Systems
The substrate the fleet runs on. Cloud Opus writes the code; local hardware (Strix Halo + Ollama) does bounded non-code work; the harness gates, coordinates, and records everything.
Dated ground-state note; every claim cites file:line or command output.
Human-approved plan: in/out of scope + a runnable verification step.
Branch/worktree + an in-progress TASKS claim (the file lock is live).
A new test observed to FAIL for the right reason.
The stage-3 test passes; no out-of-scope files touched.
Behavioral evidence captured — screenshot/200s or test output.
The ordered test-config gate exits 0 — the actual exit code, never a paraphrase.
Reviewed by a non-implementing agent; author is never the certifier.
Human-merged; the body embeds the evidence chain.
Swaps a ~700KB base64 image for a ~2KB local-OCR description (qwen3.5:2b).
File-level lock: blocks edits to files another session has claimed in TASKS.md.
Routes web fetches through local lfetch (robots-free, localhost, --js/--strip modes).
Local meta-search: SearXNG + Wikipedia, reranked by qwen3.5:2b, optional cited synthesis.
Blocks platform-wrong commands and names the right SSH target (Linux vs Mac).
Logs every tool call to usage.db — the source of the /telemetry page.
Flags any tool response ≥100KB into anomalies.log + a desktop notification.
Passive: logs which .claude/rules globs an edited file matches (consumed on demand).
Opt-in GATE: runs the project's test-config commands and blocks stop on any failure.
Surfaces workspace: rules, skills, active TASKS state, git, cross-machine sync staleness.
- qwen3-coder-next51 GB
- qwen3-coder:30b32 GB
- qwen3:30b-instruct32 GB
- qwen3-vl:8b-instruct9.8 GB
- qwen3-vl:8b6.1 GB
- qwen3.5:9b6.6 GB
- qwen3.5:2b2.7 GB
- nomic-embed-text274 MB
lreviewruns a different model family from the one writing the code — an independent reviewer that does not share the implementer's blind spots.