M5a task 04 stopped twice on the chat-template tags the check refuses.
The first time, the model read them in a file and took them for its own
template. The second time, it spelled them out in its own reasoning.
Measured on straylight: the think and tool-call tags are one token
whatever the tokenizer's flag. Inside any chat message, the turn
markers become control tokens, so a tool result can forge a system
turn.
Adds the rule to AGENTS.md, and proposes neutralizing the tags in
loopd as an open decision.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The session did the task and stopped on make verify-device. The cause
was not the server: run-plan.sh and the device tests both read
BOXMAKER_MODEL, and the run's value (OpenCode's provider-prefixed id)
reached llama-server as the model name. With it unset, verify-device
passes 6 of 6. The copied files are identical to the plan's, and the
gate prints gate: ok, 936 tests.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Task 04's second session stopped too. It checked its code against the
markup test cases in its reasoning, and the tag it spelled out ended
its reasoning. Building tags from pieces keeps them out of files, not
out of the model's thoughts.
check.rs is now given whole, with the tags from pieces, and task 04
fills only entries.rs. The three later tests that used a tag only as
something the check refuses now use an instruction phrase or an
invisible character. They pass against the reference.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Task 04's session read the given test and the check.rs skeleton, both
holding the tags the check refuses, and Ornith took them for its own
template. It lost track of the file and made no change. The tags are
now built with concat! and format! in check.rs and in the four given
tests that held them, and the spec describes them in words. The
rewritten tests pass against the reference, are rustfmt-clean, and
still catch the case-folding mutation.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Follows the owner's other projects (sift, tether): nixpkgs pinned to
nixos-26.05, x86_64-linux, the version from the commit. `packages.default`
builds all six binaries with rustPlatform (tests stay in `make gate`),
`packages.tools-image` is deploy/tools-image.nix, and `devShells.default`
has Rust 1.95.0, clippy, rustfmt, cargo-deny and make. No NixOS module
yet: that is M8 (P22).
Checked on straylight: `nix flake check` passes; `nix build` gives
brokerd, bxctl, gatewayd, inferproxy, loopd and toolkit, each printing
its usage; `nix build .#tools-image` and `nix develop` work.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Owner's direction, 2026-09-25: one agent per host, each with its own
identity, memory and credentials, sharing fleet memory through Honcho;
inference from a configured llama-server-compatible endpoint; a last
milestone that deploys to NixOS hosts and other Linux hosts. Not applied
to the brief yet.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
memoryd (files, the check, edits, journal, index, recall, serving),
the memory tools and built-in grants in brokerd, the memory block in
loopd, bxctl reindex and memory, and the device checks with memoryd.
Each task was built as a reference and gated. Each was replayed on its
own at its end state (854 to 936 tests) and mutation-checked. Each
skeleton compiles and fails its tests. Every todo comment was read
beside its signature.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The memory socket in brokerd.toml is optional, and bxctl's flag is
--memory-socket. A note's header carries its class. Built-in grants are
kept apart from the owner's, and builtin- ids are reserved. The session
state is carried through Decision and RunSpec, labels through
run_labelled. loopd's assemble_with and MemorySource. Baseline
measured at 881 and 2,432 tokens.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
nobody_else_and_nothing_else_answers_it read five replies after
waiting only for the first, so it failed whenever the other four were
not posted yet. It failed once in a full `make gate` run, when many
test binaries ran at once. The helper now waits for as many posts as
the test reads.
The M4b plan's copy is left as it was handed over.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The research feeds M5b's sanitizer. One finding is a bug in the M5a
draft: category Cf misses variation selectors, the combining grapheme
joiner, and the Hangul and Braille fillers.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Writes from sessions that have read untrusted content go to the
archive, labelled untrusted, and never reach the prompt. Every write is
checked, announced in its thread, journalled and can be undone.
Session-scoped approvals for other tools become an open decision.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The owner's review of 2026-09-24. P17: the memory tool gains
consolidate, which moves entries out of memory.md into
memory/archive.md, and from M5b into Honcho, instead of losing them.
P18: memoryd owns memory/ and the SQLite index, so every memory tool
runs in one place.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The owner's direction of 2026-09-24:
- Memory is three local files, as Hermes has them: a soul written by
the owner, and memory and user files the agent edits.
- Honcho is the fleet's memory. Each Boxmaker writes only to its own
workspace, and can read other workspaces at a trust level the owner
sets.
- A new role, memoryd, does the syncing, because brokerd never holds
conversation history and loopd has no network.
- M5 splits into M5a (files and local search), M5b (memoryd and
Honcho) and M5c (the scheduler).
The brief is not changed until the owner approves P17 to P20.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The two posts caught up at restart are two top-level posts, so two
sessions, and their turns run on their own threads. gatewayd starts
them in order, but which one reaches loopd first is up to the
scheduler. Under load the test failed about 1 run in 60; task 18's
session hit it once. mm_json.rs already checks the order of catch-up
(reversing the sort fails there), so this test now compares the two
turns as a set. It passed 100 of 100 runs under load.
The M4a plan's copy is left as it was handed over.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
kyleandClaude Opus 5.5