mcp-agent-handoff 0.9.1

Portable MCP server for agent handoff state and review tracking.

Agent Handoff MCP

Portable MCP server for agent handoff state, review findings, exports, close checks, and lane-scoped worker/orchestrator coordination.

Scope

This package owns handoff state and related workflow primitives:

• task state, decisions, blockers, and next actions
• review findings and review runs
• generated CURRENT_TASK.json snapshots
• lane registration, lane activity, worker reports, and lane messages
• artifact indexing and derived metrics snapshots

It does not include unrelated repo-intel or UI helpers.

Runtime State

By default, runtime state lives under the workspace you point the CLI at:

• SQLite DB: .task-state/handoff.db
• exports: .task-state/exports/
• generated markdown: CURRENT_TASK.json

Override these with CLI flags or AGENT_HANDOFF_* environment variables.

Installation

From PyPI (recommended)

pip install mcp-agent-handoff
# or, as an isolated tool:
uv tool install mcp-agent-handoff
# or, ad-hoc without installing:
uvx mcp-agent-handoff --workspace-root /path/to/workspace serve-stdio

From the monorepo source tree (development)

From this package root inside agentic-protocol-monorepo:

cd packages/mcp-agent-handoff
python -m pip install -e ".[dev]"

Development

All package-local development commands are intended to run from the package root:

make lint-handoff
make fix-lint-handoff
make format-handoff
make mypy-handoff
make test-handoff
make check-handoff

Prefer the package-local Make targets when running checks from editors or agent sessions. Workspace-level interpreter discovery can point at a minimal environment that lacks this package's dev extras such as pytest, while make test-handoff and make check-handoff keep execution anchored to the package root and honor PYTHON=/path/to/python3 overrides when you need to pin a specific interpreter.

The Makefile automatically adds a sibling ../codex-subagent-bridge/src to PYTHONPATH when that checkout exists. In a fully extracted repo, that fallback is unnecessary once dependencies are installed normally.

Direct commands also work:

PYTHONPATH=src python -m ruff check src tests
PYTHONPATH=src python -m mypy src
PYTHONPATH=src python -m pytest tests -q

Python API

The package can be used directly as a library — this is the primary fallback when MCP tool calls are unavailable.

Always import from the package root (agent_handoff_mcp), never from submodules like .config, .decisions, or .core. Submodules are internal and may change.

from pathlib import Path
from agent_handoff_mcp import (
    RuntimeConfig,
    configure_runtime,
    get_handoff_state,
    search_handoff,
  validate_decision_id,
    record_event,
    review_findings,
    list_review_findings,
    set_handoff_state,
    update_task_status,
    get_verified_tests,
    render_handoff,
    record_file_touch,
    get_touched_files,
)

# Configure runtime before any read/write call
configure_runtime(RuntimeConfig.for_repo(Path("/path/to/workspace")))

# Read state
state = get_handoff_state(sections="identity")

# Search decisions
results = search_handoff(queries=["slice_complete"], record_types=["decision"], limit=5)

# Preflight a slice-complete decision id before writing
preflight = validate_decision_id(
  decision="codex_slice_complete_plan0004_contract-pinning-and-docs",
  decision_kind="slice_complete",
)

# List open findings
findings = list_review_findings(status="open")

Without installation (monorepo source tree):

PYTHONPATH=packages/mcp-agent-handoff/src python3 -c "
from pathlib import Path
from agent_handoff_mcp import RuntimeConfig, configure_runtime, get_handoff_state
configure_runtime(RuntimeConfig.for_repo(Path('.')))
state = get_handoff_state(sections='identity')
print(state['data']['active']['task_ref'])
"

CLI Usage

Run the MCP server over stdio:

mcp-agent-handoff --workspace-root /path/to/workspace serve-stdio

Run the MCP server over HTTP:

mcp-agent-handoff --workspace-root /path/to/workspace serve-http

FastMCP's current HTTP defaults are:

• host: 127.0.0.1
• port: 8000
• path: /mcp

Useful CLI checks:

mcp-agent-handoff --workspace-root /path/to/workspace doctor
mcp-agent-handoff --workspace-root /path/to/workspace state
mcp-agent-handoff --workspace-root /path/to/workspace validate-decision-id --decision codex_slice_complete_plan0004_contract-pinning-and-docs --decision-kind slice_complete
mcp-agent-handoff --workspace-root /path/to/workspace review-findings --operation list
mcp-agent-handoff --workspace-root /path/to/workspace handoff-close-check

Slice-complete Decision IDs

Do not hard-code the slice-complete regex in client prompts or scripts. Treat get_handoff_state(sections="identity").data.limits.write.slice_complete_decision_id as the authoritative registry, and use validate_decision_id(decision=..., decision_kind="slice_complete") for side-effect-free preflight when composing ids outside close_slice(author_tag=..., work_ref=..., slug=...).

Valid: cdx_slice_complete_plan0004_contract_pinning_and_docs

Invalid: cdx_slice_complete_plan0004_contract-pinning-and-docs

Review Intake

Preferred path when agent-orchestrator-mcp is available:

1. get_latest_slice_review_packet
2. get_review_findings_summary or review_findings --operation list as needed

Handoff-only fallback:

1. load_session
2. search_handoff(queries=["slice_complete"], record_types=["decision"], limit=1)
3. get_verified_tests(task_ref=..., commit_sha=...)
4. review_findings(review={"operation":"list","status":"open"})

Use the MCP read surfaces above before inspecting .task-state/handoff.db directly. Drop to raw SQLite or filesystem inspection only when the required MCP tool is unavailable or when you are debugging MCP transport or serialization failures.

Source-tree execution without installation:

PYTHONPATH=src python -m agent_handoff_mcp --workspace-root /path/to/workspace serve-stdio

Token-Efficient Usage

The v2 envelope is already compact, but callers still save the most tokens by shaping read responses deliberately:

• Use get_handoff_state(sections="identity") for routine task checks instead of a full state fetch.
• Use detail="summary" on read surfaces such as get_handoff_state, load_session, review_findings, search_handoff, and artifacts when truncated text is acceptable.
• Use top_n_*, limit=, and fields= to cap read size instead of trimming large payloads client-side. load_session accepts top_n_touched_files (default 20, max 200) to bound the additive touched_files list.
• Read from the canonical data block — result["data"]["active"] etc. The legacy top-level mirror was removed in 0.3.0 and never returns.

Package-local guidance and examples live in docs/guides/token-efficient-usage.md.

Wire format note (≥0.3.0)

Starting in agent-handoff-mcp 0.3.0, MCP tool responses are native JSON objects on the wire, not JSON strings inside structured_content.result. Every handler is annotated -> dict and returns a real dict via _envelope(); FastMCP serialises it once. If you previously did json.loads(handoff_tool(...)) to parse a JSON string return value, drop the json.loads — the call returns a dict directly. If you previously read result.content[0].text from the MCP wire payload and parsed it, read result.structured_content directly instead.

The envelope field set is unchanged (ok, schema_version, tool, scope, data, mutation, artifacts, warnings, task_ref) and schema_version stays at 2. Only the wire-encoding moved from JSON-string-inside-JSON to a native nested object. See CHANGELOG.md for the full migration notice.

Client Adapter Shape

Installed console-script adapter:

{
  "name": "altcontext-mcp",
  "command": "mcp-agent-handoff",
  "args": ["--workspace-root", "/path/to/workspace", "serve-stdio"]
}

Source checkout adapter:

{
  "name": "altcontext-mcp",
  "command": "python3",
  "args": [
    "-m",
    "agent_handoff_mcp",
    "--workspace-root",
    "/path/to/workspace",
    "serve-stdio"
  ],
  "cwd": "/path/to/mcp-agent-handoff",
  "env": {
    "PYTHONPATH": "/path/to/mcp-agent-handoff/src"
  }
}

Tool Surface

mcp-agent-handoff now exposes a single unified MCP surface.

Legacy --tool-profile all|core|extended inputs are still accepted for compatibility, but they all expose the same tool set:

mcp-agent-handoff --workspace-root /path/to/workspace --tool-profile core serve-stdio

Unified surface

CLI name	MCP tool
state	get_handoff_state
set	set_handoff_state
validate-decision-id	validate_decision_id
event	record_event
next-actions	next_actions
review-findings	review_findings
review-runs	review_runs
handoff-close-check	handoff_close_check
render-handoff	render_handoff (kind=current_task/dashboard)
(no CLI)	load_session
(no CLI)	close_slice
audit-decisions	audit_decision_ids
export	export_handoff_state
import	import_handoff_state
archive	archive_task_state
task-status	update_task_status
artifacts	artifacts
handoff-search	search_handoff
get-verified-tests	get_verified_tests
get-archived-task	get_archived_task
record-file-touch	record_file_touch
get-touched-files	get_touched_files

record_event uses a typed event payload with event_kind="decision" | "test_result" | "blocker" so each variant keeps its own required fields. next_actions uses operation="list" | "add" | "update" | "complete" | "skip". review_findings uses operation="record" | "batch_record" | "update" | "list". review_runs uses operation="record" | "list" | "coverage".

Branch/worktree drift is surfaced through warnings on write responses. If you set AGENT_HANDOFF_ENFORCE_BRANCH=1, branch mismatches against the active task's target_branch fail before mutation on enforceable branches. Direct Python callers see BranchMismatchError; MCP clients receive the standard v2 error envelope with data.expected_branch and data.actual_branch.

CLI-only extras (not part of MCP registry): artifact-list, artifact-terms.

Surface classes:

• action: mutates canonical state; do not blind-retry
• query: read-only inspection of canonical state; usually safe to retry
• generator: derived output such as close checks, markdown renders, and metrics summaries

Troubleshooting

The fastest first check is still:

mcp-agent-handoff --workspace-root /path/to/workspace doctor

If startup succeeds but calls fail, check the workspace paths first: --workspace-root, --state-dir, --current-task-path, and --exports-dir must all point at the same workspace state.