| name | ops-rollback |
| description | The rollback runbook-class playbook for the guild:operations router — producer devops-incident-runbook, the guild.ops.v1 + guild.incident.v1 outputs, the hard INTERACTIVE-always (never autonomous) posture, and the REFUSE-on-non-force-pass-Quality-block rule. Invoked by the guild-operations router after it selects class=rollback. TRIGGER for "run the rollback runbook", "rollback class", "roll back the deploy", "revert to previous version", "back out the release". DO NOT TRIGGER directly at the ops phase (the guild-operations router selects the class and dispatches here), for the other runbook classes (release/monitoring/incident/maintenance own theirs), for the quality/release-readiness computation (guild:quality owns it), for reverting a single skill version (guild:rollback-skill owns that), for authoring the runbook content (devops-* produce it), or for closing the initiative (InitiativeCloseout owns D8 closure). |
| when_to_use | Invoked by the guild-operations router when class==rollback — run the discovered rollback runbook INTERACTIVELY (never autonomous) under the router's safety rails + pre-flight and populate guild.ops.v1 + guild.incident.v1. |
| type | meta |
ops-rollback
Per-class playbook for rollback, dispatched by the guild:operations
router after ClassSelect resolves class == rollback. rollback is
INTERACTIVE always — never autonomous (rail 2, router-enforced and
hard-coded). The router has already enforced the 4 safety rails + mandatory
pre-flight; this playbook runs the rollback producer and populates the outputs.
When to use it
When the guild:operations router routes a selected rollback class here — to
execute the discovered rollback runbook interactively (revert a deploy/release
to a known-good state).
When not to use it
- Other classes →
ops-release / ops-monitoring / ops-incident /
ops-maintenance (an active outage with no known revert target is
incident, not rollback).
- Reverting a single Guild skill version →
guild:rollback-skill (different
surface entirely).
- Direct invocation at the ops phase → the router selects the class.
- Quality / release-readiness →
guild:quality.
- Authoring runbook content →
devops-incident-runbook.
- Initiative closure →
InitiativeCloseout.
Required inputs
- The selected
rollback runbook + its wiki trust root
.guild/wiki/standards/runbooks/<name>.md (router-validated).
- The consumed
guild.quality.v1 recommendation
(.guild/runs/<run-id>/quality/<run-id>.md).
- Frozen
guild.ops.v1 / guild.incident.v1
(../../guild-operations/operations-contract.md §"guild.ops.v1 fields", pointer only).
Output format
guild.ops.v1 always + guild.incident.v1 (class ∈ {incident, rollback}),
populated by pointer to ../../guild-operations/operations-contract.md §"guild.ops.v1 fields" (zero re-spelled
fields); per-step evidence under .guild/runs/<run-id>/ops/evidence/.
Workflow steps
producer
Producer devops-incident-runbook — class→producer map per
../../guild-operations/operations-contract.md §Posture (by pointer). Posture: INTERACTIVE
always — never autonomous (rail 2), regardless of runbook approval or
--auto-approve. Active specialists ≤ producer + challenger.
KB integrity check (optional, defense Layer 5)
When the rollback target is .guild/wiki/ state itself (a suspected
poisoned/tampered knowledge-base page — distinct from a deploy/release
rollback), this class also has a restorative security check available:
prompt-injection-defenses.md §Layer 5's KB snapshot & rollback helper. It is
opt-in — invoke it only when the runbook's target is the wiki, never for a
plain deploy/release rollback.
# One-time (or periodic) snapshot of a known-clean wiki state:
npx tsx ${GUILD_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-$HOME/.local/share/guild/dist/claude-code}}/scripts/lib/kb-snapshot.ts snapshot \
--wiki-dir <repo-root>/.guild/wiki --dest-dir <repo-root>/.guild/kb-snapshots --id <snapshot-id>
# At rollback time: verify current state against that snapshot (detection only).
npx tsx ${GUILD_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-$HOME/.local/share/guild/dist/claude-code}}/scripts/lib/kb-snapshot.ts verify \
--manifest <repo-root>/.guild/kb-snapshots/<snapshot-id>.json
# If tampered, get a plan-only diff (added/removed/tampered) — this NEVER
# restores automatically; you drive the actual restore (git checkout / copy
# from backup) per the plan:
npx tsx ${GUILD_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-$HOME/.local/share/guild/dist/claude-code}}/scripts/lib/kb-snapshot.ts rollback \
--manifest <repo-root>/.guild/kb-snapshots/<snapshot-id>.json
rollback is plan-only by default — it lists what changed, it never mutates
.guild/wiki/ itself. This is the class's never-autonomous posture applied
to KB restoration too: a human reviews the diff and performs the restore.
execute
Each step records op_class; every hard-set step shows
autonomy: prompted_inline (rail-3 proof). Populate guild.ops.v1 head +
guild.incident.v1 by pointer to ../../guild-operations/operations-contract.md §"guild.ops.v1 fields".
Evidence requirements
- producer is
devops-incident-runbook; posture interactive (never
autonomous); per-step op_class; hard-set step prompted_inline;
guild.incident.v1 populated by pointer; active set ≤ producer + challenger.
Escalation rules
rollback is never autonomous — always interactive (rail 2).- A non-force-pass Quality
block makes rollback REFUSE with route-back
to Quality (rollback is in the router's refuse set, unlike incident); only
a recorded human force_pass proceeds.
Safety constraints
- Rail 2 (never autonomous) is absolute for this class; approval lowers only
the SOFT gate, never the hard set (rail 3).
- No frozen
guild.ops.v1 / guild.incident.v1 field text reproduced — all by
pointer through ../../guild-operations/operations-contract.md. No cloud-build task (GR-7).
Eval cases
rollback under --auto-approve=all → still interactive (rail 2);
hard-set steps prompted_inline (rail 3).rollback, Quality block, no force_pass → REFUSE with route-back
(rollback is in the refuse set, unlike incident).rollback, Quality block, recorded human force_pass → proceeds.rollback → emits guild.ops.v1 + guild.incident.v1; active set never
exceeds producer + challenger.
