v1.0.7 -- "生命中，再无聊的时光，也都是限量版。"

• chore(release): bump version to 1.0.7

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01GXxfHB38J4dvd7j4vEnWGA

• refactor: strip dead code and the redundant CoreInterface layer

• An ablation pass over lib/ (core/ excluded): remove each candidate, then

• require flutter analyze and flutter test to stay at baseline.

• CoreInterface was a mixin declaring 35 abstract methods that

• CoreHandlerInterface then implemented one for one, and nothing in the tree

• ever used it as a type. Merged it into CoreHandlerInterface, keeping the

• four lifecycle methods the platform backends supply.

• Also gone: two files nothing imported (common/archive.dart, common/store.dart),

• 27 unreferenced top-level declarations, and ~25 dead members. Three of those

• were hiding defects rather than just weight -- putProfileDisabledRule wrote

• RuleScene.added instead of disabled, Script.saveWithPath copied a file onto

• itself, and importUri was a bare alias for importText. Dropping

• getMaterialColorShades made _createPrimarySwatch, value32bit, the four *8bit

• getters and lighten unreachable; dropping Windows.registerTask left

• String.encodeUtf16LeWithBom with no caller but its own test, so that pair went

• too (hence 862 tests, down from 864).

• Kept, because the ablation showed they carry weight: the six abstract

• interfaces under core/desktop each pair a production implementation with a

• test fake, and NodeCodec has eight implementations.

• Verified beyond the suite by installing a debug APK on an arm64 emulator --

• core start/stop/restart, theme switching, node import and the settings screens

• all behave as they did before.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01XSQDyYNbJPMa7e3Bkxhgrt

Download based on your OS:

OS	Download
Android
Windows
macOS
Linux

v1.0.6 -- "到最后遗憾的是我，还是你。"

• fix(ci): survive the flaky hdiutil detach that fails the macOS dmg

• The v1.0.6 build produced Avalon.app fine and then died in appdmg's last

• steps: hdiutil detach /Volumes/Avalon returned "No such file or directory",

• appdmg treated that as fatal and deleted both the temporary image and the

• target dmg. Twelve seconds passed between "Blessing image" and the failure,

• so appdmg had been retrying: the first detach reported an error, the volume

• went away underneath it anyway, and every retry after that found nothing left

• to unmount. The image was complete the whole time.

• This is the runner, not the build - a volume held a moment too long by a

• background process, or unmounted asynchronously by DiskArbitration. appdmg

• gives us no way to tell it that an already-unmounted volume is a success, so

• put a shim in front of hdiutil on the PATH of the packaging process. It only

• touches detach: an absent mount point exits 0, a busy one is retried, and

• the last attempts sync and force it out so the image content is still

• flushed. Five failed attempts still exit non-zero, so a genuinely stuck

• volume is not swallowed.

• The shim is only prepended for macOS packaging, so no other platform's

• environment changes.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_013ChdTt9BG4wAfqY35MDr1o

• chore(release): bump version to 1.0.6

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01NSAj59bHG8PDrWLncUW8fC

• fix(nodes): default udp on, the real cause of the WebRTC leak

• The virtual adapter was never the problem. With the tun up and its routes

• installed, a STUN probe still left through the physical interface carrying the

• real address, and 1.0.4's strict-route did nothing about it.

• mihomo gates UDP on a per-outbound flag: adapter/outbound/*.go declares it as

• UDP bool with the proxy:"udp,omitempty" tag, so a proxy that omits the key

• is parsed as not supporting UDP. tunnel.go's match() then skips every rule that

• resolves to such an outbound and, once the rules run out, returns

• proxies["DIRECT"]. The packet leaves through the physical interface with the

• real source address and nothing is logged. WebRTC's STUN probes and QUIC both

• take that path, which is why ipcheck.ing kept reporting the real address while

• everything else went through the proxy.

• Share links do not carry udp; only Clash YAML subscriptions write it. So this

• hit exactly the nodes this app exists for - the ones imported by URI and

• combined into relay chains - and left subscription users untouched.

• Default it to true, in the URI codec so the value is visible in the node editor

• and travels with exported links, and again in effectiveNodeConfig and

• effectiveStoredNodeConfig so nodes already in the library are covered without a

• migration. Profile proxies get the same treatment in the chain runtime: a

• subscription that omits the key leaks the same way and never reaches the node

• library. Only the fifteen types the core actually reads the key for; hysteria,

• hysteria2, tuic and shadowquic set UDP in Base themselves, http and ssh are

• always false. The default is applied before the overlay merge, so an explicit

• udp: false and an explicit removal both still win.

• A server that cannot carry UDP now fails to dial instead of quietly going

• direct, which is the failure this needs.

• Drop the plain udp flag from the relay-hop compatibility check. It is the

• default now and carries no signal; leaving it in would have put a warning on

• every multi-hop chain, and setup.dart logs each one on every apply.

• Hide the strict-route switch on macOS. sing-tun implements it in tun_linux.go

• as unreachable rules and in tun_windows.go as WFP filters; tun_darwin.go never

• reads the field. The switch and its description promised a protection that does

• not exist there. The field, the bridge and the config writer stay - they are

• real on the other two desktops.

• Turn the tun switch back off when authorization fails. The state stayed at

• unauthorized, _getEffectiveTunEnable rewrote tun.enable to false for the core,

• and the switch kept showing on: the user believes traffic is going through the

• virtual adapter while only the system proxy is running. Log it too, because the

• state was otherwise invisible from outside the process. Authorization is still

• asked for once per app lifecycle.

• Stop escaping spaces in checkIsAdmin's core path. It goes to Process.run as an

• argv element, not through a shell, so an escaped path makes stat miss the file

• and the app reports itself as unauthorized.

• Verified on device: importing a share link without a udp parameter now stores

• udp: true, while nodes imported before this change still have the key absent.

• For anyone reading the 1.0.4 notes: "stop DNS and WebRTC leaking the real

• address" was wrong about WebRTC. strict-route is a no-op on macOS and the leak

• had a different cause entirely. The IPv6 takeover and fake-ip-range6 changes

• from that release are unrelated and still stand.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01NSAj59bHG8PDrWLncUW8fC

Download based on your OS:

OS	Download
Android
Windows
macOS
Linux

v1.0.5 -- "就像那句狠话：有一天，我一定会后悔的。"

• chore(release): bump version to 1.0.5

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01LKYUwNq8JMjiGfvPQWekST

• fix(nodes): send XHTTP transport settings to xhttp-opts

• Importing a VLESS or VMess link with type=xhttp put path and host into

• ws-opts. mihomo reads XHTTP settings from xhttp-opts only (XHTTPOptions

• in adapter/outbound/vless.go), so path was lost, mode was dropped on

• the floor, and host fell back to the TLS servername. The node could not

• connect at all, and nothing in the app ever wrote xhttp-opts, so only

• raw YAML nodes could carry an XHTTP transport.

• Route them to xhttp-opts on import, keep mode, and read them back on

• export so the URI round-trips. httpupgrade keeps using ws-opts, which

• is where mihomo reads it from.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01LKYUwNq8JMjiGfvPQWekST

• fix(tun): put the IPv6 takeover behind an explicit switch

• The virtual adapter took over IPv6 whenever it was enabled, overriding

• the IPv6 switch the user had turned off in settings. That also handed

• ::/0 to the tun, so direct traffic started depending on the host having

• a working IPv6 uplink — the shape of failure where a half-broken v6

• (address and default route present, packets dropped) makes dual-stack

• sites time out.

• Add tun.ipv6, on by default, and let it decide. With it off,

• inet6-address is not written either, so sing-tun never installs the

• ::/0 route, and the global ipv6 switch is left exactly as the user set

• it. Taking over ::/0 is safe to keep as the default: verified on a

• device with deliberately broken IPv6 that domestic sites still load,

• because mihomo restores the domain in preHandleMetadata and re-resolves

• it, so the outbound address family is mihomo's choice and

• tcp-concurrent races the dead v6 leg away. Xray honours what the client

• picked and has no such fallback, which is why the same setup breaks

• there.

• Android builds the adapter in VpnService, where tun.enable is always

• false, so the takeover lives in the VPN IPv6 switch instead. That

• switch never reached the core: VpnService pointed ::/0 at the adapter

• while the core still had ipv6 off, so the route was taken over but no

• AAAA was ever resolved. Both platforms now decide through

• SetupAction._tunTakesIpv6, and the result is passed into

• makeRealProfileTask rather than derived there — that runs in a compute

• isolate and should not reach for the platform singleton.

• Also move the fake-ip-range6 gate out of the overrideDns branch. A

• profile shipping its own dns block skips that branch entirely, so a

• subscription setting ipv6: true silently overrode the local DNS IPv6

• switch. Outside the branch the dns map may still be the read-only one

• the profile arrived with, so copy it before writing.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01LKYUwNq8JMjiGfvPQWekST

• chore(release): bump version to 1.0.4

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01YWcTvHimUrxTUtVBRVTcZn

• fix(tun): stop DNS and WebRTC leaking the real address

• With the virtual adapter on, DNS queries and WebRTC candidates still

• exposed the real address. mihomo implements the defences for both, but

• the app could not reach them: the Tun model, the config writer and the

• Go bridge each carried the same six tun fields, so strict-route had no

• way out of the app.

• Add strict-route, on by default. On Windows sing-tun installs WFP

• filters that permit only the tun interface and this process and block

• remote port 53 everywhere else, which is what Windows' multihomed name

• resolution needs; on Linux it adds the unreachable ip rules.

• Make the tun carry IPv6. mihomo clears tun.inet6-address whenever ipv6

• is false (config.parseIPV6), so a tun that does not take over IPv6

• leaves the system routing it out of the physical interface. Force ipv6

• on whenever the tun is enabled and write inet6-address explicitly.

• Whether AAAA is resolved stays with dns.ipv6, so this does not change

• which address family applications prefer.

• Both settings also travel through the bridge, because core_manager

• patches the running config instead of rewriting the profile: enabling

• the tun after launch would otherwise recreate it without the IPv6

• address it needs.

• Add dns.fake-ip-range6 and tie it to dns.ipv6. Under fake-ip, AAAA is

• answered by the withFakeIP middleware, which returns an empty answer

• when there is no v6 pool and never reaches the dns.ipv6 check in

• withResolver, so that switch did nothing in the default mode. It now

• provisions the pool. Hosts without global IPv6 are unaffected:

• parseIPV6 runs before parseDNS and parseTun and clears both the v6 pool

• and the tun address.

• Verified by parsing the generated config with mihomo's own config.Parse

• in both environments. With IPv6 the tun keeps fdfe:dcba:9876::1/126 and

• FakeIPPool6 is created; without it both are dropped and parsing still

• succeeds.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01YWcTvHimUrxTUtVBRVTcZn

• chore(release): bump version to 1.0.3

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01THm54YvuV1sjPqEMGFrgxq

• feat(profiles): give each tab its own action button

• The Profiles scaffold showed one Add Profile button for all three tabs,

• while the node and chain actions sat inline above their lists. The

• button that matters depends on the tab you are looking at.

• Drive the floating action button from the tab index: Add Profile on

• Profiles, Create node and Import node on Nodes, Create chain on Chains.

• The inline buttons in the node toolbar and above the chain list are gone

• now that the actions live in the scaffold; the empty states keep theirs

• as a call to action.

• The two node buttons sit at opposite edges, which the end-aligned slot

• cannot express, so CommonScaffold gained a floatingActionButtonLocation

• and that tab uses centerFloat with the edge inset added back.

• To reach the actions from the scaffold, the node import/create handlers

• and the chain create handler become top-level functions in their own

• files; the views delegate to them.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01THm54YvuV1sjPqEMGFrgxq

• feat(nodes): import nodes without a preview step

• Removing only the overwrite-or-copy radios left the preview dialog in

• place, so importing still interrupted the user to confirm a list they

• did not ask to review. Importing a link means "add this": parse it, add

• what is valid, and report what is not.

• Drop the preview entirely from both import paths. NodeLibraryService

• already skips drafts that fail to parse and returns their issues, so the

• Nodes tab now surfaces those issues instead of showing them in a list

• nobody needed to approve. Binding follows the caller as before: imports

• bind to the current profile when there is one.

• Delete NodeImportPreview, NodeImportSelection and their test, along with

• the importPreview and bindCurrentProfile messages that only that dialog

• used. _previewAndCommitNodes becomes _commitNodes, since it no longer

• previews anything.

• Verified on an Android emulator: a node deep link and the Nodes tab

• import both create the node with no dialog, and importing the same link

• twice yields two copies.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01THm54YvuV1sjPqEMGFrgxq

• feat(nodes): always add on import instead of asking about duplicates

• The import preview asked the user to choose between updating the

• existing node and creating a copy. The choice is not worth a prompt:

• importing a link means "add this", and re-importing the same link should

• simply yield another copy.

• Drop the radio group and commit imports with createCopy, so the preview

• only asks what it still needs, which drafts to keep and whether to bind

• them to the current profile.

• Clean up what the choice left behind:

• • NodeImportSelection.createCopy, now that nothing reads it.
• • The createCopy parameter on importNodeText and _previewAndCommitNodes,

• which no caller ever passed.

• • The updateExisting message, whose only reference was the removed radio.

• NodeLibraryService.commit keeps its createCopy parameter: the node menu

• still duplicates a node with it, and manual creation still dedupes by

• fingerprint without it.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01THm54YvuV1sjPqEMGFrgxq

• fix(android): keep warm-start deep links out of Flutter routing

• Deep links are handled by app_links, but the Flutter engine's own deep

• link handling is on by default and pushes the incoming URI into the

• navigation system. This app uses a plain MaterialApp with no route

• generator, so a link delivered to an already running instance raised

• "Could not find a generator for route" and never reached LinkManager.

• Cold start worked only because getInitialLink reads the intent

• directly. Warm start is the common case here, since the app is meant to

• stay running in the background while a subscription link is opened from

• a browser or chat app.

• Set flutter_deeplinking_enabled to false so the intent reaches

• app_links in both cases.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01THm54YvuV1sjPqEMGFrgxq

• fix(nodes): treat every http(s) URI as a subscription

• isHttpProxyUri classified any http(s) URI with an explicit port as an

• HTTP proxy node, so...

v1.0.4 -- "那些过去，从未过去。"

• chore(release): bump version to 1.0.4

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01YWcTvHimUrxTUtVBRVTcZn

• fix(tun): stop DNS and WebRTC leaking the real address

• With the virtual adapter on, DNS queries and WebRTC candidates still

• exposed the real address. mihomo implements the defences for both, but

• the app could not reach them: the Tun model, the config writer and the

• Go bridge each carried the same six tun fields, so strict-route had no

• way out of the app.

• Add strict-route, on by default. On Windows sing-tun installs WFP

• filters that permit only the tun interface and this process and block

• remote port 53 everywhere else, which is what Windows' multihomed name

• resolution needs; on Linux it adds the unreachable ip rules.

• Make the tun carry IPv6. mihomo clears tun.inet6-address whenever ipv6

• is false (config.parseIPV6), so a tun that does not take over IPv6

• leaves the system routing it out of the physical interface. Force ipv6

• on whenever the tun is enabled and write inet6-address explicitly.

• Whether AAAA is resolved stays with dns.ipv6, so this does not change

• which address family applications prefer.

• Both settings also travel through the bridge, because core_manager

• patches the running config instead of rewriting the profile: enabling

• the tun after launch would otherwise recreate it without the IPv6

• address it needs.

• Add dns.fake-ip-range6 and tie it to dns.ipv6. Under fake-ip, AAAA is

• answered by the withFakeIP middleware, which returns an empty answer

• when there is no v6 pool and never reaches the dns.ipv6 check in

• withResolver, so that switch did nothing in the default mode. It now

• provisions the pool. Hosts without global IPv6 are unaffected:

• parseIPV6 runs before parseDNS and parseTun and clears both the v6 pool

• and the tun address.

• Verified by parsing the generated config with mihomo's own config.Parse

• in both environments. With IPv6 the tun keeps fdfe:dcba:9876::1/126 and

• FakeIPPool6 is created; without it both are dropped and parsing still

• succeeds.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01YWcTvHimUrxTUtVBRVTcZn

Download based on your OS:

OS	Download
Android
Windows
macOS
Linux

v1.0.3 -- "宇宙是蚂蚁的梦。"

• chore(release): bump version to 1.0.3

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01THm54YvuV1sjPqEMGFrgxq

• feat(profiles): give each tab its own action button

• The Profiles scaffold showed one Add Profile button for all three tabs,

• while the node and chain actions sat inline above their lists. The

• button that matters depends on the tab you are looking at.

• Drive the floating action button from the tab index: Add Profile on

• Profiles, Create node and Import node on Nodes, Create chain on Chains.

• The inline buttons in the node toolbar and above the chain list are gone

• now that the actions live in the scaffold; the empty states keep theirs

• as a call to action.

• The two node buttons sit at opposite edges, which the end-aligned slot

• cannot express, so CommonScaffold gained a floatingActionButtonLocation

• and that tab uses centerFloat with the edge inset added back.

• To reach the actions from the scaffold, the node import/create handlers

• and the chain create handler become top-level functions in their own

• files; the views delegate to them.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01THm54YvuV1sjPqEMGFrgxq

• feat(nodes): import nodes without a preview step

• Removing only the overwrite-or-copy radios left the preview dialog in

• place, so importing still interrupted the user to confirm a list they

• did not ask to review. Importing a link means "add this": parse it, add

• what is valid, and report what is not.

• Drop the preview entirely from both import paths. NodeLibraryService

• already skips drafts that fail to parse and returns their issues, so the

• Nodes tab now surfaces those issues instead of showing them in a list

• nobody needed to approve. Binding follows the caller as before: imports

• bind to the current profile when there is one.

• Delete NodeImportPreview, NodeImportSelection and their test, along with

• the importPreview and bindCurrentProfile messages that only that dialog

• used. _previewAndCommitNodes becomes _commitNodes, since it no longer

• previews anything.

• Verified on an Android emulator: a node deep link and the Nodes tab

• import both create the node with no dialog, and importing the same link

• twice yields two copies.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01THm54YvuV1sjPqEMGFrgxq

• feat(nodes): always add on import instead of asking about duplicates

• The import preview asked the user to choose between updating the

• existing node and creating a copy. The choice is not worth a prompt:

• importing a link means "add this", and re-importing the same link should

• simply yield another copy.

• Drop the radio group and commit imports with createCopy, so the preview

• only asks what it still needs, which drafts to keep and whether to bind

• them to the current profile.

• Clean up what the choice left behind:

• • NodeImportSelection.createCopy, now that nothing reads it.
• • The createCopy parameter on importNodeText and _previewAndCommitNodes,

• which no caller ever passed.

• • The updateExisting message, whose only reference was the removed radio.

• NodeLibraryService.commit keeps its createCopy parameter: the node menu

• still duplicates a node with it, and manual creation still dedupes by

• fingerprint without it.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01THm54YvuV1sjPqEMGFrgxq

• fix(android): keep warm-start deep links out of Flutter routing

• Deep links are handled by app_links, but the Flutter engine's own deep

• link handling is on by default and pushes the incoming URI into the

• navigation system. This app uses a plain MaterialApp with no route

• generator, so a link delivered to an already running instance raised

• "Could not find a generator for route" and never reached LinkManager.

• Cold start worked only because getInitialLink reads the intent

• directly. Warm start is the common case here, since the app is meant to

• stay running in the background while a subscription link is opened from

• a browser or chat app.

• Set flutter_deeplinking_enabled to false so the intent reaches

• app_links in both cases.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01THm54YvuV1sjPqEMGFrgxq

• fix(nodes): treat every http(s) URI as a subscription

• isHttpProxyUri classified any http(s) URI with an explicit port as an

• HTTP proxy node, so a subscription link served on a non-default port

• was imported as a node instead of creating a profile. Self-hosted

• panels commonly use ports such as 8443 or 2096, and the

• avalon://install-config deep link routes through the same path, so its

• explicit intent was overridden by the heuristic.

• http and https are no longer node schemes, matching what the deep link

• dispatcher and the Android intent filters already assumed. Drop

• HttpCodec, its https alias and the port heuristic, and let importText

• treat any http(s) URI as a subscription URL.

• Raw Mihomo configs carrying type: http proxies keep working; only the

• URI codec is gone.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01THm54YvuV1sjPqEMGFrgxq

• perf(telemetry): gate UI telemetry on a unified activity state

• Traffic sampling ran on a fixed 1s global timer that only stopped when

• the proxy stopped, so a hidden window still paid for Dart scheduling,

• two IPC round trips and provider updates every second.

• Introduce a process-wide AppActivityState fed by the Flutter lifecycle,

• window visibility and window focus, and drive telemetry from it:

• • SetupAction picks 1s, 5s or no sampling from the current activity,

• the dashboard traffic widgets and the tray title need, resamples

• immediately on resume, and skips a tick while a request is in flight.

• • The tray speed title is macOS-only, so its need is platform-gated;

• otherwise showTrayTitle (default true) kept background sampling alive

• for a consumer that does not exist off macOS.

• • Pages and the app root wrap their subtree in TickerMode so animations

• stop when the page or the window is inactive.

• • NetworkSpeed and TrafficUsage drop their provider subscription while

• inactive; Logs and Requests buffer through ActiveSnapshotMixin and

• commit one snapshot once page and window are both active again.

• • Tray titles dedupe on the formatted string, and getTraffic /

• getTotalTraffic bypass the debug RPC timing logs.

• Frame pausing now distinguishes an on-screen window from a hidden one:

• blur only gates telemetry and tickers, while hide and minimize pause

• frames after 300ms instead of 5s. Losing focus must not stop frames, as

• an unfocused window is still scrollable and would appear frozen.

• Verified on an Android emulator: 1 tick/s in the foreground, zero ticks

• across 25s backgrounded, and an immediate resample on resume.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01THm54YvuV1sjPqEMGFrgxq

• fix(ui): show chain names and align node actions

Download based on your OS:

...

OS	Download
Android
Windows
macOS
Linux

v1.0.2 -- "经过上千年的修炼，你学到的就是这个吗？"

• ci: retry Flutter SDK setup after transient downloads

• ci: use Maven Central mirror for Android builds

• chore(release): bump version to 1.0.2

• Start Flutter splash from a blank first frame

• Add animated Flutter startup splash

• Refresh primary icon and status generator

• Update Avalon branding assets across platforms

Download based on your OS:

OS	Download
Android
Windows
macOS
Linux

v1.0.1 -- "很多人宁可否认事实，也不愿面对真相。"

• chore(release): prepare signed v1.0.1

• ci: keep release notes out of the repository

Download based on your OS:

OS	Download
Android
Windows
macOS
Linux

v1.0.0 -- "无心忆本失心意，忆心失本心无意。"

• fix(windows): resolve installer locale path

• ci: drop unsupported build architectures

• chore: organize branding sources

• chore(release): bump version to 1.0.0

• ci: optimize release workflow

• feat: refresh cross-platform branding assets

• fix: update project links and About page

• fix: correct packaging paths and platform naming

• docs: rewrite README and sync translations

• chore: license Avalon additions under AGPLv3

• fix(android): update JNI exports for Avalon package

• refactor: drop the dead crash-reporting plumbing and unused strings

• Removing Firebase left a settings field, a shared-state field and four

• localized strings behind, kept alive only by a localization test. This drops

• them end to end: AppSettingProps.crashlytics/crashlyticsTip, SharedState.crashlytics

• on both the Dart and Kotlin sides, and the crashlytics, crashlyticsTip,

• dataCollectionTip and dataCollectionContent strings in all four locales. The

• shared-state provider now watches two fields instead of three.

• Also removes chainLibrary, importNodeDesc and chainHopHint, three strings that

• were never referenced, and renames plugins/rust_api/CLAUDE.md to NOTES.md so the

• flutter_rust_bridge integration notes survive without carrying an assistant

• config filename.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_018PvMz7SJGTDgSFovM66xPm

• chore: remove the remaining agent tooling files

• AGENTS.md and CLAUDE.md only routed into the deleted .agents directory, and

• .mcp.json and opencode.jsonc configured a codegraph MCP server this project

• no longer uses.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_018PvMz7SJGTDgSFovM66xPm

• chore: remove the JetBrains run configuration

• .run/main.dart.run.xml only provides a ready-made Flutter run entry for

• JetBrains IDEs and takes no part in building, testing or packaging.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_018PvMz7SJGTDgSFovM66xPm

• chore: stop ignoring docs/

• The directory now holds the translated READMEs, so new files under it must

• be trackable. Also drops the .claude/settings.local.json rule, which no

• longer has a directory to apply to.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_018PvMz7SJGTDgSFovM66xPm

• docs: move the translated READMEs under docs/

• Keeps the Chinese README at the repository root and relocates the seven

• translations to docs/. Language navigation, image sources and the LICENSE,

• NOTICE and CHANGELOG links are rewritten so every relative path still

• resolves from its new depth.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_018PvMz7SJGTDgSFovM66xPm

• chore: remove the .codegraph directory

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_018PvMz7SJGTDgSFovM66xPm

• chore: remove agent tooling configs

• Drops the .gemini, .codex, .claude, .agents and .zed directories, which

• carry per-tool assistant configuration rather than anything the project

• builds or ships.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_018PvMz7SJGTDgSFovM66xPm

• chore(release): bump version to 0.8.97

• Artifact file names come from the pubspec version while the release body

• template builds its download links from the tag, so the two have to match.

• v0.8.96 is already taken by an upstream FlClash tag, hence 0.8.97 for

• Avalon's first release.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_018PvMz7SJGTDgSFovM66xPm

• ci: put a hitokoto quote in the release title

• Every release is named <tag> -- "<quote>", matching how emailbox names

• its releases. The quote is fetched from v1.hitokoto.cn with a 5s timeout and

• falls back to a fixed string when the API is unreachable or returns something

• jq cannot parse. The value is stripped of newlines and quotes and capped at 80

• characters so remote content cannot inject extra step outputs or break the

• title.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_018PvMz7SJGTDgSFovM66xPm

• refactor!: rename FlClash fork to Avalon and trim the release pipeline

• Rebrand this FlClash fork as Avalon, repoint every project-owned identifier,

• and reduce the inherited release pipeline to building installers and attaching

• them to a GitHub Release.

• Rename:

• • Dart package fl_clash -> avalon
• • application id com.follow.clash -> com.masteralanlab.avalon, with the Kotlin

• package directories moved to match

• • executables FlClashCore / FlClashHelperService -> AvalonCore /

• AvalonHelperService, keeping the Windows pipe name, unix socket path and

• helper protocol header in sync between the Dart and Rust sides

• • custom URL scheme flclash:// -> avalon:// and generated chain group prefix

• _flclash -> _avalon

• • a fresh Inno Setup app id so the Windows installer is a distinct application

• Upstream references are deliberately left untouched: the Clash.Meta submodule

• branch, the flutter_distributor git ref and the chen08209 forks in pubspec.yaml

• still point at the upstream projects.

• Docs:

• • rewrite the README for Avalon in 8 languages, replacing README_zh_CN.md
• • add NOTICE carrying the GPL-3.0 attribution to FlClash and the list of

• modifications made in this fork

• • state in every README that the project is for study and research only and

• that any illegal use is prohibited, and add a recommended-resources section

• CI:

• • drop the F-Droid, Homebrew tap and Telegram publishing steps
• • skip Android signing when no keystore secret is configured instead of writing

• an empty keystore over the build

• • export ANDROID_NDK from the setup-ndk output, which the Go core build tool

• requires under that exact name

• • fall back to the runner default Xcode when the pinned version is absent

• Android:

• • remove Firebase Crashlytics and Analytics; crash-loop protection now reads

• ActivityManager.getHistoricalProcessExitReasons on API 30+ and degrades to

• no protection below it

• BREAKING CHANGE: the application identifiers changed, so Avalon installs

• alongside FlClash instead of upgrading it. Settings, profiles and the WebDAV

• backup directory are not migrated.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_018PvMz7SJGTDgSFovM66xPm

• Fix three codec outputs the Core rejects

• Device testing surfaced three configs that the Dart layer considered valid

• but the bundled Core refuses at apply time:

• • SOCKS nodes were emitted as type: socks. Mihomo registers socks5

• only, so adapter.ParseProxy failed with "unsupport proxy type: socks"

• and no SOCKS node could start. SOCKS4/4a keep their variant in

• version, which _socksScheme already uses to restore the scheme.

• • scheme://base64(user:password)@host share links kept the undecoded

• blob as the username and left the password empty, so servers rejected

• the auth. Decode it in a shared _credentialParts, but only when the

• decoded value carries a separator, so a literal username that happens

• to be valid base64 is left alone. SOCKS and HTTP shared the bug.

• • h2 transport was written to http-opts, and header values were

• written as plain strings. Mihomo splits these: HTTPOptions takes

• path as a list and headers values as lists, HTTP2Options takes

• host as a list with a plain path, and WSOptions takes plain

• strings throughout. Fixed on URI parse, VMess JSON parse and export.

• None of these were caught by the existing suite because it validates only

• within Dart. Two of them were in fact pinned as expected behaviour by

• codec tests asserting type == 'socks'. Those assertions are corrected

• here, and hub_validate_test.go now pins the SOCKS naming contract on the

• Go side.

• Found by feeding parsed sample URIs for every supported protocol to the

• real handleValidateConfig; the other 15 protocol/transport combinations

• tested clean.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01GuY5bCnM5DDp2Qb5LHKCkS

• Replace the node import entry with create-from-chain

• Add Profile offered "Import node", which duplicated the button already on

• the Nodes tab, while there was no way at all to create a profile without a

• subscription URL or a config file. Nodes and chains are global, but binding

• either one needs a profile, so the no-subscription single-node flow was

• unreachable from the UI.

• Add Profile now lists the chain library instead: picking a chain creates a

• stub profile, binds the chain as default and sets the chain entry group.

• The stub ships an empty PROXY group on purpose. _attachChainEntry appends

• the generated selector and keeps the first member in place, so an empty

• group leaves the chain as the only member and therefore the default

• outbound; a DIRECT placeholder would silently win instead. Core validation

• accepts a select group with no members, verified against the bundled Core.

• Verified on an Android emulator: the entry lists the chain library, creates

• the profile and makes it current.

• Co-Authored-By: Claude Opus 5 noreply@anthropic.com

• Claude-Session: https://claude.ai/code/session_01GuY5bCnM5DDp2Qb5LHKCkS

• Fix R1-R9 from the node library and chain review

• • R1: keep old files, database and settings on restore failure
• • R2: preserve SOCKS4/4a version across export round-trip
• • R3: include chains and generated groups in export
• • R4: show diagnostics and confirmation be...