Files
stack/packages/mosaic/framework/tools/wake/test-wake-digest-hmac.sh
mosaic-coder d3318c5910
All checks were successful
ci/woodpecker/pr/ci Pipeline was successful
feat(wake): W3 — cumulative-state digest renderer + non-circular HMAC signer
Builds on the merged W2 store/ack (EPIC #892). Adds A3 (digest.sh) and A5
(sign.sh) under packages/mosaic/framework/tools/wake/, honoring CONVERGED-DESIGN.

A3 — digest.sh (cumulative-state digest renderer), §2.1:
- CUMULATIVE-STATE: renders the FULL unacked set since consumed_seq from the
  durable pending-inbox (state-since-CONSUMED, not an event delta).
- TWO-TIER TRUST: orientation tier (who/lane/board-head + changed-obligation
  list with observed_seq + locators) decides the no-op case with ZERO tool
  calls; actionable tier renders every consequential fact ONLY as a
  CLAIM-TO-VERIFY, point-in-time-at-seq — never auto-actioned.
- HARD LOCATORS: every actionable claim must carry repo+issue#/40-char SHA/
  file:anchor; a missing locator is fail-loud (exit 4, nothing emitted).
- BOUNDING/INJECTION/SECRETS: source free-text is quoted only inside a
  delimited, length-capped, ANSI/bidi/zero-width-stripped untrusted block;
  secret-canary redaction over any inlined content; embeds the W2 ack line.

A5 — sign.sh (non-circular HMAC signer), §2.5:
- wake_id generated INDEPENDENTLY at emit (not derived from, and not a member
  of, the signed field-tuple); wake_mac = HMAC(key, wake_id || agent_identity
  || mission_generation || observed_seq || emit_ts || content_hash) — over
  wake_id PLUS the fields, genuinely non-circular (the MAC is never its own
  input). Fills the `hmac` placeholder W2 left in store entries.
- Key resolved BY NAME from the credential store, never inlined, never echoed;
  no flag accepts key material. Same-uid threat boundary documented; off-uid
  signer named as a future gate.

RED-FIRST tests (test-wake-digest-hmac.sh, wired into test:framework-shell):
cumulative-state, hard-locator fail-loud, two-tier (zero-call orientation +
claim-to-verify), scrub (secret-canary + ANSI/bidi/zero-width), non-circular
HMAC (independent wake_id + tamper-breaks-MAC), key-by-name-never-inline. Each
verified to go RED on a targeted regression. shellcheck clean; operator-agnostic
(XDG/env only). manifest.txt bumped to 0.2.0.

Part of #892

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0158NZqN2n2ymKFeJAZ4GUCb
2026-07-25 19:19:10 -05:00

279 lines
15 KiB
Bash
Executable File

#!/usr/bin/env bash
# test-wake-digest-hmac.sh — RED-FIRST invariant harness for W3 (EPIC #892):
# the cumulative-state digest renderer (digest.sh, A3) + the non-circular HMAC
# signer (sign.sh, A5).
#
# Each test asserts ONE CONVERGED-DESIGN invariant and is designed to go RED if
# that invariant regresses:
# D1 CUMULATIVE-STATE: two still-pending changes both render (not just the
# latest delta) — a delta would silently drop the older change. (§2.1)
# D2 HARD-LOCATOR enforcement: an actionable claim with no precise locator
# FAILS LOUD (non-zero exit, nothing emitted). (§2.1)
# D3 TWO-TIER trust: orientation decides the no-op case with ZERO tool calls;
# an actionable fact is a CLAIM-TO-VERIFY, never auto-actioned. (§2.1)
# D4 SCRUB: a secret-canary + ANSI/bidi/zero-width in SOURCE free-text is
# stripped/neutralized in the rendered digest. (§2.1)
# H1 NON-CIRCULAR HMAC: wake_id is NOT a member of the signed field-tuple and
# is independently generated; wake_mac is not its own input; tampering ANY
# signed field (or wake_id) breaks the MAC. (§2.5)
# H2 KEY BY-NAME, NEVER INLINE: the key is resolved by NAME from the credential
# store; no flag accepts key material; the key never leaks into output; the
# same-uid threat boundary is documented. (§2.5)
#
# Isolated: every test runs against a fresh temp state / credential file.
set -uo pipefail
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
STORE="$SCRIPT_DIR/store.sh"
DIGEST="$SCRIPT_DIR/digest.sh"
SIGN="$SCRIPT_DIR/sign.sh"
command -v jq >/dev/null 2>&1 || {
echo "SKIP: jq not available" >&2
exit 0
}
command -v openssl >/dev/null 2>&1 || {
echo "SKIP: openssl not available" >&2
exit 0
}
TMP_ROOT="$(mktemp -d)"
trap 'rm -rf "$TMP_ROOT"' EXIT
# Failures recorded to a marker FILE (not a shell var): each test runs in a
# subshell for env isolation and a subshell cannot mutate a parent variable, so
# a var-based counter would silently swallow failures.
FAILFILE="$TMP_ROOT/failures"
: >"$FAILFILE"
pass=0
fail_msg() {
echo " FAIL: $*" >&2
echo "x" >>"$FAILFILE"
}
ok() { pass=$((pass + 1)); }
fresh_state() {
local d="$TMP_ROOT/$1"
rm -rf "$d"
mkdir -p "$d"
printf '%s' "$d"
}
# A 40-hex sha, for hard locators.
SHA40="abcdef0123456789abcdef0123456789abcdef01"
# fresh_cred NAME KEYNAME KEYVALUE — write a temp credential store with a wake
# HMAC key at .wake.hmac_keys.<KEYNAME> and echo its path.
fresh_cred() {
local f="$TMP_ROOT/$1.cred.json"
jq -cn --arg k "$2" --arg v "$3" '{wake:{hmac_keys:{($k):$v}}}' >"$f"
printf '%s' "$f"
}
# ===========================================================================
echo "== D1: CUMULATIVE-STATE — two pending changes BOTH render (not just latest) =="
(
WAKE_STATE_HOME="$(fresh_state d1)"
export WAKE_STATE_HOME
unset WAKE_AGENT WAKE_LANE
# Two distinct obligations observed since the last CONSUMED. A delta-style
# renderer would show only the newest; cumulative-state must show BOTH.
"$STORE" enqueue --seq 1 --class actionable --locators '{"repo":"r","issue":11}' >/dev/null
"$STORE" enqueue --seq 2 --class actionable --locators '{"repo":"r","issue":22}' >/dev/null
out="$("$DIGEST" render)" || fail_msg "D1: render exited non-zero"
echo "$out" | grep -q 'issue=#11' || fail_msg "D1: OLDER change (issue #11) dropped — digest is a delta, not cumulative state"
echo "$out" | grep -q 'issue=#22' || fail_msg "D1: newer change (issue #22) missing"
echo "$out" | grep -q 'seq 1' || fail_msg "D1: seq 1 not listed in cumulative set"
echo "$out" | grep -q 'seq 2' || fail_msg "D1: seq 2 not listed in cumulative set"
# A coalescing digest-class entry is STATE (full), not a delta: a later digest
# subsumes the earlier, but the cumulative unacked set (both actionables) stays.
echo "$out" | grep -q 'pending=2' || fail_msg "D1: cumulative pending count wrong (expected 2)"
) && ok
echo "== D2: HARD-LOCATOR enforcement — actionable claim with no locator FAILS LOUD =="
(
WAKE_STATE_HOME="$(fresh_state d2)"
export WAKE_STATE_HOME
unset WAKE_AGENT WAKE_LANE
# An actionable claim carrying NO hard locator (no repo+issue / sha / file).
"$STORE" enqueue --seq 1 --class actionable --locators '{"claim":"mergeable=true"}' >/dev/null
if out="$("$DIGEST" render 2>/dev/null)"; then
fail_msg "D2: malformed digest (actionable, no locator) must FAIL, but render succeeded"
fi
rc=0
"$DIGEST" render >/dev/null 2>&1 || rc=$?
[ "$rc" -eq 4 ] || fail_msg "D2: expected fail-loud exit 4, got $rc"
# Nothing emitted on stdout when malformed (no silent partial send).
[ -z "${out:-}" ] || fail_msg "D2: a malformed digest emitted body instead of failing loud [$out]"
# A present-but-imprecise sha (not 40 hex) does NOT satisfy the hard locator.
WAKE_STATE_HOME="$(fresh_state d2b)"
export WAKE_STATE_HOME
"$STORE" enqueue --seq 1 --class actionable --locators '{"claim":"ci=green","sha":"abc123"}' >/dev/null
rc=0
"$DIGEST" render >/dev/null 2>&1 || rc=$?
[ "$rc" -eq 4 ] || fail_msg "D2: imprecise sha (not 40-hex) must not satisfy the hard-locator gate (got exit $rc)"
# The SAME claim WITH a precise 40-hex sha renders fine.
WAKE_STATE_HOME="$(fresh_state d2c)"
export WAKE_STATE_HOME
"$STORE" enqueue --seq 1 --class actionable --locators "$(jq -cn --arg s "$SHA40" '{claim:"ci=green",sha:$s}')" >/dev/null
"$DIGEST" render >/dev/null 2>&1 || fail_msg "D2: a well-located actionable claim must render"
) && ok
echo "== D3: TWO-TIER — orientation no-op with ZERO tool calls; actionable = claim-to-verify =="
(
WAKE_STATE_HOME="$(fresh_state d3)"
export WAKE_STATE_HOME
unset WAKE_AGENT WAKE_LANE
# Tool-call tripwire: shim git/gh/tea/curl on PATH; any invocation touches a
# marker. The orientation no-op path must decide with ZERO of them.
bin="$TMP_ROOT/d3bin"
mkdir -p "$bin"
for t in git gh tea curl wget; do
{
printf '#!/usr/bin/env bash\n'
printf 'touch "%s/TOOL_CALLED"\n' "$WAKE_STATE_HOME"
printf 'exit 0\n'
} >"$bin/$t"
chmod +x "$bin/$t"
done
# No pending obligations => no-op common case.
out="$(PATH="$bin:$PATH" "$DIGEST" render --lane build)" || fail_msg "D3: no-op render failed"
echo "$out" | grep -q 'NO-OP' || fail_msg "D3: empty inbox must render an explicit NO-OP orientation"
[ -e "$WAKE_STATE_HOME/TOOL_CALLED" ] && fail_msg "D3: orientation no-op made a live tool call (must be ZERO)"
# Now an actionable, consequential fact. It must be a CLAIM-TO-VERIFY, never
# rendered as a trusted assertion or an auto-action.
"$STORE" enqueue --seq 1 --class actionable \
--locators "$(jq -cn --arg s "$SHA40" '{repo:"r",issue:7,claim:"ci=success",file:"pkg/x.ts",sha:$s}')" >/dev/null
out2="$(PATH="$bin:$PATH" "$DIGEST" render)" || fail_msg "D3: actionable render failed"
# Rendering itself still makes ZERO live calls (it only lays out claims).
[ -e "$WAKE_STATE_HOME/TOOL_CALLED" ] && fail_msg "D3: rendering an actionable claim made a live call (must defer to the consumer's gate)"
echo "$out2" | grep -q 'CLAIM@seq' || fail_msg "D3: consequential fact not framed as CLAIM@seq"
echo "$out2" | grep -qi 'VERIFY LIVE' || fail_msg "D3: claim not marked for live verification"
echo "$out2" | grep -qi 'do NOT act on this line' || fail_msg "D3: claim missing do-not-auto-action framing"
# The consequential fact must NOT appear as a bare trusted directive.
echo "$out2" | grep -qiE 'merge now|go ahead and (merge|deploy)|safe to merge' &&
fail_msg "D3: digest auto-actioned a consequential fact (imperative present)"
# And its hard locator + one-call re-verify hint are present.
echo "$out2" | grep -q "re-verify (ONE call)" || fail_msg "D3: actionable claim missing one-call re-verify locator"
true
) && ok
echo "== D4: SCRUB — secret-canary + ANSI/bidi/zero-width in source free-text neutralized =="
(
WAKE_STATE_HOME="$(fresh_state d4)"
export WAKE_STATE_HOME
unset WAKE_AGENT WAKE_LANE
# Build hostile source free-text: ANSI CSI, bidi override (U+202E), zero-width
# space (U+200B), ZWNBSP (U+FEFF), a C0 control (BEL), and two secret canaries.
nasty="$(printf 'good\x1b[31mRED\xe2\x80\xaeEVIL\xe2\x80\x8bZW\xef\xbb\xbfBOM\x07BEL ghp_0123456789abcdefghijABCDEFGHIJ012345 AKIAIOSFODNN7EXAMPLE')"
loc="$(jq -cn --arg s "$nasty" --arg sha "$SHA40" '{repo:"r",issue:3,claim:"changed",sha:$sha,summary:$s}')"
"$STORE" enqueue --seq 1 --class actionable --locators "$loc" >/dev/null
out="$("$DIGEST" render)" || fail_msg "D4: render failed"
# ANSI escape / CSI must be gone.
printf '%s' "$out" | LC_ALL=C grep -q "$(printf '\x1b')" && fail_msg "D4: ANSI ESC survived the scrub"
# bidi/zero-width/BOM UTF-8 sequences must be gone.
printf '%s' "$out" | LC_ALL=C grep -qP '\xe2\x80[\x8b-\x8f\xaa-\xae]|\xef\xbb\xbf' &&
fail_msg "D4: bidi/zero-width/BOM survived the scrub"
# C0 control bytes (except tab/newline) must be gone.
printf '%s' "$out" | LC_ALL=C grep -qP '[\x00-\x08\x0e-\x1f\x7f]' && fail_msg "D4: a C0 control byte survived the scrub"
# Secret canaries must be redacted, never inlined.
printf '%s' "$out" | grep -q 'ghp_0123456789' && fail_msg "D4: GitHub-token canary LEAKED into the digest"
printf '%s' "$out" | grep -q 'AKIAIOSFODNN7EXAMPLE' && fail_msg "D4: AWS-key canary LEAKED into the digest"
printf '%s' "$out" | grep -q 'REDACTED-SECRET' || fail_msg "D4: secret redaction marker absent — canary may not have been scrubbed"
# Free-text is quoted inside a DELIMITED untrusted block, framed as NOT instructions.
printf '%s' "$out" | grep -q 'BEGIN UNTRUSTED DATA' || fail_msg "D4: source free-text not placed in a delimited untrusted block"
# The 40-hex git SHA locator must NOT be mangled by the secret scrubber.
printf '%s' "$out" | grep -q "$SHA40" || fail_msg "D4: legitimate 40-hex SHA locator was wrongly scrubbed"
true
) && ok
echo "== H1: NON-CIRCULAR HMAC — wake_id independent + not a signed field; tamper breaks MAC =="
(
MOSAIC_CREDENTIALS_FILE="$(fresh_cred h1 default 'key-material-h1')"
export MOSAIC_CREDENTIALS_FILE
WAKE_AGENT=agentA
WAKE_MISSION_GENERATION=3
export WAKE_AGENT WAKE_MISSION_GENERATION
unset WAKE_HMAC_KEY_NAME
env1="$("$SIGN" sign --observed-seq 5 --content-hash "$SHA40")" || fail_msg "H1: sign failed"
# wake_id is NOT a member of the signed field-tuple (it is prepended to the MAC
# input, not one of the signed fields whose authenticity the MAC establishes).
echo "$env1" | jq -e '.signed | has("wake_id") | not' >/dev/null || fail_msg "H1: wake_id must NOT appear inside the signed field-tuple"
# wake_mac is not its own input: it is not a member of `signed` either.
echo "$env1" | jq -e '.signed | has("wake_mac") | not' >/dev/null || fail_msg "H1: wake_mac must not be part of its own signed inputs (self-reference)"
echo "$env1" | jq -e 'has("wake_id") and has("wake_mac")' >/dev/null || fail_msg "H1: envelope must carry independent wake_id + wake_mac"
# Independently generated: two emits over IDENTICAL fields yield DIFFERENT
# wake_ids (id not derived from the tuple) AND different MACs (id is bound in).
env2="$("$SIGN" sign --observed-seq 5 --content-hash "$SHA40")"
w1="$(echo "$env1" | jq -r .wake_id)"
w2="$(echo "$env2" | jq -r .wake_id)"
[ -n "$w1" ] && [ "$w1" != "null" ] || fail_msg "H1: wake_id missing"
[ "$w1" != "$w2" ] || fail_msg "H1: wake_id not independently generated (identical fields produced identical id — derived, not fresh)"
m1="$(echo "$env1" | jq -r .wake_mac)"
m2="$(echo "$env2" | jq -r .wake_mac)"
[ "$m1" != "$m2" ] || fail_msg "H1: MAC did not bind the independent wake_id (same MAC despite different wake_id)"
# Untampered verifies.
echo "$env1" | "$SIGN" verify >/dev/null 2>&1 || fail_msg "H1: untampered envelope failed to verify"
# Tampering ANY signed field breaks the MAC.
for mut in \
'.signed.agent_identity="attacker"' \
'.signed.mission_generation="9"' \
'.signed.observed_seq="6"' \
'.signed.emit_ts="0"' \
'.signed.content_hash="deadbeef"'; do
if echo "$env1" | jq "$mut" | "$SIGN" verify >/dev/null 2>&1; then
fail_msg "H1: tampering [$mut] did NOT break the MAC"
fi
done
# Tampering the independent wake_id also breaks the MAC (it is bound in).
if echo "$env1" | jq '.wake_id="wake_forged"' | "$SIGN" verify >/dev/null 2>&1; then
fail_msg "H1: forging wake_id did not break the MAC (id not bound)"
fi
# A wrong key fails verification (the MAC genuinely depends on the key).
wrong="$(fresh_cred h1b default 'DIFFERENT-key')"
if echo "$env1" | MOSAIC_CREDENTIALS_FILE="$wrong" "$SIGN" verify >/dev/null 2>&1; then
fail_msg "H1: verify passed under a DIFFERENT key (MAC not key-dependent)"
fi
) && ok
echo "== H2: KEY BY-NAME, never inline; no key leak; same-uid boundary documented =="
(
MOSAIC_CREDENTIALS_FILE="$(fresh_cred h2 signing-key 'SUPERSECRET-KEYVALUE-XYZ')"
export MOSAIC_CREDENTIALS_FILE
WAKE_AGENT=agentB
WAKE_MISSION_GENERATION=1
export WAKE_AGENT WAKE_MISSION_GENERATION
# Resolve BY NAME (the credential-store key name), never the key material.
env1="$(WAKE_HMAC_KEY_NAME=signing-key "$SIGN" sign --observed-seq 1 --content-hash "$SHA40")" ||
fail_msg "H2: by-name key resolution failed"
echo "$env1" | jq -e .wake_mac >/dev/null || fail_msg "H2: no MAC produced from by-name key"
# The key VALUE must never appear in the signed output.
echo "$env1" | grep -q 'SUPERSECRET-KEYVALUE-XYZ' && fail_msg "H2: key material LEAKED into the signed envelope"
# A signed store-entry (hmac placeholder filled) must not leak the key either.
entry="$(printf '{"observed_seq":1,"locators":{"repo":"r","issue":1},"class":"actionable","emit_ts":1700000000,"hmac":""}' |
WAKE_HMAC_KEY_NAME=signing-key "$SIGN" sign-entry)"
echo "$entry" | grep -q 'SUPERSECRET-KEYVALUE-XYZ' && fail_msg "H2: key material LEAKED into the signed entry"
echo "$entry" | jq -e '.hmac != "" and .hmac != null' >/dev/null || fail_msg "H2: sign-entry did not fill the hmac placeholder"
# No flag may accept key MATERIAL inline — only a key NAME. An attempt to pass
# a literal key must be rejected as an unknown option (never silently honored).
for badflag in --key --secret --hmac-key --key-value; do
if WAKE_HMAC_KEY_NAME=signing-key "$SIGN" sign --observed-seq 1 --content-hash "$SHA40" "$badflag" 'INLINE-KEY' >/dev/null 2>&1; then
fail_msg "H2: sign.sh accepted an inline-key flag [$badflag] (key must be by-name only)"
fi
done
# An unknown key NAME fails loud (never signs unsigned / with an empty key).
if WAKE_HMAC_KEY_NAME=does-not-exist "$SIGN" sign --observed-seq 1 --content-hash "$SHA40" >/dev/null 2>&1; then
fail_msg "H2: signing with an unresolvable key name must FAIL LOUD"
fi
# The same-uid threat boundary + off-uid follow-up must be DOCUMENTED in the tool.
grep -qi 'same-uid' "$SIGN" || fail_msg "H2: same-uid threat boundary not documented in sign.sh"
grep -qi 'off-uid' "$SIGN" || fail_msg "H2: off-uid future signer not named in sign.sh"
) && ok
echo
if [ -s "$FAILFILE" ]; then
echo "wake digest/hmac harness: FAILED ($(grep -c . "$FAILFILE") assertion(s))" >&2
exit 1
fi
echo "wake digest/hmac harness: all invariants passed ($pass groups)"