LAYER 01embedded
APK shell
Rootfs tarball and default extensions embedded in app assets
aura-harness
by
Joseph Itoro
Joseph ItoroBuild the following to your best abilities
Comments (0)
Sign in to leave a comment
Project Tasks
32 planning tasks#1
Generate system requirement document
2m 21s0.2 cr used
Done#2
Generate personas & user flows
0m 20s0.2 cr used
Done#7
Create flow for Android Developer (Agent Operator)
0m 18sCredits in parent
Done#8
Create flow for Full-Stack Web Developer
0m 17sCredits in parent
Done#9
Landing
35m 3sCredits in subtasks
Done#34
Fix Landing review findings: Hero core covers readable letters in the FUSION CORE… (+1 more)
0m 49s0.4 cr used
Done#32
Fix Landing review findings: Q1 as reviewed now: Current… (+1 more)
4m 21s0.4 cr used
Done#31
Fix Landing review finding: Q1 as reviewed now: Regression in…
2m 55s0.4 cr used
Done#30
Fix Landing review findings: Q1 as reviewed now: Current Landing Fusion Core Hero.jsx… (+1 more)
2m 21s0.4 cr used
Done#29
Fix Landing review finding: frontend/src/components/Landing Fusion Core Hero.jsx renders…
1m 31s0.4 cr used
Done#27
Landing / Navigation
0m 0s0.85 cr used
Done#22
Landing / Fusion Core Hero
1m 4s0.85 cr used
Done#23
Landing / Runtime Stack
1m 16s0.85 cr used
Done#24
Landing / Telemetry Strip
1m 9s0.85 cr used
Done#25
Landing / Capability Grid
1m 38s0.85 cr used
Done#26
Landing / Initiate Harness
3m 45s0.85 cr used
Paused#28
Navigation
0m 21sCredits in parent
Done#33
Retry First Page
3m 37s0.4 cr used
Done#10
Login
Credits in subtasks
Backlog#11
Agent Workspace
Credits in subtasks
Backlog#12
Models
Credits in subtasks
Backlog#13
Agent Chat
Credits in subtasks
Backlog#14
Permissions
Credits in subtasks
Backlog#15
Editor
Credits in subtasks
Backlog#16
Terminal
Credits in subtasks
Backlog#17
Web Server
Credits in subtasks
Backlog#18
Web Preview
Credits in subtasks
Backlog#19
Database Browser
Credits in subtasks
Backlog#20
Scaffolding
Credits in subtasks
Backlog#21
Snapshots
Credits in subtasks
Backlog#5
Architecture
0.2 cr needed
Backlog#6
Workspace task plan
0.2 cr needed
Backlog
FIVE LAYERS FROM APK TO LAMP
The aura-harness runtime descends from an embedded APK shell through the native Rust core, the Debian userland and the model providers, landing on a local LAMP stack served at localhost:8080.
LAYER 02native
Rust core
aarch64-linux-android reasoning, tool executor and Fusion state via JNI/UniFFI
LAYER 03extracted
Linux userland
Debian AArch64 rootfs extracted to /data/data/<package>/files/usr on first launch
LAYER 04routed
Model provider
OpenAI-compatible REST, Anthropic Messages, Gemini, and local llama.cpp GGUF
LAYER 05localhost:8080
LAMP stack
Apache2 + PHP-FPM + MariaDB orchestrated behind one Start Web Server action
Locallocalhost:0
Fusion telemetry
RSS0.0 MBTokens0.0x Loop0.0 ms
0.0MB
Memory22–30 MB envelope · peak 30 MB
0.0x ↓
Token budgetfusion prefix · O(1) growth
0tools
Tool callsnative command loop
Everything the harness runs on, instrumented
The accepted runtime surface of aura-harness — bounded memory, a 17-skill agent ecosystem, an embedded Debian AArch64 userland, a unified model provider layer and a persistent foreground service — presented as one lit cluster.
Fusion Model
Immutable prefix, bounded patches
Instead of accumulating conversational history, the Fusion Model holds an immutable system prefix and applies explicit structured JSON state patches per step, so prompt size stays flat as a session grows.
Token use14.8x ↓
Prompt growthO(1)
Core RSS22–30 MB
Agent modes
PLAN / AGENT switch
PLAN keeps the panel cyan and read-only: file-write tools stay struck through until you unlock the loop.
- fs.readread
- fs.writewrite
- fs.patchwrite
- shell.execread
- git.commitwrite
- web.fetchread
17-skill ecosystem
Self-healing tool loop
Nineteen built-in tools grouped into six skill families, driven by autonomous dependency resolution and self-healing edits with undo.
- Fileread, write, patch
- Shellexec, pipe, cwd
- Gitstatus, diff, commit
- Webfetch, parse, browse
- Dependencyresolve, install
- Code intelligenceAST, LSP, refactor
Linux userland
Debian AArch64 rootfs
Extracted to /data/data/<package>/files/usr on first launch — no root, no network download.
| Binary | Version | Role |
|---|---|---|
| php | 8.2+ | PHP-FPM application runtime |
| composer | 2.x | PHP dependency manager |
| mariadb | 10.x | Relational database server |
| apache2 | 2.4 | HTTP server, localhost:8080 |
| node | v22+ | Node.js JavaScript runtime |
| python3 | 3.12+ | Python interpreter |
| gcc | 12+ | AArch64 C/C++ toolchain |
Model provider
One abstraction, four backends
A unified Rust ModelProvider routes tool/function calls across cloud and local backends behind a single contract.
- RESTOpenAI-compatible REST endpoints
- MSGAnthropic Messages API
- GEMGemini generative API
- GGUFLocal llama.cpp GGUF models
- phpMyAdmin
- localhost:8080
Editor
LSP-aware workspace
CoreCodeMirror 6 editing surface
Languages40+ language highlighting
DiagnosticsLSP errors, warnings and hovers
CompletionGhost-text inline completion
Persistence
Zero-Termux foreground service
AgentForegroundService keeps the harness alive across aggressive out-of-memory background termination, holding persistent loops without any Termux dependency.
- START_STICKYService restarts after an OOM kill
- PARTIAL_WAKE_LOCKCPU stays awake for the loop
- WIFI_MODE_FULL_LOW_LATENCYStreaming provider transport
On-device agent harness
Fusion Core
aura-harness keeps the agent loop, the AArch64 Rust core and the embedded Linux userland running entirely on the device, so the harness is initiated from the phone it already lives on.
Initiate Harness
Opens the accepted access surface at /login.
Diagram Error: Diagram could not be rendered.
FIVE LAYERS FROM APK TO LAMP
The aura-harness runtime descends from an embedded APK shell through the native Rust core, the Debian userland and the model providers, landing on a local LAMP stack served at localhost:8080.
LAYER 01embedded
APK shell
Rootfs tarball and default extensions embedded in app assets
LAYER 02native
Rust core
aarch64-linux-android reasoning, tool executor and Fusion state via JNI/UniFFI
LAYER 03extracted
Linux userland
Debian AArch64 rootfs extracted to /data/data/<package>/files/usr on first launch
LAYER 04routed
Model provider
OpenAI-compatible REST, Anthropic Messages, Gemini, and local llama.cpp GGUF
LAYER 05localhost:8080
LAMP stack
Apache2 + PHP-FPM + MariaDB orchestrated behind one Start Web Server action
Locallocalhost:0
Fusion telemetry
RSS0.0 MBTokens0.0x Loop0.0 ms
0.0MB
Memory22–30 MB envelope · peak 30 MB
0.0x ↓
Token budgetfusion prefix · O(1) growth
0tools
Tool callsnative command loop
Everything the harness runs on, instrumented
The accepted runtime surface of aura-harness — bounded memory, a 17-skill agent ecosystem, an embedded Debian AArch64 userland, a unified model provider layer and a persistent foreground service — presented as one lit cluster.
Fusion Model
Immutable prefix, bounded patches
Instead of accumulating conversational history, the Fusion Model holds an immutable system prefix and applies explicit structured JSON state patches per step, so prompt size stays flat as a session grows.
Token use14.8x ↓
Prompt growthO(1)
Core RSS22–30 MB
Agent modes
PLAN / AGENT switch
PLAN keeps the panel cyan and read-only: file-write tools stay struck through until you unlock the loop.
- fs.readread
- fs.writewrite
- fs.patchwrite
- shell.execread
- git.commitwrite
- web.fetchread
17-skill ecosystem
Self-healing tool loop
Nineteen built-in tools grouped into six skill families, driven by autonomous dependency resolution and self-healing edits with undo.
- Fileread, write, patch
- Shellexec, pipe, cwd
- Gitstatus, diff, commit
- Webfetch, parse, browse
- Dependencyresolve, install
- Code intelligenceAST, LSP, refactor
Linux userland
Debian AArch64 rootfs
Extracted to /data/data/<package>/files/usr on first launch — no root, no network download.
| Binary | Version | Role |
|---|---|---|
| php | 8.2+ | PHP-FPM application runtime |
| composer | 2.x | PHP dependency manager |
| mariadb | 10.x | Relational database server |
| apache2 | 2.4 | HTTP server, localhost:8080 |
| node | v22+ | Node.js JavaScript runtime |
| python3 | 3.12+ | Python interpreter |
| gcc | 12+ | AArch64 C/C++ toolchain |
Model provider
One abstraction, four backends
A unified Rust ModelProvider routes tool/function calls across cloud and local backends behind a single contract.
- RESTOpenAI-compatible REST endpoints
- MSGAnthropic Messages API
- GEMGemini generative API
- GGUFLocal llama.cpp GGUF models
- phpMyAdmin
- localhost:8080
Editor
LSP-aware workspace
CoreCodeMirror 6 editing surface
Languages40+ language highlighting
DiagnosticsLSP errors, warnings and hovers
CompletionGhost-text inline completion
Persistence
Zero-Termux foreground service
AgentForegroundService keeps the harness alive across aggressive out-of-memory background termination, holding persistent loops without any Termux dependency.
- START_STICKYService restarts after an OOM kill
- PARTIAL_WAKE_LOCKCPU stays awake for the loop
- WIFI_MODE_FULL_LOW_LATENCYStreaming provider transport
On-device agent harness
Fusion Core
aura-harness keeps the agent loop, the AArch64 Rust core and the embedded Linux userland running entirely on the device, so the harness is initiated from the phone it already lives on.
Initiate Harness
Opens the accepted access surface at /login.
Diagram Error: Diagram could not be rendered.
No comments yet. Be the first!