handoffpack 0.1.0

Deterministic structured handoff-card codec + MCP translator server for agent handoffs, session compaction, and instruction compilation (client-side sampling for semantics; zero server-side LLM cost).

handoffpack — MCP structured-card translator

Deterministic codec + guards + MCP server for HP1 structured handoff cards. Turns agent handoffs, session logs, and recurring-instruction files into compact structured cards. All deterministic work runs locally; the server never calls an LLM API and holds no API keys.

Install

Python 3.11+ required.

# From PyPI (once published):
pip install handoffpack

# Pre-publish / local: from the package directory
pip install .
# or editable, with test deps:
pip install -e .[dev]

Core codec/guards tests run with stdlib + pytest only; the mcp package is needed only to run server.py.

Register the MCP server (Claude Code)

python -m handoffpack.server                       # stdio transport, run directly
claude mcp add handoffpack -- python -m handoffpack.server

Other clients (Cursor, etc.): use the same command in their MCP server registration UI. A handoffpack console script is also installed as an alias for python -m handoffpack.server.

License key (paid add-ons only)

The 4 MCP tools (encode / decode / compact_to_l2l / verify) are FREE and unlimited. Only paid add-ons (e.g. the full hp-instructions conversion run) require a key. The verifier is fully offline (embedded Ed25519 public key, no phone-home). Key discovery precedence:

export HP_LICENSE=HP1.xxxxx...        # macOS/Linux
set HP_LICENSE=HP1.xxxxx...           # Windows
# or place the key on the first line of ~/.handoffpack/license
# or pass --license <KEY> to hp-instructions

hp-instructions --check (drift scan) is free; a full conversion run requires a plus-tier key.

Layout

mcp-translator/
├── pyproject.toml
├── src/handoffpack/
│   ├── legend.json    # public card schema (typed slots + kv pairs)
│   ├── codec.py       # assemble/parse/validate + chars/4 token estimator
│   ├── guards.py      # breakeven passthrough, BAN-field rule, nuance heuristic
│   ├── license.py     # offline Ed25519 license verifier (paid add-ons)
│   ├── adapter_instructions.py  # hp-instructions CLI (CLAUDE.md -> HP1 cards)
│   └── server.py      # MCP server: encode / decode / compact_to_l2l / verify
└── tests/

Run tests

pytest        # 114 tests

Build & publish (maintainer)

python -m build              # produces dist/*.whl + dist/*.tar.gz
python -m twine upload dist/*

If build is unavailable offline, an equivalent wheel can be produced with python -m pip wheel . -w dist --no-deps --no-build-isolation.

Architecture notes (caller-does-semantics)

• Deterministic work (card assembly, parsing, schema validation, guards, token estimation) is local and pure. The server never calls an LLM API and holds no API keys (zero-server-cost invariant).
• Semantic work (slot extraction, compaction keep/drop judgment) belongs to the CALLING model:
  • encode(slots={...}) — the caller extracts TASK/ROLE/IN/DO/OUT/LIM/BAN from its own context and passes them directly (mode: "slots", confidence: "caller"). Pass text too for token metrics.
  • compact_to_l2l — two-call protocol. Call 1 (session_text only) returns a per-segment scaffold + fill_slots_and_recall + a crude deterministic fallback. Call 2 adds segment_slots={"<seg>": {slots} | "TASK string"}.
  • decode / verify — deterministic.
• Freeform encode(text=...) without caller slots returns a deterministic skeleton card plus fill_slots_and_recall guidance for a second call.
• MCP sampling (session.create_message) survives only as an OPTIONAL, capability-detected upgrade path (deprecated upstream: MCP 2026-07-28 spec RC, SEP-2577; replaced by Multi Round-Trip Requests, SEP-2322; Claude Code never implemented it, anthropics/claude-code#1785). It is OFF unless the operator sets HP_ENABLE_SAMPLING=1 AND the client advertises the capability; results are always deterministically re-validated. Nothing depends on sampling.
• Token estimates use ceil(chars/4). Crude by design; never used for claims or billing.

Guard behavior

• Inputs under ~70 estimated tokens are returned unchanged with a reason (encoding overhead exceeds savings below breakeven).
• BAN slots accept plain forbidden actions only. Conditional negation ("X unless Y", "only outside Z") is rejected; express the condition as a positive LIM constraint instead.
• Prose with high conditional-clause density gets an advise-skip verdict instead of a lossy card.