Skip to content

chekusu/mokutori

v0.3.6

Receive pages scanned with the Mokutori app for iPhone and iPad straight in this agent: pair with a QR code, then read each batch of corrected page images and OCR text.

GitHub

Changelog

[Unreleased]

  • Claude Code can install the plugin from this repository as a marketplace (/plugin install mokutori --marketplace chekusu/mokutori-agent-plugin); .claude-plugin/marketplace.json lists the plugin at the repository root.
  • The README and manifests describe the Mokutori app for iPhone and iPad (it is universal since 1.0) and link to the App Store.

[0.3.6] - 2026-11-07

  • Batches no longer go missing late in a long scan. A claimed job carries a transcript of its conversation, and the gateway reads every earlier message of the conversation to build it, so late in a long conversation of page batches a claim took longer than the engine's 15-second request limit. The client gave up while the gateway had already started the job; ten minutes later the expired claim was marked read and the batch never reached this folder (pages 113–125 in one scan, found by recover). Claims now wait up to 170 seconds, and a claim whose answer is lost is received at once from the gateway's own copy of that message (read with the owner credential, checked as a claimed job is). The receipt notes lease_lost until the gateway ends that claim.
  • inbox no longer stops with a lease error on a job that started at the gateway without a claim here (the same lost answer, seen on a later run): it saves that batch from the gateway's copy and goes on with the rest.
  • inbox claims for about 90 seconds and then reports how many jobs still wait in waiting_at_gateway; next says to run inbox again. A host that cuts long commands loses nothing already saved.

[0.3.5] - 2026-11-07

  • recover no longer times out on a long history. The gateway reads every kept body of a conversation (about 100 in a full one) from its store before it answers, which took longer than the client's 15-second request limit, so 0.3.4's recover failed on a real agent. A history request now has 170 seconds, and a run checks conversations for about --seconds (default 40; at least one) and keeps its place, so a host that cuts long commands loses nothing: run recover again until remaining_conversations is 0. Conversations that were open while a gap in this folder's page numbers opened are checked first, so the missing batches usually come back in the first run. --restart checks everything again; a finished pass starts over on the next run.
  • The report adds checked_now (conversation, time, turns, seconds), remaining_conversations, and per device conversations_checked, turns_at_gateway, batches_here and gaps_here (page numbers still missing in this folder). missing_pages appears once every conversation of the device is checked.

[0.3.4] - 2026-11-07

  • New recover command for pages the owner sent that the agent does not have. It compares what each paired device sent through the gateway (the owner's view of its conversations, read-only) with the data folder, and saves every batch the gateway still keeps (30 days by default) but the folder never got or lost, as an unread batch for inbox. This covers batches marked read by another run or host whose files did not survive, batches that ended at the gateway before they were read, and deleted files. Each batch is opened with the bot's key and verified as inbox verifies a job; nothing changes at the gateway. Per device it reports missing_here, waiting_at_gateway, files_missing_here, unreadable and missing_pages: page numbers no batch at the gateway or here carries, which the app never delivered. other_senders_waiting shows who sent the messages inbox ignores as unpaired, when, and with how many files. --check only reports; recover MESSAGE_ID ... saves named batches again.
  • inbox adds ignored_note when it ignores messages from unpaired senders, pointing to recover --check.

[0.3.3] - 2026-11-06

  • inbox --ack no longer acknowledges batches the agent has not seen. With pages sent continuously, the next batch often arrived between an agent's inbox and its inbox --ack; the --ack run saved it and acknowledged it in the same step, so it never showed in a later inbox and whole batches went missing from the middle of a scan. --ack now clears only the batches it names (--ack MESSAGE_ID ...), or with none named the batches an earlier inbox printed. Anything else is printed under batches as still unread, and next says to read it and acknowledge it. The output adds acknowledged_ids; batches in an --ack run now lists the batches still unread, not the ones just acknowledged.
  • inbox prints a long backlog a part at a time: at most --limit batches and about 20,000 characters, so an agent host that cuts long tool output (Claude Code at 30,000 characters) no longer drops batches from the middle of what --ack then clears. remaining says how many more wait, and batches are shown in arrival order.

[0.3.2] - 2026-11-02

  • A second agent of the same owner sets up with the owner's email. The official gateway lets one email register one owner, so a Muse set up after a Grok Bot stopped at email_in_use, and the only way round it copied the owner credential, which administers every bot, to the new machine. setup --email-code now adds the new bot to that owner with the same code (Yobiru 0.7.20 POST /v1/enroll/join), and the new data home keeps only a manager credential for its own bot. pair-accept, pair-decline, rename, wake, unpair and recovery use it unchanged; it cannot reach the owner's other bots. status shows credential: "manager" for such a setup. Existing setups, including those with setup.owner_home, behave as before.
  • The skill and README no longer suggest copying an owner directory to another machine (--owner-home is for an owner directory on the same machine).
  • Update the bundled engine to Yobiru 0.7.20, which knows manager credentials and the join route. The full suite runs against the pinned 0.7.20 gateway, plus a new join test with email verification on.

[0.3.1] - 2026-11-01

  • Add wake set --url-file F --secret-file F, wake status, wake test and wake clear for Yobiru 0.7.17 hosted wakes: the gateway calls a routine's webhook when a pairing request or pages arrive, so a hosted agent such as Grok Bot is woken without a machine running a connector. Registering sends a note telling the agent to run pair-requests and inbox (the gateway's own text names no client) and a test wake unless --no-test. wake status shows failures, the next allowed wake and a route the gateway turned off. The URL and key are read from private files and never printed; gateways without hosted wakes are reported plainly.
  • Update the bundled engine to Yobiru 0.7.17. It opens sealed fields of up to 16 MiB, so a full 8 MiB batch from Mokutori can be read once the app sends one.

[0.3.0] - 2026-11-01

  • Pairing now waits for the owner. A verified pairing request is recorded, not admitted: pair-check, pair-requests and inbox list it in pair_requests, the agent asks the owner, and pair-accept PAIRING_ID admits the device (grants chat,files and accepts the request). pair-decline PAIRING_ID rejects the request at the gateway, so Mokutori can tell the owner it was declined.
  • A pairing request counts when it reached the gateway before its code expired, however late the agent reads it. Before, an agent that checked after the 15 minutes silently dropped a request that had arrived in time, and Mokutori waited forever.
  • The skill tells agents to keep checking with a short scheduled task after pair, and to accept only after the owner agrees. Pairing records saved by 0.2.x keep their behaviour.

[0.2.5] - 2026-10-28

  • Add the optional agent brand a to both pairing links: grok-bot for --agent grok and muse for --agent muse; other agents omit it. Setup now records its agent, and setup --agent on an existing installation records a missing value locally without touching the gateway.

  • Require setup --name, the bot display name Mokutori shows for this agent (at most 80 characters). A missing name, or "Mokutori", is refused instead of silently naming every bot "Mokutori". The skill tells agents to use the name their owner knows them by, and to ask when unsure.

  • Add rename --name NEW to update the bot's gateway display name in place; status reports name and agent. Bots still named "Mokutori" should be renamed.

  • Add an optional Muse bridge audit log (MOKUTORI_MUSE_AUDIT_PATH). It is off by default and records only allowlisted MCP methods, tool names, batch IDs, success flags and returned image hashes; never keys, headers, sessions, document text or response bodies.

[0.2.4] - 2026-10-27

  • Return a clickable pair_link alongside the QR image, using https://mokutori.com/pair#... with all pairing parameters in the fragment and no HTTP query secrets.
  • Keep the existing pair_url and QR custom-scheme payload for installed scanning builds. Both entry points use the same one-time secret, expiry and pairing proof.
  • Require the skill to show both image and clickable link, and only report success after a nonempty pair-check.paired result. Document installation fallback and the absence of trusted online presence.
  • Preserve 0.2.3 owner-home migration, historical pairing records and encrypted inbox recovery.

[0.2.3] - 2026-10-26

  • Restore existing setup.owner_home installations for pairing, inbox recovery and unpairing. Reuse the original owner credential without copying it or creating symlinks; keep private-file validation.
  • Restore setup --owner-home for new receivers under an existing Yobiru owner, and report the selected owner directory in status.
  • Preserve legacy pairing records that lack modern proof/job fields. Report non-secret warnings instead of crashing, replaying permissions or marking an incomplete pairing successful.
  • Keep malformed or expired pairing entries from blocking new codes while retaining their history.
  • Add migration tests covering external owner reuse, credential-file restrictions, old pairing state, encrypted receiving and acknowledgement.

[0.2.2] - 2026-10-26

  • Fix the bundled QR generator importing on Python 3.8. QR image content and the pairing protocol are unchanged.

[0.2.1] - 2026-10-26

  • Included the optional self-hosted Muse MCP bridge for scan listing, image retrieval and local acknowledgement.
  • Added setup --agent muse while preserving the existing pairing and encrypted-claim recovery behavior.
  • Added connector tests and a synthetic end-to-end CLI-to-MCP scan test.
  • Added public source and security links and corrected installation guidance to retain the full plugin folder.

[0.2.0] - 2026-10-26

  • Renamed the public pairing URI to mokutori://pair and the proof prefix to mokutori-pair/1.
  • Added the bundled QR generator so pair --png works without a package install.
  • Added the public-bot engine sources, pairing checks, encrypted inbox handling and local end-to-end tests needed to install the plugin in Grok Bot, Codex or Claude Code.
  • Documented the Grok Bot upgrade path and the boundary between the plugin's local cache and the gateway's encrypted retention.

[0.1.0] - unreleased

  • Initial internal Mitori agent plugin prototype.