activeing123/mcptoon v0.8.36

v0.8.36 — the footer describes the reader, not the machine

on GitHub

[0.8.36] - 2026-11-11

Added

• DSH is a supported host. mcptoon sync --self now writes the skill pointer
  into $DSH_HOME/AGENTS.md. The contract was already there — the shell-capability
  table reads DSH_SHELL (verified set to 1 inside a real DSH session), and the
  pointer block was written with it in mind ("Codex (and DSH) mount no MCP, so the
  handshake instructions channel never reaches them") — but there was no row in
  sync._AGENT_SPECS, so a DSH user got no pointer, no sync --self and no
  takeover, and had to hand-write the catalog line into their own instruction file.
  The path follows DSH_HOME rather than a vendor constant, because DSH is
  relocatable and one machine can run several instances with different homes
  (measured: port 18085 → E:\dsh-next\home, port 18090 → its own); ~/.dsh is the
  fallback. mcptoon sync --dry now scans 13 hosts and names DSH.
• CLI discovery is counted. manifest, inspect, search and select now
  record a usage line under the reserved mcptoon server, the way the gateway's
  own tools already did. Nothing counted them before, so on a CLI machine — the leg
  whose design leans entirely on the model asking for the catalog instead of
  receiving it — mcptoon usage --json reported native_calls: 0 on a machine with
  522 routed calls, and there was no way to tell "nobody asked" from "nobody
  counted the asking". mcptoon activity now shows mcptoon:manifest beside the
  routed calls.

Changed

• The per-turn footer describes the reader, not the machine. On a machine whose
  gateway is wired into other hosts while their direct entries remain
  (serve-alongside), a reader that can run a shell was told 未接管·未省 on every
  turn — a statement about four other applications, printed at an agent that was
  already on the cheapest route there is (no MCP mounted, zero resident). Measured
  on the machine this was found on: 16 days installed, 527 routed CLI calls,
  resident_avoided still 0, and the headline never named the reader's own cost. A
  reader whose environment identifies it as shell-capable (DSH_SHELL,
  CLAUDECODE, CODEX_HOME, CURSOR_TRACE_ID) now gets its own figures —
  CLI 路线 · 0 常驻 · 省 655,518/轮 · 539 次 — while an unidentified or GUI reader
  keeps the machine-global verdict, which is the true one for it. Unknown is
  deliberately not treated as "shell-capable": guessing would print a CLI credit to
  a host whose real cost is a serve figure. The /轮 suffix was added after the
  bare 省 655,518 read as a running total; the cumulative figure is a different
  caliber (mcptoon report measures it as a lower bound on the serve caliber and
  says so).
• The not-saving nudge names the subject. It used to read 现在还没省:跑 mcptoon sync --takeover, which is read as a claim about whoever is reading it —
  how a CLI-leg session got told it was saving nothing. It now says 另有宿主仍直连 32 条,每次照旧直接加载, so the entries are attributed to the hosts that hold
  them.

Fixed

• Nothing said what to do when nothing is being saved. The tools line reported
  未省, which is a status; a status does not tell a reader what to type. When the
  mode is direct or serve-alongside, the footer now carries one pasteable
  command (mcptoon sync --takeover or mcptoon quickstart), throttled to at most
  once per 24 hours and five times in total so it reads as a nudge rather than
  nagging. --json stays read-only and does not spend the allowance, so a script
  polling the figures cannot burn a human's.
• The test suite no longer writes into a developer's real DSH home. Adding the
  DSH row made sync write pointers through $DSH_HOME, and tests/conftest.py
  isolated HOME, %APPDATA% and XDG_CONFIG_HOME but not that variable — so a
  suite run appended the skill-pointer block to the live
  $DSH_HOME/AGENTS.md (measured 1628 → 2368 bytes), i.e. the agent instructions of
  whoever ran pytest. conftest.py now points DSH_HOME at a temp path and clears
  the agent marker variables (DSH_SHELL, CLAUDECODE, CODEX_HOME, ...) for the
  session, so which agent launched pytest can no longer change what the suite
  asserts. Verified: the file's SHA-1 is unchanged across a full run.