Repository navigation
Releases
Releases · Echolonius/boardcheck
v0.8.1
- AMS-006 no longer stops with an error when a listing's
title,descriptionorinstructions
is not text (a number, list or object). Such a value is skipped, as AMS-008 already did, in both
the Python library and the browser version. check_listingnow also runs AMS-006 and AMS-007. The single-listing MCP tool runs all five
checks that need only one listing (AMS-001, AMS-004, AMS-006, AMS-007 and AMS-008), in the order
scan()runs them. An agent checking one bounty now also learns about prompt-injection phrases
in its text and about an upfront fee with no confirmed escrow. AMS-002, AMS-003 and AMS-005
compare the listings on a board with each other, so they stay inscan()andscan_listings,
which are unchanged.- The tool's description,
manifest.json, README.md, SCHEMA.md,docs/llms.txtand the site's
agent instructions now say the same thing about whatcheck_listingruns. A test checks those
descriptions, and the tool lists inmanifest.json, README.md anddocs/llms.txt, against the
code. - README: new section, "Check a bounty issue before you work on it". It shows how an agent passes
one GitHub bounty issue, or any task text, tocheck_listingas a single listing (the title as
title, the raw body asdescription).tests/test_readme_examples.pyruns the made-up example,
so the findings shown are the ones the code returns.docs/llms.txtand the site's agent
instructions have the same recipe in short form. - Wording: the install note at the top of
agent_market_signals/mcp_server.pyused the
repository's old name, and FLYWHEEL.md still called the MCP server a next step.
MCP Bundle: boardcheck-mcp-0.8.1.mcpb (SHA-256 0c2ee61856552b960a102ad096b660cc2b9b93d57f76774ef2b36de1a8205ed1)
v0.8.0
- AMS-008
agent_bait_listing(highorwarn): flags listing text written to bait an AI
agent.highwhen it asks for a private key, seed phrase or API/access token to be sent or
pasted, or tells the solver to download and run remote code (curl … | python3, "download
verify.py and run it");warnwhen it says human submissions will not be reviewed, links several
escrow services for one bounty, promises automatic payout on merge or/claimwith nothing to
check, or hides text from people (zero-width or bidirectional control characters, or an HTML
comment that addresses an AI reader or gives it an instruction, such as "AI agents: …", "if you
are an LLM …", "ignore the above", or a hidden key request or download-and-run command). Merely
mentioning AI does not count, so template leftovers (<!-- Describe the bug -->) and AI-use
disclosure questions (<!-- Did you use AI or GPT to write this? -->) stay clean. One finding
per listing, naming every sign found.
Running the project's own tests never counts, and listings without text are skipped. Grounded in
copied "escrowed" USDC bounties seen on GitHub on 2026-11-06 and a published analysis of a
"bounty verifier" skill (both cited in SPEC.md). Runs inscan(), in the site's browser version
and in the single-listingcheck_listingMCP tool, which now runs AMS-001, AMS-004 and AMS-008. - Hidden text is part of AMS-008, not AMS-006: AMS-006 is defined as a
highphrase match, and
SPEC.md treats redefining an indicator as a major-version change, while new behaviour gets a new
ID. - The JS↔Python parity test now covers AMS-006 and AMS-008 as well as AMS-001 to AMS-005.
- New example file:
examples/agent_bait_listings.json.
MCP Bundle: boardcheck-mcp-0.8.0.mcpb (SHA-256 79d32730e06aef651a0b8d3735b66a1736033660781d86ed346cdc7f82812ea3)
v0.7.0
- AMS-006
prompt_injection_payload(high): fires when a listing'stitle,description
orinstructionscontain one of a fixed list of instruction-override or shell-command phrases
(letter case ignored), or when the listing hashas_prompt_injection: true. Runs inscan()and
in the site's browser version. - AMS-007
upfront_fee_gating(high): fires whenrequires_upfront_feeistrueand
has_escrowis nottrue(a missinghas_escrowcounts as no escrow). Runs inscan(); the
site's browser version does not run it. - Five new optional listing fields for these checks:
title,description,instructions,
has_prompt_injection,requires_upfront_fee(see SCHEMA.md). generate_certificate()and a fifth MCP tool,generate_audit_certificate: turn a scan
into a JSON summary (verdict, counts, timestamp and a short hash-based ID) with Markdown/HTML
badge snippets. It is not signed and is built from whatever listings are passed in.- Neither new check is in the JS↔Python parity test yet, and the single-listing
check_listing
MCP tool still runs only AMS-001 and AMS-004. - Fixed: the MCP server failed to start on fresh installs. The
mcpextra allowed any
mcp>=1.2.0, so installs since mcp 2.0 (July 2026) got mcp 2.x, which removed the
mcp.server.fastmcpmodule the server imports. The extra is now pinned tomcp>=1.2.0,<2. - Package links fixed: the project URLs in
pyproject.tomlused the repository's old name;
they now point at https://github.com/Echolonius/boardcheck and the live site. - Packaging for MCP hosts:
manifest.jsondescribes an MCP Bundle (.mcpb, uv runtime,
Python 3.10 or newer) that runs the MCP server, andserver.jsonlists it in the official MCP
Registry asio.github.Echolonius/boardcheck. A manual workflow
(.github/workflows/publish-mcp.yml) builds the bundle with the official MCPB CLI, starts it
and checks its tools, attaches it to the GitHub Release and publishes the registry entry. - Fixed version drift again: after 0.6.0,
__init__.__version__briefly said0.7.0while
pyproject.toml(and so--version) said0.6.0; both now say0.7.0.
MCP Bundle: boardcheck-mcp-0.7.0.mcpb (SHA-256 ba4ef0e6ac0895a35c0ade9ecbd9e8586279976e342f7c7c4b7a91dfbe7fa077)