rduffy.uk — writing

The first half of the budget saga, where the Claude proxy undercounted, overcounted, compensated for its own bugs, and finally matched ccusage after deduplicating repeated JSONL usage entries.

Building a Codex implementation workflow with separate configuration, visible usage, isolated worktrees, and review before merge.

Quên went deeper on memory trust.

How a home-lab AI console grew from a viewer into an instrument panel: live knowledge maps, review coverage, writable ideas, voice capture, supervised agent sessions, approvals, and read-only terminal access.

A Thursday where the platform was already noisy before the console existed: vLLM admission failures, indexer livelock, a retired arch-review pipeline, YouTube ingestion, frontmatter corruption, and the first review-tracker cockpit.

After the first agent audit, I had to give the vault, the agent config, and the memory graph the same custody I expect from production code.

After the overnight audit found 173 de-duplicated bugs, the next job was proving the fixes held in production, catching a benchmark that lied, and writing down what changed.

Most agent-memory demos prove they can remember a name.

Season 4 opens with a dynamic workflow that sent 23 slices / ~46 agent roles through finder-verifier audit work, paused for usage limits, resumed, and came back with 173 de-duplicated bugs.

A month where I tore the platform's repository down to the studs and rebuilt it as a 16-package workspace without ever letting main break — while restoring a CI pipeline that had been silently dead for four days, fighting a month-long embedding-vs-GPU con…

I rebuilt rduffy.uk into one design system and published 20 of my 77 architecture decision records — a sortable, filterable decision log with the context, the call, and what each cost.

A weighted token-budget governor for Claude Code + Codex — and how I found it was metering a number Anthropic never confirmed.

April was the month the platform stopped being one fast thing and started being four uneven ones.

A Monday that converted the whole RAG pipeline from sync to async (fast search 9 seconds to under 150 milliseconds), then watched the platform's own automation find a path traversal in its own code — bracketed by the discovery that an upstream auto-update…

Every AI coding agent signs its commits with a forgeable plain-text line.

A weekend search-quality sprint that took benchmark scores from 0.67 to 0.74, an MCP transport migration that cut cold starts by 96%, ten bugs found along the way, and a CVSS 9.4 Harbor CVE caught by an automation pipeline deployed the same day — with the…

A false-alarm audit on Wednesday patched 37 CVEs, reverted a cluster upgrade in fourteen minutes, then cascaded into ten more tickets on Thursday.

How the second knowledge graph went from intermittent to dependable — Python AST plus FalkorDB plus Leiden community detection — and the moment the Kafka consumer pattern from the journal pipeline generalised into a fleet of code-intelligence workers.

How a single score threshold became the difference between a bot that hallucinates and a bot you can trust — and why the architectural decision to refuse synthesis is the same shape as a circuit breaker, a feature flag, and a dead-letter queue.

I exposed my MCP bridge to the internet so Claude.ai could search my vault remotely.

The automated journal entries were fiction.

3.5 months of silence hid 450 commits, a platform rebrand, 2TB migration, and the transformation from side project to platform engineering discipline.

BT SRE interview Oct 17th morning (strong networking/Linux, exposed K8s gaps), followed by 5.5 hours implementing Anthropic patterns after work.

Working with Claude, we shipped three infrastructure systems in 48 hours.

54 minutes of active work across 14 hours (3 sessions with gap-aware timeline).

18.5 hours across 48 hours.

5 hours debugging temporal search to validate Season 1.

The final episode: Using ConvoCanvas to document its own creation, then publishing with AstroPaper and seamless episode navigation.

From hand-coded Mermaid diagrams to automated architecture visualization.

Saturday morning, October 5, 2025.

K3s was already running.

A weekend implementing ChromaDB, a Monday demoing it at work, and discovering that 1,142 files need more than folders - they need semantic search.

From context window frustration to local control.

From vault design to working code.

The context window error that sparked everything.