stuck-playbook — guardrail-recovery escape-hatches (HIMMEL-211)

You hit a himmel guardrail and need the recovery rule. The full symptom→action
playbook lives in the repo at docs/internals/stuck-playbook.md — read it
now and apply the section matching your symptom.

Read docs/internals/stuck-playbook.md

(In a worktree, the repo root is the worktree dir; the path is the same.)

The one rule that overrides every workaround

Never reshape a command to dodge a guardrail. The guardrails are structural
on purpose (HIMMEL-195: structural > instructional). If a write is still denied
after applying the matching playbook section, that denial is correct — defer
to the operator. Prefer a structural fix (a new auto-approve-safe-bash case,
a CLI flag) over a cleverer command.

Symptom index (detail in the playbook doc)

• Bash command stopped by a permission prompt (interactively it waits;
  headless/auto it DENIES at rc=0 and continues — a silent no-op) → the native
  matcher bails on $var / $(…) / backticks / compound operators. Prefer
  literal single commands. (HIMMEL-203 / HIMMEL-1969)
• Jira write fell through to the classifier and was DENIED → command-SHAPE
  problem, not a write-permission problem. Prefer literal node …/jira … (bare
  or cd-prefixed); multi-line bodies via --comment-file / --desc-file.
  (HIMMEL-205 / 209)
• Pre-push gate failed on a missing attestation trailer → put the trailer in
  the FIRST commit; never reactive --amend (HARD-blocked in auto-mode). If
  already pushed, add it to the PR body.
• /worktree refused the branch ("PR already MERGED") → deliberate reuse
  bypasses with REUSE_MERGED_BRANCH_OK=1 set in the LAUNCHING shell; prefer a
  fresh type/slug branch. Lingering merged-PR worktrees are flagged read-only
  by /himmel-doctor C7 → run /clean.
• Every Write/Edit under /tmp refused: "cannot determine branch for
  '/tmp'" → a stray .git in an ANCESTOR of the edited path (not the file
  itself) makes block-edit-on-main treat that ancestor as a repo root.
  /himmel-doctor C36-stray-tmp-git catches it; rmdir it if empty, stop and
  name it (never delete contents) if not. (HIMMEL-2739)
• An ad-hoc Bash line containing the literal word "enable" is refused in a
  worktree-pinned session → the harness's own worktree-isolation guard, not
  a himmel script; independent of git involvement. Do not reshape the
  command to dodge it (the rule above still applies) — rephrase to avoid the
  word where possible, or defer to the operator. (HIMMEL-2739)
• A late fix after plain pre-commit run is gated against stale content
  → it stashes every unstaged file repo-wide and lints the STAGED index;
  --all-files does NOT stash. git add the fix again before re-running.
  (HIMMEL-2739)
• An outward-facing command (gh pr create / gh pr comment / git push)
  is denied with Stage 2 classifier error → do not retry verbatim
  back-to-back; end the turn or do one unrelated read, then retry ONCE,
  delayed, with the same head: gh pr create/gh pr comment via
  --body-file instead of inline; git push has no body flag, so retry the
  exact same command once — the delay and intervening read are themselves
  what make it non-identical, not a contrived alternate spelling. Any denial
  after that one retry → stop, BLOCKED to the console with the exact text;
  the console never runs it (laundering). [Out-of-Place Publication] gets
  no retry at all, first time seen or not — escalate immediately the same
  way. (HIMMEL-3020)

Why these are a load-on-trigger playbook and not CLAUDE.md rules:
docs/internals/stuck-playbook.md § Why this is a playbook, and memory
feedback_no_operational_rules_in_claudemd.