matrym-hashchain

A tamper-evident, append-only, hash-chained ledger. Dependency-free (stdlib only).

Each entry carries a SHA-256 over its own payload and the previous entry's hash, so any later edit, reorder, or deletion of a past record is detected the next time the log is read. It's a file-simple, zero-dependency primitive for audit trails, transaction logs, and chain-of-custody records - the kind of thing a compliance, finance, or records system needs and shouldn't hand-roll.

Install

pip install matrym-hashchain

Released to PyPI via trusted publishing (GitHub OIDC, no stored token).

Use

from pathlib import Path
from hashchain import append, read, verify, HashChainError

ledger = Path("audit.jsonl")

append(ledger, {"event": "created", "who": "alice"})
append(ledger, {"event": "approved", "who": "bob"})

for entry in read(ledger):          # read() verifies the whole chain as it goes
    print(entry.seq, entry.payload)

verify(ledger)                      # True -- the history reads clean end to end

# If anyone edits, reorders, or deletes a past record...
ledger.write_text(ledger.read_text().replace("alice", "mallory"))
verify(ledger)                      # False
read(ledger)                        # raises HashChainError: record 0 was tampered ...

The store is plain JSON Lines (.jsonl) - one record per line, human-readable, greppable, and portable. No database, no daemon, no dependencies.

Command line

The same three verbs, from one terminal line (installed as matrym-hashchain, or run python -m hashchain):

matrym-hashchain append audit.jsonl '{"event": "created", "who": "alice"}'
matrym-hashchain append audit.jsonl '{"event": "approved", "who": "bob"}'
matrym-hashchain read   audit.jsonl        # 0<TAB>{"event": "created", ...}
matrym-hashchain verify audit.jsonl        # "clean: 2 record(s), head e7c1f28a3595"; exit 0

verify exits 0 on a clean chain and 1 on a broken one, naming the tampered record on stderr, so it drops straight into a shell gate or a CI step. append also reads the payload object from stdin when you omit it (... | matrym-hashchain append audit.jsonl).

What it guarantees (and what it doesn't)

This proves integrity by default, and closes the two classic gaps with two optional features:

Any break fails loud with HashChainError rather than returning a dishonest history.

Signing and truncation, honestly

KEY = b"a secret only writers hold"
append(ledger, {"event": "created"}, key=KEY)   # HMAC-signed
verify(ledger, key=KEY)                          # True; verify(ledger) alone -> False (no key)

anchor = head_hash(ledger, key=KEY)              # stash this off-ledger
# ... later, after someone drops the last record ...
verify(ledger, key=KEY, expected_head=anchor)    # False -- the tail is gone

Honest label: HMAC is a symmetric shared-secret MAC. It proves a record was written by a holder of the key; it is not public-key non-repudiation (anyone with the key can both sign and verify). The same key must be passed to every append/read/verify/head_hash over a signed ledger.

One more honest bound: this is a single-writer primitive. append is not guarded against two processes writing the same ledger at once (no file lock), so concurrent appends can interleave. Serialize writers (one process, or your own lock) if that's a risk; read/verify are read-only and safe alongside a single writer.

Performance

append is O(1) in ledger size: it reads only the ledger's tail (a backward seek), not the whole chain, so appends stay fast as the log grows without bound. That claim ships with evidence, not just an assertion, measured head to head against the pre-0.1.1 O(n) behavior:

Across a 32x ledger growth the old append slowed 36.6x (linear); the current one stayed flat. Full method, table, and reproduction: benchmarks/RESULTS.md (make bench).

API

All functions take an optional keyword-only key: bytes (sign/verify with HMAC; omit for plain integrity).

• append(path, payload, *, key=None) -> Entry - validate, hash-chain, and append one record. Reads only the tail (O(1)), validating that record before extending, so you can't quietly append onto a tampered tail.
• read(path, *, key=None) -> list[Entry] - read every record, verifying as it goes. Empty/missing store -> [].
• verify(path, *, key=None, expected_head=None) -> bool - True if the ledger reads clean end to end; pass expected_head to also catch truncation.
• head_hash(path, *, key=None) -> str - the last record's content hash (or GENESIS if empty); the value you anchor off-ledger to detect truncation.
• content_hash(mapping, *, key=None) -> str - the canonical digest the chain uses: SHA-256 unkeyed, HMAC-SHA256 when a key is given (sorted keys, no whitespace).
• Entry(seq, payload, prior_hash, content_hash) - one link; HashChainError - the loud failure.

Test

pip install -e ".[dev]"   # pytest, ruff, mypy, coverage, pip-audit
make check                # ruff + mypy --strict + pytest at 100% coverage (the full gate)
make security             # pip-audit the project's dependency closure (proves "zero deps")
make bench                # the O(1)-append benchmark

The suite pins both acceptance and refusal: a clean chain reads end to end and verify returns True, while every tampered case (an edited payload, a reorder, a deleted middle record, a mislinked record with a valid own hash, an append onto an already-broken log) fails loud with HashChainError rather than returning a dishonest history. CI runs it on every push across Python 3.10-3.13, at 100% line + branch coverage, and audits the dependency closure.

Provenance

This is a MatrymLabs Hardware Store part: a pattern first proven in the CodeForge MUD engine (the "Chronicle" - the game's tamper-evident memory), then reused across the fleet (a federal guidance-check ledger; an AI-triage decision audit trail). Published standalone so any project can reuse the same tested primitive - one well-made part, many jobs.

License

MIT