ansi_escape_codes

The CHANGELOG has carried a 4.0.1 section since the meta floor came down;
this is the pubspec catching up to it, on the owner's word to publish. The
release carries one change and no API movement: the constraint that made
Flutter 3.27 unable to resolve a package built for Dart 3.6.

Co-Authored-By: Claude Opus 5 (1M context) [email protected]

Changed:

• The meta constraint reaches down to 1.15.0 again, so that the package resolves everywhere its own SDK constraint says it does. Flutter pins meta to an exact version out of its SDK, and Flutter 3.27 — the line carrying Dart 3.6, this package's floor — pins 1.15.0; against ^1.16.0 that was a version solving failure, which made the real Flutter floor 3.29. Nothing in the API or its behaviour changes: the annotations used are @immutable, @internal and @visibleForTesting, all older than 1.15.0.

The owner asked for everything except publication to move into their list,
so the member-signature gate, the cheap backward walk and the gate against a
constant multiplier now sit in docs/backlog.md. Each entry carries the facts
an agent needs to start — the missing transitTo in the snapshot, the 1.7ms
against 178ms, what complexity_guard does not measure — rather than a
pointer into the handoff, which is rewritten every wave and would decay.

The handoff drops its own copies of all three. An item kept in two files
drifts, and the copy nobody edits is the one that ends up lying; what stays
there is the state of the tree and the findings the owner did not take.

Writing into the backlog is otherwise forbidden to agents by AGENTS.md. This
edit is the owner's instruction, not an agent's initiative.

Co-Authored-By: Claude Opus 5 (1M context) [email protected]
Claude-Session: https://claude.ai/code/session_01Uox8hLVY5Ho8K6fofJd9g7

Added:

• State, Style, Stack and Styles now model the standard primary and nine alternative fonts, fraktur, curly/dotted/dashed underline, proportional spacing and five ideogram renditions. Existing enum case names and style APIs remain source-compatible.
• insertBefore and insertAfter on Parser and StackedParser, with the ansiInsertBefore and ansiInsertAfter string extensions. Text put into a styled string takes the style of the place it lands in and gives it back, so the rest of the string is left as it was: the style it opened of its own is closed after it, and the hyperlink it landed inside — links do not nest, and the inserted text may have superseded it with one of its own — is opened again behind it. Text inserted outside every link is closed off instead, so what follows it stays outside whatever the insertion pointed at.
• The hyperlink a string has open, read back the way the style is: linkAt and finalLink on Parser and StackedParser, and link on Match, beside its state, for a walk over the matches. A link is state but not style — it carries no rendition, and SGR 0 does not close it — so it travels on a channel of its own and is answered on its own: linkAt(pos) for the link the character at pos sits inside, finalLink for what the string leaves open.
• The independent control functions, ESC Fs: constants for all ten of them, ControlFunctionsEscFs, EscCommon — the entity a switch over them matches — and resetTerminal for RIS.
• DEL, which the C0 set was missing.
• useAlternateScreen and useMainScreen, the screen a full-screen program draws on.
• Types for the control sequences that carry something worth reading: CursorUp, CursorDown, CursorRight, CursorLeft, CursorNextLine, CursorPrevLine, CursorHPos, ScrollUp and ScrollDown carry n; CursorPos and CursorHVPos carry row and col; EraseInPage and EraseInLine carry an ErasePart; ShowCursor, HideCursor, UseAlternateScreen and UseMainScreen stand for the four private modes this package writes itself.
• ControlString, what the five strings of the standard have in common, with Dcs, Sos, Pm and Apc beside the Osc that was there already, and terminated to ask whether one of them got the terminator that ends it. Osc is a ControlString now rather than a direct child of EscapeCode, which leaves a switch matching it by name as it was, and the four new entities carry UnrecognizedEscapeCode like the other codes this package keeps without reading.
• ansiHasUnderlineColor and ansiRemoveUnderlineColor: the extensions knew the foreground and the background but not the underline colour.
• ansiRemoveControlCodes, the other half of ansiHasControlCodes: the C0 set, DEL and the eight-bit C1 taken out, with exclude for the ones to keep — a text that is to stay in lines keeps its LF. ESC is one of them, so the escape codes come out first or their bodies are left behind as text. exclude names members of ControlFunctionsC0 and so cannot spare an eight-bit C1: this package gives those no names, on the reasoning below.
• Colors implements Comparable, so a list of them can be sorted.
• ansiShowControlFunctions and ansiOptimizeControlFunctions, which used to live in the tests though the README and the examples took them for public.
• The control function types the API returns — ControlFunctionsSGR, ControlSequencesFunctions and the rest — are exported from the main entry point.
• ansi_escape_codes.dart brings the String extensions as well, so it is one import for all the string work: the ready-to-use strings, the styles, the parser, the state, the control function tables and the extensions. Two things stand outside it. ansi.dart always did — the ready-to-use strings are built from its raw byte tables, and neither import brings the other. And utils.dart does, because tabs and currentCursorPos talk to a terminal in person through dart:io, and nothing else in the package touches a platform library: bringing them in through the umbrella would tag the whole package native-only for the sake of two names, and take the web and WebAssembly away from the thousand that run anywhere. A test walks the directives of each entry point and holds that line. extensions.dart and style.dart still bring their smaller namespaces, which is what they are for — not for reaching something the main import lacks.

Verification:

• Entry-point signature closure is now backed by an exact exported-namespace snapshot, and neither is asked anything until lib/ has been swept for analysis errors and warnings: a library that fails to analyse still has an element model, only a smaller one. The generator preflights its eight registered marker zones before writing any generated file.
• Stable CI gates 95.0% hand-written lib/ coverage while retaining the full coverage artifact, and runs a separate warmed complexity guard. The ordinary complexity assertions in test/performance_guards_test.dart remain deterministic and timer-free.
• The complexity guard times the single implementation each scenario side has, checks the work that implementation produced both before and after the timing series, and holds the band against the median of the paired ratios rather than against a ratio of separately sorted medians.

Performance:

• The scanner finds the next escape code by indexOf rather than by the regex engine: text with no escape codes at all is parsed and stripped well over a hundred times faster (about 1 ms down to under 9 µs), a coloured page parses roughly 45-50 % faster across matches, removeAll, optimize and showControlFunctions, and a page that is mostly escape codes is not worse — about a fifth faster rather than a wash. ansiHasEscapeCodes and friends answer a clean string with contains(ESC) outright, without touching a pattern, and lengthWithoutEscapeCodes counts without building the cleaned string — the same walk over the same matches, so what is saved is the copy and not the time: a 5 MB page is measured with megabytes less at the peak — some 8 to some 20, depending on the page — and takes about as long as it did.

• substring and the insert seams keep their place the way stateAt always did, instead of walking from the start each time: slicing a 200-line document through one parser is about three times faster (3.95 ms down to 1.33 ms), and the shape of the cost is linear now, not quadratic — the guard that doubles the input and watches the time saw a cost of ×3.73 where it now sees ×1.65, a linear cost being ×2.

• An escape code is told apart by its second byte instead of four named regex groups, a simple SGR function comes from a cached table instead of being rebuilt, and the matched text is read out of the match once: a simpler, more correct hot path, though its own saving lands inside the numbers above rather than as one of its own — isolated, it measures at noise level once the scanner and slicing fixes are in. One visible side effect: two SgrSimpleFunctions built for the same code used to be two separate objects, and were never == to each other either, since the class defines neither == nor hashCode; now they are the same cached instance, so both identical() and == see them as one and the same.

• A full parse retains the match list once, not twice, and a Text piece cuts its own substring out of the input only the first time something reads it, so a piece nobody reads keeps no copy of its own — stateAt, which never asks a piece for its string, is the concrete beneficiary: a walk that never reads a piece's string leaves tens of megabytes on the table against one that reads every piece, on the benchmark's 5000-line page. A scenario that forces everything to materialize up front, the way prepare does, is unaffected — full materialization was never what this bought. The saving is not free of a cost, though: a Text kept alive past its Parser pins the whole original input in memory for as long as the Text itself lives, whether or not its string is ever read.

• What has been read of a string is kept, instead of being read again by every question asked of it. Call prepare when there are many.

• A control sequence is looked up in a map rather than by walking the list.

• A Stack keeps its histories as frames with a shared tail rather than as lists. Immutable lists had to be copied to be grown — twice, once to build the new one and once to seal it — so a push cost the whole depth so far, and a parse, which keeps every state it passed, kept every one of those copies with it. Text that switches attributes without resetting them, which is what ls --color and most syntax highlighters write, made that quadratic in time and in memory both: 320 kB of it took 15 s and 7.6 GB through StackedParser, where Parser took 51 ms and nothing above the floor. The same string now takes 55 ms, level with Parser, and 2.5 MB of that shape takes 299 ms. A push and a pop are one small allocation each, and every version of a stack shares the whole of its own tail with the versions it came from. Nothing a Stack answers has moved: its histories were only ever asked what was on top and whether they were empty.

• flush on all four printers: it writes out whatever is being held back without ending the line. Printer and StackedPrinter hold a line until a writeln ends it, and until now nothing could make them let go of one --- a write with no writeln behind it was lost for good. SinkPrinter and StackedSinkPrinter hold the tail of a write that stopped in the middle of a sequence, and this says that the write which would have finished it is not coming.

Fixed:

• ansiHasSgr and ansiRemoveSgr disagreed with the parser about the same sequences: '\x1B[1<m'.ansiHasSgr was false while the parser read it as a rendition, and ansiRemoveSgr left behind what ansiRemoveEscapeCodes took out. They read sgrPattern, whose parameter class excluded <, =, > and ? everywhere rather than only in the first place, where alone they make a sequence private use. The pattern now says what the parser says.

• ansiShowEscapeSequences wrote a dangling separator where a control string had no terminator to name: \x1B]0;title showed as [OSC 0;title ]. The separator goes with the name now, so it shows as [OSC 0;title] and a terminated one is unchanged.

• Parser.substring refused a maxLength too large to add to start. Asking for everything from anywhere but the beginning --- substring(1, maxLength: <a very large number>) --- took the sum round through the negatives and came back a RangeError for a slice that was only asking for the rest of the string. A length reaching past the end is the rest of it, and stays so where no sum can hold it.

• An SGR carrying a private byte past the first --- CSI 1 < m, CSI 99 ; < m --- was written to the terminal twice. Only the first byte of a parameter string makes a sequence private use, so these are ordinary CSI ... m whose parameters cannot be read: the parser puts them in the opaque rendition branch and the branch writes them again, while the output also copied their bytes over as they came. Two places were answering "is this a rendition?" --- the parser where it reads the sequence, and a pattern whose parameter class knew nothing of a private byte past the first --- and the answers could differ. Only the parser answers it now.

• A write to SinkPrinter or StackedSinkPrinter that stopped in the middle of a sequence corrupted it. Every write is dressed on its own and the dressing opens with a reset, so that reset landed between the halves and the terminal read it followed by the rest of the sequence as text: \x1B[31m cut anywhere inside it showed [31m and no colour, a hyperlink cut inside its url showed the url, and a surrogate pair cut between its halves came out as two replacement characters. A write is now cut where no sequence is open across it and what is left waits for the write that finishes it, so the same bytes read the same however the writes fall across them. An unterminated control string is part of this: it now waits for the write that goes on with it, where before it went out at once and was ended by the next write's reset --- which made write('a' + title) then write('b') show a b that the same bytes written in one go make part of the title.

• ESC 8 with no ESC 7 in front of it left a non-default defaultStyle behind. DECRC without DECSC clears the rendition, taking the terminal to its own defaults rather than to the printer's, and the printer went on believing its default style was still on --- so the transition for the text after it wrote nothing and that text came out bare. The printer's own Parser read its own output as saying so. The model now goes where the terminal goes and writes the default style back on.

• ESC 7 / ESC 8 lost their saved rendition, hyperlink and opaque SGR at every printer line or sink-write boundary, and a restore with no preceding save incorrectly used the previous chunk's seeded state. All four printers now keep one non-consuming, replaceable cursor save slot for their session; sink prepare rolls a probed slot back with its other carry.

• Style.call read its inner ANSI through Stack, so a selective reset after two setters revealed the earlier inner value instead of returning to the caller's default style. Style wrappers now use the same terminal reset semantics as Printer; the explicitly selected Stack and Stacked* APIs keep their hierarchical pop contract.

• optimize, substring, insertions and all printers silently discarded an ordinary SGR function that State did not model, and collapsed decorated underline to a single line. Known standard functions now have typed state; truly unknown SGR is carried by a private ordered residual channel through cuts, branches and printer resets until a real SGR 0 clears it.

• ESC 7 and ESC 8 carried no style. A terminal saves the rendition along with the cursor and restores both, so ${fgRed}ESC7${fgBlue}ESC8 shows red where the parser said blue — and every question asked after it was answered from the wrong state. The hyperlink travels in that same bundle: a terminal keeps it among the attributes it saves, so what a restore brings back is clickable again exactly where the save was, and a save made where no link was open puts that away as readily — the restore leaves no link behind it, rather than the one the string was started inside.

• An insertion left a hyperlink open. OscLink carries no style, and the closing was worked out from the style alone, so text inserted with an unclosed OSC 8 swallowed everything after it.

• An unfinished escape sequence in the inserted text swallowed the original tail: a truncated OSC consumed it whole, while a truncated CSI took its first byte as the missing final byte. Insertions now preserve the same text model as optimize, substring and the printers, without rewriting completed escape codes.

• cursorDown moved the cursor left, and the cursor functions built sequences out of any number, -1 included.

• runZonedStackedPrinter printed only the first line.

• currentCursorPos left the terminal in raw mode, read the answer as one chunk it does not always arrive in, and took the first CSI that came for the report — an arrow key is a CSI as well, so a key pressed while the terminal answered threw away a report that had in fact arrived. The report is looked for in what comes now, rather than assumed to be at the front of it.

• tabs looped forever on a tab width that never advances, and wrote over the line it was called on.

• Stack threw where a reset had no style to pop, and took a colour it cannot hold in underlineColor.

• faint stood for bold instead of dim.

• An ESC sequence was cut after two characters, and one carrying intermediate bytes was shown without them: ESC ( B and ESC ) B came out alike. A bare ESC at the end of a string was swallowed as text — shown as nothing and counted in the length — and is a code of its own now: it shows as [ESC], and Parser('abc\x1B').length says 3 where it said 4.

• optimize and substring dropped every code that was not SGR, and they, with isClosed, ignored the state the parser started from.

• The printers dropped them as well: prepare('${cursorUp}x') gave back [CSI 0 SGR]x, the cursor movement gone, and a hyperlink came out as the bare text it was written over. What is not a style passes through them untouched now.

• An empty sub-parameter threw the whole sequence away, an RGB colour cancelled the rest of it, a broken colour took the rest with it, and CSI 4:0 m switched the underline on.

• An OSC string ended at the wrong place, or nowhere; a URL carrying ; was refused.

• ESC P, ESC X, ESC ^ and ESC _ — DCS, SOS, PM and APC — were read as finished two-character escape sequences, so the body the standard gives each of them came back as text. Each of the four opens a control string that runs to its ST, the way an OSC does, and one that never got a terminator ends at the next ESC or at the end of the text. What such a string carries — a sixel image, a DECRQSS answer, a termcap reply — is part of the escape code now rather than of the plain text, and is no longer counted in the length: Parser('a\x1BPq#0;2;0;0;0\x1B\\b').length says 2 where it said 13, and neither a slice nor an insertion cuts through the body, an unfinished code ending the string included. ST ends all five; the BEL that ends an OSC is xterm's and not the standard's, and it ends none of the other four, so a DCS whose body happens to end in one is unterminated still — and one left unterminated is held back and given its terminator the way an unterminated OSC is.

• SaveCursor, RestoreCursor and OscLink carried a reset as their text, so all three were equal to one another — an Entity compares by what it is written with — and none of them equalled the same entity read back by the parser.

• ansiHasForeground, ansiRemoveForeground and their background pair only saw a colour that was the whole parameter list. CSI 1;31 SGR, which is the shape this package's own optimize writes, matched nothing, and neither did the colon form CSI 38:5:196 SGR the parser has always read. Removal now keeps the functions standing beside the colour: CSI 1;31 SGR becomes CSI 1 SGR rather than going whole.

• A colour held by a Stack named itself ?256Red, where the same colour from a Parser said fg256Red. A Style written as a constant said the same, and a colour set on one target and then held in another slot answered under the target rather than the slot — bg256Red for the colour of the text. The slot names it now, whichever way the style was built.

• NoStyle passed for the colours of the terminal.

• A private-use sequence was reported as an unknown one.

• Color256.rgb and Color256.gray checked their arguments in an assert only, which release builds leave out.

• The superscript and subscript pair picked its winner the other way round from every other pair.

• DEL counted as a control code but was never shown as one.

• Entities and functions described themselves wrongly in toString.

• On Windows the terminal modes were put back in an order the console refuses — echo first, line mode still off — so currentCursorPos threw and left the terminal raw. Line mode now comes back first, and each mode is restored even when the other throws. Turning them off is guarded the same way now: when a stdin refuses one change, the one already made is undone instead of being left behind.

• substring cut a hyperlink in two and kept neither half right: a slice that began inside one came out unclickable, the opening having been left behind on the other side of the cut, and a slice that ended inside one left it open, so everything printed after the slice was clickable on the slice's URL. A slice is self-contained now, the way it always was in the style: one that began inside a link opens that link again in front of its first piece of text, and close: true closes at the end what the slice has open. With close: false the link is left open, as the style is. Cutting a document into lines this way gives lines that are each clickable on their own.

  The opening is written again in the bytes it came in, parameters and all: a link opened BEL-terminated stays BEL-terminated, and an id= — which is what OSC 8 gives for a link a line break cuts in two — travels with it. The close written is OSC 8;; ST whatever form the opening took; terminals take either.

• Once a link passed through them at all, the printers had the gap the slice had: a printed line that opened a hyperlink left it open, and everything printed after was part of it. A line now closes the link it leaves open, and the line after opens it again — in the bytes it was opened with, as the slice does it — so a link a line break falls inside of goes on being one link, which is what the id= of OSC 8 is for. SinkPrinter and StackedSinkPrinter take a write at a time and one line may be composed of several, so there an open link is carried across the writes and closed where the line really ends — at a writeln, or at a '\n' in what is written. A styled call goes through a printer and changed with them: Styles.red('…') now closes a link its text left open, and in a multi-line string the link reaches the end of the text instead of ending with the first line of it.

• optimize left a hyperlink open where substring closed one. With close: true it now ends the string outside every link as well as in the default style: a string that opened a link and never closed it comes back closed, so that what is printed after it is not clickable. With close: false both are left as the string leaves them.

• Copying a control string out of the string it was read from could swallow the text behind it. One that never got its terminator — an OSC, a DCS, an SOS, a PM or an APC, a hyperlink opening no less than a window title or anything else the terminal answers to — runs on to the next ESC or to the end of the text, which is how the parser reads it on purpose, and written again in front of text that had not followed it there, by a slice, by optimize or by a printed line, it read that text as part of the sequence and showed nothing. The terminator it lacks is supplied where text follows it. That is of the codes copied over as they stand: there an escape code following leaves the bytes exactly as they came, the ESC of what stands behind being terminator enough. At the edge of an output that closes — a slice or an optimize with close: true, a printed line — the terminator is written although nothing follows it there, for the reason the hyperlink close is written in the same place: what is printed after must not be read as more of the sequence. With close: false the bytes are left as they came. SinkPrinter and StackedSinkPrinter pay the same debt where the line really ends — at a writeln, or at a '\n' in what is written — and owe nothing at the end of a write the line goes on past. An opening written again for a slice or a line that began inside the link carries its terminator whatever follows it.

• The insertions reached the same mechanism last. insertAfter goes past the codes standing at the seam, and where the string ended inside a sequence that never finished it went past those bytes as well — Parser('aa\x1B]0;title').insertAfter(2, 'X') handed back a string whose plain text was still aa, the X having become part of the window title, and a hyperlink opening swallowed it no differently. A bare ESC turned the insertion into an SOS and a CSI with no final byte into an ECH. Both insertions now stand in front of such a sequence rather than inside it, and the tail is copied on as it came: no byte of the input is invented, which is why no terminator is supplied here as it is for a slice. A finished code ends the run and is passed along with what stands before it — the run stood in front of is the one reaching the text, not everything unfinished in the string.

  A sequence still waiting for the byte that ends it hands the bytes it waits through back as text — the parameters of a truncated CSI are the case worth naming, but a LF, a DEL or a letter outside ASCII breaks off the pattern of a bare ESC and of an ESC on an intermediate byte the same way — and a position among them has no right answer: in front of the sequence is before characters counted in front of it, and where it was asked for is inside the sequence. Both insertions refuse it with an UnfinishedSequenceException, which carries the position asked for and the offset of the sequence the text would have been read as part of. Before this the same position quietly ate what stood there: Parser('aa\x1B[31').insertAfter(3, 'X') answered a string whose plain text was aa1, the 3 having become a parameter. insertBefore was no better, though the backlog had it down as safe everywhere.

  Unfinished codes come in runs, and the seam is in front of a whole run rather than in a gap between two of them — a gap between two of them is the inside of the first. Parser('aa\x1BPpay\x1B[31').insertAfter(2, 'X') answers 'aaX\x1BPpay\x1B[31', where before it answered 'aa\x1BPpayX\x1B[31', whose plain text was aa31: the X had gone into the body of the DCS.

  Where such a run begins behind a piece of text a sequence in front of it is still reading, the place before the run is where that sequence's ending would be written, so the seam has no end to serve and is refused along with everything past it. This takes back answers that used to come: Parser('aa\x1B[31\x1BPpay\x1B').insertAfter(4, 'X') throws where it answered 'aa\x1B[31\x1BPpayX\x1B' — plain text aa31, the X swallowed by the body of the DCS. A code that stands finished between the text and the run gives the run a seam of its own, and that one is served.

• insertBefore and insertAfter could put text between the halves of a surrogate pair and hand back a string that is no longer valid UTF-16. A position inside a pair now shifts to its edge — insertBefore to the front, insertAfter past it. Positions, length and the paddings are UTF-16 code units, as String counts them, and the docs now say so instead of promising what is seen.

• ansiHasSgr and ansiRemoveSgr counted private control sequences ending in m — xterm's modifyOtherKeys, SGR mouse reports — as SGR, and removing styles removed them too. The pattern now takes digits, ; and : only, the way the parser classifies them.

• ansiRemoveForeground and its background and underline siblings ate the parameters after a colour cut short: \x1B[38;2;1;2m lost its bold and dim along with the broken colour. A colour missing arguments now gives up only its introducer and kind, the way the parser reads it — and the same goes for a kind the package does not know. It reads the other direction too: the parameter after a bare 38, 48 or 58 is that colour's kind and goes with it, so \x1B[38;41m loses its 41 and \x1B[38;4:3m its curly underline where both used to be left standing. The ansiHas* answers moved together with the split — \x1B[38;41m no longer has a background.

• Leading zeroes hid a colour from the same functions: removing the colour from \x1B[38;05;196m removed everything but it. Parameters are now read as numbers, as ECMA-48 allows them to be written.

• A style operation with nothing to change built a new object anyway, and a NoStyle asked for a pointless reset came back a Style that writes: NoStyle().resetItalic('x') opened with a reset. Nothing to change now answers itself, as State promised all along.

• NoStyle().transitTo(Style.terminalColors) wrote a reset between two surfaces that are both the terminal's own. A transition between equal surfaces is empty.

• A Printer given defaultStyle: NoStyle() still opened every line with a reset and unwound it at the end. It now imposes nothing: the line goes out as it came, its own codes included — ansiCodesEnabled: false remains the way to take those out.

• The style entry point returned types it could not name: ControlFunctionsSGR and its four control-function siblings were reachable from the entities but undefined to the importer. The five exports are now part of the point, and every entry point carries an exports test. The extensions point had the same gap: ansiRemoveControlCodes takes a Set<ControlFunctionsC0> its own importer could not name, so the enum is now part of the point.

• prepare on SinkPrinter and StackedSinkPrinter coloured the writes that came after it. The piece it is asked about never reaches the sink, and the link open in the output, the link open in the text and the terminator an unterminated control string owes were all put back for that reason — the style the piece ended in was not, so prepare('${bold}asked') left the printer reading the next write as if the bold had been sent. All four carries are put back now.

• The eight-bit forms of the C1 controls — the bytes 0x80 through 0x9F — went through the control-code extensions unseen. ansiHasControlCodes answered false for a string made of them, ansiRemoveControlCodes left them where they stood and ansiShowControlCodes showed nothing, so a string cleaned or spelt out for display still carried bytes a terminal prints as rubbish. All three know them now, and the display writes them as the number of the byte in every style: they have neither an abbreviation nor a Unicode picture of their own, and this package gives them no names, since a name would suggest it reads them as sequences. It does not, and that has not changed: Parser reads 0x9B as text rather than as a CSI, and ansiRemoveEscapeCodes leaves it standing. What is parsed here are decoded Dart strings rather than byte streams, where a genuine eight-bit C1 does not survive UTF-8 decoding and terminals emit the seven-bit ESC [ form anyway. 0xA0 and above are not controls and are untouched.

• substring(close: false) took an attribute off that was meant to survive. A slice left open is written by asking transitTo for the reset half alone — it unwinds what the string took off by the cut and does not put on what belongs to the character after it. But CSI 22 takes bold and dim off together, so transitTo writes it wherever one of the pair goes off and leans on the other half to bring the survivor back: CSI 22;1. That 1 belongs to the reset rather than being a set of its own, and going out without it left \x1B[1;2mAB\x1B[22;1m sliced open at \x1B[22m — a slice standing in neither its own state nor the string's. skipSet leaves it in place now. The four other pairs are unaffected: 24, 25, 54 and 75 are written only where the far end carries nothing at all, so a change from one kind to the other is a plain set and an open slice goes on leaving it out.

• optimize, substring and the printers swallowed the text behind a code the parser could not finish. All four held a code back only where it was an unterminated control string, while three other shapes wait for a byte just as surely — a bare ESC, a CSI with no final byte, an ESC left on an intermediate byte. In the string each of those was ended by the ESC of whatever stood behind it, and where that was an SGR these loops do not copy it but write it again as a transition — which, for a redundant SGR, writes nothing at all. The code then stood against the text and read it as its own: Parser('\x1B[3\x1B[0m1m!').optimize() gave \x1B[31m!, three characters of text turned into a colour, and a truncated CSI went on eating until it found a final byte. A redundant SGR is what optimize exists to remove, so the defect was the feature working. All four now hold back whatever the parser could not finish and supply an ST where what follows would otherwise be swallowed — an ST is an ESC and a \, so its ESC breaks off the waiting sequence exactly as the string's own did, and an ST that closes nothing does nothing. What removeAll calls the text is what comes out of all four; see docs/records/2026-08-13[6] for the invariant and for what it costs on a truncated CSI, where this package's reading of the input and a terminal's already differed.

• link and linkBel wrote the address into the body of an OSC 8 unchecked. An ESC there ends the sequence where it stands, so a url carrying one handed the rest of itself to the terminal as codes of its own: link('https://ok\x1B\\\x1B[2J…') cleared the screen, and the parser read the result as seven entities where three were meant. Urls in a command-line tool arrive from git remotes, HTTP answers and registries, so the bytes are rarely the caller's. Both functions percent-escape what an OSC 8 cannot carry — the C0 controls and DEL — and nothing else: an address that carries none, which is every address that is one, comes out byte for byte, its own percent-escapes untouched. Uri.encodeFull, which the OSC 8 note asks for, escapes the % as well and would turn an already-encoded address into %2520. The eight-bit C1 are deliberately not escaped: one of them is a single code unit in a Dart string and two bytes in UTF-8, so a single-byte escape would name the wrong byte, and this package does not read them as control codes anyway. The text of a link is written as it came — styling it is what the codes are for — and where none is given the encoded address stands for it.

Renamed:

• Match is Piece, Matches is Pieces, and parser.matches is parser.pieces. The old name shadowed dart:core.Match, and shadowed it silently: an explicit import outranks the implicit one, so the compiler never asked which was meant. Ordinary code written beside this package — for (final Match m in RegExp(r'\w+').allMatches(s)) — failed with two errors that named no package, and the advice this README gave for the Flutter names did not cover it, because the Flutter names do raise the question and this one did not. Piece is the word the package already used for the thing: the class dartdoc opened with "one piece of a parsed string", and _pieceAt, nextPiece and takePiece were there before the rename. There is deliberately no typedef Match<S> = Piece<S> to ease the move — it would reintroduce the shadowing this removes. A test holds the name open from the outside: it uses dart:core.Match beside a single import of this package, and stops compiling if the name is ever taken back.
• rgb and gray are rgb256 and gray256. Both answer with an index into the 256-colour table --- the 6×6×6 cube and the 24-step grey ramp, taking 0..5 and 0..23 --- and stood one name away from fgRgb and its pair, which take a truecolour triple of 0..255 and write it into the sequence itself. fg256(rgb(255, 0, 0)) is the mistake the old names invited, and it throws rather than showing the wrong colour, but the new names say which of the two kinds of red is being asked for. They are also two very general words to have been taking out of a caller's namespace. Color256.rgb and Color256.gray keep their names: a named constructor says whose they are.
• The hyperlink entity is OscLink, not Link. Link shadowed dart:io.Link --- a symbolic link --- and shadowed it the silent way Match used to shadow dart:core.Match: an explicit import outranks the implicit one, so a command-line tool, which almost always imports dart:io, got two errors that named no package. The new name says which sequence it is, the way EscCommon and CsiCommon do. A test holds dart:io.Link open from the outside and stops compiling if the name is ever taken back.
• RESERVED to RESERVED_5F, named after its byte rather than claiming a word that plain in the namespace this package exports.
• toStringAsEscapeSquences to toStringAsEscapeSequences, which was missing a letter.
• The standart_colors directory is spelt standard_colors.

Removed — every name deprecated in an earlier release, and some that never were:

• The style constants renamed in 2.0.0: faint, resetBoldAndFaint, italicized, resetItalicized, singlyUnderlined, doublyUnderlined, resetUnderlined, slowlyBlinking, rapidlyBlinking, resetBlinking, negative, resetNegative, concealed, resetConcealed, crossedOut, resetCrossedOut, framed, encircled, resetFramedAndEncircled, overlined, resetOverlined, superscripted, subscripted and resetSuperAndSubscripted. Use dim, italic, underline, blink, inverse, invisible, strikethrough, frame, encircle, overline, superscript, subscript and their reset… counterparts.
• The string extensions without the ansi prefix: hasEscapeCodes, hasCsi, hasSgr, hasForeground, hasBackground, removeEscapeCodes, removeCsi, removeSgr, removeForeground, removeBackground, showEscapeCodes and showControlCodes. Use the ansi… names.
• The typedefs left behind by the renaming in 3.0.0: AnsiParser, AnsiPrinter, SgrState, SgrPlainState and SgrStackedState. Use Parser, Printer, State, Style and Stack.
• Style.defaults and Stack.defaults. Use terminalColors, or NoStyle where nothing at all should be written.
• Parser.stateAtPos and runZonedAnsiPrinter. Use stateAt and runZonedPrinter.
• foregroundPattern and backgroundPattern, the regular expressions that encoded the assumption a colour is the whole parameter list. Nothing else used them.
• MatchingState, MatchesResult and ParserIterator, which the parser passes to and gets back from its own private methods and nothing else could reach. Matches and Match are unchanged.
• IntensityStyle left the public API. It is the element a Stack's intensity history holds; nothing public takes or returns it, and bold and dim — unlike the other pairs — can be on at once, so no getter could honestly answer with one of them.
• The parsing entry point. After 4.0.0 made it byte-identical to style, one of the two names had to go: import package:ansi_escape_codes/style.dart — the same 81 names — or the umbrella ansi_escape_codes.dart.

Breaking changes:

• Stack.underlineColor takes an ExtendedColor where it took a Color. SGR 58 carries a 256-colour index or a truecolour triple and has no 16-colour form at all, so a Color16 was a colour the sequence could not be written with — Style.underlineColor had always taken the narrower type, and the two now agree. Narrowing a parameter is source-breaking: stack.underlineColor(Color16.red) no longer compiles, and Color256.red, whose index is the same colour, is what it becomes. The Fixed list below mentions the change as part of the bug it belongs to; it is named here because the compiler will name it first.

• OscLink(url) — Link(url) before the rename above — is no longer const. It percent-escapes a control byte in the address, as link does and for the same reason, and a const initializer admits neither a function call nor a contains — so the address could there be neither encoded nor so much as checked. A const Link('…') has to lose the keyword along with the name; nothing else about it moves. OscLink.url reads back the encoded address rather than the bytes handed in, so that it agrees with a parse of OscLink.string and with the equality an Entity takes from those bytes.

• The named control sequences print as themselves: CursorUp(4), CursorPos(3, 7), EraseInPage(ErasePart.all) where Csi([CSI 4 CUU]) was written before. Nothing reads toString but a person and a golden test.

• Csi, Esc and EscapeCode are sealed, and this release adds types under them. A switch that covers them exhaustively has to name the new ones. is checks, casts and the identifiers entities are shown by are unchanged.

• Color.withPrefix(String) is gone, and nothing public stands in its place: a colour is named by the slot of the state it is held in, so Style(background: c).backgroundColor?.id is what c.withPrefix('bg256') was for. The string was a way to be wrong — withPrefix('bg256') gave bg256256Gray5 — and it let the name of a target be written out by hand, which is how the colour of the underline came to call itself underlineColor256Red where the constant is underline256Red. ColorTarget names the three slots now, and takes the name from the SGR function that sets them, so there is one place for it.

• The colours on Style — red, bgYellow, rgb531 and the rest of that table — now come from an extension, StyleColors, rather than from the class itself. Written the usual way they behave as they did; what an extension cannot do is answer a dynamic receiver.

• The predefined styles are constants of one class, Styles, and there are 783 of them: the fifteen properties — Styles.bold, Styles.italic — and the 256-colour table three times over, Styles.red for the colour of the text, Styles.bgRed for the colour behind it, and Styles.underlineRed for the colour of the underline, which had a name in neither the 530 nor the getters. Being constants, a style can be held in one: const error = Styles.red.

  The 530 top-level names are gone, among them foreground, background and underlineColor. A chain that starts at nothing at all starts at Style.terminalColors; the three functions are the constructor, Style(foreground: c), or the methods of the same name on a style.

  The colours are still getters as well — the StyleColors extension — so Styles.red.bold.bgYellow is one chain, as red.bold.bgYellow was.

  This is what took the 31 names that style.dart and ansi_escape_codes.dart both claimed out of the way, and the second now exports the styles as well: one import where there were two.

add operators <, <=, >, >= for Colors

• Add operators <, <=, >, >= for Colors enum.

fix toString for NoStyle

add NoStyle, rename defaults to terminalColors

The style doesn't handle the string when it's empty

add ansiHasControlCodes and lengthWithoutEscapeCodes

add open and close to Style

• Add open and close to Style.

fix padLeft and padRight

• Fix multiline output by Style.

• Fix SDK constraint.
• Add Color256.rgb and Color256.gray.

• Fix README.
• Add predefined colors: Color256.rgb123, etc.
• Add Color256.rgb and Color256.gray.
• Update example/colors256.dart.

• Add methods indexOf, lastIndexOf, contains, startsWith, endsWith that work with a plain string.

• Shift package focus to parsing and standardization.
• Add AnsiParser.
• Add AnsiPrinter.
• Add stacked AnsiPrinter.
• Add their corresponding functions to intercept the print function using zones: runZonedAnsiParser.

Breaking changes:

• The names of some constants have changed: italic to italicized, blinking to slowlyBlinking. All constants of the form not… are renamed to reset…. (fg/bg/underline)Bright… ara renamed to (fg/bg/underline)High….
• Removed methods: handle…, all…. Use AnsiParser instead.

• Add handleEscapeSequences and handlePlainText.

• Add analysis escape sequences (showEscapeSequences).
• Add all control codes (0x00-0x1F).

• Add dart doc comments.
• Update README.
• Refactor.

• Add saveCursor and restoreCursor.
• Add progress_example.dart.
• Refactor.

• Initial version.