Two existing contracts the envelope broke when every result got every field:
a resource and its tool say the same (books://settings is books.settings,
checked by the_kept_resources_read), and a tool with no warnings has no
warnings key (books.add_account, defect 8). So head and next_steps are on
every tool that can change the books, next_steps is on a read only when
something follows it, guards is on a refusal, and the envelope adds no
warnings of its own (a tool's stay as they are).
The happy-path spec test now exercises books.journal_history (it fails on
a tool it never called), the demo company gets its books.start size check
(Done when 3), and the As-built block of P5 is written.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
Agent reason: the envelope is only worth having if an agent can do its work
by following it. crates/ledger-cli/tests/agent_evals/ holds six scripted
transcripts, no model: onboarding to a first post, a receipt-backed post, a
draft that waits for its receipt and posts when it arrives, a reversal, a
reconciliation, a month-end close. agent_evals_follow_only_next_steps plays
each against a real books server: after books.start a step may call only a
tool (or fetch only a prompt) the previous result's next_steps named, with
the arguments the step gave and the ones the agent decides; setup calls (the
chart and the books an agent finds) are not the agent's moves.
agent_surface.rs holds the over-the-wire tests of the rest: the journal in
words (order, subject and type filters, paging, a bad subject refused),
source corrections through the three tools, the prompts rendered against
their files and refused when malformed, the resource templates advertised
and read, books.start v2 whole and under 8 KiB with the self-report in the
access log only, a new book told to onboard, and a mutating call answering
with the head the journal has and a refusal naming its remedy.
Docs: testing.md's P5 section (every test, and the For-P12 paragraph naming
what a mutant would flip), parity.md (no feature moved; one row), and
architecture.md (the envelope and the guards); mcp-interface.md's overview
sections now describe the surface as built.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
Agent reason: one name for one thing. S4 and S6 kept the old spellings "until
P5" so an agent that learned them did not break; P5 removes them, so the
schema an agent reads has one parameter for each fact and a call cannot name
the same thing twice with different answers.
books.create_invoice takes customer and books.create_bill vendor (a record's
id, its party's id or its name; required), no party_id.
books.receive_payment and books.pay_bill take invoice or bill and
deposit_to or paid_from, no document_id or cash_account; a payment names its
document and its account or is refused saying which to pass. A number that is
one record's id and another's name is still refused, never guessed (INVARIANT
6).
The over-the-wire tests that used the old spellings now use the new
(document_id becomes invoice or bill, cash_account becomes deposit_to or
paid_from, party_id becomes customer or vendor given as the id's text); the
one test that named the alias, party_id_alias_still_works, is now
a_customer_by_id_and_by_name_is_one_record and checks the same two spellings
of a customer, by id and by name, name one record.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
Agent reason: the first call of a session should say who is asking, what is
waiting and what to do, in under 8 KiB, from the same lists the viewer
draws. books.start now carries book.id and book.url (the book's path on the
server, never a path on a disk), you {principal, kind, scopes, client_id},
drafts {building, awaiting_evidence}, reconciliation {tracked,
never_reconciled, behind}, backup {last_at, through_seq, stale}, the
prompts, and needs_attention (the twelve most urgent items, problems first,
with every item counted by kind), beside next_steps now in the envelope's
shape (tool or prompt, why, required; attention problems are required). A
new book is told to run the onboard prompt first, then to record the
company's facts and set up the chart; an active one is offered a draft
entry. The tool carries _meta anthropic/alwaysLoad, so a client that defers
tools loads it up front. The size fit shortens attention items and waiting
drafts after the convention summaries and before the lists, as before.
agent {model, reasoning_effort, harness} is accepted and written to the
call's access record (the access log shows it as agent), self-reported and
never checked, and used for nothing else: not in the answer, not in any
decision. Who asked is the request's alone: it is cleared when the request
ends, so a later resource read on the same thread cannot see it.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
Agent reason: the procedures an agent follows (onboarding, reconciling a
month, closing a month, the CPA hand-off, the year's compliance, an audit)
should be one name away, written against this server's own tools, instead of
rebuilt from the handbook each session. The server now advertises the
prompts capability (also in the bridge's cold-start initialize, which uses
the same capabilities function) and serves prompts/list and prompts/get to a
token that may read: onboard, reconcile {account, month}, month_end_close
{month}, cpa_handoff {year}, annual_compliance {year}, audit {scope}. Each
text is a file in crates/ledger-mcp/prompts/ with the company's name and the
arguments filled in. migrate, process_inbox and receipts_from_email are not
built (D-46, D-28, D-17). audit points at books.verify, books.control_check
and the digest; P8 fills in its own tools.
An argument is one token (letters, digits, - _ . :, at most 40) of the shape
the prompt asks for (a month, a year, a word from a list), so what a caller
passes can never rewrite the text; an unknown prompt, a missing, extra or
malformed argument is refused. A test checks every prompt names only tools
that exist.
The five resource templates of plan 4.5 are advertised (resources/templates/
list): books://entries/{id}, books://attachments/{id}, books://journal/{seq},
books://reports/{kind}{?start,end,as_of} and
books://digest{?since,since_seq,as_of}; the reports now take their dates
from the query (they read only today's or the year to date before), and the
digest reads since_seq (what was posted since a journal line, backdated
entries included).
Tests: the_prompts_name_only_tools_that_exist,
an_argument_is_one_token_of_the_right_shape.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
Agent reason: an agent traces a figure back to what produced it by reading
the journal, but it must never read the file. books.journal_history is the
one way: each line in order with its seq, time, type, a sentence saying what
it did to what, its subject and the entry it belongs to, filtered by
after_seq, type (exact, or a group like entry.) and subject (entry:12,
invoice:3, bill:3, account:10, party:5, asset:2), paged with more, with the
torn tails set aside on open listed beside them. It reads the verified
journal through read_journal_and_format, so a damaged chain is an error, not
a story.
The sentences come from ledger-store's describe.rs, which quotes only
identifiers, dates, amounts and the short names a person typed: never a
setting's value (settings.set names its keys) and never a full TIN (masked,
B-26). Every event type the store writes has a sentence (tested against
EVENT_TYPES). The tool is read-only (books:read); the books://journal/{seq}
and books://entries/{id} resources read through it.
Departure from the plan: the viewer's own sentences (its describe arms in
ledger-gui, which carry icons, routes and today's party and account names)
stay where they are; moving a thousand lines of egui text behind a
formatting trait was not worth the risk to the viewer's tests, so the server
has the shared describer and the viewer keeps its own.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
Accounting reason: a convention, a decision or an open item carries who
decided it, when and on what authority. When that is recorded wrongly the
honest fix is to say so on the record, not to restate the policy: a policy
revision bumps the convention's revision and reads as a change of the rule,
and a new decision supersedes the old one. A source correction changes only
the authority and says why, so the audit trail shows the words were never
altered.
New event types convention.source_revised, decision.source_revised and
open_item.source_revised {id, source, reason} under the journal feature
source_revisions: an older version would refuse the types, or show the old
source as the record's authority, so it stops at the declaration instead
(rules.md B-27) and books that have not had the owner's upgrade refuse the
write with format_upgrade_pending, journaling nothing. Replay sets the
source and nothing else (the text, the revision and the journal line stay);
the projection updates the rows' existing source columns, so no schema
changed. A correction must be one: other words, the same source, no reason
or an unknown record is refused.
books.set_convention takes reason (the text exactly as it stands),
books.record_decision and books.raise_item take revises with reason. Without
reason the old refusal stands and now says how to correct only the source.
8.0.6: the hand-written lines of the three shapes in
books_written_under_older_rules_still_open; fixture old-books/p5_before,
written by b97234ac (a revised convention, two decisions, two open items, a
draft) through old-books-writer/src/bin/p5_before.rs, opens to that
version's figures, so no figure moved; books_p2_served_cut_over_whole
refuses the correction before consent and takes it after; the fuzzer and the
viewer's journal know the types. The live books refuse the new write until
the owner gives the go-ahead; merging does not need it.
Tests: a_source_only_revision_is_journaled,
a_source_revision_must_change_only_the_source,
source_revisions_wait_for_the_upgrade, source_revisions_is_its_own_feature.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
Agent reason: an agent should never have to guess what follows a call. One
place (call_tool, after the tool answers) now gives every tool result the
same four fields: next_steps (what to do next, each with why and whether it
is required), guards (the named refusal and the tool that remedies it),
warnings, and, on any tool that is not read-only, head (where the journal
ends after the call's last write). A tool's own field of the same name is
kept, never overwritten; the text block says what the structured content
says.
Next steps come from the tool's own result (a preview gets its confirm step
with the digest; a post is followed by showing the entry and, only if it
was in error, its reversal; a draft by its lines; a statement by marking
its lines and reconciling; a period list by the close) and, after a change
to an entry or an invoice, from the viewer's own Needs-attention items about
that subject, so the agent and the human read one list.
guards.rs is the registry of refusals (stale_preview,
format_upgrade_pending, receipt_required, approval_required, party_shares,
party_required, open_drafts, untagged_control_line, duplicate_vendor_ref,
first_number_locked), each naming a tool that exists. A refusal whose text
carries a guard gets it as structuredContent.guards; its text is unchanged.
Tests: every_mutating_result_has_head_and_next_steps (the whole router),
the_envelope_keeps_what_a_tool_said_and_carries_a_refusal_guard,
every_guard_names_an_existing_tool, a_refusal_becomes_its_guard,
every_step_a_tool_names_exists.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
GitHub's check on f043db35: bookspec's `rules_md_carries_every_id`
refuses "DECISION" in docs/spec/rules.md (bookspec's word is CHOICE,
P13 step 3); P4's B-43 wrote "DECISION D-47" twice. The portable
macOS job: `a_socket_with_no_other_end_names_no_process_and_isnt_us`
uses `SOCK_CLOEXEC` and Linux's answer for a peerless socket; it is
gated to Linux, which is what it tests.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
The Linux half of the packaging trial, run on this machine and in Debian,
Fedora and Arch containers, and the recipes it produced (docs/packaging.md
holds the table and what it found):
- mise.toml pins cargo-packager 0.11.8 and cargo-generate-rpm 0.21.1, now
that `just package-linux` calls them. cargo-packager-updater is a
library, not a tool, so it is not pinned (update.rs does the checking
with minisign-verify, which is what the updater uses).
- deploy/packaging/: packager.toml (.deb, AppImage, the pacman tarball and
its PKGBUILD), the rpm table in crates/ledger-cli/Cargo.toml, the AUR's
own PKGBUILD (repacks the release tarball), build-linux.sh,
repo-indexes.sh (the apt and dnf indexes, unsigned until the signing
step), local-release.sh (a throwaway-keyed feed to try `update`),
test-linux-packages.sh (linux_packages_install_and_upgrade: install
0.1.0, open books as an ordinary user, upgrade to 0.1.1, check the host
restarted onto the new file, the books' last journal line unchanged and
`update` refuses; passes on Debian trixie, Fedora and Arch). A package
marks itself package-managed with share/wanderbooks/install-method.
- The icon is a placeholder square (the AppImage cannot be built without
one).
- .github/workflows/release.yml: on a version tag, build on ubuntu-24.04
(the oldest glibc the packages claim: a program linked on this machine
needs GLIBC_2.43 and does not start on Debian 13), attach the packages to
the release, and add them to the apt and dnf repositories on gh-pages.
No signing and no macOS or Windows job yet.
- The release profile strips symbols.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
Accounting reason: a confirm must act on the thing that was previewed.
Every tool with confirmed (post_entry, reverse_entry, run_schedules,
dispose_asset, reconcile, close_period, create_invoice, create_bill,
write_off_invoice, purge_tin, consolidate_duplicates, and the S6
settlement tools receive_payment and pay_bill, which gain confirmed and
preview on a clone of the ledger) returns preview (SHA-256 of the
canonical preview) and refuses a confirm whose preview is missing or
stale (stale_preview, before anything is written). A draft's digest is
its own state and the settings seq; an operation's is the tool, its
arguments as sent, the settings seq and the journal head seq, so another
session's write forces a re-preview (D-50).
New tools books.list_drafts, books.set_draft_state, books.discard_draft,
books.preview_entry; reports carry as_of_status (drafts open on or before
their end, unreconciled accounts, settings problems) and take strict; the
CPA package is refused while a draft dated in it is open unless
allow_open_drafts gives a reason, journaled as a decision;
books.digest takes since_seq (what was posted since a journal line,
backdated entries included: defect 6) and lists zero_net_accounts,
drafts_by_state and the reconciliation split. Needs attention gains
'draft awaiting receipt' beside 'draft not posted'; the drafts handbook
is rewritten. Every over-the-wire harness confirms the way an agent does.
Docs: mcp-interface, rules B-43, parity, testing (with the For-P12
paragraph), architecture, and the unit's As-built block.
Tests: drafts.rs (a_draft_waits_for_its_receipt_and_posts,
the_gate_moved_and_the_mcp_error_is_unchanged,
a_reversal_needs_no_receipt_of_its_own, discard_unblocks_the_close_over_the_wire,
report_warns_on_draft_before_as_of, cpa_package_refuses_open_draft,
stale_preview_refused, every_confirm_tool_takes_a_preview,
digest_since_seq_is_posting_time, legacy_reversed_replays_identically,
gui_and_server_gate_agree), attention_parity.rs, format_upgrade.rs.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
Accounting reason: a draft that waits on its receipt, and one thrown
away, are facts about the books' working papers that must survive a
restart and replay identically, or a discarded draft would come back and
hold its period open. entry.draft_state_set and entry.draft_discarded are
drafts_v2's (an older version stops at the declaration instead of
misreading them); no figure moves. Schema 19 (taken at merge as the next
after the current last): entries.posted_seq, and a draft_states table
(drafts are not rows of entries, so the plan's entries.draft_state column
has nowhere to live). posted_seq is the journal line that posted each
entry (entry.posted, entry.reversed, the line that closed a period or
posted a schedule month), identical live and on replay.
8.0.6: a hand-written line of each new shape in
books_written_under_older_rules_still_open; fixture old-books/p4_before
written by 25b52580 (hand entries posted late, a reversal, a document, a
schedule month, a recurring draft, open drafts, a close) opens to the
same figures (serialized entries omit a building state, so no golden
moved); drafts_v2_is_its_own_feature; the fuzzer and GUI journal know the
types.
Tests: draft_states_replay_and_project, drafts_v2_waits_for_the_upgrade,
posted_seq_is_the_posting_lines_seq, an_older_version_stops_at_the_draft_events.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
Accounting reason: an entry that is finished but waits on its receipt is
a different thing from one still being built, and a draft that will
never post must have a way out, or the period that holds it open can
never close (the Lean hypothesis hdrafts). JournalEntry gains
DraftState { Building, AwaitingEvidence { note } }; Ledger gains
set_draft_state (a note: no line, balance or report moves),
discard_draft (a draft only; a posted entry is never removed, B-8; ids
are never reused; what was recorded about the draft goes with it) and
posted_seq (the journal line that posted an entry, for the digest's
since_seq). Lean Step.discard with its preservation proof and three
theorems (a discard moves no figure, removes only its draft, and
unblocks a close), listed in Audit.lean.
Tests: discard_unblocks_close (ledger-core invariants).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
Deploy recipe loses its D-Bus dance, install.sh its macOS skip, README and
site their macOS caveat; packaging, architecture and testing records follow;
U0e As built filled in (D-63).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
Step 1 now covers the issue: preview first (nothing written, the DRAFT PDF,
an abandoned preview uses no number), confirm only after the human has
seen it, the filed issued copy that a later look never replaces,
quantities and prices, the PO, and the snapshot of the customer at issue.
Step 6 is the CPA's list: agings tied to 1200 and 2000, sequential numbers
with voids visible, bills on their real accounts with the vendor's number
and no duplicate, card payments identified for 1099-NEC, deposits as
liabilities, light governance by decision.
A {documents_v2} placeholder (handbook::section_with, the server's
documents_v2 feature test) says "your books have not declared documents_v2
yet" before the owner's upgrade and what is on after it; no other section
changes. Parity, testing (with the For-P12 paragraph) and mcp-interface
notes for S4.5.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
The page a customer pays from must show what the books issued, and only
that. InvoiceDoc gains each line's quantity and unit price ("3 x $33.33"
beside the line's own amount, never recomputed), the customer's PO, memo,
footer, the ways to pay (a QR code through labels::qr_svg for a method
with a link that asks for one), the seller's email and website, and a
Ship-to block. InvoiceVersion separates the issued copy (no stamp, no
settled rows, no "re-created" note) from the current copy, whose
"Re-created on" note prints the as_of it was given, never the wall clock,
so the bytes stay deterministic (B-40 item 3). Every amount reaches the
page through report::amount_parts; every string is data.
ledger_package builds all three copies (issued, DRAFT, current) from one
function over the invoice's own snapshots (A-13, D-55: number, terms,
bill-to contact block, shipping address), falling back to the party record
only for a document made before snapshots, so the three cannot disagree.
invoice_document drops its now-unused `today` argument.
rules.md gains B-39 (terms) and B-40 (numbers and the issued copy) from
the 8.3.13 spec blocks, CONVENTION until their sources are vendored.
No golden moved: the render test never held a hash, it compares two
renders for equal bytes, and still does.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
bookspec/spec holds the first edition: the model and the five laws
(00-model.md) and one file per area, each B-rule a section with its
status, a MUST/MAY statement, sources pinned either to a vendored source
with a verbatim quotation or to an exact external paragraph, its symbolic
form, its examples and its applicability. B-39, B-40, B-43 and B-45 are
reserved sections (allocated, not yet stated in rules.md); B-44 is stated
in rules.md and now carries its bookspec id and CONVENTION status.
Eight new examples show the laws that had none, each figure worked by
hand: an unbalanced entry is refused (Pacioli ch. 36, Pub. 583), a
negative or zero line is refused, a fraction of a cent is refused (ISO
4217), a child account takes its parent's type, the trial balance and the
accounting equation on one small book (cash 6550.25 = loan 2000 + capital
5000 + (sales 750.50 - rent 1200.25)), a reversal leaves the original
standing and a second reversal is refused, a close sweeps revenue and
expenses into Retained Earnings (3999) and the closed quarter takes no
late entry, and cents add exactly. The refusal vocabulary gains
unbalanced, negative_amount, zero_amount, too_few_lines,
finer_than_minor_unit, parent_type_mismatch, already_reversed and
period_closed; a refused entry may carry a non-positive amount.
The checker's spec.rs fails the build on a rule without its section, a
statement without MUST/SHOULD/MAY, a pin that is not verbatim in its
vendored source, examples that differ from the catalogue, or more than 60
pages.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
D-61 (Q-9): the writer sheds the typesetter and the crate map is layered
again. The PDF/XLSX reader (extract.rs, plan P6) is now the plain library
crate ledger-reader, moved whole with git mv and no behavior change.
ledger-store depends on it instead of ledger-import, so the store no
longer builds Typst or the fonts; ledger-import re-exports it as
ledger_import::extract so gui, mcp, host and the fuzz targets compile
unchanged, and still never depends on the store. New crate carries only
the dependencies extract.rs imports (hayro, hayro-interpret, kurbo, zip,
quick-xml, thiserror); fuzz/Cargo.lock gains the new package.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
The agent path was Linux-only because tokens had a keyring store only
there (every other OS hit NOT_YET, so `connect` and the bridge failed).
D-56's destination is one installation for macOS, Windows and Linux.
Both new stores sit behind the same Items trait as the Secret Service,
so INVARIANT 4 (one account, one item, locked is an error) is kept once
and tested once through the file keyring. Windows credentials have a
unique target name, so a probe stands in for a search.
New dependencies (ledger-host only, target-gated): apple-native-keyring-store
(feature keychain) and windows-native-keyring-store, the stores the
keyring-core 1 API we already use is published with. No ledger-core change.
The `portable` job builds and tests the workspace on macos-latest and
windows-latest. Windows is allowed to fail: the rest of the host is
Unix-only (inventory in docs/packaging.md). Not verified locally: no
macOS or Windows target is installed; CI is the evidence.
Also pins cargo-packager, cargo-packager-updater and cargo-generate-rpm
in mise.toml for the trial, and records that the trial is not yet run.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
The owner, 2026-10-29, the question tool: the reader (`extract.rs`)
moves into a small library crate of its own, `ledger-reader`, that
`ledger-store` and `ledger-import` both depend on, so the writer sheds
the typesetter and the crate map is layered again; half a day for the
implementing team once S4.5 is in. 8.2 gains the unit U0d in the
template (its Done-when is `cargo tree`), 8.0.1 its row, 8.0.3 the
crate name (D-53's closed `ledger-extract` subprocess stays closed),
10.1 D-61, 10.3 closes Q-9. STATUS.md, HANDOFF.md and architecture.md
say the same; the arrow stands as built until U0d lands.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
The plan and the status documents say what is true after the merges
of 2026-10-29: served.md section 0, the 8.0.1 rows and unit Status
lines for S4.4, S5, S6, S7 and P6 (landed, with commits), the 8.0.2
resume paragraph turned into the record of what the merges found,
P6's At-merge bullet (the write-off and dedup interaction, the five
tests its interrupted gate never reached, schema 18), P9a's Done-when 1
verified by the green two-user job, D-58 and D-59 in order.
D-60 (the owner): per merge a compile and the unit's own tests; the
exhaustive gates once per batch, at the end. AGENTS.md's two gate
sentences, HANDOFF.md's merging procedure and the plan carry it, with
the reason: five full gates around five merges took twelve hours and
the one gate at the end found the same five tests.
Q-9 (open, for the owner): P6's `ledger-store` -> `ledger-import`
dependency, with the options and the recommendation; architecture.md
marks the arrow as under review. testing.md records the merge's test
changes under P6. STATUS.md gains "Now (2026-10-29)" with the gate
figures (170 suites, 2,944 tests, 77 minutes), the CI state and the
next steps; HANDOFF.md, roadmap.md, gui.md and the parties doc drop
the paused-branch wording; README counts 129 tools and about 2,950
tests.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
The last of the paused branches, from its "not for merge" checkpoint:
extraction in ledger-import (text, pages, page images; capped, caught),
journaled as attachment.text_extracted; dedup by hash with
books.find_attachment and books.consolidate_duplicates; annotations and
books.set_entry_settles; the company record categories; the two fuzz
targets; feature attachment_text. Nineteen files conflicted with the
four P4b merges, resolved as plan 8.0.2 says: P6 retakes schema 18 (it
had taken 16, which S4.3 holds; S5 holds 17); the tool table is 129 and
the destructive count 31; both sides' appends kept with their seams
restored; one B-27 paragraph naming every feature; in the old-books
fixture P6's hand-written reimbursement is entry 7 (its ids collided
with S4.3's invoice); the replay fuzzer's list gains S7's
invoice.written_off; S4.3's issued-PDF test expects the extraction
record P6 journals right after a filing.
One integration defect between S7 and P6, found by the arap property
test and fixed here: under P6's dedup, a write-off whose evidence bytes
were filed already (the same collection letter for a second write-off of
the same invoice) was refused only after its entry had been journaled,
at the final link to the invoice ("already linked"). The write-off now
resolves the dedup before anything is journaled and keeps a link the
record already has to the invoice instead of refusing it; an explicit
existing_attachment_id already on an entry is still refused (S7's
one-receipt-once rule). P6's gate never ran before the pause; it runs on
main after this merge, and its ledger-store -> ledger-import dependency
is the open review point (plan 8.0.1).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
The fourth of the paused branches: Ledger::write_off and the bad_debts
control role, invoice.written_off with its evidence filed in the same
store call, books.write_off_invoice, the write-off on the timeline and
in Needs attention, Lean writeOff_within_bounds, rules.md B-41 and the
bookspec examples. Twenty-eight files conflicted with the S4.4, S5 and
S6 merges; resolved as plan 8.0.2 says, keeping both sides' appends with
their seams restored, and: the scope table at 125; S7's law numbered 37;
B-41 placed before B-42; the handbook's step 4 and recovery procedures
folded into the one invoices-and-bills section; S7's wire tests in their
own module with the confirm S4.4's contract needs; S7's copy of the
payment steps dropped (S6 replaced them with Ledger::settle) and its
write-off journaled through S6's journal_auto_lines; in the old-books
fixture S7's hand-written write-off moved to a second invoice, since
S6's hand-written payment settles invoice 1 in full.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
The owner, 2026-10-29: mutation testing is for local runs, not CI. The
weekly workflow's mutation shards (core, store, import, config, MCP) are
removed; they failed every week on the survivors and timeouts D-33
deferred to P12, and their red hid the run's real findings (a
fuzz_store_replay crash and a fuzz_settings_replay build break, both
being fixed). just mutants-* stays for the desktop; P12 runs the pass
whole. Recorded as D-59; the docs that said the shards keep running are
corrected.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
The last of the paused branches, from its "not for merge" checkpoint:
extraction in ledger-import (text, pages, page images; capped, caught),
journaled as attachment.text_extracted; dedup by hash with
books.find_attachment and books.consolidate_duplicates; annotations and
books.set_entry_settles; the company record categories; the two fuzz
targets; feature attachment_text. Nineteen files conflicted with the
four P4b merges, resolved as plan 8.0.2 says: P6 retakes schema 18 (it
had taken 16, which S4.3 holds; S5 holds 17); the tool table is 129 and
the destructive count 31; both sides' appends kept with their seams
restored; one B-27 paragraph naming every feature; in the old-books
fixture P6's hand-written reimbursement is entry 7 (its ids collided
with S4.3's invoice); the replay fuzzer's list gains S7's
invoice.written_off; S4.3's issued-PDF test expects the extraction
record P6 journals right after a filing.
One integration defect between S7 and P6, found by the arap property
test and fixed here: under P6's dedup, a write-off whose evidence bytes
were filed already (the same collection letter for a second write-off of
the same invoice) was refused only after its entry had been journaled,
at the final link to the invoice ("already linked"). The write-off now
resolves the dedup before anything is journaled and keeps a link the
record already has to the invoice instead of refusing it; an explicit
existing_attachment_id already on an entry is still refused (S7's
one-receipt-once rule). P6's gate never ran before the pause; it runs on
main after this merge, and its ledger-store -> ledger-import dependency
is the open review point (plan 8.0.1).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
The fourth of the paused branches: Ledger::write_off and the bad_debts
control role, invoice.written_off with its evidence filed in the same
store call, books.write_off_invoice, the write-off on the timeline and
in Needs attention, Lean writeOff_within_bounds, rules.md B-41 and the
bookspec examples. Twenty-eight files conflicted with the S4.4, S5 and
S6 merges; resolved as plan 8.0.2 says, keeping both sides' appends with
their seams restored, and: the scope table at 125; S7's law numbered 37;
B-41 placed before B-42; the handbook's step 4 and recovery procedures
folded into the one invoices-and-bills section; S7's wire tests in their
own module with the confirm S4.4's contract needs; S7's copy of the
payment steps dropped (S6 replaced them with Ledger::settle) and its
write-off journaled through S6's journal_auto_lines; in the old-books
fixture S7's hand-written write-off moved to a second invoice, since
S6's hand-written payment settles invoice 1 in full.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
The third of the paused branches: Ledger::settle with discounts and the
processor's fee (kernel::settle, Kani settle_is_exact, Lean
settle_posts_a_document_step), books.receive_payment and books.pay_bill
with applications, books.bills_due, the agings' detail and tie-out, the
aging exports and the CPA package's aging CSVs, the handlers in
tools/settlements.rs. Twelve files conflicted with the S4.4 and S5
merges, resolved as plan 8.0.2 says: both routers; both test appends
(the shared match arm in arap_standing.rs spliced into both tests; the
shared `.unwrap()` in store.rs); the scope table at 124; one
invoices-and-bills handbook (S5's send step 1 with S6's steps 2, 3, 5
and procedures) registered once; docs/mcp-interface.md with S4.4's
create docs and S6's payment docs; S5's law renumbered 35 and S6's 36
(S1's is 34); the parity rows and sections of both. S6's wire tests
create their bills as S4.4's contract now requires (confirmed, the
vendor's number, a line account, the vendor's invoice uploaded) and live
in their own module beside S4.4's tests.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
The second of the paused branches. invoice.sent, Ledger::mark_sent (with
its Lean step), books.mark_sent in its own group, the two attention
items, schema 17. Conflicts, resolved as plan 8.0.2 says: the shared
over-the-wire test file crates/ledger-cli/tests/documents.rs (S4.4's
file kept whole, S5's test appended, its create call given the
confirmed: true the S4.4 contract now needs), types.rs (both appends
kept), docs/parity.md (the review's oracle row and S5's new row).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
The first of the five paused branches (plan 8.0.2, "Resuming the paused
wave"). books.create_invoice and books.create_bill preview without
confirmed and post with it (A-1), the issued PDF filed in the same
store call, the party's full contact block (D-55) on books.update_party
and the invoice's bill-to. The only conflict was docs/parity.md, where
the 2026-10-28 review had written the S1-S4.3 notes on the A/R and A/P
rows and S4.4 its own: both are kept.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
The planning session read plan/served.md, every design document and the
five branches the owner paused on 2026-10-27 against main (44ee4083),
with dry-run merges touching no branch. plan/2026-10-28_DRIFT_REPORT.md
lists every divergence and judges every recorded departure (all kept;
three owed back; one needs a decision: P6 made ledger-store depend on
ledger-import).
The plan: section 0 "What changed on 2026-10-28"; 8.0.1 and the unit
Status lines name S4.4, S5, S6, S7 and P6 as built but unmerged; 8.0.2
records the destination (D-56: the installed app, P11a starts now) and
the resume order with the known conflicts (the shared documents.rs
test file, the types.rs appends, P6 retaking schema 18, the law numbers
34 to 37); 8.0.3 lists the names the EM approved during the wave; P11a
gains step 0 (the macOS and Windows keyring stores, the Windows data
root and pipes, a cross-platform CI build: today the agent path runs on
Linux alone); Q-5 is due and Q-8 asks whether D-47 to D-54 stand; the
Landlock sentences in section 2 and server.toml are struck.
STATUS.md and HANDOFF.md say where things stand and how to resume;
roadmap.md and gui.md are rewritten to today. The docs lose their
stale facts: 122 tools, 2,700 tests, the parity floor at 21 with the
P4b notes the slices never wrote, the ten crates and their dependency
graph (jiff added to the core's allowed list), the cutover of
2026-10-20 recorded in capability-parity.md and parallel-run.md, the
P4b S1 law and its two Kani harnesses added to verification.md (law
34), the proof counts (51 and 9), and README, the site and install.sh
honest about macOS (the viewer runs; connecting an agent waits for
P11a's keyring store, and the installer now says so instead of failing).
just check on main at 44ee4083: exit 0 (167 suites, 2,871 passed, 0
failed, 7 ignored, the server-only build clean). No code changed.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp
The As-built block records step 4 built after U0b, its two items and
their tests; docs/testing.md names them and what P12 must mutate (the
number exemption, the open/partial check, sends.is_empty(), the two day
comparisons with their checked date arithmetic, the last-send max);
docs/parity.md says sent tracking also feeds Needs attention.
S4.4 is working in tools/documents.rs, so the send tool moves beside it
rather than into it: `crates/ledger-mcp/src/tools/sent.rs`, declared like
every other group and named in docs/architecture.md's layout. The two
helpers the document read tools share (`sent_out`, `last_sent`) go to
tools/helpers.rs, which the groups share; the handler and its own
argument readers stay private to `sent`.
No tool name, schema or behavior changes: books.mark_sent is where it
was on the wire, only in a file of its own.
rules.md B-27 lists `documents_v2`, the one feature S4 to S7 share
(D-52), including `invoice.sent`; receipt-store.md lists the event;
parity.md gains the sent-tracking row (done: implemented, tested, and
golden-covered by the trial balance of the fixture book around a send);
testing.md records the tests and what P12 must mutate, and names the
two attention predicates as deferred with step 4 until U0b lands.
AGENTS.md's formal-methods bar makes every new Ledger command a Lean
Step with a preservation proof, so `Ledger::mark_sent` gets one even
though the unit expected no Lean change (the As-built block records
that). `Documents.lean`'s `Books` gains the sends recorded, and
`Step.markSent` adds one on a document the lines carry. A send touches
no line, so every control balance, every document's sum and what is
unattached are unchanged through any day
(`Step.markSent_changes_no_figure`, no axioms), and `Reachable.tagged`
covers the new step.
`Audit.lean` lists the theorem; rules.md B-38's model now covers the
record side of a document's life as well as its figures.
`books.mark_sent` records the send the human or a delegate made: the
invoice (`invoice:7` or its id), the day, who sent it (the token's name
unless given), how (email by default, or mail, portal, hand, another
channel), to whom (the bill-to email, else the customer's), whether it
was the original or a reminder (original on the first send, reminder
after), and the channel's own id. It answers with every send on the
invoice, newest last. `books.documents` rows gain `last_sent` and
`books.document` shows the sends, so "did we send it, and when" is one
read away, while a bill (which we receive) has none.
The schema lint refuses a parameter named `to` (a path-name rule), so
the recipients parameter is `recipients`; the journal event and the
result keep the plan's `to`. The tool needs `books:write`: it changes
the record, never a figure.
The handbook gains the invoices-and-bills section (S4.5 had not landed
yet) with the send procedure: the books never send, the PDF goes to the
human, and each send is recorded afterwards.
Over the wire, mark_sent_records_who_and_how: two sends (the second a
reminder with a message id), the list's `last_sent`, the document's
`sends`, and the invoice still owing its total. The P2 cutover refuses
a send before consent and takes it after.
The CPA's first tie-out is that the aging equals the control account, so
books.export and the CPA package now write ar_aging.csv and ap_aging.csv
from one builder: one row per open document (number, party, dates, terms,
total, paid, discounted, outstanding, bucket, days overdue) and a TOTAL
line carrying the outstanding sum, the GL account and balance, and the
difference (always 0.00; B-23, B-38 INVARIANT 1). XLSX and PDF lay the
same aging out as a statement with the tie-out beneath it.
The package README states the tie-out, the gross discount method (A-8)
and the direct write-off method (A-5). The chart templates gain 6995
Purchase Discounts so no template puts another account at a control
role's code (4100 and 6060 already match). The viewer's Money out list
gains the discount-deadline column and colours a due-soon or
discount-open date; the screenshot is re-rendered.
Test: the_aging_csvs_tie_to_the_control_check in ledger-package; the
existing package file-order test and mcp_journeys see the two new files.
books.receive_payment and books.pay_bill take applications, each
document's cash with an optional discount and reason, and a customer
payment may carry the processor's fee; the result names every document it
settled and what reached the account. books.bills_due lists what to pay
and the running total by due date. The agings take detail and return the
control account's tie-out, and books.export writes ar_aging/ap_aging.
The handlers live in tools/settlements.rs (the EM's split, so S4.4's
documents.rs stays free); documents.rs keeps the documents, the void and
the control check. docs/mcp-interface.md documents every changed tool and
the new one; docs/architecture.md names the group; docs/testing.md
records the tests and P12's mutation targets.
Old names (document_id, cash_account) still work; the output keeps
remaining/status for one document. A fee's processor is named in the
description, not tagged as the entry's vendor: ledger tags are entry-wide
and tagging would move the customer's party totals and 1099-NEC.
The handbook's invoices-and-bills section gains the payment, discount,
fee and bounce-replacement procedures (S4.5 can extend it). Over the wire:
one_payment_settles_two_invoices, late_discount_needs_a_reason,
a_fee_posts_gross_to_receivables_and_net_to_the_bank,
card_paid_bill_is_not_1099_nec,
multi_document_payment_counts_once_for_1099_and_party_totals,
overpayment_refused_with_the_excess,
bills_due_orders_by_due_and_discount and the_aging_export_ties_out.
Terms cannot yet be journaled on a document (S4.3) and a bill line's
account is not an MCP field (S4.4), so bill_on_terms_with_discount,
equipment_bill_payment_is_investing_cash and the discount half of
bills_due are core tests until those land; the unit's As-built block
records this.
One payment had to name one document and could take no discount or
processor fee, so paying several invoices from one deposit meant several
entries and the fee had nowhere to go. This adds the kernel's exact
settlement arithmetic (applications total, net deposit), a Ledger::settle
that posts one entry for a customer's or vendor's documents, and the gross
method and gross fee A-8/A-4 chose.
- kernel::settle: sum(cash)+sum(discounts)=applications, sum(cash)-fee=
deposit, None on overflow or a negative deposit. Kani settle_is_exact
plus the Decimal bridge.
- Ledger::settle / settlement_lines: one party and one kind per
settlement, one application per document, each cash+discount within
payment_room; a discount within the terms (discount_max and
discount_by) needs no reason, a larger or later one does; a bill with
an asset line is refused with D-49's remedy; a fee is a customer
payment's only and posts gross.
- Rows are payment lines in application order, then discount lines, the
fee line and the cash line, so a journaled event's line indices are
stable; replay tags exactly the line an event names and leaves an
unmatched line unattached (never refuses).
- receive_payment_with and pay_bill_with become one-application calls.
- Roles: sales_allowances (4100), purchase_discounts (6995) and
payment_processing_fees (6060), constants used on demand but not added
to control_accounts() (existing charts may use those codes for other
things).
- Lean: settle posts through Step.document, so settle_posts_a_document_step
and settle_unattached, no new constructor; both in Audit.lean.
- Core tests: bill_on_terms_with_discount, late_discount_needs_a_reason,
discount_on_an_asset_bill_is_refused_with_the_remedy,
a_fee_posts_gross_to_receivables_and_net_to_the_bank,
a_settlement_names_its_lines_and_refuses_a_document_twice, and
equipment_bill_payment_is_investing_cash in the cash-flow suite.
- Bookspec B-42 is stated, with its id BS-arap-6; the catalogue and index
now record it as CONVENTION.
just check and just prove green.
rules.md B-41 states the write-off: the direct method the owner chose
(A-5), Dr 6990 / Cr 1200 with the receivable tagged, at most what is
owed and at most the least owed from that day on (a partial write-off as
§1.166-3(a) allows), refused on cash-basis books (§1.166-1(e)), its
evidence filed in the same call (§1.166-2(a)); bills are never written
off (ASC 405-20-40-1); and a recovery is not a new movement (D-49): the
same year, reverse the write-off and record the payment; a later year, a
plain entry to a bad-debt-recoveries revenue account, the invoice staying
written off (IRC §111).
bookspec states B-41 for the first time and gives B-38 its example: two
new examples, bad-debt-direct-write-off (two partial write-offs of one
invoice, the balances and the trial balance worked by hand) and
arap-hand-control-posting-refused (a hand line an earlier version wrote
replays, the same line written today is refused). The example schema
gains the dated write_off event, lets a refused entry say so (an entry
posts whole or not at all; older and refused never combine), and its
refusal vocabulary gains untagged_control_line, B-38's control guard.
The registers: docs/architecture.md names the new tool group;
docs/receipt-store.md lists invoice.written_off and what replay does with
it; docs/parity.md records the slice (no row changes status, the floor
stands); docs/testing.md records every test and what P12 must mutate;
docs/verification.md the two Lean theorems and what they cover.
No figure moves: no golden file changed.
One tool for the bad debt (rules.md B-41): it writes off an invoice that
will not be collected, Dr 6990 Bad Debt Expense / Cr 1200 Accounts
Receivable, the receivable line tagged with the invoice so the document
reads written_off and its figures stay derived (B-38).
The evidence is the point, and the tool takes exactly one of
upload_id (an upload, filed with the write-off) and
existing_attachment_id (one already filed, linked to the write-off's
entry and the invoice); neither or both is refused. amount defaults to
what the invoice owes and may be partial.
Without confirmed it is a preview: the engine runs on a clone (every
refusal the post would give) and the answer is the figures the write-off
would leave, entry_id 0, nothing journaled.
The tool lives in its own group file, tools/write_offs.rs (plan U0a), so
S4.4's document tools and this one land apart; it is destructive, needs
books:write to preview and books:post to post (scopes.rs), and is
documented under "Tools — A/R and A/P". The destructive-tool count, the
tool tables and the journey test's tool count each move by one.
Tests over the wire: write_off_then_recovery, write_off_needs_evidence.
An invoice is issued, not edited. Without `confirmed` the create tools now
preview: the posting (with the accounts' codes, which the live chart may
not hold yet), the number it will take, its due date, terms and early
discount, warnings, and for an invoice the DRAFT copy of the PDF, handed
over as a file. Nothing is written and no number is used up. With
`confirmed` the invoice posts, takes the number the preview showed, and
files the issued copy with itself in the same call: `receipt.attached` on
its entry plus the link to the invoice, its SHA-256 in the event (A-1,
A-7, B-40). The copy the customer received is therefore the copy in the
books, and `books.document`'s re-created page can never be mistaken for
it. The result also carries a `mailto` link for a human to send it; the
books never send (A-4).
A bill keeps what the plan says a bill is: the vendor's own invoice
number, the vendor's invoice filed on its entry as its evidence, and, in
books that have declared `documents_v2`, an account on every line, so no
bill can land in General Expenses by default any more (A-9, A-10, A-13).
A vendor number entered twice is refused unless the caller says why, and
the reason is journaled with the second bill (B-40 item 4). Books that
have not declared the feature keep recording bills exactly as before, so
the owner's books keep working between units (D-39, 8.0.4).
Terms are resolved in one place and in one order: the due date given,
else the terms given, else the customer's or vendor's own default, else
the books' `invoicing.default_terms`; a document given only a due date
stores no terms, so an old-shape invoice is still written byte for byte
as before (B-39, A-13).
D-55: a party carries the owner's full contact block (phone, contact
person, billing and shipping addresses, billing email). It rides the same
`party.updated` event behind `documents_v2`, and an invoice's bill-to is
snapshotted at issue with the billing address and billing email in place
of the general ones when the party has them — `PartySnapshot::from_party`
is the one place that rule lives, so a later edit cannot change what was
sent, and the next invoice takes today's details.
Tests: the wire tests are the new `crates/ledger-cli/tests/documents.rs`
(preview writes nothing, the DRAFT copy, the issued copy's SHA-256 equals
`books.get_receipt`'s, numbers run, a bill's file and terms, the duplicate
refusal, the line-account rules, the terms fallback, the `party_id` alias
and the ambiguity it refuses, the contact block's set and clear, and the
bill-to at issue); `format_upgrade.rs` gains
`documents_v2_refused_until_consent` and the cut-over test now refuses and
then makes a document carrying a new field (8.0.6 item 3); the store
tests are `the_contact_block_round_trips_and_replays` and
`an_invoice_keeps_the_billing_details_it_was_issued_with`, and
`books_written_under_older_rules_still_open` gains a hand-written
contact-block event (8.0.6 item 1).
Merge does not need the owner's go-ahead: the live books refuse a write
that needs `documents_v2` until he gives it, and take every other write as
before (8.0.6 item 6). No figure moved: S4 moves none.
S4.1 gave documents a number, terms, line accounts and snapshots, and S4.2
gave parties an email and default terms, but the store journaled none of
them: a document carrying one was refused by a temporary guard, because a
public write that replay loses would break I-3. This slice makes the
journal carry them.
- `invoice.created` and `bill.created` write a number, terms, bill-to and
seller snapshots, references, memo and a duplicate's reason, and each
line its account, quantity and unit price, every key only when set. A
document naming none is journaled byte for byte as every earlier version
wrote it (`no_new_fields_journals_byte_for_byte`), so the owner's live
books keep working.
- All of it is behind `documents_v2` (B-27): a write that names one waits
for the owner's upgrade, refused with `format_upgrade_pending` and
journaling nothing; a write that names none never waits
(`documents_v2_waits_for_the_upgrade`).
- `invoicing.first_number` is settable only while no numbered invoice
exists (A-2, B-40): the series has begun, and a number is never reused or
renumbered. Refused at `set_setting`, before anything is journaled.
- The issued copy is filed in the issuing call (A-7, B-40 item 3):
`create_invoice_issued` journals the entry, the invoice, then
`receipt.attached` on that entry (`INV-0007.pdf`, `application/pdf`,
"Issued invoice") and `attachment.linked` to the invoice. There is no
`issued_pdf` field: the attachment id is allocated after the entry exists
(CD-A6), and the link is how the copy is found and how its SHA-256 is
checked. Every refusal happens before the first journal line; the PDF's
bytes are stored before it, so a refused issue leaves an unreferenced,
deduplicated blob at worst.
- A preview runs the issue on a clone and journals nothing: the posting,
the total, the number it would take and the due date (`preview_invoice`,
`preview_bill`). Books that haven't upgraded keep the legacy shape: the
copy is filed as "Invoice {id}.pdf" and no number is taken (migration
item 5).
- Schema 16 adds the document columns; a line's new fields ride
`lines_json` as its project does. `differs_from` already compares whole
invoices, bills and parties, so it sees every new field.
8.0.6: (1) `books_written_under_older_rules_still_open` hand-writes both
new document shapes through JournalWriter and still opens; (2) a new
fixture `old-books/p4_s43_documents` was written by main d5385a7e (S4.2,
the version before this slice) through the old-books-writer harness, with
its expected.json; this version opens it and reports the same figures;
(3) see the note on R-A step 4 in the PR (the over-the-wire document write
needs S4.4's arguments; S4.2's invoicing-setting entry already covers the
feature's cut-over); (4) `documents_v2_is_its_own_feature` is green; (5) no
figure moves: no golden changed; (6) merge does not need the owner's
go-ahead; the live books refuse the new writes until he gives it.
just check green (no kernel function, Lean theorem or Ledger command, so no
`just prove` change).
Review finding: `--http` parsed whatever address it was given and bound
it. The door is the plaintext MCP service with a static bearer token;
plan 3.9 and the P9a model put a TLS proxy or tunnel in front of a
loopback listener, and no decision authorises exposing the books
unencrypted. `--http 0.0.0.0:7411` would have done exactly that.
`local::tunnel_address` now accepts only a loopback address
(127.0.0.0/8 or ::1; `127.0.0.1:7411` stays the default) and refuses
anything else with a message naming loopback and why. `server` refuses
while parsing, before it takes the lock or binds a socket; `serve`
refuses before it starts a host, so nothing is left half-open. The
bind-failure behaviour for a loopback address is unchanged (said, not
fatal, 3.9), as is the unit's hardening.
The test `the_tunnel_door_binds_loopback_only` refuses 0.0.0.0, a public
sample address and [::] through `server`, checks no socket or host was
left behind, refuses through `serve`, then starts a loopback listener and
reaches `books.start` with the bearer.
No accounting code, report or figure changes.
The books can live on a machine of the owner's own (plan 3.8, 8.2 P9a,
D-51). Three things make that honest instead of nominal:
- `--isolation same_user|enforced` sets `server::Isolation`, carried
through `Served` and `BooksServer` to `books.start`. Locally the books
stay `same_user (advisory)`; under the self-host unit the server is told
`enforced` and reports `service_user (enforced)`. The server never
guesses, and the literal it used to report is gone.
- `--http [ADDR]` binds the same MCP router on loopback (default
127.0.0.1:7411) beside the Unix sockets, before the ready line, for the
TLS proxy or tunnel in front of the server. It is the same bearer door
and the same graceful close; a failed bind is reported and is not fatal
(3.9). `serve` carries both flags through the host, so the system unit
can ask for them once.
`deploy/wanderbooks-server.service` is the unit: a dedicated account, a
0700 data root, `ProtectSystem=strict`, no seccomp or Landlock (D-32), and
loopback-only networking. `deploy/README.md` is the recipe: install, setup,
move the books by `books adopt` with its head check (D-31: a copied book
is the transport), mint a bearer token, put Caddy or Tailscale Funnel in
front of the loopback door, and point Claude Code and Codex at it.
The two-account CI job (`.github/workflows/two-user.yml`) makes the
deployment and runs `crates/ledger-cli/tests/two_user.rs` as another
account: `agent_uid_cannot_read_tenant` proves the tenant root and its
journal are refused, then reads, writes, uploads, files, exports and
verifies over HTTP with the bearer (in a second book, so the moved books'
head is not advanced); `self_host_adopt_same_head` compares the moved
books' head and event count with the source clone's.
`isolation_label_is_enforced_under_the_unit` runs everywhere, including
the unit-file read that keeps the recipe from drifting.
Accounting reason for the shape: nothing here changes an entry, a report
or a figure. It changes who can reach them: the move is lossless and
verifiable at the head, and the boundary is the operating system's, not a
promise in our text. No journal code, kernel function, Kani harness or
Lean step is touched, so `just prove` is not required.
No go-ahead needed: no new journal feature or event; the live books are
unaffected.
books.start passed Integrity::default() to the attention engine, so it
asserted "database matches, no file failures". On a book whose database
differs from the journal, or whose stored file is missing or corrupt, the
viewer showed the database/stored-files items and the server did not: the
two lists diverged exactly where they matter most.
Call find, which computes the check: every distinct stored blob is read
and hash-checked once (a damaged blob is still reported per attachment
filed from it), and Store::disk_check now makes the viewer's comparison
through a snapshot of its data directory instead of reporting a match.
The parity test now damages its own fixture (removes the receipt's blob,
then changes a projected account row) and asserts both lists still agree
item for item and both name the damage. On the demo books (138
attachments, 138 distinct blobs, debug build) books.start takes about
0.26 s warm, 0.29 s cold, so the real check costs no decision.
swaitsandClaude Fable 5.1