hermes-flow-v2

name	hermes-flow-v2
description	Run long or multi-checkpoint repository work through Hermes Kanban with autonomous OpenAI author/reviewer routing.
version	1.0.0
platforms	["linux"]
metadata	{"hermes":{"tags":["coding","kanban","review","codex"],"related_skills":[]}}

Hermes Flow v2

Load this skill for repository work with more than one checkpoint/PR or an expected duration over 30 minutes. Do not continue such work through a chain of chat --resume sessions.

Hard contract

1. Use Hermes Kanban as the durable DAG/checkpoint ledger: preflight → author → commit → reviewer → CI/PR → merge → cleanup. Store artifacts under /home/uap/swarm-out/<mission>/ and attach their paths to task results.
2. Before any write, run /home/uap/swarm-bin/flow_contract.py delivery-route with the installed policy/signals and then guard-repo for the exact repository, remote, branch, and disposable worktree. An activated profile must be schema v3 with three to seven bounded correction retries after the initial candidate; migrate it only while its timer is stopped. The canonical contract rejects any review cycle above eight.
3. Use only the exact standing-approved OpenAI route returned by the policy. Never substitute another provider or a local/GPU model. Ordinary Luna/Sol/Terra use and subscription spend do not require owner confirmation.
4. Run the author with the exact model returned by the route. The author edits/tests only; the orchestrator stages an explicit allowlist and creates the commit after re-running checks.
5. Run the reviewer in a separate session with codex exec --sandbox read-only. Immediately before it starts, create an attest-source artifact and include its printed marker once in the review prompt. Summarize each Codex JSONL together with its matching local rollout and source artifact; runtime provider, model, sandbox, cwd and source SHA must match the resolved route and role. For same_provider_independent, the exact reviewer model must differ from the author model. The reviewer reads the real diff, runs checks, and produces verification.json; it never edits, commits, pushes, or merges.
6. Validate summary.json and verification.json against the persisted canonical route decision, current HEAD and green required CI. No manual model/review waiver exists.
7. Use only the profile's bounded correction budget (three to seven retries after the initial attempt). Failed author gate, independent review rejection or required-CI failure increments the same routing signal and automatically escalates the next OpenAI route. Persist only redacted bounded author diagnostics and bounded CI check-name/outcome metadata, never raw logs. Reconcile lost initial-push, PR-create or repair-push responses only when the exact branch/candidate/base identity matches. Exhausted pre-commit author gates retain only redacted bounded diagnostics and terminate after cleanup without PR evidence. After a PR exists, final failure requires a live claim and its exact durable number/head/base. Preserve an open exact failed PR/branch as bounded evidence; if already closed, lease-delete only its unchanged branch/SHA. Clean local disposable state and publish terminal failure. Complete a successful root card only after merge, default-branch proof, remote/local branch deletion, disposable worktree removal, and a green terminal-check.
8. When the task belongs to a central Hermes mission, use the installed mission_adapter.py ingress/sync boundary and preserve its mission_id/tenant. Never bypass its idempotency key. --allow-dispatch additionally requires the configured profile, an explicit assignee and a non-scratch workspace.

The full command reference and artifact schemas are in runbooks/hermes-flow-v2.md in PavelLizunov/unified-agent-platform.