(v2.11.1 — Negative Facts & Next-Action Tracking)
OverCR (Operational Vigilance, Execution, Recovery, Command & Routing) is a Hermes-first portable AI orchestration substrate.
It is not a chatbot, a SaaS wrapper, or a vertical product. It is the governed operating layer that AI workloads run inside.
OverCR provides identity, doctrine, boot continuity, filesystem-first state, subagent routing, packet validation (L1-L6), workflow choreography, audit trails, approval gates, semantic memory, operator TUI, knowledge management, controlled web ingestion, execution sandboxing with kernel isolation, integration hardening, release validation, and vault-grounded guidance facts (negative facts + next-action tracking).
| Version | Layer | Description |
|---|---|---|
| v1.0.0 | Substrate Core | Identity, doctrine, boot, subagent routing, L1-L6 packet validation, task lifecycle, approval gates, audit trails |
| v2.1.0 | Semantic Memory | Governed advisory memory — informs but never authorizes. Rule-gated promotion, conflict detection, deterministic retrieval |
| v2.2.0 | Operator TUI | Terminal-first observatory — task views, DAG visualization, packet inspection, audit streaming, approval queue. Reads from FS, never mutates |
| v2.3.0 | Workflow Library | Template-based reusable workflows with registry, loader, executor, isolated contexts |
| v2.4.0 | Knowledge Layer | Governed research subsystem — source registry, classification, document ingestion, knowledge index, provenance tracking, contradiction detection |
| v2.5.0 | Web Ingestion | Operator-auditable single-page fetches with URL validation, prompt injection scanning, content normalization, robots policy |
| v2.6.0 | Execution Sandbox | Approval-gated command execution — 14-command allowlist, 9-check policy, filesystem guard, network guard (default-deny), rollback snapshots, execution receipts |
| v2.7.0 | Kernel Isolation | Backend abstraction with LocalBackend, BubblewrapBackend, FirejailBackend. Auto-selects strongest available. Isolation profiles, resource limits |
| v2.8.0 | Workflow Composition | Composite DAGs — condition evaluation, branching, escalation policy, retry policy, routing policy, subworkflow loading with cycle detection, state machine |
| v2.9.0 | Integration Hardening | 8 validators — schema registry, system validation, replay determinism, state consistency, release integrity, migration checking, compatibility matrix, recovery verification |
| v2.10.0 | Stable Release | Release preparation — semantic compatibility, install validation, reproducible builds, operator readiness, version matrix, release manifests |
| v2.11.0 | Vault Memory | Obsidian vault integration — reads structured facts fences, builds index, injects relevant knowledge into task context at creation time. No vector DB. No embeddings. Pure CAG. |
| v2.11.1 | Guidance Facts | Negative facts (kind:rejected) and next-action tracking (kind:next_action) — agents stop retrying dead ends and know what to do next. Optional fact kind prefix, backward compatible. |
overcr/
README.md # This file
CHANGELOG.md # Full change history v0.1 → v2.10.0
INSTALL.md # Installation and quick start
RELEASE.md # Release notes
LICENSE.md # MIT license
soul.md # Identity, personality, rules, workflow
soul_reference.md # Integrity-check copy of soul.md
boot.sh # Cold-start boot script
config/ # YAML configuration
inference_routing.yaml # Inference routing policy
model_policy.yaml # Model capability/class/sovereignty policy
model_routing.yaml # Model routing configuration
workflow_choreography.yaml # Workflow choreography definitions
configs/ # Config templates (v1-era, fill placeholders)
prompts/ # Boot prompt templates
skeleton/ # Directory scaffold for new instances
memory/ # v2.1.0 — Semantic memory layer
tui/ # v2.2.0 — Operator interface (observatory)
workflow_library/ # v2.3.0 — Reusable workflow templates
knowledge/ # v2.4.0 — Governed research subsystem
web_ingestion/ # v2.5.0 — Controlled web ingestion gateway
sandbox/ # v2.6.0-v2.7.0 — Execution sandbox + kernel isolation
workflow_composition/ # v2.8.0 — Composite DAGs, branching, escalation
integration/ # v2.9.0 — Integration hardening & validators
release/ # v2.10.0 — Release preparation & validation
validation/ # v2.10.1 — Soak tests, perf baselines, fuzzing
subagents/
coder/ # Advisory code analysis & patch planning
cryer/ # Reconnaissance & signal intelligence
knower/ # Knowledge, research, observation, evaluation
pyper/ # Execution planning & safety review
runtime/ # Core runtime (v1.0.0 foundation)
orchestration/ # Task lifecycle, packet examples
tools/ # validate_packet CLI validator
tests/ # 27 test suites
examples/ # 30+ demo scripts
scripts/ # 15+ check/build/release scripts
docs/ # Governance, runtime, repo docs
security/ # Threat model, security review
references/ # 35+ architecture reference docs
dist/ # Release archives
- Filesystem truth is authoritative. Chat history is ephemeral; filesystem state is canonical.
- Model output is untrusted. Outputs must be sanitized and validated (L1-L6) before state advancement.
- OverCR is sovereign. Subagents never route directly to each other. All routing passes through OverCRRuntime.
- No autonomous outbound contact. External contact requires explicit operator approval.
- No autonomous filesystem mutation. CodER and PypER produce plans, not unsupervised host actions.
- Semantic memory is advisory. Memory informs decisions but never authorizes state mutation.
- TUI is an observatory. The operator interface reflects truth from disk; it never creates it.
- Hermes-first, runtime-agnostic. Hermes is the validated reference runtime.
| Subagent | Purpose | Packet Types |
|---|---|---|
| KnowER | Knowledge, Observation, Wisdom, Evaluation & Research | claim_review, myth_fact, research, assessment |
| CryER | Reconnaissance & signal intelligence | engagement_signal, recon, reputation_signal, hiring_growth |
| CodER | Code analysis, patch planning, diagnostics | patch_plan, diagnostic, completion, blocked |
| PypER | Execution planning, safety review, receipts | execution_plan, execution_receipt, execution_refusal |
PypER always operates with approval_required=true and execution_authority="none". All PypER output routes to the operator, never to another subagent.
- Copy
overcr/to your desired location (e.g.,$HOME/overcr) - Copy skeleton contents and fill config templates:
cp -r skeleton/* . sed -i 's|{{OVERCR_ROOT}}|/your/path/to/overcr|g' configs/*.tpl # Replace remaining {{PLACEHOLDER}} values, then: for f in configs/*.tpl; do mv "$f" "${f%.tpl}"; done
- Run
./boot.shand launch Hermes with the printed command.
cd tests && python run_all.py27 test suites covering every v2.x subsystem: memory, TUI, workflows, knowledge, web ingestion, sandbox backends, integration hardening, release candidate, workflow composition, and more.
| Version | Date | Type | Notes |
|---|---|---|---|
| v0.0.3–v0.0.5 | 2026-05-09/10 | Core | Clean core, subagents, orchestration |
| v0.1.0 | 2026-05-10 | Runtime | Filesystem task runtime, approval gates, audit trail |
| v0.2.0–v0.2.4 | 2026-05-10 | Workers | First live workers, routing, testing, packaging |
| v0.3.0–v0.4.3 | 2026-05-10 | KnowER | KnowER live worker, inference governance, real inference |
| v0.5.0–v0.6.0 | 2026-05-10 | CryER, CodER | CryER live inference, CodER advisory patch planning |
| v0.7.0–v0.8.0 | 2026-05-10/11 | PypER, Workflows | Execution planning, cross-worker workflow choreography |
| v0.9.0-rc1 | 2026-05-11 | RC | Threat model, security review |
| v1.0.0 | 2026-05-11 | Stable | Hermes-first portable orchestration substrate |
| v2.1.0 | 2026-05-11 | Memory | Governed semantic memory layer |
| v2.2.0 | 2026-05-11 | TUI | Terminal-first operator observatory |
| v2.3.0 | 2026-05-11 | Workflows | Template-based workflow library |
| v2.4.0 | 2026-05-11 | Knowledge | Governed research & knowledge subsystem |
| v2.5.0 | 2026-05-11 | Web Ingestion | Controlled web ingestion gateway |
| v2.6.0 | 2026-05-11 | Sandbox | Controlled execution sandbox |
| v2.7.0 | 2026-05-11 | Isolation | Kernel isolation backends (bubblewrap, firejail) |
| v2.8.0 | 2026-05-11 | Composition | Composite workflow DAGs, branching, escalation |
| v2.9.0 | 2026-05-11 | Hardening | Integration validators, replay, state consistency |
| v2.10.0 | 2026-05-16 | Stable RC | Release preparation, semantic compatibility, reproducibility |
| v2.11.0 | 2026-06-08 | Stable | Vault-grounded memory — reads Obsidian vault facts fences, injects relevant knowledge into every task context. No vector DB. No embeddings. |
| v2.11.1 | 2026-06-10 | Stable | Negative facts & next-action tracking — kind: prefix on bullet facts, agents stop retrying dead ends. |
MIT License. See LICENSE.md.