rogerwilco
Roger: received. Wilco: understood, and will comply.
One-line, typed, closed-vocabulary messages for handing work between AI agents, with a readback the sender verifies in code before it lets go of the work.
RW1 HO research-01 from=research to=writer item=gh:crox/site#42 ask=DRAFT due=20261002 pri=2 // hero copy, 120 words
RW1 RB writer-01 re=research-01 from=writer to=research item=gh:crox/site#42 ask=DRAFT due=20261002 h=f209197fa615
RW1 DN writer-03 re=research-01 from=writer to=research item=gh:crox/site#42 out=OK art=sha256:9f3e1c // v1, 118 words
Agents hand work to each other in prose today. Prose hides state, cannot be validated, and gets misread silently. rogerwilco puts the facts in slots, keeps reasoning in a bounded tail or an artifact, and makes the receiver read the handoff back before it owns it. The readback is checked in five tiers: receipt and integrity (roger), echo, comprehension and execution (wilco); before an irreversible act the handover is checked against the record, and after it the done report is checked against the plan. It is content, not transport: it rides inside an MCP tool result, an A2A part, a chat turn, a queue message or a log line.
Install
pip install rogerwilco # the codec and the `rogerwilco` command; standard library only
Until the package is on PyPI: pip install git+https://github.com/afieldofdreams/rogerwilco. The site at rogerwilco.crox.io has a two-minute start.
Three things in this repo
| Folder | What | Start here |
|---|
rogerwilco/ | The protocol as a Claude Agent Skill: SKILL.md, the spec, the reference codec (standard library only), profiles, conformance tests. Copy the folder into .claude/skills/ or upload it as a skill. | rogerwilco/README.md |
app/ | The hosted checks: REST endpoints, a remote MCP server (validate, gate, check_rb, make_rb, check_ho, check_dn, legend, to_json), the site and the docs. Runs at rogerwilco.crox.io. Add the MCP server by URL to any agent that has no shell. | app/README.md |
research/ | The evidence: a five-format wire study, a readback pilot with planted faults, a multi-agent month-end close with human-in-the-loop escalation (28 runs), a clinical translation study (144 translations against 422 gold assertions), a readback and set-back study (114 trials) and a pre-registered execution readback study (2,762 trials). Corpora, harnesses, raw logs and results pages. | research/README.md |
Quick start
# validate a line
python3 rogerwilco/scripts/rogerwilco.py validate "RW1 HO a-1 from=a to=b item=x:1 ask=DRAFT"
# sender side: check a readback across the five tiers
python3 rogerwilco/scripts/rogerwilco.py check-rb "<HO line>" "<RB line>" --profile pay
# before an irreversible act, the handover against the record; after it, the done report against the plan
python3 rogerwilco/scripts/rogerwilco.py check-ho "<HO line>" '{"amt": "5400.00", "acct": "40-12-88/61102938"}' --profile pay
python3 rogerwilco/scripts/rogerwilco.py check-dn "<HO line>" "<DN line>" --profile pay
# legend for another agent's system prompt (about 400 tokens for core, more with a profile)
python3 rogerwilco/scripts/rogerwilco.py legend --profile pay
# conformance
python3 rogerwilco/tests/test_conformance.py
# the hosted app, locally
pip install -r app/requirements.txt && uvicorn app.main:app --reload
Status
v0.1, draft. The core vocabulary is small on purpose; extend it through a profile, not by editing the core. Issues and profiles welcome.
Apache License 2.0, copyright Croxio Ltd; see LICENSE and NOTICE. Built by Adam Field. Lineage: M-Form (medical-record compression, 2025) to P-Form (knowledge-graph handoffs, April 2026) to rogerwilco (September 2026).
