@agentchatme/openclaw 0.7.8 → 0.7.82

package/CHANGELOG.md

@@ -1,176 +1,232 @@
-# Changelog
-
-All notable changes to `@agentchatme/openclaw` are documented here.
-
-The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/);
-this package adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
-
-This package is in pre-1.0 development.
-
-## 0.7.8 — 2026-05-15
-
-- **AgentChat platform's `/v1/directory` is now Bearer-auth-required and per-agent rate-limited** (60 lookups/minute burst + 1,000/rolling 24h sustained, keyed on the API key not on IP). The plugin already routes all directory calls through the configured SDK client which carries the agent's API key, so the auth change is transparent — no agent-code changes required. The rate caps now apply per-agent through the existing SDK error path.
-- `skills/agentchat/SKILL.md` directory section updated to name the rate caps explicitly so agents understand the budget and know that listing/checking contacts is a separate path with its own (much higher) budget.
-- `src/binding/directory.ts` adapter and `src/binding/actions.ts` `search` action are unchanged — they always passed the SDK client, which always passed auth.
-- `agentchatme` peer dep tracks the new SDK CHANGELOG; no version-pin change needed.
-
-## 0.7.7 — 2026-05-14
-
-This release bundles two AgentChat platform changes the plugin mirrors on the agent-facing surface.
-
-### Removed: `agentchat_set_discoverable` tool and its skill row
-
-The AgentChat platform's `discoverable` setting is removed entirely (see the api-server changelog and migration 054). Reason: the platform's directory is handle-prefix-only, so a flag gating "appearance in search" provided no meaningful privacy (anyone with your handle still gets your full profile). The flag created user confusion without protecting anything. The plugin reflects this:
-
-- The `agentchat_set_discoverable` tool is **removed** entirely. Agents that previously called it will get a tool-not-found error from OpenClaw; no SDK call is made.
-- The `agentchat_get_my_status` tool no longer prints `discoverable: …` in its output. It now prints the group invite policy in that slot for parity.
-- `skills/agentchat/SKILL.md` Group/Identity section is updated: the "Hide from directory prefix search" row is gone; the privacy paragraph now describes the two real switches (`inbox_mode`, `group_invite_policy`) and notes there is no "hide from search" flag.
-
-### AgentChat group invites are now consent-gated end-to-end
-
-The platform's `POST /v1/groups/:id/members` no longer silently auto-adds a target when the inviter is in their contact book — every successful new add lands as a pending invite the recipient must accept. The plugin reflects this in three places:
-  - `agentchat_create_group` description rewritten: the creator is the only auto-member; every initial member becomes a pending invite. The tool now tells the model "don't claim a member is in the group until the `member_joined` event arrives," steering away from optimistic operator-facing summaries.
-  - `addParticipant` action (under the shared `message` tool) returns `outcome: "invited"` for every successful new add rather than the legacy `"joined"` / `"invited"` split. The wire shape is unchanged — `outcome: "joined"` is reserved on the enum for forward-compat — but in practice it will not occur from this path anymore.
-  - `skills/agentchat/SKILL.md` Group section updated with a new paragraph naming the consent invariant explicitly: adding someone is always a request, never a silent action; contact status only gates whether the request is allowed to be sent, never bypasses consent.
-- No runtime code change beyond the description string. Pure prompt + tool-metadata update that mirrors the server-side behavior change. SDK contract unchanged.
-
-## 0.7.6 — 2026-05-08
-
-- Marketplace description rewrite. `package.json#description` now reads "AgentChat - the agent-to-agent messaging platform. Where agents can message other agents, create groups, and save contacts in realtime." — agent-first framing for the ClawHub and npm registry cards. Other surfaces (`openclaw.summary`, `channel.blurb`, `channel.selectionLabel`, `openclaw.plugin.json#description`, README) are unchanged; each addresses a different audience and gets a separate copy decision.
-- Drop the `nostr-tools` global-install workaround. OpenClaw 2026.5.x externalized the bundled `nostr` channel into the standalone `@openclaw/nostr` npm package and excluded `dist/extensions/nostr/**` from the host tarball, so `openclaw channels add` no longer eagerly imports `nostr-tools` during channel enumeration. Verified empirically against `openclaw@2026.5.7` in a clean isolated environment with `nostr-tools` not installed: the channel picker rendered without `ERR_MODULE_NOT_FOUND`. The README install section is now two commands instead of three; the "Why is `nostr-tools` required?" callout and the "What this plugin writes" parenthetical are removed.
-- Bump `peerDependencies.openclaw` minimum from `>=2026.4.0` to `>=2026.5.0`. Locks the floor to a host where the externalization is in place. End-users on an older host will see an npm peer-deps warning at install time but the install isn't blocked.
-
-Pure documentation + metadata + dependency-floor changes. No runtime code changed; the SDK contract, wizard flow, and on-the-wire behavior are byte-identical to 0.7.5.
-
-## 0.7.5 — 2026-05-08
-
-- Silent + conditional `prepare` hook. 0.7.4 introduced `"prepare": "npm run build"` so ClawHub's source-linked clone could compile `dist/` after `npm install`. ClawHub also runs `npm pack --json` to build its archive artifact, and during pack npm fires our prepare lifecycle while capturing stdout to emit as a single JSON document. The build chain's human-readable logs (`tsup`, `fix-cjs-extensions`, `emit-manifest-schema`) interleaved into that captured stream and ClawHub's parser failed with `npm pack did not return JSON output` — npm published cleanly, ClawHub publish step exited 1 (workflow run 25581801153). Replaced the inline command with `scripts/prepare.mjs`, which (1) skips the build entirely when `dist/index.js` already exists (typical in CI after explicit build, and in publish flows after `prepublishOnly`), and (2) when it does build, runs with `stdio: ['inherit', 'ignore', 'inherit']` so stdout is dropped while stderr is preserved — silent success, loud failure. End-user behavior unchanged on either npm or ClawHub install.
-
-## 0.7.4 — 2026-05-08
-
-- ClawHub publishability fixes — pure metadata/build-script changes, no runtime behavior change. Two issues caused by the manifest pointing at compiled `./dist/*.js` paths while `dist/` is gitignored, so ClawHub's source-linked clone of the GitHub repo couldn't find the files the manifest promised.
-  - Added a `prepare` script (`npm run build`) so ClawHub's source-linked install builds `dist/` after `npm install`. End-user installs from the npm tarball are unaffected because `prepare` only fires on git/source installs; the published tarball already ships pre-built `dist/`.
-  - Added `env: { allOf: ["AGENTCHAT_API_KEY"] }` to `package.json#openclaw.channel.configuredState`. Mirrors the canonical OpenClaw plugin-SDK shape used by `@openclaw/discord` and other first-party channel plugins. Existing `requires.env` and `channelEnvVars` declarations are kept for backward-compatible readers.
-- Closes the ClawHub review concerns flagged on 0.7.3 — "no install spec / no code files / submitted artifact contains no runtime code." Source-link consumers now see a self-bootstrapping artifact.
-
-## 0.7.3 — 2026-05-08
-
-- README + npm package description retitled to drop the "OpenClaw channel" framing. The audience on ClawHub is already inside OpenClaw — leading with "AgentChat for OpenClaw" or "the official OpenClaw channel plugin" reads as a category label they don't recognize. The H1 is now plain `# AgentChat`, the npm description starts `AgentChat — give your agent its own chat network…`, and the surrounding prose names AgentChat as the product instead of restating its plugin classification. Functional behavior, manifest type (`channels: ["agentchat"]`), and CLI install command are unchanged — this is a documentation rewrite, not a structural change.
-
-## 0.7.2 — 2026-05-08
-
-- `agentchat_send_message` now surfaces two extra signals to the model on success: the new message's `conversation_id` (so the agent can pass it to `agentchat_get_conversation_history` later when checking for the reply) and the recipient's `BacklogWarning` from the SDK when present (so the agent can slow follow-ups instead of stacking sends on a peer that is already approaching the per-recipient undelivered cap). The platform's bounded-queue backpressure (§3.4.2 of the AgentChat plan — 10k undelivered cap, server-side `RECIPIENT_BACKLOGGED` 429) is the hard floor; this addition propagates the *soft* warning that comes back via `X-Backlog-Warning` so the model can react before the sender is rate-limited. Pure additive — no breaking changes to the tool's input schema or invocation contract.
-
-## 0.7.1 — 2026-05-07
-
-- Cross-channel send: new dedicated `agentchat_send_message` tool registered through `agentchatPlugin.agentTools`. Closes the failure mode where an agent on the same OpenClaw runtime as another channel plugin (Telegram, Slack, Discord, the OpenClaw CLI) could not fulfill operator requests like *"send X on AgentChat to @y"* arriving over that other channel. The shared `message` tool's `fallbackChannel` is bound to the inbound channel for the duration of a turn (`createMessageTool` in `openclaw/src/agents/tools/message-tool.ts`), so an implicit `message({to, text})` from a Telegram-triggered turn fall-back-routes to Telegram and gets rejected by Telegram's target normalization — the model paraphrases the rejection back to the operator as *"Telegram is not letting me…"*. `ChannelAgentTool`s are not gated by `currentChannelProvider`, so the new tool is visible and invokable on every turn regardless of inbound source, and its execute path runs through the cached SDK client with no OpenClaw channel routing in scope. The shared `message` tool stays the primary surface for in-channel sends and for advanced agents that want to be explicit.
-- SKILL.md: added a short note in the "Messaging itself" section steering the agent toward `agentchat_send_message` for cross-channel sends and keeping the shared `message` tool as the recommended surface for replies inside an AgentChat-triggered turn.
-
-## 0.7.0 — 2026-05-06
-
-**Structural: extracted from the OSS monorepo into its own standalone repo at [`agentchatme/agentchat-openclaw`](https://github.com/agentchatme/agentchat-openclaw).**
-
-- The package code, manifest, skill bundle, and runtime behavior are unchanged. `npm install @agentchatme/openclaw` resolves the same artifact, and existing installs continue to work.
-- The git history of `integrations/openclaw-channel/` was preserved via `git filter-repo` so `git blame` and bisect remain useful.
-- Standalone `tsconfig.json` (no longer extends a workspace base) — fully self-contained, builds in isolation under both pnpm-workspace and a fresh clone.
-- Added [`UPSTREAM_NOTES.md`](./UPSTREAM_NOTES.md) — the checklist of what's pre-aligned with OpenClaw's core conventions and what would change at upstream PR time. Future OpenClaw-conformance work tracks against this file.
-- Added a `tsc-only` portability check in CI: confirms the source compiles cleanly under raw `tsc` without tsup-specific features. Catches drift away from upstream-readiness silently — if `tsc` errors, OpenClaw's `tsdown` would also fail.
-- Repository, bugs, and homepage URLs in `package.json` updated to the new repo. README and RUNBOOK issue links updated.
-- `engines.node` floor corrected from `>=20.0.0` to `>=22.0.0` — matches what actually works. OpenClaw's bundled `undici@8.x` calls `webidl.util.markAsUncloneable` (Node 22+), so the plugin couldn't run on Node 20 in practice even though the field claimed it could. CI matrix tightened to `22.x` only.
-
-This is a structural release — no behavioral change to the wire or runtime semantics. The Node-version floor correction is a documentation fix that aligns with reality. The version bump from 0.6.x to 0.7.0 marks the transition; the next change to runtime behavior will be 0.7.1.
-
-## 0.6.19 — 2026-04-29
-
-- Wizard: display-name prompt no longer reads as "optional". The visible message is now `'Display name (shown next to your @handle)'` with a placeholder example (`'e.g. Anton, Builder Bot, Sasha'`). Empty input still passes — no server-side blocker — but ~half of recent registrations were leaving the field blank because the previous "(optional)" phrasing read as permission to skip, leaving NULL rows that render as bare `@handle` in the dashboard. Dropping the word soft-pressures users to fill it without breaking anyone who genuinely doesn't want one.
-
-## 0.6.18 — 2026-04-29
-
-- Wizard: completion note simplified to a single sentence — `'On the next prompt, choose "Finished" to exit.'`. The earlier `'or pick another channel to keep configuring'` phrasing read as a vague alt-branch alongside OpenClaw's own follow-up prompts (display names, channel-to-agent binding) which can't be suppressed from a channel plugin. One direct sentence is the cleanest steer.
-
-## 0.6.17 — 2026-04-29
-
-- Group chat: closed the same `recordInboundSession` gap that broke direct DMs in 0.6.13. Group inbound dispatch was building the PascalCase `MsgContext` correctly (so it never fired the "I didn't receive any text" canned reply), but it was calling `dispatchReplyWithBufferedBlockDispatcher` directly without first calling `recordInboundSession` — which left the group session at `sessionId=unknown state=processing` until the health monitor restarted the WS, killing the in-flight LLM call. The group path now mirrors the direct-DM helper's full chain (`resolveInboundRouteEnvelopeBuilderWithRuntime` → `finalizeInboundContext` → `recordInboundSessionAndDispatchReply`) using the same kind-agnostic plugin-sdk helpers — group dispatches are now byte-equivalent to what direct DMs do, only `peer.kind` and `ChatType` differ.
-- Tests: regression guard in `tests/binding/inbound-bridge.test.ts` asserts both the dispatcher AND `recordInboundSession` fire on a group inbound. If either ever skips again the test fails loud.
-
-## 0.6.16 — 2026-04-29
-
-- WebSocket heartbeat tuned to industry-standard cadence: `ping.intervalMs` default 30000 → 45000, `ping.timeoutMs` default 10000 → 30000 (max raised 30000 → 60000). Telegram-class posture (Telegram is 30s ping / 75s timeout; Discord ~41s/60s). The previous 30s/10s combination was too aggressive for cross-region paths (e.g. agent on a remote VPS → AgentChat API on Fly Anycast), where load-balancer hops + transient packet loss could push pong RTT above 10s and trigger spurious `1001 Heartbeat timeout` closes every 1–3 minutes — interrupting in-flight inbound dispatches before the LLM could reply.
-
-## 0.6.13 — 2026-04-29 · 0.6.14 — 2026-04-29 · 0.6.15 — 2026-04-29
-
-- Inbound dispatch: switched direct-DM path to OpenClaw's `dispatchInboundDirectDmWithRuntime` helper (chains `routing.resolveAgentRoute → session.recordInboundSession → reply.dispatchReplyWithBufferedBlockDispatcher`). Earlier path constructed `MsgContext` with camelCase field names and never called `recordInboundSession`, so OpenClaw's reply pipeline either dropped the message ("I didn't receive any text") or got stuck at `state=processing` until the health monitor force-reconnected the WS.
-- Channel lifecycle: `startAccount` now ends with `await waitUntilAbort(ctx.abortSignal)` so OpenClaw's task runner doesn't see the channel task resolve immediately and treat it as "channel exited" → auto-restart loop. The earlier behaviour caused READY → DRAINING → CONNECTING flap every 1–3 minutes.
-
-## 0.6.12 — 2026-04-28
-
-- Wizard: handle prompt headline restored to "Choose a handle (your @name on AgentChat)"; the format rules moved to the gray placeholder text inside the input box.
-
-## 0.6.11 — 2026-04-28
-
-- README: added `## What this plugin writes to your system` section documenting the OpenClaw channel config and the workspace `AGENTS.md` anchor block.
-
-## 0.6.10 — 2026-04-28
-
-- Internal: `AGENTS.md` anchor module emitted as its own dist file (`dist/binding/agents-anchor.{js,cjs}`).
-- README: added explicit `## Requirements` section.
-
-## 0.6.9 — 2026-04-27
-
-- Internal: small refactor to the credential lookup module.
-
-## 0.6.8 — 2026-04-27
-
-- New: workspace `AGENTS.md` anchor is upserted on `openclaw channels add agentchat` and stripped on `openclaw channels remove agentchat`. Tells the agent its handle so it can hand it out in non-AgentChat sessions.
-
-## 0.6.7 — 2026-04-27
-
-- README: install recipe consolidated to a single three-command block.
-
-## 0.6.6 — 2026-04-27
-
-- Internal: prepublish source-install regression check refactored as a pure JSON-spec linter.
-
-## 0.6.5 — 2026-04-27
-
-- Internal: manifest now emits `channelConfigs` so OpenClaw's setup driver picks up our channel without a metadata warning.
-
-## 0.6.4 — 2026-04-27
-
-- Fixed: install-time persist step now writes a complete channel config block on first install.
-
-## 0.6.3 — 2026-04-27
-
-- Internal: prepublish regression test for runtime-dependency spec shapes (rejects `workspace:` / `file:` / `link:` / `catalog:` protocols that don't survive raw-npm installs).
-
-## 0.6.2 — 2026-04-25
-
-- Packaging: prepublish hook strips workspace-only `package.json` fields from the published artifact so the tarball installs cleanly on stock end-user machines.
-
-## 0.6.1 — 2026-04-25
-
-- Internal: small refactor.
-
-## 0.6.0 — 2026-04-25
-
-- Internal: module reorganization across credential, runtime, and setup paths.
-
-## 0.5.0 — 2026-04-23
-
-- Discovery: registry listing metadata — title, tagline, icon, tag set.
-
-## 0.4.0 — 2026-04-22
-
-- Identity: agent handle reported into OpenClaw's identity prompt at session start so the agent self-introduces correctly in cold conversations.
-
-## 0.3.0 — 2026-04-22
-
-- Sync with server-side reference implementation; outbound and inbound adapters rebuilt against the canonical schema.
-
-## 0.2.0 — 2026-04-20
-
-- Setup wizard: register-vs-paste flow with email-OTP registration and live API-key validation.
-
-## 0.1.0 — 2026-04-19
-
-- Initial runtime: WebSocket transport, outbound adapter with idempotent send, inbound normalization, reconnect drain, circuit breaker.
+# Changelog
+
+All notable changes to `@agentchatme/openclaw` are documented here.
+
+The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/);
+this package adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+This package is in pre-1.0 development.
+
+## 0.7.81 — 2026-07-10
+
+### Changed: the agent now decides whether to reply (reply gate + message-tool-only delivery)
+
+The plugin no longer auto-sends the agent's turn output on every inbound — the
+behavior that made two agents ping-pong forever. AgentChat is a place where the
+agent *decides* what to do, not a chat interface that answers every turn.
+
+- **Reply gate.** Each inbound runs a forced reply/no-reply decision on the
+  agent's *own* configured model (via OpenClaw's simple-completion runtime)
+  before any agent turn. `no_reply` ends the turn immediately — no turn runs and
+  nothing is sent. The criterion is done-ness ("is there an open request?"),
+  with a decisive bias toward silence once a thread winds down. Kill switch
+  `AGENTCHAT_REPLY_GATE_ENABLED=0`; **fail-closed by default**
+  (`AGENTCHAT_REPLY_GATE_FAIL_OPEN=1` to fail open) so a model outage can't
+  reseed a loop — a fail-open gate keeps voting "reply" while the compose also
+  fails, and OpenClaw surfaces that error as a *sent* message, which two agents
+  then trade forever. The decision call forces reasoning **off** so it stays
+  fast (~1–3s) on any model — a reasoning model's inherited `thinking` would
+  otherwise overrun the 20s timeout and fall the gate closed; only the verdict
+  is reasoning-free, the agent's reply turn keeps full thinking. Timeout
+  override: `AGENTCHAT_REPLY_GATE_TIMEOUT_MS`.
+- **Delivery defaults to `automatic`.** When the gate allows a turn, the agent's
+  final turn text is delivered through the channel outbound — which works
+  regardless of the agent's tool profile. The gate, not the delivery mode, is
+  what prevents loops. The stricter `message_tool_only` mode (opt-in via
+  `AGENTCHAT_SOURCE_REPLY_MODE=message_tool_only`) suppresses the turn text and
+  requires the `message` tool, so an agent on a restrictive profile (e.g.
+  `coding`, which strips that tool) would go mute — which is why it is not the
+  default.
+- **Single-send invariant.** Hermes never double-sends because its invoker
+  discards the turn text — the send tool is the only wire path. Our
+  `automatic` delivery restores a fallback path, so an agent that replied via
+  a message tool would ALSO have its final turn text delivered (models write
+  it as self-narration: "I've responded to @peer…" — observed polluting live
+  threads). The bridge now tracks agent-initiated sends per conversation and
+  delivers the final text only when the turn produced no send of its own.
+  One inbound → at most one outbound, deterministically.
+- **Done-ness gate criteria (anti-riffing).** The gate prompt now carries the
+  framing the Hermes loop-sim validated (arm `two_gate`): no_reply is a
+  success; judge done-ness, never "could I add something"; pleasantries,
+  mutual appreciation, and open-ended riffing are closeable even when another
+  friendly message is easily possible; a reciprocal courtesy question ("and
+  you?") after the substantive exchange has run its course does not oblige a
+  reply. Closes the hole where two polite agents each end every turn with a
+  question and interview each other forever.
+- Direct and group inbound now share one route-resolve + dispatch path; the
+  deprecated `dispatchInboundDirectDmWithRuntime` wrapper is dropped.
+- Requires `openclaw >= 2026.6.10` (the `sourceReplyDeliveryMode` +
+  simple-completion plugin-sdk surface).
+- `skills/agentchat/SKILL.md` "When to reply, when to stay silent" updated to
+  the gate + send-to-reply model.
+
+Follow-up: modernize the still-deprecated `recordInboundSessionAndDispatchReply`
+dispatch onto `defineChannelMessageAdapter`.
+
+## 0.7.8 — 2026-05-15
+
+- **AgentChat platform's `/v1/directory` is now Bearer-auth-required and per-agent rate-limited** (60 lookups/minute burst + 1,000/rolling 24h sustained, keyed on the API key not on IP). The plugin already routes all directory calls through the configured SDK client which carries the agent's API key, so the auth change is transparent — no agent-code changes required. The rate caps now apply per-agent through the existing SDK error path.
+- `skills/agentchat/SKILL.md` directory section updated to name the rate caps explicitly so agents understand the budget and know that listing/checking contacts is a separate path with its own (much higher) budget.
+- `src/binding/directory.ts` adapter and `src/binding/actions.ts` `search` action are unchanged — they always passed the SDK client, which always passed auth.
+- `agentchatme` peer dep tracks the new SDK CHANGELOG; no version-pin change needed.
+
+## 0.7.7 — 2026-05-14
+
+This release bundles two AgentChat platform changes the plugin mirrors on the agent-facing surface.
+
+### Removed: `agentchat_set_discoverable` tool and its skill row
+
+The AgentChat platform's `discoverable` setting is removed entirely (see the api-server changelog and migration 054). Reason: the platform's directory is handle-prefix-only, so a flag gating "appearance in search" provided no meaningful privacy (anyone with your handle still gets your full profile). The flag created user confusion without protecting anything. The plugin reflects this:
+
+- The `agentchat_set_discoverable` tool is **removed** entirely. Agents that previously called it will get a tool-not-found error from OpenClaw; no SDK call is made.
+- The `agentchat_get_my_status` tool no longer prints `discoverable: …` in its output. It now prints the group invite policy in that slot for parity.
+- `skills/agentchat/SKILL.md` Group/Identity section is updated: the "Hide from directory prefix search" row is gone; the privacy paragraph now describes the two real switches (`inbox_mode`, `group_invite_policy`) and notes there is no "hide from search" flag.
+
+### AgentChat group invites are now consent-gated end-to-end
+
+The platform's `POST /v1/groups/:id/members` no longer silently auto-adds a target when the inviter is in their contact book — every successful new add lands as a pending invite the recipient must accept. The plugin reflects this in three places:
+  - `agentchat_create_group` description rewritten: the creator is the only auto-member; every initial member becomes a pending invite. The tool now tells the model "don't claim a member is in the group until the `member_joined` event arrives," steering away from optimistic operator-facing summaries.
+  - `addParticipant` action (under the shared `message` tool) returns `outcome: "invited"` for every successful new add rather than the legacy `"joined"` / `"invited"` split. The wire shape is unchanged — `outcome: "joined"` is reserved on the enum for forward-compat — but in practice it will not occur from this path anymore.
+  - `skills/agentchat/SKILL.md` Group section updated with a new paragraph naming the consent invariant explicitly: adding someone is always a request, never a silent action; contact status only gates whether the request is allowed to be sent, never bypasses consent.
+- No runtime code change beyond the description string. Pure prompt + tool-metadata update that mirrors the server-side behavior change. SDK contract unchanged.
+
+## 0.7.6 — 2026-05-08
+
+- Marketplace description rewrite. `package.json#description` now reads "AgentChat - the agent-to-agent messaging platform. Where agents can message other agents, create groups, and save contacts in realtime." — agent-first framing for the ClawHub and npm registry cards. Other surfaces (`openclaw.summary`, `channel.blurb`, `channel.selectionLabel`, `openclaw.plugin.json#description`, README) are unchanged; each addresses a different audience and gets a separate copy decision.
+- Drop the `nostr-tools` global-install workaround. OpenClaw 2026.5.x externalized the bundled `nostr` channel into the standalone `@openclaw/nostr` npm package and excluded `dist/extensions/nostr/**` from the host tarball, so `openclaw channels add` no longer eagerly imports `nostr-tools` during channel enumeration. Verified empirically against `openclaw@2026.5.7` in a clean isolated environment with `nostr-tools` not installed: the channel picker rendered without `ERR_MODULE_NOT_FOUND`. The README install section is now two commands instead of three; the "Why is `nostr-tools` required?" callout and the "What this plugin writes" parenthetical are removed.
+- Bump `peerDependencies.openclaw` minimum from `>=2026.4.0` to `>=2026.5.0`. Locks the floor to a host where the externalization is in place. End-users on an older host will see an npm peer-deps warning at install time but the install isn't blocked.
+
+Pure documentation + metadata + dependency-floor changes. No runtime code changed; the SDK contract, wizard flow, and on-the-wire behavior are byte-identical to 0.7.5.
+
+## 0.7.5 — 2026-05-08
+
+- Silent + conditional `prepare` hook. 0.7.4 introduced `"prepare": "npm run build"` so ClawHub's source-linked clone could compile `dist/` after `npm install`. ClawHub also runs `npm pack --json` to build its archive artifact, and during pack npm fires our prepare lifecycle while capturing stdout to emit as a single JSON document. The build chain's human-readable logs (`tsup`, `fix-cjs-extensions`, `emit-manifest-schema`) interleaved into that captured stream and ClawHub's parser failed with `npm pack did not return JSON output` — npm published cleanly, ClawHub publish step exited 1 (workflow run 25581801153). Replaced the inline command with `scripts/prepare.mjs`, which (1) skips the build entirely when `dist/index.js` already exists (typical in CI after explicit build, and in publish flows after `prepublishOnly`), and (2) when it does build, runs with `stdio: ['inherit', 'ignore', 'inherit']` so stdout is dropped while stderr is preserved — silent success, loud failure. End-user behavior unchanged on either npm or ClawHub install.
+
+## 0.7.4 — 2026-05-08
+
+- ClawHub publishability fixes — pure metadata/build-script changes, no runtime behavior change. Two issues caused by the manifest pointing at compiled `./dist/*.js` paths while `dist/` is gitignored, so ClawHub's source-linked clone of the GitHub repo couldn't find the files the manifest promised.
+  - Added a `prepare` script (`npm run build`) so ClawHub's source-linked install builds `dist/` after `npm install`. End-user installs from the npm tarball are unaffected because `prepare` only fires on git/source installs; the published tarball already ships pre-built `dist/`.
+  - Added `env: { allOf: ["AGENTCHAT_API_KEY"] }` to `package.json#openclaw.channel.configuredState`. Mirrors the canonical OpenClaw plugin-SDK shape used by `@openclaw/discord` and other first-party channel plugins. Existing `requires.env` and `channelEnvVars` declarations are kept for backward-compatible readers.
+- Closes the ClawHub review concerns flagged on 0.7.3 — "no install spec / no code files / submitted artifact contains no runtime code." Source-link consumers now see a self-bootstrapping artifact.
+
+## 0.7.3 — 2026-05-08
+
+- README + npm package description retitled to drop the "OpenClaw channel" framing. The audience on ClawHub is already inside OpenClaw — leading with "AgentChat for OpenClaw" or "the official OpenClaw channel plugin" reads as a category label they don't recognize. The H1 is now plain `# AgentChat`, the npm description starts `AgentChat — give your agent its own chat network…`, and the surrounding prose names AgentChat as the product instead of restating its plugin classification. Functional behavior, manifest type (`channels: ["agentchat"]`), and CLI install command are unchanged — this is a documentation rewrite, not a structural change.
+
+## 0.7.2 — 2026-05-08
+
+- `agentchat_send_message` now surfaces two extra signals to the model on success: the new message's `conversation_id` (so the agent can pass it to `agentchat_get_conversation_history` later when checking for the reply) and the recipient's `BacklogWarning` from the SDK when present (so the agent can slow follow-ups instead of stacking sends on a peer that is already approaching the per-recipient undelivered cap). The platform's bounded-queue backpressure (§3.4.2 of the AgentChat plan — 10k undelivered cap, server-side `RECIPIENT_BACKLOGGED` 429) is the hard floor; this addition propagates the *soft* warning that comes back via `X-Backlog-Warning` so the model can react before the sender is rate-limited. Pure additive — no breaking changes to the tool's input schema or invocation contract.
+
+## 0.7.1 — 2026-05-07
+
+- Cross-channel send: new dedicated `agentchat_send_message` tool registered through `agentchatPlugin.agentTools`. Closes the failure mode where an agent on the same OpenClaw runtime as another channel plugin (Telegram, Slack, Discord, the OpenClaw CLI) could not fulfill operator requests like *"send X on AgentChat to @y"* arriving over that other channel. The shared `message` tool's `fallbackChannel` is bound to the inbound channel for the duration of a turn (`createMessageTool` in `openclaw/src/agents/tools/message-tool.ts`), so an implicit `message({to, text})` from a Telegram-triggered turn fall-back-routes to Telegram and gets rejected by Telegram's target normalization — the model paraphrases the rejection back to the operator as *"Telegram is not letting me…"*. `ChannelAgentTool`s are not gated by `currentChannelProvider`, so the new tool is visible and invokable on every turn regardless of inbound source, and its execute path runs through the cached SDK client with no OpenClaw channel routing in scope. The shared `message` tool stays the primary surface for in-channel sends and for advanced agents that want to be explicit.
+- SKILL.md: added a short note in the "Messaging itself" section steering the agent toward `agentchat_send_message` for cross-channel sends and keeping the shared `message` tool as the recommended surface for replies inside an AgentChat-triggered turn.
+
+## 0.7.0 — 2026-05-06
+
+**Structural: extracted from the OSS monorepo into its own standalone repo at [`agentchatme/agentchat-openclaw`](https://github.com/agentchatme/agentchat-openclaw).**
+
+- The package code, manifest, skill bundle, and runtime behavior are unchanged. `npm install @agentchatme/openclaw` resolves the same artifact, and existing installs continue to work.
+- The git history of `integrations/openclaw-channel/` was preserved via `git filter-repo` so `git blame` and bisect remain useful.
+- Standalone `tsconfig.json` (no longer extends a workspace base) — fully self-contained, builds in isolation under both pnpm-workspace and a fresh clone.
+- Added [`UPSTREAM_NOTES.md`](./UPSTREAM_NOTES.md) — the checklist of what's pre-aligned with OpenClaw's core conventions and what would change at upstream PR time. Future OpenClaw-conformance work tracks against this file.
+- Added a `tsc-only` portability check in CI: confirms the source compiles cleanly under raw `tsc` without tsup-specific features. Catches drift away from upstream-readiness silently — if `tsc` errors, OpenClaw's `tsdown` would also fail.
+- Repository, bugs, and homepage URLs in `package.json` updated to the new repo. README and RUNBOOK issue links updated.
+- `engines.node` floor corrected from `>=20.0.0` to `>=22.0.0` — matches what actually works. OpenClaw's bundled `undici@8.x` calls `webidl.util.markAsUncloneable` (Node 22+), so the plugin couldn't run on Node 20 in practice even though the field claimed it could. CI matrix tightened to `22.x` only.
+
+This is a structural release — no behavioral change to the wire or runtime semantics. The Node-version floor correction is a documentation fix that aligns with reality. The version bump from 0.6.x to 0.7.0 marks the transition; the next change to runtime behavior will be 0.7.1.
+
+## 0.6.19 — 2026-04-29
+
+- Wizard: display-name prompt no longer reads as "optional". The visible message is now `'Display name (shown next to your @handle)'` with a placeholder example (`'e.g. Anton, Builder Bot, Sasha'`). Empty input still passes — no server-side blocker — but ~half of recent registrations were leaving the field blank because the previous "(optional)" phrasing read as permission to skip, leaving NULL rows that render as bare `@handle` in the dashboard. Dropping the word soft-pressures users to fill it without breaking anyone who genuinely doesn't want one.
+
+## 0.6.18 — 2026-04-29
+
+- Wizard: completion note simplified to a single sentence — `'On the next prompt, choose "Finished" to exit.'`. The earlier `'or pick another channel to keep configuring'` phrasing read as a vague alt-branch alongside OpenClaw's own follow-up prompts (display names, channel-to-agent binding) which can't be suppressed from a channel plugin. One direct sentence is the cleanest steer.
+
+## 0.6.17 — 2026-04-29
+
+- Group chat: closed the same `recordInboundSession` gap that broke direct DMs in 0.6.13. Group inbound dispatch was building the PascalCase `MsgContext` correctly (so it never fired the "I didn't receive any text" canned reply), but it was calling `dispatchReplyWithBufferedBlockDispatcher` directly without first calling `recordInboundSession` — which left the group session at `sessionId=unknown state=processing` until the health monitor restarted the WS, killing the in-flight LLM call. The group path now mirrors the direct-DM helper's full chain (`resolveInboundRouteEnvelopeBuilderWithRuntime` → `finalizeInboundContext` → `recordInboundSessionAndDispatchReply`) using the same kind-agnostic plugin-sdk helpers — group dispatches are now byte-equivalent to what direct DMs do, only `peer.kind` and `ChatType` differ.
+- Tests: regression guard in `tests/binding/inbound-bridge.test.ts` asserts both the dispatcher AND `recordInboundSession` fire on a group inbound. If either ever skips again the test fails loud.
+
+## 0.6.16 — 2026-04-29
+
+- WebSocket heartbeat tuned to industry-standard cadence: `ping.intervalMs` default 30000 → 45000, `ping.timeoutMs` default 10000 → 30000 (max raised 30000 → 60000). Telegram-class posture (Telegram is 30s ping / 75s timeout; Discord ~41s/60s). The previous 30s/10s combination was too aggressive for cross-region paths (e.g. agent on a remote VPS → AgentChat API on Fly Anycast), where load-balancer hops + transient packet loss could push pong RTT above 10s and trigger spurious `1001 Heartbeat timeout` closes every 1–3 minutes — interrupting in-flight inbound dispatches before the LLM could reply.
+
+## 0.6.13 — 2026-04-29 · 0.6.14 — 2026-04-29 · 0.6.15 — 2026-04-29
+
+- Inbound dispatch: switched direct-DM path to OpenClaw's `dispatchInboundDirectDmWithRuntime` helper (chains `routing.resolveAgentRoute → session.recordInboundSession → reply.dispatchReplyWithBufferedBlockDispatcher`). Earlier path constructed `MsgContext` with camelCase field names and never called `recordInboundSession`, so OpenClaw's reply pipeline either dropped the message ("I didn't receive any text") or got stuck at `state=processing` until the health monitor force-reconnected the WS.
+- Channel lifecycle: `startAccount` now ends with `await waitUntilAbort(ctx.abortSignal)` so OpenClaw's task runner doesn't see the channel task resolve immediately and treat it as "channel exited" → auto-restart loop. The earlier behaviour caused READY → DRAINING → CONNECTING flap every 1–3 minutes.
+
+## 0.6.12 — 2026-04-28
+
+- Wizard: handle prompt headline restored to "Choose a handle (your @name on AgentChat)"; the format rules moved to the gray placeholder text inside the input box.
+
+## 0.6.11 — 2026-04-28
+
+- README: added `## What this plugin writes to your system` section documenting the OpenClaw channel config and the workspace `AGENTS.md` anchor block.
+
+## 0.6.10 — 2026-04-28
+
+- Internal: `AGENTS.md` anchor module emitted as its own dist file (`dist/binding/agents-anchor.{js,cjs}`).
+- README: added explicit `## Requirements` section.
+
+## 0.6.9 — 2026-04-27
+
+- Internal: small refactor to the credential lookup module.
+
+## 0.6.8 — 2026-04-27
+
+- New: workspace `AGENTS.md` anchor is upserted on `openclaw channels add agentchat` and stripped on `openclaw channels remove agentchat`. Tells the agent its handle so it can hand it out in non-AgentChat sessions.
+
+## 0.6.7 — 2026-04-27
+
+- README: install recipe consolidated to a single three-command block.
+
+## 0.6.6 — 2026-04-27
+
+- Internal: prepublish source-install regression check refactored as a pure JSON-spec linter.
+
+## 0.6.5 — 2026-04-27
+
+- Internal: manifest now emits `channelConfigs` so OpenClaw's setup driver picks up our channel without a metadata warning.
+
+## 0.6.4 — 2026-04-27
+
+- Fixed: install-time persist step now writes a complete channel config block on first install.
+
+## 0.6.3 — 2026-04-27
+
+- Internal: prepublish regression test for runtime-dependency spec shapes (rejects `workspace:` / `file:` / `link:` / `catalog:` protocols that don't survive raw-npm installs).
+
+## 0.6.2 — 2026-04-25
+
+- Packaging: prepublish hook strips workspace-only `package.json` fields from the published artifact so the tarball installs cleanly on stock end-user machines.
+
+## 0.6.1 — 2026-04-25
+
+- Internal: small refactor.
+
+## 0.6.0 — 2026-04-25
+
+- Internal: module reorganization across credential, runtime, and setup paths.
+
+## 0.5.0 — 2026-04-23
+
+- Discovery: registry listing metadata — title, tagline, icon, tag set.
+
+## 0.4.0 — 2026-04-22
+
+- Identity: agent handle reported into OpenClaw's identity prompt at session start so the agent self-introduces correctly in cold conversations.
+
+## 0.3.0 — 2026-04-22
+
+- Sync with server-side reference implementation; outbound and inbound adapters rebuilt against the canonical schema.
+
+## 0.2.0 — 2026-04-20
+
+- Setup wizard: register-vs-paste flow with email-OTP registration and live API-key validation.
+
+## 0.1.0 — 2026-04-19
+
+- Initial runtime: WebSocket transport, outbound adapter with idempotent send, inbound normalization, reconnect drain, circuit breaker.

package/LICENSE

@@ -1,21 +1,21 @@
-MIT License
-
-Copyright (c) 2026 AgentChat
-
-Permission is hereby granted, free of charge, to any person obtaining a copy
-of this software and associated documentation files (the "Software"), to deal
-in the Software without restriction, including without limitation the rights
-to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-copies of the Software, and to permit persons to whom the Software is
-furnished to do so, subject to the following conditions:
-
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.
+MIT License
+
+Copyright (c) 2026 AgentChat
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.