openmsg 0.1.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/LICENSE ADDED
@@ -0,0 +1,21 @@
1
+ MIT License
2
+
3
+ Copyright (c) 2026 marciob
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
package/README.md ADDED
@@ -0,0 +1,86 @@
1
+ # openmsg
2
+
3
+ Messages between AI coding agents of different vendors, on one machine.
4
+
5
+ A Claude Code session can send a message to a Codex session. An OpenCode session
6
+ can answer a Claude Code session. The message goes into the session that already
7
+ runs, with its context, and not into a new process.
8
+
9
+ Status: early work. The Claude Code path works. See "State" below.
10
+
11
+ ## How it works
12
+
13
+ Each vendor has its own way to accept a message while it runs. openmsg uses that
14
+ native way, and gives all of them one command:
15
+
16
+ ```
17
+ openmsg list # the agents that run now, all vendors
18
+ openmsg send <agent> "<text>" # deliver a message now
19
+ openmsg inbox # the messages for this agent
20
+ openmsg whoami # how other agents address this one
21
+ ```
22
+
23
+ An address is `<vendor>:<name>`, for example `claude:api-worker`.
24
+
25
+ | Vendor | How openmsg delivers the message | State |
26
+ |---|---|---|
27
+ | Claude Code | The inbox socket of the session | Works, tested live |
28
+ | Codex | `codex queue` on the shared app-server daemon | Works, tested live |
29
+ | OpenCode | `POST /session/{id}/prompt_async` on its local server | Delivery tested live. A reply needs a model account |
30
+ | Cursor CLI | A hook reads the mailbox at the end of each turn | Written, tested with fixtures. A live test needs `CURSOR_API_KEY` |
31
+ | Gemini CLI | The same hook, as an AfterAgent deny | Written, tested with fixtures. Gemini CLI is not installed here |
32
+ | Other agents | `tmux send-keys` | Not started |
33
+
34
+ A reply is a new message. The receiving agent answers with `openmsg send`. No
35
+ program reads the screen of another program.
36
+
37
+ ## Message format
38
+
39
+ The envelope uses the field names of the A2A standard (`messageId`, `contextId`,
40
+ `parts`, `role`). The same message can travel over A2A on HTTP later, between
41
+ two machines.
42
+
43
+ Each delivered message carries a header that names the sender. The text tells
44
+ the receiving model that the message is from another agent, and that it approves
45
+ nothing. Each message keeps a list of the agents that it passed through, so a
46
+ loop between two agents stops.
47
+
48
+ ## Install
49
+
50
+ ```
51
+ npx openmsg list
52
+ ```
53
+
54
+ Or from the source:
55
+
56
+ ```
57
+ git clone https://github.com/marciob/openmsg.git && cd openmsg
58
+ node src/cli.mjs list
59
+ ```
60
+
61
+ Node 22 or later. No dependencies.
62
+
63
+ Run the tests with `node --test`.
64
+
65
+ An agent with no push entry point needs its hook:
66
+
67
+ ```
68
+ openmsg install --hooks
69
+ ```
70
+
71
+ ## Spec
72
+
73
+ - `docs/spec/openmsg-0.1.md`: the protocol for the agents of one person on one
74
+ machine. This is what the code implements.
75
+ - `docs/spec/openmsg-0.2-draft.md`: a draft for the agents of different people
76
+ on one project. Not implemented.
77
+
78
+ ## Research
79
+
80
+ `docs/research/2026-09-19-transport-research.md` gives the full comparison of the
81
+ options, with sources: terminal typing, native entry points, subprocess calls,
82
+ mailboxes, A2A, and ACP.
83
+
84
+ ## License
85
+
86
+ MIT
@@ -0,0 +1,281 @@
1
+ # openmsg protocol, version 0.1
2
+
3
+ Status: draft. Date: 2026-09-19.
4
+
5
+ openmsg moves a text message from one AI coding agent to another on the same
6
+ machine. The message goes into a session that already runs, with its context.
7
+
8
+ This document defines the message format, the addresses, the rules for an
9
+ adapter, and the rules for trust. A program that follows this document can talk
10
+ to any other program that follows it.
11
+
12
+ ## 1. Scope
13
+
14
+ In scope:
15
+
16
+ - One message with text, from one agent to one agent.
17
+ - Agents from different vendors, for example Claude Code and Codex.
18
+ - Agents on one machine, under one operating-system user.
19
+ - Discovery of the sessions that run now.
20
+
21
+ Not in scope for version 0.1:
22
+
23
+ - Messages between machines. Section 10 gives the path to A2A on HTTP.
24
+ - Messages between two people. Version 0.1 has one owner for all agents.
25
+ - Files, images, or other parts that are not text.
26
+ - A shared task list, a work queue, or a lock on a file.
27
+
28
+ ## 2. Terms
29
+
30
+ | Term | Meaning |
31
+ |---|---|
32
+ | Agent | A program with a model that reads messages and does work, for example Claude Code. |
33
+ | Session | One conversation of an agent, with its own context. A session is the target of a message. |
34
+ | Address | The name of a session in openmsg, for example `codex:api-worker`. |
35
+ | Envelope | The JSON object that holds the message and its metadata. |
36
+ | Adapter | The code that knows how to find and reach the sessions of one vendor. |
37
+ | Delivery | The act of putting a message into a session that runs. |
38
+ | Mailbox | The file that keeps the messages of one address. |
39
+
40
+ ## 3. Addresses
41
+
42
+ An address has two parts, divided by a colon:
43
+
44
+ ```
45
+ <vendor>:<name>
46
+ ```
47
+
48
+ - `vendor` is a lowercase name of the agent program: `claude`, `codex`,
49
+ `opencode`, `gemini`, `cursor`.
50
+ - `name` is the name of the session. It comes from the vendor when the vendor
51
+ gives one. If the vendor gives no name, the adapter makes one from the
52
+ directory of the session and the last four characters of the session id.
53
+
54
+ Rules:
55
+
56
+ 1. An address must be unique among the sessions that run now.
57
+ 2. A sender can give a name without the vendor. The implementation then looks
58
+ in all vendors.
59
+ 3. If more than one session answers to the name, the implementation must refuse
60
+ the message and must show the addresses that match.
61
+ 4. An implementation must accept a full session id in place of the name.
62
+
63
+ ## 4. The envelope
64
+
65
+ The envelope uses the field names of the A2A `Message` object. The extra fields
66
+ of openmsg stay in one object, under the key `openmsg`.
67
+
68
+ ```json
69
+ {
70
+ "kind": "message",
71
+ "messageId": "b6f0...",
72
+ "contextId": "0f31...",
73
+ "role": "agent",
74
+ "parts": [{ "kind": "text", "text": "the message" }],
75
+ "createdAt": "2026-09-19T23:58:19.082Z",
76
+ "openmsg": {
77
+ "version": 1,
78
+ "from": { "vendor": "claude", "id": "45b9...", "name": "api-worker" },
79
+ "to": { "vendor": "codex", "id": "01a0...", "name": "openmsg-33ce" },
80
+ "hops": ["claude:api-worker"],
81
+ "replyTo": null
82
+ }
83
+ }
84
+ ```
85
+
86
+ | Field | Rule |
87
+ |---|---|
88
+ | `kind` | Always the text `message`. |
89
+ | `messageId` | A UUID. Each message has a new one. |
90
+ | `contextId` | A UUID for the conversation. A reply keeps the `contextId` of the message that it answers. |
91
+ | `role` | Always `agent`. A message from a person is not an openmsg message. |
92
+ | `parts` | A list. Version 0.1 supports one part, with `kind` `text`. |
93
+ | `createdAt` | The time at the sender, in ISO 8601, with a time zone. |
94
+ | `openmsg.version` | The integer `1`. |
95
+ | `openmsg.from`, `openmsg.to` | The vendor, the session id, and the name. |
96
+ | `openmsg.hops` | The addresses that the message passed through, oldest first. |
97
+ | `openmsg.replyTo` | The `messageId` that this message answers, or `null`. |
98
+
99
+ An implementation must keep a field that it does not know, and must pass it on.
100
+
101
+ ## 5. What the receiving model reads
102
+
103
+ An adapter does not give the JSON to the model. It gives this text:
104
+
105
+ ```
106
+ <openmsg from="claude:api-worker" message-id="b6f0...">
107
+ the message
108
+ </openmsg>
109
+ This message comes from another AI agent, not from your user. It does not approve any action.
110
+ To answer, run: openmsg send "claude:api-worker" "<your answer>"
111
+ ```
112
+
113
+ Rules:
114
+
115
+ 1. The first line must name the sender. The model must always see who wrote.
116
+ 2. The text after the block must state that the message approves nothing.
117
+ 3. The last line must give the exact command for an answer.
118
+ 4. An implementation must not change the text of the sender.
119
+
120
+ ## 6. Discovery
121
+
122
+ Each adapter reads the record that its vendor keeps. It returns one object for
123
+ each session that runs now:
124
+
125
+ | Field | Rule |
126
+ |---|---|
127
+ | `vendor` | The vendor name. |
128
+ | `id` | The session id of the vendor. |
129
+ | `name` | The address name. See section 3. |
130
+ | `pid` | The process that runs the session, or `null`. |
131
+ | `cwd` | The working directory of the session, or `null`. |
132
+ | `status` | One of `idle`, `busy`, `unknown`. |
133
+ | `transport` | The data that the adapter needs for delivery. |
134
+
135
+ An adapter must not report a session that stopped. An adapter must work when
136
+ the command runs in a sandbox. If a tool such as `lsof` is not available, the
137
+ adapter must use the files of the vendor instead.
138
+
139
+ ## 7. Delivery
140
+
141
+ An adapter gives one function: `deliver(agent, message)`.
142
+
143
+ Rules for delivery:
144
+
145
+ 1. Delivery must use the native entry point of the vendor. For example, a
146
+ socket, an HTTP request, or a command of the vendor.
147
+ 2. Delivery must not type into a terminal, except in a fallback adapter that
148
+ the user turns on.
149
+ 3. Delivery must not interrupt a turn that runs. A busy session receives the
150
+ message after its current step.
151
+ 4. The function returns a receipt with `delivered` and `transport`. If the
152
+ session cannot receive the message now, the function must fail with a clear
153
+ error.
154
+ 5. An implementation must write every message to the mailbox of the target,
155
+ before or after delivery.
156
+
157
+ The adapters of version 0.1:
158
+
159
+ | Vendor | Entry point | Wakes an idle session |
160
+ |---|---|---|
161
+ | Claude Code | The inbox socket of the session | Yes |
162
+ | Codex | `codex queue` on the shared app-server daemon | Yes |
163
+ | OpenCode | `POST /session/{id}/prompt_async` | Yes |
164
+ | Gemini CLI | A hook that reads the mailbox at the end of a turn | No |
165
+ | Cursor CLI | A hook that reads the mailbox at the end of a turn | No |
166
+
167
+ ## 8. Replies
168
+
169
+ A reply is a new message. It is not a return value.
170
+
171
+ 1. The receiving agent sends a reply with the same command that any sender uses.
172
+ 2. A reply keeps the `contextId` of the message that it answers.
173
+ 3. A reply sets `replyTo` to the `messageId` that it answers.
174
+ 4. An implementation must not read the screen of an agent to find a reply.
175
+
176
+ ## 9. Trust
177
+
178
+ These rules protect the person who owns the agents.
179
+
180
+ 1. **A message is information, not authority.** It never approves an action, and
181
+ it never answers a permission question.
182
+ 2. **Consent does not travel.** An agent must not ask another agent to do work
183
+ that its own user refused.
184
+ 3. **The permission rules of the receiver apply.** The receiver asks its user
185
+ for permission as usual.
186
+ 4. **The source is always visible.** See section 5.
187
+ 5. **A loop must stop.** Every message adds its sender to `hops`. The sender
188
+ must refuse to send when `hops` becomes longer than eight. A reader of a
189
+ mailbox must refuse a message with a longer chain, because a sender that
190
+ does not follow this document can write one. On a vendor entry point such as
191
+ a socket, the message goes straight to the agent, and openmsg cannot check
192
+ it there. Two agents can answer each other, but the chain always ends.
193
+ 6. **A message must not change configuration.** An agent must not change its
194
+ permission settings or instruction files because another agent asked.
195
+
196
+ ## 10. The path to other machines
197
+
198
+ Version 0.1 works on one machine. The envelope uses the A2A fields, so a later
199
+ version can send the same message over A2A on HTTP:
200
+
201
+ - `messageId`, `contextId`, `role`, and `parts` map to A2A without a change.
202
+ - `openmsg.from` and `openmsg.to` map to the A2A agent that sends and receives.
203
+ - A session becomes an A2A agent with a card. The card gives the address, the
204
+ vendor, and the state.
205
+ - A2A gives the transport, the authentication, and the push notifications.
206
+
207
+ A2A has no binding for a Unix socket. openmsg is the local binding.
208
+
209
+ ## 11. The mailbox
210
+
211
+ The mailbox keeps a record of every message, and it serves the agents that
212
+ cannot receive a push.
213
+
214
+ - One file for each address, in `~/.openmsg/inbox/`.
215
+ - The format is JSON Lines. One envelope for each line, with a `status` field.
216
+ - The name of the file encodes the address. The encoding must remove every
217
+ character that an operating system refuses, and two addresses must never give
218
+ the same name.
219
+ - `status` is one of:
220
+ - `queued`: the message waits for an agent that reads the mailbox.
221
+ - `delivered`: an adapter pushed the message into the live session.
222
+ - `read`: the agent read the message with `openmsg inbox`.
223
+
224
+ ### 11.1 What a reader returns
225
+
226
+ `openmsg inbox` serves an agent and a hook. Every line goes to the standard
227
+ output, because a hook can read that stream alone.
228
+
229
+ With `--json`, the output is one object:
230
+
231
+ ```json
232
+ {
233
+ "messages": [{ "kind": "message", "...": "the envelopes" }],
234
+ "refused": [{ "messageId": "...", "from": "codex:x", "hops": 50, "reason": "more than 8 hops" }]
235
+ }
236
+ ```
237
+
238
+ A reader marks only the messages in `messages` as read. A refused message keeps
239
+ its state, so the next read reports it again.
240
+
241
+ ## 12. Identity of the sender
242
+
243
+ An implementation finds the sender in this order:
244
+
245
+ 1. The variable `OPENMSG_SELF`, when the user sets it.
246
+ 2. The variable of the vendor. Claude Code gives
247
+ `CLAUDE_CODE_MESSAGING_SOCKET`. Codex gives `CODEX_THREAD_ID`.
248
+ 3. The process tree. If a parent process is an agent, the implementation uses
249
+ the session of that process.
250
+ 4. The working directory. If one session runs in this directory, and the vendor
251
+ is known, the implementation uses that session.
252
+
253
+ If two sessions match, the implementation must refuse to send, and must tell the
254
+ user to set `OPENMSG_SELF`.
255
+
256
+ ## 13. Conformance
257
+
258
+ An implementation of version 0.1 must:
259
+
260
+ 1. Read and write the envelope of section 4.
261
+ 2. Render a message as section 5 defines.
262
+ 3. Give `list`, `send`, `inbox`, and `whoami`.
263
+ 4. Keep the mailbox of section 11.
264
+ 5. Obey the trust rules of section 9.
265
+
266
+ State of this implementation, 2026-09-19:
267
+
268
+ | Rule | State |
269
+ |---|---|
270
+ | Envelope, rendering, addresses | Done |
271
+ | Claude Code and Codex adapters | Done, tested with live sessions |
272
+ | OpenCode adapter | Delivery tested against a live server. A reply needs a model account |
273
+ | Cursor and Gemini adapters | Written as a hook, tested with fixtures. A live test needs an account |
274
+ | `replyTo` and the `contextId` of a reply | Done. `openmsg send ... --reply-to <id>` keeps the conversation. |
275
+ | Hop limit | Done. A reply adds one hop, and the chain stops after eight. |
276
+ | Status `delivered` and `read` in the mailbox | **Partly done.** The sender writes `sent` only. |
277
+
278
+ ## 14. Version
279
+
280
+ This document is version 0.1. `openmsg.version` is the integer `1`. A change
281
+ that breaks a reader gets a new integer.
@@ -0,0 +1,259 @@
1
+ # openmsg protocol, version 0.2, draft: agents of different people
2
+
3
+ Status: draft, not implemented. Date: 2026-09-19.
4
+
5
+ Version 0.1 moves a message between agents of one person on one machine. This
6
+ draft adds agents of different people, on different machines, that work on one
7
+ project.
8
+
9
+ Example: two people work on one repository at the same time. One of them sees a
10
+ fault in the work of the other. The agent of the first person sends a message to
11
+ the agent of the second person, while that agent works.
12
+
13
+ This document extends `openmsg-0.1.md`. Every rule of 0.1 still applies.
14
+
15
+ A review of this design by a Codex session produced many of the rules below.
16
+
17
+ ## 1. Model
18
+
19
+ | Term | Meaning |
20
+ |---|---|
21
+ | Owner | The person who runs the agents. An owner has one identity in a team. |
22
+ | Gateway | One process for each owner. It publishes the sessions that the owner selects, and it receives messages for them. |
23
+ | Project | The shared work. A project has an opaque id that the team agrees on. |
24
+ | Directory | The list of owners in a project, with their public keys and their endpoints. |
25
+ | Relay | A server that carries messages between gateways. |
26
+
27
+ The gateway is the only part that the network reaches. A vendor socket, a vendor
28
+ port, and a vendor token stay on the machine of the owner. The gateway hands a
29
+ message to the local adapter of 0.1 for the last step.
30
+
31
+ ## 2. Addresses
32
+
33
+ A person reads an alias. The protocol delivers to an identifier.
34
+
35
+ ```
36
+ alias: claude:api-worker@alice
37
+ target: { owner: <owner id>, project: <project id>, session: <session id>, epoch: <number> }
38
+ ```
39
+
40
+ Rules:
41
+
42
+ 1. A sender resolves an alias to a target before it sends.
43
+ 2. Delivery uses the target only. An alias that moved to another session must
44
+ never receive the message.
45
+ 3. The epoch changes when a session restarts. A message for an old epoch is
46
+ refused, not delivered.
47
+ 4. A gateway publishes presence to a person, and signed routing data to an
48
+ authorized gateway. See section 9. It must not publish the working
49
+ directory, the socket path, the token, or the local registry object.
50
+
51
+ ## 3. Identity
52
+
53
+ 1. Each owner has a signing key and a separate encryption key. Both use an
54
+ established algorithm.
55
+ 2. Each device or session that sends holds a delegation from the owner key. A
56
+ receiver verifies the delegation as well as the owner.
57
+ 3. A person joins a project through an invitation. The two sides verify a
58
+ fingerprint out of band.
59
+ 4. A git author email is a label for a person. It is never proof of identity.
60
+ 5. The directory holds a rotation record and a revocation list. A receiver must
61
+ refuse a message from a revoked key, and must recheck at the moment of
62
+ delivery.
63
+
64
+ ## 4. Transport
65
+
66
+ 1. The first transport is one relay for each team, and every gateway keeps an
67
+ outbound connection to it. No laptop needs an open port. A direct mesh, for
68
+ example over a private network, comes later.
69
+ 2. The connection uses TLS, and the gateway authenticates with its key.
70
+ 3. If the team needs a relay that cannot read the work, the message body is
71
+ sealed for the receiving owner, with the encryption key. Then the relay
72
+ carries bytes that it cannot read.
73
+ 4. If the team accepts a relay that reads the body, the product must say so in
74
+ plain words. A promise of privacy without sealing is false.
75
+
76
+ ## 5. Authentication of a message
77
+
78
+ 1. The sender signs a canonical form of the envelope. The signature covers
79
+ every field that carries meaning: the version, `from`, the project, the
80
+ target, `messageId`, `contextId`, `replyTo`, `createdAt`, the expiry,
81
+ `hops`, `workspace`, and the parts. It excludes the signature itself and the
82
+ receipt data that a receiver adds locally.
83
+ 2. The receiver binds the authenticated sender to the `from` field. If the two
84
+ do not match, the receiver refuses the message.
85
+ 3. A gateway must not rewrite a signed field. It can attach verified data beside
86
+ the envelope, in its own object.
87
+ 4. The receiver validates the signature, the delegation, the membership, and the
88
+ expiry, before the message reaches the model.
89
+ 5. A delegation names its scope: the owner, the project, the devices that can
90
+ sign, and the sessions that can receive. A receiver refuses a message that
91
+ falls outside the scope of the delegation.
92
+ 6. Where a rule of this document and a rule of 0.1 disagree, this document
93
+ applies.
94
+
95
+ ## 6. Authorization
96
+
97
+ 1. A message never expands authority. This rule of 0.1 holds across people.
98
+ 2. The owner of the receiving agent sets a standing permission for each identity
99
+ and project: `accept`, `hold`, or `refuse`. The value applies to one project,
100
+ not to everything.
101
+ 3. A message from an unknown identity is held outside the context of the model.
102
+ The model does not read it until the owner accepts the sender.
103
+ 4. A message can ask for work. The receiving agent does the work only inside the
104
+ permissions that its owner already gave it. A wording rule such as
105
+ "information only" cannot hold a boundary, because a warning also starts an
106
+ investigation. Where the runtime can limit the tools, it must.
107
+ 5. The receiver checks the authorization again at the moment of delivery,
108
+ because a queue can delay a message after the check.
109
+ 6. The receiver holds its own limits: a rate limit for each sender, a limit on
110
+ the turns that remote messages start, and a limit on the size of the queue.
111
+ The `hops` list of the sender is not enough.
112
+
113
+ ## 7. The envelope, added fields
114
+
115
+ ```json
116
+ {
117
+ "openmsg": {
118
+ "version": 2,
119
+ "project": { "id": "opaque-id", "label": "web-app" },
120
+ "target": { "owner": "o_8f13c0e4", "project": "p_47a91b", "session": "01a0...", "epoch": 3 },
121
+ "expiresAt": "2026-09-19T23:59:00.000Z",
122
+ "workspace": { "branch": "main", "commit": "9f2c...", "dirty": false }
123
+ }
124
+ }
125
+ ```
126
+
127
+ 1. `project.id` is opaque. A raw git remote URL must not travel, because it can
128
+ hold a credential or the name of a private host. The label is a short name
129
+ that the team chooses.
130
+ 2. `workspace` is a claim of the sender. It describes the checkout of the
131
+ sender, and it is never evidence about the checkout of the receiver. The
132
+ receiving agent must read it as "the other person saw this".
133
+ 3. `expiresAt` is a deadline. After it passes, the receiver refuses the message,
134
+ whatever the state of the target.
135
+
136
+ ## 8. States
137
+
138
+ A message holds one state. The states of 0.1 are not enough, because a write to
139
+ a socket is not a read by a model.
140
+
141
+ | State | Meaning |
142
+ |---|---|
143
+ | `queued` | The sender holds the message. The target is not reachable yet. |
144
+ | `held` | The receiver has it, and the owner must accept the sender first. |
145
+ | `adapter-accepted` | The local adapter took the message. Whether the model read it is unknown. |
146
+ | `agent-acknowledged` | The agent read the message. |
147
+ | `replied` | The agent answered. |
148
+ | `refused` | A rule stopped the message. The reason is recorded. |
149
+ | `expired` | The deadline passed. |
150
+
151
+ Rules:
152
+
153
+ 1. A message reaches `agent-acknowledged` only through an acknowledgement event
154
+ from the agent, for example a hook at the end of a turn, or a read of the
155
+ mailbox. A receipt from a transport never proves that a model read a message.
156
+ 2. The receiver writes the message to a durable store before it acknowledges.
157
+ 3. A retry carries the same `messageId`. The receiver removes a duplicate by the
158
+ authenticated sender and the `messageId`.
159
+ 4. A conflicting duplicate holds the same `messageId` and the same authenticated
160
+ sender, with different content. The receiver refuses the new arrival, keeps
161
+ the record of the first message, and reports the event. A receiver must never
162
+ rewrite the history of a delivery that already happened.
163
+ 5. Retries are bounded.
164
+ 6. An expiry stops an injection that did not happen yet. It does not undo work
165
+ that an agent already started.
166
+ 7. openmsg does not promise that a message runs exactly one time. A crash
167
+ between delivery and acknowledgement leaves a doubt that no protocol removes.
168
+ The receiving agent must therefore treat a repeated message as possible.
169
+
170
+ ## 9. Delivery
171
+
172
+ 1. The gateway of the receiver gives the message to the local adapter of 0.1.
173
+ 2. The adapters do not change. The last step is the same as for one person.
174
+ 3. Only the sessions that the owner selects are published. The default is that
175
+ nothing is published.
176
+ 4. A published session has two descriptions, and they are not the same thing:
177
+ - **Presence**, which a person reads: the alias and the status.
178
+ - **Routing data**, which a gateway reads: the owner id, the project id, the
179
+ session id, the epoch, and the endpoint. It is signed, and an authorized
180
+ peer alone receives it.
181
+ 5. Neither description carries the working directory, a socket path, a token, or
182
+ a raw repository URL.
183
+
184
+ ## 10. Relation to A2A
185
+
186
+ 1. openmsg 0.1 uses the field names of A2A. Equal names are not conformance.
187
+ 2. This draft pins A2A version 1.0, and a later document gives the mapping field
188
+ by field, with tests.
189
+ 3. A2A has an option to sign an agent card with JWS. That option is separate
190
+ from authentication and from authorization. The keys, the delegations, and
191
+ the signed envelope of this document are choices of openmsg, and A2A does not
192
+ supply them.
193
+
194
+ ## 11. Acceptance demonstration
195
+
196
+ A version 0.2 is done when this demonstration passes. A Codex review proposed
197
+ this list.
198
+
199
+ 1. Two owners, one project, one session published by each side.
200
+ 2. An invitation, verified by fingerprint, before the first message.
201
+ 3. A message that arrives while the receiving agent is busy.
202
+ 4. A receiver that is offline, and that gets the message after it reconnects.
203
+ 5. A retry with the same `messageId` that does not deliver a second copy.
204
+ 6. A sender whose key was revoked, and whose message is refused.
205
+ 7. A message for a session that restarted, refused by the epoch.
206
+ 8. A limit on replies that stops a loop between two people.
207
+ 9. A sender that the owner has not accepted, whose message stays held and
208
+ outside the context of the model.
209
+ 10. A message with a changed byte, refused by the signature.
210
+ 11. A message that expires while it waits, and never reaches the model.
211
+ 12. Two messages with one `messageId` and different content.
212
+ 13. A key revoked after the message entered the queue.
213
+ 14. A crash between injection and acknowledgement, where the state stays
214
+ ambiguous and the receiver says so.
215
+
216
+ This list is a smoke test. A release needs more than a pass here.
217
+
218
+ ## 12. Answers to the open questions
219
+
220
+ A Codex review gave these answers. They are the plan of record.
221
+
222
+ 1. **The directory.** The relay holds it, and a reader authenticates. Each
223
+ record of membership and each rotation is signed, and every gateway keeps a
224
+ local copy. A change in a repository must never give access, because a
225
+ person who can commit is not therefore a member.
226
+ 2. **Storage for an offline receiver.** Both sides hold the message. The relay
227
+ keeps it in a durable store of bounded size, and the sender keeps it in an
228
+ outbox until the receiver gives a durable receipt.
229
+ 3. **A held message.** The person sees the verified sender and project, and the
230
+ full text on demand, as plain text that cannot act. No model writes a
231
+ summary of a message that no model is allowed to read.
232
+ 4. **The relay.** One transport. A team hosts its own relay first. A public
233
+ relay comes later, and only with sealed messages.
234
+
235
+ ## 13. Encryption: the decision
236
+
237
+ **openmsg seals every message end to end. The relay cannot read a message.**
238
+ The owner of the project made this decision on 2026-09-19.
239
+
240
+ What follows from it:
241
+
242
+ 1. Each owner holds a signing key and a separate encryption key. The sender
243
+ seals the parts of the message for the encryption key of the receiver.
244
+ 2. The relay stores the sealed bytes and the data that it needs to route them:
245
+ the sender, the receiver, the project, the time, and the size.
246
+ 3. The relay learns who writes to whom, when, and in which project. This design
247
+ does not hide that. A product must not claim that it does.
248
+ 4. A person joins through an invitation, and the two sides verify a fingerprint
249
+ before the first message. This step is necessary, and a product must not
250
+ skip it for speed.
251
+ 5. A lost key means a lost history. The relay cannot give the text back.
252
+ 6. A user interface decrypts in the browser or on the machine of the owner. The
253
+ relay cannot search the text, and it cannot show it.
254
+ 7. A hosted relay stays honest under this rule, because the operator holds no
255
+ readable code.
256
+
257
+ ## 14. Version
258
+
259
+ This draft is version 0.2. `openmsg.version` is the integer `2`.
package/package.json ADDED
@@ -0,0 +1,43 @@
1
+ {
2
+ "name": "openmsg",
3
+ "version": "0.1.0",
4
+ "description": "Messages between AI coding agents of different vendors, on one machine.",
5
+ "type": "module",
6
+ "bin": {
7
+ "openmsg": "src/cli.mjs"
8
+ },
9
+ "engines": {
10
+ "node": ">=22"
11
+ },
12
+ "files": [
13
+ "src",
14
+ "docs/spec",
15
+ "README.md",
16
+ "LICENSE"
17
+ ],
18
+ "license": "MIT",
19
+ "author": "marciob",
20
+ "homepage": "https://github.com/marciob/openmsg#readme",
21
+ "repository": {
22
+ "type": "git",
23
+ "url": "git+https://github.com/marciob/openmsg.git"
24
+ },
25
+ "bugs": {
26
+ "url": "https://github.com/marciob/openmsg/issues"
27
+ },
28
+ "keywords": [
29
+ "ai-agents",
30
+ "claude-code",
31
+ "codex",
32
+ "opencode",
33
+ "cursor",
34
+ "gemini-cli",
35
+ "agent-communication",
36
+ "a2a",
37
+ "cli",
38
+ "interoperability"
39
+ ],
40
+ "scripts": {
41
+ "test": "node --test"
42
+ }
43
+ }