Side page · Generated, then verified

What I built
this week

August 24–30

The week's largest single artefact was a stacked-branch recovery merge on Hold the Line that finally landed the milestone-one command-loop work alongside the V2 Relief reboot's scenario, Living Front, visual production pass, and Cockpit revisions. In parallel, the portfolio site grew a measured second agent — Codex now sits alongside Claude Code on numbers and week, with a corrected counting rule verified against an independent source — and an editorial hierarchy pass reshaped the home page and the numbers ledger. Ask-gpt reached Windows and now ships as a Claude Code plugin after several rounds of adversarial review.

01 · HOLD THE LINE

The stacked branch chain carrying milestone one recovered into a single merge that landed the command-loop stages and the V2 Relief reboot together. Playtest-driven work made assaults read as waves of marks emerging from captured ground, moved allegiance into blue and red trench-map ink, and reframed the battlefield around a Command Phase where a paused snapshot receives orders that commit as one replayable bundle. The V2 reboot added a scripted Relief scenario with the game's first win condition, turned neighbouring formations into combatants with mechanical levers, executed the long-deferred visual production pass, and put the command card on the map.

19 commits · 6 PRs merged · 364.3M tokens read

02 · THIS SITE

The main thread was giving the site a measured second AI engine and reshaping how it presents itself. Codex was added alongside Claude Code with a corrected token-counting rule verified against a source the collector never reads, per-agent breakdowns on numbers and week, and a build story for a Codex-only project. An editorial pass quieted the chrome, promoted ask-gpt into a featured band, cut the hero, and re-sorted the numbers ledger largest-first by tokens. Smaller work added a plaintext guard on generated summaries, retired the principles route, rebuilt the social card from HTML, raised the day-job section to the ceiling Tom supplied, and dropped the OpenAI Sites deploy target.

45 commits · 8 PRs merged · 13 sessions · 257.0M tokens read

03 · ASK-GPT

Windows support arrived through repeated adversarial-review rounds: a filesystem-safety layer for symlinks and locking, owner-only protection carried through a legacy thread migration, and slash commands fixed so the tool now ships as a Claude Code plugin. Alongside that, a per-project ledger and thread identity keyed by repo root landed, an independent full review closed three security findings and hardened the ledger and prompts against ReDoS and boundary forgery, and a claims-cleanup pass retired stale documentation and bounded Codex compatibility for the beta. The reviewer itself was hardened against repository-borne prompt injection.

103 commits · 1 PR merged · 1 session · 245.6M tokens read

04 · READING HELPER

An engine-review pull request went up for review this week and remains open; nothing landed on the main branch.

0 commits · 0 tokens read

THE METER

commits
167
merged PRs
15
tokens read
866.9M
tokens written
1.8M
model calls
3,139
working sessions
14

97.9% of model input this week was cached context being read again.

View all weeks