The owner, 2026-10-18: one app to simplify installing and updating.
One wanderbooks executable whose modes run as separate processes; the
books open while the app, an agent session (its stdio bridge), a serve
terminal or a CLI command holds them, and close after the last (still
no background service); a signed, notarized dmg and a per-user Windows
installer that update themselves, AUR, apt, dnf and a self-updating
AppImage on Linux. New section 3.9 (leases, the helper's clean host
start, local sockets instead of TCP, restarts, one version at a time,
the server's format-upgrade gate); P2, P11 and the decisions follow.
Five review rounds, the last clean.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T3gXMVt7XJPi6ens1or9Dj
The owner, 2026-10-18: no background service for now. wanderbooks serve
runs in a terminal (server plus archive process) until stopped; nothing
is installed to run in the background. P2's scope, cutover, tests,
limitations, user flows and P11 follow.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T3gXMVt7XJPi6ens1or9Dj
The first company's P1 cutover found the K-1 worksheet printing a
member's share as typed: "50%" read from a settings file printed
0.50, the same share kept by the books as 0.5 printed 0.5. Shares now
print at least two places whatever their spelling.
Every company document now goes into the books (D-20, owner,
2026-10-17), and formation papers and plans print the company's EIN,
which the filing scan refused as a printed TIN. It now passes the
company's own EIN, as its settings hold it, printed 2-7, and nothing
else; a vault TIN is still refused (rules.md B-26). Its refusal no
longer suggests converting a file to get past it.
plan/served.md: D-20; correcting a record's source without changing
its words (P5).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T3gXMVt7XJPi6ens1or9Dj
Many test helpers removed their temp dir only before the test, or only
on pass, so runs left them behind: 117,000 filled /tmp (RAM here) and
stopped a mutation run. Each now holds a guard that removes the dir on
drop, unwind included; servers started on a test dir hold its guard
until they stop. A full cargo test --workspace leaves 0 dirs (859
before). just sweep-test-tmp stays for runs killed mid-test.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T3gXMVt7XJPi6ens1or9Dj
Settings, conventions (posting templates, approval rules, policies,
close steps, calendar, mappings, notes), decisions, open items and
entry approvals are journal events (journal format 2). Books that keep
their settings never read wanderbooks.toml or config.json; the server
and the viewer take the vault key from --vault-key-file instead.
books.import_settings moves a settings file in once, exactly as the
server reads it. Approval rules, when enforced, gate every posting
path. Replay refuses what it can't read as newer, never as corrupt.
Lean: a record changes no figure. Mutation on the whole diff: every
survivor killed by a test but two equivalent (docs/testing.md). The
Company page says a late open item is overdue in words, not colour
alone.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T3gXMVt7XJPi6ens1or9Dj
CI failed on a_file_swapped_after_the_check_is_never_read: its filesystem
handed the deleted checked file's inode number to the file made in its
place, so device and inode matched for another file. The opened file
must now also match the checked file's size, modification time and change
time. The test now models a real swap (the replacement exists before the
original goes), the reused-inode case, an in-place rewrite, and that the
unchanged file is read.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T3gXMVt7XJPi6ens1or9Dj
Owner, 2026-10-17. D-19 settled: one installer and updater per
system; server, viewer and CLI ship together but stay separate
programs; no launcher; the viewer gains back up, export, restore and
new books (the last two through the server). P2 gains verified,
non-destructive restore onto new or existing installs, and restore
drills.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T3gXMVt7XJPi6ens1or9Dj
plan/served.md: defect 21 (found in P0: the CPA package's
general-journal.iif ignores as_of, so an as-of package includes entries
dated after it; its xlsx and csv files respect it) joins the defect
count, P3 beside defect 7, and the P10 re-run; 5.6 records that the
first company's goldens are kept privately in its own books dir
(export_cpa_package zips plus a canonical JSON form made by its
goldens/canon.py, which normalizes per-export text, with the comparison
protocol in its README), not in this public repo, by decision.
plan/STATUS.md: P0 is in final review.
Review round 3: STATUS says what is true (P0 in final review, the
mutation runs done).
Adds defect 22 (depreciation through an asset's final month plugs to
the full amount mid-month) to P3 and to the P10 re-audit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T3gXMVt7XJPi6ens1or9Dj
docs/mcp-interface.md: the connect text and its budgets, the handbook
(tool and resources), annotations, failing closed, import_local_file's
roots and refusals, the set_w9 path rule and where TIN-bearing files may
go. docs/testing.md: the P0 laws and the tests that check them.
docs/architecture.md: the journal format gate and the server's
fail-closed store lock.
Round-2 review (2026-10-16): mcp-interface.md lists all 26 destructive
tools (unlink_receipt, unlink_attachment and correct_asset were missing)
and openWorldHint false on every tool; the connect text states the
parties rule and the exact receipt rule, and how it fits its budget; the
lint covers every refusal; the import roots refuse /, the home folder and
folders holding the books or a key, the open is a component walk, every
missing or outside path gets one refusal, and the TIN-file registry and
text scan (and what they can't catch) are described. architecture.md:
the frozen envelope, the lenient next-line prev, no repair of newer
books. testing.md: the new laws and tests. verification.md and
parity.md: P0 notes (no new law to prove, no status moved).
The TIN text check, narrowed (2026-10-16; rules.md B-26,
mcp-interface.md): text is refused for a TIN the vault holds, in its
common renderings, or a stand-alone printed SSN or EIN shape, so bank
exports go in; the oracle a vault comparison gives is recorded as a
known limit for P7. testing.md: the P0 mutation run of the final diff,
per crate, and how each survivor was resolved (killed by a named test,
removed by simpler code, or argued equivalent).
The probe bound (rules.md B-26, mcp-interface.md): rule (a) answers
for at most two candidates a filing, so finding a TIN from its last four
takes on the order of 50,000 refused filings; the open limit now names
that path, closed by P2's access log and P7's human-only TIN channel.
testing.md: the law and its tests; tin_files.rs mutated again, 74 of
74 caught.
Review round 3: rules.md B-28 (four-digit years, amounts a line can
carry, checked products, bounded recurring runs) and B-26/B-27 updates
(BOM-less UTF-16, when set_w9 records); mcp-interface.md matches the
code (25 destructive tools, the five idempotent ones, no reverse_entry
receipt checklist, strictly above the threshold, no tool text names
wanderbooks.toml, dot names and roots inside the books refused, the new
settings lists); testing.md lists the new laws and this round's
mutation run.
Review round 4: rules.md B-28 says the bounds are for new commands
only, replay accepts every line an earlier version accepted, and derived
figures refuse rather than panic; mcp-interface.md documents plan_error
and the nullable totals; testing.md lists the old-books contracts and this round's
mutation run (core, store, MCP, config, GUI).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T3gXMVt7XJPi6ens1or9Dj
- The 13K-character instructions (clients cut them short) become a
connect text of at most 1,400 characters whose first 512 stand alone,
plus books.handbook and books://handbook/{section}: the old guidance by
section, with no helper commands, home paths, pull requests or
file-reading suggestions (tool_text_lint).
- Every tool has a title and read-only/destructive hints; no
requiresUserInteraction anywhere (owner decision 2026-10-16).
- books.import_local_file reads only regular files inside the import
roots (the downloads folder by default, or --import-root), never the
data dir, books dir, vault or the key's folder, no symlink escapes or
hard links, and the file opened must be the file checked.
- The store lock is a tokio mutex taken off the runtime threads.
- get_info refreshes settings first; a bare books://digest reads.
Review fixes (P0 review, 2026-10-16):
- Fail closed: a tokio mutex never poisons and spawn_blocking caught the
panic, so after a panic mid-mutation the server kept serving a Store
behind its journal. A panic (the task, or while holding the store)
now stops the shared store; every later request on every session is
refused with a plain message; a restart re-verifies the journal. The
join error goes through redact_refusal and never carries the panic's
text.
- books.set_w9 w9_file_path read any path (a link named w9.pdf to the
vault key was stored and get_w9 wrote it out): it now goes through the
import roots' lockdown, its type judged on the canonical path.
- Sealed even inside a root: the key file in force, where a linked one
leads, both folders, and the default key folder whatever the settings
say. Refusals check the roots first, so outside them every path gets
one message (no probing what exists); inside, other books-dir files say
to copy them out first.
- The open is O_NOFOLLOW | O_NONBLOCK (a pipe swapped in after the check
can't hang a thread) and must match the checked file's device and
inode. New dependency: libc (unix only), for those two open flags,
which std doesn't name.
- Default import roots are the downloads folder only (the plan); files
with full TINs (get_w9, vendor_1099 include_tins, the CPA package with
TINs) are never written into an import root, and the server refuses to
import any file it wrote with TINs (by SHA-256).
- set_tin and set_w9 overwrite vault records: destructive_hint = true;
annotations_complete pins the key tools' hints. The read-only hint is
checked in behaviour on every spec_mcp call (journal and data-dir files
unchanged).
- Text: the connect text names the settings file as the one local file
to edit, says every entry balances on real accounts and names its
party, says invoice/bill/payment/payroll tools post unchecked, rewords
the exemptions, and shows the company name as one plain line; the
handbook takes section "all", lists its ids in the schema and says to
read contract first; books.settings drops the config.json path and no
error names the data dir or the key file. tool_text_lint covers the
whole tool JSON and resources and bans git/commit/WL-0001; tool
descriptions over 700 characters are pinned to today's eight (P5).
- digest: an empty since= is bare; the default date is pinned on every
day of a month.
Round-2 review fixes (2026-10-16):
- TIN files (B-26): the in-memory, per-session set of TIN-file hashes
became one registry beside the vault (vault/tin-files.sha256, hashes
only), shared by every session and every later server. It holds the
W-9 set_w9 reads (recorded as read), get_w9 copies, 1099 worksheets
with TINs, CPA packages with TINs and each file in one that holds a
TIN; a file is recorded before it is written, and a failure refuses.
import_local_file, attach_receipt and file_document (base64 was never
checked) refuse a recorded file, and a text file (by type, name, or
UTF-8/UTF-16 decoding) holding a full TIN. Not caught: an altered
binary copy; the docs say so (P2/P7 remove the path tools).
- The text check is narrow (owner/engineer decision, 2026-10-16: bank
exports are filed for reconciliation, and has_full_tin's "nine digits
however separated" refused their routing numbers, long references and
dates run into amounts). Text is refused when it holds (a) a TIN the
vault holds for a party, anywhere, as nine digits, 3-2-4 joined by
hyphens, spaces or dots, or 2-7 joined by a hyphen or space, compared
on the server against the decrypted vault and never returned or
logged (a vault that can't be opened leaves (b)); or (b) a stand-alone
printed SSN (123-45-6789) or EIN (12-3456789) shape: no digit, and no
hyphen or dot followed by a digit, beside it. Full-width digits and
Unicode hyphens and spaces read as ASCII. has_full_tin still guards
refusals and results (redact_refusal). Known limit, for P7: a refusal
under (a) tells the agent a text it chose holds a vault TIN.
New dependency: zeroize in ledger-mcp (already in the workspace
through ledger-vault, same major version), so the vault's TINs decoded
for the scan are wiped from memory on drop as the vault's own are.
- Mutation testing of the final P0 diff (docs/testing.md): tests for
each law a survivor broke (handbook titles, a dropped default root, a
.jpg W-9, the waker, the walk's folder flags against a link and a
pipe, the 40-link limit, save paths refused as such, the package
marking only its TIN files, a registry that can't be read, text types,
the short receipt rule's counts and a 12-character code). Simpler
code where only a race could tell: inspect places an existing path by
canonicalize itself (no second comparison), read_checked's read limit
is its one size check, openat fails on -1 as specified. The 25 MiB
upload test removes its 50 MB of books however it ends.
- The GUI's stale-rule prompt no longer says to open a pull request: it
asks the agent to find the figure from the rule's official authority,
confirm it with the human, pass it explicitly, and tell the human so it
can be reported upstream (as the handbook says); a lib test pins it and
bans files, git, GitHub and pull requests from every attention prompt.
- Import roots (local_files.rs INVARIANT 0): XDG_DOWNLOAD_DIR="$HOME/"
made $HOME (or /) the root. No root may be /, the home folder or hold
it, or hold the books, data dir, vault or a key: a named one stops the
server at startup, a default one is dropped; the default falls back to
~/Downloads only if that is a real folder other than $HOME; with no
root, an import says so.
- The read is a component walk from / with openat(O_NOFOLLOW |
O_DIRECTORY) per folder and O_NOFOLLOW | O_NONBLOCK for the file, so a
folder swapped for a link after the check fails; no step may land on a
protected inode (one renamed into a root); dev/ino and nlink stay. No
/proc. Portable to macOS. (libc's openat, two small unsafe blocks.)
- One refusal for missing, dangling and outside paths: resolved()
follows every link, dangling ones too, so a link's target existing or
not no longer changes the answer.
- Sealing: the key files (in force, default) and their targets always;
their folders only when not $HOME, not a root, not holding a root; and
~/.config/wanderbooks whatever XDG_CONFIG_HOME says.
- Agent text (I-12): RuleNotKnown names the authority (IRS, with its
address) and says to tell the human, never a repo path, PR or GitHub;
the digest's stale_rules drop the repo catalog and path for it; vault,
outside-books, STOPPED and local-file refusals name no git, path or
flag. tool_text_lint bans docs/, .md, /home, data dir, --data-dir,
--import-root, vault.key, .jsonl, sqlite, .gitignore, github, pull
request, and also lints every error template (scanned from source),
built error texts, and the server's refusal constants.
- The receipt rule told (ReceiptPolicy::rule) is the one enforced: the
exact threshold in the base currency (no "$", no rounding), required
and exempt accounts, reversals exempt; the connect text fits 1,400
characters by dropping reasons, then codes, never the rule.
- The connect text follows the parties setting; "offer books.digest"
everywhere. shown_name drops Unicode format characters (bidi, zero
width).
- read_only_tools_change_nothing compares file contents, the database
by every row, the vault and the settings file (raw database bytes did
not change either; rows are the stable comparison).
- import_local_file takes the name and type from the canonical path;
the digest's journal repair says when the tail was set aside, not the
journal's file name; "version-history messages" back in the TIN rule;
openWorldHint false on every tool, pinned in annotations_complete.
- Outside the MCP crate, for the same texts: ledger-core's DatedRule
gains `authority` (the statute or publication with its address) and
RuleNotKnown carries it; ledger-store's "sqlite error" reads "database
error"; ledger-config's ReceiptPolicy::rule takes the base currency.
New dev-dependency: rusqlite in ledger-mcp (already in the workspace,
same version and features as ledger-store), to read the database's
rows in read_only_tools_change_nothing.
- The filing scan is no TIN oracle (tin_files.rs INVARIANT 4): text
holding more than two TIN-shaped numbers (any rendering rule (a)
reads) that end in a known TIN's last four (the books' TIN and W-9
facts, via known_tin_last4, before the vault is opened; then the
vault's own) is refused with TIN_PROBE, the same words whether or not
one is the vault's TIN; only one or two candidates are compared
against the vault. Tests: unit
text_probing_for_a_tin_is_refused_alike_whether_or_not_it_holds_it,
one_or_two_candidates_are_compared_against_the_vault; over HTTP
text_probing_for_a_tin_is_refused_alike_by_every_filing_tool.
Plan: plan/served.md P0, section 6 "Immediate hole (P0)", 4.1, 4.3, 4.9.
Review round 3. S1: a panic stops the server for every session, so no
input may reach one. Dates are four-digit years (new ledger_core::
calendar, 0001-01-01..9999-12-31): the MCP's parse_date is strict
YYYY-MM-DD, periods are bounded, last_day_of_month returns a refusal
(server and store); schedules, assets, recurring entries, periods,
documents, disposals and entries refuse dates outside the calendar, and
replay refuses such journal dates. Every amount a command takes is at
most a line's (draft lines by size, document lines, payouts, payroll,
statements, shares, tax basis, proceeds, recurring templates), on
replay too, so sums are exact; the 1099 vendor part and the K-1
property share use checked products (TooLargeToCompute); a recurring
run makes at most 1,000 drafts. tests/hostile_args.rs drives every tool
with schema-derived hostile arguments and every resource with generated
URIs, asserting the server never stops; the audit's panics are pinned.
S2: a `.` or `..` path name is refused as NO_SUCH_FILE up front. S3:
NUL-interleaved UTF-16 without a byte-order mark is scanned both ways.
S4: resource-read refusals are TIN-redacted. S5: set_w9 records its W-9
only once the call passes every check, just before the vault write.
S8: idempotentHint only on verify, run_depreciation, run_schedules,
run_recurring, set_party_roles (a repeat leaves journal and vault
alone; spec_vault checks every candidate twice); every tool's hints
pinned in one table. C1: books.settings lists receipt_required_accounts
and receipt_exempt_accounts uncut. C2: the default company name reads
"a new company". C3: the parties rule names no-party reasons. C4/C5:
the handbook names the settings file only as books.settings' books_toml,
and restores "check books.digest and offer it". C6: an import root
inside the books, data dir or vault is refused at startup. The lint
catches `just` however it ends and where the books are;
digest_resource_bare_reads reads the server's month.
Review round 4 (2026-10-17): round 3's bounds ran in replay too, so
journals main wrote and reopened (an entry dated 0000-06-01 or
+10000-01-01, a draft line of 2e15, an asset ending past 9999, a basis
or share past the line limit, sub-cent document and payroll lines) no
longer opened. Posted is forever: the bounds are the commands' only.
- Replay applies main's rule exactly: parse_date takes any date chrono
reads; replay_add_line(_fx) has no size bound (add_line/add_line_fx
do); replay_register_asset/correct_asset (chrono overflow only),
replay_set_asset_tax_basis (check_journaled), replay_set_entry_parties
(check_shares_journaled), replay_register_recurring
(validate_recurring_journaled), replay_register_schedule (months 1 to
600), replay_create_period (order only), replay_dispose_asset, and no
line, date, payment or payroll range checks in replay_invoice/bill,
their payments and replay_payroll_run.
- Derived figures over such books refuse (TooLargeToCompute,
DateOutOfRange), never panic: amortization plans, depreciation into
chrono's last month (and the true-up for lives past 9999), intervals,
document totals and aging, net pay and payroll plans, money flows and
share resolution, the asset register, receipt-policy debits; the MCP
tools and the viewer show what can't be computed.
- tests/fixtures/old-books: the review's 13 journals, its control and 15
more edges, written by main's Store API with main's figures
(expected.json); every one opens in the writer and the viewer to those
figures, every derived figure computes or refuses, every MCP tool and
GUI page answers; hostile_args restarts the server after each
property; each bound has a command/replay test pair.
Final review: a schedule's plan refuses a total of 10^26 or more
(TooLargeToCompute) instead of letting Decimal round its cents silently,
and a null `remaining` in the schedule output is always explained by
plan_error. Replay is unaffected: no journaled schedule month can belong
to a total that large (the line limit bounds every posted month).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T3gXMVt7XJPi6ens1or9Dj
An unknown event type on a whole, chained line, or a journal.format_set
declaring a newer journal format, now raises StoreError::NewerJournal
(update wanderbooks) instead of JournalCorrupt. Corruption (bad hash,
seq, or JSON) is still corruption. The gate runs in the store, the
Snapshot and the GUI, after each line's parse, seq and hash checks, and
never skips a line it can't read. Adds the journal.format_set event
(format 1 today; nothing writes it yet) and a registry of every event
type this version reads, tested against the replay and projection arms.
Review fixes (P0 review, 2026-10-16):
- A line's own hash is checked only by the next line's prev, and the
gate used to refuse a line before reading the next: a flipped bit in a
middle line's type (or a format_set edited from 1 to 2) was reported
as "newer, the books are not damaged". Now the reader settles every
refusal against the next line (format::settle): a next line that
doesn't chain makes it corrupt; one that chains makes it newer and
verified; with no next line the message says "newer or damaged; update,
then run verify", never "not damaged" (NewerJournal.verified).
- A newer format number is judged newer before its features are parsed:
a newer format may shape them differently.
- Snapshot treats a blank complete line as corrupt, as the writer does.
- docs/spec/rules.md B-27, which the code cites, now exists.
Round-2 review fixes (2026-10-16):
- The envelope {seq, ts, type, payload, prev} and its hash (SHA-256 of
the canonical JSON) are frozen across formats (rules.md B-27); new
fields go in payload. settle reads only the next line's prev, leniently,
so a newer format's next line (a float, a string seq, keys of its own)
is never called corrupt: no readable prev means unverified newer.
- A key beside the envelope's was silently dropped (and the chain then
broke as "corrupt"): it is now NewerJournal, hashed over the raw
canonical JSON so the next line still checks it; JournalLine::parse
refuses such a line.
- "the books are not damaged" overclaimed: a verified refusal says line N
is intact and this version can't check the lines after it.
- The writer read the journal with BufRead::lines (invalid UTF-8 was an
I/O error); it now splits bytes on newlines as the Snapshot does, so a
non-UTF-8 line is JournalCorrupt in both.
- Store::open truncated a torn tail before the gate ran, altering newer
books (perhaps mid-append) before refusing them. The gate now runs over
the complete lines first; the tail is set aside (and journaled) only
when they pass.
- Mutation testing (docs/testing.md, P0): the viewer and the writer say
the same words about the same damaged line (the newline is no part of
it); read_journal_and_format returns every line and the format in
force; the journal page tells a format declaration.
Plan: plan/served.md P0 and 5.1.
Review round 3 (S6, S7). Store::open replays every complete line in
memory (the first line's currency, then the ledger) before a cut-off
last line is set aside, so a complete line the books can't apply
refuses the open with the journal byte for byte as it was and nothing
set aside (B-27 SCENARIO 3; test
a_bad_complete_line_before_a_torn_tail_leaves_the_journal_as_it_was);
catch_up now only projects. The viewer
applies the writer's first-line rule (created_currency, shared) instead
of guessing USD: any other first line, or a store.created without a
sound base_currency, is corrupt to both (test
the_first_line_is_store_created_for_the_viewer_and_the_writer). fuzz_store_replay varies the
first line (14,696 runs clean).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T3gXMVt7XJPi6ens1or9Dj
mise.toml: pin cargo-mutants 27.1.0 and cargo-fuzz 0.13.2 (the latest
published, and what is installed) instead of "latest", so mutation and
fuzz runs repeat (plan/served.md I-13, P0).
docs/mcp-interface.md: the spec had two Transports sections, and the
first claimed bearer-token auth on Streamable HTTP. There is none: the
server uses rmcp's default config, whose only guard is a Host check for
localhost, 127.0.0.1 and ::1, with no token and no Origin check. One
section now says so plainly and points to plan/served.md P2 for auth.
ledger-cli AGENTS.md template: "the journal is the evidence - read it,
don't guess" led an agent to cat the journal file instead of asking the
books. The template now says to trace a figure through books.get_entry,
books.search_entries, or books.export (journal or general_ledger), and a
new bullet says the books are reached only through the wanderbooks MCP
tools: never open, list or parse the books' files, not even to look; if
a tool can't answer, say so. The other lines that sent the agent to the
data dir go too: onboarding is triggered by an empty books.chart instead
of a missing journal.jsonl, the CPA paragraph drops the ledger-gui
--data-dir books-data command, and the backup note no longer names the
journal's path. Test agents_md_reaches_the_books_only_through_tools
pins this; agents_md_carries_the_contract no longer expects journal.jsonl.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T3gXMVt7XJPi6ens1or9Dj
Records where the Served work stopped on 2026-10-16 (owner's usage
credits): which P0 pieces are done, which are work in progress in which
jj workspace, how to resume, and where the before-snapshots of the first
company's reports live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T3gXMVt7XJPi6ens1or9Dj
The 2026-10-16 audit of the first company's books found an agent reading
the journal file directly, the instructions truncated by clients, and 20
engine defects. plan/served.md is the plan to fix all of it: the books
become a server that owns its data, every agent reaches them only
through the MCP from any directory or client, settings and conventions
move into the journal, and the scaffold goes away. Phases P0-P10, ending
with a full re-audit through the MCP alone.
Chosen from three independent designs by a three-judge panel, then
revised for the owner's decisions: books kept as the user in a private
data root (advisory isolation, labelled so; enforcement is the self-host
deployment), the viewer reads the data directly, no added permission
prompts, self-hosting now and multi-tenant hosting later.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T3gXMVt7XJPi6ens1or9Dj
Every crumb above the page you're on links to where it names: the
section goes to its home (Parties › Stephen Waits → Parties), and a
reconciliation's trail runs Accounts › its account › it. Same look; a
pointer and "Go to …" on hover. A crumb naming the current page isn't a
link.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
The party page's "TIN on file" and the W-9 card's TIN copy the full,
formatted number on click (right-click: digits only). The viewer reads
it from the vault only at that moment, with the server's key file in
force, never creates a key, refuses a key inside the books dir, and
never draws more than the last four. A locked vault copies nothing and
says why.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Regulators and registries that don't tax the company (a division of
corporations, a licensing board), distinct from tax_agency. Core Role,
MCP descriptions, viewer label/noun/icon, docs, and tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
An independent recompute of the real books found no numeric
disagreements; review found 15 defects, all fixed with spec-first
regression tests:
- k1: a prepaid credited straight to capital is property; a member
paying a vendor through 2100 is cash. Docs and B-13 match.
- parties: a reversal of a reversal counts on its root's side;
liability-to-liability payments (member pays the card) count as paid.
- viewer: party money uses the server's cash accounts and full calendar
year; per-party "What for" scales multi-party entries.
- reconciliations: refuse a withdrawn statement before journaling; new
books.void_statement; reconciled lines can't be un-cleared; one shared
reconcilable-accounts rule for digest and viewer.
- CPA package includes receipts linked to entries; a file can't be both
filed on and linked to one entry.
- parties = "require" covers payroll, tax remittance (agency) and
disposal (buyer); no-party marks follow reversal chains and clear when
a vendor is tagged.
- derived member shares are frozen to exact amounts when posted, so a
settings change never re-splits history.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Entries may name several parties, each with an exact amount, a percent,
or a share derived from the members' capital lines or [[company.members]]
shares (entry.parties_set); shares always sum exactly. The 1099
worksheet counts a vendor's share. A member's party page is their
partner capital statement, from the same ledger_core::k1 source as the
CPA package.
books.set_tin stores any party's TIN (members' SSNs for K-1s) in the
vault (party.tin_set, last four in the books). The CPA package carries a
K-1 member worksheet (k1-members.csv, with a members' total that ties to
Schedule L) and the 1099 worksheet with full TINs, saved outside the
books only. files_jointly_with notes spouses' two K-1s land on one 1040.
Adversarial review fixes: reversed contributions no longer inflate K-1
contributions and distributions; a K-1 row never takes another party's
TIN; TIN detection covers dot/slash/dash separators; refusals redact
TINs; vendor tagging can't silently replace shared parties; reversals
follow re-apportioning.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
ledger_core::parties::entry_flow: cash lines count; liability lines count
except the part paired against cash in the same entry, which counts once
as paid to the liability's creditor (or received from a lender). A
reimbursement to the members no longer nets to zero. What-for breakdown
comes from the same function and sums to the net exactly.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
MCP: books.link_attachment, unlink_attachment, attachments; link_to on
attach_receipt and file_document; record_statement, statements;
books.reconcile previews and, confirmed against a recorded statement,
records the reconciliation; void_reconciliation, reconciliations;
import_bank_csv links its source file; digest lists unreconciled
accounts.
TIN vault: books.set_w9 takes the TIN and the form into the encrypted
vault (last four in the books); books.vendor_1099 include_tins and
save_csv_to for the CPA's copy; books.get_w9; books.purge_tin. A leak
test scans every tool output, the journal, the database, and the blobs.
[books] parties = off | suggest | require (default suggest): every
preview names the entry's party or suggests likely ones; require refuses
to post until it names one or is marked as having none (entry.no_party).
Viewer: documents on every subject page; statements and reconciliation
history per account; reconciliation pages; reconcile and
no-counterparty attention items; party money by period (paid out,
received in, net, what for, by month); Needs attention gets its own
full-width section.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Parties: one directory for vendors, contractors, customers, employees,
members, platforms, agencies, banks, landlords (ledger-core parties.rs).
W-9/W-4 facts without the number (last four only; any nine-digit run in a
party field is refused, Treas. Reg. 301.6109-4). Vendor and customer
records link to parties on replay; core tags every bill, invoice, payment,
and reversal with its counterparty. Journal events party.*,
entry.party_set; payroll lines name the employee party.
MCP: books.add_party, update_party, set_party_roles, set_w9, set_w4,
parties, party, set_entry_party; draft_entry party; run_payroll employee;
show party. Viewer: Parties directory and party pages, search, journal
text, W-9 attention, demo data.
One definition of a party's money (ledger_core::parties::activity) for
MCP and viewer. 1099 amounts count payments from cash and financing
accounts (a member paying for the company counts; a card doesn't; a
prepayment counts when made): vendor_1099_report_with, B-19.
Store: attachment links to any subject, statements, reconciliations
(schema v11), with replay and projection. ledger-vault: encrypted TIN and
W-9 records (ChaCha20-Poly1305, key file outside the books).
Spec-blind tests for parties (core, store, MCP, viewer) and attachments.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
- core: TaxBasisRule (cost IRC 1012, carryover IRC 723, personal-use conversion Treas. Reg. 1.168(i)-4(b)), TaxBasis with the contributor's adjusted basis, Asset::tax_basis and ::built_in_gain (IRC 704(c)); the register gains tax rule, contributor basis, tax basis, built-in gain, total tax basis, missing count (rules.md B-25; sources vendored in docs/spec/sources/irs-partnership-basis.md)
- store: asset.tax_basis_set journaled, replayed, projected (schema v10)
- mcp: books.set_asset_tax_basis; register_asset takes tax_basis_rule and contributor_basis, checked before anything registers (found by the spec-blind suite); list_assets and the register export (json, csv) carry tax basis; instructions say to record it for every asset
- import: the register's XLSX and PDF gain the federal tax basis table with citations
- viewer: tax basis on the Assets page, the asset page (a Tax basis card), the report, the journal, and Needs attention for assets without one
- just: check and the mutation recipes sweep this user's test book dirs older than an hour (killed runs had left 23 GB in RAM-backed /tmp)
- docs: the parties, attachments, TIN vault, completeness engine, and updates designs (docs/parties-and-attachments.md)
- tests: spec-blind tax basis suites in core, store, MCP; register PDF/XLSX and viewer tests
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
The acquisition link was added after the viewer's asset work and showed its raw event name in recent activity. The new test reads the event types from the store's replay arms, so the next new event without text fails it (checked: it fails without this fix).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
- core: asset tags (generated PREFIX-0001 or the human's own; unique ignoring case; never reused by another asset) and serial numbers (rules.md B-24); a journaled link from an asset to the entry that acquired it (posted, unreversed, debiting the cost account); fixed_asset_register(as_of), reconciled to the GL (unposted depreciation, unregistered cost)
- store: asset.tagged, asset.serial_set, asset.acquisition_linked journaled, replayed, projected (schema v9); differs_from now also compares control roles and tag history
- config: [books] asset_tag_prefix, default the company's initials without legal-form words
- import: Avery 5160/8160 label sheets with QR codes (geometry vendored from Avery, docs/spec/sources/avery-5160.md); the register as XLSX and PDF
- mcp: register_asset tags (and links) automatically; books.label_asset, books.asset_labels; asset tools take a tag; books.show asset / assets page / register report; export fixed_asset_register (json, csv, xlsx, pdf) and in the CPA package
- viewer: Assets page (register, reconciliation), asset page (identity, schedule posted vs not, entries, documents), Fixed asset register report, global search by tag/serial/description, journal text, needs-attention items
- fix: verify compared snapshot amounts as strings; a zero total summed in hash-map order could print "0" or "0.00", so agreeing books failed verify at random. Now by value (regression test)
- tests: spec-blind suites for tags, register, prefix, store, MCP; label and register PDF/XLSX tests; viewer tests
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
- core: ControlAccount keys, Ledger::assign_role/check_assign_role/role_account (rules.md B-23): one account per role, of the role's type, balance sheet roles move only when empty; invoices, bills, payroll, remittance, and closes post to the assigned account; an assigned role frees its default code for renumbering
- store: role.assigned journaled and replayed (strict account_id)
- mcp: settings.rs reloads by content at the start of each request, keeps the last good settings on a broken edit; books.settings with chart problems; digest settings_problems; instructions say when to edit the toml
- viewer: journal text for role.assigned
- cli: full example wanderbooks.toml, tested to load with every example uncommented
- tests: spec_roles.rs (25, spec-blind), store roles.rs (6), MCP reload/problems/roles, settings unit tests
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
A payroll tax remittance was journaled as a plain entry, so the viewer
labeled it "Journal entry" and, at a 0.00 threshold, flagged it as
missing a receipt (the server posts it without one, like a payroll
run). The store now names it with payroll.taxes_remitted, as it names
runs with payroll.run; replay and the projection pass over it, and the
viewer's journal page describes it.
tests/view_model.rs: the cash history is never empty when cash moved
and ends on the last day it did; the prior month runs to the same day;
burn averages over only the months the books cover, half away from
zero; cash accounts found by name when the settings name none; card
accounts; each entry's accounts debits first, once each; only what
arrived since the last refresh is fresh; integrity counts files again
without rereading them; the invoice's customer, the bill payment's
vendor, payroll and remittance origins.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
profit_and_loss_by_class re-added every posted entry in the range,
closing entries included: after closing January, its by-class P&L for
January showed an unclassified row of the whole period's net income,
negated. Closing entries are left out, as profit_and_loss_for_period
leaves them out.
tests/oracle_mcp.rs (24 tests): every export (CSV, JSON, XLSX, PDF,
IIF) against ledger-core's reports on the same books; the CPA package's
contents; P&L by class; the 1099 worksheet; disposals; depreciation;
payroll remittance; the digest against the journal; books.show's viewer
commands; resources; receipts and documents round-tripping; the 25 MiB
upload limit; bank import sides; reference resolution; partial FX lines;
preview balances. Written against the mutation run's 126 survivors.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
year_earnings counted every closing entry dated this year against this
year's income, but a close takes all unclosed income through its end,
earlier years included. A partnership closing January after an unclosed
December printed "Net income (loss)" -969 for a year earning +730, and
the closed December as "Undistributed earnings, prior years". The
unclosed part is now the income after the last close, from no earlier
than the fiscal year's start (closes are cumulative).
Unit tests: the 25 MiB upload limit, YYYY-MM periods, config_mentions
for every setting naming a code, report members, year earnings before
and after a close, and which entries may lose their last receipt.
tests/startup.rs: --help, the default data dir under HOME, settings
read from beside the books, and a held lock refused after the wait
instead of waited on forever.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Kpis gets a spec (cash, liabilities, receivables, month-to-date against
the same days of last month, fiscal year to date, burn). tests/view_model.rs
computes each from ledger-core on random books (month ends, leap days,
any fiscal start) along with the period ranges, cash history, account
labels, journal count, receipt rule, origins and parties, the blob
root, and integrity counts. Burn rounds half away from zero, like the
rest of wanderbooks (it was banker's rounding).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Company details (EIN, state number, address, every profile value), an
entry's number and amount, an account's code and balance, a document's
file name and SHA-256 copy on click: a pointer and copy icon on hover,
a check for a moment after. Amounts copy as their exact plain decimal
(fmt::plain: no "$", commas, or parentheses; "-" for negatives); where a
click already drills down (report cells, register and transaction rows,
general ledger lines), a right-click menu copies the amounts, the entry
number, or the description. The sidebar's chain check copies the
journal head and data folder.
tests/copy.rs: every display style copies back to the exact value
(property), and clicking the EIN, an entry's amount, and its number puts
exactly that on the clipboard.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
just mutants* points TMPDIR at ~/.cache for cargo-mutants' tree copies;
the tests inherited it and synced every journal append to disk, several
times slower per mutant. Cargo's [env] reaches the tests it runs, not
cargo-mutants, so the copies stay in the cache.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Read in the ASC basic view and vendored as short quotes with paragraph
numbers (docs/spec/sources/fasb-asc.md). Depreciation: any systematic and
rational allocation of cost less salvage (360-10-35-4); prepaids are
current assets with no prescribed expensing method (340-10-05-4,
210-10-45-1/2); cash flows gross and classified three ways, owners'
advances financing (230-10-45-7/10/14(b)/15(b)); a principal's revenue
is gross (606-10-55-37B). Citation fixed: 230-10-45-14(b), not (2).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
tests/read_only.rs: with every file and folder of the books read-only on
disk, the viewer loads, follows live, and renders every page (receipts
included), and every file keeps its bytes, size, and mtime; on writable
books nothing appears beside the database either.
Found: with no writer running, the snapshot's read-only open of the WAL
database had to create -wal/-shm (a write into the books) and failed on
read-only books, showing a false "database unreadable". With no -wal
file no writer has it open, so check_disk now reads it immutable.
just mutants-gui sweeps only what the viewer must get right: the view
model's numbers (books.rs) and their formatting (fmt.rs); 179 mutants.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
cargo-mutants copies and builds the tree per job in TMPDIR; on Arch /tmp
is tmpfs (RAM, up to half of it), and a GUI run put 20 GB there. Every
recipe now uses $XDG_CACHE_HOME/wanderbooks/mutants (~/.cache/...).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
tests/http_concurrency.rs: six HTTP clients at once against one store
(main.rs serve_http shares it): every bare entry refused at threshold
0.00, every receipted one posted, each contested draft posted exactly
once, books verify and balance, a restart replays the same head.
spec_mcp.rs: the response timeout is 180 s (a mutation run's jobs
contend for a synced disk; the baseline timed out at 30 s), and
MCP_CASES sets the state machine's cases (just mutants-mcp uses 16).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
tests/spec_gui.rs (31 spec-blind tests): exact money text round-trips,
file names never escape the folder or hit a Windows device name, history
against a browser model, fuzzy search against a subsequence model, axis
ticks, the view model against ledger-core, and the attention list.
Found and fixed:
- fuzzy matched greedily, so "ac c" scored 1 for "c" (the "c" inside
"ac") instead of 5 for the word "c"; it now scores the best alignment.
- ticks accumulated float error (a last tick at -5.55e-17 for a top of
0), gave more than n + 2 ticks at large magnitudes with tiny spans,
and overflowed for spans near f64's limit. Ticks are now whole
multiples of the step, checked, with [lo, hi] when f64 has no nice step.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
The ranks of n parts are a permutation of 0..n-1 (being ahead is a strict
total order; pigeonhole by hand, core Lean has no subperm), so exactly
u parts get a unit and one gets the fraction. apportion_sum and
apportion_sum_of_shortfall; the audit shows only the standard axioms.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
ledger-config: parse gets a sourced spec. A fiscal year starts on the
first of a month (IRS Pub. 538); tax_classification is read by whole
words (Treas. Reg. 301.7701-3, Schedule C instructions; "business corp"
no longer reads as an S corp, "corporation or S corporation" is refused);
duplicate or blank members and shared capital accounts are refused; a
negative or non-grammatical receipt threshold is refused; codes trimmed.
tests/spec_config.rs (65 spec-blind tests, 14 properties), fuzz target
fuzz_config_parse (weekly). Mutation: 55 of 55 viable caught (from 25/38).
The receipt policy moves to ledger-config, with one loader
(effective_receipt_policy) and one decimal grammar, so the server and
the viewer enforce and show the same rule (the viewer had its own: only
expense-like debits, a silent $25 default, and no config.json).
ledger-mcp: a module spec (MCP 2025-06-18 tools, JSON-RPC 2.0, the
operating contract). Amounts in are exactly -?digits[.digits]; a number
that is one account's code and another's id (vendors, classes alike) is
refused, not guessed; the threshold is validated at startup. Fixed:
add_line with a bad class journaled the line before refusing;
profit_and_loss_by_class ignored `class`; aging buckets said "0". A wire
in front of rmcp answers malformed JSON with -32700 and invalid requests
with -32600 and their id (rmcp dropped the first, omitted the id).
tests/spec_mcp.rs (26 spec-blind tests: every tool's result against its
own outputSchema, a random-call state machine checking the journal is
untouched by every refusal, amount grammar, previews, the receipt policy).
ledger-gui: money, accounting, and dollars never round (they rounded to
cents, banker's style); export file names avoid Windows device names,
trailing dots and spaces, empty names, and 255-byte overflow; specs on
attention, history, fuzzy search, and axis ticks.
ledger-core: the lifetime profit_and_loss is renamed unclosed_earnings
(docs/testing.md #3).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Ledger::renumber_account changes only the code (entries name accounts by
id): refuses an unknown account, a blank or used code, and a control code
for an account of the wrong type or a placeholder. Journaled as
account.renumbered (with the previous code), replayed, projected, shown
in the viewer's journal. The MCP tool lists wanderbooks.toml settings
that still name the old code. The server instructions now warn that the
engine can't tell a same-type account with another purpose at a control
code (a hosting expense at payroll's 6100).
Tests: core property (only the code changes; trial balance equal), core
refusals, the store's command property, MCP end to end with restart.
Mutation on the new code: 8 of 8 viable caught.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
report_to_xlsx gains INVARIANTS 4-6 (header block, right-aligned amount
headings, spacer and account spacing, footer one empty row down, rules
exactly where rule_above says, exact text grouped in threes), tested by
reading the workbook's styles, merges and column widths. The general
ledger's footer now sits one empty row down like the other sheets.
Also killed: IIF transliteration arms, tax summary's optional sections,
and day-equals-month bank dates. `just mutants-import`, weekly in CI.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
tests/spec_import.rs, written from the import crate's specs and sources
alone: every statement foots as printed on random books (P&L, balance
sheet with and without members, cash flow, trial balance, members'
equity), format_amount round-trips, bank CSV round-trips in both layouts
with date-order scenarios, IIF transactions sum to zero with every
account declared and ASCII fields, XLSX cells read back through a small
zip/DEFLATE reader, PDF labels with Typst markup render as text and the
bytes are deterministic, chart CSV and templates.
Its two findings, both fixed:
- "1/15/69" parsed as 2069: chrono's %y cuts over at 70, POSIX strptime
at 69 (69-99 are 1900s). Two-digit years are computed by POSIX now.
- An amount finer than a cent went into an XLSX as rounded text ("1.01"
for 1.005); it is written as its exact text now.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
- Starter charts put "Platform and Payment Processing Fees" at 6100, the
code payroll posts wages to, so a template company that ran payroll
would book its wages as processing fees; their own payroll accounts
(7000, 7050, 2300) were never used. Now 6100 Salary & Wage Expense and
6110 Payroll Tax Expense (the engine's control accounts, with tax
categories), fees at 6060, and the unused accounts gone (the
withholding payables 2110-2150 are created when payroll first runs).
New tests: every template imports with categories valid for its
return; no template puts another account at a control code; a
partnership names every partner and its shares sum to 100%.
- format_amount printed a negative zero ("(0.00)" for -0.004); it rounds
half away from zero first now, like amount_parts.
- Specs (GOAL/SOURCE/INVARIANTS) for statement finishing, P&L, balance
sheet, cash flow, trial balance, members' equity (each must foot as
printed), format_amount, and the templates.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
From the IRS form instructions (now quoted in docs/spec/sources/
irs-tax-lines.md):
- On Forms 1065 and 1120-S, Schedule K takes "only taxable portfolio
interest", and charitable contributions are among the items the entity
"must state separately". The tax summary counted both in ordinary
income and deductions; they now go in TaxSummary::separately_stated,
outside book_net and tax_net, and the report, MCP export, and viewer
list them under "Separately stated to owners (Schedule K)".
- Schedule C: "Do not include charitable contributions". The category no
longer applies to Schedule C (a sole proprietor's gift is personal, an
owner draw).
- interest_income's note: interest earned in the business itself is
ordinary income (other_income).
XLSX: a cell number is a binary double and Excel honors 15 significant
digits (Microsoft, "Excel specifications and limits"), so an amount over
$9,999,999,999,999.99 or finer than a cent is written as its exact text,
right-aligned, instead of a number Excel would show wrong. The old
conversion also fell back to 0.0 silently; gone. Property test: every
whole-cent amount up to 15 digits round-trips exactly.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
New vendored sources (docs/spec/sources/): Intuit's bank CSV upload and
IIF formats, ECMA-376 / Excel number storage, and the IRS form lines the
tax categories map to.
IIF export (rewritten to Intuit's IIF Import Kit):
- declares every account first (!ACCNT NAME, ACCNTTYPE, ACCNUM): an
undeclared account is created in QuickBooks "as a new bank account";
the MCP export types them (cash accounts BANK, the control accounts AR
and AP, registered fixed assets FIXASSET, the rest by type);
- subaccounts by colon path, parents first; colons in a name replaced;
- ASCII only (IIF is ASCII text), dashes and quotes transliterated;
- the kit's 2007+ journal layout and M/D/YYYY dates; an entry using an
undeclared account is refused.
Bank CSV:
- the 4-column layout is money out then money in by position, which
matches Intuit's "Credit, Debit" and banks' "Debit, Credit" both (the
doc's "Expense, Income" labels were wrong, the parsing was not);
- a slash date's order comes from the file (a part over 12 settles it)
or the caller's new date_order, never a guess: Intuit recommends
dd/mm/yyyy, US banks write mm/dd/yyyy;
- fractions of a cent and double negatives ("(-5)") are refused; the
header is recognized on the first non-blank row.
Chart CSV: a blank code or name is refused with its row (found by the
new fuzz_import_csv target), and the core now refuses a blank account
code or name too (replay keeps older books).
MCP:
- books.import_chart_csv and books.import_bank_csv rehearse the whole
file on a copy of the ledger first: a bad row anywhere refuses the
file (before, a duplicate within the file, a parent after its child,
or a date in a closed period left earlier rows behind);
- bank rules: the longest matching pattern wins (they arrive in a map,
so "first match" changed from run to run).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Final store mutation run: 357 caught + 1 timeout of 365 viable (from 82%
this morning). Three of its seven survivors now have tests:
- replay refuses a re-chained journal whose schedule posting differs
from what it re-derives (check_replayed_entry guards that event);
- a disposal refused for an account's type says the type it needed;
- a database whose last pages are damaged while its settings still read
fails the integrity check and is rebuilt (verified to fail with the
check disabled: "database disk image is malformed").
The four left are equivalent: sort_value's two arms, OpenFlags | vs ^,
and writing the long form of a sweepless retained-earnings close.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Store mutation rerun: 345 caught of 373 viable (92.5%, from 82%). The
survivors, triaged:
- Removed code the survivors showed could never matter: set_line_class's
hand-written checks (its rehearsal refuses first), ensure_on (the
engine creates missing control accounts itself), and verify's
live-vs-replay halves of the asset, class, and payroll comparisons
(the whole-ledger comparison above them already checks that; the
SQLite read-back halves stay).
- New tests: disposal account types; depreciation periods must be real
months; search_entries (filters, newest first, paging, no drafts);
the database's invoice amount paid and status; replay refusing a
re-chained journal whose closing entry differs from what it
re-derives; a projection missing a table.
- Equivalent: sort_value's arms (serde_json's maps are already sorted
without preserve_order, which nothing enables; the function stays as
the guard if anything ever does), OpenFlags | vs ^ (disjoint bits), and
writing the long form of a sweepless retained-earnings close (replay
reads both).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Store::open refused books whose database recorded a base currency other
than the journal's, with JournalCorrupt, though the journal was fine and
the database is derived (a mutation-testing survivor: nothing tested
the case). Now the currency comes from the journal, and a database that
fails its integrity check or disagrees on currency goes through the same
path: kept aside for forensics and rebuilt (open SCENARIO 6).
Also tests: the viewer tells a database Behind the journal from one
Ahead of what it has read; the table check counts missing and
unexpected rows; a new journal creates its directory; a stored blob
exists and an unknown one doesn't.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
From the first store mutation run (324 caught of 396 viable):
- receive_payment and pay_bill repeated the engine's refusals by hand
after rehearsing the engine itself, so those checks could never fire
(their mutants survived). Removed: the rehearsal is the check.
- New hand-computed tests: a disposal's final catch-up after some
depreciation was posted (200 of 300, gain 100); a disposal on the
purchase date (allowed) or the day before (refused); a retained
earnings close with a sweep replays it; a float inside an array is
refused; the viewer's journal and head hash.
- The concurrency test's reader stops when the writer thread ends, not
only when it says it's done, so a mutant that kills the writer fails
the test at once instead of spinning to the five-minute timeout.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
`just mutants-store` (64 random cases per mutant, temp books on a RAM
disk in CI) runs beside the core's weekly; fuzz_store_replay joins the
weekly fuzz matrix. docs/testing.md lists the spec-blind store suite,
the refusal property's converse law, and the replay fuzz oracles.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
The refusal property gained the converse law (an accepted command
journals something; only batch runs may find nothing to do) and checks
the store's class and payroll listings. It found a schedule registered
with an out-of-range total whose first month was refused after its entry
was begun: the live ledger kept a draft the journal never had.
- Core: compound commands (post_schedule_month, reverse_entry,
close_period_to, create_invoice, create_bill, receive_payment,
pay_bill, run_payroll, remit_payroll_taxes) run inside
Ledger::atomically, so a refusal partway leaves the ledger exactly as
it was.
- Core: a schedule total no line could carry, or finer than a cent, and
an asset cost beyond the line range, are refused at registration.
Replay keeps older schedules (replay_register_schedule).
- Store: add_line, add_line_fx, mark_cleared, set_fx_rate,
post_schedule_month and close_period_to changed the live ledger before
journaling; they now run on a copy, journal what it produced, then
swap it in, so a failed write leaves the live books as journaled.
New spec-blind suite tests/spec_store.rs (57 tests written from the
store's specs alone). Its two findings, adjudicated:
- A file filed twice under different names listed the first name and
type for both (the projection kept them per blob). Schema v8 gives
each attachment its own; older databases backfill and heal by verify.
- Withdrawn attachments stay in the store's listing, marked (the MCP
listing hides them): the spec and doc said "omit" and are corrected.
Also removes journal::truncate_to_last_good_seq: unused, and it dropped
everything after the first bad complete line, which the books now treat
as corruption for a person, never repaired automatically.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
benches/store.rs (criterion, in `just bench`): open 249 ms, viewer load
616 ms, full verify 809 ms, post one entry (four synced appends) 62 ms.
The books are built on a temp dir and copied under target/ so every
sync hits a real disk. Baselines in docs/testing.md; no budget, by
decision.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Each close into members' capital rounded on its own, ties to the
earlier member, so twelve monthly 50/50 closes of odd-cent income could
leave one member several cents ahead of half the year (docs/testing.md
#12, found with the members' equity fix).
- CloseTo::Members now splits cumulatively over a run (closes back to
back with the same members and shares): each close gives a member
allocate(run total so far) - allocate(run total before). Periods
record their close order and a cumulative close's members and net
(Period::close_seq, member_close) so the next close continues the run
and replay recomputes it exactly. A close to anything else, or other
shares, ends the run.
- Journaled "rounding": "cumulative". The per-close rule is kept as
CloseTo::MembersPerClose for closes journaled "largest_remainder";
older closes still replay with the last member absorbing it.
- Lean: runPart_telescopes (a member's total from a run is their
allocation of the run's total) and closePart_total (each close hands
out exactly its net), audited to the standard axioms.
- Tests: hand-computed three one-cent months (per close A 0.03 / B 0;
cumulative A 0.02 / B 0.01), a run ending, a property that each member
stays within a cent of their share (fails on the per-close rule), and
a store contract that each rule journals and replays as itself.
Wanderland's books have no closes yet, so all of theirs use the new rule.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
New concurrency test: one thread writes entries, drafts and receipts as
fast as it can while a Snapshot (the viewer) refreshes in a loop. At
every refresh the snapshot's books must balance, never go backward, and
never see the writer's database as differing from the journal; at the
end it must equal the writer in full. It failed 30 runs in 50:
check_disk read the applied position and the tables in separate
statements, so a write landing between them made a consistent database
look like it "differs", which the viewer shows as a problem in Needs
attention.
- check_disk reads in one read transaction (one consistent view).
- A database that has applied lines the viewer hasn't read yet is now
DiskCheck::Ahead ("viewer catching up"), not Behind.
- Without the transaction the test fails 14 runs in 20; with it, 50/50.
Also: members' equity spec 3b states the exact bound (a member's period
income is within two cents of their share, one cent of rounding at each
end; the running balance chains exactly), and the spec-blind property
checks that bound. docs/testing.md gains the concurrency layer and item
12: each close into members' capital still rounds on its own (pending).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
A capital account is one running balance (Form 1065 Schedule K-1
instructions, Item L: beginning and ending capital "based on the
partnership's books and records"), but the statement split prior
earnings and the period's income separately, so a member's ending on one
statement could differ by a cent from their beginning on the next.
Now each member's share of undistributed earnings is allocated as of a
date (largest remainder, as before): beginning as of the day before the
period, ending as of its last day. A member's period income is the
roll-forward's remainder (their share within a rounding cent); incomes
still sum to net income. New property
each_members_capital_chains_across_statements (random books, odd-cent
shares, closes to members or not, any split date) fails on the old code
and passes now. The hand-computed test is recomputed: the old split gave
two cents off an even 50/50 to date.
Also: rules.md B-13 quotes Item L; docs/testing.md #7 and #8 marked
changed; the CPA-question list is replaced by the rule that wanderbooks
settles questions from the sources and makes real choices per-company
settings.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
From ASC 230 (via Deloitte DART and PwC Viewpoint), replacing two
formerly assumed behaviors:
- Financing and investing are gross (230-10-45-7): per account, cash in
and cash out are separate lines (contributions apart from
distributions, borrowing apart from repayment), netted only within
one entry. Lines come in code order, in before out.
- Investing includes making and collecting loans and buying investments
(230-10-20, 45-13), not only registered fixed assets:
CashFlowParams gains investing_accounts.
- The roles are per-company settings: [books] financing_accounts and
investing_accounts in wanderbooks.toml (beside cash_accounts). The MCP
tools (cash_flow, export, CPA package) default every list to them and
share one path; the viewer reads them too, falling back to its name
guess for debt. A due-to-owners liability listed as financing gives
PwC 6.9.3.1's constructive treatment of expenses an owner paid.
rules.md B-14 quotes the sources; new tests cite them; docs/testing.md
#8 is resolved.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Each formerly ASSUMED behavior is now CONFIRMED, CONVENTION, DESIGN, or
CHANGE (pending), with its source (Reg. S-X, ASC 230 via Deloitte DART,
IRS K-1 instructions, OpenStax). Two need engine work: members' equity
should allocate cumulatively so each member's beginning equals the prior
ending (#7), and the cash flow should show financing gross and classify
investing by account role (#8). Two questions for the CPA are listed.
The test doc comments that relied on them now cite the verdicts.
Core mutation run: 1,011 caught + 13 timeouts of 1,024 viable, 0 missed.
No benchmark budget, by decision.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
CI's run of the store refusal property found a payroll run that posted
$65.27794: the employer tax was passed in unrounded and nothing refused
it, though rules.md B-3 already says a ledger in one money holds no
fractions of its smallest unit. Sums of it then differed in scale
between journal replay and the SQLite read-back.
- Ledger::post_entry refuses any line finer than its currency's ISO 4217
minor unit (FinerThanMinorUnit), whatever made the line: payroll,
closes, disposals. add_line refuses it up front. Older journals replay
as booked through replay_post_entry. Spec tests for the kernel's
ten-place limit now run on the replay path.
- The property's body is now run_ops, so failing sequences become named
regression tests (regression_payroll_with_a_fraction_of_a_cent). The
generator rounds payroll tax to the cent, as callers must.
- Ledger derives PartialEq and gains differs_from; Store::verify compares
the live ledger with the journal replay in full (drafts, documents,
schedules, periods, settings), not just balances, and names the parts
that differ. The property checks the ledger its commands built against
that replay and against the reopened books.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
A crash or power loss mid-append can leave a partial last line. The
writer acknowledges an append only after its whole line is synced, so
that line was never acknowledged: opening now saves those bytes to
journal.jsonl.torn.<time>, truncates to the last newline, and journals
journal.tail_repaired (replay and projection ignore it; the GUI shows
it; books.digest reports it in journal_repairs). Damage before the last
newline is still corruption and still refuses to open. Tested at every
byte of the last append.
The MCP server now waits up to ten seconds for the writer lock: a
restarted server (a client reconnecting) raced the exiting one and died
with Locked, which also made mcp_schedules_and_recurring_over_stdio
flaky. The test harness now waits for each server to exit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
The refusal property now starts from real books and adds compound
commands with valid arguments (book an entry, register an asset or
schedule, raise an invoice or bill, close a month to members), so every
command's success path is exercised. Compound commands checkpoint after
each step, so the law is checked against the one command refused.
STORE_CASES sets the case count (256 by default).
It found two bugs:
- run_recurring created drafts dated inside a closed period (B-12).
generate_recurring now refuses them, atomically; the store leaves
them ungenerated and books.run_recurring reports them in
skipped_closed_period for a human to decide. Replay uses
replay_generate_recurring so older journals still open. The store
also journals before mutating the ledger here, as everywhere else.
- Disposing an asset for no proceeds (scrapping it) journaled a
catch-up depreciation entry and a half-built draft with a 0.00 line,
then refused. Zero lines are now left out (Store::disposal_lines,
shared with the MCP preview), and the whole disposal is rehearsed on
a clone before anything is written.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Max-testing pass on ledger-store, fault injection (rank 1). A full disk
is simulated with no code seam: the test re-runs its own binary as a
child under bash with the file size capped (ulimit -f, SIGXFSZ
ignored), so writes past the cap fail as on a full disk. Two bugs, both
of which left the books unopenable:
1. SQLite side: applying one journal line runs several statements (rows,
events mirror, last_applied_seq) in autocommit. When the disk filled
partway, the rows committed but last_applied_seq didn't, and the next
open's catch-up re-applied the line into "UNIQUE constraint failed:
entries.id". Each line now applies inside a savepoint: all or
nothing. (docs/receipt-store.md promised exactly this heal; it
wasn't atomic.)
2. Journal side: an append that failed partway left a partial line, and
any later line would land after it, mid-journal. append now cuts the
file back to its length before the write on any error.
Tests (tests/faults.rs, its own binary because the child re-runs it):
a_full_disk_mid_command_leaves_the_books_whole (the database fills
first) and a_full_disk_mid_append_leaves_only_whole_journal_lines (the
journal writer alone). Both shown to fail without their fix.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Max-testing pass on ledger-store (security property, rank 2). Blob ids
come from the journal, which is untrusted input on disk. BlobStore built
paths from any string: an id under four bytes (or with multi-byte
characters) panicked on slicing, and exists() would report on any path
an id like "../../x" reached. get() never served such bytes (it serves
only content whose SHA-256 equals the id), but it read them.
Now only a 64-hex-digit id names a blob; anything else is an error from
get() and false from exists(), touching nothing. Test
a_blob_id_never_reaches_outside_the_blob_store (empty, short, unicode,
traversal, absolute, non-hex, 63/65 digits; uppercase hex still works).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Max-testing pass, step 8. The final mutation run left 112 survivors
(89% caught). Three agents triaged them by file, by the playbook's
table: spec the boundary then test it, a law rather than one more
pinned value, assert the real value, never weaken a test. Every expected
value is hand-computed.
- reports.rs (35): 28 killed, chiefly cash flow (inclusive start date,
each line's sign and label, a loss on disposal moving to investing,
financing designation), general ledger and members' equity date
boundaries, and the balanced() checks (a real report nudged by a cent
must stop balancing). 7 equivalent: four >/>= after an early return on
zero; two in dead code (removed in this commit); one sign of a zero
cleaned on output.
- ledger.rs, kernel.rs, journal.rs, money.rs (52, plus one new): 49
killed, including every replay_* mutant at the core level (ids,
accumulation, status, vendor tags, out-of-order refusal), FX rates,
0%/100% deductible boundaries, a dividend sweep with no income, the
legacy member rounding at close, a bill paid on its own date, a
liability-account reconciliation, mixed-currency entries (via serde),
ISO 4217 minor units, Display/Debug. 3 equivalent: zero refused before
a </<= check; a period whose end is in a closed period refused earlier;
columns(0).
- schedules, tax, analysis, payouts, assets, form1099 (25): 23 killed
(schedule length 1-600, recurring end on its start, frequency parsing
and stepping, cost of goods in tax net, a zero payout, disposal on the
purchase date, quarter starts, labels); 2 equivalent (zero adds zero;
days_held is provably >= 1).
Two small production cleanups the triage surfaced: kernel::columns(0)
returned a negative-zero credit that would print "-0" (regression
a_zero_movement_has_no_negative_zero_column, shown to fail before); and
cash_flow computed a book value it never used.
Spec lines added where a boundary wasn't specified (cash flow's
inclusive start and line signs, members' equity start date, payout with
no deposit, disposal on the purchase date, bill paid on its own date).
`just mutants` now runs core and store tests (the store's replay tests),
not the whole workspace, which rebuilt the viewer for every mutant.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Max-testing pass on ledger-store, found during the survey: nothing
stopped two processes from opening the same books for writing. Two
agent sessions (two MCP servers) on one books dir would each extend the
hash chain from the same last line; one write from each broke the
journal ("seq gap: expected 3, found 2") and the books would not open.
The store now holds an exclusive lock on books-data/wanderbooks.lock
(std's File::try_lock, no new dependency) from before it reads the
journal until it is dropped; the OS releases it if the process dies. A
second writer gets StoreError::Locked saying another process has the
books open. The viewer reads through Snapshot and never takes the lock.
The lock file is git-ignored in the books scaffold.
Regression: a_second_writer_is_refused_while_the_first_has_the_books
(refused while open, allowed after close, books whole).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Max-testing pass, ranks 4-6 and step 9 for ledger-core.
Fuzzing (rank 4): fuzz_op_sequence now drives balanced pairs, foreign
lines, periods, closes, and amounts from raw decimal bits, and asserts
the books stay valid (TB balances, posted entries validate, no draft in
a closed period); new fuzz_core_serde (every persisted core type
round-trips) and fuzz_store_open (a corrupt journal is an error, never a
panic); hangs count (-timeout=10). 54M executions; one crash:
- fuzz_journal_line: a number too big for an integer parsed as a float,
and its canonical form (which the hash chain covers) re-canonicalized
differently. The journal never holds floats (R-2), so the reader now
refuses them (JournalLine::parse, used by every journal reader);
regression a_journal_line_with_a_float_is_refused. Checked first: the
real books' journal has none. Rerun: 13.2M executions, clean.
Seeds in fuzz/seeds/ (synthetic, minimal); the grown corpora stay out of
git. `just fuzz <target> <seconds>`.
Miri (rank 4): the core's unit, bridge, and invariant tests under Miri
(our code has no unsafe; rust_decimal and chrono do): no undefined
behavior. `just miri`. A pinned nightly (nightly-2026-10-01, with miri)
is now in mise.toml.
Benchmarks (rank 6): criterion (new dev-dependency: the playbook's Rust
benchmark tool; none existed) in benches/books.rs; baselines in
docs/testing.md. The store's per-command rehearsal clone costs ~1 ms at
10k entries. No budget yet, so nothing gates on them. `just bench`.
CI: .github/workflows/check.yml runs `just check` on every push;
scheduled.yml runs proofs, the GnuCash oracle, fuzzing, and mutation
weekly. Tools come from mise.toml.
docs/testing.md: every layer, its oracle rank, command, runtime, and
gate; benchmark baselines; the ASSUMED invariants for Stephen to reject;
the gap list with triggers. docs/oracle.md's fuzzing paragraph had
claimed persistent corpora that didn't exist; corrected.
Also: ledger-import now inherits the workspace lints (nightly cargo
warned).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Max-testing pass, rank 3. tests/model.rs transcribes the Lean ledger
state machine (verify/lean/Wanderbooks/Ledger.lean) into the plainest
Rust with the same observable behavior: lists of accounts, drafts,
posted entries, and periods. Random command sequences (add account,
draft, add line, a balanced pair of lines, post, reverse, create period,
close), weighted toward commands that interact, run against both; after
every command the engine and the model must make the same accept/refuse
decision and agree on the posted log line for line, the chart, drafts,
periods, ids, and every account's balance. 2,048 sequences of up to 60
commands; about 3s.
Honest about its first version: it agreed vacuously. Counting accepted
commands (MODEL_STATS=1) showed zero posts and zero reversals: random
lines never balance. A balanced-pair command fixed it (now ~150 posts,
~100 reversals, ~200 closes per 512 sequences). Shown to have teeth:
letting the model reverse an entry twice fails in 0.03s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Three spec decisions from Stephen (2026-10-13), raised by the spec-blind
test writers, each test-first:
- A payment dated before its invoice or bill is refused
(PaymentBeforeDocument); a real prepayment is a deposit. Test:
a_payment_never_predates_its_document. Replay is unaffected (it goes
through replay_invoice_payment / replay_bill_payment).
- Aging as of a date shows only documents issued by then, less only the
payments made by then. The ledger now records each document's payment
entries (live and on replay), and aging counts those dated by as_of.
Test: aging_as_of_a_date_sees_only_what_existed_then; the A/R and A/P
conservation properties' aging oracles updated to the confirmed rule
(they had assumed aging ignored dates).
- A schedule through a mid-month date prorates that month by days; the
code was right and its doc wrong. The spec-blind test that disagreed is
rewritten to the confirmed spec and no longer ignored.
Also found while wiring payments: a payment to an unknown cash account
created the payment's draft and then failed, leaving it behind; the cash
account is now checked before anything is created.
The unreviewed golden van_schedule.json is gone: every row was first
checked against an independent hand computation (all 61 matched), then
replaced by van_schedule_is_the_hand_computed_one, which asserts that
computation directly (548.39, 59 x 1,000.00, 451.61; total 60,000.00).
Tests: 452. No spec-disagreements left ignored.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Max-testing pass, steps 2-4.
Spec-blind tests (rank 1-2): four agents, each given only the specs and
the public interface (doc comments and signatures, every body removed,
never src/ or the existing tests), wrote tests/spec_reports.rs (62),
spec_ledger.rs (72), spec_arap.rs (42), spec_assets_payroll.rs (50):
hand-computed examples per scenario and a proptest law per invariant
(conservation, partition, invariance, scaling, reversal cancels,
"a refusal leaves the books unchanged" over random invalid commands).
They found:
1. A refused payroll run changed the books: an all-zero run passed
validation, then failed at post after creating seven payroll
accounts and using up an entry id. validate_payroll_lines now
refuses a run with no gross pay.
2. A refused add_account used up an account id (empty code, duplicate,
unknown parent). The id is now taken only when the account is added.
3. Stale prose: close_period_to and members_equity still described the
old last-member rounding.
4. Open (left #[ignore]d pending Stephen): depreciation_schedule's doc
says "through the month containing" the date; the code prorates that
month by days.
GnuCash oracle (rank 3, an oracle never the spec): `just oracle` builds
GnuCash 5.10 in a container pinned by base digest and package versions
(oracle/gnucash/) and runs tests/oracle_gnucash.rs: 200 seeded random
books built through wanderbooks (entries, foreign lines, reversals,
monthly closes), every account's balance at month ends and random dates
and each month's P&L, compared exactly: 15,800 comparisons. First run:
4,724 differences, all from foreign lines; without them, zero.
Adjudicated from principles (rules.md B-3): wanderbooks booked
foreign * rate unrounded ($108.34567); a ledger in one money holds no
fractions of its smallest unit, and its cash must match a bank. Stephen
decided: book in the base currency's minor unit, half away from zero.
add_line_fx now rounds; minor units per ISO 4217 list one (2026-09-17,
SIX), generated into Currency::minor_units and recorded in
docs/spec/sources/iso-4217.md. Replay now takes the journaled base
amount (replay_add_line_fx), so lines booked before replay exactly as
booked (contract: books_written_under_older_rules_still_open). Now:
zero differences, foreign lines included.
Tests: 449, from 223.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Max-testing pass, step 2: every domain decision in ledger.rs and
reports.rs gets a GOAL / SOURCE / INVARIANTS / SCENARIOS block in its
doc comment: post_entry, try_create_entry, close_period_to,
mark_cleared, reconciliation_report, account_balance, create_invoice,
create_bill, receive_payment, pay_bill, register_asset, dispose_asset,
run_payroll, remit_payroll_taxes, set_line_class, set_entry_vendor;
trial_balance(_as_of), profit_and_loss_for_period, balance_sheet_as_of,
general_ledger, members_equity, cash_flow, profit_and_loss_by_class,
snapshot. Sources trace to docs/spec/rules.md (and through it to
Pacioli, the IRS, FASB); what no document settles is marked CONVENTION,
DECISION, or ASSUMED for Stephen to confirm. The specs are the input to
spec-blind test writing (the next step): tests derived from the spec
alone, never from the code.
Also: rustdoc now runs in `just check` with -D warnings; fixed the seven
existing broken intra-doc links and one bare URL.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Three spec decisions from Stephen (2026-10-13), each test-first:
- A child account has its parent's type (ParentTypeMismatch): a child
subdivides its parent. Test: a_child_account_has_its_parents_type.
- Periods nest (a month in a quarter in a year) but never partially
overlap (PeriodsOverlap; a refusal uses no id). Test:
periods_nest_but_never_partially_overlap.
- A line's account must exist and take postings when the line is added,
not only at post (UnknownAccount / PlaceholderAccount from add_line
and add_line_fx); posting still re-checks. Two older tests that pinned
the post-time refusal now check both (the post-time check is reached
through the unchecked replay path, as old journals can).
Compatibility: replay reads history through new Ledger::replay_*
commands (add_account, add_line, add_line_fx, create_period) that skip
checks added later, so books written before these rules still open.
The contract test books_written_under_older_rules_still_open writes such
a journal directly and found, on its first run, that the SQLite
cross-check rebuild used the checked commands and would have failed
verification ("sqlite projection can't be read back"); it now replays
too. Documented in docs/spec/README.md; decisions recorded in rules.md
(B-1, B-4, B-12).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Stephen (2026-10-13): when what wanderbooks knows is stale, check the
date, research it or ask the owner or CPA, and update the rule for
everyone with a pull request (best) or an issue (worst).
- ledger_core::rules: a registry of rules set by law that change over
time. Each says the years it's known for, its rules.md entry, its
vendored official source, and a review date. stale_rules(today) is
pure; the engine has no clock.
- The 1099 threshold is the first: $600 through 2025, $2,000 in 2026,
then indexed for inflation (26 U.S.C. §6041(h)). default_threshold
and vendor_1099_report now return Result; for 2027+ they refuse with
LedgerError::RuleNotKnown, whose message says what to do: get the
figure from the official source or the owner/CPA, pass it explicitly,
then open a PR (or issue) at github.com/swaits/wanderbooks. Before,
2027+ silently used $2,000.
- MCP: books.vendor_1099 relays that refusal; books.digest gains
stale_rules; the server instructions tell the agent the procedure (and
how to handle control-account conflicts: explain, never renumber).
The CPA package writes a NOT-INCLUDED note instead of a guessed 1099
worksheet.
- Viewer: "Needs attention" shows each stale rule with a ready prompt
for the agent; the 1099 page explains instead of guessing.
Review date for the 1099 rule: 2026-11-01, when the IRS usually
publishes next year's inflation adjustments (CONVENTION, marked as such
in the registry).
Tests: an_unknown_year_threshold_is_never_guessed,
dated_rules_go_stale_on_their_review_date (dates passed in, so no clock
in the test), MCP stdio: 2027 refused with guidance, explicit threshold
accepted, digest carries stale_rules (asserted date-independently).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Invoices, bills, payments, payroll, and closes post to accounts the
engine finds by code (1200, 2000, 4000, 6999, 3999, 6100-2150). If an
account with that code existed it was used whatever its type: a 1200
"Customer Deposits" liability would have taken receivables. Now
(Stephen, 2026-10-13: never force an imported chart into our numbering)
the engine refuses with ControlAccountMismatch, naming the code, the
account, its type, the role, and the type needed, so the agent can work
it out with the human. A placeholder is refused too.
Every command checks all its control accounts before changing anything
(Ledger::check_controls, spec + SOURCE in its doc comment), so a refusal
never leaves a half-made account behind. One table of control accounts
(ledger_core::RECEIVABLES ... PAYROLL) replaces the codes duplicated
across the engine and the store.
Tests (spec-derived, all assert the books are untouched after the
refusal): an_invoice_refuses_a_1200_that_is_not_receivables,
an_invoice_refuses_a_placeholder_1200,
an_invoice_creates_nothing_when_its_second_account_is_wrong,
a_bill_refuses_a_2000_that_is_not_payables,
a_close_refuses_a_3999_that_is_not_equity.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Stephen asked that specs come from recognized documents, not from
anyone's memory. There is no single US bookkeeping bible; authority is
layered, and this records it:
- docs/spec/sources/: verbatim excerpts (whitespace normalized), each
with URL, retrieval date, and the SHA-256 of the fetched file, of
public-domain texts: Pacioli's 1494 treatise in Geijsbeek's 1914
translation (double entry, the trial balance, one money, write only
forward, correct by opposite entry, close into capital); IRS Pub. 583,
538, 946, 15; the 1099-MISC/NEC instructions (2025 and 2026); 26 U.S.C.
§6041, §6041A, §704; Treas. Reg. §1.263(a)-1(f) and -4(f); Notice
2015-82; Rev. Proc. 98-25; Form 1065 instructions; 17 U.S.C. §105. The
excerpts were extracted by matching each passage against the fetched
file, so none can contain text the source doesn't.
- docs/spec/rules.md: the rule catalog, B-1 to B-22. Each rule says
AUTHORITY, CONVENTION, or DECISION; cites its sources (FASB's
Codification and Concepts Statement 8 and OpenStax quoted briefly,
never copied: copyright and CC BY-NC-SA); and names the code that
enforces it and the tests and proofs that check it.
- Spec blocks in the code now carry SOURCE: lines into the catalog;
AGENTS.md makes sourcing rule 0 of the formal-methods bar.
Corrections the sources turned up: the de minimis limit is $500 in the
regulation and $2,500 by Notice 2015-82 (both now cited); 1099
thresholds after 2026 are inflation-indexed, so the engine's flat
$2,000 for 2027+ is an open question (B-19). Unverified ASC paragraphs
(360-10-35-4, 340-10, 230-10-45-10, 606-10-55-37B, 210) are listed in
docs/spec/README.md until someone with a FASB basic-view login reads
them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Max-testing pass, step 2: two spec decisions Stephen confirmed
(2026-10-13), each test-first.
Aging (arap::aging_bucket): a document due today was "1-30 days past
due"; it is Current, since past due starts the day after (QuickBooks'
convention). Spec in the doc comment. Boundary examples at every edge
(due today, 1, 30, 31, 60, 61, 90, 91 days) fail on the old code. A
property: the balance lands in exactly one bucket, and time only ages
it.
Member splits (periods::allocate): the last member absorbed every odd
cent, so the same partner always got or lost it. Now largest remainder:
each share rounded toward zero to the cent, leftover cents to the
largest remainders, ties to the earlier member, a loss the negation of
the gain. The core is a new kernel function, kernel::apportion,
allocation-free so Kani proves it (n <= 3): apportion_is_exact (it
succeeds exactly when the shortfall is under a unit per part; then the
parts sum to the whole, each gains 0 or 1 unit, one may take the
sub-unit fraction) and apportion_is_fair (never a unit to a part while
a larger remainder goes without). Kani found an off-by-one in the first
version: a shortfall of n units plus a fraction was accepted. Native
regression apportion_refuses_a_shortfall_of_a_unit_per_part_or_more.
Properties at Decimal (2,000 cases, up to 5 members): conservation,
within a cent of each exact share, fairness, sign symmetry; they fail on
the old rule. Bridge: apportion agrees on i64 and Decimal.
Contract: closes journaled before this change replay exactly. New
member closes journal "rounding": "largest_remainder"; journal lines
without it replay as CloseTo::MembersLastAbsorbs, the old rule
(periods::allocate_last_absorbs, pinned by
closes_journaled_under_the_old_rule_replay_with_it). The members'
equity statement isn't journaled and uses the new rule.
Open gap: Lean proof of apportion's sum law for any n (needs a
permutation lemma on the ranking); Kani covers n <= 3, properties n <= 5.
The Lean closing theorems already hold for any split summing to the
net.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Max-testing pass, step 2 (rank 1 regressions + a rank 2 law). Two ways
to make the books unopenable, both found by test, both fixed:
1. Reversing into a closed period: the store journaled `entry.reversed`,
then the engine refused it at post. Replay could never apply that
line, so Store::open failed ("period 1 is closed"). Regression:
reversal_into_a_closed_period_is_refused_before_journaling.
2. An oversized invoice (or any command whose hand-written store
pre-checks lagged the engine, here the new line range) journaled,
then failed in the engine. Found by the new property
refused_commands_never_reach_the_journal within 4s.
The fix is structural, not another hand-written check: every
journal-first store command now rehearses on a scratch clone of the
ledger (`Store::rehearse`), so whatever the engine refuses is refused
before the journal, now and after any future engine change. The
create_entry rehearsal uses the checked constructor: no draft is ever
dated in a closed period (the Lean invariant Inv.draftsOpen, which the
store had bypassed).
The property (1,000 random command sequences, heavy on refusals)
asserts, after every command: a refusal leaves the journal
byte-identical; every posted entry validates; the trial balance
balances; no draft sits in a closed period; and the books reopen and
verify at the end. Shown to have teeth: removing the create_entry
rehearsal makes it fail and shrink to open-period, close, draft.
Engine (spec confirmed by Stephen, 2026-10-13), Ledger::check_reversal,
a pure pre-check that reverse_entry and the store share:
- an entry is reversed at most once (EntryAlreadyReversed); reversing
the reversal re-instates it;
- a closing entry is never reversed (ClosingEntryIrreversible);
- a refused reversal leaves no trace (it used to leave a draft and
burn an id).
Specs go in doc comments as a ```text GOAL/INVARIANTS/SCENARIOS block.
New dev dependency: proptest for ledger-store (already the workspace's
property library).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
The money arithmetic every invariant rests on now lives in one kernel,
ledger-core/src/kernel.rs: signing a line, totals, net, per-account
movements, reversal, normal balances, trial-balance columns, closing
lines, and the rounding plug for member splits. It is generic over a
checked `Amount` trait. The engine runs it on Decimal (validate,
movements, reverse_entry, account_balance, close_period_to, allocate,
trial_balance): there is one copy of the arithmetic, and it is the one
under proof.
Three layers, docs/verification.md maps every law to its proofs:
- Kani (src/proofs.rs): 17 harnesses prove the kernel at i64 for every
input up to 4 lines over 3 accounts. An entry is accepted exactly
when it is two or more positive, in-range lines whose debits equal
credits; accepted entries net to zero; grouping by account loses
nothing; a reversal mirrors and cancels its original; trial-balance
columns balance iff the books net to zero; closing lines zero their
accounts and a close balances however the net is split; the plug
makes parts sum to the whole. No panics, no overflow. kani::cover!
shows the proofs are not vacuous.
- Lean 4 (verify/lean): 24 headline theorems for books of any size. The
ledger is a state machine (add account, draft, post, reverse, open,
close), and in every reachable state the books net to zero, the trial
balance balances (as of any date), and the accounting equation holds.
Posting only appends ("posted is forever"), closed periods are
frozen, a close zeroes revenue and expenses and credits exactly the net
income, and period statements add up. Audit.lean shows standard axioms
only, no sorry.
- Bridge (tests/kernel_bridge.rs): the Decimal and i64 kernels agree.
What the proofs found (each one missed by every test, fuzzer, and
mutant):
- The engine accepted negative line amounts; only MCP refused them. Now
NegativeAmountLine: direction is the side's job.
- Summing lines near Decimal::MAX panicked. Lines are now capped at one
quadrillion and ten decimal places (AmountOutOfRange), which keeps
every realistic sum exact, and the kernel's sums are checked.
- FX conversion (foreign_amount * fx_rate) could panic the same way; it
is now a checked multiply.
- Kani itself: 0.68 mis-models symbolic-length copy_from_slice
(model-checking/kani#4904) and reported a reversal that doesn't net to
zero. The counterexample, replayed natively, passes, and a 15-line
standalone crate reproduces the bug. Harnesses avoid bulk copies, and
a should_panic canary will flag the upstream fix.
`just prove` runs it all (Kani memory- and time-capped, fails on an
unsatisfiable cover or a non-standard Lean axiom). AGENTS.md now makes
formal coverage part of every core change.
Also, docs: GnuCash is *an* oracle, never *the* oracle (Stephen). The
spec is accounting principles, proved. Golden files are recomputed from
principles. External implementations (GnuCash, the old Sheet, QBO/Xero)
are cross-checks whose disagreements get adjudicated, never obeyed.
docs/oracle.md gains "Who is right"; AGENTS, README, architecture,
roadmap, parity, and STATUS are reworded to match.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
Rust (with clippy and rustfmt), just, Node (for deploys), Kani, Lean,
cargo-mutants, and cargo-fuzz, so `mise install` is the whole bootstrap
and nothing gets installed by hand. Lean is 4.34.1, the newest release
past mise's minimum release age; verify/lean/lean-toolchain matches.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
The general ledger draws its own per-account tables, so it had no PDF
button. It now has one: each account a section, a line per posting
(date, entry, description, the other side) with debit, credit, and
running balance, its beginning balance, and totals with the ending
balance, typeset through the same statement template.
The PDF plumbing (button, status, background job) is shared by both
paths now. The ledger's rows also measure and clip their text and
scroll sideways in narrow windows, like every other table.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GWVo7x3UHw5kwCUaWxiejX
swaitsandClaude Opus 5.5