d01e2b832e91167abbae6b473e99f1a6191d8ef7
7
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
  adminandClaude Opus 5.5
|
d01e2b832e |
Let the agent edit animations, frames, tiles and tilemaps (#429)
After stage 5a the agent could build behaviors and scenes, but not animate an object, cut a texture into frames or paint a tilemap. Stage 5b adds three areas with 30 commands, 115 in all. - animation (12): set the root, add, retarget and make rows relative, set, move, copy and delete keys, easing, insert or remove frames, and curve points. - textures (8): add, change and delete frames, slice a grid, nine slice, sorting line and collision outline. - tiles (10): the images of a tile, painting tilemaps, layers (add, copy, move, change) and the border of a placed tilemap. - get_asset names the targets of an animation with their animatable properties. A tilemap up to 400 cells shows its cells, a bigger one only their count. - After such a call the editor opens the tab of the animation, curve, texture, tile or tilemap. A frame edit selects the frame there. - One paint step takes rectangles and single cells. Rectangles go first, so a cell can cut a hole into a rectangle. Cells outside the grid, ids that are no tile and more than 20 000 cells are refused. The grid is the tilemap's largest, a grown placement included. Left out on purpose: Generate collision, because it reads the pixels of the texture, and pasting keys, which goes through the clipboard. Found on the way: - Nine slice without fill now keeps the fill the image has. The model's default is repeat, not stretch as first described. - Painting ran the other way round, rectangles overwrote cells. Tests: - agent-art (new, 11) and the other agent tests green, 64/64. 5 bends of the code, all red. - e2e 71 on atx: 23/23, the new check opens the agent's animation in its tab and reads its targets. - Gate on atx: 321 files, 4969 tests. - Live on a test project: a Spin animation of 360 degrees in 30 frames matched 12 degrees per frame to 0.01 frames in the game. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> |
||
|
adminandClaude Opus 5.5
|
0d8fd6ff52 |
Let the agent use the graph, scene and project commands (#429)
The agent knew 23 of the editor's commands. It could not point a Variables node at a variable, group nodes, edit ports, transitions, containers, colliders or the project settings. Stage 5a brings the catalog to 85 commands in five areas and keeps every answer short. - The catalog lists names and descriptions only. With a query (an area or command names) it adds the arguments. All schemas at once would pass the 60 000 character limit of one answer. - add_port points a Variables node at a variable. The port takes the variable's kind, and an unknown name is refused with the list of the ones that exist. - add_group and add_comment take nodes instead of coordinates. They follow the nodes after the arrange step, in the same undo step. - The rule for which list a port picks from moved into portChoices.ts. Inspector and agent now share it. - The frame of a group around nodes moved into annotationDrag.ts, so the button and the agent get the same frame. - list_assets returns the project settings. get_asset shows the content of placed containers with the ids overrides need. get_graph shows declared ports and frame colors. - A new node is placed at the height of its port, in the free spot whose links cross the fewest others. A source goes left of the node it feeds. Before, new nodes always went right of the rest. - Arrange breaks a tie in crossings by putting the flow on top, and exec links pull harder when a column is sorted. - A comment over a group keeps clear of the group's title bar. - The agent's instructions and lint_graph ask to look at crossings too and to call arrange_graph when there are more than needed. Left out on purpose: file assets (no upload), paste from the clipboard, internal wrapper commands and switching plugins on. Found on the way: - Variable commands on a Struct asset did not run through WithStructTypes. New fields did not reach variables of that type. - A query "variables" also matched set_variables_scope. A word that names an area now means only that area. Tests: - agent-catalog (new, 10), agent-edit, graph-arrange (3 new) and relay green. 11 bends of the code, all red. - e2e 71 on atx: 22/22, the new check sees port speed and a group around all 4 drawn nodes. - Live on a test project: a speed variable of 300 moved the player to x 150.03 in 0.5 s. A re-placed Variables node landed below Axis input with 4 crossings instead of 7. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> |
||
|
adminandClaude Opus 5.5
|
1943539a7a |
Let the agent play the game and read its state through MCP (#429)
An agent could build a game object but not see whether it works. It had to ask the user to press the keys. Now it can play the game in the editor itself, send keys and mouse input and read what happened. - play_input sends steps into the running game: hold or tap a key, keep it down or let it go, click, press, release or move the mouse, turn the wheel, wait. Holding and waiting count game seconds. - A mouse target is a point in world units or an object by id or name. A click on an object reports whether it landed on it. - The answer names the game time used, the keys still held, the mouse and the new console lines. - One tool instead of press_keys and pointer from the plan. A step list can then mix both, like holding Shift during a click. - get_runtime reads the running game: objects with id, type and position, the own variables of their behaviors, the Variables assets, the camera, keys, mouse and the console. Named objects come with all their properties. - A hidden editor tab gets no frames. play_input says so at once. If game time stands still for 25 s it stops, names the step and lets go of every key it held. - get_view and play say when the scene has no active camera. The first agent test missed that and only saw it on a screenshot. - The relay instructions tell the agent to test what it built. Found in the live test on a real project: - A click on an object by its name reported hit false. The hit test looked the target up by id only. - get_runtime listed the built-in switches of a behavior (enabled, runInEditor) as its variables. Tests: - agent-play (new), agent-edit, agent-view, relay green. - 5 counter-checks red: hold counts one frame, key never released, click ignores the object, hidden tab not refused, no variable filter. - e2e 71: 21/21. The agent holds ArrowLeft 0.5 s (x -120 to -270) and clicks the object (hit true). - Gate: CHECK EXIT 0, 319 files, 4944 tests. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> |
||
|
adminandClaude Opus 5.5
|
97e0d93168 |
Let the editor follow the agent's edits on screen (#429)
The user asked to see what the agent is working on through MCP. Until now the agent had to call open for that, and mostly did not. - After run_commands the editor shows what the call worked on last. A graph opens its tab with the touched nodes selected, deleted ones left out. A scene or container shows with its objects selected. Any other asset shows in the inspector. - Objects in another scene switch the scene inside the same undo step. A switch of its own would be its own step, and Ctrl+Z would take back the switch before the agent's edit. - A graph tab that is open already moves its camera after the nodes. It stays put while they are in view, moves only as far as needed with the zoom kept, and frames them like Fit only when they do not fit. This is revealBoxes in framing.ts and the state's reveal event. - A tab that has just opened frames itself first and moves after that. - The account menu has "Follow agent" below "Agent", on by default and kept in the browser. While the game plays the editor never follows. - The run_commands description and the relay's instructions tell the agent about it. Tests: - agent-edit.test.ts: 3 new tests, each red with its bent-back fix (scene switch outside the undo step, deleted nodes selected, switch ignored) - graph-framing.test.ts 29/29, 6 of them new - e2e 71 on melektaus-atx: ALL PASSED, 20 checks. 71g4 adds a node 3000 px to the right of an open tab: it ends 24 px inside the edge, zoom still 1 - Tried live on a test project through a real MCP session: the tab opened, the nodes and the object were selected. With the old build a far node stayed out of view, after the reload the camera went there - Gate on melektaus-atx: CHECK EXIT 0, 318 files, 4931 tests Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> |
||
|
adminandClaude Opus 5.5
|
b9d1602b94 |
Arrange graph nodes for the agent and with a button (#429)
An agent that built a graph through MCP left every new node in one row. In the first real test that put nine nodes on a line 2080 px long, with the Owner's links running straight through all the others. Placing nodes by hand cost nine move_node calls and guessed heights. The agent now gives no coordinates, and the editor lays the nodes out. - arrange.ts is a pure function. It puts the flow into columns from left to right and sets an Owner or Get right before the first node it feeds. It orders each column for the fewest crossings and lines nodes up by their linked ports without overlaps. - Node sizes come from the drawn NodeView in the browser and from an estimate in unit tests. Both share the constants in nodeBox.ts. - run_commands arranges the nodes that came without x and y when the call ends, inside the same undo step. Nodes that were there stay. - New command arrange_graph and new tool lint_graph. The lint counts overlaps, crossings, links through foreign nodes, backward links and the longest links, with example ids. - The relay instructions tell the agent not to give coordinates and to fetch far values with a second Get or Owner node. - An Arrange button next to the snap magnet in all eight graph tabs. It arranges the selected nodes, or all of them, as one undo step. Found while testing on a real project: - lint_graph missed a crossing where two links swap rows exactly mirrored. Both curves then share a sample point in the middle, and only a crossing between two sample points was counted. Links that meet on a point count now (segmentsMeet). - The Arrow-key graph with one Owner for four Moves cannot get below 8 crossings. The Owner feeds object in row 2 of Move, between exec and dt of the key, so it always crosses one of them. Tests: - graph-arrange, agent-edit, snap-toggle: 33/33. The swapped-rows test was red before the fix (0 instead of 1). - Four counter-checks that bend arrange or the edit plan, all red. - e2e 71 on melektaus-atx: 18/18. Lint after the build shows 0 overlaps, 0 links through nodes, 0 backward. Arrange takes a tangled graph from 3 backward links and 2 crossings to 0, and one Ctrl+Z restores it. - Gate on melektaus-atx: CHECK EXIT 0, 318 files, 4917 tests, measured before the lint fix. - Live through MCP on the user's project: lint_graph reports 8 crossings after Arrange, matching the screenshot. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> |
||
|
adminandClaude Opus 5.5
|
25af064249 |
Let an AI agent read and edit the open project through MCP (#429)
Stage 2 of #429. Until now an agent could only look at the editor. Now it reads the project and changes it through the same commands the user's own edits run, so every change shows up live and one Ctrl+Z takes back a whole agent call. - Read tools: list_assets, get_asset, get_graph and catalog. The catalog lists the commands with their arguments, the nodes of a graph from the same palette the user sees, asset types and variable kinds. - run_commands runs its steps as one undo step named "Agent: ...". A failing step rolls back the steps before it. The answer names the step and the reason the editor gave. - A step can name what it made ("as") and later steps use "$name". Each step is built after the one before has run, so one call can make a behavior and fill it. - 23 commands cover assets, variables, scene objects and graphs. Textures, fonts, sounds and videos need a file and are left out. - open, select, play and undo/redo drive the editor. - Deleting asks the user with the same text as their own delete. The tab sends "wait", and the relay then gives the call 2 minutes instead of 30 seconds. - tools.json is the only source of the argument schemas. argCheck checks every call against them before anything runs. - An open menu now names its first entries in get_view, instead of an empty title. - The new host functions live in editorToolHost.ts, so Editor.ts does not grow further. Found while testing: e2e 71 hung for 30 s in Page.navigate after a failure. The beforeunload guard of an unsaved project held the page and hid the real error. The module now reports the error first and waits for the save before it navigates. Tests: - Gate on melektaus-atx: CHECK EXIT 0, 317 files, 4903 tests. - agent-edit 14/14, agent-view 10/10, relay 24/24. - e2e 71 16/16 on atx and against the live site. One call builds a sprite with an arrow-key behavior, play moves it from x 0 to -125, and one undo restores the assets exactly. - Counter-checks turned red: no rollback, a partial undo, no delete question, a relay that ignores "wait". Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> |
||
|
adminandClaude Opus 5.5
|
48ec722660 |
Let an AI agent see the open editor through MCP (#429)
Claude Code should be able to work inside the user's open editor tab. This first stage connects the two and lets the agent look: what the user sees and a picture of it. Changing the project comes in stage 2. - An agent token per account opens only /mcp, no API route. The dialog "Agent tokens..." in the account menu creates it once, shows the ready "claude mcp add" line and revokes it. The server keeps only its sha256 (api/lib/AgentTokens.php). - Wrong tokens count into the sign-in brake from #427. Changing the password, renaming or deleting an account ends its tokens. - The relay (tools/mcp/) is a small Node service without packages, WebSocket server included. It runs in the container feuerwerk-mcp because PHP-FPM cannot hold connections. Whether a token or a tab ticket is valid is still decided by the PHP API, through two routes that answer only with the relay key. - The tab joins over WebSocket on the same origin with the normal session, an Origin check and a one-time ticket. It connects only after "Agent" is switched on in the account menu. One tab per account is active, a new one takes over. Guests get nothing. - Tools: get_view (project, tab, document, selection with names, camera, play state, open dialog) and screenshot. The resource feuerwerk://view can be subscribed. Claude Code does not act on that reliably, so a changed view is also noted on the next answer. - A badge next to the account menu shows the connected agents. It counts only agents with an open stream or a request in the last two minutes. - CI starts the relay in front of php -S (tools/ci/serve.sh), and e2e 71 runs the whole path in the browser. Found on the live test: the badge showed four agents because old sessions of revoked tokens were still counted. Tests: - Gate on atx: 316 files, 4888 tests - relay protocol test with fake API and fake tabs: 23/23 - get_view, tools, socketUrl: 16/16 - PHP agents 29, rights 103 checks - e2e 71 on atx 10/10, against the live site 10/10, no CSP violation - real Claude Code against the live site: get_view answered - ZAP against /mcp: nothing at High or Medium - every new check fails when the guard it covers is taken out (6) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> |