v0.2.3

Changed

• Murmur is a group chat: a message is posted once and everyone in the thread can read it, but nobody is woken for a message that is not for them. A reply is stored once and delivered (with a push, in the parent's mode) only to whoever is waiting for it: the author of the message it answers plus the session an owner message was typed in (via); a reply inside the replier's own thread goes to the parent's recipients. Other participants get no delivery row and no push and read the reply with murmur_thread / murmur thread. Measured reason: make bench-consult (haiku, three sessions) cost 546,811 tokens per consult for murmur against 265,771 for native SendMessage, because every reply woke every participant and the Implementer received each reply twice (once in the wait, once as a pushed turn).
• murmur_wait returns the text of the replies addressed to the waiting handle (replies: id, seq, sender, verdict, delivered body with the truncation pointer). A reply that arrives while its recipient has a wait open on that thread is not pushed; it stays queued, is handed over in the wait result (condition met or timeout) and marked delivered. A reply that arrives with no wait open is pushed as before.
• murmur_thread takes an optional messageId (CLI murmur thread <id> --message <m>): the result is only that message and the replies to it, with full delivered bodies; a message not in the thread is a bad_request. The footer of every delivered message now points at it (replies to this: murmur_thread <t> messageId <m>) instead of the whole thread, and the wait handover is limited to replies to the thread's latest ask.
• The MCP instructions, the murmur_reply and murmur_wait descriptions, the opencode and Hummin adapter descriptions and the docs describe this routing.
• Claude Code never defers murmur's tools: murmur install claude writes "alwaysLoad": true into mcpServers.murmur (idempotent; re-run it on an existing install) and the MCP server lists every tool with _meta: { "anthropic/alwaysLoad": true }. Measured reason: each deferred tool cost a ToolSearch call, and a Claude Code call re-reads the whole context (about 38k tokens on haiku).
• Teaching the reply: the footer on every delivered message and the MCP instructions say plainly that the answer goes through murmur_reply with that messageId and that plain text never reaches the sender.
• murmur_send takes optional wait (all_replied or any_replied) and timeoutMs (same default and cap as murmur_wait): it sends and waits on that thread in the same call and returns { sent, wait } with states, verdicts and the replies' text. Asking is one tool call.
• to also accepts a session's display name (its CLI title, case-insensitive) or a name whose slug is a handle (Peer A resolves to peer-a). Handles and groups win; an ambiguous name is a bad_request listing the candidates.
• The /murmur command file for Claude Code and zcode sets disable-model-invocation: true: it is the owner's command, and a model that invoked it through the Skill tool sent nothing. Re-run murmur install claude / zcode to get it.
• make bench-consult counts a run as completed only when the Implementer's final answer holds both peers' verdicts and their answers really reached it, and reports API calls and ToolSearch calls per session. Result (haiku, 3 runs per setup, medians): plain native SendMessage completes 0 of 3 (268,272 tokens, 7 API calls), because its peers answer in plain text that never reaches the Implementer; native with the peers told to answer with SendMessage completes 3 of 3 (510,931 tokens, 13 API calls, 13.8 s); murmur completes 3 of 3 (362,879 tokens, 9 API calls, 10.1 s), 29 percent fewer tokens than the native setup that finishes.

Fixed

• A reply held back for an open murmur_wait is now only one the wait hands over (a reply to the thread's latest ask); a reply to an older ask, or to a reply, is pushed as usual instead of staying queued with no path to its recipient. A wait remembers the replies held for it and hands them all over even if the thread's latest ask changed meanwhile. Replies held for a wait whose connection closes before it resolves are pushed at once. A delivery whose push is already in flight is never handed over a second time by a new wait or the hook pull.
• The live rig (scripts/live/lib.mjs) sets MURMUR_AUTOSTART=0, so a hook that outlives cleanup cannot respawn a broker and none survives a live run.

Added

• make bench-consult (scripts/live/bench-consult.mjs): three real Claude sessions per run, plain native SendMessage, native with the peers told to reply with SendMessage, and murmur, with tokens, API calls, copies received and wall time from the sessions' own transcripts; results in docs/bench/consult.md and consult.json.
• Docs and repository files: README hero, demo and linked CLIs; a Diataxis docs navigation (getting started, using murmur, consult, design, concepts, reference); CODE_OF_CONDUCT.md, SUPPORT.md, MAINTAINERS.md, CODEOWNERS, issue forms, a pull request template, Dependabot, an OpenSSF Scorecard workflow, a docs CI workflow (markdownlint, typos, lychee) and SHA-pinned GitHub Actions.