Files
stack/packages/mosaic/framework/tools/wake/detector.sh
T
mos-dt-0andMos 8d1d6e5e76
ci/woodpecker/push/publish Pipeline was successful
ci/woodpecker/push/ci Pipeline was successful
feat(wake): #958 A11 preimage.sh — durable provenance for the operator-side preimage definition (#964)
Closes #958.

The preimage definition (source-adapter.sh) is the single most consequential file in the wake pipeline — every observed_hash is a sha256 of what it emits — and it was UNVERSIONED: no git history, no backup. When it was edited at 07:27 on 2026-07-30, attribution was recoverable only because an agent transcript happened to still be on disk. A11 gives that file durable provenance (option (2) of #958: recorded content-addressed, not in-band).

DESIGN, per the pre-registration:
- Provenance is OUT-OF-BAND (never on the adapter's stdout) — an in-band record would advance observed_hash for every source at once and manufacture the re-baseline it exists to explain (B3).
- The obligation never depends on the provenance path: a missing/corrupt store cannot halt the detector or swallow a wake (#940 advisory-fields precedent; B4).
- DESC_FMT=d1 is NOT the provenance record — the tag versions the descriptor FORMAT; a behaviour change that keeps descriptor shape re-baselines every hash and leaves the tag unchanged (B6).

CREDENTIAL HARD GATE (rebuilt after the first verdict FAILED it): byte capture is now RECORD-ONLY BY DEFAULT (extras opt in via WAKE_PREIMAGE_CAPTURE), not allow-by-default-refuse-on-shape — because a shape list can only refuse the secrets someone already enumerated, and the tool's own usage text recommended adding detector.env (where HMAC material lives). Deny is evaluated on BOTH raw and resolved path forms with resolved anchors, ordered before allow — closing the realpath-before-deny ordering defect that let a renamed symlink target through.

VERIFICATION (reviewer, mos-dt, independent of the author's claims):
- Seven decoy cases by planted-marker-then-grep-whole-state-dir: known cred path / same-name symlink / RENAMED-target symlink / prefixed secret / prefixless secret / opted-in-symlink-to-DENIED-target all REFUSED; opted-in-symlink-to-ALLOWED-target CAPTURED (positive control that the harness can capture at all, and that C was not closed by breaking every symlink).
- Polarity-completeness self-test RE-RUN with the shape list stubbed always-allow AND both deny lists stubbed — case D still safe: the flip is complete, the shape list is not load-bearing. Each stub proven live first (a stub that silently fails to apply reports the dangerous state as safe).
- B11: rm-then-change fails LOUD (rc=1), refuses to re-baseline, leaves the ledger absent; absent-with-emptied-objects still first-installs cleanly (absent-is-not-corrupt not paid for by breaking first install).
- RED-first reproduced exactly P13-P16 pre-fix; each refusal corroborated three ways (loud stderr, ledger row captured:false WITH a hash so attribution survives refusal, objects/ holding only the adapter).

KNOWN RESIDUAL (filed #969, non-gating): the deny check is both-forms but the suite needles only the resolved form — a deny reduced to resolved-only survives 17/17 green and would leak a renamed-symlink case. No reachable leak at this head (shipped code correct on all seven decoys); it constrains a FUTURE edit. Doctrine: a both-forms fix needs a needle per form; a fixture that satisfies its assertion through a DIFFERENT rule is testing the rule it did not mean to test.

Authored by pepper (sb-it-1-dt); independently reviewed by mos-dt (sb-it-1-dt) under diff-blind pre-registration (7242688b1, predating first read) — NOT CLEAR on the first verdict (B2/B11 failed by decoy), CLEAR at 8aff7d8 after the polarity rebuild. Manifest version 0.7.0.

Co-authored-by: mos-dt-0 <[email protected]>
2026-07-30 21:48:06 +00:00

573 lines
25 KiB
Bash
Executable File

#!/usr/bin/env bash
# detector.sh — A1 of the wake canon (EPIC #892, W4): the per-host,
# single-instance, delta-gated DETECTOR daemon.
#
# CONTRACT ANCHORS (docs/scratchpads/heartbeat-planning/CONVERGED-DESIGN.md):
# §1.1 Placement — ONE detector instance PER HOST, serving that host's LOCAL
# panes only (no cross-host waking). A long-lived SERVICE, NOT a timer:
# retiring unconditional timers is the whole point. Generalizes the
# proven delta-gated poller (gitea-pr-watch: poll -> hash -> deliver only
# on delta; 0-wasted) and adds (a) repo-section/anchor-scoped hashing so
# human-decision FILE edits are caught, not just API-visible state.
# §1.2 observe -> enqueue into the W2 durable store (the store ALLOCATES the
# observed_seq — single store-side allocator, #908).
# §2.4 Cursor semantics: `observed_seq` is a monotonic int, AUTHORITATIVE, now
# allocated SOLELY by store.sh enqueue (#908 unified the two former
# allocators). The detector no longer keeps a private counter: it enqueues
# and captures the store-assigned seq. Source SHAs are DESCRIPTORS, not the
# cursor. A per-watch LAST-OBSERVED HASH is compared each poll so a revert
# (A->B->A across polls) is caught as a delta.
# §4/G2a FAIL-LOUD source semantics: a source/target failure OR a 401 / 403 /
# privacy-404 / partial / ambiguous-empty response MUST fail loud, MUST
# NOT advance the authoritative cursor (`observed_seq`), and MUST NOT be
# silently treated as "no change". An error, or an empty response that
# might mean "hidden", is NEVER "no change".
# Gate B (#869 discipline): the watch-list `schema_version` is validated
# against the component manifest's supported [schema_min, schema_max]
# range; an out-of-range watch-list is REJECTED (fail loud), never
# silently coerced.
#
# SCOPE: this tool OBSERVES sources and ENQUEUES deltas to the W2 store
# (store.sh, merged). It does NOT reimplement the store, ack, digest, or signer
# (W2/W3 — it only CALLS their public APIs). FN-oracle/reconciler (W5), off-host
# beacon (W6) and installer (W7) are explicitly OUT OF SCOPE.
#
# SOURCE ADAPTER CONTRACT (operator-agnostic firewall: the detector bakes in NO
# git/gitea/HTTP/file specifics — the operator supplies the adapter):
# WAKE_DETECTOR_SOURCE_CMD is invoked as: <cmd> <kind> <id>
# with the source definition JSON on STDIN. It reports the CURRENT observed
# state of that source:
# exit 0 + NON-EMPTY stdout = the raw source descriptor (bytes to hash).
# exit 0 + EMPTY stdout = AMBIGUOUS-EMPTY -> FAIL LOUD (an empty that
# might mean "hidden/403'd" is never "no change").
# exit non-zero = source error (network failure, 401/403,
# privacy-404, partial/truncated) -> FAIL LOUD.
# The detector applies anchor-scoping (feature (a)) to the raw descriptor when
# the source defines an `anchor`, then hashes the scoped region.
#
# Operator-agnostic: all state via XDG/env; no operator paths/names/secrets/hosts.
set -uo pipefail
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
# shellcheck source=./_wake-common.sh disable=SC1091
. "$SCRIPT_DIR/_wake-common.sh"
STATE_DIR="$(wake_state_dir)"
STORE_SH="$SCRIPT_DIR/store.sh"
MANIFEST="$SCRIPT_DIR/manifest.txt"
# Detector-local state lives UNDER the store's STATE_DIR in its own subdir so it
# never collides with the store's cursor/inbox files.
DET_DIR="$STATE_DIR/detector"
_need_jq() {
command -v jq >/dev/null 2>&1 || {
echo "detector.sh: jq is required" >&2
exit 3
}
}
# _hash — sha256 of stdin, first field only (portable across sha256sum/shasum/
# openssl). Used for the per-watch last-observed hash (delta gate).
_hash() {
if command -v sha256sum >/dev/null 2>&1; then
sha256sum | awk '{print $1}'
elif command -v shasum >/dev/null 2>&1; then
shasum -a 256 | awk '{print $1}'
elif command -v openssl >/dev/null 2>&1; then
openssl dgst -sha256 | awk '{print $NF}'
else
echo "detector.sh: no sha256 tool (sha256sum/shasum/openssl) available" >&2
exit 3
fi
}
usage() {
cat >&2 <<'EOF'
Usage: detector.sh <command> [options]
Commands:
poll-once Run EXACTLY ONE delta-gated polling pass
over every source in the watch-list.
Enqueues to the W2 store on a delta only.
Exits non-zero if ANY source failed loud
(G2a) — the failing source does NOT
advance observed_seq and is NOT treated
as "no change".
run [--once] Long-lived per-host SINGLE-INSTANCE
service: acquire a non-blocking flock
(a 2nd instance REFUSES), then loop
poll-once every WAKE_DETECTOR_INTERVAL
seconds. --once does a single guarded
cycle then exits (still flock-guarded).
validate Load + validate the watch-list and its
schema_version against the manifest;
exit 0 if usable, non-zero (loud) if not.
cursors Print the store cursors (observed_seq is
the store's single source of truth, #908;
the detector keeps no private counter).
Environment:
WAKE_WATCH_LIST Path to the operator watch-list JSON (required).
WAKE_DETECTOR_SOURCE_CMD Source adapter command (see header contract; required
for poll-once/run).
WAKE_DETECTOR_INTERVAL run-loop poll interval seconds (default 30).
WAKE_DETECTOR_LOCK flock path (default <state>/detector/detector.lock).
WAKE_STATE_HOME/WAKE_AGENT store namespace (see store.sh).
EOF
}
# --- watch-list load + Gate B schema-version validation --------------------
_manifest_val() {
# _manifest_val KEY — echo VALUE for KEY=VALUE in the manifest (blank if none).
local key="$1"
[ -f "$MANIFEST" ] || return 0
sed -n "s/^${key}=//p" "$MANIFEST" | head -n1 | tr -d '[:space:]'
}
# _load_watchlist — validate the watch-list path + JSON + schema_version range.
# FAIL LOUD (non-zero) on anything unusable. Echoes the validated JSON on stdout.
_load_watchlist() {
_need_jq
local wl="${WAKE_WATCH_LIST:-}"
if [ -z "$wl" ]; then
echo "detector.sh: WAKE_WATCH_LIST is not set (no watch-list to observe)" >&2
return 2
fi
if [ ! -f "$wl" ]; then
echo "detector.sh: watch-list not found: $wl" >&2
return 2
fi
local json
if ! json="$(jq -e . "$wl" 2>/dev/null)"; then
echo "detector.sh: watch-list is not valid JSON: $wl" >&2
return 2
fi
# Required shape.
if ! printf '%s' "$json" | jq -e 'has("schema_version") and has("watches")' >/dev/null 2>&1; then
echo "detector.sh: watch-list missing required 'schema_version' or 'watches'" >&2
return 2
fi
# Gate B: schema_version MUST be an integer inside the manifest's supported
# [schema_min, schema_max] range. Out of range => REJECT (never coerce).
local ver smin smax
ver="$(printf '%s' "$json" | jq -r '.schema_version')"
smin="$(_manifest_val schema_min)"
smax="$(_manifest_val schema_max)"
case "$ver" in
'' | *[!0-9]*)
echo "detector.sh: watch-list schema_version must be an integer (got '$ver')" >&2
return 2
;;
esac
if [ -z "$smin" ] || [ -z "$smax" ]; then
echo "detector.sh: manifest is missing schema_min/schema_max (cannot validate Gate B)" >&2
return 2
fi
if [ "$ver" -lt "$smin" ] || [ "$ver" -gt "$smax" ]; then
echo "detector.sh: FAIL LOUD (Gate B) — watch-list schema_version $ver is OUTSIDE the supported range [$smin, $smax]; refusing to run against an incompatible watch-list" >&2
return 2
fi
printf '%s' "$json"
}
# --- observed_seq: the store is the SOLE allocator (#908) -------------------
#
# The detector NO LONGER keeps a private observed_seq counter. That private
# counter was the shared root of three defects (burn-before-enqueue, W5
# dual-allocator aliasing, and the migration-restart silent-swallow); see the PR
# for #908. observed_seq is now allocated EXCLUSIVELY by store.sh enqueue (which
# reads its own cursor and returns the assigned seq). The detector calls enqueue
# WITHOUT --seq and captures the store-returned seq for logging/locators — it
# never allocates a seq itself, so the seam is gone.
# --- per-watch last-observed hash (§2.4 revert/ABA-at-rest detection) --------
_hash_file() {
# _hash_file KIND ID — deterministic per-source hash-state filename.
local kind="$1" id="$2" key
key="$(printf '%s\037%s' "$kind" "$id" | _hash)"
printf '%s/watch-%s.hash' "$DET_DIR" "$key"
}
# --- anchor-scoped hashing (§1.1 feature (a)) -------------------------------
# _scope_anchor ANCHOR (raw content on stdin) — extract ONLY the region of the
# content delimited by ANCHOR: from the first line CONTAINING the anchor marker
# up to (but excluding) the next markdown-style heading line, or EOF. This lets
# a file edit OUTSIDE the lane's anchored section not wake the lane, while an
# edit INSIDE it is caught — human-decision FILE edits, not just API state.
# An empty scope (anchor not present) is surfaced to the caller as empty output.
_scope_anchor() {
local anchor="$1"
awk -v a="$anchor" '
index($0, a) > 0 && !inzone { inzone=1; print; next }
inzone && /^#/ { exit }
inzone { print }
'
}
# --- one poll of one source -------------------------------------------------
# _poll_source KIND ID DEF_JSON CLASS — returns:
# 0 processed (delta enqueued OR no-op no-change OR silent first-seen baseline)
# 1 FAIL LOUD (source error / ambiguous-empty) — cursor + hash left UNTOUCHED
_poll_source() {
local kind="$1" id="$2" def="$3" class="$4"
local anchor
anchor="$(printf '%s' "$def" | jq -r '.anchor // empty')"
# --- observe via the operator source adapter (fail-loud contract) ---------
# The source definition is handed to the adapter on STDIN via a temp FILE
# (not a pipe): an adapter that ignores stdin must not take SIGPIPE and read
# back as a spurious "source error" under `pipefail`.
local raw rc deftmp metatmp
mkdir -p "$DET_DIR"
# _wake_tmp_prefix is provided by _wake-common.sh (sourced above).
# shellcheck disable=SC2154
deftmp="$(mktemp "$DET_DIR/${_wake_tmp_prefix}defXXXXXX")" || return 1
printf '%s' "$def" >"$deftmp"
# fd 3 is the OUT-OF-BAND snapshot-metadata channel (#940): the adapter MAY
# write one JSON object {"snapshot_sha": "<git commit sha>", "snapshot_ts":
# <epoch>} there. It must stay out of stdout because EVERYTHING on stdout is
# hashed by the delta gate — an in-band tip-commit SHA would advance
# observed_hash on every unrelated push. Adapters that never write fd 3 leave
# the file empty: byte-identical legacy behavior.
metatmp="$(mktemp "$DET_DIR/${_wake_tmp_prefix}metaXXXXXX")" || {
rm -f "$deftmp"
return 1
}
raw="$("$WAKE_DETECTOR_SOURCE_CMD" "$kind" "$id" <"$deftmp" 3>"$metatmp" 2>/dev/null)"
rc=$?
rm -f "$deftmp"
# Slurp + remove the metadata file NOW so every return path below is clean;
# it is parsed only after the fail-loud gates (metadata from a FAILED
# observation is meaningless and must not be trusted or diagnosed).
local rawmeta=""
[ -s "$metatmp" ] && rawmeta="$(cat "$metatmp")"
rm -f "$metatmp"
if [ "$rc" -ne 0 ]; then
echo "detector.sh: FAIL LOUD (G2a) — source '$kind/$id' errored (adapter exit $rc: network/401/403/privacy-404/partial). observed_seq NOT advanced; NOT treated as 'no change'." >&2
return 1
fi
if [ -z "$raw" ]; then
echo "detector.sh: FAIL LOUD (G2a) — source '$kind/$id' returned AMBIGUOUS-EMPTY (empty-that-might-mean-hidden is never 'no change'). observed_seq NOT advanced." >&2
return 1
fi
# --- snapshot metadata (fd 3, #940) — ADVISORY, validated, never gating ----
# Malformed metadata is dropped with a LOUD diagnostic but cannot fail the
# poll or suppress the wake: the obligation never depends on optional dating.
local snap_sha="" snap_ts="" snap_json=""
if [ -n "$rawmeta" ]; then
if snap_json="$(jq -ce '.' <<<"$rawmeta" 2>/dev/null)"; then
snap_sha="$(jq -r 'if (.snapshot_sha|type) == "string" then .snapshot_sha else "" end' <<<"$snap_json")"
snap_ts="$(jq -r 'if (.snapshot_ts|type) == "number" then (.snapshot_ts|floor|tostring) else "" end' <<<"$snap_json")"
if [ -n "$snap_sha" ] && ! printf '%s' "$snap_sha" | grep -Eq '^[0-9a-f]{7,64}$'; then
echo "detector.sh: source '$kind/$id' snapshot_sha rejected (not a 7-64 char lowercase-hex git sha) — snapshot metadata DROPPED, poll continues (#940)." >&2
snap_sha=""
snap_ts=""
fi
# A ts must be a sane positive epoch BEFORE any arithmetic touches it: a
# negative or absurdly large value would make the shell integer comparison
# below error out and silently KEEP the bad ts — validate first, compare after.
if [ -n "$snap_ts" ] && ! printf '%s' "$snap_ts" | grep -Eq '^[0-9]{1,12}$'; then
echo "detector.sh: source '$kind/$id' snapshot_ts rejected (not a sane positive epoch) — snapshot_ts DROPPED, poll continues (#940)." >&2
snap_ts=""
fi
# ts is only meaningful anchored to a revision the consumer can re-verify —
# a bare number with no sha behind it is the weakest possible attestation,
# so ts requires a valid sha (#940 review §2).
if [ -n "$snap_ts" ] && [ -z "$snap_sha" ]; then
echo "detector.sh: source '$kind/$id' snapshot_ts without a valid snapshot_sha — snapshot_ts DROPPED (unverifiable dating), poll continues (#940)." >&2
snap_ts=""
fi
# A FUTURE ts renders a stale snapshot fresher-than-fresh (negative age) —
# wrong in the reassuring direction, the exact failure class #940 fixes.
# Cross-host NTP skew of a few seconds is the NORMAL case, so allow a small
# slack; beyond it, drop the ts (the sha stays: independently verifiable).
if [ -n "$snap_ts" ]; then
# The OPERATOR's knob gets the same discipline as the adapter's ts: it
# is interpolated into $((...)) under set -u, so a non-numeric value
# ('300s', '5m', 'abc') would be FATAL to the poll — the one thing this
# block must never be. Resolve it ONCE, validate, fall back loudly.
local now_s slack slack_ok
slack="${WAKE_SNAPSHOT_TS_FUTURE_SLACK:-300}"
# NOT grep: grep is LINE-oriented, so ^...$ anchors bind per line and a
# multi-line value ($'300\n8') passes the regex yet is FATAL in $((...)).
# The case pattern matches the WHOLE string, newlines included. (snap_ts
# is immune: jq's number type-check above cannot emit an embedded newline.)
case "$slack" in
'' | *[!0-9]*) slack_ok=1 ;;
*) [ "${#slack}" -le 9 ] && slack_ok=0 || slack_ok=1 ;;
esac
if [ "$slack_ok" -ne 0 ]; then
echo "detector.sh: WAKE_SNAPSHOT_TS_FUTURE_SLACK='$slack' is not a plain non-negative integer of at most 9 digits (seconds) — falling back to 300, poll continues (#940)." >&2
slack=300
fi
# Shape validation is not radix validation: bash reads a leading zero as
# OCTAL, so '08'/'09' pass the shape check yet are FATAL in $((...)), and
# '0300' silently means 192. Force base-10 so the knob means what the
# operator wrote (safe: the case pattern above guarantees pure digits).
slack=$((10#$slack))
now_s="$(date +%s)"
if [ "$snap_ts" -gt $((now_s + slack)) ]; then
echo "detector.sh: source '$kind/$id' snapshot_ts is beyond the ${slack}s future-skew allowance (ts=$snap_ts now=$now_s) — snapshot_ts DROPPED, poll continues (#940)." >&2
snap_ts=""
fi
fi
else
echo "detector.sh: source '$kind/$id' wrote UNPARSEABLE snapshot metadata on fd 3 — DROPPED, poll continues (#940)." >&2
fi
fi
# --- anchor-scope (feature (a)) then hash ---------------------------------
local scoped
if [ -n "$anchor" ]; then
scoped="$(printf '%s' "$raw" | _scope_anchor "$anchor")"
if [ -z "$scoped" ]; then
echo "detector.sh: FAIL LOUD (G2a) — source '$kind/$id' anchor '$anchor' not present in the observed content (ambiguous: cannot tell 'section removed' from 'hidden'). observed_seq NOT advanced." >&2
return 1
fi
else
scoped="$raw"
fi
local h hf last
h="$(printf '%s' "$scoped" | _hash)"
hf="$(_hash_file "$kind" "$id")"
mkdir -p "$DET_DIR"
if [ ! -f "$hf" ]; then
# First-seen: establish the baseline WITHOUT a wake (generalizes the
# proven deliver-on-NEW poller; startup parity is the reconciler's job,
# W5, out of scope). Flagged as a deliberate choice in the PR body.
printf '%s' "$h" | _atomic_write "$hf"
return 0
fi
last="$(tr -d '[:space:]' <"$hf")"
if [ "$h" = "$last" ]; then
# DELTA GATE: unchanged -> NO enqueue (0-wasted). Cursor untouched.
return 0
fi
# --- DELTA: enqueue and let the STORE allocate observed_seq (#908) ---------
# (a revert A->B->A lands here because $h != the stored $last). The store is
# the sole allocator: we call enqueue WITHOUT --seq and capture the seq it
# assigns (for logging/locators/emit). We do NOT allocate a seq ourselves.
local locators emit_ts
emit_ts="$(date +%s)"
# NB: `def` is a reserved word in jq — the source-definition arg is `$sdef`.
# snapshot_sha/snapshot_ts (#940) join only when the adapter attested them.
locators="$(jq -cn \
--arg kind "$kind" \
--arg id "$id" \
--argjson sdef "$def" \
--arg hash "$h" \
--arg ssha "$snap_sha" \
--arg sts "$snap_ts" \
'{kind:$kind, id:$id, observed_hash:$hash}
+ (if $ssha != "" then {snapshot_sha: $ssha} else {} end)
+ (if $sts != "" then {snapshot_ts: ($sts|tonumber)} else {} end)
+ ( $sdef | {repo, path, anchor, remote, branches} | with_entries(select(.value != null)) )')"
local args=(enqueue --locators "$locators" --emit-ts "$emit_ts")
# Pass class through only when the source declares one; absent => the store
# defaults to `actionable` (fail-safe, §2.3). Never guess a coalescible class.
[ -n "$class" ] && args+=(--class "$class")
local seq
if ! seq="$("$STORE_SH" "${args[@]}")"; then
echo "detector.sh: store enqueue FAILED for '$kind/$id' (observed_seq NOT advanced; hash NOT advanced)" >&2
return 1
fi
# Only AFTER a durable enqueue do we advance the per-watch last-observed hash.
# (store-allocated seq $seq is available for logging/emit if needed.)
: "$seq"
printf '%s' "$h" | _atomic_write "$hf"
return 0
}
# --- one full pass over the watch-list -------------------------------------
cmd_poll_once() {
local json
json="$(_load_watchlist)" || exit $?
if [ -z "${WAKE_DETECTOR_SOURCE_CMD:-}" ]; then
echo "detector.sh: WAKE_DETECTOR_SOURCE_CMD is not set (no adapter to observe sources)" >&2
exit 2
fi
_wake_init_dir "$STATE_DIR"
# Detector-run maintenance stale-tmp reap (#927). Stale-tmp cleanup was moved
# OFF the per-enqueue hot path (it clobbered concurrent in-flight writes); the
# detector poll tick is a natural once-per-pass maintenance point that bounds
# crash-left tmp accumulation on the long-running co-feed daemon. Age-scoped,
# so it can never delete a live in-flight enqueue write (its own or a
# concurrent reconciler's).
_wake_clean_stale_tmp "$STATE_DIR"
mkdir -p "$DET_DIR"
local failed=0
# #958 preimage provenance: check the OPERATOR-SIDE preimage definition (the
# source adapter file, the watch-list, operator-declared extras) BEFORE
# observing any source. A changed preimage re-baselines EVERY source at once;
# running the check first means its first-class cause line is enqueued at a
# LOWER observed_seq than the N per-source deltas it explains, so the digest
# shows the cause, not just the flood. An infrastructure failure of the check
# is LOUD and marks this pass failed (G2a discipline — never read as "no
# change"), but source observation still proceeds: provenance must not be
# able to starve wake delivery.
if ! "$SCRIPT_DIR/preimage.sh" check --enqueue; then
echo "detector.sh: FAIL LOUD — preimage provenance check failed (see preimage.sh above); source observation continues but this pass exits non-zero." >&2
failed=1
fi
# Iterate the DECLARED source-coverage inventory (§4/G3): only sources listed
# in watches[].sources[] are polled. An omitted source is not observed (and so
# cannot make anything pass vacuously); a referenced-but-undefined source is a
# malformed watch-list -> FAIL LOUD.
local pairs
pairs="$(printf '%s' "$json" | jq -r '
[ .watches[].sources[] | "\(.kind)\t\(.id)" ] | unique | .[]')"
if [ -z "$pairs" ]; then
echo "detector.sh: watch-list declares no sources under watches[].sources[]" >&2
exit 2
fi
local kind id def class
while IFS=$'\t' read -r kind id; do
[ -n "$kind" ] || continue
# Resolve the source definition from its top-level collection by id.
local coll
case "$kind" in
repo) coll="repos" ;;
board_file) coll="board_files" ;;
lane_anchor) coll="lane_anchors" ;;
*)
echo "detector.sh: FAIL LOUD — unknown source kind '$kind' in watch-list" >&2
failed=1
continue
;;
esac
def="$(printf '%s' "$json" | jq -c --arg c "$coll" --arg id "$id" \
'(.[$c] // []) | map(select(.id == $id)) | .[0] // empty')"
if [ -z "$def" ]; then
echo "detector.sh: FAIL LOUD (G3 parity) — watch references '$kind/$id' but no such entry is declared in '$coll'" >&2
failed=1
continue
fi
class="$(printf '%s' "$def" | jq -r '.class // empty')"
if ! _poll_source "$kind" "$id" "$def" "$class"; then
failed=1
fi
done <<EOF
$pairs
EOF
[ "$failed" -eq 0 ] || exit 1
}
# --- long-lived single-instance service ------------------------------------
cmd_run() {
local once=0
while [ $# -gt 0 ]; do
case "$1" in
--once)
once=1
shift
;;
*)
echo "detector.sh run: unknown option '$1'" >&2
exit 2
;;
esac
done
mkdir -p "$DET_DIR"
local lock="${WAKE_DETECTOR_LOCK:-$DET_DIR/detector.lock}"
mkdir -p "$(dirname "$lock")"
# PER-HOST SINGLE-INSTANCE (§1.1): a non-blocking exclusive flock. A second
# instance CANNOT acquire it and REFUSES loudly rather than double-observing.
exec 9>"$lock"
if ! flock -n 9; then
echo "detector.sh: FAIL LOUD — another detector instance already holds $lock; refusing (per-host single-instance)." >&2
exit 1
fi
# Ready marker so a supervisor/test can observe the lock is held before racing.
printf '%s\n' "$$" >"$lock.pid" 2>/dev/null || true
local interval="${WAKE_DETECTOR_INTERVAL:-30}"
case "$interval" in
'' | *[!0-9]*) interval=30 ;;
esac
while :; do
# A single failing source must not kill the long-lived service; poll-once
# already failed loud on stderr for it. The loop keeps serving healthy ones.
cmd_poll_once || true
# W6 SEAM (§1.3, off-host dead-man beacon): emit ONE monotonic liveness
# beacon per poll cycle. This is the ONLY W6 call site in the detector — a
# single, clean, opt-in seam. It is INERT unless the operator wires a beacon
# sink (WAKE_BEACON_SINK_CMD); wiring + install-validating that target is W7's
# job, not the detector's. Liveness is SPLIT from work-triggering, so a beacon
# emit failure NEVER kills the detector loop — but it is loud (not silent),
# and the off-host monitor's absence check (beacon.sh check) is the real
# safety net regardless of what this dying host does.
if [ -n "${WAKE_BEACON_SINK_CMD:-}" ]; then
"$SCRIPT_DIR/beacon.sh" emit >/dev/null || \
echo "detector.sh: WARN — off-host liveness beacon emit failed (see beacon.sh); the off-host absence check remains the authoritative dead-man." >&2
fi
[ "$once" -eq 1 ] && break
sleep "$interval"
done
}
cmd_cursors() {
# observed_seq has a SINGLE source of truth now (#908): the STORE cursor. The
# detector keeps no private counter, so cursors simply reports the store's
# authoritative cursors (observed_seq / consumed_seq / pending_depth).
"$STORE_SH" cursors 2>/dev/null || true
}
cmd_validate() {
local json
json="$(_load_watchlist)" || exit $?
local n
n="$(printf '%s' "$json" | jq -r '[.watches[].sources[]] | length')"
echo "detector.sh: watch-list OK ($n declared source reference(s))"
}
main() {
[ $# -ge 1 ] || {
usage
exit 2
}
local cmd="$1"
shift
case "$cmd" in
poll-once) cmd_poll_once "$@" ;;
run) cmd_run "$@" ;;
validate) cmd_validate "$@" ;;
cursors) cmd_cursors "$@" ;;
-h | --help | help) usage ;;
*)
echo "detector.sh: unknown command '$cmd'" >&2
usage
exit 2
;;
esac
}
main "$@"