>~/posts_
Notes from the work as it happens: daily logs, discoveries and longer studies.
first posts in writing
scroll~/daily
Short logs of what was built, broken and fixed that day.
- The floor every repository has2026-11-21
- A deny list nobody has to write2026-11-20
- bin/sync and the pointer pull request2026-11-20
- Fresh code first, once per session2026-11-20
- hakoniwa, 2026-11-13: plugins and executors2026-11-13
- hakoniwa, day one: from ADRs to a world2026-11-09
- A clock in the prompt cost 62% per turn2026-10-29
- Knowledge round 4: small gains, honest numbers2026-10-27
- Reading a whole message up front2026-10-24
- Triage calibration: 81.5 to 96.22026-09-26
~/discoveries
Something that behaved differently than expected, and why.
- From a repeated problem to a catalogued lint2026-11-21
- The 1 ms animation that never finished2026-11-21
- The theme reveal that jumped at its end2026-11-21
- A gate the model passes by doing the work2026-11-20
- What a knowledge collector may merge by itself2026-11-20
- The payment double that rescanned everything2026-11-09
~/studies
Longer research with measurements, alternatives and a conclusion.
- Evals that decide2026-11-21
- Reading notes: Elixir/OTP and durable workflows2026-11-21
- Reading notes: enrichment, backfills and late data2026-11-21
- Reading notes: ScyllaDB's shard-per-core design2026-11-21
- Sagas by hand vs durable execution2026-11-21
- Elicitation is not approval2026-11-20
- How tool definitions get pinned2026-11-20
- One center, three satellites2026-11-20
- heya playground: what ran on one workstation2026-11-18
- Channel APIs: what a 29-product matrix did not decide2026-11-17
- A plugin hop that did not pay for itself2026-11-13
- Durable execution, explained through nagare2026-11-13
- What exactly-once over a journal costs2026-11-13
- A research method rewritten from its own failures2026-11-12
- Testing with a frozen clock2026-11-09
- Where the wall time went2026-11-09
- How agents talk to CLIs and to each other2026-11-08
- A transcript is an event log2026-11-07
- Why the durable engine sits behind a port2026-11-07
- What MCP solves, and where it stops2026-11-05
- McNemar for paired evals2026-10-27
- Wilson intervals for pass rates2026-10-27
Meanwhile, read the work history or who I am.