Seven Additions and One Revision

Our resource catalog now lists 230 entries: seven new tools for recording, verifying, or structuring agent work, one revised paper, and no removals.

  • Tracekit: a signer outside the agent allows, denies, or holds intercepted tool calls for human approval and signs each run into an offline-verifiable Ed25519 hash chain. Calls bypassing its hooks, adapters, and gateway go unrecorded; in default dev mode, a compromised agent can reach the key. Code.
  • SealedRun: a recorder holding keys outside the agent signs proxied LLM, MCP, and A2A calls and SDK self-reports into an offline-verifiable hash chain with delegation and approval checks. Steps skipping the recorder go unrecorded; self-reports remain the agent's account. Code.
  • meilynx-verify: an offline verifier for exported Meilynx audit logs that checks hashes, the SHA-256 chain, Ed25519 signatures, and RFC 3161 timestamps, and reports the first break. An intact chain does not show completeness. Code.
  • mnema: a CLI and MCP server that commits a coding agent's decisions to the repository as a hash chain with Ed25519-signed checkpoints. It is tamper-evident, not tamper-proof: cutting the newest events with their checkpoint goes undetected. Code.
  • deja-vu: a local CLI that indexes coding-agent transcripts stored on disk, lists at most three sessions that edited or mentioned each changed file, covering the first ten files in path order by default (file evidence, not proof of authorship), and reports credentials in transcripts. Code.
  • Agent Role Contracts: an offline checker for declared agent roles, task scopes, review separation, and handoffs. A pass grants no execution authority. Code.
  • Ponder: a Node.js CLI that logs coding-agent reviews in an append-only ledger and refuses to stop while declared units remain ungraded; its convergence verdicts are advisory. Code.

The arXiv 2607.13304 record now uses the revised title, Query Language Sensitivity in Automated Sentiment Scores of LLM Responses. The revision finds its variance decomposition highly sensitive to including English answers, identifies no causal English effect, adds no human validation, and drops earlier brand-ranking and query-budget advice. Paper.

Pull requests to the companion list proposed all seven: #47 (suirindo), #48 (Quinnod345), #49 (LoopGlitch26), and #50 (glanitcom), merged November 5, and #51 (vshulcz), #52 (mlx-cassio), and #55 (felipesauer). Each submitter is the project's author, maintainer, or disclosed affiliate. Inclusion does not certify a tool or its security claims. This date marks catalog publication, not tool releases. Updates appear in our RSS / Atom feed.