Cherry-picks the recency contract from #885 so stale items cannot lead ranks, including cluster representatives and the final sort.
Co-authored-by: bekonyn <noyanjeanbean@gmail.com>
Cherry-picks arctic listing fallback from #960 and keeps failed shreddit lanes honest when arctic recovers only part of the request.
Co-authored-by: technicianofthesacred <technicianofthesacred@users.noreply.github.com>
Adds an opt-in `amazon` source backed by the Bright Data CLI: product
discovery with live ratings and prices, plus a capped sample of recent
reviews woven in as buyer voice.
The signal it exists for is drift -- an all-time rating set against the
average of only the reviews inside the last 30 days. The emoji footer
names products and the direction they moved rather than counting
inventory, following the Polymarket shape:
📦 Amazon: 12 products │ Spirit E-325 Gas 4.4★→4.8★, Genesis E-325 Gas 4.4★ quiet
Off by default and dual-gated: the CLI must be on PATH and logged in, and
the run must explicitly request the source. It never auto-fires from
inferred intent. A drift arrow requires n>=5 dated in-window reviews,
since a live date census showed Amazon's top-reviews ordering is only
weakly recency-biased.
Includes doctor registration, honest setup reporting, SKILL.md trigger
guidance with worked examples, and CONFIGURATION.md coverage.
Reviewed by three independent agents plus Greptile (5/5). Fixes landed
for: the API key being visible in argv, --amazon-query leaking into
competitor sub-runs, a lane deadline that did not bound wall clock, brand
inference that silently disabled the competitor guard, and a saved-report
source list that dropped whole sources from the per-source dump.
Verified with three live end-to-end runs against a funded account.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`.get("engagement_score", 0)` only substitutes 0 for an *absent* key, so a
present-but-None value reached `max(None, existing)` on the update branch and
raised TypeError ("'>' not supported between instances of 'float' and
'NoneType'"). Guard `new_engagement` with `or 0`, symmetric with the
existing-row side already guarded on the same line.
The standard `findings_from_report` pipeline coerces None->0 at the producer,
so this hardens the public `store_findings(List[Dict[str, Any]])` boundary for
arbitrary callers rather than fixing a reachable pipeline crash. Adds a
regression test in tests/test_store.py.
Multi-line untrusted snippets could break out of Evidence indentation and
mint structural markdown headings inside EVIDENCE FOR SYNTHESIS. Indent
continuation lines and escape leading ATX markers at emit time.
Fixes#874
Passing --trustpilot-domain (or a plan-level trustpilot_domain) was accepted
and then silently ignored unless INCLUDE_SOURCES/--search already opted in.
Treat an explicit domain pin as intent and activate the source for the run.
Fixes#873
Stop shared Unreleased edits that conflict across PRs, and cut releases via
Prepare release (towncrier + every plugin/marketplace version bump) then
auto-tag on merge, with agent-oriented CONTRIBUTING and PR template guidance.
Matt Van Horn