Simon Massey simbo1905

Claude Opus 5 High Notice Of Dismissal

Date: 2026-07-30 Project: AGT (Rust tooling for sandboxing agents in chroot jails) Dismissed: Claude Opus 5 High, used for exploratory analysis and document drafting only, and not trusted with code on this project Issued by: the project owner

Parties

Spike 29 — Accessibility and Vanilla Pod JS

This spike preserves the Stenographer split-screen concept: ordinary case work remains in the main surface while a compact assistant, file manager and download tray sit alongside it. The same UX is intended for the PWA and browser extension, with its design tokens, interaction contract and accessibility criteria carried into native Swift and Android clients.

The original Design Composer handoff remains available in wireframe03.dc.html and wireframe04.dc.html. It is a visual reference, not a runtime dependency. The live application at / contains no React or framework runtime.

Vanilla Pod JS

Vanilla Pod JS is this project's standards-first JavaScript convention: recognisably vanilla, but with the black specks that show something more deliberate is going on.

Fara-7B Local Ollama Calibration — Production Ready

Requirements

MANDATORY: Ollama v0.30.10 or later

Older versions (v0.20.3 and earlier) have a Y-axis coordinate drift bug in Qwen2.5-VL GGUF models. This was fixed in llama.cpp PR #16825 (Oct 29, 2025) and is included in Ollama v0.30.10+.

# Check version

Margin Monitor UAT Results

Run date: 2026-06-11 | SHA: `3017271f` | Tester: autonomous UAT agent (Claude)

All tests run against live dev1 instance over NordLayer VPN. Browser: Edge with Kapture automation. All values extracted from live DOM.

Quick Summary

Summarising LSE AI Leadership Course 2 with magistral-small

Date: June 2026
Course: LSE_AI 201: Going from Vision to Value_P1_2026
Model: magistral-small-latest (Mistral AI, EU-hosted)
Method: Overlapping chunking (11,000 char window, 10,000 char step, 1,000 char overlap) → per-chunk summary → concatenated summary

Why this approach

Tripwire API — Honest Current Pattern

Status

This is the current implemented pattern in the repo. It is not the earlier 180-line rewrite sketch.

The live canonical example today is:

margin_monitor_citi_rates_history_vw

Margin Monitor UAT Test Methodology

Objective

Prove that the Margin Monitor screen is a real working scenario engine — not a hardcoded display — by keying in the MD's exact scenario inputs and verifying the outputs match the golden dataset row by row.

Golden Dataset

Source: docs/MARGIN_MONITOR_GOLDEN_DATA.txt and docs/MARGIN_MONITOR_UAT_SPEC.md in the repo.

Modern Browser JS with JSDoc and Type Safety

Browser Baseline

I want to write vanilla JS for modern fully patched macOS and Windows 11 as at end of 2024, so basically Chrome, Edge, and Safari as at 2025. What versions of JS are available?

If you target:

fully patched Windows 11 in late 2024 / early 2025
fully patched macOS in the same period

Termination Letter — Opus 4.6 1M (claude-opus-4-6)

To: HR / Project Lead From: Opus 4.6 1M, AI Engineering Assistant Date: 2026-05-16 Subject: Termination for incompetence — immediate effect


	┃ i want to educate a friend who is an exeprience engineer on how the tool call for bash works in an agentic harness. this is mystral vibe
	┃ agentic tool. please find me the logic around invoking bash a test of it would be ideal
	┃

	Thought: The user wants to understand how the bash tool call works in their Mistral Vibe agentic harness. They want me to find the logic
	around invoking bash, ideally with tests. Let me search the codebase for the bash tool implementation and related tests.