The interfaces everyone builds against, so mobile/ and bridge/ can be built in parallel
without blocking each other. Change something here first, then ping the other two —
don't let the wire format drift out of sync with what's actually implemented.
- Mukeshkumar M — app shell,
git/,editor/, escalation logic (when to callbridge/) - Delfi A —
engine/— on-device inference, native swap (CPU today, Hexagon NPU if Option A in../finals-30hr/EVENT_PLAN.mdis taken) - Nambert Jones L —
bridge/— the laptop server, toolchain execution, model escalation
One interface, four things it does. Same model, same interface, whether it's answering
inline or the request got escalated to bridge/ first.
interface CodeEngine {
suspend fun write(instruction: String, fileContext: String): String
suspend fun review(diff: String): List<Finding>
suspend fun debug(stackTrace: String, fileContext: String): String
suspend fun explain(snippet: String): String
}
data class Finding(val line: Int, val severity: Severity, val message: String)
enum class Severity { INFO, WARNING, BUG }Two implementations:
OfflineEngine— regex/heuristic fallback, no model needed. Build the UI against this first.NativeEngine— real GGUF inference via the salvaged llama.cpp JNI bridge (salvaged/native/). JNI export is already bound tocom.iqforge.engine.NativeEngine.nativeGenerate— keep the package/class name if you move the file, or update both sides together.
interface RepoManager {
suspend fun clone(url: String, into: File): Repo
suspend fun pull(repo: Repo)
suspend fun commit(repo: Repo, message: String, paths: List<String>)
suspend fun push(repo: Repo)
}Backed by JGit. This is new — nothing to port from the old repo, which only ever fetched a single diff via the GitHub REST API.
Escalate to the bigger model — POST http://<laptop-ip>:8000/escalate
{ "task": "review" | "debug" | "explain" | "write", "context": "...", "instruction": "...", "effort": "low" | "medium" | "high" | "extra" | "max" }{ "result": "..." }Run a toolchain command — POST http://<laptop-ip>:8000/exec (built — path-restricted to
ALLOWED_EXEC_ROOTS, with a timeout)
{ "command": "npm test", "cwd": "/path/on/laptop/to/the/cloned/repo" }{ "stdout": "...", "stderr": "...", "exit_code": 0 }Ground a prompt with web results — POST http://<laptop-ip>:8000/search
{ "query": "Android 16 KB page compatibility", "max_results": 5 }{
"query": "Android 16 KB page compatibility",
"results": [
{ "title": "...", "url": "https://...", "snippet": "..." }
]
}Search is explicit and disabled by default in the phone UI. If enabled, results are appended to the on-device prompt context; provider failure never blocks the local engine response.
Discover installed real models — GET /models
The response is generated from Ollama's live /api/tags result. Models that are merely named in
configuration but not installed are not returned.
{ "models": [{ "id": "qwen2.5-coder:1.5b", "parameter_size": "1.5B", "quantization": "Q4_K_M", "capabilities": ["completion"], "selected": true }] }Select an installed model — POST /models/select
{ "model": "qwen2.5-coder:1.5b" }The endpoint rejects names that Ollama does not report as installed.
Probe real connectors — GET /connectors
The bridge checks Ollama, authenticated GitHub access through the GitHub CLI OS keyring, and the live web-search provider. A failed probe is returned as unavailable and never as connected.
Field names and casing must match exactly on both sides. If the shape changes, update this file first.
Run and observe the deployment demo — POST /deploy, then GET /deploy/status
The phone creates a deployment review first; this endpoint is called only after its preflight checklist is complete. The bridge currently provides the hackathon build/test/deploy/live simulation. The mobile client polls the status response and turns completed bridge stages into verified checklist ticks.
{ "repo": "iqforge", "commit_sha": "abc123", "message": "Staging deployment from iQForge", "gated": true }{
"deploy_id": "d34db33f",
"stage": "build" | "test" | "deploy" | "live" | "queued",
"stage_label": "Building",
"percent": 25,
"logs": ["Installing dependencies..."],
"done": false,
"stage_complete": false,
"repo": "iqforge",
"commit_sha": "abc123",
"message": "Staging deployment from iQForge",
"started_at": 0
}The initial Start Deployment Review action is phone-side and never releases code. Production
release remains an explicit action in the Deploy stage. With gated: true, the bridge runs only
the Build stage. Each subsequent phone approval calls POST /deploy/advance with the deployment
ID and the exact next stage (test, deploy, then live); skipping a stage is rejected.
The NPU (or its CPU fallback) handles anything that's reasoning over code. bridge/
handles anything that needs an operating system — installing dependencies, running a
build, running a test suite.