Skip to content

docs(rig): move the verify queue out of tree to $AC_HOME - #422

Merged
mkovero merged 2 commits into
mainfrom
rig-queue-restore
Sep 3, 2026
Merged

mkovero merged 2 commits into
mainfrom
rig-queue-restore

Conversation

@mkovero

@mkovero mkovero commented Sep 1, 2026

Copy link
Copy Markdown
Owner

Follow-up to #421, which restored two blocks a branch had silently overwritten. This removes the mechanism that allowed it.

The queue is per-rig coordination state, not per-branch state. Living in the repo gave every branch its own copy, and any branch could overwrite it during a routine docs sync — twice, in #421's case, losing the only record of how to close two open QA findings.

Moved to $AC_HOME/rig-verify-queue.md, beside handoff/ and session/, which are out of tree for the same reason. Content is byte-identical to the file this deletes, restored blocks included:

d8d651c9aff30e79984bc352a52e3dc8b6dc997953583f0b8d736fa240af7a55  $AC_HOME/rig-verify-queue.md
d8d651c9aff30e79984bc352a52e3dc8b6dc997953583f0b8d736fa240af7a55  main:rig/rig-verify-queue.md

The file is already in place on this machine; merging this only removes the in-tree copy and repoints the references.

References repointed

TESTING.md, .agents/qa.md, .agents/rig.md (×3), CLAUDE.md, audit/rig-session-3/negative-lag-rule.md, work/rig/rig-verify-125-results.md, work/planning/usable-ir-path-issues.md (×2). No rig/rig-verify-queue.md string remains in the tree.

Two of those cited rig-verify-queue.md:339-366. Line numbers cannot survive a file with no revisions to diff against, and they were already wrong — the block they pointed at was deleted and restored at a different offset. Both now name the block.

$AC_HOME is now a git repo

The first draft of this PR said the move costs recoverability, since nothing versioned $AC_HOME. That is no longer true, and the third commit corrects the two notes that claimed it.

$AC_HOME is its own repository as of c2e3ef6 (local, no remote). It tracks the ~1.4 MB of markdown — handoffs, session records, the queue — and ignores:

ignored why
wt/ git worktrees of this repo, ~30 GB, already versioned in ac/.git; a worktree's .git is a pointer file that would commit as a broken tree
target/, target-*/ cargo output, ~42 GB — includes bin/common.sh's per-branch AC_TARGET dirs and the nested target/pr-420/, so both patterns are unanchored and match at any depth
log/ 136 MB of append-only agent transcripts, regenerated per run, and the likeliest place for a credential to sit in cleartext
scratch-*/ per-run scratch, mode 700

So an overwrite of the queue is still recoverable, just from a different repo: git -C "$AC_HOME" log -p -- rig-verify-queue.md. What is not recoverable is an edit never committed there — nothing commits automatically. .agents/rig.md now says to commit at the end of a rig session rather than the start of the next one, and to mark blocks executed rather than deleting them.

No remote, so this protects against an overwrite or a deletion, not against losing the disk. $AC_HOME/README.md says so rather than leaving it to be found out.

Docs only. No code, no test, no wire format touched.

🤖 Generated with Claude Code

https://claude.ai/code/session_017tb982HsRpbLenwEq2cay1

mkovero and others added 2 commits September 1, 2026 11:33
The queue is per-rig coordination state, not per-branch state, but living
in the repo gave every branch its own copy — and the previous commit
shows what that costs: two branches each overwrote their own queued
measurement with main's copy during a routine docs sync, and the entries
lost were the only record of how to close an open QA finding.

Moves it to `$AC_HOME/rig-verify-queue.md`, alongside `handoff/` and
`session/`, which are out of tree for the same reason. Content is
byte-identical to the version this commit deletes (sha256
d8d651c9…af7a55), restored blocks included. One copy, no branch can
diverge from it, and the failure mode the previous commit repairs cannot
recur.

Every in-tree reference is repointed at `$AC_HOME/rig-verify-queue.md`:
TESTING.md, `.agents/qa.md`, `.agents/rig.md` (three), CLAUDE.md,
`audit/rig-session-3/negative-lag-rule.md`,
`work/rig/rig-verify-125-results.md`, and
`work/planning/usable-ir-path-issues.md` (two).

Two of those cited `rig-verify-queue.md:339-366`. Line numbers cannot
survive a file that no longer has revisions to diff, and they were
already wrong — the block they point at was deleted and restored at a
different offset. Both now name the block instead.

`.agents/rig.md`'s entry says what the move costs as well as what it
buys: nothing versions `$AC_HOME`, so a deleted entry is gone with no
`git show` to recover it. That is the one real regression here, and the
reason the role note now says to mark blocks executed rather than remove
them. CLAUDE.md carries the same warning for `$AC_HOME` generally.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017tb982HsRpbLenwEq2cay1
…imed

The two notes added a commit ago said nothing versions `$AC_HOME`, so a
deletion there is permanent. That was true when written and is no longer:
`$AC_HOME` is now its own git repository (`c2e3ef6`), tracking the ~1.4 MB
of markdown and ignoring the git worktrees under `wt/`, the cargo target
directories — `target/`, the per-branch `target-<tag>/`, and the nested
`target/pr-420/` — the agent transcripts under `log/`, and the mode-700
scratch directories.

So the trade this PR described is smaller than stated: moving the queue
out of `ac` still means no branch can diverge from it, and an overwrite
is still recoverable — just from a different repo. What does not survive
is anything edited and never committed there, since nothing commits
automatically, which is what the rig note now says.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017tb982HsRpbLenwEq2cay1
@mkovero
mkovero merged commit 4763916 into main Sep 3, 2026
@mkovero
mkovero deleted the rig-queue-restore branch September 3, 2026 04:03
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant