Skip to content

Releases

v0.3.11

Choose a tag to compare

@github-actions github-actions released this 25 Oct 17:37
b7cc2b8

Patch Changes

  • 271792c: A guest campaign can run with the allowlisted loadout, and create_chat always sends the first message.

    • Campaign loadout. A [[guest_campaigns]] entry takes loadout = "allowlisted" to give its keyed guests the tools an allowlisted DM gets under [security], browser included, instead of the reduced guest surface. The key's expiry, its caps and its campaign context still apply, and the guest section of the prompt stops describing a reduced tool surface those guests do not have. The default, "guest", is unchanged.
    • create_chat sends its first message. IMCore mints a new chat on this Mac only, and nobody else sees it until a message goes out, so a create with no message reported success for a group that existed on no one's phone. text is now required, the result names the chat it went to, and the description says the person asking is not added unless their handle is included.
  • 271792c: Edmund sees a message that arrives while he is working, and can answer it then instead of when the task ends.

    A message for a conversation with a turn in flight used to wait in a queue until the turn finished, unless the model thought to look. Over 30 days, 330 turns ran past 3 minutes (57 in groups, 35 of them past 10 minutes), 75 had someone waiting, and handoff_current_work was never called. In a group, the person waiting was often not the one who asked for the work.

    The CLI accepts a message mid-turn and hands it to the model at its next tool step, and in those long turns a step came every 4.9 s at the median. So the daemon now writes the message to the running worker as [SYSTEM — a new message arrived while you're working], with the sender and message id, and the prompt tells the model to answer it with send_message before its next tool call and carry on. In a probe on Opus, a second person's question was answered 19.7 s after they sent it, while the first person's three-step job ran to the end. Before, they would have waited for the whole job.

    • The worker writes a message only while a tool step is running, so it always joins the turn; the CLI's echo of it tells the worker the model saw it. If one misses the turn anyway, the worker is retired before the CLI can start a turn for it that the next message would be mistaken for.
    • A message the model saw mid-turn gets no follow-up pass and no turn of its own afterwards; one it did not see, or one parked by a turn that failed, is answered as before.
    • Photos and files still wait for the end of the turn, since only text can join it; a voice note goes as its transcript.
    • Scheduled and wake-up turns are unchanged: a message still waits for them (14 of 499 mid-turn arrivals in 30 days).

v0.3.10

Choose a tag to compare

@github-actions github-actions released this 25 Oct 16:29
f081dbb

Patch Changes

  • ed3c185: Sub-agents are stopped when they go quiet, not when they turn fifteen minutes old, keep what they found when stopped, and report to their parent once.

    • The reaper judges output, not age. It stopped every agent 15 minutes after spawn. Of the 35 it reaped, the 33 with logs had all written output within the last 151 seconds (29 within the last minute), so it was killing work in progress. The runner now stamps a heartbeat on each line its CLI prints, and the reaper stops an agent after 15 minutes without output or at a 90-minute cap. Finished agents ran at most 18.3 minutes and went at most 482 seconds without output. The in-runner team sweep no longer fails teammates that are still working.
    • A stopped agent keeps its partial result. The runner died on SIGTERM without writing anything. It now writes its last output with the reason it stopped, and the parent is told that reason.
    • One report per agent, one per team. The runner's exit, its SIGTERM handler, the reaper, a spawn failure and cancel could each settle the same agent and wake the parent. The row is now claimed atomically and only the claimant reports; a team's team-done is claimed the same way. A runner whose row was settled before it started does not start the model.
    • Cancel stops the CLI too. cancel_agent, cancel_team and edmund agents cancel signalled the runner's pid only; they now stop its process group.
    • Completions that land together wake the parent once. 85 of 234 completion wake-ups since May arrived within two minutes of another for the same chat, up to 19 in a row, each a full turn. A completion now waits 20 seconds for others (at most 90 seconds after the first) and they arrive as one event. Replayed over those 234: 49 fewer turns.
    • A sub-agent's background job runs inline. Its wake-up went to an agent: session that does not exist (262 times, all skipped) and the agent polled check_bg_job instead. Every such job finished within 62 seconds.
    • A bad agent_model is named as such. Ten agents failed at start on a model the CLI could not use, one of them the misspelled claude-sonnet-5.5. A malformed id is refused before anything spawns, and an unavailable one is reported as the configuration problem it is, not as a failed task.
  • ed3c185: Follow-ups to the sub-agent lifecycle change.

    • Dashboard cancel works again. It still called a store method the lifecycle change removed, so cancelling from the dashboard returned an error and left a pending agent running. It now takes the same path as cancel_agent, and the overview counts stuck agents by the reaper's own limits.
    • A follow-on the daemon spawns gets the configured agent settings. When the reaper settled a team's last member, the deep-research synthesizer it spawned ran on the default model, without plugins and under the wrong host-access policy, because the daemon's environment lacks the settings only MCP children get. The daemon now starts it from the same environment an MCP child of that conversation would have.
    • A stopped agent's result always says it was stopped. When the CLI's exit reached the runner before its own SIGTERM, the result was written without the note, and the parent heard "stopped" while reading what looked like a complete result.
    • A team's synthesizer is always reported. Two members settling at once could both see the team finished; one spawned the synthesizer while the other sent team-done before the synthesizer existed, and the synthesizer's result was then never reported. The follow-on and team-done decision is now taken under one lock.
    • Stopping kills the right process. The reaper and cancel stopped the pid they had read before settling, which could be empty for an agent that had just started; they now stop the process the settle found, and the reaper's "never started" no longer lands on an agent that has started.
    • A long tool call does not look like a hang. A sub-agent prints nothing while a tool runs, and its background tools now run inline (a video render polls for up to ten minutes), so the reaper could have stopped a healthy agent for silence. The agent's heartbeat is kept while its tool call runs.
  • c1d5e08: A conversation that Claude's safeguards refuse now moves to a fresh thread and gets answered, instead of failing every turn, and a failed /compact is reported as failed.

    • A refused Claude thread is dropped. The CLI already retries a flagged turn once and then falls back to the previous Opus. When the fallback's safeguards flag it too ("Opus 4.8's safeguards flagged this message"), the judgment is about the thread's whole history, so resuming it fails the same way. That error used to classify as unknown, so recovery retried the same thread; one DM failed four turns running. It now classifies as thread_refused, the class the Codex safety refusal already used. runModel forgets a refused thread for both CLIs, shuts down the warm worker still holding it, and keeps the caller from restoring the id. Recovery then answers on a fresh thread with the recent chat.
    • No retry on the refused thread. The Claude runner's in-band heal no longer drops the thread and then resumes that same id. The thread_refused healer also evicts the warm worker, because the pool reuses a session's worker whatever id the next spawn asks for.
    • A failed /compact is a failure. The CLI ends a failed compaction with a success result whose text is Error during compaction: ..., so the pool logged /compact ok and auto-compact logged "applied". Seven compactions of one DM failed that way while its context grew from 209k to 259k. The worker now reports that text as a failure, and a refused compaction re-anchors the thread instead of retrying it every turn.
  • b6d3ad1: Messages of about 42 bytes decode exactly, and group chatter Edmund was never asked about no longer counts as a reply he owes.

    • The decoder no longer reads the type tag as a length. macOS writes the visible string as NSString 01 94 84 01 2B <len> <utf8>. The decoder scans offsets after the marker and keeps the longest valid slice, and the + (0x2B) tag is itself a valid one-byte length of 43. Read that way it yields the length byte plus the first 42 bytes of text, which ties a 43-byte message and beats a 42-byte one, and also beats any longer message that trims below 43 characters. The fallback notice read back as "+hit a snag getting back to you, still on i". Candidates whose prefix follows the tag are now preferred; payloads without it keep the old rule. Run over the 50,531 rows of a live chat.db, 842 decode differently, all corrections: 802 had a framing character glued to the front, and 40 short messages ("? ") had been replaced by the attribute token "iI". No row decodes worse.
    • A group row with no routing record is nobody's. ownsRow treated an unrecorded row as owned by the main session, a default meant for old DM rows. In a group, every accepted row is recorded (live path, catch-up, address check), so an unrecorded one is the group talking among themselves. It was fed to recovery turns and counted toward the fallback notice, which sent "hit a snag" into a group nobody had asked anything. Group sessions now require the record; DMs are unchanged.
  • ed3c185: A resumed conversation sees today's memory, recall no longer stalls the daemon, and the docs stop describing a write guard that never ran.

    • Resumes render the current prompt. Since about CLI 2.1.287 a conversation recorded its system prompt on the first request and resent that record on every resume until it compacted, ignoring the prompt the harness passed. The person file and SOUL.md are in that prompt, so over 7 days (2,330 turns) the median turn ran on memory 1.7 h old, the 75th percentile 13.6 h, and 13% of turns on memory over a day old. Workers now pass --system-prompt-snapshot off; an unchanged prompt still reads from cache, and a persona edit reaches a cold resume on its next turn again.
    • The daemon keeps its recall cache. It was evicted after 10 idle minutes and reloaded in one synchronous block, up to 31 s from a cold disk: 19% of turns over 14 days waited more than 5 s, every other chat with them. The daemon now keeps it and loads it at boot in slices (longest block 0.9 s on the live index). The cache no longer holds row text, which only the returned hits need: 0.6 GB instead of 1.4–1.6 GB, and searches run in 21–109 ms instead of 80–323 ms.
    • No phantom guard. .claude/settings.json declared a write guard and a media-delivery hook. Workers read project settings only from their own sandbox, so neither ever ran, while the docs, comments and the prompt ("A path guard rejects Write/Edit/Bash mutations outside the sandbox") said it did. Enabling it would have blocked about 1,200 of 20,000 worker writes a month, much of it ordinary work in /tmp, and the delivery hook would have double-sent media the model already sends. Both are removed and the text describes the real boundary: disallowedBuiltinTools and the MCP server's path checks.
    • Ghost outcomes read every chat of a person. The outcome sweep looked for a reply only in the chat row the session recorded, so a reply from a person's other address scored the fire as ignored.
    • A bad agent_model shows at boot. The startup self-check validated only the main model; a malformed sub-agent model surfaced as one failed spawn_agent after another.
  • ed3c185: The ghost looks at chats once a conversation has settled, spreads its sweeps across every enabled chat, and its pre-screen judges with the ghost's own calibration.

    • Settled ticks instead of post-reply ticks. A reply used to schedule a ghost tick 60-120 seconds later. Over 2026-10-09..23, 288 of 394 ticks landed within 2.5 minutes of Edmund's own message, and the pre-screen declined nearly all of them as mid-conversation. The tick n...
Read more

v0.3.9

Choose a tag to compare

@github-actions github-actions released this 22 Oct 16:20
1a9d7e6

Patch Changes

  • d6f53a4: SMS conversations stop looking empty to the history tools and proactive outreach, scheduled reminders into a group say who is in it, and people without a contact entry can be named.

    • History tools read the SMS transcript. get_thread_context, search_history, get_message and catch_me_up asked chat.db, which has no rows for a Twilio conversation, so an SMS group holding dozens of messages answered "no earlier messages in this chat". They now read the SMS store for sms: sessions. list_attachments says SMS media is not downloaded instead of reporting none.
    • Proactive outreach reads it too. The ghost judged an SMS group a "fresh session with zero message history"; it now gets the transcript, and an SMS DM no longer borrows that person's iMessage reactions and attachments.
    • A cold wake-up into a group carries the roster. Scheduled, proactive and recovery wake-ups had the recent thread but not the Participants: line an inbound turn gets, so a reminder could address someone who had only been talked about. Inbound turns and wake-ups now share one roster helper.
    • name_person records what someone goes by when no contact entry names their number. Rosters and transcripts then show Pat (+1…) instead of a bare number. A learned name is display-only: it never answers displayName, which decides SMS admission. It cannot reuse a contact's name or name Edmund's own addresses, and it is vetted to letters, spaces and basic punctuation.

v0.3.8

Choose a tag to compare

@github-actions github-actions released this 22 Oct 15:39
f18d425

Patch Changes

  • 1bf4358: Offer a kitchen on the first food question, explain it at setup, and follow up through the first week.

    • The offer is actually made now. A chat with no kitchen used to get an error from every kitchen tool, carrying an offer gated on "the second or third time" somebody asked. Between 2026-08-17 and 2026-10-20 that error came back 43 times across about 30 food questions from 7 chats, and the offer was made twice. Now the tools answer normally: answer the question in full first, then offer in the same message, with what it is, what they get and what it takes.
    • Never hold the answer back for setup. The old text said not to answer "from nothing", and the one time it was followed somebody waited 46 minutes for a bake time.
    • Asked at most twice, and a no is final. data/kitchen/prospects.json records food questions, offers and a no (kitchen_onboard action:"decline"). An unanswered offer is made once more after a week, then never. An offer that chat.db shows never reached the chat does not count.
    • Small groups. A group of two or three is offered a kitchen only when Edmund judges they share one, such as a couple; a bigger group never is. start in a group also links each member's own chat, unless they already belong to a kitchen.
    • How it works, at setup. When the link goes out, Edmund explains the one habit (a photo of the receipt), that they can just ask what to cook, and that there is nothing to keep up.
    • First-week check-ins. start queues three wake-ups in the chat where setup happened: day 1 a dinner from their shelves, day 4 a receipt nudge, day 7 a short recap with an offer of dinner texts. Each reads the ledger when it fires (kitchen_onboard action:"checkin") and stays quiet when it is no longer needed.
  • 1bf4358: Security updates for three advisories at high: @modelcontextprotocol/sdk 1.32.1 (experimental tasks were not tied to the session that created them; one request within the default size limit could exhaust server memory during validation), and launch-editor 2.14.2 via override, reached only through @changesets/cli (an existence check bypassable on Windows). The main, computer-use and ghost MCP servers publish byte-identical tool lists under the new SDK.

v0.3.7

Choose a tag to compare

@github-actions github-actions released this 12 Oct 15:00
964c192

Patch Changes

  • 0539756: Pin imcore-bridge 0.2.4, so a fresh install gets message edits that land on macOS 26. The daemon declares an edit's compatibility text as NSAttributedString and silently dropped edits that sent an NSString; the bridge now passes the class the daemon declares.

v0.3.6

Choose a tag to compare

@github-actions github-actions released this 12 Oct 14:31
c1feeca

Patch Changes

  • 5aeffc2: Recover from a thread the provider refuses, stop overlapping recovery sweeps, and load Claude Code plugins into workers.

    • A refused thread is dropped, not retried. "This request was blocked by our safety systems" now classifies as thread_refused. The Codex runner forgets the thread the moment a turn is refused, and a healer drops one that was stored anyway, so the next turn starts fresh instead of failing the same way on every resume.
    • A cold recovery turn carries the recent chat, the way scheduled wake-ups do, so a turn whose thread was dropped knows what was being discussed.
    • One recovery per session at a time. A sweep that started during another's recovery turn used to answer the same messages again once the lock freed.
    • Recovered messages lose their inbound acks, so a restart inside the ack window cannot replay them as "you were offline".
    • [claude] plugin_dirs loads Claude Code plugins into every Claude worker and sub-agent that has a shell. Sandboxed and guest workers get none. A directory without .claude-plugin/plugin.json fails the boot.
    • Security updates for four advisories at high or critical: @modelcontextprotocol/sdk 1.31 (OAuth client could send credentials to a server-chosen authorization server), proxy-addr 2.0.8 (IP spoofing through IPv4-mapped IPv6), sharp 0.35.5 (librsvg CVE), and shell-quote 1.12 via concurrently 10.0.6 (command injection in quote()). Every MCP server publishes byte-identical tool schemas under the new SDK.

v0.3.5

Choose a tag to compare

@github-actions github-actions released this 25 Sep 03:51
261fd21

Patch Changes

  • e4773eb: MCP servers can be added to the assistant's workers from config, and the sky camera now uses one.

    • [mcp_servers.<name>] in config.toml. Workers run with --strict-mcp-config, so a server added with claude mcp add never reached them.

      • A configured server supports http, sse or stdio, and headers for a bearer token.
      • It goes to operator sessions by default. Add tiers = ["operator", "contact"] to reach contacts too.
      • Guests never get one, and names the harness uses itself are refused.
      • The Claude and Codex runners choose the loadout through one function.
      • The generated data/mcp*.json files are now mode 0600, since they can hold tokens.
    • The skycam skill uses the openskycam MCP. The sky camera's recordings now live on openskycam (SkyVision). The rewritten skill covers:

      • a photo of now from the newest recording, one to three minutes behind;
      • photos of past moments, footage, detections, and workflow runs.

      It downloads a frame and sends the file rather than pasting a signed link. The old skystream skill and the snap.sh / sky.sh scripts are removed.

v0.3.4

Choose a tag to compare

@github-actions github-actions released this 25 Sep 03:14
97ceabd

Patch Changes

  • 83e95c9: The group missed-name check now wakes the assistant less often by mistake, and a string of misfires can't feed itself.

    • Messages that open with another member's first name. Jev is told when a message starts with another member's first name. It gets only that fact, never the name. Senders reach Jev as letters, so it had no way to know "Sam, who's playing Sunday?" was said to Sam.
    • Right after he has spoken, both scores must be high enough. "Wants a reply" must reach 0.7 (it was 0.6), and "said to him" must now also reach 0.5. Over 840 labels, wrong wakes fell from 29 to 8, and real messages caught went from 87 to 82 of 100.
    • A budget per run of un-named wakes. Each wake right after he has spoken costs 1 minus how often a message with that score was really for him. The rate comes from a measured table: Jev's scores are not calibrated on this question, and a 0.75 was right about half the time. A run stops once it reaches streak_budget, 0.6 expected wrong wakes. A confident back-and-forth barely spends it. The budget resets when someone says his name or swipe-replies to him. The streak is read back from chat.db and the decision log, so a restart loses nothing.
    • He sees the run. His note says how many times in a row he has been woken without his name.
    • Calibration. scripts/addressing-calibrate.ts writes the measured table for the current question wording.

    New config keys: addressed_floor and streak_budget. The approach is in docs/group-addressing.md.

v0.3.3

Choose a tag to compare

@github-actions github-actions released this 25 Sep 02:36
779ad32

Patch Changes

  • 76c7713: After a restart, live messages are answered while the boot catch-up is still running. Before, the watcher started only after every catch-up turn had finished. On 2026-09-24 two long turns in one DM kept every other chat unanswered for 87 minutes.

    Catch-up now:

    • reads the backlog and writes a durable ack for each row;
    • hands the watcher its cursor at once, and runs the turns in the background;
    • merges each chat's orphans and backlog into one turn;
    • adds a live message to its chat's catch-up turn if that turn hasn't started, so the chat is answered once, oldest first.

    The recovery loops still start only once catch-up has drained.

  • 76c7713: Scheduled fires for one chat no longer hold up other chats. Before, the scheduler waited for each fire, a whole model turn, before starting the next. On 2026-09-25 a 38-minute background-job turn in one DM delayed the mirror's severe-weather check by 34 minutes, so it was skipped as stale. Each session's fires still run one at a time, in order.

  • 96fa970: Experimental [group_addressing], off by default. It lets a group message through when it's meant for the assistant but doesn't say his name. Code picks the candidates: a swipe-reply to one of his messages, a word within two letters of one of his names, or a message soon after he spoke. Jev (typesafe/jev-1.13 on OpenRouter) decides whether each one is for him. A message it lets through reaches his turn marked as un-named, and he still decides whether to answer. "shadow" mode only records decisions, to data/addressing.jsonl.

    On the hand labels, a 0.6 threshold caught 37 of 49 missed messages and woke him wrongly for 17 of 708 other messages. scripts/addressing-calibrate.ts reruns that measurement after the wording or model changes.

    Swipe replies now carry their parent. iMessage stores the parent in thread_originator_guid, which the watcher never read, so every inline reply arrived with no parent: 91 of 91 in 90 days.

  • 76c7713: The remember_about_subject outcome field now says it takes one word: worked, rejected, mixed or untested. The story goes in learned. The remember_about_person section field now lists its five sections and says preferences go in what-ive-learned. Edmund had been writing a sentence as the outcome (8 times since 2026-09-19) and naming a "preferences" section that doesn't exist.

  • 76c7713: The recall index's keyword (FTS) table is repaired only by the daemon, and the repair runs as one transaction. Before, every process that opened the store, including each MCP server and each generated image, rebuilt the table whenever the counts disagreed. Those rebuilds ran without a lock timeout and outside a transaction, and they interleaved with the daemon's writes. On 2026-09-24 the rowid map held 933 entries for 104,658 rows, so updates added duplicate entries and each semantic_search took 5–34 s.

    What changed:

    • The store opens through openDb, which sets busy_timeout.
    • The boot check also catches a same-sized map that points at the wrong documents.
    • The indexer skips pending.json, which the kitchen drain rewrites every pass, so it no longer re-embeds three files a minute.
  • 76c7713: A turn is no longer ended by a result that belongs to a turn the CLI started itself. When a session is resumed after its last process was killed with a background shell running, the CLI first delivers a "stopped" notification as its own turn and prints a result with num_turns 0. That happens before it reads our message. On 2026-09-25 that result ended a turn after 1 second. The model kept working untracked, and the memory governor evicted the "idle" worker in the middle of a render.

    Workers now:

    • pass --replay-user-messages;
    • stamp each message with a uuid;
    • skip a result with num_turns 0 that arrives before that message's echo.

    This applies to both the resident worker and the per-turn process.

  • 96fa970: The spend ledger now books each turn's actual cost. The Claude CLI's total_cost_usd is the running total for the whole model session, carried across process restarts. Turns, cron fires and proactive fires resume a conversation, and they were booking that total as each turn's cost. One DM logged $99.03 for a turn that cost $0.11, and the ledger summed to 4-11x the real spend.

    Those callers now pass the total as sessionTotalUsd, along with the model session id, and the ledger books the rise since that session's last total. When that can't be known, the cost is recorded as unknown (null):

    • the total dropped, because a killed process restored an older total;
    • or a resumed session has no earlier total in the ledger.

    The log field is now session_total instead of cost. scripts/spend-ledger-repair.ts repairs rows booked before the fix; it dry-runs by default, and --apply backs up spend.db first.

v0.3.2

Choose a tag to compare

@github-actions github-actions released this 24 Sep 20:43
295aac0

Patch Changes

  • 078c0e4: Run the SOUL.md size gate, and send know-how to subject notes instead of SOUL.md.

    The gate for SOUL.md was written on 2026-08-28 and never called by the daemon, so the file grew from 31 KB back to 43 KB of every system prompt. It now runs in the boot sweep and after every self-note. remember_about_self requires a section and is limited to character and facts about Edmund's life; recipes, tool quirks and mistakes go to remember_about_subject, which is now in the tool catalog. The output contract states once that everything in the final text ships and that KEEP_QUIET is the only way to send nothing.

  • e2d7027: A standing dinner text asks for a written recipe page on the day's first send again. The check read "any stored receipt" as a retry, and stored receipts still hold the last day it fired, so the page was requested once (2026-08-17) while the text promised it on 17 more days.

  • b07d6ed: Scheduled and proactive wake-ups that start without the session's model conversation now carry the recent thread, as inbound cold starts do. The runner's persona-edit log says what happens (the session resumes and picks up the new prompt when its worker next starts) instead of announcing a fresh start that never happened.