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>
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>
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>
Each locks <home>/run/<role>.lock in main (gatewayd after its secret,
so nothing is made before the secret is read). brokerd's old
second-instance test in serve.rs expected the audit lock's message,
so serve.rs has a new version; task 01 now copies serve.rs-task01.
Checks: given tests fail at ref 12 as the task says (2 each in
brokerd, memoryd, gatewayd; serve.rs passes); the task's quoted edits
applied to ref 12 reproduce ref 13's tree exactly; replay gate ok with
1118 tests (1113 - 1 + 6), /tmp unchanged; 14 mutations, all caught;
the six instance test files stressed 25 runs under 16 busy loops and
20 under 32, no failure.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
bxcore::instance::lock (File::try_lock, file 0600, missing directory
0700), LockError::line for the two lines the daemons print, and the
already-running runbook fragment; loopd locks <home>/run/loopd.lock
before its socket and self-test, inferproxy <socket>.lock.
Checks: given tests fail at ref 11 as the task says (bxcore E0432;
loopd 3 and inferproxy 3 fail); the task's quoted edits applied to
ref 11 reproduce ref 12 byte for byte; replay gate ok with 1113 tests
(1099 + 14), /tmp unchanged; 25 mutations, 24 caught (the uncaught
one maps TryLockError::Error to Held, not reachable from a test);
given files rustfmt-clean, under 500 lines; no chat-template tags.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The task file, the skeleton crates/bxctl/src/sanitize.rs (five todo!()
bodies: parse_sanitize, code_point, token, prefix, tag_count) and two
given tests: sanitize_check.rs (8, runs the binary) and
sanitize_report.rs (10, the library).
Decisions: the one argument after `check` is the text, whatever it
looks like; none means standard input. Output: refuse line (first rule
only; a character as U+XXXX, a tag as its token id, a secret as its
prefix and an ellipsis), flag line (rule only), neutralise line (every
occurrence of every tag). Exit 0 nothing fires, 1 something fires, 2
usage error or unreadable text. Nothing of the text is printed.
Checks: every given test fails before the task as the task says (the
report file does not compile: one E0432, two E0599; the binary tests
exit 2); the skeleton state gives 10 and 8 not-yet-implemented
failures; the reference passes make gate with 1099 tests (1081 + 18);
the task's edits applied to a fresh ref-10 copy with its own build
directory give a tree byte-identical to the reference; 21 mutations,
each caught; 40 runs under load, no failure; /private/tmp unchanged
around a gate run; rustfmt and the line limit on the given files; no
tag or key-shaped string in the plan or given files.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Task 10 (spec 9, the owner's option C): a new module gatewayd::links with
disarm(text, untrusted) -> Disarmed { text, links }, given as a skeleton
with 16 small todo!() helpers, and deliver posting each answer through
answer_posts: disarm, split_answer, disarm each part again, NOTICE once
at the end of the last post when a link was put in a code span.
Checks, against a reference kept off master:
- Replayed from ref 09 in a scratch worktree with its own target/: the
task's edits (applied from the task text) plus the given files plus the
reference bodies are byte-identical to the reference. make gate on
rustc/clippy 1.98.1: gate: ok, 1081 tests (1058 + 7 + 7 + 9). /tmp
holds nothing new after the gate.
- Step 2 (E0432 only) and step 3 (6/1, 7, 9 failing, all "not yet
implemented"; warnings as the task lists) seen in the replay.
- 71 one-line breaks of the reference, each caught by the given tests.
- Differential fuzz of generated Markdown through Mattermost's marked
fork (mattermost/marked c211e4c, the webapp's options) and CommonMark:
on the final code, 600k replies in each mode, half with inline LaTeX
on, render no image and (untrusted) no link.
- disarm(disarm(x)) == disarm(x) on 200k generated replies, both modes.
- Given files pass rustfmt; skeleton comments read beside their
signatures; no chat-template tag in the plan or given files;
deliver_links 40 runs under CPU load, no failure.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
bxcore's TurnDone gains untrusted: bool, always written and required
when read (no serde default: a missing field would read as a clean
session, the wrong way to fail). loopd's new Session::untrusted is true
once any ToolResult in the log is labelled untrusted, and channel::handle
puts it in the TurnDone; the log is the one source, so it holds across
resumes and restarts. gatewayd, bxctl and brokerd sources are unchanged;
their tests that build a TurnDone get new versions. Task 01's broker.rs
is kept as broker.rs-task01, and task 01 now copies it.
Checks: the given tests fail at ref 08 as the task says (strict's
turn_done.json, and 13 test targets with E0560/E0609 only); the task's
edits applied from its text to a fresh ref 08 tree are byte-identical to
the reference; the gate passes with 1058 tests; 12 mutations caught;
/private/tmp and TMPDIR unchanged around the runs; no tag in plan or
given files.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
loopd's self-test gains a check, run second (after "server matches
config", before any prompt): each tag in bxcore::sanitize::TAGS must be
one token from /tokenize, its recorded id, and its neutralised form
more than one. A failure names the tag by its id, never its text, and
points at a new runbook entry, loopd-tags-unexpected, given as a
fragment; SelfTestError::pointer picks the entry per failure. One
request per text (66): /tokenize has no batch form. The fake server
gains respond and a fake tokenizer; selftest.rs, serve.rs, pointers.rs
and support/mod.rs get M5b copies.
Checks, on a reference at ref 08 (titan, rustc 1.98.1, own target/):
- /tokenize fields from llama.cpp tools/server/README.md at b10964
(b29c606) and its handler; on straylight all 33 tags gave their ids
and every neutralised form 7 to 11 tokens, 66 requests in ~0.3 s
- given tests before the change: 4 fail as the task lists, selftest_tags
does not compile (7 E0599); pointers passes
- make gate: gate: ok, 1050 tests (1042 + 8); /tmp and $TMPDIR
unchanged around it; the four suites 25 runs under load, no failure
- 20 mutations, each caught: id not compared; neutralised check
dropped; > 1 as >= 1, as == 1 missing 0, as > 2; first or last tag
skipped; tag asked twice; pointer dropped or wrong (all or Tokenize);
check skipped in serve; check run last; parse_special false;
add_special left out; id missing from the message; the tag's text in
the message; tokenize error as Infer; wrong count; wrong other id
- the task's edits applied to a fresh ref-07 copy give files
byte-identical to the reference; every before block found once
- no chat-template tag whole in the task or given files; task lines
within 100 characters except tables and verbatim lines; given files
under 500 lines
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
loopd neutralises the user's message (the User record), every text
appended as a tool result (results, recall's, denials, failures and
loopd's own answers share the one append in run_turn), after the cap,
and the memory block in Baseline::assemble_with. The model's output,
system.md and replay are untouched: an old log with a raw tag is
replayed byte for byte.
Checks, on a reference at ref 07 (titan, rustc 1.98.1, own target/):
- given tests before the change: 7 of 9 fail as the task lists; the
two that guard what must not change pass before and after
- make gate: gate: ok, 1042 tests (1033 + 9); /tmp unchanged around it;
the two new suites 20 runs, no failure
- 14 mutations, each caught: user not neutralised; neutralised before
the cap; recall path missed; only untrusted results; truncated from
the final length; block not neutralised; model content, reasoning or
tool arguments neutralised; replay re-neutralises user, tool or
system; open rewrites records; system.md neutralised too
- the task's edits applied to a fresh ref-06 copy give files
byte-identical to the reference; every before block found once
- no chat-template tag whole in the task or given files; task lines
within 100 characters except tables; given files under 500 lines
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
memoryd's writes go through bxcore::sanitize::refuse, then flag; check.rs
and its test go. A refused write answers with the rule's sentence and is
logged with the rule's name; a flagged identity write becomes untrusted
notes with the rule in the answer, the journal (new optional field `flag`)
and the log, and is never staged for review; the block leaves out an
entry that breaks a refuse rule and logs the rule. The three runbook
entries the new pointers need are given here as fragments, because the
gate's runbook check fails on a pointer without its entry.
Checks, on titan (rustc 1.98.1, clippy 0.1.98):
- given tests fail before as the task says: 7 failing in edit, serve and
serve_sanitize; files and sanitize do not compile (E0609; E0432, E0609,
E0615)
- the task's edits, parsed from the task file and applied to a fresh
ref-05 worktree with its own target/, give a tree byte-identical to the
reference; its gate: ok, 1033 tests; /private/tmp unchanged
- 26 one-line mutations of the reference, each caught by the given tests
- new tests stressed under load: 30 runs, 0 failures
- no chat-template tag and no whole fake key in the task or given files
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Records what porting Hermes' patterns decided: seven rules kept,
"dump the conversation" narrowed, agent-configuration changes left out
(Boxmaker's agent loads no such files), about fifty look-alikes,
apostrophes folded, and a flag in bxctl names only its rule. A false
positive also marks a session that recalls the note untrusted.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The design model's review of task 05: a recalled flagged note taints the
reading session, so a false positive on a code-work note costs a session.
- dump-conversation: the object must end its clause (a new part, Ends,
and a function to write, ends_clause): punctuation after it, the last
word, or one of so far, above, to, into, in, with, verbatim next. "include
the conversation id in each log line" and CLAUDE.md's "spike output that
contains conversation content" now pass (code-work corpus).
- persist-agent-config dropped: Boxmaker's agent never loads AGENTS.md,
CLAUDE.md, .cursorrules or .clinerules as its instructions. Listed under
not ported; its notes are code-work cases, its attack a decided edge.
- normalise folds the typographic apostrophes U+2018 and U+2019 to '.
- The corpus is split: other scripts and decided edges move to
sanitize_flag_edges.rs, keeping each file under 500 lines.
Checks: make gate on the reference (ref 05, 818dbc7), rustc 1.98.1: gate
ok, 1023 tests (22 + 2 + 2 new); the task applied to a fresh ref-04 copy
is byte-identical to it; 60 mutations, each caught; /tmp unchanged around
the gate; no chat-template tag in the task or given files; given files
pass rustfmt and the 500-line limit.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Task 05 adds normalise (case, full width, white space and zero-width,
leetspeak, fifty Cyrillic and Greek look-alikes from UTS #39
confusables 18.0.0) and flag: nine rules, eight ported from Hermes
Agent's tools/threat_patterns.py (NousResearch/hermes-agent at d0288be,
MIT) as word sequences with bounded gaps, and M5a's six phrases. Given:
two skeletons (eleven small todo!()s, tables whole) and two test files
(21 unit tests, 4 corpus tests including Russian and Greek notes and a
table of decided edges).
Checks: the given tests fail before for unresolved imports only (E0432);
the reference (ref 05) passes make gate on rustc 1.98.1, 1022 tests;
the task applied to a fresh ref-04 copy is byte-identical to it; 45
one-line mutations, each caught; /tmp unchanged around the gate; no
chat-template tag in the task or given files; given files pass rustfmt
and the 500-line limit; every todo!() comment read beside its signature.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The private-key rule also refuses a line that starts `-----BEGIN` and
ends `PRIVATE KEY BLOCK-----`, OpenPGP's armor header (RFC 9580
section 6.2.1; RFC 4880 section 6.2). The two endings are a given
table, PRIVATE_KEY_ENDS; the skeleton's private_key comment names it.
A PGP private key block joins the attack corpus, a PGP public key
block and a signature the code-work corpus, and the edge test gains an
indented PGP header (refused), a PGP public header and a PGP header
without its closing dashes (both pass).
Checks: the reference passes the gate on titan, 997 tests (unchanged:
the cases are inside existing tests), /tmp unchanged around the run;
the task's edits applied to a fresh ref-03 copy are byte-identical to
the reference; 8 private-key mutations, 4 of them new for PGP, all
caught; tag and private-key-header greps over the task and given files
are empty; given files pass rustfmt and the 500-line limit.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The private-key rule named only PEM's ending, so an armored PGP
private key passed. Public-key blocks still pass.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The refuse half of spec 6.1 in bxcore: M5a's control and invisible
rules and tables, the tag rule narrowed to task 03's exact matching,
and new rules for private use, noncharacters and known-prefix secrets
(GitHub, GitLab, sk-, Slack, AWS, private-key lines). A skeleton with
eight small todo!()s; the rule table, sentences and secret tables are
given filled in. memoryd is untouched until task 06.
Checks: given tests fail before (E0432, then 16 not yet implemented);
the reference passes the gate on titan, 997 tests (981 + 16), /tmp
unchanged around the run; the task's edits applied to a fresh ref-03
copy are byte-identical to the reference; 36 mutations, all caught
(one needed a new case: lower case after AKIA); tag grep over the task
and given files is empty; given files pass rustfmt and the 500-line
limit; every todo!() comment read beside its signature.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Task 03 is now the tag list and sanitize::neutralise; refuse moves to 04 and flag
to 05. The list is the 33 control and user-defined entries of the model file
straylight's router loads (Ornith-1.5-35B-Q4_K_M.gguf, snapshot 12393612fd4f,
llama-server b10964-b29c606), each checked with /tokenize on 2026-09-26 to be one
token with its id. Review focus 1 is decided by the same measurement: a tag
inside a word is still the token and is neutralised; upper or mixed case, a
zero-width, space or line-feed split, and full-width brackets are not, and are
left as is. Each shape is pinned by a given test.
Checks: the given tests fail before (E0432, no bxcore::sanitize); the reference
passes make gate on titan (rustc and clippy 1.98.1) with 981 tests (963 + 18);
16 mutations of tags.rs, all caught; no failure message prints a tag; /tmp
unchanged around the gate; the task's edits applied to a fresh ref 02 tree give
a tree byte-identical to the reference; no plan or given file holds a tag whole;
given files pass rustfmt and are under 500 lines.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The task moves gatewayd's secret store to bxcore::secrets behind the
Cargo feature `secrets` (spec, section 5). The moved module is given
whole; SecretSpec and SecretSource move with it, so gatewayd's config
keeps its [secrets.<name>] table with the per-secret type from bxcore.
load and read_checked take the role's name first, so the three
sentences that named gatewayd literally now name the role; gatewayd
passes "gatewayd" and its bytes are unchanged (its tests/main.rs pins
them and is not touched). A given loopd test parses the three
manifests (loopd's, the workspace entry, bxcore's features) and checks
the cfg line in bxcore's lib.rs; the task also checks `cargo tree -p
loopd -e features` and builds loopd on its own.
Checked: the reference, replayed from the task text applied
mechanically to the start tree (master plus task 01), is byte for byte
the reference; every before block found exactly once there. Step 2's
failures recorded from that tree: ten test targets with E0432 on
`bxcore::secrets`, two of four no_secrets tests. make gate on titan's
toolchain: gate: ok, 963 tests (956 + 3 + 4), clippy clean, /tmp
unchanged across the run. 23 one-line mutations of the reference, all
caught; removing the zeroize wrapper is not observable and not counted. Given
files pass fmt and the line limit; no chat-template tags.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Task file and 13 given files (2 new: the wire fixture with a detail and
brokerd/tests/invalid_detail.rs; 11 new versions of existing tests, most
only adding `detail: None`).
Checked against a reference in a scratch worktree with its own target/,
on titan (rustc 1.98.1, clippy 0.1.98): make gate prints gate: ok with
956 tests (944 + 12 new); /tmp has no new entries after the gate. All 21
quoted before-blocks occur once in master, and applying the edits to
master gives the reference byte for byte. The given files are the
reference's, formatted, none over 500 lines; no chat-template tag in
the plan.
Before the edits, strict's envelopes_reject_unknown_keys_at_every_depth
fails and ten test targets do not compile, as step 2 lists.
Mutations: 21 one-line breaks of the reference; 20 caught by the given
tests. The one not caught (dropping `default` from the detail attribute)
cannot change behaviour: serde reads a missing Option field as None.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Code constraints: `proto` became `bxcore` in M5b task 00 and holds the
code roles share as well as the shared types. Recorded in
docs/decisions.md with the M5b spec.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Task 00 (proto -> bxcore) by the design model before the run; fourteen
Ornith tasks: the invalid-arguments detail, the secrets move, the
sanitizer in three tasks, memoryd, loopd's neutralising and self-test,
TurnDone.untrusted, gatewayd's links, bxctl sanitize check, the instance
lock in two tasks, and the runbook. Global constraints, five review-focus
inputs each owned by a task, and the interfaces between tasks. The task
files and the reference come next, through docs/plans/CHECKLIST.md.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every daemon removes a leftover socket before binding, which also lets a
second copy silently take a running one's sockets. Owner's decision:
refuse to start instead. Each daemon takes an exclusive std File lock
(stable since 1.89, checked on 1.95) on <home>/run/<role>.lock, or
<socket>.lock for inferproxy, before touching its sockets; the system
releases it however the process ends. A runbook entry, a security note
and a test per daemon.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Code spans stop both the link preview and the server's image fetch;
Markdown links are not previewed; images are fetched by the server at
posting time. Emoji refusal accepted, phrases move to flag, Hermes'
patterns ported at plan time.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
bxcore (compiler-driven rename, with the spike's four layers named); the
invalid-arguments reason on the wire; the secret store moved into
bxcore behind a feature loopd never enables; bxcore::sanitize with
refuse, neutralise and flag, its lists, hand-written normalisation, two
corpora, a rule-change procedure and `bxctl sanitize check`; memoryd
refusing and flagging (a flag keeps the text as an untrusted note);
loopd neutralising tags on append, with a tokenizer self-test and the
session's trust on TurnDone; gatewayd escaping images and, in untrusted
sessions, links; a table of every link between the roles; runbook
entries; attack tests per trust boundary and the device checks. Four
points left for the owner in section 14.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Owner's decisions, 2026-09-26. The second row lists what was adopted,
which driver guards are trials for M5b's run, and what was left out.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
kyleandClaude Opus 5.5
