Research-Stack/6-Documentation/docs/gcl/HermesAgentFieldOperatorBridge.md

7 KiB

Hermes Agent — Field Operator Bridge for Research Stack

Status: Integration Proposal
Date: 2026-05-02
Version: 1.0.0-Phase0
Source: https://github.com/allaunthefox/hermes-agent
Lean Module: 0-Core-Formalism/lean/Semantics/Semantics/HermesAgentIntegration.lean


Core Thesis

Hermes Agent is the field-operator layer for Research Stack: a messenger, scheduler, skill runtime, and subagent launcher that can perform repeatable GCL/Warden workflows while remaining subordinate to Lean, receipts, and Warden promotion gates.

Hermes is not the brain. The canonical model, lawfulness filter, and proof authority remain in OTOM/GCL/Lean. Hermes sits outside the proof core, executing tasks that Lean delegates but never verifying its own outputs as truth.


Role Separation

System Role Authority
Hermes Observes, routes, schedules, executes Workflow layer only
GCL Classifies, compresses, binds Coding-space authority
Lean Verifies, proves, is source of truth Proof authority
Warden Promotes, holds, blocks Promotion gate
ENE Persists, distributes, syncs Swarm substrate

Warden-Safe Doctrine

Hermes may execute.
Hermes may remember.
Hermes may suggest.
Hermes may schedule.
Hermes may not promote without receipts.

Critical Rules

  1. If Hermes skill writes or updates doctrine and no source/provenance/receipt boundary is emitted: mark HOLD.

  2. If Hermes self-improvement modifies a skill used for promotion: require SkillMutationReceipt + AdversarialTrial.

  3. All scheduled audit outputs default to HOLD until external receipt validates them.

  4. Hermes is autopoietic at the workflow layer, not authoritative at the truth layer.


Integration Surfaces

Hermes Feature Research Stack Binding Warden Boundary
Skills Repeatable GCL procedures Skill mutation must be receipted
Memory Project/user continuity Memory is context, not evidence
Cron Scheduled audits Scheduled output defaults to HOLD
Messaging Gateway Remote command interface High-risk commands require confirmation
Subagents Parallel review/search/build jobs Subagent agreement is not evidence
MCP Connectors/tools surface Tool output needs provenance
Terminal Backends Local/SSH/Docker/Modal execution Command allowlist + sandbox receipts
RL/Trajectories Training data for workflow agents Trajectory compression is not proof

Command Surface

Hermes exposes these read-only and receipt-producing commands:

/run-gcl-audit          → GCL classification audit
/summarize-receipts    → Warden receipt ledger summary
/queue-deepseek-review → Adversarial review bundle assembler
/check-lean-sorries    → Lean sorry scan + classification
/search-provenance     → Provenance database search
/build-cff-entry      → CFF citation entry generator
/warden-status         → HOLD/BLOCK/CANDIDATE/REVIEWED report
/skill-run <name>      → Execute named skill with receipt emission

Skill Specifications (Phase 1)

gcl-provenance-cff

  • Task: Ingest DOI / article / source → create CFF entry → add Warden boundary
  • Sources: doi, article_metadata, provenance_db
  • Receipts Required: sourceAudit, humanReview
  • Failure Modes: missing_doi, malformed_cff, no_provenance

lean-sorry-audit

  • Task: Scan Lean files → list sorry locations → classify gaps
  • Sources: 0-Core-Formalism/lean/Semantics
  • Receipts Required: leanBuild
  • Failure Modes: build_failure, sorry_increase, missing_todo_comment

adapter-spec-writer

  • Task: Turn source cluster into GCL bridge doc
  • Sources: source_cluster, gcl_schema
  • Receipts Required: humanReview, deltaPhiAudit
  • Failure Modes: missing_bind, no_cost_function, float_in_hotpath

deepseek-review-bundle

  • Task: Collect docs → enforce adversarial convergence prompt → save artifact
  • Sources: review_docs, adversarial_prompt_template
  • Receipts Required: adversarialTrial, humanReview
  • Failure Modes: missing_convergence_prompt, artifact_too_large, no_receipt_emitted

warden-triage

  • Task: Classify outputs as HOLD / CANDIDATE / BLOCK / REVIEWED
  • Sources: output_artifact, receipt_ledger
  • Receipts Required: wardenEmission
  • Failure Modes: missing_status, no_receipt_boundary, self_promotion_detected

Scheduled Audit Jobs (Phase 2)

Daily

  • Check Lean build (lake build)
  • Check sorry count (trend vs. baseline)
  • Check uncommitted docs in docs/gcl/
  • Check CFF malformed entries
  • Check provenance gaps

Weekly

  • Run DeepSeek bundle lint
  • Summarize new GCL docs
  • Report Warden HOLD/BLOCK items
  • Check MMR divergence / attestation staleness

Promotion Phases

Phase 0 — Read-Only Bridge (Current)

  • No writes, no autonomous mutation
  • Hermes can: search docs, summarize Warden state, run read-only git status, list Lean sorries, assemble review bundles
  • Output: this document + HermesAgentIntegration.lean

Phase 1 — Receipt-Producing Skills

  • Allow artifact creation, no promotion
  • Every skill emits: SkillRunReceipt + SourceReceipt + WardenStatus

Phase 2 — Scheduled Warden Jobs

  • Cron-driven recurring audits
  • All outputs default to HOLD

Phase 3 — Controlled Write Authority

  • Hermes may open PRs or commits
  • Each commit must include: source boundary, Warden state, rollback path, no self-promotion

Prior Art

  • Nous Research Hermes Agent: Self-improving agent with 4-layer memory, session lineage, FTS5 search, skill creation/patching, schema-versioned SQLite storage
  • Awareness Date: 2026-04-05
  • Clean Room Status: Independence confirmed for Research Stack's dual-storage SQLite + JSON, procedural/episodic separation, skill patch pattern, and prompt memory cap
  • Adaptations from Hermes: Session lineage (parent_session_id + recursive CTE), FTS5 virtual table, pre-reset memory saving on EMERGENCY regret threshold, schema versioning, WAL mode with retry jitter
  • Record: scratch/exploit_recovery/audit/sessions/prior-art-hermes-agent-nous-research-20260405.json

Lean Formalization

The integration boundary is formalized in HermesAgentIntegration.lean:

  • HermesCommand — inductive command surface (8 constructors)
  • HermesSkill — skill specification with phase, receipts, failure modes
  • HermesStatus — promotion state: HOLD / CANDIDATE / REVIEWED / BLOCKED
  • SkillRunReceipt — execution receipt convertible to ReceiptCore.Receipt
  • canPromote — gate function requiring receipt validation
  • readOnlyCannotPromotetheorem (proven) — Phase 0 skills never promote
  • defaultStatusIsHoldtheorem (proven) — default status is HOLD

The One Rule

Hermes is autopoietic at the workflow layer, not authoritative at the truth layer.