feat(autopilot): make Hermes implementer delegation lifecycle-safe

[ritsuKai2000]

Summary

• Keep the full implement-issue lifecycle in one canonical skill and reduce the Claude/Hermes differences to thin runtime adapter references.
• Route Hermes coordinator sessions and depth-needing stages through a Jinn-owned stateless launcher, while leaving the installed upstream Hermes runtime unchanged.
• Launch stages 1, 3, 4, and 5 as fresh Hermes OS roots so each stage gets its own native delegation-depth budget and can fan out.
• Carry the selected adapter, bare model, explicit provider, installed Python, and Autopilot package path consistently from dispatcher to stage runner.
• Fail closed at boot if the configured Python cannot import Hermes, the model is org-prefixed, or the provider is not openai-codex.
• Add contract tests for the canonical skill boundary, launcher session context, runtime argv/env, fresh-root stages, runtime/billing guards, and Claude compatibility.

Why the stateless launcher exists

Hermes Agent v0.18.2 forces top-level model delegation into background mode. A finite hermes chat -q process can therefore exit before its delegated child result returns. There is no released CLI/config switch that changes that capability. The checked-in launcher binds the normal Hermes session context with async_delivery=False, invokes the standard Hermes CLI, and clears the context afterward. This makes finite sessions synchronously join native delegates without patching or forking upstream Hermes.

Upstream context:

• Exact one-shot delegation issue: [Bug]: delegate_task results are lost when called from Kanban workers; parent exits even with goal_mode NousResearch/hermes-agent#63169
• General background delegation issue: [Bug]: Async delegation completion events lost in cron jobs NousResearch/hermes-agent#53027
• Proposed upstream repair: fix(kanban): keep delegated results in worker turn NousResearch/hermes-agent#63218

Verification

• cd packages/autopilot && yarn typecheck — exit 0.
• cd packages/autopilot && yarn test — 65 test files, 618 tests passed; exit 0.
• Installed-Hermes import probe (gateway.session_context + hermes_cli.main) — exit 0.
• git diff --check origin/next...HEAD — exit 0.
• Final whole-branch review — ready to merge, with no remaining findings.

Direct native-delegation smoke used the checked-in launcher with the installed Hermes Python, normal HERMES_HOME, bare gpt-5.6-sol, and explicit openai-codex provider:

HERMES_HOME="$HOME/.hermes" "$HOME/.hermes/hermes-agent/venv/bin/python" \
  packages/autopilot/bin/jinn-hermes-stateless.py chat -q \
  'Delegate exactly one task using delegate_task. In that child task, instruct the child to return exactly HERMES_SYNC_CHILD_OK and nothing else. Wait for the delegated result. After receiving it, output exactly: PARENT_RECEIVED: HERMES_SYNC_CHILD_OK' \
  -Q --yolo --accept-hooks --model gpt-5.6-sol --provider openai-codex

Result: exit 0 in 16 seconds with exact marker PARENT_RECEIVED: HERMES_SYNC_CHILD_OK.

Live Autopilot outcome

Issue #1822 was still parked solely from the original Hermes one-shot failure, had no implementation PR/new external work, and returned classification: clear. It was restored for one bounded retry: #1822 (comment)

The authorized command was:

cd packages/autopilot
JINN_DISPATCHER_IMPLEMENTER_RULES='[{"effort":"Low","implementer":"hermes"}]' \
  yarn autopilot --once --cap 1

The boot log correctly named the Hermes coordinator, bare model, explicit provider, and installed Python. However, the cycle dispatched no issue: existing shared eng-loop work consumed the cap (skipped (throttle): 6). The shared dispatcher was repeatedly retrying #1816, #1829, and #1833, whose session logs ended on Claude session-limit messages. Concurrent drift also changed #1822 during observation without producing a #1822 worktree, updated session log, commits, or PR.

To prevent unintended pickup, #1822 was safely parked at Blocked on: Human / Status: Human: #1822 (comment)

Therefore the direct native-delegation proof passed, but the full nine-point live issue pipeline was blocked before Hermes dispatch. No live Hermes session or implementation PR was created, and this PR does not claim that end-to-end proof.

Scope guard

The unrelated local change to docs/runbooks/swe-rebench-v2-public-testnet.md was not staged, committed, pushed, or included in this PR.