The ledger cannot promise an item is never sent twice: a crash between a send and its sent line,
or an owner removing that line, sends it again. A decided line holds a skip and its reason, as
task 07's ledger does. Turns from before sync began are skipped (owner's decision, 2026-09-28).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
memoryd.toml gains an optional [honcho] section, [[honcho.read]] and the
Honcho keys in [secrets.<name>], checked at start into
memoryd::honcho::config::HonchoConfig (13 reasons, runbook entry
honcho-config-invalid given as a fragment), and [sockets] state for
brokerd's state.sock. Keys are checked, never read; no [honcho] leaves
memoryd as it was. memoryd enables bxcore's feature secrets.
Checks: a reference at ref 05 + this task passes make gate with 1188
tests (1171 + 17); /tmp unchanged around the gate; the given tests fail
before as the task says (two targets do not compile, 4 start tests
fail); 49 one-line mutations of the reference, all caught (one gap, a
URL with a user and no password, found and closed); the new tests 40
times under CPU load, no failure; a replay from the task text on a fresh
ref-05 worktree with its own target is byte-identical and passes the
gate, and every before block was found once; given files pass fmt and
the 500-line limit; no chat-template tag in any file.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
brokerd serves a third socket, [sockets] state (default
<home>/run/memory-broker/state.sock), bound by the same listen (dir 0700,
stale socket removed, socket 0600), one thread per connection, one
question per connection. The answer is StateStore::read: the stored taint
and flag, the store's default (private, trusted) for a session with no
file. A damaged state is an internal error with the store's text, never an
answer; other kinds are forbidden; malformed frames get read_request's
error. A new tests/serve.rs waits for an answer before reading the
startup notices (a race that failed about one run in ten under load,
before this task too).
Checks: every before block found once at ref 04 (20 edits); the new file
in the task is byte-identical to the reference; the given tests fail
before as the task says (config and state_socket do not compile;
serve_state 5 of 5 fail); gate ok with 1171 tests on the reference and on
a replay from the task text in a fresh ref-04 worktree with its own build
dir (byte-identical to the reference; removed); /tmp unchanged around the
gate; 25 mutations, 24 caught, the one missed (state bound last) changes
no tested behaviour; the socket tests 25 runs under load, 0 failures;
no task-text line over 100 characters outside tables and code.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The speaker goes on every turn of the chat: interactive, --say and --say's retry. Without the
flag it is None. A name that is empty or holds whitespace or a control character (Unicode), or a
second --speaker, is a usage error: the usage text on stderr and exit status 2. The usage text
gains the flag and one line explaining it. Given: chat_speaker.rs (new) and new versions of
chat.rs, chat_approvals.rs and cli.rs. Task 01 now copies chat.rs-task01, the older chat.rs.
Checks: every before block found once at ref 03; replayed from the task text on a fresh ref-03
worktree with its own build directory, byte-identical to the reference, gate ok with 1158 tests
(1148 + 10); /tmp unchanged around the reference's and the replay's gate runs; 20 of 20
mutations caught; the new tests passed 40 of 40 runs under CPU load; given files pass fmt and
the 500-line check; no chat-template tag in the plan or the given files.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A batch carries the Mattermost user id of each message; the turn's speaker
is the username of the one user who wrote all of it (Client::username,
kept in a map on the event loop), and None for two or more people or a
failed lookup. The turn's text is unchanged.
Checks: reference replayed from the task text on a fresh ref-02 worktree
with its own build dir, byte-identical; make gate ok with 1148 tests
(1138 + 10); /tmp unchanged around both gate runs; 21 mutations, all
caught; new tests stressed 40 times under load, no failure; every before
block found once at ref 02; no chat-template tags.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The owner's hosts (titan, talos) run 1.98.1; nixos-26.05's Rust is
1.95.0, so builds and lints differed by host. The flake gains a second
input, oxalica's rust-overlay (following our nixpkgs, so host configs
can still follow theirs), and both the package build and the dev shell
use its stable 1.98.1 with clippy and rustfmt. The workspace's
rust-version rises to 1.98 to match; AGENTS.md, docs/dependencies.md
and the M5c-1 plan follow.
Checked on straylight: nix build gives all seven binaries; the dev
shell has rustc, cargo and clippy 1.98.1, rustfmt 1.9.0, cargo-deny
0.19.6; make gate in it passes with 1122 tests, as on titan.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
loopd copies the Turn's speaker onto the User record it logs, as it
came (not checked, not neutralised), and never into a request; after a
finished turn it sends SessionUpdated with the turn's number, the count
of User records in the session's log (Session::turns), so the first
turn is 1, a failed turn that was logged counts, and the count goes on
after a reopen or a restart. Failed turns are still not told.
Given: speaker_turn.rs (8 tests) and new versions of loopd's memory.rs,
turn_broker.rs, end_to_end.rs and support/turn.rs (run_turn gains
speaker, MemorySource::updated gains turn). Task 01 now copies
memory.rs-task01, its own version.
Checks: the reference passes make gate with 1138 tests (1130 + 8); the
task replayed from its text on a fresh ref-01 worktree with its own
target is byte-identical to the reference and passes the gate; /tmp
unchanged around both gate runs; 15 one-line mutations (speaker not
logged or dropped by the channel or neutralised, speaker in the request
at log time or at replay, turn 0-based or +1, counting Assistant
records, all records or only speakers' turns, a global or per-process
counter, told before logging or on failure, MemoryPort sending 0) all
caught; speaker_turn run 60 times under load, no failure; no
chat-template tag in the plan or given files; task lines within 100
characters but tables.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Task 01 changes bxcore's wire and log types for M5c-1: Turn.speaker and
LogRecord::User.speaker (optional, skipped when None, so old frames and logs
keep their bytes), SessionUpdated.turn (required, refused when missing), and
the new SessionState / SessionStateReply messages. The other crates only
construct the new fields (speaker: None, turn: 0) and brokerd's kind() names
the two new kinds; later tasks give them values. 20 given files.
Checks, against a reference at 7b40686 with its own build directory:
- every "before" block found exactly once at 7b40686 (and at master, which
changed no crate since);
- step 2's failures as the task names them: two strict tests, fifteen test
targets with E0559/E0560/E0432/E0599 only;
- make gate: gate: ok, 1130 tests (1122 + 8); /tmp unchanged around it;
- replay from the task text in a fresh worktree of 7b40686 with its own
target/: crates/ byte-identical to the reference, gate ok, 1130; removed;
- 23 mutations: 21 caught; the two that drop `default` from an Option
field cannot change behaviour (serde reads a missing Option as None);
- no chat-template tag whole in the task or the given files; lines over
100 characters only in tables, fixtures and verbatim strings.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
`.envrc` runs `use flake` when Nix is present, so on the NixOS hosts
(programs.direnv.enable brings direnv and nix-direnv) cd'ing into a
clone gives the pinned Rust toolchain, clippy, rustfmt, cargo-deny and
make; on titan, which has no Nix, it does nothing. nix-direnv's cache,
.direnv/, is ignored. Checked on straylight in a throwaway copy:
cargo 1.95.0, cargo-deny 0.19.6, rustc from the Nix store.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
29 request/response pairs, each request exactly as roncho 1.0.0's
blocking client sends it, recorded in a throwaway workspace
(boxmaker-plan-probe) created for it and deleted afterwards; no other
workspace was touched. For the fake Honcho server of task 10. Found:
creation is get-or-create (201, then 200); adding messages to a missing
session creates it; a 422 echoes the whole input; a deleted conclusion
answers 204 with an empty body, and 404 the second time.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The owner's Honcho client, blocking, with TLS: default features off,
feature tls. It adds only itself to the lock file (checksum a5ccccc2…,
as checked against crates.io in review). cargo deny passes. The
dependency and the egress to rift are documented; the egress table also
lists the design model's probe of rift in a throwaway workspace.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Memory items 4 and 5: two workspaces per agent and one fleet workspace
for untrusted notes; the private ceiling, secret sessions never synced;
trusted memory writes mirrored as conclusions; the agent's peer not
observed; no rebuild, corrections by contra-fact or deleting our own
conclusion; the fleet workspace always untrusted. Recorded in
docs/decisions.md.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A turn without complete speaker information is not sent; every peer in
the fleet workspace is a member, so fleet notes are messages that one
search finds, with a retraction message for an undone note (Honcho
cannot delete a message); the fleet workspace derives nothing
(reasoning, summaries and dreaming off); API keys per workspace are
first-class, with a runbook entry for a refused key.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Designed with the owner on 2026-09-27: per agent a trusted and an
untrusted workspace plus one fleet workspace for untrusted notes;
private ceiling, secret sessions never synced; no rebuild (contra-facts,
or deleting our own conclusion); the owner and other people observed,
the agent not; every trusted memory write mirrored as a conclusion; a
decision ledger that doubles as the outbox; recall over the local index
and the fleet; ask_memory by workspace chat. Grounded in roncho 1.0.0
and Honcho 3.2.0 (source read at v3.2.0) and rift's live metrics.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
decisions: M5b settles the invalid-arguments reason, tag
neutralisation and the shared secret store (their open rows go), and
the verdict on the run's two driver trials. CLAUDE.md's current state:
M5b done, M5c (Honcho, roncho 1.0.0) next; straylight's clone is
boxmaker-m5b.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Found in the M5b device check on straylight: the self-test asked about
each of the 33 tags and each neutralised form, 66 requests, one
connection each, and inferproxy's token bucket (burst 10) refused the
ninth, so loopd would not start. Now it asks about all the tags at once,
one per line, and then all their neutralised forms: the answer to the
first must hold each tag's recorded token in order, the second none of
them. Each tag is a special token, split out before anything else, so
the lines between cannot change it; and a neutralised form is only
dangerous as a tag's own token, so that is what the second answer must
not hold. The errors name a tag by its id, as before.
The fake tokenizer reads text with several tags in it; the tests cover
each tag as ordinary tokens, as another listed or unlisted token, a
duplicated token in the middle and at the end, a form holding each tag,
and both requests failing. Spec section 8 and the runbook entry follow.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The review of tasks 01 to 14 in the implementer log (task 14's row
renamed to its file's name), and four tips for writing tasks: what to
do with a skeleton's comments, golden files for scanners, naming the
log row, and running the implementer without plugins.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
M5b review, findings 3 and 4: the instruction comments tasks 03 and 04
left inside filled bodies, and a space of indent task 11 added to four
lines of bxctl's usage text. No behaviour changes.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
M5b review, findings 1 and 2. url_len ended a clean reply's URL at any
`<` outside parentheses, not at a `<` before white space, a bracket,
`<`, `>` or the end; closes used a count of characters as a byte
offset, so a closer indented with a no-break space did not close its
fence. Both bodies, and fence_end's (an index replaced by `get`), are
the reference's again. tests/links_review.rs pins both cases with the
reference's output.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Text from outside the model (the user's message, every tool result, and the
memory block) is neutralised the moment it is first logged, so every later
request replays the same bytes and a tag in that text cannot reach the model as
one of its own control tokens. The model's own output, replay, and the system
prompt are left byte-for-byte.
Co-Authored-By: ornith-1.5-35b-a3b <agent@wntrmute.dev>
Three changes from the M5b run (implementer lessons T47, T48):
- each session runs `opencode run --pure` with
OPENCODE_DISABLE_CLAUDE_CODE=1, so no global plugin or borrowed rules
file reaches the implementer; PLUGINS=1 restores both;
- with RESUME=1, a session stopped at SESSION_LIMIT_MIN with work left
uncommitted is resumed once, as one that ended by itself already was;
an idle stop is still not resumed;
- the prompt names the log row's Task cell, so a task's heading cannot
mislead the row the driver then looks for.
Checked against a stub opencode in scratch repositories (17 checks:
the flags and variable, PLUGINS=1, a resume after the limit, no resume
without work or without RESUME).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Open item from M5b task 09: loopd reads a session's trust from its
epoch-0 log, so compaction must carry it into the new epoch.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The task table follows the files (03 tags, 04 refuse, 05 flag) and
gives each task's gate count and mutations; the interfaces are the ones
the tasks built (task 02's role argument and SecretSpec, Refusal.at,
Flag, TurnDone.untrusted required, disarm, LockError::line); the review
focus names the tasks that pin each item; which task adds each runbook
entry, and the -taskNN convention for replaced given copies.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The last runbook entry, the rule-change procedure of spec 6.5; the
entries memory-write-refused and memory-write-flagged now send the
reader to it instead of to the spec. Tasks 06, 08 and 12 added the
other five, because the gate needs an entry for every pointer.
Checks: the edits, applied from the task's text to ref 13, give a
runbook byte-identical to the reference; each before block found once;
the gate passes on the reference with 1118 tests; /private/tmp unchanged.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
kyleandClaude Opus 5.5
