Imported from meow-sci/gatOS (.agents/skills/gatos/SKILL.md). Install upstream with npx skills add meow-sci/gatOS --skill gatos. Copyright stays with the author.

gatOS /sim programming

gatOS exposes live KSA simulation state as a filesystem. Programs read game data and control vehicles by reading and writing files under /sim — cat a sensor, echo a value to a control file and it actuates the game (returning a real Linux errno on failure). The exact same surface is served over HTTP /v1 and MQTT, so a program can run inside the guest (read /sim directly) or on the host (HTTP). No custom RPC — the files are the API.

For an AI agent, gatOS also provides a first-class MCP server. It projects the same snapshots and command pipeline as concise logical JSON resources/tools (world, celestial, vessel, kitten, runtime, direct controls, and typed same-tick/timed batches), rather than mirroring every /sim leaf. Its public contract is SPEC_MCP.md. Do not invent a filesystem-path tool or use /sim/display: that terminal-video stream is deliberately outside MCP v1.

The complete, authoritative catalog is SPEC_9P_FILESYSTEM.md at the repo root — every path, format, unit, read/write semantic, command action key, errno, and HTTP route. This skill is the orientation; the SPEC is the reference. When you change the /sim surface, update the SPEC in the same change (it has its own constitution).

Read these as needed

In-repo working references: examples/gogogo-rs (minimal Rust control panel — throttle + ignite), examples/land-o-matic (full Rust G-FOLD+UPFG landing autopilot), examples/sdk-ts (typed TypeScript/Bun SDK over both transports).

Mental model in 90 seconds

• Reads are SENSOR files: cat /sim/<path> → one value + \n. Formats: G9 doubles, 0/1 flags, space-separated x y z vectors and x y z w quaternions, verbatim strings, NDJSON streams. The atomic per-vessel doc vessels/<id>/telemetry is one self-consistent JSON snapshot — prefer it for control loops.
• Writes actuate the game: ctl/… (per-vessel control — incl. ctl/translate, bang-bang RCS translation by body-axis signs: 1 0 0 = thrust along the nose, latches until 0 0 0; and its sibling ctl/rotate, bang-bang RCS torque by body-axis signs: +x roll right, +y pitch up, +z yaw right — full authority needs attitude_mode=manual), per-module files (engines/<n>/active, lights/<n>/on, …), and debug/… (cheats: teleport/impulse/refuel/warp/switch). A write is line-buffered and actuates on the newline; failures return EINVAL/EACCES/EBUSY/ETIMEDOUT/…
• A write blocks until the game thread executes it — one command per frame. Sequential writes can therefore never land in the same tick (a formation teleport smears ~100 m per frame gap at orbital speed). To make N writes execute in the same tick, write them as one group to /sim/ctl/batch: one <path> <value> line per command, then a commit line (SPEC §3.10). Atomic, in order, one phase per batch, ≤64 commands. ⚠ Since KSA 2026.8.22.5348, teleport/impulse are Solver phase, so a batch mixing either with a Frame-phase line (e.g. ctl/throttle) now fails EINVAL; an all-teleport formation batch is still correct. To spread writes over time instead of collapsing them into one tick, commit the same lines — each prefixed with an absolute ms offset — to /sim/ctl/timed_batch; they become a host-owned player under /sim/ctl/schedules/<id>/ you can pause/scrub/re-rate/loop/stop (SPEC §3.10). Phase mixing is allowed there.
• Two attitude paths: write a named mode to ctl/attitude_mode (Prograde, Retrograde, …) and the onboard autopilot steers (warp-correct, no math) — or compute a Body→CCI quaternion and write ctl/attitude_target for a custom direction. Attitude/burn writes are solver-phase (take effect next solver step; run ~1× warp for closed loops) — and since KSA 2026.8.22.5348 so are debug/…/teleport and debug/…/impulse, which moved Frame → Solver (revs 5331/5339).
• CCI is the working frame: position/cci, velocity/cci, attitude/quat, debug/teleport, debug/…/impulse (default frame), and ctl/burn Δv are all Celestial-Centered Inertial about the parent body (Z = north pole, X = vernal point, equatorial X–Y plane). Constants come from bodies/<parent>/{mu,radius,rotation_rate}.
• Pace in sim time, not wall time: block on time/alarm (or GET /v1/time/wait); gate on time/sim_dt==0 (paused) and time/warp>1 (warping).

MCP operating loop for an AI agent

MCP is operation-shaped even where a tool uses the shared value/values/token/aux wire slots. Never fill every optional slot speculatively. Choose the operation, read its parameter descriptions and capability entry, and send only the slots that operation consumes.

1. Call gatos.get_capabilities once per session and again after a feature/configuration change.
2. Call gatos.get_world(detail:"summary"), discover raw ids with a list tool, then read the target.
3. Check controllable, current parent, module ordinals, feature gates, and degraded health before acting.
4. Use the narrowest logical tool. Use gatos.command only as the canonical-action backstop.
5. Use gatos.execute_batch for same-tick commands of one derived phase; use gatos.schedule_batch for absolute millisecond offsets that may mix phases.
6. Save the pre-action snapshot_sequence, call gatos.wait(after_sequence:...), and re-read the affected state. An accepted command is not immediate read-back.
7. Branch on structured errno and retryable. Never blindly retry stage, ignite, undock, decoupler fire, release, remove, clear, or another one-shot/lifecycle trigger.

The public operation-by-operation calls and recovery playbooks live under site/src/content/docs/mcp/; the versioned contract is SPEC_MCP.md.

Maintaining the MCP projection

gatOS deliberately uses the official C# SDK's low-level primitives behind a custom, stateless, POST-only Streamable HTTP host. Preserve its Host/Origin validation, lack of sessions/GET/SSE, request limits, cancellation, and real-SDK client tests. Do not add a second command model: logical tools must map to SimCommand and immutable reads must project SimSnapshot/the shared stores.

An MCP surface change updates together: gatOS.Mcp/ and tests, CommandCatalog where the shared action changes, SPEC_MCP.md, site/src/data/mcp-reference.ts, the affected site/.../mcp/ prose, and this skill when agent behavior changes. Multipurpose tool docs must collocate each operation's exact required fields, enum/range/unit, full example, phase/gate, and retry hazard.

Access at a glance

# In the guest (the /sim mount):
cat   /sim/vessels/active/telemetry            # read the atomic vessel doc
echo 0.5 > /sim/vessels/active/ctl/throttle    # actuate (returns errno on failure)
echo "6578100 0 0 0 7784 0" > /sim/debug/vessels/Hunter/teleport   # CCI state vector
echo "10 0 0 body dv" > /sim/debug/vessels/Hunter/impulse          # +10 m/s kick off the nose

# On the host (HTTP /v1, default 127.0.0.1:4242; guest: $GATOS_HTTP):
curl  http://127.0.0.1:4242/v1/vessels/active/telemetry
curl -X POST http://127.0.0.1:4242/v1/command \
     -H 'content-type: application/json' \
     -d '{"vessel_id":"Hunter","action":"debug.teleport","values":[6578100,0,0,0,7784,0]}'

The generic write is POST /v1/command with {vessel_id, action, ordinal?, value?, values?, token?} — see the action-key catalog in SPEC §5. Its file twin is POST /v1/fs/<path> with the raw value; both mirror the /sim files leaf-for-leaf.

Where things live (quick index of /sim)

time/{ut,warp,sim_dt,warp_speeds,auto_warp,alarm}
system/{name,home,sun}
bodies/<id>/{id,class,parent,children,mass,radius,mu,soi,rotation_rate,
             position/ecl, velocity/ecl, orbit/*, atmosphere/*, ocean/*, focus}
vessels/active/…  (alias of the controlled vessel)   vessels/by-id/<id>/
    id name situation parent controlled controllable com scale always_render telemetry stream
    position/{cci,ecl,lat,lon}  velocity/{orbital,surface,inertial,cci}
    attitude/{quat,rates}  altitude/{barometric,radar}  mass/{total,dry,propellant}
    orbit/*  navball/*  environment/*  battery/*  power/*
    engines/<n>/*  tanks/<r>/*  rcs/<n>/*  solar/<n>/*  generators/<n>/*
    lights/<n>/*  docking/<n>/*  decouplers/<n>/*  animations/<n>/*  encounters
    srb/<n>/{engine,part,substance,grain,grain_shape,segment_count,valid,error,active,
             propellant,mass,mass_initial,mass_unburnable,mass_burnable,fraction,
             mass_flow,burn_time,burning_area,chamber_pressure,chamber_temp,
             exit_pressure,exit_temp,area_ratio,
             segments/<m>/{part,substance,grain,mass,mass_initial,mass_unburnable,
                           fraction,radius,length,volume,burn_depth}}
                                        (solid rocket motors — READ-ONLY. Solid propellant is NOT a
                                         tank, so tanks/ never shows a booster; srb/<n>/fraction and
                                         /burn_time are the booster gauges. Ignite via the engine
                                         surface: srb/<n>/engine names the engines/<n> entry. A lit
                                         solid cannot be throttled or shut down.)
    parts/json                          (the whole part/subpart tree as ONE JSON doc — cat + jq it)
    parts/<n>/{instance_id,id,display_name,template,is_root,subpart_count,position,
               subparts/<m>/{instance_id,id,display_name,template,position}}
                                        (parts + subparts; welds anchor picker; telemetry_vessel_parts)
    ctl/{ignite,shutdown,engine,stage,throttle,lights,rcs,translate,rotate,
         attitude_mode,attitude_frame,attitude_target,burn,focus}
events
status/{game_version,sampler,accessors,transports}
ctl/batch                               (atomic same-tick command groups — SPEC §3.10)
ctl/timed_batch                         (the same control surface on a clock; schedule_enabled=true)
ctl/schedules/{count,clear,help,
               <id>/{kind,group,state,t,duration,pending,dropped,clock,last_error,
                     pause,scrub,rate,loop,stop,remove}}      (the live-player registry)
camera/{status,info,target,playback,last_error,enabled,release,mode,follow,tidal,
        map/scope, track/<name>, play,set,stop,
        pose/{position,frame,anchor,geo,orbit/{radius,azimuth,elevation},rotation,
              aim,aim_target,aim_offset,aim_frame,aim_up,roll,fov,ortho,ortho_height,
              smoothing,reset}}         (programmable cinematic camera; camera_enabled=true)
audio/{file/<name>,play,set,stop,status,info}         (userland audio; audio_enabled=true)
debug/                                  (only when debug_namespace=true)
    vessels/<id>/{teleport,impulse,refill_fuel,refill_battery,docking/<n>/pushoff_impulse,
                  weld,weld_here,unweld}
    welds/{clear,count,<source>/{target,part,offset,rotation,lock_rotation,enabled}}
    thug_life/{add,clear,count,<id>/{vessel,part,position,rotation,size,visible,remove,spec}}
    iva/{enabled,run_outside_iva,adopt,adopt_all,clear,count,stats,interior,help,
         <id>/{vessel,part,name,template,position,velocity,angular_velocity,mass,shape,size,
               asleep,nudge,release,spec}}                   (OFF by default — see below)
    always_render_iva   time/warp   focus   control_vessel
    engineplume/{help,templates/<id>/{core,absorption,emission,mach_diamonds,noise,quality}/*,json,reset}
    plumetrail/{help,render/*,json,clear,reset}       (one GLOBAL trail renderer)
    clouds/{help,bodies/<body>/{shared/*,layers/<n>/{*,two_d/*,raymarch/*,types/<m>/*},json,reset}}
    terrain/{help,wireframe,bodies/<body>/{min_height,max_height,slope_roughness_deg,hapke_albedo,
                                           biomes/*,tessellation/*,json,reset}}

Each leaf's format, units, archetype (read-only vs control vs trigger), and backing command action are in SPEC_9P_FILESYSTEM.md. Per-module dirs appear only when the vessel actually has that module; bodies/detail dirs depend on telemetry config gates.

Cheats (/sim/debug, ported from the sibling unscience mod): always_render_iva forces interior (IVA) meshes to render outside the IVA camera; welds rigidly attach one vessel to another vessel's part (weld = explicit pose, weld_here = capture the current relative pose, unweld; the registry + clear/enabled live under welds/); thug_life anchors a flat, world-space textured quad (the "thug life" sunglasses meme) to a part on a vehicle, tracked each frame (gatOS's first custom GPU render — add/clear + per-entry position/rotation/size/visible/remove). For welds and thug_life, discover the anchor part from the target's parts/<n>/instance_id (pass 0 to anchor to the body/CoM frame). A weld anchor may also be a subpart — use parts/<n>/subparts/<m>/instance_id; an animated subpart anchor (robotics/landing-leg segment) tracks the animation. (thug_life anchors remain top-level parts only.) Full arg shapes, action keys, and errnos are in SPEC §3.7; a worked weld_here example is in recipes.md §9 and a thug_life example in recipes.md §10. The render-side internals (how the quad is drawn into KSA's scene) are documented in the ksa skill's quad.md.

IVA cabin physics (/sim/debug/iva) — OFF by default: free-floating objects inside a vessel's interior, with real inertial physics — weightless and drifting while coasting, slammed aft when the engines light, flung around by RCS rotation, and colliding with the actual interior surfaces. enabled is a master switch that starts and ends the whole feature: while it is 0 (the shipped default) no simulation exists at all, and writing 0 puts every object back at its exact rest pose and tears everything down. Turn it on, then cut props loose:

echo 1 > /sim/debug/iva/enabled
echo "Gemini7 4 Sardine" > /sim/debug/iva/adopt_all   # the 4 smallest props matching "Sardine"
echo "Gemini7 12"        > /sim/debug/iva/adopt_all   # or the 12 smallest loose props, any template
cat /sim/debug/iva/0/position                         # each object is /sim/debug/iva/<id>/
echo "0.3 0 0" > /sim/debug/iva/0/nudge               # flick it across the cabin
echo 1 > /sim/debug/iva/clear                         # everything back where it was

Objects drive real shipped IVA prop SubParts — so pass a subpart instance_id (parts/<n>/subparts/<m>/instance_id, or cat parts/json | jq), never a top-level part: KSA saves a top-level part's transform but not a SubPart's, which is exactly why this can never touch your save file. Objects park (frozen) under time warp, in the VAB, and outside the IVA camera unless run_outside_iva=1. Emits iva.impact / iva.escape / iva.release events — wiring iva.impact to /sim/audio gives you clunks. cat /sim/debug/iva/help is a full readme; the surface, tuning knobs and errnos are in SPEC §3.7 (iva).

FX editors (/sim/debug/{engineplume,plumetrail,clouds,terrain}): the game's four built-in render editors as filesystems — one writable leaf per knob, each with a fixed inclusive range (out of range ⇒ EINVAL before it reaches the game), live read-back, a per-entity json document for discovery, and a per-entity reset that restores the values from before gatOS first wrote them. Scopes differ and this is the thing to get right: engineplume is per template (an edit repaints every nozzle using that template), plumetrail is a single global renderer (plus a one-shot clear that deletes existing trails), clouds is per body → layer → cloud type, terrain is per body (only bodies with a live render slot) plus a global wireframe. All writes are Frame-phase, cheap enough to animate at 10–60 Hz — group simultaneous ones through /sim/ctl/batch. Values are stored as 32-bit floats (read-back is single-precision) and the surface is session-scoped (never persisted; restored at unload). Every field, range and unit is in SPEC §3.7; cat <family>/help in-guest.

Audio playback (/sim/audio, gated by audio_enabled=true — NOT a debug cheat): play real audio (mp3/ogg/wav/flac) through the game's speakers. Upload with cat clip.mp3 > /sim/audio/file/clip.mp3 (in-memory, playable once the cat finishes; rm evicts), then echo 'clip.mp3' > /sim/audio/play — optional key=value tokens start=/end= (ms range), vol= (0..1), loop=0|1, group=sfx|music|ui (which in-game volume slider governs it), id= (a handle for later set/stop; auto #N otherwise), pan=, pitch=. Live control: echo 'bgm vol=0.15' > /sim/audio/set (also pause=1/resume=1/seek=ms); stop with echo 'bgm' > /sim/audio/stop (or all). cat /sim/audio/status lists live channels; audio.finished events land on /sim/events (grep -m1 for completion instead of polling). Keeps playing at any time-warp (deliberate — alarms must not mute). Host-side binary upload: curl -T clip.mp3 http://127.0.0.1:4242/v1/audio/file/clip.mp3. Full grammar, caps, errnos: SPEC §3.9.

Timed sequences (/sim/ctl/timed_batch + /sim/ctl/schedules, gated by schedule_enabled=true): ctl/batch makes N writes land in one tick; a timed batch spreads them over a timeline instead. Same paths, same values, each line prefixed with an absolute offset in milliseconds (fractional allowed — never deltas), optional @id/@clock/@rate/@loop/@group directives first, then commit:

cat > /sim/ctl/timed_batch <<'EOF'
@id      launch-seq
@clock   render                          # render | wall | ut
0        vessels/active/ctl/throttle  1
1200     vessels/active/ctl/ignite    1
commit
EOF
cat /sim/ctl/schedules/launch-seq/state   # pending|running|paused|done|failed
echo 0.25 > /sim/ctl/schedules/launch-seq/rate      # also pause/scrub/loop/stop/remove

The host owns the clock, so timing survives the guest going away and replays identically. render accumulates the game's clamped frame delta (lags after a hitch, never catches up — right for footage), wall is true elapsed time, ut is sim time (diverges under warp — right for mission events). Phase mixing is allowed (unlike ctl/batch). Commit is non-blocking and all-or-nothing (ENOENT bad path, EINVAL bad value/directive/cap). On catch-up every trigger fires in order while state controls coalesce to the last write per path (counted at <id>/dropped) — which is why a densely generated script is cheap. schedule.* events land on /sim/events. Full grammar + semantics: SPEC §3.10, or cat /sim/ctl/schedules/help in-guest; worked example in recipes.md §12.

Programmable camera (/sim/camera, gated by camera_enabled=true — NOT a debug cheat): take the game's main camera and fly it from a shell. echo 1 > /sim/camera/enabled captures the live camera, parks the game's controller and makes gatOS the only writer; everything you then write is an override over that snapshot, and echo 0 > enabled (eased) or echo 1 > camera/release (hard cut) puts it all back. A pose is anchor (pose/anchor: vessel:/body:/part:<vessel>/<iid>/none)

• frame (pose/frame: ecl|cce|bodyfixed|enu|lvlh|chase) + a placement — pose/position (cartesian), pose/geo (lat lon alt [body:<id>], altitude above terrain), or pose/orbit/{radius,azimuth,elevation} (a non-zero radius wins). Orientation is normally pose/aim <target> [off x y z] [frame <f>] [up world|target|velocity|free] [roll <deg>] — the offset is measured on the subject and re-resolved every frame, which is what glues it to a moving hull. pose/{fov,ortho,ortho_height,roll,smoothing} are the lens. Anchored smoothing filters only the camera-to-anchor component: the anchor's current-frame translation passes through exactly, and an aim target remains centred because look-at is an exact constraint rather than a smoothed rotation.

echo 1 > /sim/camera/enabled
echo "vessel:apollo11" > /sim/camera/pose/anchor
echo "bodyfixed" > /sim/camera/pose/frame     # +X nose, +Y right, -Z up (so "up" is negative)
echo "-40 0 -6"  > /sim/camera/pose/position
echo "vessel:apollo11 off 0 0 -1.2 up world" > /sim/camera/pose/aim
echo 0.35 > /sim/camera/pose/smoothing        # critically-damped filter, seconds
cat /sim/camera/status                        # pose channels + applied_position_ecl/applied_rotation
echo 0 > /sim/camera/enabled                  # eased hand-back

Three layers composite per channel — track ?? your override ?? the ownership baseline — so a timed_batch (or a bare echo) can pull focus while a JSON track interpolates position. A track (cp shot.json /sim/camera/track/shot; echo shot > /sim/camera/play) registers as a player at /sim/ctl/schedules/camera/, so it needs schedule_enabled too and answers the same pause/scrub/rate/loop/stop. Every track channel has a pose/ leaf, so the JSON is never the only route. Gotchas: mode/follow/tidal and the player's camera keys are inert while gatOS owns the camera (use pose/anchor + pose/aim_target); pose/ortho_height is the one change gatOS cannot undo; the camera is floored at surface + 0.5 m; an EVA kittenaut walks relative to the shot while you hold the camera; and iva/map cannot be gatOS ownership contexts at all (only fixed is — camera/map/scope is the map knob that does ship). Full catalog, errnos and the track JSON schema: SPEC §3.11; frames in coordinate-frames.md §8; worked shot in recipes.md §13.

First-class per-vessel nodes (NOT under /sim/debug; also ported from unscience): vessels/by-id/<id>/scale — write any finite value > 0 to uniformly rescale the whole vessel model one-shot (echo 50000 > scale = planet-sized; echo 1 > restores the live transform; 0/negative → EINVAL). The live write changes transforms without refreshing physics, unlike the physical in-game editor gizmo (clamped 0.5×–2×, quantized to 0.25 m). KSA saves top-level part scale: loading can rescale colliders, mass, tanks and nozzles, including every non-root part since 2026.9.10.5438 (rev 5434). Restore the intended scale before saving a cosmetic experiment. Staging/undocking need not reset it; subpart and EVA avatar transforms have distinct lifetimes. vessels/by-id/<id>/always_render — write 1 to keep that vessel rendered at any distance (bypasses the sub-pixel cull that normally hides far vessels; the mark survives scene rebuilds and auto-drops when the vessel despawns; EVA kittens are unaffected). Both work on any vessel by id even with control_all_vessels=false (deliberately authority-exempt). SPEC §3.4.1 has the full semantics.

Gotchas that cause silent failures

• A vessel's name is its id (Hunter, Polaris are literal ids). Ids in /sim paths are sanitized (non-[A-Za-z0-9._-] → _).
• control_enabled=false → all writes EACCES. control_all_vessels=false → only the controlled vessel is commandable (camera.focus, vessel.scale, vessel.always_render and debug.* stay exempt). debug_namespace=false → /sim/debug/** is gone. All default on.
• /sim/debug/iva/enabled is off by default and is a separate switch from debug_namespace — the iva/ directory is there, but adopting anything before echo 1 > enabled answers EOPNOTSUPP, and nothing is simulated until you do. adopt takes a subpart instance_id (a top-level part is ENOENT by design) and refuses anything bigger than iva_max_object_size (0.5 m) so you cannot cut a hull panel loose.
• debug.teleport sets a CCI state about the vessel's current parent body — the vessel must already be in the intended body's SOI. See SPEC §6. Since KSA 2026.8.22.5348 both teleport and impulse are Solver phase (they land on the next solver step, and cannot share a ctl/batch with Frame-phase actions).
• debug/…/impulse defaults to newton-seconds (Δv = J ÷ live mass), not m/s — append the dv keyword to apply the vector directly as Δv, and body to read it in the vessel frame (+X = nose): echo "10 0 0 body dv" > impulse. Same CCI-about-current-parent frame as teleport otherwise.
• ctl/stage is per-module and no longer touches RCS (KSA 2026.8.22.5348, rev 5329). It walks the staged part's subtree and activates only ISequenced modules — exactly engines and decouplers — whose own sequence matches. ThrusterController is not one, so rcs/<n>/active does not flip as a side effect any more: toggle RCS explicitly. Engines/decouplers on sub-parts now stage where they used to be skipped, and a part with modules in different sequences needs one ctl/stage press per sequence.
• engines/<n>/min_throttle on a multi-engine stack inverted (KSA 2026.8.22.5348, rev 5317). The write still sets that engine's own floor, but the flight computer now folds the active engines with Max (seed 0) instead of Min (seed 1) — the most restrictive engine sets the effective floor, not the least. One high value now raises the whole stack's floor.
• navball/deltav and navball/twr were corrected in KSA 2026.8.22.5348 (rev 5318) — a part in sequence 0 used to zero both. Same path, format and units; re-baseline stored expectations.
• ctl/burn timing moved (KSA 2026.8.22.5348, rev 5317) — same interface and CCI Δv, but burn duration, throttle profile and the point at which the FC abandons an auto burn all differ from earlier builds. Don't hard-code a burn duration.
• debug/terrain/.../tessellation/range_m re-baselined (KSA 2026.8.22.5348) — the engine default moved 220 → 50 m and the shader's displacement falloff moved from range×0.1…0.95 to range×0.75…0.975. Read the live value before writing; old tunings look wrong.
• debug/plumetrail/render/trail_color is gone (KSA 2026.9.7.5402). The renderer's global debug tint was removed — trail colour/density/lifetime are per plume-trail template asset now. Reading or writing it answers ENOENT; the other ten render/* fields are unchanged.
• parts/<n>/display_name is the authored template name since KSA 2026.9.7.5402 ("Parachute Bay", "Drogue Radial A"…) and therefore not unique — key on instance_id, never on the name.
• Crashes can spawn debris/fragment vessels (KSA 2026.9.7.5402). Parts now have crash tolerances; a hard contact can split a vessel into <id>_1, <id>_2… fragments plus single-part debris, all of which appear under vessels/ (controllable=0, one part, orbit events frozen), and control moves to the largest controllable fragment on its own. Programs that enumerate vessels/ should tolerate ids appearing mid-flight; there is no debris flag yet.
• ctl/stage also arms parachutes and fires cut modules (KSA 2026.9.7.5402) — the new ParachuteDeploy/ParachuteCut modules are ISequenced, so a stage press behaves exactly like the stage key on a chute bay. animations/<n>/goal on a chute-bay door fights the chute state machine.
• POST /v1/command / gatos/command require vessel_id even for globally addressed actions — the whole camera.*, schedule.*, audio.* families and debug.warp/debug.thug_life_*/… name no vessel, so send "vessel_id": "". Omitting it (or sending null) is 400 EINVAL.
• camera/play/set/stop need [schedule] schedule_enabled as well as camera_enabled — a camera track is a /sim/ctl/schedules player (kind = camera-track, id camera). With scheduling off they answer EOPNOTSUPP; every camera leaf still works.
• A malformed camera track fails on close, which carries no errno — the cp looks like it worked. Read /sim/camera/last_error (or upload over HTTP, where the POST does carry the EINVAL).
• Timeline units are mixed on purpose: timed_batch offsets and ctl/schedules/<id>/t are milliseconds; a track's t/duration and camera/play at are seconds.
• Mass is kg; gravity is μ/r² (never 9.8); ground-referenced velocity is v_cci − ω×r.
• Don't substitute a generic quaternion library for the Body→CCI attitude math — use KSA's exact convention (transform(+X, q) == thrust_direction); see coordinate-frames.md.

Example: "teleport Hunter to a 120 km Earth orbit, Polaris 50 m ahead"

This is fully worked (a copy-pasteable Bun/TS program) in recipes.md §1. The shape: read bodies for Earth's mu/mean_radius, compute r = radius + 120000 and v = sqrt(mu/r), teleport Hunter to CCI [r,0,0, 0,v,0] (equatorial circular), and teleport Polaris to the same orbit advanced in true anomaly by Δθ = 50/r. Everything you need to derive it is in the SPEC + frames doc.

Paint vehicles and EVA kittens

Paint is session-only and explicitly opt-in. Enable vehicle shader paint with echo 1 > /sim/paint/parts/enabled, then set a normalized sRGB triple and enable the desired rule. Whole-vessel example: write 0.15 0.5 1 then 1 under /sim/vessels/by-id/<id>/paint/parts/{color,enabled}. An adjacent part/subpart paint/ directory is the individual override; it requires the parts telemetry gate. Precedence is instance, live vessel, template, global.

Enable safe EVA clones with echo 1 > /sim/paint/kittens/enabled. Shared rules live under /sim/paint/kittens/{shared,materials}; individual rules under the EVA vessel's paint/kitten/{default,materials}. Material names are discovered live (body, fur, helmet, visor, mmu, with numeric suffixes). Disabling either master restores stock but retains rules; write its clear trigger to remove rules. HTTP and MQTT use the normal field mirrors; MCP uses gatos.paint_control or canonical gatos.command.

Draw ground clutter with your own images

/sim/paint/textures re-points a stock ground-clutter texture (rocks, trees, grass, shrubs) at an image you upload. An ordinary sRGB PNG renders as authored — the default faithful bind mode corrects for the clutter shader, so no pre-processing and no colour maths are asked of the user. Session-only, in-memory, restored on unbind. There is no master switch — the feature is inert until something is bound.

The flow is upload → bind → teardown:

cat /sim/paint/textures/clutter          # texture-id slot w h mips used_by ecotypes
                                         # texture-id is a content-relative asset PATH, e.g.
                                         # Textures/Planets/Earth/GroundClutter/Grass_Diffuse.ktx2
cat rock.png > /sim/paint/textures/file/rock.png    # png|jpeg|bmp|hdr|dds|ktx|ktx2
echo 'Textures/Planets/Earth/GroundClutter/Grass_Diffuse.ktx2 rock.png' > /sim/paint/textures/bind      # 'faithful' by default
echo 'Textures/Planets/Earth/GroundClutter/Grass_Diffuse.ktx2 rock.png raw' > /sim/paint/textures/bind  # …or byte-for-byte, like a stock texture
cat /sim/paint/textures/applied          # <id> <file> pending|applied|failed w h mips vram error
echo 'Textures/Planets/Earth/GroundClutter/Grass_Diffuse.ktx2' > /sim/paint/textures/unbind  # one; 'all' = everything
echo 1 > /sim/paint/textures/clear       # same global teardown as 'unbind all'; uploads survive
rm /sim/paint/textures/file/rock.png     # evict an upload (unbinds it first)

Rules a program must follow:

• A plain sRGB PNG just works. Bind's optional third token is the render mode and defaults to faithful, which rewrites the decoded pixels so the image renders at its authored colours, in every biome. Nothing has to be pre-corrected, and a user does not have to be told to author against mid-grey.
• Discover, never guess, target ids. Read clutter and use its first column. A <texture-id> is the asset's content-relative path (e.g. Textures/Planets/Earth/GroundClutter/Grass_Diffuse.ktx2) — install-independent, unique per asset and space-free, so it drops straight into the space-separated line. It is not a symbolic name and cannot be constructed by hand. Bind takes <texture-id> <file> [faithful|raw] — the same shape a bindings row reads back (mode column included), so a listing line can be echoed straight back to re-create a binding.
• The write's success is not the GPU's. bind only queues a Frame-phase command. Poll applied (or status, which carries bound applied retiring vram_bytes revision error) for the real outcome; a failed decode leaves the stock texture drawn and puts the reason in the row.
• A binding replaces a texture asset. used_by > 1 in the clutter listing means every one of the listed ecotypes changes together. Check before binding.
• Errno vocabulary: EINVAL bad name / unparseable line / unrecognised container; ENOENT unknown texture id or missing upload; EBUSY the upload has not committed yet; ENOSPC file, byte, or binding cap; EFBIG per-file cap (mid-write, so the failing write(2) carries it). Images larger than paint_texture_max_dimension are downscaled, not rejected.
• Uploads over the network are the one non-uniform surface. HTTP uses PUT /v1/paint/texture/file/<name>[?offset=N&complete=0|1] and answers 413/EFBIG above the 1 MiB request cap, so chunk any real PNG. MCP uses gatos.paint_texture(operation:"upload") with base64 (which inflates 4/3 against the 24 MiB frame). MQTT carries no binary upload, though it mirrors every control leaf. Everything else rides /v1/fs/paint/textures/... and gatos/sim/paint/textures/... normally; MCP binds with gatos.paint_control(operation:"texture_bind", target:"<texture-id>", file:"<name>", value:0|1) — value is the mode (0 = faithful, 1 = raw).
• raw is the like-for-like option. Clutter diffuse maps are modulation maps, not albedo: the shader computes albedo = 2 * decode(t.rgb, t.a) * mix(meanLum, instanceColor, t.a) / meanLum, so the texel is doubled and alpha is not opacity — it selects sRGB (0) vs linear (1) decoding and the blend toward the per-instance terrain tint. faithful (the default) cancels both: RGB is scaled by 2^(-1/2.2) (white 255 → 186, round-trip error < 0.2%) and alpha is cleared to 0. raw uploads the decoded bytes untouched, so the image is read exactly as one of KSA's own clutter textures — mid-grey 0.5 neutral, biome-tinted at A=255 — which is what you want when replacing a stock texture like-for-like. A decode that is not RGBA8 (some ktx/dds/hdr) cannot be corrected: the faithful binding lands failed in applied, its error naming raw as the fix. Real cutout opacity is a separate opacity slot either way. Mips are generated automatically and are mandatory — without them the texture aliases badly at range.

Spray your own images onto the world (stickers)

/sim/paint/stickers projects an uploaded image onto a vehicle part, onto terrain, or onto the ground clutter standing on it. There is no sticker upload surface — a sticker's image is an entry of /sim/paint/textures/file/, so upload once and reuse it for as many stickers as you like. Session-only, no master switch: with nothing placed the feature is one branch per frame.

cat meow.png > /sim/paint/textures/file/meow.png    # the shared image store (png|jpeg|bmp|...)
echo 'meow.png w=2 h=2' > /sim/paint/stickers/spray  # aim=camera by default; aim=cursor also works
cat /sim/paint/stickers/last        # 0 vessel Kitten-1 part 41 hit 8.42m   (or: no hit within 2000m)
cat /sim/paint/stickers/status      # <id> <image> vessel|body <target> live=0|1 texture=ready|missing|...
cat /sim/paint/stickers/info        # enabled= stickers= stickers_max= live= images= vram_bytes= patch= renderer= max_view_distance_m=
echo 0.4 > /sim/paint/stickers/0/alpha        # size "w h" | depth | rotation | alpha | brightness | visible | image
cat /sim/paint/stickers/0/spec      # meow.png vessel Kitten-1 41 0.1 0.5 -1.4 0 1 0 roll=15 w=2 h=2 d=0.3 alpha=0.4 brightness=1
echo 1 > /sim/paint/stickers/0/remove         # one; 'echo 1 > clear' for all

Place by coordinates instead of aiming, when you want it exact and scriptable:

iid=$(cat /sim/vessels/by-id/Kitten-1/parts/0/instance_id)      # needs telemetry_vessel_parts=1
echo "meow.png vessel Kitten-1 $iid 0 0.5 -1.4 0 1 0 roll=15 w=0.6 h=0.3" > /sim/paint/stickers/place
echo 'meow.png body Mun 12.03 -41.88 heading=90 w=5 h=5'                   > /sim/paint/stickers/place

Rules a program must follow:

• Two grammars, one shared tail. spray takes <image> [aim=camera|cursor] [range=] [roll=] [w=] [h=] [d=] [alpha=] [brightness=]; place takes <image> vessel <vessel_id> <part_iid> x y z nx ny nz [roll=…] or <image> body <body_id> <lat> <lon> [heading=…]. w/h ∈ (0,1000] m default 1; d ∈ (0,100] m default 0.3 on a vessel, 1 on a body; alpha ∈ [0,1] default 1; brightness ∈ (0,8] default 1; range ∈ (0,1e6] m default 2000; roll/heading any finite degrees, default 0. The rotation key is anchor-specific — roll for a vessel, heading for a body — so neither spelling can be used against the wrong anchor. Duplicate, unknown or out-of-range keys are EINVAL on the write(2), before anything reaches the game.
• spray can miss. No hit within range ⇒ ENOENT, and last reads no hit within <range>m. It hits vehicle parts first and the terrain behind them; ground clutter cannot be aimed at (it exists only on the GPU) — but the decal box then projects onto the rock anyway, which is the point. On spray, roll= adds to the orientation that reads upright from where you are, and omitting d= is not the same as passing a number: the default is chosen after the ray says what it hit.
• Ids are the smallest free slot and are reused after remove/clear. Never assume the id of a create — read last, or wait for the paint.sticker_placed event on /sim/events (its detail is the same line, and vessel_id is set only for a vessel anchor).
• <id>/spec is exactly a place line. Echo it back to clone the sticker under a new id. That is also the save game: for d in /sim/paint/stickers/[0-9]*; do cat "$d/spec"; done > ~/specs, plus a copy of the images, replayed at boot. Nothing is persisted host-side.
• Dormant, not deleted. A despawned vessel, a staged-away part or an evicted image gives live=0 (texture=missing) and the entry survives with its settings and spec intact; it draws again when the anchor or the image comes back. Re-uploading an image under the same name hot-swaps it under every sticker using it.
• Sub-parts are valid anchors. part_iid may come from parts/<n>/subparts/<m>/instance_id, and spray naturally anchors to the sub-part it hit — which is why a decal on a gimballing bell or a robotics segment tracks it.
• What the renderer will and will not do. Main viewport only (not crew portraits or extra windows). Beyond paint_stickers_max_view_distance_m (5000 m) a sticker is not drawn at all. The decal fades out at grazing angles (normal cutoff) and its lighting is an approximation of the surface it sits on — use brightness ∈ (0,8] to compensate rather than expecting exact scene lighting. echo 1 > /sim/paint/stickers/debug draws every sticker as a magenta checker of its projection box: use it first whenever a sticker is missing or misplaced.
• Health, not exit status. A create only queues a Frame-phase command. info carries patch=0|1 renderer=idle|active|degraded and last_error carries any renderer/texture fault; the latches are paint.sticker_renderer and paint.sticker_texture in /sim/status/accessors.
• Caps and gates: paint_stickers_max_count (256) — a full registry is EINVAL naming the limit, not ENOSPC. The subtree exists only when both paint_stickers_enabled and paint_textures_enabled are on. Gate string control_enabled + paint stickers.
• Transports. Everything above mirrors to /v1/fs/paint/stickers/... and gatos/sim/paint/stickers/... normally. MCP uses one tool, gatos.paint_sticker(operation:"place"|"spray"|"set"|"remove"|"clear"|"list"|"debug") — set takes id plus exactly one knob (width+height, depth, roll/heading, alpha, brightness, image, or value for visibility) — plus the paint_stickers runtime feature document. Images still upload through gatos.paint_texture.