ContextAssembler v6.x — the Hermes Agent plugin version of the Context Assembler: keeps the context window dense, cache-friendly and cheap by spending as little as possible to feed the cloud LLM a context with the highest mutual-information density per token.
This is the Hermes (Python) version. The design has been carried forward into Context Assembler DSH V0.99 — the DeepSeek Harness (dsh) plugin version, which is under active development. This repo is maintained as the reference implementation of the original Hermes design.
- E-stage — write-on-disk: every session turn is captured as it arrives (5 hooks), including tool calls, LLM calls and think traces (decision 44: block-level OODA tagging).
- F-stage — background local 4B daemon turns raw turns into facts:
Fct(OODA four phases) +Hdl(handles), with multi-affair transaction frames. - Topic management — real-time topic-block segmentation (forced phrases → confirmation turns → Jaccard with water-pressure: the more context accumulates, the more aggressively topics split; at peak it force-splits).
- Grading & cache freezing — on topic switch, old blocks are graded ACT/REL/FAR against the current question (embedding centroid radius); grades freeze until the next switch so the context prefix stays stable and cloud prompt caches keep hitting.
- A-stage assembly —
_build_conv_history_v6: tail hard-protection (2 user turns verbatim), ACT/REL/FAR tiered downgrading, FAR thought/tool removal; topic-block summary versions built on switch (2–6 strands per block via one 4B call). - Reality merge & inject — strands are merged into "realities" by work affinity (question-cloud / work-cloud + 4B acceptance); on new topic blocks, related realities are injected at the head (θ ≤ 0.5, top-15 pool → 4B pick; a legal empty pick means no injection — "rather missing than wrong").
- Tool summarizer — deterministic per-tool structured summaries of
toolCall/toolResult(bash exit code/stderr/key lines, read path+lines, edit/write paths, MCP JSON key fields). - L-stage — idle refinement daemon: merge review → detail regeneration → cross-validation → health scoring → knowledge subgraph; L1 code (0 tokens) → L2 local 4B → L3 cloud.
ca/ # plugin implementation (stage modules, store, reality, topic mgmt, ...)
topic_manager.py # topic segmentation engine
tests/ # pytest suites (unit / stage / store / plugin / audit)
scripts/ # offline ops: reality migration, reprocess pipeline, wiki sync, ...
docs/ # dual-dimension wiki — see docs/README.md
├── architecture/ # current system components (01-overview … 14-idle-refinement + design notes)
└── decisions/ # decision tree (01–45, numbered ADRs)
plugin.yaml # Hermes plugin manifest
The plugin runs inside a Hermes Agent profile. plugin.yaml declares the required
pip dependencies (numpy, urllib3, sentence-transformers, pyyaml) and the hooks it
registers. Point the plugin root of your profile at this repo, set the CA_* environment
overrides documented in ca/config.py (local 4B LLM endpoint, embedding endpoint, topic
thresholds, context budgets), then start Hermes.
Run the test suite:
python3 -m pytest tests/- docs/README.md — the dual-dimension wiki entry (architecture + decisions)
- docs/architecture/ — current system components
- docs/decisions/ — decision records (ADR), 01–45
- The authoritative design intent and the mapping to Context Assembler DSH V0.99 live in Context Assembler DSH docs/DESIGN.md