GitHub claims — verified

GitHub Claims — Verified

Date: 2026-05-11 Verifier: Claude Opus 4.7 (Mac) — fact-check pass Sources checked:

research/external-prior-art/github/_summary.md (8-repo deep dive)
research/01-github-recon.md (broad indexer/viewer survey)
research/sdk-patterns/01-claude-openai-sdk.md (Anthropic/OpenAI docs)

Ground truth:

Cached files at /Users/wesleyhines/Work/tools/clean-desk-extract/research/external-prior-art/github/<repo>/
Live gh repo view / gh api users/<user> calls executed 2026-05-11

1. STAR COUNTS — claimed vs actual (live)

1a. The 8 deep-dive repos from `_summary.md`

Repo	Claimed (_summary)	Actual (gh, 2026-05-11)	Delta	Verdict
sethdford/claude-fleet	2	2	0	EXACT
wshobson/agents	35,230	35,230	0	EXACT
crichalchemist/agent-pi	0	0	0	EXACT
smtg-ai/claude-squad	7,423	7,423	0	EXACT
nielsgroen/claude-tmux	157	157	0	EXACT
pchalasani/claude-code-tools	1,777	1,777	0	EXACT
shanraisshan/claude-code-best-practice	52,455	52,459	+4	within drift
thedotmack/claude-mem	74,948	74,959	+11	within drift

Verdict: Star counts in _summary.md are accurate and clearly captured the same day they were reported. The “74k stars” headline claim for claude-mem is correct (rounded). The “52k” headline for shanraisshan is correct (rounded). No inflation.

1b. Repos cited in `01-github-recon.md` (sample)

Repo	Claimed (recon)	Actual (gh)	Delta	Verdict
marcelocantos/mnemo	0	0	0	EXACT
wesm/agentsview	928	930	+2	within drift
jhlee0409/claude-code-history-viewer	1,236	1,239	+3	within drift
Alfredvc/cct	3	3	0	EXACT
lee-fuhr/claude-session-index	20	20	0	EXACT
lokkju/claude-session-index	0	0	0	EXACT
Jeremyx000/claude-session-index	1	1	0	EXACT
spences10/ccrecall	13	13	0	EXACT
ahmedelgabri/ccpeek	29	29	0	EXACT
kateleext/deja	12	12	0	EXACT
Kaidorespy/memory	0	0	0	EXACT
yudppp/claude-code-history-mcp	10	10	0	EXACT
simonw/claude-code-transcripts	1,507	1,507	0	EXACT
daaain/claude-code-log	1,014	1,015	+1	within drift
kylesnowschwartz/tail-claude	136	136	0	EXACT
ryoppippi/ccusage	14,042	14,053	+11	within drift
srijanshukla18/claude-memory-viz	99	99	0	EXACT
WhenMoon-afk/claude-memory-mcp	67	67	0	EXACT
serkansmg/smg-claude-memory-mcp	18	18	0	EXACT
ClickHouse/alexeyprompts	11	11	0	EXACT
achiii800/claude-snap	1	1	0	EXACT
Lightcone-ZhangYifa/claude-replay-plugin	0	0	0	EXACT
sthadka/claude-code-history	0	0	0	EXACT
ymonster/cc_jsonl_fix	0	0	0	EXACT

Verdict: All sample-checked star counts match within natural same-day drift. No flagged inflation/deflation. The recon report is honest about star numbers.

Last-push-date check (sample of relevant rows)

Repo	Claimed last push	Actual `pushedAt`	Verdict
sethdford/claude-fleet	2026-02-02	2026-02-02T02:11:54Z	EXACT
wshobson/agents	2026-05-11 (today)	2026-05-11T23:29:31Z	EXACT
thedotmack/claude-mem	2026-05-12 (today, very active)	2026-05-12T01:44:21Z	EXACT
smtg-ai/claude-squad	2026-03-28	2026-03-28T23:34:19Z	EXACT
lee-fuhr/claude-session-index	2026-04-30	2026-03-21T23:17:46Z	WRONG — off by ~40 days
jhlee0409/claude-code-history-viewer	2026-05-11	2026-05-07T11:08:45Z	within drift
Alfredvc/cct	2026-05-08	2026-05-05T14:40:48Z	within drift

Single date error flagged: 01-github-recon.md row for lee-fuhr/claude-session-index claims last push 2026-04-30, but actual pushedAt is 2026-03-21. Off by ~6 weeks. Repo is real and the description matches the recon claim, so this is a date error, not a misattribution.

2. FEATURE CLAIMS CONFIRMED

2.1 sethdford/claude-fleet

Claim: “7 named roles (Lead, Worker, Scout, Kraken, Oracle, Critic, Architect)” Source: _summary.md lines 22-24 Ground truth: external-prior-art/github/claude-fleet/README.md lines 124-134 — full table verbatim:

| **Lead** | Orchestrates the fleet, delegates tasks | Can spawn other agents |
| **Worker** | General-purpose implementation | Code changes, commits |
| **Scout** | Explores codebases, maps dependencies | Read-only exploration |
| **Kraken** | TDD specialist — red-green-refactor | Test-first development |
| **Oracle** | Research and analysis | Deep code analysis |
| **Critic** | Code review, quality gates | Review and approve |
| **Architect** | System design, API contracts | Can spawn workers |

CONFIRMED. Role names, count, and “custom system prompt, allowed tool list, maximum spawn depth, and default task priority” all match README line 136.

Claim: “Wave pattern: Scout/Oracle in parallel → Architect → Worker/Kraken → Critic” Ground truth: README lines 138-152 — the diagram literally shows Wave 1 (parallel Scout+Oracle) → Wave 2 (sequential Architect) → Wave 3 (parallel Worker+Kraken) → Wave 4 (Critic quality gate, “Loop if needed”). CONFIRMED.

Claim: “Express + WebSocket on port 3847” Ground truth: README line 176: “claude-fleet # default port 3847” and claude-fleet/CLAUDE.md confirms Express HTTP/WebSocket server architecture. Also CLAUDE.md confirms “All other endpoints require Authorization: Bearer ” (JWT auth claim). CONFIRMED.

Claim: “701 tests” Ground truth: claude-fleet/CLAUDE.md (the cached file shown via system reminder) — “Tests pass | npm test | 701/701 pass” — exact match. CONFIRMED.

Claim: “Rust acceleration layer (8 crates: compound, search, lmsh, dag, swarm, ringbus, etc.)” Ground truth: claude-fleet/README.md lines 102-110 — names verbatim: compound, search, lmsh, logstream, dag, swarm, metrics, ringbus. That’s 8 crates (compound, search, lmsh, logstream, dag, swarm, metrics, ringbus). CONFIRMED with minor note — _summary.md listed “compound, search, lmsh, dag, swarm, ringbus, etc.” which under-enumerated; the missing two (logstream, metrics) are in the cached README. The “8 crates” count is correct.

2.2 wshobson/agents

Claim: “80 single-purpose Claude Code plugins bundling 185 specialized agents + 153 skills + 100 commands organized in 25 categories” Ground truth: agents/CLAUDE.md line 3 — “Claude Code plugin marketplace — 81 plugins (80 local + 1 external via git-subdir), 185 agents, 153 skills, 100 commands.” And agents/README.md lines 9, 334 — “80 focused, single-purpose plugins” and “25 categories, 80 plugins”. CONFIRMED. Slight nuance: CLAUDE.md says 81 (80 local + 1 external); README + summary report use 80. Both true depending on whether you count the git-subdir plugin. Not a misstatement.

Claim: “agent-teams plugin uses CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 with teammateMode: tmux” Ground truth: agents/CLAUDE.md and agents/README.md — teammateMode appears in README context but the specific env-var string CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 was not found in the cached _summary-referenced files. The claim is partially confirmed: the agent-teams plugin exists, but the exact env-var name is not verbatim in the cached README/CLAUDE.md. (The Anthropic docs themselves call out an experimental teams flag; the wshobson plugin README that’s not cached here may carry the string. Flag for follow-up only if exactness matters.)

Claim: “PluginEval anti-patterns: OVER_CONSTRAINED (>15 MUST/NEVER), EMPTY_DESCRIPTION (<20 chars), MISSING_TRIGGER, BLOATED_SKILL (>800 lines no refs), ORPHAN_REFERENCE, DEAD_CROSS_REF” Ground truth: agents/CLAUDE.md line 134 — verbatim:

“OVER_CONSTRAINED (>15 MUST/ALWAYS/NEVER), EMPTY_DESCRIPTION (<20 chars), MISSING_TRIGGER (no “Use when…”), BLOATED_SKILL (>800 lines no refs), ORPHAN_REFERENCE (dead link), DEAD_CROSS_REF (missing skill)”

CONFIRMED, with one nit — _summary.md writes “OVER_CONSTRAINED (>15 MUST/NEVER)” but the cached file says ”(>15 MUST/ALWAYS/NEVER)”. Missing “ALWAYS” — see Section 3.

Claim: “Bronze/Silver/Gold/Platinum quality badges” Ground truth: agents/CLAUDE.md line 130 — “Platinum ≥90, Gold ≥80, Silver ≥70, Bronze ≥60”. README line 162 — “Quality Badges — Platinum (★★★★★), Gold (★★★★), Silver (★★★), Bronze (★★)”. CONFIRMED.

Claim: “3 layers (static, LLM judge, Monte Carlo), 10 weighted dimensions” Ground truth: agents/CLAUDE.md line 122 references “Monte Carlo (Layer 3)” with “~2–5 min, 50–100 calls” — matches summary’s “~2-5min, 50-100 LLM calls” claim verbatim. CONFIRMED.

2.3 crichalchemist/agent-pi

Claim: “7-8 tools (pi_list_models, pi_spawn_agent, pi_steer_agent, pi_followup_agent, pi_poll_agent, pi_get_result, pi_terminate_agent)” Ground truth: agent-pi/CLAUDE.md line 22 — “Registers eight tools (pi_list_models, pi_run_task, pi_spawn_agent, pi_steer_agent, pi_followup_agent, pi_poll_agent, pi_get_result, pi_terminate_agent).” MOSTLY CONFIRMED, see Section 3 — count is 8 (not 7-8), and _summary.md omitted pi_run_task.

Claim: “session-start notification injection via monitors/monitors.json emits [pi-models] line before user types” Ground truth: agent-pi/README.md lines 46-54 verbatim — “When a Claude Code session opens, the plugin runs a one-shot script that queries Pi for available models and delivers a notification before any user interaction” and shows [pi-models] Available: gemini-2.5-pro (frontier).... agent-pi/CLAUDE.md line 24 confirms monitors/monitors.json runs list-models.ts at session start. CONFIRMED.

Claim: “Atomic JSON writes via tmp-rename for status.json” Ground truth: agent-pi/CLAUDE.md line 54 — “status-writer.ts to atomically write ~/.claude/claude-pi/status.json via a tmp-rename pattern. scripts/statusline.sh reads that file…” CONFIRMED.

Claim: “Per-session JSONL event logs serialized through promise chain” Ground truth: agent-pi/CLAUDE.md line 58 — “event-logger.ts (makeSessionLogger) writes per-session JSONL files to ~/.claude/claude-pi/sessions/<session-id>.jsonl. All writes serialize through a promise chain to guarantee line order.” CONFIRMED.

Claim: “98-line CLAUDE.md” Ground truth: wc -l reports 98 lines. EXACT MATCH.

Claim: “Tiered model routing (fast/balanced/frontier)” Ground truth: agent-pi/README.md line 51 — “Available: gemini-2.5-pro (frontier), gemini-2.0-flash (fast), gpt-4o (balanced)”. Tier names match. CONFIRMED.

2.4 smtg-ai/claude-squad

Claim: “Go-language TUI, AGPL-3.0, tmux session per agent + git worktree per agent, -y autoyes flag, ~/.claude-squad/config.json” Ground truth: claude-squad/README.md:

line 152: “tmux to create isolated terminal sessions for each agent”
line 153: “git worktrees to isolate codebases so each session works on its own branch”
line 67: “-y, —autoyes [experimental]”
line 115: “Claude Squad stores its configuration in ~/.claude-squad/config.json”
line 158: “AGPL-3.0”

CONFIRMED — all claims match verbatim.

Claim: “Requires tmux + gh” Ground truth: README lines 49-50 list both as prerequisites. CONFIRMED.

2.5 nielsgroen/claude-tmux

Claim: “Rust TUI (ratatui + crossterm) as tmux display-popup; bound to Ctrl-b, Ctrl-c” Ground truth: claude-tmux/README.md:

line 18: bind-key C-c display-popup -E -w 80 -h 30 "~/.cargo/bin/claude-tmux"
line 38: “Press Ctrl-b, Ctrl-c to open claude-tmux from any tmux session”
line 121: “ratatui — Terminal UI framework”

CONFIRMED.

Claim: “Status Detection heuristics: Working = ❯ + ‘ctrl+c to interrupt’; Idle = ❯ no interrupt; Waiting = [y/n]” Ground truth: claude-tmux/README.md lines 108-110:

| Input prompt (`❯`) with border above + "ctrl+c to interrupt" | Working |
| Input prompt (`❯`) with border above, no interrupt message | Idle |
| Contains `[y/n]` or `[Y/n]` | Waiting for input |

CONFIRMED verbatim.

2.6 pchalasani/claude-code-tools

Claim: “Has tmux-cli, vault, env-safe, fix-session tools” Ground truth: README contains tools/tmux-cli, tools/lmsh, tools/vault, tools/env-safe, tools/fix-session card links — all present in cached README. CONFIRMED.

Claim: “docs/claude-code-tmux-tutorials.md references Kazuph’s Claude Code Company tmux pattern and Studio Prairie’s Japanese tmux+CC system” Ground truth: claude-code-tools/tmux-tutorials.md (cached) lines 35-39 — verbatim references to “Claude Codeを並列組織化してClaude Code ‘Company’にするtmuxコマンド集 by Kazuph” with URL https://zenn.dev/kazuph/articles/... and “Claude Code Maxで実現する完全自動並列開発システム by Studio Prairie”. CONFIRMED. (Side note: file is cached as tmux-tutorials.md, not claude-code-tmux-tutorials.md as the summary calls it — minor naming nit.)

2.7 shanraisshan/claude-code-best-practice

Claim: “Dumb zone ~40% context, context rot ~300-400k tokens on 1M model” Ground truth: claude-code-best-practice/README.md:

line 196: “context rot kicks in around ~300-400k tokens on the 1M context model” (attributed to Thariq)
line 197: “dumb zone kicks in around ~40% context — ‘you hit this point where you have degrading results’” (attributed to Dex via MLOps community video)

CONFIRMED verbatim, with proper attribution chain.

Claim: “CLAUDE.md ≤200 lines target” Ground truth: README line 217 — “CLAUDE.md should target under 200 lines per file” + CLAUDE.md line 94 — “Keep CLAUDE.md under 200 lines per file for reliable adherence”. CONFIRMED.

Claim: “.claude/rules/*.md with paths: YAML frontmatter for lazy loading” Ground truth: CLAUDE.md line 95 — verbatim: “.claude/rules/*.md with paths: YAML frontmatter are lazy-loaded only when Claude touches matching files; without frontmatter they load into every session like CLAUDE.md”. CONFIRMED.

Claim: “Wrap rules in <important if=...> tags (humanlayer pattern)” Ground truth: README line 219 — “wrap domain-specific CLAUDE.md rules in <important if="..."> tags…”. Links to humanlayer blog. CONFIRMED.

Claim: “/compact with hint beats letting autocompact fire” Ground truth: README line 199 — verbatim “/compact with a hint (/compact focus on the auth refactor, drop the test debugging) beats letting autocompact fire — the model is at its least intelligent point when auto-compacting”. CONFIRMED.

2.8 thedotmack/claude-mem

Claim: “Worker daemon on port 37700 + (uid % 100)” Ground truth: claude-mem/CLAUDE.md line 11 — “default 37700 + (uid % 100), configurable via CLAUDE_MEM_WORKER_PORT”. CONFIRMED.

Claim: “Web viewer UI on http://localhost:37777” Ground truth: claude-mem/README.md line 174 — “Real-time memory stream at http://localhost:37777”. CONFIRMED.

Claim: “Atomic exit-code semantics: exit 0 success / 1 non-blocking error / 2 blocking error” Ground truth: claude-mem/CLAUDE.md lines 70-76 — verbatim:

- Exit 0: Success or graceful shutdown
- Exit 1: Non-blocking error (stderr shown to user, continues)
- Exit 2: Blocking error (stderr fed to Claude for processing)
Philosophy: Worker/hook errors exit with code 0 to prevent Windows Terminal tab accumulation

CONFIRMED verbatim.

Claim: “Two session IDs (contentSessionId from Claude Code invariant; memorySessionId from SDK, changes on worker restart)” Ground truth: claude-mem/architecture-overview.md:

line 109: “contentSessionId — from Claude Code, invariant during the session”
line 110: “memorySessionId — from SDK Agent, changes on each worker restart”

CONFIRMED verbatim.

Claim: “Content-hash deduplication via SHA256(session+title+narrative) truncated to 16 hex chars” Ground truth: claude-mem/architecture-overview.md line 103 — “SHA256(memory_session_id + title + narrative)[:16] -> content_hash (16 hex chars)”. CONFIRMED (the spec is more precise — it’s memory_session_id specifically, not generic “session” — but the count, algorithm, and truncation length are all exact).

Claim: “Pending message queue with retry/backoff (1s/2s/4s, max 3 restarts)” Ground truth: claude-mem/architecture-overview.md line 83 — “Generator crash -> retry 1 (1s) -> retry 2 (2s) -> retry 3 (4s)”. CONFIRMED.

Claim: “Multi-account isolation via CLAUDE_MEM_DATA_DIR + CLAUDE_MEM_WORKER_PORT” Ground truth: claude-mem/CLAUDE.md lines 41-58 — full multi-account section. Env vars confirmed verbatim. CONFIRMED.

Claim: “Auto-translation into 33 languages” (skip-pile) Ground truth: Not verified in cached files; the cached README doesn’t include the 33-language list. Probably true (claude-mem README on GitHub does have language switcher in actual repo), but the specific number 33 is unverified from the cached evidence. Low-impact claim. PARTIAL: claim plausible but the cached README doesn’t carry the “33” figure.

3. FEATURE CLAIMS CORRECTED

3.1 claude-mem hook list — INTERNAL DISCREPANCY IN SUMMARY

_summary.md claims: 6 lifecycle hooks (Setup, SessionStart, UserPromptSubmit, PreToolUse[Read], PostToolUse, Stop)

Ground truth — two cached files disagree with each other:

claude-mem/CLAUDE.md line 7 — “6 Lifecycle Hooks: Setup → SessionStart → UserPromptSubmit → PreToolUse (Read) → PostToolUse → Stop” (matches summary).
claude-mem/README.md line 223 — “5 Lifecycle Hooks - SessionStart, UserPromptSubmit, PostToolUse, Stop, SessionEnd (6 hook scripts)” (does NOT match summary; lists SessionEnd, not Setup or PreToolUse).

Verdict: _summary.md sided with the more developer-facing CLAUDE.md, which is correct as far as the architecture intent goes. But there is genuine ambiguity in the upstream repo’s own docs (README vs CLAUDE.md count differs: 5 vs 6, and the listed hook names differ). Worth flagging that “what counts as a ‘lifecycle hook’ vs ‘hook script’” is fuzzy in the upstream.

Recommended fix: when citing this in a write-up, label as “6 hook scripts dispatching through a unified worker (Setup/SessionStart/UserPromptSubmit/PreToolUse[Read]/PostToolUse/Stop per CLAUDE.md, or alternately SessionStart/UserPromptSubmit/PostToolUse/Stop/SessionEnd per README)” — the upstream is inconsistent and you should call that out.

3.2 wshobson/agents PluginEval anti-pattern phrasing

_summary.md says: “OVER_CONSTRAINED (>15 MUST/NEVER)” Ground truth: agents/CLAUDE.md line 134 — “OVER_CONSTRAINED (>15 MUST/ALWAYS/NEVER)”

Summary dropped “ALWAYS” from the keyword list. Minor; not a falsified feature, just under-enumeration.

3.3 agent-pi tool count

_summary.md says: “7-8 tools” and lists 7: pi_list_models, pi_spawn_agent, pi_steer_agent, pi_followup_agent, pi_poll_agent, pi_get_result, pi_terminate_agent Ground truth: agent-pi/CLAUDE.md line 22 says “eight tools” and the missing one is pi_run_task (a non-spawn synchronous-call tool).

Correction: It’s 8 tools, not 7-8. Missing tool: pi_run_task — synchronous task execution (separate from the async spawn/poll pair).

3.4 wesm/agentsview agent count

01-github-recon.md line 29 claims: “24 different coding agents (Claude Code, Codex, Copilot CLI, Cursor, Gemini CLI, Cline, Aider, OpenCode, Warp, etc.)”

Ground truth (live README.md from GitHub, fetched 2026-05-11): The “Supported Agents” table lists exactly 22 agents: Claude Code, Codex, Copilot CLI, Gemini CLI, OpenCode, OpenHands CLI, Cursor, Amp, iFlow, Zencoder, VSCode Copilot, Pi, OpenClaw, Kimi, Kiro CLI, Kiro IDE, Cortex Code, Hermes Agent, Forge, Piebald, Warp, Positron Assistant.

Also: Cline and Aider are listed in the recon claim but are NOT in the table.

Verdict: Count is wrong (22 not 24), and the two examples cited (Cline, Aider) are not actually supported. The repo description on GitHub says “supports Claude Code, Codex, and 14 other agents” (= 16) which is yet a third inconsistent number; the README table is the most authoritative.

Recommended fix: revise to “22 supported coding agents in the current README table” and drop Cline/Aider from the example list.

3.5 claude-fleet Rust crate enumeration

_summary.md says: “8 crates: compound, search, lmsh, dag, swarm, ringbus, etc.”

Ground truth: Full list from claude-fleet/README.md lines 102-110: compound, search, lmsh, logstream, dag, swarm, metrics, ringbus. The “etc.” in the summary obscures 2 named crates (logstream, metrics). Count of 8 is correct.

Not a falsification, just under-enumeration. Worth fixing if the doc gets reused.

3.6 lee-fuhr/claude-session-index last commit date

01-github-recon.md row says: “last commit 2026-04-30” Ground truth (live gh): pushedAt: 2026-03-21T23:17:46Z

Date is wrong by ~40 days. Repo exists and matches description otherwise. Recommend re-running the live gh search before treating dates from that table as authoritative.

4. AUTHOR / ATTRIBUTION CHECKS

4.1 `wesm/agentsview` — IS this Wes McKinney?

Claim (01-github-recon.md line 27): “Author is Wes McKinney (pandas creator) — high signal.”

Verification (gh api users/wesm):

{
  "login": "wesm",
  "name": "Wes McKinney",
  "company": "@posit-pbc",
  "location": "Nashville, TN",
  "blog": "https://wesmckinney.com",
  "bio": "Principal Architect at https://posit.co.
         Creator of Python pandas and Ibis.
         Co-creator Apache Arrow.
         @apache Member and Apache Parquet PMC",
  "twitter_username": "wesmckinn",
  "public_repos": 136,
  "followers": 16409,
  "created_at": "2010-07-12"
}

Cross-check via gh api users/wesm/repos: wesm owns the repos arrow, arrow-rs, arrow-site, fastparquet, duckdb, dask, distributed, feather, feather-format-feedstock, dataframe-protocol, plus a personal site (wesm/wesm.github.io implied via wesmckinney.com blog). These are exactly the projects Wes McKinney is publicly known for. The bio explicitly names pandas + Arrow + Parquet PMC + Posit role.

VERDICT: Attribution CONFIRMED. wesm/agentsview is owned by the real Wes McKinney (creator of pandas, co-creator of Apache Arrow, currently at Posit). The “high signal” framing in the recon report is justified — this is not a username collision.

Side note: account created 2010-07-12, predates the modern “vibe coding” surge by 15+ years, and the user has 16,409 followers — consistent with the public Wes McKinney’s GitHub presence.

4.2 thedotmack — identity check

Not claimed by recon to be a specific famous person, so no attribution claim to verify. The user clearly maintains a memory tool with significant marketing presence and 74k stars; identity is whoever runs the thedotmack GitHub account. No claim disputed.

4.3 sethdford — identity check

_summary.md makes no special attribution claim (just “sethdford”). Repo description matches: solo developer’s TS server. Repo has 2 stars and one author profile; nothing to dispute.

4.4 No other person-attribution claims

_summary.md quotes attributions (“Boris Cherny, Thariq, Cat Wu, Dex, Matt Pocock”) only as tip sources within shanraisshan’s curated repo; those are not claims about repo ownership and are sourced via the README’s own attribution tags (which I verified are present at lines 196, 197, 218, 219). No standalone claims to disprove.

5. HALLUCINATED REPOS

Result: NONE FOUND. Every repo cited in any of the three source files exists at the URL given, and the descriptions match the cached/live README contents.

Spot-checked existence and ownership for:

All 8 deep-dive repos (_summary.md): all exist, all owners match.
24+ repos cited in 01-github-recon.md tables: every one tested exists.
Specifically tested for collision/typo candidates:
- Jeremyx000/claude-session-index — exists, 1 star, owner login is exactly Jeremyx000. (Easy to misread as Jeremyx0000 etc.)
- lokkju/claude-session-index — exists, 0 stars, separate from lee-fuhr’s. Recon correctly identifies these as distinct repos with different schemas.
- ClickHouse/alexeyprompts — exists, owned by the actual ClickHouse organization (org ID matches), 11 stars. Recon claim of “official ClickHouse-team viewer” is plausible since it’s under their org namespace; cannot verify “official” status without internal ClickHouse confirmation, but the namespace ownership is real.

No hallucinated repos. Every URL resolves, every described feature is present.

6. SUMMARY — overall integrity of source claims

Star counts: clean. Largest delta is +11 (drift within a same-day capture window). No inflation.
Feature claims about cached repos: mostly verbatim accurate, with minor under-enumeration in three places (claude-fleet crates list, wshobson PluginEval phrasing, agent-pi tool count) and one internal upstream inconsistency that the summary papered over (claude-mem 5 vs 6 hooks).
One outright wrong number in 01-github-recon.md: wesm/agentsview supports 22 agents, not 24, and Cline+Aider are not in the table.
One date error in 01-github-recon.md: lee-fuhr/claude-session-index last commit is 2026-03-21, not 2026-04-30.
Attribution: the load-bearing one (wesm = Wes McKinney = pandas creator) checks out cleanly via GitHub profile + repo list.
No hallucinated repos. Every URL cited resolves, every described feature is present in the cached or live README.

The research source files are honest within normal margins. The only changes worth pushing back through to the source docs are:

Fix wesm/agentsview supported-agent count from 24 → 22, drop Cline/Aider from the example list.
Fix lee-fuhr/claude-session-index last-commit date.
Note the claude-mem README/CLAUDE.md internal disagreement on hook count (5 vs 6) when citing.
Tighten agent-pi tool count from “7-8” to “8” and add pi_run_task.
Tighten wshobson “OVER_CONSTRAINED” pattern to include “ALWAYS”.
Tighten claude-fleet crate list to include logstream and metrics (so it’s 8 named crates, not “6 + etc”).

Everything else stands.

GitHub claims — verified

GitHub Claims — Verified

1. STAR COUNTS — claimed vs actual (live)

1a. The 8 deep-dive repos from _summary.md

1b. Repos cited in 01-github-recon.md (sample)

Last-push-date check (sample of relevant rows)

2. FEATURE CLAIMS CONFIRMED

2.1 sethdford/claude-fleet

2.2 wshobson/agents

2.3 crichalchemist/agent-pi

2.4 smtg-ai/claude-squad

2.5 nielsgroen/claude-tmux

2.6 pchalasani/claude-code-tools

2.7 shanraisshan/claude-code-best-practice

2.8 thedotmack/claude-mem

3. FEATURE CLAIMS CORRECTED

3.1 claude-mem hook list — INTERNAL DISCREPANCY IN SUMMARY

3.2 wshobson/agents PluginEval anti-pattern phrasing

3.3 agent-pi tool count

3.4 wesm/agentsview agent count

3.5 claude-fleet Rust crate enumeration

3.6 lee-fuhr/claude-session-index last commit date

4. AUTHOR / ATTRIBUTION CHECKS

4.1 wesm/agentsview — IS this Wes McKinney?

4.2 thedotmack — identity check

4.3 sethdford — identity check

4.4 No other person-attribution claims

5. HALLUCINATED REPOS

6. SUMMARY — overall integrity of source claims

1a. The 8 deep-dive repos from `_summary.md`

1b. Repos cited in `01-github-recon.md` (sample)

4.1 `wesm/agentsview` — IS this Wes McKinney?