Handbook: meta/plans/
owns this directory's conventions.
This directory holds documents that are not part of the
description of how the system works today:
plans for work that is not finished,
records of plans that came true,
and records of designs that never will.
This file is what names the contents,
so nothing under plan/ is an orphan.
What is current is owned by the
handbook —
walking down from its root finds the leaf for any topic —
with .claude/skills/ holding the routines'
machine-loaded playbooks.
A document here may describe something that has since changed;
when the two disagree, the handbook leaf is right.
| To read about | Read |
|---|---|
Simplifying the vendoring pipeline: one verdict store, tooling from main, a docs tree |
PLAN-vendoring-simplification.md |
Implementing dbSendQueryArrow() and the DBI Arrow API |
PLAN-dbSendQueryArrow.md |
| A producer thread for streaming results: overlapping engine production with R-side conversion, and spillable materialized results | PLAN-streaming-thread.md |
| Making Ctrl+C reach a DuckDB call blocked in a network wait, and why the obvious escalation is refused | PLAN-query-cancellation.md |
| Opening the v2.0 series out of the preview line it forks from, and what becomes of the preview line | PLAN-v2-series-open.md |
The handbook states the outcome as fact; these are kept for the reasoning that got there.
| To read about | Read |
|---|---|
| The CRAN-safe storage-location policy, and the work that implemented it | done/PLAN-storage-locations.md |
Kept for the reasoning, not as a description of the system; each says so in its own first lines. They live in their own directory because the distinction is the point: a plan may still come true, a superseded design never will.
| To read about | Read |
|---|---|
The agentic-loop design that preceded the series loop (its measurements now live in experiments/) |
superseded/vendoring-loop.md |
Not a plan and not a design: a record of something that happened, kept because closing it would otherwise lose a fact. Each says in its first lines what it is and which document owns its topic today.
| To read about | Read |
|---|---|
What the TIMESTAMPTZ session-TimeZone fix (#2401) was verified against, scenario by scenario, and the limitations it left |
history/2026-05-timestamptz-icu.md |
| Findings verified against the code by the closed handbook wave, and the issues its defects became | history/2026-08-handbook-wave-salvage.md |
| The glue adaptations the August 2026 forward of all three series replayed, where each modification was placed, and what the run found out about the routine | history/2026-08-series-forward-glue.md |
A plan goes in plan/ as PLAN-<topic>.md, and opens with a line saying
what it is and which document owns its topic today; add a row above.
A file under plan/ that this table does not name is an orphan, which is
the one thing the tree's structure exists to prevent.
Where a plan leaves depends on what happened to it:
one that came true moves to done/, one overtaken by events to
superseded/, and its row moves with it.
A record of work that is over rather than a proposal for work goes in
history/, named for when it happened, and never moves again.
A measurement is not a plan at all — it belongs in
experiments/, one directory per run.