Dicklesworthstone/agentic_coding_flywheel_setup v0.8.0

v0.8.0 — the installer works again

on GitHub

ACFS is installed with curl | bash, so this release is the script itself, not a set of binaries. There are intentionally no downloadable assets — the deliverable is that this tag installs.

Why this release exists

v0.7.0 (2026-06-26) is 432 commits behind and its installer does not work: it fails its own internal integrity check before doing anything. Until now, v0.5.0 — two releases back — was the newest ref that actually installed, so anyone following the README's version-pinned form got a broken install with no obvious way to tell why.

Verified, not assumed

Checked at this exact commit (f659b12a) in a clean ubuntu:22.04 container on linux/amd64, using the exact README command — jsDelivr-first with the || fallback, piped, never saved to a file:

✓ [0/9] Pre-flight validation passed

The following tools will be installed from upstream:
  Oh My Zsh · Powerlevel10k · Bun · Rust · uv · Claude Code (native)
  NTM · MCP Agent Mail · UBS · Beads Viewer · CASS · CM · CAAM · SLB

INTEGRITY failures: 0
install failures:   0

Both --print and --yes --mode vibe --dry-run complete cleanly.

The pipe form matters: saving the script to disk first sends the installer down its local-checkout code path and produces a different, self-inflicted failure that is easy to mistake for the real bug.

What was actually wrong

checksums.yaml / internal_checksums.sh is a committed artifact that must be regenerated whenever a covered file changes. Commits were landing without regenerating it, so the installer broke continuously and healed only by accident. Across five probes the failure count went 2 → 0 → 4 → 2 → 0; the passing run in the middle was an incidental good window, not a fix.

Two distinct failure kinds were being conflated:

• checksum mismatch — the ledger is stale relative to a covered file. Fixed by regeneration; prevented going forward by a pre-commit hook.
• missing — the ledger asserts membership the bootstrap archive can never satisfy. install.sh arrives via curl | bash; it is the thing doing the bootstrapping, so it is structurally not inside the archive it downloads. No amount of regeneration could fix that one.

Bumping ACFS_VERSION in install.sh is itself a change to a covered file, so this release regenerates the ledger in the same commit — otherwise the version bump would have reproduced the exact failure the release exists to fix.

Known limitations

• The pre-commit hook is not retroactive. An existing clone keeps its old .git/hooks/ and is unguarded until the hook is reinstalled. On a machine with many concurrent committers that is most clones.
• The hook catches content drift, not membership drift. A ledger set == archive set assertion is the missing complement.
• A future commit that changes a covered file without regenerating will break main again. This tag is frozen and stays good regardless — which is the point of cutting it.

Tracked in #345 and #346.