agenthropic

Methodology

Mandate

Ivan’s instruction was explicit: before any action, produce a very detailed, holistic, merciless analysis. The two vendor reports (Claude-Code-Agent-Dashboards_Due-Diligence.docx and its v2.0 extended) are treated here as an input to be audited, not a conclusion to be trusted. Where the report’s claims and the source code diverge, the source wins.

Scope

Six candidates were audited:

1. simple10/agents-observe
2. hoangsonww/Claude-Code-Agent-Monitor
3. ek33450505/claude-code-dashboard (aka “CAST”)
4. disler/claude-code-hooks-multi-agent-observability
5. NirDiamant/claude-watch
6. davila7/claude-code-templates — not in the reports; added because the panel missed it (see market-landscape.md).

Process

1. Report extraction. Both .docx files were unzipped and their word/document.xml parsed to clean text (split on </w:p>, extract <w:t> runs). The v1 report carried the A−/B+/B+/B−/C+/F grades and the “adopt hoangsonww” call; v2.0 added §4 the scoring model, §9 per-project deep dives, §8.5 OPCⁿ fit, §10 gap analysis, §11 recommendations.
2. Source audit. Each repo was shallow-cloned and read at source by an independent auditor, in parallel. Auditors were told to verify — not assume — every headline claim: subagent hierarchy, DAG reality, test depth, security posture, license, maintenance, extractable value.
3. Fact verification. Every quantitative claim (stars, forks, last-push, LOC, table count, test/assertion count) was re-pulled from the GitHub API on 2026-07-03 via gh. Graph libraries and security patterns were confirmed by grep (d3, cytoscape, dagre, reactflow; auth/CORS/bind/spawn).
4. Grading. Independent letter grades were assigned from the source findings, then compared against both the panel’s letters (v1) and its weighted /5 scores (v2).

Tools

• gh CLI for live repo metadata and issue/PR verification
• shallow git clone + LOC counting + test-file / expect() (assertion) counting
• grep sweeps for graph libraries, bind/0.0.0.0, auth guards, execSync, spawners
• .docx → text extraction (Python, zipfile + regex over word/document.xml)

The verified fact base (2026-07-03)

Project	Stars	Forks	Last push	License (real)
hoangsonww	(per report, verified in range)	—	active	MIT + LICENSE file
disler	1,475	385	stalled 2026-02-08	none (private:true, no license)
simple10	607	58	2026-06-29	MIT + LICENSE (© 2025 Joe Johnston)
cast	3	—	active	README badge only, no LICENSE file
nirdiamant	—	—	active	MIT declared, no LICENSE file
claude-code-templates	28.4k	—	active	MIT + LICENSE
(context) claude-hud	26.1k	—	active	MIT
(context) ccusage	16.8k	—	active	MIT

Where the report’s “verified fact base” cracked

• simple10 was reported as “blocked by API rate-limiting, last activity 5 Jun.” It returned on the first call: 607★/58/MIT, pushed 29 Jun. The report’s numbers were stale and its “no true DAG” claim (built on that stale read) is false.
• disler was reported at “~1,400★ / 372 forks” — real is 1,475 / 385.
• disler hook scripts were reported to “protect .env/keys” — the guard is commented out (pre_tool_use.py:324-327).

Limits & caveats

• These repos move fast; star counts and last-push dates drift. Re-verify before a final commitment.
• Pricing tables inside cost features (e.g. cast’s delegation-savings) are hardcoded and may be stale — re-check model rates before trusting dollar figures.
• Subagent edges are, in every candidate, derived from the event stream at render time, not persisted as first-class rows. This matters for a global/historical DAG and is called out per-project.
• The audit reads code, not running instances; runtime behaviour (esp. arm64 native bindings) should be smoke-tested on the actual Mac Mini before commitment.

This site is open source. Improve this page.