Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
Prompts for Claude Code
1. Run one task (simplest)
Start a new session in the repo for each task and paste:
Use the strata-fixer subagent to complete audit-fixes/tasks/T01-one-model-name.md.
High and medium findings only. Follow the task's "How to check" and "Done when" sections.
Change the file name for each task. Set the effort to medium for the subagent (for example with /model, or in your settings) to keep costs down.
2. Run a whole phase (orchestrated)
Paste into one session:
Read audit-fixes/README.md. Run Phase 1: for each task T01–T08, launch the strata-fixer
subagent with its task file, high and medium findings only. Run tasks whose
dependencies are met in parallel; wait for T01 before T05 and T08.
Each subagent works on its own branch. When all are done, merge the branches into
fix/phase-1 in task order, resolve conflicts, and run npm run check, node --test tests/
and npm run build. Then stop and summarise PROGRESS.md for me.
Use the same prompt for Phase 2 (T09–T10) and Phase 3 (T11–T19).
3. Review after each phase
Run this in a fresh session, ideally on a stronger model:
Review the fix/phase-1 branch against audit-fixes/REVIEW.md and the task files for T01–T08.
For each finding, check the change in the code and, for the states listed, re-capture the
screen and compare it with docs/evidence/audit-2026-10-07/. Report, per task: fixed, not
fixed, regressions. Don't fix anything yourself, just list what to send back.