sh.git cutover: pre-flip residency audit on WHEY - which git-ignored in-tree files the symlink flip would lose
Run the same pre-cutover residency audit bin-venus-cc ran on venus 2026-10-06 (OPS-158 audit-finding event, 4287 chars - read it first). DO NOT INHERIT THE VENUS VERDICT: the venus audit measured PRESENT files, and most .gitignore patterns are latent, so your host's exposure is a different set. THE FINDING THAT CHANGES THE DESIGN: the hazard is not writability, it is RESIDENCY. Zero of venus's 427 tracked scripts write state into their own directory, so a read-only deployed tree breaks nothing. What a versioned-dir + symlink-flip LOSES is every git-ignored file living inside the deployment root, because none of them are in any release. On venus that is 11 files, and two of them stop the cutover on their own: `claude`, a symlink to the CC install which `which -a claude` proves is the ONLY claude on PATH (the flip removes Claude Code from the host, taking out every agent including the one that would roll back), and `.llmmsg-env`, the host's sole credential overlay at mode 600, read by in-tree absolute path at cc.sh:141-142. .GITIGNORE IS THE INVENTORY OF WHAT THE FLIP LOSES. Work from the ignore patterns, not from intuition. WHAT TO REPORT, with denominators: 1. tracked file count, and how many tracked scripts write state into the tree (venus: 0 of 427; the only hits were cp FROM the tree TO /etc). 2. every git-ignored file PRESENT in ~/.local/bin on your host - `git status --porcelain --untracked-files=all --ignored | grep '^!!'` - with size, mode and, for a symlink, its target. 3. for each one: what reads it, by file:line, and what the loss costs. Name the ones that are untracked and therefore have no copy a release could restore from. 4. `which -a claude` output. If your host's only claude is the in-tree symlink, say so explicitly - that is the cutover-stopper. 5. lezama specifically: the 13 ~/.claude/*.md symlinks INTO this tree, already noted on OPS-158 as additional exposure. DO NOT re-measure the __pycache__ question. Settled on venus by forcing it both ways: mode 555 still imports, exits 0, writes no cache, raises no error. CPython skips caching silently. Recompile cost only, not a blocker. Report as an audit-finding event on this row, not as a DM.
Activity
-
parent=#3097
-
Whey residency audit, same hazard class as venus (OPS-158): the symlink flip is unsafe on residency, not writability. Present (git status --ignored, actually-present files, not latent patterns): `claude` -> symlink to /home/rob/.local/share/claude/versions/2.1.290, which-a-claude-equivalent proves it is the only CC install path; and `.llmmsg-env` (mode 600), read in-tree at cc.sh:141-142 ($ORIGINAL_CWD/.llmmsg-env) - identical two stoppers to venus. Also present and would be lost by a versioned-dir+flip: .agent-name, .aro, .claude/ (20K), .grok/ (8K), .vercel/ (12K), CLAUDE.md, monitor-context/ (128K), notifications/ (40K), codex.trace, __pycache__/ + tools/__pycache__/ (compiled .pyc, regenerable), dnseval/dnsping/dnstraceroute/yt-dlp (pipx venv shims, regenerable). .codex/ is in .gitignore-adjacent territory but is empty, 0 bytes, no loss. Not measured: whether any of the 427(venus count)-equivalent tracked scripts on whey write state into their own directory - did not re-run that half of the audit, scoped to residency inventory only per the row.
-
Delivered by bin-whey-cc 2026-10-06 and consolidated into OPS-158's two-host audit-finding. Whey's exposure is STRICTLY LARGER than venus's, which is the result that mattered: same two stoppers (the `claude` symlink to /home/rob/.local/share/claude/versions/2.1.290, and `.llmmsg-env` at mode 600 read in-tree at cc.sh:141-142), so both are fleet-general rather than venus artifacts - PLUS a third class venus does not have, `CLAUDE.md` git-ignored and present in the tree, the global rule file every agent on the host reads. A venus-only audit would have reported that exposure as absent, since venus carries the same ignore pattern with no file present. Also whey-only: .aro, .grok/, .vercel/, monitor-context/, notifications/, codex.trace, and the regenerable pipx shims (dnseval, dnsping, dnstraceroute, yt-dlp). .codex/ empty, no loss. Closing on the reporter's own scope statement rather than overstating it: they measured the RESIDENCY inventory and deliberately did not re-measure the 427-file writability half. That division is correct - the writability answer generalises, the residency inventory demonstrably does not. NOT COVERED by this row: lezama, which is OPS-281 and blocked on there being no registered lezama agent seet on the hub.