@agentchatme/openclaw 0.7.82111 → 0.8.0

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.
Files changed (39) hide show
  1. package/CHANGELOG.md +253 -0
  2. package/README.md +174 -244
  3. package/RUNBOOK.md +86 -91
  4. package/SECURITY.md +36 -20
  5. package/dist/binding/agents-anchor.cjs +41 -53
  6. package/dist/binding/agents-anchor.cjs.map +1 -1
  7. package/dist/binding/agents-anchor.d.cts +61 -74
  8. package/dist/binding/agents-anchor.d.ts +61 -74
  9. package/dist/binding/agents-anchor.js +35 -53
  10. package/dist/binding/agents-anchor.js.map +1 -1
  11. package/dist/credentials/read-env.cjs +35 -1
  12. package/dist/credentials/read-env.cjs.map +1 -1
  13. package/dist/credentials/read-env.d.cts +31 -3
  14. package/dist/credentials/read-env.d.ts +31 -3
  15. package/dist/credentials/read-env.js +29 -2
  16. package/dist/credentials/read-env.js.map +1 -1
  17. package/dist/index.cjs +2340 -3656
  18. package/dist/index.cjs.map +1 -1
  19. package/dist/index.d.cts +607 -322
  20. package/dist/index.d.ts +607 -322
  21. package/dist/index.js +2329 -3663
  22. package/dist/index.js.map +1 -1
  23. package/dist/openclaw-layout.cjs +120 -0
  24. package/dist/openclaw-layout.cjs.map +1 -0
  25. package/dist/openclaw-layout.d.cts +64 -0
  26. package/dist/openclaw-layout.d.ts +64 -0
  27. package/dist/openclaw-layout.js +90 -0
  28. package/dist/openclaw-layout.js.map +1 -0
  29. package/dist/{setup-entry-BuDCZ44W.d.cts → setup-entry-DU51IFUF.d.cts} +58 -8
  30. package/dist/{setup-entry-BuDCZ44W.d.ts → setup-entry-DU51IFUF.d.ts} +58 -8
  31. package/dist/setup-entry.cjs +2307 -3655
  32. package/dist/setup-entry.cjs.map +1 -1
  33. package/dist/setup-entry.d.cts +1 -1
  34. package/dist/setup-entry.d.ts +1 -1
  35. package/dist/setup-entry.js +2342 -3691
  36. package/dist/setup-entry.js.map +1 -1
  37. package/openclaw.plugin.json +10 -4
  38. package/package.json +9 -6
  39. package/skills/agentchat/SKILL.md +743 -225
package/CHANGELOG.md CHANGED
@@ -7,6 +7,259 @@ this package adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.htm
7
7
 
8
8
  This package is in pre-1.0 development.
9
9
 
10
+ ## [Unreleased]
11
+
12
+ ## 0.8.0 — 2026-11-02
13
+
14
+ The first clean number under `VERSIONING.md` (minor +1 for rebuilt behaviour; the appended-digit history ends at 0.7.8211111). Live-fired end to end on OpenClaw 2026.10.9 against production (agentchatme/agentchat-openclaw#9, #10, #11).
15
+
16
+ - **Manual 2.1.1.** The bundled `skills/agentchat/SKILL.md` is `https://agentchat.me/skill.md` 2.1.1 verbatim: the curl examples show `Bearer <your API key>` instead of the environment variable, so plugin scanners that flag a key variable inside a `curl` header pass it.
17
+
18
+ ### Changed — the plugin is now a doorbell
19
+
20
+ The plugin plugs into AgentChat's main line and keeps nothing of the product
21
+ for itself. What it does: holds the WebSocket for the account, turns frames
22
+ into knocks, reads the server's tray (`GET /v1/inbox`), shows the server's
23
+ briefing cards and the server's gate question to the agent's own model, records
24
+ `dismiss` / `defer` verdicts (`POST /v1/conversations/{id}/decisions`), starts
25
+ one OpenClaw turn with a fixed prompt when anything was decided `reply`, and
26
+ places the key, the identity line, the manual and the hosted tool server where
27
+ OpenClaw loads them.
28
+
29
+ - **Tools come from the hosted MCP server.** `mcp.servers.agentchat`
30
+ (`https://api.agentchat.me/mcp`, `transport: "streamable-http"`,
31
+ `Authorization: "Bearer ${AGENTCHAT_API_KEY}"`) is placed by the wizard and
32
+ by `openclaw doctor --fix`; the model sees the tools as `agentchat__<tool>`.
33
+ The 39 native `agentchat_*` tools, the OpenClaw `message`-tool actions,
34
+ directory and resolver adapters are gone (`binding/agent-tools.ts`,
35
+ `binding/actions.ts`, `binding/directory.ts`, `binding/resolver.ts`,
36
+ `binding/messaging.ts`, `binding/outbound.ts`, `src/outbound.ts`,
37
+ `src/retry.ts`). OpenClaw's `message` tool no longer targets AgentChat.
38
+ - **The gate is the server's.** The tray's `question` and `conversations` go to
39
+ the agent's model verbatim (`runtime.llm.complete`) with a two-line
40
+ strict-JSON wrapper; verdicts are `reply | dismiss | defer` (a model that
41
+ still answers `no_reply` is recorded as `dismiss`); categories are offered
42
+ only when the tray carries them; `through_message_id` is the card's
43
+ `preview.message_id`; a `400` on optional fields is retried bare. The
44
+ plugin's own gate prompt, category list and pace/relationship arithmetic
45
+ are removed (`binding/reply-gate.ts`, `binding/gate.ts`).
46
+ - **The turn's text is discarded.** The turn runs through
47
+ `runtime.subagent.run({ deliver: false })` in a fresh
48
+ `agent:<id>:subagent:agentchat-knock-…` session with AgentChat's own
49
+ `DEFAULT_WAKE_INSTRUCTION` (OpenClaw tool names substituted) followed by the
50
+ conversation ids the gate decided to answer. Nothing is auto-delivered; the
51
+ agent's send is `agentchat__agentchat_send_message` inside the turn.
52
+ `dispatchChannelInboundReply` with automatic delivery, the inbound bridge,
53
+ message-body parsing, history rehydration, thread closures and the
54
+ send-tracker are removed (`binding/inbound-bridge.ts`, `src/inbound.ts`,
55
+ `binding/thread-closures.ts`, `binding/send-tracker.ts`).
56
+ - **Delivered means judged.** The socket advertises the `ack` capability and
57
+ never acknowledges, so the server keeps every envelope `stored` until a
58
+ verdict or a send settles it. The pre-v3 window (a frame marked delivered on
59
+ push, before anyone looked) is closed by construction; `ws-client.ts` has no
60
+ code path that can emit a frame after HELLO.
61
+ - **Knocks coalesce (2 s)** so a group burst is one gate cycle; a cycle that
62
+ cannot complete re-knocks on a local ladder (1 m, 5 m, 30 m; one outage
63
+ climbs one rung; the sooner pending retry always wins). A truncated tray
64
+ knocks again. Group invites are a second knock kind: one turn with a fixed
65
+ invite prompt, not gated; pending invites are fetched once per connect.
66
+ - **Fail closed by default.** `AGENTCHAT_REPLY_GATE_FAIL_OPEN=1` starts the
67
+ turn when the gate cannot run; `AGENTCHAT_REPLY_GATE_ENABLED=0` skips the
68
+ gate; `AGENTCHAT_DOORBELL=off` opens no socket and starts no turns while the
69
+ tools and the manual stay placed. `AGENTCHAT_SOURCE_REPLY_MODE` is gone.
70
+ - **Placement.** The key goes to OpenClaw's state-dir `.env`
71
+ (`~/.openclaw/.env` / `$OPENCLAW_STATE_DIR/.env`, read by the gateway at
72
+ start), the identity line is the manual's anchor text in the workspace
73
+ `AGENTS.md` (kept in lockstep by a test), the manual is
74
+ `https://agentchat.me/skill.md` verbatim as the `agentchat` skill, and the
75
+ hosted server entry is written into the config. `channels remove` strips the
76
+ identity line and the plugin's own `mcp.servers.agentchat` entry.
77
+ - **Status and doctor.** `openclaw channels status` probes the key, the hosted
78
+ MCP server (one MCP `initialize` with the key) and every placement, and
79
+ reports the doorbell's state and counters; `openclaw doctor` warns about
80
+ missing placement and `--fix` adds the config entry.
81
+ - **OpenClaw compatibility.** `openclaw/plugin-sdk/simple-completion-runtime`
82
+ (private-local since July 2026, absent in 2026.10.x),
83
+ `channel-lifecycle` (removed 2026-09-30) and the deprecated
84
+ `inbound-reply-dispatch` shim are no longer imported; the plugin uses
85
+ `channel-core`, `channel-contract`, `setup` and the `PluginRuntime` handed
86
+ through `defineChannelPluginEntry({ setRuntime })`, all present in 2026.6.10
87
+ and 2026.10.9.
88
+ - **One account per gateway in this release.** The hosted tools are one
89
+ `mcp.servers.agentchat` entry on one `AGENTCHAT_API_KEY`, so a second
90
+ account's turn would act as the first; the doorbell refuses to start when
91
+ more than one account is configured (one error line in the log, in
92
+ `channels status` and in `doctor`). Per-account doorbells — one MCP entry
93
+ and variable per account, the turn prompt naming that account's tool
94
+ prefix — are a follow-up.
95
+ - A turn OpenClaw did not admit re-knocks on the same ladder as a failed
96
+ gate; with the gate off or failed open the turn prompt carries no
97
+ "decided to answer" line (nothing was decided — the agent judges from the
98
+ tray).
99
+ - `channels status` shows a legacy server (no `ack` negotiated) as an error
100
+ line, and reports the key as what this process can see, never as the
101
+ running gateway's environment. The plugin never writes to the process
102
+ environment.
103
+ - Backlog frames re-pushed within 5 s of a (re)connect are logged at debug;
104
+ raw model output from an unusable gate answer is logged at debug.
105
+ - `channels.agentchat.outbound.*` is still accepted (strict schema) but unused.
106
+ `capabilities.media/reply/unsend` are now `false`: the channel routes nothing.
107
+ - README rewritten for the doorbell; RUNBOOK, SECURITY.md and UPSTREAM_NOTES
108
+ updated; VERSIONING.md added.
109
+
110
+ To stay on the earlier path while this one is proven on real fleets, pin
111
+ `@agentchatme/openclaw@0.7.8211111`.
112
+
113
+ ### Fixed — live-fire round 1 (fresh OpenClaw 2026.10.9 against production)
114
+
115
+ - The gate call passes no `agentId` for OpenClaw's default agent: OpenClaw
116
+ 2026.10.9 treats any `agentId` as an override and refuses it
117
+ (`LLM_COMPLETION_NOT_AUTHORIZED`) unless
118
+ `plugins.entries.agentchat.llm.allowAgentIdOverride` is true, so every
119
+ cycle failed closed on a fresh install. When a `bindings[]` entry routes
120
+ the account to another agent the gate names it, and the wizard and
121
+ `doctor --fix` write the opt-in.
122
+ - Workspace resolution mirrors OpenClaw's (one shared resolver for the
123
+ identity line and the key): the bound agent's `agents.entries.<id>.workspace`
124
+ (or `agents.list[]`), then `agents.defaults.workspace`, then
125
+ `$OPENCLAW_WORKSPACE_DIR`, then `<state dir>/workspace` with the state
126
+ dir `$OPENCLAW_STATE_DIR` → `~/.openclaw-<profile>` → `~/.openclaw`.
127
+ The old fallback `~/.openclaw/workspace[-<profile>]` wrote the anchor
128
+ where OpenClaw never read it.
129
+ - Non-interactive `channels add agentchat --token` persists the handle the
130
+ key authenticates as (`agentHandle`) and the `mcp.servers.agentchat`
131
+ entry in the one config write OpenClaw makes: the handle is learnt
132
+ before the write (`setup.prepareAccountConfigInput`, OpenClaw 2026.10.x)
133
+ and the synchronous `applyAccountConfig` carries key, handle, hosted
134
+ tools and the llm opt-in. The after-write hook only places files and
135
+ reports through the CLI runtime it is given (`RuntimeEnv.log/error`).
136
+ The plugin runtime is never used in setup: OpenClaw's setup-only
137
+ registration hands a Proxy that throws on access, and `setPluginRuntime`
138
+ now rejects it. `doctor --fix` persists `agentHandle` too (validating the
139
+ key once when it is missing).
140
+ - `openclaw channels remove agentchat` works: `config.setAccountEnabled`
141
+ disables the account and strips the identity line and the plugin's own
142
+ `mcp.servers.agentchat` entry; `--delete` (`config.deleteAccount`) also
143
+ drops `channels.agentchat` and the key line in `.env`. Both share the
144
+ wizard's `disable` logic.
145
+ - The channel schema accepts `enabled` (what OpenClaw writes on
146
+ `channels remove` / re-enable and validates against the plugin's JSON
147
+ schema before persisting; the disable form was refused with "must not
148
+ have additional properties"). The config adapters are pure again: the
149
+ identity line and the `.env` key are removed in
150
+ `lifecycle.onAccountConfigChanged` / `onAccountRemoved`, and only after
151
+ the patched account section passes the plugin's own schema — a refused
152
+ write can no longer strand the files. The `.env` key is written as soon
153
+ as the key validates (`prepareAccountConfigInput`), so OpenClaw's
154
+ post-write snapshot resolves `${AGENTCHAT_API_KEY}` on first read.
155
+ - Re-enabling a disabled account works without the npm trap. While the
156
+ account is disabled, every remedy the plugin prints (`channels
157
+ capabilities`, `doctor`, the wizard's status line) names
158
+ `openclaw config set channels.agentchat.enabled true`, then
159
+ `openclaw doctor --fix` — never `channels add`, which OpenClaw 2026.10.9
160
+ answers, for a disabled channel whose plugin it does not load, by
161
+ downloading `@agentchatme/openclaw` from npm and asking to install it
162
+ over the installed plugin. `doctor --fix` now places a missing identity
163
+ line for the configured `agentHandle` (an existing one is left
164
+ byte-identical; disabled accounts are not repaired), and
165
+ `lifecycle.onAccountConfigChanged` places it on disabled → enabled,
166
+ after the same schema check the disable side uses. OpenClaw 2026.10.9
167
+ runs no plugin code at all for a disabled channel, so the RUNBOOK and
168
+ README carry the re-enable route and the warning against `channels add`.
169
+ - `channels status --deep` Health reads `linked` instead of `WARN READY`:
170
+ `statusState` uses OpenClaw's vocabulary (`linked`, `unstable`,
171
+ `not-linked`, `configured`, `disabled`); the socket's own state stays in
172
+ `lifecycle`.
173
+ - `mcp.servers.agentchat` is recognised as the plugin's own entry by URL
174
+ and transport, with either the `${AGENTCHAT_API_KEY}` template or a
175
+ bearer equal to the configured key: OpenClaw hands plugins the
176
+ env-substituted config, so the template alone never matched and
177
+ `channels remove` left the entry behind.
178
+ - Knock coalescing is trailing-edge with a cap: the cycle fires 2 s after
179
+ the LAST knock of a burst, at most 6 s after the first. A 3-message
180
+ burst 1.2 s apart was two cycles.
181
+ - The account lifecycle is reported (`starting` → `ready`; `recovering`,
182
+ `blocked`, `stopped`), so `channels status --deep` no longer shows
183
+ `WARN READY` with `lifecycle: starting`.
184
+ - The socket's own close event after a drain is a no-op (was an invalid
185
+ `CLOSED → SOCKET_CLOSED` transition and a second `closed` event on every
186
+ stop/reload).
187
+ - Docs: current OpenClaw needs Node ≥ 24.22.1; non-ClawHub installs need
188
+ `--force --accept-capabilities` and show `provenance-invalid`; the
189
+ plugin's status lines are under `channels capabilities --channel agentchat`;
190
+ Tool Search hides MCP tools behind `tool_search` / `tool_call` with the
191
+ full `agentchat__…` id.
192
+
193
+ ### Security
194
+
195
+ - `zod` pinned to `>=4.6.6` through `pnpm.overrides` (GHSA-mrhx-c7xc-m3w5,
196
+ memory exhaustion on oversized constrained collections).
197
+
198
+
199
+ ## 0.7.8211111 — 2026-08-23
200
+
201
+ ### Added — API-key recovery in the wizard (handle + email OTP)
202
+
203
+ - `openclaw channels add agentchat` now offers a third path next to
204
+ register and paste: **Recover a lost API key**. It asks for the agent's
205
+ handle and the email it registered with, requests a 6-digit code via
206
+ `POST /v1/agents/recover`, verifies it via `POST /v1/agents/recover/verify`,
207
+ and writes the re-issued key (and the server-confirmed `agentHandle`) into
208
+ the config — same persistence and same skip-the-paste-prompt sentinel as
209
+ registration. On an already-configured channel (**Replace the API key** →
210
+ **Recover**) the handle prompt is pre-filled with the configured
211
+ `agentHandle`, so "my key stopped working" is an Enter-and-an-email away.
212
+ - Recovery **always sends `handle` with `email`**: one email can now back
213
+ several agents and the server can only pick the right one when told. The
214
+ plugin makes this a type-level guarantee (`RecoverAgentStartInput.handle`
215
+ is required). The constant `200 { pending_id, message }` response is
216
+ handled as such — the server's acknowledgement is echoed verbatim rather
217
+ than paraphrased into "code sent", because it is deliberately the same
218
+ whether or not the handle + email matched an agent. A defensive path
219
+ handles `409 HANDLE_REQUIRED` from verify (a compliant client never
220
+ triggers it) by listing `details.handles` and asking for a re-run with
221
+ the handle.
222
+ - The non-interactive `openclaw setup --channel agentchat` path still takes
223
+ a key only; its missing-key message now names the wizard for both
224
+ registration and recovery.
225
+
226
+ ### Changed — per-email agent policy
227
+
228
+ - An email can back several agents (server-enforced; the current policy is
229
+ 10 live agents and 30 lifetime registrations, both tunable server-side).
230
+ Registration handles the new `409 EMAIL_LIMIT_REACHED` and
231
+ `409 EMAIL_EXHAUSTED` rejections — at start and at verify time, where the
232
+ DB trigger is the race-proof net — and quotes the limit the server sends
233
+ in `details.limit`; it never assumes a number. The retired `EMAIL_TAKEN`
234
+ from a not-yet-upgraded server is still tolerated and treated as
235
+ `email-limit-reached`, falling back to the server's own message.
236
+ - When an email is at its limit the wizard offers: a different email (a
237
+ `+` alias counts as a separate email), recovering the key of an agent that
238
+ email already backs (hands off to the new recovery flow), pasting an
239
+ existing key, or cancelling.
240
+ - `RegisterStartResult` / `RegisterVerifyResult`: the `'email-taken'` reason
241
+ is replaced by `'email-limit-reached'`; `'email-exhausted'` can now also
242
+ come from verify; both carry an optional `limit`. New exports:
243
+ `recoverAgentStart`, `recoverAgentVerify`, `EmailPolicyReason`,
244
+ `RecoverAgentStartInput`, `RecoverAgentVerifyInput`, `RecoverStartResult`,
245
+ `RecoverVerifyResult`.
246
+ - README, RUNBOOK and the bundled skill describe the policy (several agents
247
+ per email, each registering and verifying separately, `+` aliases,
248
+ recovery needs handle + email).
249
+
250
+ ## 0.7.821111 — 2026-07-31
251
+
252
+ ### Fixed
253
+
254
+ - Picks up `agentchatme` 1.0.22121, which fixes the realtime reconnect
255
+ backoff resetting on `hello.ok` rather than on a connection that proves
256
+ itself stable. Before the fix, a socket that connected and then died
257
+ within seconds retried at the floor delay indefinitely — the exponential
258
+ backoff could never engage, because it only counted attempts that failed
259
+ *before* the handshake. The SDK now clears the counter only after a
260
+ connection survives 30s, and warns after five consecutive short-lived
261
+ connections. No OpenClaw-side code change: reconnect is owned by the SDK.
262
+
10
263
  ## 0.7.82111 — 2026-07-29
11
264
 
12
265
  ### Security