagons instead of squares. A square grid fine enough to resolve a pitch’s own footprint (about a quarter foot across) checkered into isolated single-count cells scattered well outside any real cluster, since that’s finer than the spread between two pitches thrown inches apart; hexagons sized a little wider than a ball pool that spread into one cell instead, and tile without the seams a coarser square grid would show. Empty hexagons are left undrawn (mincnt=1) rather than shaded at the palette’s own floor, so the surface still reads as a figure over the background rather than a wall-to-wall grid.
• The headline/dek pairing above a chart is bolder and better separated: 17px bold instead of 15px semibold for the headline, with more room before the 11.5px dek beneath it. Both are now positioned a fixed distance from the figure’s own edges in inches, rather than at a fixed fraction of it, so the spacing holds regardless of how tall a cropped frame ends up.
• The tunnel plot’s frame was quietly narrower than it needed to be: fixed at the same figure size plot_zone() used, whose box didn’t match TUNNEL_Z_RANGE’s taller aspect ratio, so set_aspect("equal") was letterboxing roughly half a foot of blank margin onto each side to keep the scale honest. It now sizes its own box to the frame it actually draws, the same geometry plot_zone() uses, so the full width is real plot instead of margin.

Added

• plot_zone_panels(), for a comparison assembled by hand rather than one collection’s own facets — a before/after, two non-adjacent windows, a grid of pitch type by outing. plot_zone() gives a single chart a headline, dek and source line; split_by gives the same to a figure faceted from one collection’s column. Every multi-panel figure built outside those two paths used to lose both: a bare plt.subplots() and fig.suptitle(), no dek, no source line, and no frame shared between panels. plot_zone_panels() takes (label, collection) pairs instead of a column’s distinct values, wraps into a grid with ncols, and draws every panel against the same data-driven vertical range the rest of this release adds. scripts/make_docs_images.py and the two example scripts that used to hand-roll this are rebuilt on it.

Added

• Pitch tunneling, measured rather than eyeballed: PitchCollection.tunnels() reports how far apart two pitches were where the hitter had to commit and how far apart they finished, and plot_tunnel() draws the pair as he’d see it, with an open marker at the commit point and a filled one at the plate. Pairs each pitch with the one that followed it in the same at-bat by default, since tunneling is a question about sequence; consecutive=False and same_type=True open that up. Handed more than two pitches, plot_tunnel() ranks them and draws the best pair, which makes it a way to find an outing’s best sequence rather than only to render one already chosen.
• mound/trajectory.py and the nine fields behind it. Savant’s /gf feed has carried Statcast’s full trajectory fit all along — release position, release velocity and a constant acceleration carrying drag and Magnus together — and Mound was parsing four of the nine and dropping the rest. They describe the whole flight as a quadratic in each axis, so Trajectory answers for any point between the hand and the plate in closed form: position_at_distance(), time_before_plate(), path(). Exports gain the columns; Pitch.trajectory() is the accessor.
• The commit point is configurable and deliberately so. It defaults to 23.8 feet, which is the distance Baseball Prospectus’s tunnel work settled on and therefore the comparable one, but a fixed distance is 162 ms of a 100 mph four-seamer and 200 ms of an 80 mph curveball, and a swing decision is a reaction rather than a place — so commit_time fixes it at a number of seconds before each pitch’s own arrival instead. Neither is the right answer for every question, which is why both are there and the README says what each is for.
• docs/images/roki_semien_tunnel.png, and its recipe in scripts/make_docs_images.py: Sasaki’s forkball and four-seamer to Marcus Semien on July 24, three inches apart where Semien had to commit and seventeen by the time they arrived.
• The README is published at moundcli.com/docs rather than only linked to on GitHub, rendered from the same file the package ships to PyPI so the page and the installed version can’t disagree about what a command does. The site’s “Read the docs” button used to hand the reader to GitHub’s UI one click after the landing page, and none of the prose that would answer a search for “chase rate vs whiff rate” lived on the site’s own domain.
• Inline code in the site’s hand-written prose gets the same tinted treatment the rendered markdown already had, through a shared .code-inline class. A flag or command name set in bare green monospace at body size was hard to pick out of a sentence, which is the whole job of marking it as code.
• Repo-relative links in synced markdown now resolve to site routes where the site hosts the file: README.md to /docs, CHANGELOG.md to /changelog, and docs/examples/<name>.md to /examples/<name>, each keeping its anchor. So the Díaz walkthrough’s link into the arsenal section lands on /docs#whiff-rate-chase-rate-and-pitch-metrics instead of leaving the site mid-sentence. Anything the site doesn’t host still falls back to GitHub, which is what ROADMAP.md does, deliberately — a roadmap on a product site reads as a promise.
• .github/workflows/publish.yml uploads to PyPI when a GitHub release is published, and PUBLISH.md writes down the whole process. The upload used to happen on whichever machine ran the script, which meant a release depended on a token in someone’s shell and on that someone remembering to run it. The workflow builds its own copy from the tagged tree rather than trusting the files attached to the release, and refuses to upload if the version it built doesn’t match the tag — shipping the wrong number under the right tag is unfixable, since PyPI won’t let a version be replaced.

Changed

• The version is one literal, in mound/__init__.py. pyproject.toml declares dynamic = ["version"] and hatchling reads it from there, so the number mound --version prints and the number on the wheel can’t disagree — which two hand-maintained copies eventually do, and only visibly after the release is spent.
• scripts/publish.sh cuts the release instead of performing it: it bumps the version, opens the dated changelog section, runs ruff and pytest, commits, tags, builds and creates the GitHub release with notes read out of the changelog verbatim. It no longer uploads anything. Doing it locally and from the release is how you get a Publish run failing with “400 File already exists” against a package that already went out, so the upload lives in exactly one place.

Fixed

• Savant measures plate_x/plate_z at the middle of home plate, 17/24 feet from the point, not at its front edge. Fitting the plane empirically against Savant’s own numbers across 22,482 cached pitches lands there exactly and nowhere else; the front edge, which is the intuitive guess, is off by a tenth of an inch horizontally and three tenths vertically — small enough to look like rounding and large enough to move a pitch between zones. Nothing shipped depended on it before now, since nothing reconstructed a trajectory, but every number in tunnels() does. It’s the same 17/24 that mound.zone already uses for half the plate’s width, which is a coincidence of the plate being as deep as it is wide.

Added

• mound outing, one start broken down end to end: the shape of the outing, how each plate appearance ended, and the arsenal table — in one command instead of running mix, results and arsenal against the same --game three times. Selection is built for the question people actually ask — bare mound outing "Roki Sasaki" is the morning-after view, --date a particular day, --season his last outing of a year, --game an exact game_pk. There’s deliberately no --last, since an outing is one game; a window of starts is still what mix/arsenal/zone are for. --out adds the zone chart, opt-in rather than writing a file nobody asked for. Section headings are bold, which click strips when the output isn’t a terminal, so a piped report stays plain text.
• Two honest limits in that report worth knowing. It counts innings he appeared in, not innings pitched: a reliever who enters with two outs still shows up in that inning, and nothing in the feed counts outs, so there’s no way to render a box-score line without inventing one. And everything but --game goes through the Stats API game log, which is where the opponent in the headline comes from — a bare game_pk can’t be looked up there without guessing which season’s log to read, so --game reports the date and the game and stops. A doubleheader is the one case where --date doesn’t name an outing, and it says so, listing both game_pk values rather than silently taking the nightcap.
• plate_appearances() on PitchCollection: how the at-bats in a collection ended, counted by result. Read only from the pitch each at-bat ended on, since Savant stamps at_bat_result on every pitch of the at-bat and counting every row scores a five-pitch strikeout five times. Inherits ends_at_bat’s edges, including the one that makes a filtered count a narrower question than it looks — filter(pitch_type="splitter").plate_appearances() counts the at-bats that ended on a splitter, not every at-bat containing one.
• first_pitch_strike_rate() on PitchCollection, with the same by_pitch_type option as the other rates. Split out from strike_rate() because the count a pitcher spends the rest of an at-bat working from is mostly settled by pitch one, and a starter who can’t get it over runs up a pitch count regardless of how his stuff looks.

Changed

• mound arsenal/mound faced-arsenal gained usage rate and strike rate, and lost about 45 columns of width. The old table ran past 130 characters and wrapped in a normal terminal, most of it spent on release_extension and on spelling out horizontal_break/induced_vertical_break; those two shorten to hb/ivb and extension is gone, since it barely moves between one pitcher’s own pitches. It’s still one pitch_metrics() call away in Python, along with the full-length names. What the width bought is the two columns the table was missing: usage rate, and the strike rate that previously meant running mound results against the same query to see. The whole repertoire now fits one screen at 85 characters.
• A missing number in that table reads as - rather than NaN, which looked like a bug where it was actually information: a pitch type that never left the zone has no chase rate, and one nobody swung at has no whiff rate, neither of which is a rate of zero. Spin also loses its decimal, which was precision the tracking doesn’t have.
• Building outing is what turned this up. Composing the existing commands verbatim gave two tables keyed on the same pitch types, each spending a line on the pitch_type index name and repeating the pitch counts, to add two columns each — which was a sign the arsenal table itself was carrying the wrong columns, not that the report needed a third one of its own.
• Every published copy of the old arsenal table was re-run against the cached games rather than hand-edited, so the numbers stay real: the README’s mound arsenal "Roki Sasaki" --game 825051, both tables in the Díaz walkthrough, and moundcli.com’s landing-page hero, which is that same command. Each number the two versions share came back identical — velocity, whiff rate and chase rate to the tenth — so what moved is which columns are there, not the data behind them. The styled arsenal table beside the hero drops spin’s decimal to match the terminal’s.
• The README gained a “One outing” section, and its arsenal section now says what the short column names stand for and where the full-length ones still live. The mound arsenal sample no longer needs a sentence explaining NaN.
• The CLI’s player lookup collapsed into one helper per side, replacing four copies of the same two-branch try/except around Pitcher(name)/Batter(name). outing is what forced it: unlike every other command it needs the resolved Pitcher itself, not just the collection that comes out of it.
• moundcli.com covers outing now, with the report as printed and the four ways to name a start. The landing page also stopped advertising “eight commands and one Python object,” which was true up to 0.10.0 and is now ten commands, seven of them mirrored on the batter’s side, over Pitcher and Batter.

Fixed

• Three README examples paired Shohei Ohtani with Roki Sasaki, who are teammates — a matchup that cannot appear in any feed. Replaced with Logan Henderson, checked against the data rather than guessed: he faced Ohtani in both the last-5 and last-8 windows the examples use, so --last 5 and --last 8 each return pitches as written. Every other pairing in the docs was checked and is cross-team.
• The site’s pitch-type colors were the palette from before 0.9.0 reassigned them for perceptual distance, so a chart screenshot and the page around it disagreed on what a pitch looks like — the splitter worst of all, drawn in what is now the sinker’s clay orange while every committed figure and every terminal render it dark green. The forkball sat next to it, and the cutter, sweeper, curveball and changeup were each a shade off. All nine now come from PITCH_TYPE_COLORS unchanged, and the file that holds them says so.
• The Open Graph card had the same drift. Its illustrated pitches below the zone are changeups rather than splitters, because the splitter’s green is 2.2:1 against that card’s dark background and disappears — the same shape of pitch in a color that survives being printed on it.
• Three boolean filters no longer emit a pandas FutureWarning about downcasting an object column, which is the shape those columns take whenever a feed leaves the flag unset on a pitch: plate_appearances() and the mound pitches result column both on ends_at_bat, and whiff_rate() on is_swing. All three now cast to the nullable boolean dtype before filling, the idiom chase_rate() already used for in_zone. The masks are identical for every shape those columns arrive in — real bools, objects carrying None, all-missing and empty — so this moves no number in any table.

Added

• Pitcher.games()/Batter.games() and the mound games/mound faced-games CLI commands: a plain list of a player’s games — date, opponent, home/away, game_pk — for the last N appearances or a whole season. Reads only the Stats API’s game log, which Mound already fetched internally to decide which games to pull pitches from; exposing it directly answers “which games” without paying for a Baseball Savant round trip per game just to find out. Returns a DataFrame, so the game_pk column feeds straight into pitches(game=...) once you’ve picked which games are worth the fetch.
• --season on pitches, faced, mix, faced-mix, results, faced-results, arsenal, faced-arsenal, zone, faced-zone, video and faced-video. Pitcher.pitches(season=...)/Batter.pitches(season=...) have taken a season since the beginning, but the CLI only exposed --last/--since/--until until games/faced-games added --season — now every command that selects games shares the same four options.

Added

• mound faced, the CLI counterpart to Batter: the pitches a hitter faced, from every arm he saw, or narrowed to one matchup with --pitcher. Previously the batter side was only reachable in Python (Batter(...).pitches()) or indirectly through --batter on the pitcher-side commands. Shares the same table as mound pitches, which now names the batter and (when it varies) a pitcher column, rather than assuming every query is pitcher-first.
• mound faced-mix, faced-results, faced-arsenal, faced-zone and faced-video, closing out the rest of the CLI’s batter-side surface: the same pitch mix, strike-rate breakdown, velocity/spin/whiff/chase table, zone plot and video download that the pitcher-side commands already had, each built on a batter’s own game log instead of a pitcher’s starts.
• docs/images/skubal_arsenal_zone.png, a five-pitch-type start, so the README can show the pitch-type palette. Every other committed figure is a single pitch type or a density surface, none of which use it.
• A second worked example in docs/examples/ohtani-spin-chase.md, with examples/shohei_spin_chase.py as its runnable companion: testing a hunch from watching games (“he’s striking out a lot lately by chasing sliders and changeups away”) against 40 games from the hitter’s side. Covers what the pitcher-side walkthrough can’t — Batter and its per-game fetch cost, one row per plate appearance via ends_at_bat, the pitch each strikeout ended on, and chase rate split by pitch family and side of the plate. Also settles which sign of plate_x is the outer half, which nothing in the feed states: Ohtani’s own hit batters put a lefty’s body at +2.1 feet and a righty’s at -1.3 to -2.9, so for a left-handed hitter away is negative, and Statcast zones 11 and 13.
• examples/shohei_strikeout_supercut.py, which turns the walkthrough’s last step into one file: a clip per strikeout, each labeled with the pitcher, the pitch type in its chart color, the velocity, the zone and how the at-bat ended, joined in order behind a title card. Downloads are Mound’s; the stitching is ffmpeg, which has to be on PATH. Savant’s clips share a resolution but not a frame rate, so each one is re-encoded before being concatenated. --spin-away cuts it down to the strikeouts on breaking or offspeed pitches off the plate.

Changed

• Reassigned the pitch-type colors. The old palette grouped them by family — fastballs in blues, breaking balls in reds and purples — which put the closest colors on the pitches most likely to share a chart: across 620 pitcher-games of cached feeds, a four-seamer and a sinker appear together in 286 of them and were 9.8 apart in CIEDE2000, where about 20 is the floor for telling two scatter dots apart; a four-seamer and a cutter, together in 205, were 8.8 apart. The new assignment maximizes distance between pairs that actually co-occur, weighted by frequency and counting simulated red-green color blindness at half weight, which cuts the total penalty by about two thirds. The sinker moves to orange, the cutter to a dark blue, the splitter and forkball into the teals alongside the changeup (they share a chart with it twice in 620 games), the sweeper to a dark red and the curveball to a darker purple. The two slider variants stay close on purpose, since a sweeper is a slider. Nothing about color_by or the single-pitch-type house color changes.

Fixed

• Broadcast clips whose URL ends in base64 = padding could never be downloaded. Savant writes the padding into the page as the HTML entity =, and resolve_video_url returned the attribute verbatim, so the request went out with the entity still in it and came back 404. Found while pulling clips for the Ohtani walkthrough, where two of eight pitches failed. The URL is now unescaped before it’s fetched.

Added

• ends_at_bat on every pitch, marking the one each at-bat ended on. Savant repeats at_bat_result and description on all of an at-bat’s pitches, so neither says which pitch produced them and a five-pitch strikeout reads as five strikeouts. Nothing in the feed answers this — result_code is pitch-level, and a walk’s fourth ball looks like its first three — so it’s derived from the highest pitch number in each at-bat, checked against 10,876 at-bats of cached feeds where pitch numbers run 1..n with no earlier pitch ending anything. Exports gain a column.

• mound pitches --ends-at-bat and filter(ends_at_bat=True), one row per plate appearance. The flag is computed when the feed is parsed rather than from whatever a filter left behind, so --pitch changeup --ends-at-bat returns the changeups that ended at-bats, not each at-bat’s last changeup. Two edges: an at-bat still being pitched marks nothing, since a pitcher mid-count hasn’t ended anything, and an at-bat ended by a throw rather than a pitch (a runner caught stealing for the third out, about 1 in 500) still marks its last pitch, which is where the record ends even if that pitch didn’t decide it.

• zone on every pitch: Statcast’s numbered zones as Baseball Savant draws them, 1-9 across the strike zone and 11-14 for the quadrants outside it, with no 10. Derived from the pitch’s own coordinates rather than read from the feed’s zone field, the same way in_zone already was, so the two can’t drift apart. Matching Savant exactly takes three details: the grid is cut from the zone grown by one ball radius (a pitch an inch above sz_top is zone 1, not 11), the thirds come from that grown rectangle rather than the strike zone proper, and membership still uses the sphere overlap, whose corners are round, so a pitch clipping a corner diagonally reads as outside. Agrees with Savant’s own zone on all 42,538 pitches in a local cache of 146 games.

• mound pitches --zone 5 / --zone 11,12,13,14, and filter(zone=...) taking a number or a list.

• plot_zone(kind="zones") (mound zone --kind zones), a heatmap binned into Statcast’s numbered zones instead of a 25x25 histogram, with each cell labeled by its count and its number. Only the nine in-zone cells are shaded: 11-14 run out to wherever a pitch landed, so they gather more pitches than any single cell almost by definition, and shading them on the same ramp darkened the border and flattened the nine cells the chart is about. They print their counts instead. In place of the strike zone box the other kinds draw, the grid carries a heavy line one ball radius outside it, since that is the edge the numbering is actually cut on. Counts come from each pitch’s own zone, measured against the batter it was thrown to, while the cells are drawn from the panel’s average zone — the same averaging every strike zone in these plots already does, so a pitch can be counted in a cell its own dot wouldn’t sit in.

• plot_zone(grid=True) (mound zone --grid), drawing the 3x3 grid under a scatter, heatmap or KDE surface, for reading a plot against the zones --zone would return. Clipped to the drawn strike zone rather than to the grid’s own outer edge, which sits a ball radius further out and would read as a second, wrong zone box.

Fixed

• The half-plate constant was rounded to 0.708 feet, five hundredths of an inch short of the true 17/24. Small enough to look harmless and large enough to matter: it put 4 pitches per 42,538 in the wrong zone and disagreed with Savant’s own isInZone on 2, so the README’s “zero mismatches” claim was off by two. Now exact, and in_zone matches Savant on all 42,538. Drawn strike zones move by half a thousandth of a foot — a fraction of a pixel, visible in the committed images only as antialiasing along the box edge.

Changed

• The mound pitches table gained at-bat, count, batter and zone columns, and prints at_bat_result only on the pitch that produced it. The inning alone can’t separate three at-bats in the same inning, which is what made a repeated result confusing to read in the first place; the count also shows where a pitch-type filter is hiding pitches, since a jump from 0-1 to 1-2 means a slider went by in between.
• The same table pays for that width by stating whatever doesn’t vary once, in a headline above the rows, instead of down a column: the pitcher, the date of a single outing, the hitter in a matchup, the type behind --pitch splitter. Pitch types show as Statcast codes in the rows and by name in the headline, zone replaces in_zone because 1-9 versus 11-14 says the same thing in two characters and adds the location, and a truncated table now says Showing 20 of 77 pitch(es). rather than reporting a total it isn’t showing. What counts as unvarying is read from the whole collection, so a --limit landing mid-at-bat can’t promote a column that actually differs.

Added

• mound zone --color-by {pitch_type,stand,none}, exposing a plot_zone() argument the CLI couldn’t reach before, and a color lookup for batter handedness so color_by="stand" actually renders lefties and righties as two colors instead of one undifferentiated gray. Handedness draws the house green against the splitter’s orange: an earthier clay looked better but collapsed into nearly the same olive as the green under red-green color blindness, where this pair keeps a 20-point lightness gap. It’s the one-panel counterpart to --split-by stand: color holds both sides against the same axes, splitting gives each its own strike zone.
• The legend key labels handedness as “vs LHB”/“vs RHB”, matching what split_by already titles its panels, and orders the two the same way (left, then right). Pitch types still read most-common-first, so the key doubles as a pitch mix.
• scripts/make_docs_images.py regenerates every plot committed to docs/images/, each pinned to the window it was first made from rather than a relative one like last=4, so a styling change can be re-rendered without the figures quietly sliding forward to last night’s start and contradicting the prose around them.

Changed

• Density surfaces (plot_zone(kind="heatmap") and kind="kde") run yellow-to-green instead of cream-to-red, so a chart embedded on moundcli.com no longer fights the page around it. The ramp is ColorBrewer’s 7-class YlGn, whose stops already step down in even increments of perceived lightness, plus one darker stop: YlGn ends at a medium-dark green that left the hottest cell short of the punch a peak wants, and the site’s own darkest green supplies it. The yellow low end earns its place by keeping a one-pitch bin visible against the plot background, which a green that faint wouldn’t be.
• A single-color scatter (color_by=None, or a value the active palette doesn’t recognize) draws in green rather than neutral gray, for the same reason.
• A scatter showing only one pitch type now draws in that house green too, instead of that pitch’s own color. A color that separates nothing isn’t worth spending, and the headline already names the pitch; this is the same reasoning behind drawing no legend key for a single group. The trade is that a splitter is no longer orange in every chart it appears in, only in the ones where the color tells it apart from something. The rule is decided against the whole figure rather than each panel, so a split_by pair can’t end up keyed by color on one side and not the other.

Fixed

• mound zone no longer dumps a traceback for an unknown --kind, --split-by or --color-by value, matching how the command already handles an unwritable --out path.

Added

• A worked example in docs/examples/diaz-blown-saves.md, with examples/diaz_blown_saves.py as its runnable companion: fact-checking a closer’s postgame explanation (“I was throwing my fastball right in the middle”) against his pitch locations. Covers the path from a name to an answer — finding a pitcher’s recent games and their IDs, exporting every pitch, breaking down mix and arsenal by game, defining “the middle” against each batter’s own zone, and pulling the video for the pitches that got hit.

Fixed

• Games in progress are no longer cached. Caching keys on game_pk alone, on the premise that a finished game’s data never changes — but a game cached while it was still being played kept whatever partial feed existed at that moment, permanently, and a later query would silently come up short (a reliever who pitched the ninth simply missing from a game he appeared in). fetch_game_feed() now writes only feeds whose Savant game status reads final, and ignores a cached feed that doesn’t, so entries already poisoned by an earlier version repair themselves on the next run once the game is over. A game still in progress re-fetches every time.

Changed

• plot_zone(kind="heatmap") no longer draws a “Fewer”/“More” colorbar, matching kind="kde", which dropped its own in 0.4.0. Darker already reads as more pitches, and the vertical bar shrank the plot area enough to push the strike zone and home plate off-center relative to every other plot kind — a visible misalignment when heatmap and scatter panels sit side by side, and doubled in a --split-by pair.

Added

• Batter filtering, for matchup views from the pitcher’s side: PitchCollection.filter(batter=...) and Pitcher.pitches(batter=...), plus a --batter flag on mound pitches, mix, results, arsenal, zone and video. Takes a name or an MLB player ID, or a list mixing the two; names match any part of the name Savant reports, ignoring case and accents, so --batter perdomo is enough. filter(pitcher=...) is the same thing for the other side, useful once a collection spans more than one arm.
• Batter, the mirror image of Pitcher: the pitches a hitter faced, discovered from his own game log and pulled from every pitcher who faced him in those games. Batter("Geraldo Perdomo").pitches(last=5, pitcher="Roki Sasaki") and Pitcher("Roki Sasaki").pitches(last=5, batter="Geraldo Perdomo") return the same matchup from either side; the pitcher’s side fetches far fewer games, since a starter appears in a fraction of the games a hitter plays.
• PitchCollection.chase_rate() (with the same by_pitch_type option as swing_rate()/whiff_rate()): swings divided by pitches outside the zone, so a chase pitch’s real job shows up as its own number. Location comes from in_zone geometry rather than the is_strike ruling, and pitches with no plate coordinates drop out of the denominator instead of counting as strikes. mound arsenal gains a chase_rate column alongside whiff_rate.
• plot_zone() labels matchups: a plot narrowed to one hitter notes “vs. ” in its dek, and a batter-side collection headlines as “Pitch locations to ” rather than crediting the hitter with throwing them.

Changed

• mound.statsapi.pitching_game_log()/pitching_game_log_seasons() are now game_log()/game_log_seasons() with a group argument ("pitching" or "hitting"), since the same Stats API endpoint serves both sides of the ball. Internal client functions, not part of the documented Pitcher/PitchCollection API.

Fixed

• mound zone --out now creates missing parent directories before saving, matching how --export and video --out-dir already behave, instead of raising a raw FileNotFoundError.
• CLI commands no longer dump a full traceback (including every local variable, e.g. an entire pitch DataFrame) for an unwritable --out/--out-dir/--export path; zone, pitches --export, video and video-id now fail with one clear line instead.

Added

• mound --version prints the installed CLI’s version and exits, so you can check what you have installed without dropping into Python.

Changed

• Consolidated the CLI’s repeated --last/--since/--until/--game/--pitch/--stand/--at-bat/--pitch-number/--cache/--cache-dir option declarations, previously duplicated across pitches, mix, results, arsenal, zone and video, into shared Annotated type aliases in mound/cli.py — each flag’s name and help text now live in one place instead of five.
• _get_pitches()’s internal helper is now keyword-only, closing off the possibility of a silent argument-order mistake at a call site.
• video --limit now goes through a new PitchCollection.limit() method instead of constructing a PitchCollection directly from another collection’s internals.
• _fail() is now typed -> NoReturn, making explicit (for both readers and type checkers) that it always exits rather than returning.

Added

• Pitch gains spin_rate, release_extension, release_pos_x, release_pos_z, horizontal_break and induced_vertical_break, parsed from fields Savant’s /gf feed already returns but Mound wasn’t yet surfacing. All default to None rather than raising when a pitch predates or otherwise lacks tracking coverage for them. Flow through automatically to to_frame()/CSV/JSON/Parquet export; no new fetch required.
• Pitch gains is_swing and is_whiff, derived from pitch_call the same way is_strike already is. PitchCollection.swing_rate() and .whiff_rate() (each with a by_pitch_type option, matching strike_rate()) calculate the resulting percentages — whiff rate is misses divided by swings, matching Baseball Savant’s own convention, not misses divided by every pitch thrown.
• PitchCollection.pitch_metrics() averages velocity, spin rate and movement by pitch type, using the fields above. Combined with whiff_rate() in the new mound arsenal CLI command, for questions like how nasty a pitch looked in one start (--game) versus across a season (--last/--since).

Changed

• Adding is_swing/is_whiff shifts at_bat_result and description one position later in Pitch’s field order; only matters for code constructing a Pitch positionally rather than by keyword.

Added

• mound video --limit N caps how many clips a run downloads (e.g. --limit 1 for a single clip), rather than always fetching every pitch matching the other filters.
• PitchCollection.filter() / Pitcher.pitches() accept at_bat_number and pitch_number, exposed on the CLI as --at-bat/--pitch-number, to narrow down to one specific at-bat or one exact pitch (pair with game/--game, since an at-bat number is only unique within a single game). Useful on its own, and with mound video to download the clip for one particular pitch.
• mound.video.download_video_by_id() and the mound video-id <pitch_id> CLI command download a broadcast clip directly from a known pitch_id (e.g. one saved from an earlier export), with no pitcher/game lookup needed first.

Changed

• Refined kind="kde"/kind="heatmap" styling in plot_zone(): both now share a single on-brand warm gradient instead of matplotlib’s generic YlOrRd. KDE surfaces default to a fixed bandwidth (rather than scipy’s sample-size-dependent Scott’s rule, which oversmoothed small pitch samples into one shapeless blob) and a gamma-corrected color scale that keeps the true “hot zone” distinct from its faint tail. The KDE colorbar is removed — its density values are an arbitrary scale, not a pitch count, so a “Fewer”/“More” legend was either meaningless or redundant with what the color already shows.
• plot_zone()’s y-axis now marks its topmost tick with a foot mark (e.g. 4′), so the plate-height scale reads in feet without needing a full axis label.

Added

• Optional local file cache for Baseball Savant game-feed responses, keyed by game_pk. Enable with Pitcher.pitches(cache=True) / --cache (or a custom directory via cache="/some/dir" / --cache-dir, defaulting to ~/.cache/mound). Because a finished game’s data never changes, repeat queries automatically fetch only games not already cached, with no separate “update” step needed.
• plot_zone(kind="kde") renders a kernel density estimate instead of the plain 2D-histogram heatmap, for a smoother density surface on larger pitch samples. Requires the new optional scipy dependency (pip install "mound[viz]"); bandwidth is configurable via bw_method, exposed on the CLI as mound zone --kind kde --bw-method.
• Pitch.download_video() / PitchCollection.download_videos() and the mound video CLI command download a pitch’s Baseball Savant broadcast clip, resolved from its pitch_id. Captures the clip page’s default embedded angle only (in practice, the home broadcast feed).

Added

• batter_stand field on Pitch, populated from Statcast’s stand. Filter with .filter(stand="L") / Pitcher.pitches(stand=...) (accepts "L"/"left"/"LHB", "R"/"right"/"RHB", case-insensitive) or the CLI’s --stand option.
• plot_zone(split_by="stand") facets a zone chart into side-by-side vs-LHB/vs-RHB panels (each with its own strike zone and pitch count), exposed via the CLI’s --split-by option on mound zone.

Changed

• README documents PyPI installation, since mound is now published there.

Added

• Initial prototype: resolve a pitcher by name or MLB ID, retrieve Statcast pitch-level data (filterable by game, date range, last-N-starts or pitch type), calculate pitch mix and strike rate, plot pitch locations against the strike zone, and export to CSV/JSON/Parquet.
• Pitcher/PitchCollection Python API and a mound CLI (search, pitches, mix, results, zone) sharing the same underlying implementation.
• Data sourced directly from the MLB Stats API and Baseball Savant’s /gf endpoint, no pybaseball dependency.
• Pytest suite covering player resolution, game-log/pitch parsing, filtering, analysis and export, run against mocked HTTP fixtures.
• README, ROADMAP and an end-to-end example using Roki Sasaki’s splitter.