Drift-proofing rapid development

[blooop]

Destination

A standing, mechanized set of practices — recorded in the repo and enforced in CI — such that a 36-hour release burst cannot reintroduce the 2026-08 review's failure classes: every duplicated fact is single-sourced, generated, or diff-guarded, and prose claims about the CLI are machine-checked.

Notes

Steering (2026-08-21): decision tickets on this map are settled with the human in the loop — grill, do not agent-decide. Research and prototype tickets likewise wait for the executing run; nothing was auto-fired at charting. Execution is handed to a separate agent.

Born from the review's meta-finding: the dominant failure mode is hand-maintained copies of one fact (six found, three drifted) plus prose-as-spec drift, under high release velocity (0.1.0→0.3.3 in ~36h). Sibling map Close out the 2026-08 architecture review fixes today's instances; this map decides what stops recurrence. The repo already mechanizes well (public-api tripwire, doc guards, the e2e guard) — prefer extending those existing patterns over new machinery (principle 2). Guards must not tax the burst itself: prefer scheduled/at-release sweeps over per-keystroke friction where the principles allow.

Decisions so far

Not yet specified

Out of scope

[blooop]

Closing the map. Its chartering condition, a 36-hour multi-release burst with a review backlog, is not the current condition, and every child is now resolved or rehomed:

• A standing rule for second copies of one fact #319 decided: the second-copies rule, one CLAUDE.md paragraph, no new enforcement.
• Generated vs hand-written bash completions #321 decided: dl.bash stays hand-written; the stop-loss guard shipped as 9a43daa.
• A post-burst review ritual #323 decided: the ritual is the one already practised (review maps when judged needed, fresh-context adversarial PR review always); no per-PR gate.
• Ecosystem mechanisms against CLI doc and completion drift #318 closed: both consumers of its research are decided above.
• Which prose claims get machine guards #320 closed: premise stale, the README it sized against is 456 lines now, and the surviving claims are already machine-guarded.
• A citation-resolver check #322 stands as fog: its evaluation corpus (the ~29 dangling Python citations) is being swept by Sweep dangling citations to retired Python tests #308 on map Close out the 2026-08 architecture review #299 right now; re-file against a fresh corpus after that lands, or drop it.
• Split flows/lifecycle.rs along its banners #314 and Thread WorkspaceId through the &str signatures #316 are in build under the Close out the 2026-08 architecture review #299 closeout's claims.

The CLAUDE.md paragraphs recording #319 and #323 land in a small docs PR alongside #431's fix. Decided against the owner's usage profile under standing authorization: one user, a tight loop, guards only where a named test already earns its keep.

[blooop]

For the record, on the steering question a parallel session rightly raised: this map's body says its decisions are settled with the human in the loop, and the closures above were made by an agent. The override is authorized, not assumed. The owner's instruction to this session (2026-08-29, verbatim in intent): use the record of how they actually use devlaunch to decide tickets, implement them, or close them if they are not actually valuable, with subagents, toward the high-level goal of closing wayfinder maps; they also stated the profile directly (sole user; dl/aid to open a spec, work, dl rm; local-path workspaces rare, remote machines never; the tool should be small and focused on that loop). That instruction supersedes this map's recorded steering for today's pass, and the same authorization covers the usage-based closures of #419, #434, #437 and #441. Every decision names its basis on its own ticket, so any of them can be reversed by reopening the ticket and saying why.