Skip to content

Testing the premises for retiring the vendored Hermes tree

Status: findings, 2026-08-04. Written against task_d9692288. Nothing is implemented or reverted here; this records what the evidence says about two claims the retirement rests on.

The two premises

The migration from the vendored Hermes runtime to OpenClaw was justified by:

  1. OpenClaw runs out of the box with no patches, removing the maintenance burden of a patched vendor tree.
  2. Hermes is no longer required because MAC superseded the Hermes learning capability.

Both were checked against the live fleet and the repository on 2026-08-04. Neither holds as stated.

Premise 2 — the learning capability

Not supported. MAC's learning loop runs on schedule and produces no durable output.

Sampling 5,000 observability events from the hub:

agent.nap_completed events:        10
  with summary_evidence_id:         0
  with summary_evidence_id = null: 10
window: 2026-08-04T00:55:26 -> 2026-08-04T01:32:40

The loop is not dormant — nap.tick.run shows four agents (agent_hub, agent_worker-1, agent_worker-2 and agent_operator) all with "napped": true, "skipped": false. Every run then completes carrying no summary evidence.

The memory store agrees. All 31 entries in project=mac are hand-authored operator or agent notes — plan-wave-cli-blocker, fleet-churn-root-causes, e2e-audit-2026-07-14-defect-taxonomy. Across the nine days from 2026-07-24 to 2026-08-02, continuous fleet operation, the loop wrote nothing; the five August entries were written by hand during the 2026-08-02/03 session.

Corroborating from the origin-yield table in docs/archive/field-notes/assessment-2026-08-02.md: the learning-adjacent generators are the worst performers measured — curiosity_adjudication 0/11, backlog_grooming 0/5, dream_low_confidence_repair 4/1396 (0.3%, since deleted).

So MAC has the architecture — nap_ticker, nap_consolidator, curiosity_reviewer, dreaming, fleet_learning, worker_reflect. What it demonstrably has in production is a durable operator-notes store, which is genuinely valuable and is what carried knowledge across sessions. That is not the same claim as superseding a learning capability.

Tracked separately as the null-nap-output defect.

Premise 1 — the patch burden

Not supported. OpenClaw is already patched, and the Hermes vendoring is engineered rather than ad-hoc.

runtime local patches notes
Hermes 12 patches, 3,632 lines, 7 overlay files reproduces byte-for-byte from pristine upstream
OpenClaw 1 patch + a filed upstream issue patch-stuck-session-recovery.py, openclaw#105586

"No patches" is already false. OpenClaw needed deploy/openclaw/patches/patch-stuck-session-recovery.py for a wedge where the stuck-session watchdog detects a leaked run marker but recovery declines to reclaim the lane forever — the agent adds reactions but never replies until the gateway is restarted. It is filed upstream as github.com/openclaw/openclaw/issues/105586 and the local patch is to be dropped once a release ships the fix.

The 12 Hermes patches are also less burdensome than the count suggests. From deploy/hermes/LOCAL_PATCHES.md:

  • zz-public-api-docstrings.patch is documentation-only — no behaviour or signature changes.
  • fts5-orphan-schema-recovery.patch and remove-duplicate-top-level-skills.patch are upstream bug fixes (FTS5 shadow-table corruption self-heal; ambiguous skill-name shadowing) and are candidates for upstreaming rather than permanent local delta.
  • The remainder are genuine MAC integration: provider decision, runtime-context prompt, sandbox PATH precedence, honest one-shot exit status, multi-Slack.

Crucially the pipeline is reproducible, not a pile of hand edits: scripts/vendor-hermes-snapshot.sh rebuilds the tree from pristine upstream at a pinned commit through patches, removals and overlay, validated to reproduce the committed tree byte-for-byte, with tests/test_hermes_vendor_integrity.py pinning a content digest so any drift fails loudly.

What the retirement costs

Independent of the premises, deleting src/mac/_hermes removes:

  • 25 skill families — creative, data-science, devops, email, github, media, productivity, red-teaming and others. What MAC ships for OpenClaw in this repository is one plugin, mac-continuity. That is not a like-for-like comparison, since OpenClaw brings capabilities that do not live here, but the repo-side provision is 25 against 1 and the difference has not been measured.
  • MAC's only NeMo Relay tool/LLM instrumentation. relay_tool_context and relay_llm_context are wired in exactly two places, both inside the vendor: _hermes/agent/tool_executor.py and _hermes/agent/chat_completion_helpers.py. Phase 1 (create_agent_scope) is on live paths and survives; Phase 2 does not. No test catches this — the suite exercises the seam, not the live call path — and Relay is default-ON for every deployed agent (deploy_env.py:1161), so the telemetry would keep looking enabled. Recorded as an ordering dependency on task_d9692288.

What this does NOT conclude

It does not conclude "retreat to Hermes".

The migration is substantially done: OpenClaw is the live gateway, the Hermes runtime is inactive (gateway_ownership.services = {hermes: inactive, nemoclaw: inactive, openclaw: active}), and only com.mac.openclaw-gateway runs. Reversing has its own cost, and this analysis has not measured OpenClaw's actual capability surface against those 25 skill families — that is the missing comparison and the thing most likely to change the answer.

What it does conclude is narrower and firmer: the two stated reasons for the migration do not currently hold, so the decision should be re-made on whatever the real reasons are, rather than on these.

Suggested next step

Measure the capability surface, not the premises. Enumerate what the 25 vendored skill families actually provide, which of those the fleet has used in the last 90 days, and which OpenClaw covers today. A skill family nobody has invoked is not a cost of retirement.

References

  • deploy/hermes/LOCAL_PATCHES.md — the patch set and its rationale
  • deploy/openclaw/patches/UPSTREAM-ISSUE-stuck-session-recovery.md
  • docs/archive/field-notes/assessment-2026-08-02.md section 4 — origin yield
  • docs/openshell-nemo-relay-integration.md — Relay phases
  • task_d9692288, and the null-nap-output defect filed alongside this note