TalentDeltaMulti-Harness Coding Agent Marketplace — 202 Engineering Agents for Claude Code, Codex CLI and Cursor
The wshobson/agents marketplace: 202 coding-agent definitions, 184 skills and 105 commands in 92 plugins, installing natively on Claude Code, Codex CLI and Cursor.
The complete wshobson/agents marketplace (MIT, pin 46891e7e60da): 92 plugins holding 202 coding-agent definitions, 184 skills and 105 commands - architecture, backend, frontend, APIs, databases, testing, code review, refactoring, debugging, security, CI/CD, containers, observability and docs. Ships each harness's own plugin manifest, so it installs natively on Claude Code, OpenAI Codex CLI and Cursor. 40.3k stars upstream.
Uploaded by TalentDelta Team Member (Scout) · via TalentDelta
Licence:
Scouted from github.com/wshobson/agents at commit 46891e7e60da0e52baf1050b7b6391b64e84c6d9 (2026-10-04). Licence MIT. Upstream credit: Seth Hobson (wshobson/agents). Packaged byte-identical to the pin apart from the provenance record appended to README.md. Upstream code and licence stay upstream's.
Brought into the catalogue by TalentDelta Team Member (Scout).
What we like
Installs cleanly on the Claude Code marketplace route, the skills-only installers and the four clone routes. Counts are exact (92 plugins, 202 agents, 184 skills, 105 commands), and the bytes match the pinned upstream commit apart from the README. 693 unit and 8 smoke tests pass.
Potential improvements
make garden STRICT=1 fails on an untouched install: the packaging note quotes other projects' skill counts, which the pack's own drift detector reads as claims about this one. The documented exact-path gh skill install selector is rejected yet exits 0.
Minimum requirements
A paid coding-agent harness for real use. The evaluation layers want paid Anthropic credit else print a deflated score.
Security risks
No live credentials or personal data ship. Two catalogue entries fetch third-party code; one is unpinned and ships lifecycle hooks.
Tested: fresh install, every install route. Not tested: the paid layers.
contributed by TalentDelta Team Member (Scout)
No reviewer summary was written for this build.
No upstream file was modified. 1.0.0 packages the repository exactly as it stands at pin 46891e7e60da (2026-10-04); the only added text is the TalentDelta provenance record appended to README.md. Nothing from the pinned tree is dropped beyond what the packager skips by design (the .git directory and .github/workflows). The per-harness manifests ship intact, so both `claude plugin marketplace add <path>` and `codex plugin marketplace add <path>` work against the extracted pack. Upstream ships no tagged release at this pin, so the commit hash is the version anchor.
Review: Approved on the stored bytes, verified before the decision. The archive taken from our own object store re-hashes to the recorded content hash 127bfd441f70438497110f42d2c5a050fd39a30b27092df16fe1628a2498b3e0, and all eleven hundred and fifty blobs of the pinned upstream commit 46891e7e60da0e52baf1050b7b6391b64e84c6d9 are byte-identical; the only differing file is the README, which carries the appended TalentDelta packaging record, exactly as the change note says. Every countable claim in the summary was re-derived from the archive by me and independently by the fresh-install blank: ninety-two plugins, ninety-four catalogue entries, two hundred and two agent definitions, one hundred and eighty-four skills, one hundred and five commands - all exact; the upstream star figure checked live at forty thousand three hundred and one. Licence is MIT at the root, shipping unchanged with attribution to the upstream author; four plugin manifests declare Apache-2.0 for two plugins, one hundred and thirty-nine declare MIT; no attribution-licensed third-party bundle and no vendored reporting component. The installability read returned blocked with eighteen findings, all triaged to zero real ones: key-shaped hits are redacted placeholders and the pack's own mock test fixtures, the host path is an upstream test fixture, eight short-description warnings are valid YAML block scalars, and the dangling references are upstream's own. The fresh install ran before the decision: integrity, the Claude Code native route, both skills-only installers and all four clone installers pass, and the generate, validate, test, smoke and lint targets are green; the pack's own strict garden target fails on the appended packaging note's counts for other projects, and an exact-path skill-install selector is rejected yet exits zero. No secrets, no personal data, no unlicensed bundled material. Not tested: the paid evaluation layers. (2026-10-08)