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 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>
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>
One list for the design model to work through per task, gathering the
task-writing tips that apply before hand-over (T18, T19, T24 to T29,
T32 to T35, T39, T40) and three new ones from the combined Ornith
guidance: walk each given test by hand against the task's rules (T41),
grep earlier given files for anything the task changes (T42), and run
the gate's own checks on the given files (T43).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
21: the ledger raises a session's state by the stricter of the grant's
label and the result's own, so a clean session that recalls a
memory-tainted note becomes untrusted and cannot write it into
memory.md (found on straylight, 2026-09-25). 22: the answers to a note
name notes.md, and for an untrusted add the file that was not changed.
The new given tests are ledger_label.rs and new versions of memoryd's
edit.rs and serve.rs; the old copies are kept as -task06 and -task10. On
a scratch tree at the end of task 20: the named tests fail and then pass,
5 and 4 mutations are caught, and the gate passes with 944 tests.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Titan serves ornith-1.5-35b-a3b itself, so OpenCode there uses
titan/ornith-1.5-35b-a3b. The harness's device checks still use straylight.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The work moved to titan (macOS), where the gate failed for reasons older
than M5a: sockets under macOS's long TMPDIR pass its 104-byte limit,
gatewayd reads /proc/self for its uid, and /tmp is /private/tmp. Task 16
sets TMPDIR=/tmp for every process Cargo runs (.cargo/config.toml [env]),
has gatewayd take its uid from the owner of a pipe it makes (no dependency,
no unsafe), and hands over a toolkit shell.rs that compares resolved paths.
On a scratch tree the gate passes on titan and on talos with 936 tests, and
a wrong uid fails 7 gatewayd tests. Tasks 16 to 19 are renumbered 17 to 20.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
review::submit counted only the writes still waiting, so a write staged
after an accepted one got that one's id and was lost: never listed, never
acceptable. It now counts every Waiting line of the day. On a scratch tree
with task 16's line applied, both new tests fail and then pass, the old code
and a count of every line are both caught, and the gate passes (938 there;
941 after task 18).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
entry_problem refuses any line that starts with the separator, and a noted
header with an unknown last part is not a header. The new given tests are
forgery.rs and new versions of memory_args.rs and notes.rs; the old copies
are kept as -task02 and -task06. On a scratch tree at the end of task 17,
the three tests named in the task fail and then pass, four mutations of the
fix are caught, and the gate passes with 939 tests.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Task 16: memoryd takes the day and the month of a time with str::get, never a
slice. Task 17: repeated error lines in memoryd, proto, brokerd and bxctl
written once. Neither changes behaviour. Both were applied to a scratch tree
from 5f82260: the gate passes with 936 tests, and four mutations of the new
helpers are caught by the existing tests. The README runs with OPENCODE_MODEL.
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>
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 review of tasks 01 to 15 goes in the implementer log, with the
device check against the owner's Mattermost. Three follow-ups:
- 16: the refusal without a runner says why.
- 17: the texts that say who asked.
- 18: approvals from non-approvers in direct or group messages go to
an approver.
Task 14's given serve_approvals.rs becomes serve_approvals.rs-task14,
since task 18 hands over a new version. Tip T31 is added.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Task 15's session found the defect in serve/approvals.rs, a file it did not list, and stopped.
The comment in task 14 used one name for two values; task 14's tests could not reach it. Lesson T30.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Skeletons for new files only; existing files change by exact edits, so no task overwrites an
earlier one's work. Each task's tests were run against a reference built in order from master,
with the gate at each step and mutations checked; each skeleton compiles and fails its tests, and
every todo!() comment was read beside its signature. The reference is kept off this machine.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The test swaps the file between the check and the open through a `between` hook. A reference fix
passed it and the gate (769 tests) in the working tree, caught two mutations, and was removed.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A script that rewrote the skeleton's comments put `now_ms`'s text on `post`. Task 14's second
session saw the contradiction and deliberated until cut off. Both comments are fixed, each
todo!() comment in tasks 14 and 15 was read beside its signature, and both tasks now say to stop
and quote a comment that does not fit. Lesson T29.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Task 14's session read the crate to learn the APIs and planned `connect` in prose until it was
cut off. `connect`, `Gateway::new` and `catch_up` are now written; the task lists every call with
its signature and makes the copy and the failing test the first actions. Task 15's comment is
rewritten one step per item. Both checked to fail, then pass when filled from their comments.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Task 08's `header` was one todo!() with a dozen branches; Ornith planned it in its head until the
turn ran out (tip T25). `next_message` and `header` are now written as glue over seven small
helpers, and task 09's `poll` over four, each checked to pass when filled from its comments.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Each task's tests were run against a reference at its end state; the end states were replayed
from master in order with the gate at each step (650 to 762 tests); each skeleton compiles
against its tests and fails them. The reference is kept off this machine. Lessons T27 (every
wait in a test has a limit) and T28 (mutate the reference before hand-over) come from this work.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
14 moves the pipe handling out of container.rs (a pure move, replayed on its
own); 15 starts threads with Builder and bounds output collection with a 2 s
grace period; 16 escapes container errors in the log and fixes two texts; 17
fixes toolkit's thread start, casts and the egress-proxy form. Each checked
against a reference, which is not kept. Tips T24 to T26 from this run.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Review findings 1 and 2, both plan defects. curl gains --globoff and a
leading --disable; both podman runs gain --pull=never. The given fetch.rs and
the six golden files change with them. Checked on straylight with a rebuilt
image: a glob URL is one request, a missing image fails at once.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The third attempt filled five functions, then ran out of room planning all of
run in one turn. run and run_container are now given; spawn, Io::start,
Io::finish and answer are small todo!()s. Checked fillable: 11 of 11 passed.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Two sessions ended with nothing written, each out of room while planning the
whole file in one turn. The skeleton has the signatures, the constants and
the steps as comments; the task says to fill one function at a time.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Without it the field was never read in task 11, clippy's dead_code failed the
gate, and #[allow] is forbidden; the implementer stopped on the dilemma.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The first attempt at task 08 did both and stopped without a commit. Also asks
the implementer to debug inside the repository, since OpenCode refuses /tmp.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Task 04 added dependencies to toolkit and its git add line left out the lock
file, so the driver stopped on an unclean tree. The lock change is folded into
task 04's commit.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A draft spec for the owner's review and 13 offline tasks with their given
tests: shared tool arguments and host rules in proto, the sealed fetch
target (M3a finding 14), the toolkit tools and SOCKS5 egress proxy, and
brokerd's [runner], podman argument lists, runtime and proxy lifecycle. Each
task's tests were run against a reference at that task's end state (560 to
638 tests, clippy clean); the reference is not in the repository. Adds the
runner-unavailable runbook entry and tip T23 (ETXTBSY in script tests).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Tasks 01 to 22, the review, and task 23 (the review fixes, with a second,
independent review of them). The one conflict, docs/implementer-lessons.md,
had the same T18 and T19 on both sides; m3a's T20 to T22 follow them.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The review table gains findings 16 to 20, which the two review agents
reported and the first write-up dropped. The independent review of the fix
commits, and what was changed for it, is recorded; task 23's claims about
its tests are corrected. The spec and decisions record the day-long cap, the
ttl_ms bound, the socket-directory rule and the listener's retry.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The test client set its read timeout before every frame. On macOS that
fails with EINVAL once the handler has sent its final frame and closed,
so every admin test that reads a second frame failed there (40 runs of
40 at 2408e2c). The frame is already buffered, so the client now takes
that one refusal as the peer having closed and reads it. The plan's copy
changes with it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The strict.rs in files/ is the merged one and walks task 02's new wire
fixtures, so task 01's gate could not pass; the first run stopped on it.
Task 01 now copies area A's version, checked to give 3 and 5 passed at
task 01's end state.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Task files, the files they copy in (byte-identical to the reference on
m3a-ref), each area's check record, and a README with the per-task
table of what each check exposed. The handoff note is done with.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Areas A to D (audit, policy side, loopd, bxctl) are written on branches
in ~/src/boxmaker-ref-a to -d; area E (broker and approval handoff) is
not started. HANDOFF.md records each area's result, the contract
conflicts between them, the spec defects they found and the remaining
steps in order.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
All ten tasks pass the checklist, the gate, the audit and the device
checks, including a four-turn conversation with a loopd restart and no
cache loss. Reading and probing found four low defects: the busy guard
is released before the final frame on the main path but not on the
three error paths, its Drop skips a poisoned lock, an unreadable
core.md is treated as missing, and bxctl's interactive loop exits on a
failed turn. Task 11 carries the fixes with two new tests, checked
against a fixed copy of the branch.
The Model column is filled in (all Ornith) and one malformed row is
repaired. Two rules are promoted to AGENTS.md.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The tasks build the agent loop on M2a's client: channel messages and
the usage record in proto, four config tables, the tool port and
registry with find_tool and call_tool, the baseline and replay, the
session store, the turn loop with its limits and the append-only
property test, the channel server, loopd serve, bxctl chat, and the
device checks including a four-turn conversation with a restart.
Checked against a private reference implementation: the gate passes
after every task in order, the new suites pass under CPU load, and the
reference passes make verify-device on straylight with no cache loss.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The branch passes every check, including make verify-device on
straylight and repeated timing runs under load. Reading and probing
found that inferproxy does not pass an upstream close on to a client
that is still sending, and that the chunked body reader delivers a
stream only when the caller's buffer fills or the stream ends. Both
were also gaps in the tasks and tests, so tasks 14 and 15 carry the
fixes with new tests checked against the reference.
The Model column is corrected: tasks 04 to 06 and 08 to 13 were Ornith.
Lessons gain four implementer tips and five task-writing tips; three
rules are promoted to AGENTS.md.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The tasks build the inference path: emsha-backed SHA-256, inferproxy,
config, a hand-written HTTP and SSE client, request building, delta
assembly, the chat state machine, the thinking cap, the slot gate with
retry, the startup self-test and on-device verification.
Everything the tasks copy in was checked against a private reference
implementation: the gate passes after each task in order, the timing
tests pass repeatedly under CPU load, and the reference passes the
self-test and all four device checks on straylight. Expected results
for the recorded streams were derived by a separate script.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The branch passes the checklist: seven commits, copied files unchanged,
gate and audit green. Reading and probing found that AuditRecord and
ToolCall accept unknown fields, that large Timestamps panic when
formatted, and that the dependency-direction scripts miss table-form
dependencies and pass when their inputs are missing. The last two
families were gaps in the tasks, not only in the code.
Tasks 08 and 09 carry the fixes, defined by an exhaustive unknown-field
test, a bounded-Timestamp test and an extended gate-script self-test.
All three were checked against the reference implementation and fail on
the current branch.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Seven task files for the implementing model under docs/plans/M1/, with
the test files, byte-exact fixtures, Makefile, deny.toml and gate-script
self-test they copy into place. All of it was verified against a private
reference implementation: the gate passes after every task in order.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
kyleandClaude Opus 5.5