fix: guard git push by shape so feature-branch pushes run unattended

.claude/settings.json

@@ -15,7 +15,8 @@
     "filesystem": {
       "allowWrite": [
         ".",
-        "~/GitHub/bess-manager"
+        "~/GitHub/bess-manager",
+        "~/.claude/jobs"
       ]
     }
   },
@@ -87,8 +88,43 @@
     ],
     "ask": [
       "Bash(git push)",
-      "Bash(git push *)",
-      "Bash(git -* push*)",
+      "Bash(git push *--force*)",
+      "Bash(git push *-f)",
+      "Bash(git push *-f *)",
+      "Bash(git push *--delete*)",
+      "Bash(git push *-d)",
+      "Bash(git push *-d *)",
+      "Bash(git push *--mirror*)",
+      "Bash(git push *--prune*)",
+      "Bash(git push *--tags*)",
+      "Bash(git push *--all*)",
+      "Bash(git push *--branches*)",
+      "Bash(git push * +*)",
+      "Bash(git push *refs/tags/*)",
+      "Bash(git push * v*)",
+      "Bash(git push beta*)",
+      "Bash(git push * beta*)",
+      "Bash(git push *:beta*)",
+      "Bash(git push *refs/heads/beta*)",
+      "Bash(git -* push)",
+      "Bash(git -* push *--force*)",
+      "Bash(git -* push *-f)",
+      "Bash(git -* push *-f *)",
+      "Bash(git -* push *--delete*)",
+      "Bash(git -* push *-d)",
+      "Bash(git -* push *-d *)",
+      "Bash(git -* push *--mirror*)",
+      "Bash(git -* push *--prune*)",
+      "Bash(git -* push *--tags*)",
+      "Bash(git -* push *--all*)",
+      "Bash(git -* push *--branches*)",
+      "Bash(git -* push * +*)",
+      "Bash(git -* push *refs/tags/*)",
+      "Bash(git -* push * v*)",
+      "Bash(git -* push beta*)",
+      "Bash(git -* push * beta*)",
+      "Bash(git -* push *:beta*)",
+      "Bash(git -* push *refs/heads/beta*)",
       "Bash(gh api)",
       "Bash(gh api *)",
       "Bash(gh pr merge*)",

CLAUDE.md

@@ -239,16 +239,17 @@ frontend/node_modules` first) rather than installing through it. Re-running
 already contains.** Prompts are the cost, not the safety: a stalled autonomous
 run is a guaranteed loss, while anything the sandbox bounds is recoverable.
 `rm`, `git reset --hard`, `rebase`, `merge`, `git branch -D`, `git worktree
-remove` all run unattended. **Every `git push` asks**, `--force-with-lease`
-included — see below for why that one is not carved out.
+remove` all run unattended. **A `git push` naming a feature branch runs
+unattended too**; a push that forces, deletes a ref, publishes a tag, or targets
+`main`/`master`/`beta` asks — see below for how that line is drawn.
 
 What still asks is one closed list, and every entry is there because the
 **sandbox cannot contain it** — it bounds the filesystem, not the network, and
 not `.git`'s own recovery data:
 
 | Category | Rules |
 |---|---|
-| Escapes to GitHub | **every `git push`**, **every `gh api`**, `gh pr merge`, `gh release create` / `edit` / `delete` / `delete-asset` / `upload`, `gh repo edit`, `gh secret`, `gh workflow run` |
+| Escapes to GitHub | **a `git push` that forces / deletes a ref / publishes a tag / targets `main`, `master` or `beta`**, **every `gh api`**, `gh pr merge`, `gh release create` / `edit` / `delete` / `delete-asset` / `upload`, `gh repo edit`, `gh secret`, `gh workflow run` |
 | Destroys the recovery mechanism | `git gc`, `git prune`, `git repack`, `git maintenance`, `git reflog expire`, `git reflog delete`, `git update-ref`, `git tag -d` / `--delete` / `-f` |
 | Leaves the user boundary | `sudo` |
 
@@ -301,24 +302,102 @@ because leaving `rm` and `reset --hard` unattended is only defensible while the
 object database and reflog can recover them — a `gc --prune=now` that ran
 unprompted would remove the ground that argument stands on.
 
-**`git push` and `gh api` are guarded bluntly, and that is deliberate.** Both
-were enumerated by shape first, and both leaked, twice:
+**`gh api` is guarded bluntly, and that is deliberate. `git push` no longer is,
+and that took three tries to get right.** Both were enumerated by shape first,
+and both leaked:
 
 ```
 git push origin main --force          # --force not adjacent to `push`
 git push origin +beta-release-9.9     # force via refspec
 git push origin --delete release-X.Y  # destroys a shared ref
 git push origin v9.9.0                # publishes a release tag
+git push origin HEAD:main             # protected ref via colon refspec
+git push origin refs/heads/main       # bare source-side refspec: no space, no colon
+git push origin refs/tags/v1.2.3      # same, for a TAG: `*:refs/tags/*` saw only the colon form
+git push --all                        # every local branch, main included
+git push --branches origin            # exact synonym for --all
 gh api repos/o/r/pulls/N/merge -X PUT # merges, bypassing `gh pr merge`
 gh api <path> -f key=val              # any -f/-F makes it a POST
 ```
 
-The marker sits at an arbitrary argument position, and **prefix globbing cannot
-reach it.** Narrowing these two back to specific forms re-opens every line
-above, so `quality-check.sh` requires the blanket spelling rather than merely
-"some push rule exists". The cost is one prompt at Step 9 per issue, and a
-prompt during release work — which the release skill requires explicit approval
-for anyway.
+**The last three were found by the Stage 4 review of the PR that fixed the
+first five**, which is the point worth remembering: each round of this
+enumeration has been incomplete, and each hole was invisible to the round
+before. `refs/heads/main` slipped because every protected-branch rule keyed off
+the character *before* the branch name — a space in `* main`, a colon in
+`*:main` — and in `refs/heads/main` that character is `/`. The guards are now
+spelled `*refs/heads/main*` without the colon, which subsumes the
+destination-side form rather than sitting beside it. `--all` slipped because
+its neighbours `--mirror`, `--prune` and `--tags` were all covered and it
+simply was not; `--branches` is its documented synonym (`git push --help`:
+"--all, --branches") and was added in the same pass rather than waiting to
+become the fourth leak.
+
+`refs/tags/` had the same colon-shaped blind spot as `refs/heads/`, and it is
+the one entry here that **branch protection can never replace**: a tag is not a
+branch, so no protected-branch setting covers publishing a release tag. Both
+guards are now spelled without the colon (`*refs/tags/*`, `*refs/heads/main*`),
+each subsuming its destination-side form rather than sitting beside it.
+
+The marker sits at an arbitrary argument position. The first enumeration was
+**prefix-anchored** — `Bash(git push --force*)` — which genuinely cannot reach
+position 3, so it leaked every line above and was replaced by a blanket
+`Bash(git push *)`. That blanket rule was justified here by the claim that "a
+prefix glob cannot reach it", **and that claim was wrong**: `*` compiles to `.*`
+under `re.fullmatch` and spans spaces, so `Bash(git push *--force*)` matches
+`git push origin main --force` perfectly well. The very next paragraph about
+greedy globs said so all along; the two contradicted each other for four review
+rounds.
+
+**The protected-ref half of that enumeration is now GONE, and branch protection
+replaces it.** `enforce_admins` is `true` on `origin/main` and `beta/main`, so
+GitHub refuses a push to `main` server-side — including from the owner token
+every agent uses. That is a *provable* guarantee; an enumeration of command
+spellings is only ever a list of the ones somebody thought of, and that list
+leaked three times, the last two spellings (`git push origin refs/heads/main`,
+`git push --all`) found by the review of the PR meant to fix it. A fourth round
+was the expected outcome of keeping it.
+
+So `git push origin main`, `HEAD:main`, `refs/heads/main`, `-u origin main`
+and the `master` equivalents now run unattended and are rejected by GitHub
+instead of by a prompt.
+
+**`enforce_admins: true` is therefore LOAD-BEARING.** Turn it off and this repo
+has no guard on pushing to `main` at all. `quality-check.sh` asserts it on both
+remotes for exactly that reason — a false answer is an error naming the fix, an
+unreadable one is a warning, because a CI token cannot read branch protection
+and "I could not check" is not "it is off". If it ever fires, either re-enable
+the setting or restore the protected-ref patterns; do not silence the check.
+
+What stays guarded is what branch protection does **not** cover — and each of
+these is a real gap, not belt-and-braces:
+
+- **force in any position**, `--force-with-lease` included, and refspec `+`
+- **ref deletion**, `--mirror`, `--prune`, `--tags`, `--all`, `--branches`
+- **tags**: `refs/tags/` on either side, and `* v*`. A tag is not a branch, so
+  no protected-branch setting will ever cover publishing a release tag
+- **the `beta` remote**, and `beta-release-*` branches on origin. Only `main`
+  is protected, so these are still ours
+
+A push naming a feature branch runs unattended, which is what
+`implement-issue` Step 9 and `sweep-prs` actually do.
+
+`gh api` keeps its blanket rule because it is **not** separable this way: any
+`-f`, `-F` or `-X` turns a read into a mutation, so there is no lexical marker
+that isolates the safe subset. If it starts stalling runs, the fix is an
+allow-list of read-only paths, not a shape guard.
+
+`quality-check.sh` pins all of this by COMMAND STRING in three lists, never by
+rule name. Adding a rule that "looks right" without adding its command string is
+how every previous version of this section went stale.
+
+**But the gate proves "the strings we thought of are covered", not "no push
+spelling reaches a protected branch"** — and after three incomplete rounds that
+distinction is the honest summary of where this stands. The durable fix is
+normalisation: resolve a refspec's effective destination ref before matching,
+instead of pattern-matching the raw command string. That is tracked separately;
+until it exists, treat any newly-discovered spelling as expected rather than
+surprising, and add it to `MUST_BE_GUARDED` first.
 
 **Patterns match the command as written — prefix globbing, no normalisation.**
 This is the single biggest source of rules that look right and match nothing,
@@ -503,6 +582,19 @@ redundant:
   one from outside. `verify-sandbox.sh` skips its symlink check outside a
   linked worktree rather than reporting a PASS that proves nothing.
 
+- **`~/.claude/jobs` in the same `allowWrite`** — `claude agents --json` reads
+  the session list from there, and a sandboxed call silently TRUNCATES it
+  rather than failing: it returned **1 session where the truth was 12**, so
+  every other session read as dead. That is not cosmetic. `backlog-rhythm.sh`
+  keys `resume_implementation` off "no live session", and with a truncated
+  listing it told the maintainer to re-enter a worktree that a live session was
+  actively working — a second session on one branch, against commits the advice
+  itself calls the only copy. `implement-issue` Step 0 depends on the same
+  reading before it touches a resumed branch.
+
+  The failure mode is the dangerous one: not an error, an under-count. Nothing
+  about a short list looks wrong.
+
 `sandbox.excludedCommands` is **not** used and is not needed. It was tried in
 both project and user settings while the four knobs above were missing, appeared
 to do nothing, and is now moot.
@@ -555,22 +647,22 @@ and won't remove a worktree holding uncommitted or untracked files). A second
 prompt there buys nothing and costs a stall on every run — `implement-issue`
 Step 4's prune loop alone would have hit ~24 of them.
 
-`git push --force-with-lease` **used to** be excluded on the same reasoning, and
-is not any more. It is a push, and the push guard has to be blanket (above): the
-only spellings that would exempt it — `Bash(git push --force-with-lease*)` — are
-prefix-anchored, so they cannot exempt `git push origin main
---force-with-lease`, while any pattern loose enough to catch that also catches
-plain `--force`. Exempting it means re-opening the hole. It asks.
-
-**That prompt costs nothing here, which is why the hole stays closed rather
-than the rule loosened.** Neither `implement-issue` nor `release` mentions
-`rebase`, `--force` or `--force-with-lease` anywhere, and this project merges
-the target branch before a PR instead of rebasing — a merge-based flow never
-needs a force push. So the rule should lie dormant. **If it ever fires, treat
-that as the finding**: an agent has gone off-script into a rebase or an amend
-of already-pushed commits. Answer the prompt on its merits; do not "fix" it by
-narrowing the push rule — that is how the gaps PR #596 had to close were
-introduced in #588.
+`git push --force-with-lease` **still asks**, and unlike the ordinary
+feature-branch push it is not carved out. `Bash(git push *--force*)` catches it
+at any argument position, and that is intended: a lease makes the force safer
+against a *concurrent* writer, not against a wrong local history.
+
+**That prompt costs nothing here, which is why it stays.** Neither
+`implement-issue` nor `release` mentions `rebase`, `--force` or
+`--force-with-lease` anywhere, and this project merges the target branch before
+a PR instead of rebasing — a merge-based flow never needs a force push. So the
+rule should lie dormant. **If it ever fires, treat that as the finding**: an
+agent has gone off-script into a rebase or an amend of already-pushed commits.
+Answer the prompt on its merits; do not "fix" it by widening what runs
+unattended — that is how the gaps PR #596 had to close were introduced in #588.
+The feature-branch carve-out is safe precisely because it is defined by what the
+command *lacks* (no force, no `+`, no `--delete`, no tag, no protected ref),
+with every one of those shapes pinned by command string in the gate.
 
 **What the unattended set actually risks is uncommitted work**, since the
 sandbox contains writes to the repo but cannot distinguish a wanted write from