Merciless, multi-lens review of the agenthropic idea itself (not another audit of
the six rival dashboards — that lives in ../due-diligence/),
plus the build plan it produces.
Fresh session with no context? Start with PROJECT-STATE-2026-07-06.md — the single entry point: full timeline, document map with authority order, current-truth snapshot, the pending decision, and “if Ivan says X, do Y” playbooks.
Read in order: v1 analysis → external review → v2 consolidation → development plan → best-path decision → Phase-0 probe → corpus audit → red-team counter-analysis → v1/v2 roadmap (final).
| File | What’s in it |
|---|---|
| concept-analysis.md | v1 — ⚠️ superseded by v2 (kept for the audit trail). The detailed report: Senior Architect · Developer · QA · Business Analyst reviews, a brutal gap analysis, a holistic read, consolidated strengths/weaknesses, and a deep dimension-by-dimension analysis with a risk register. |
| implementation-plan.md | v1 plan — ⚠️ superseded (kept for the audit trail; uses the OLD phase numbering — see PROC-6). Part A’s D1–D7 were re-resolved as LB1/LB2 + CD-1…CD-10 in v2; Part B was replaced by development-plan.md as amended by best-path-decision.md §6. |
| external-docs-review.md | Adversarial review (5 lenses) of two externally-produced parallel reports (BASE + EXPANDED). Verdict: both sound, both pass the tripwires; BASE more precise, EXPANDED more formalized but defect-carrying. Establishes the synthesis rule for v2. |
| concept-analysis-v2.md | v2 — authoritative. Re-run of the six lenses folding in v1 + BASE + EXPANDED + the review. Resolves the two load-bearing decisions (LB1 ingest primacy, LB2 personal-first/commercial-clean) and the ten canonical decisions CD-1…CD-10 with quantified acceptance criteria. |
| development-plan.md | The build. CD-1…CD-10 decomposed into agent-distributable work packages (owner-agent, inputs/outputs, dependencies, acceptance criteria), a verified dependency-DAG with parallel waves, a critical path, and a phase-by-phase roadmap with exit gates. Open work → ../../TODO.md. |
| best-path-decision.md | Strategic decision memo — sits above the plan. “What is actually best to build, and how.” Output of an adversarial workflow (6 theses × 3 critics + judge, 25 agents) with the two load-bearing empirical claims hand-verified: moat-first greenfield spine + attributed simple10 tree; the empirical CONDITIONAL-GO correction to CD-1; keep/change edits to the plan (alerts off the critical path, delete vector-DB/dual-driver, keep outbox+backfill); a 6-risk register and honest kill-the-program dissent. Confidence 76/100. |
| phase0-probe.md | The evidence. Empirical read-only probe of the real ~/.claude/projects corpus (2026-07-04 census: 17 projects · 117 sessions · 33 subagent dirs · 148 flat + ~849 nested agent files — 85.2% nested; superseded by parser-spec.md §4.2 — 20 slugs · 141 sessions · 1855 subagent transcripts · 54 sessions with subagents — quote §4.2, not this row) — layout/version census, depth-1 hard-key join, depth-2 recoverability, token reconciliation — producing the evidence-backed CD-1 verdict with measured numbers. The do-this-now action of the best-path memo, executed. |
| animated-room-analysis.md | Feature-scoped multi-lens review (Architect · BA · QA · gap analysis · devil’s advocate) of the animated-office/room idea (DESIGN §9 Phase 1.5). Source-verifies the three reference tools (pixel-agents MIT, my-virtual-office AGPL+spawner, claw3d), and lands on adopt-not-build, gated-behind-moat, read-only. |
| corpus-audit-2026-07-06.md | Full-corpus consistency audit (self-contained for a context-free session). Stable finding register (AMEND unapplied amendments · OPEN unowned decisions · LOST dropped source material · EMP empirical gaps · PROC hygiene), a 33-decision ledger, the docs-site C1–C19 list, five senior-lens reviews incl. the first-ever UX/UI pass, and 10 ranked actions. Headline: best-path §6 is still unapplied to the plan/TODO (AMEND-1…6). |
| recovered-source-material.md | Recovered vendor material (closes corpus-audit findings LOST-1…4): the 24-capability feature matrix (v2 §12, verbatim ✓/~/✗ across all five rivals) re-framed as an agenthropic UI checklist; the adjacent-tools note (v2 §13 — Claude Code’s stable OTel query_source = main/subagent/auxiliary signal, ccusage, LiteLLM, Anthropic Console); the risk register (v2 §16 — incl. hook-schema-drift M/H and the hoangsonww CLA note); and the cost-of-ownership developer-day sizing (v2 §17). Also pins LOST-8: the “no true DAG” quote is not in the sources — do not propagate. |
| ux0-design.md | WP-UX0 design pre-work (closes corpus-audit LOST-6 + §9.5 — the first UX/UI artifact): IA map, the five question-to-screen flows, ASCII wireframes for the four views, and the uncertainty/honesty visual language (inferred edges, estimated costs, 'unknown' status always visible). |
| open-decisions.md | The OPEN-1…9 decision register — each unowned decision from the corpus audit with its recommended resolution, framed as sign-off checkboxes. Deciding them stays with Ivan. |
| red-team-audit-2026-07-06.md | Deliberately adversarial counter-analysis — attacks every premise (defused kill-condition, moat-on-rented-land, n=1 evidence, governance fiction, economics), records what survives (§9), and forces a choice between three exits (kill · brutal two-week timebox · status quo) with an explicit stop condition on further analysis (§11). Read after the corpus audit. |
| roadmap-v1-v2-2026-07-06.md | The last analysis (#9 of 9) — schedule of record. Authored on explicit owner instruction overriding red-team §11. Re-measured numbers (76 files · 141k words · 0 commits), three velocity scenarios against the 63-WP v1.0 path, kill checkpoints KC-0…KC-5 with default-death, the detailed phase-by-phase v1.0 roadmap (Exit B absorbed into Phase 0 inside CD-8, descope ladder, forbidden descopes), the earned-not-scheduled v2.0 alerts roadmap, and the analysis freeze — after it, only verdict records. |
| phase0-verdict.md | Post-freeze verdict record #1 (the WP-S7 GO/NO-GO). The Phase-0 feasibility spike (WP-S1…S7) executed end-to-end against a hostile 5-session / 224-agent corpus: verdict CONDITIONAL GO ~90% — the subagent DAG + dollar cost is mechanically reconstructable from ~/.claude/projects/*.jsonl alone, zero inference. Six-row evidence table, the 11-item gate’s final status, the three new parser MUSTs, the first velocity number, and what still needs Ivan. All numbers are self-check / PROVISIONAL until the five LABEL-ME.md trees are hand-filled. |
| parser-spec.md | Post-freeze — the parser contract distilled from the verdict. Normative, implementation-ready: the 14-item requirements gate (original 11 + three MUSTs the spike found — <task-notification> flat join, queue-operation 3rd join schema, message.id usage dedup + bucket/model pricing), the four structural join paths, token→cost rules, the self-referential depth-2 tree index, the amended EMP-1 wave-partial ordering, and the exact site-doc edits to fold into the published pages. Normative and IMPLEMENTED (2026-08-15) — see its §3 for what is implemented (14 of 14) versus what the real corpus actually exercises (11; #11 amended, #7 and N1 witnessed only by fixtures). Also the census of record (§4.2) and the duplicate-session-uuid rule (§4.3). |
| impl-review-2026-08-09.md | Post-freeze review record (owner-ordered, 2026-08-09). Reviews the shipped implementation at 2f8d103 across ten dimensions with CONFIRMED / PLAUSIBLE / UNVERIFIED-LOW verdicts; explicitly not design-analysis #10 (the 9-of-9 freeze stands); all spike/bench numbers PROVISIONAL pending LABEL-ME. |
| PROJECT-STATE-2026-07-06.md | The entry point for a context-free session. Navigation only — supersedes nothing: complete timeline 07-03→07-06, document map + authority order, current-truth snapshot (decided / unsigned / ruled-but-unapplied / open / stale), the pending decision funnel (Gate A · three exits · friction log · commit authorization), and per-request playbooks with the binding session rules. |
| closing-plan-2026-09-08.md | Post-freeze schedule record #3 — the closing plan (2026-09-08). Defines “100%” as v1.0.0 tagged by KC-4 with every RELEASE.md box ticked, [HUMAN] ones included; locates the last 15% (~5% agent code/docs · ~3% owner decisions · ~7% owner acts and measurements); the eight-item decision batch D1…D8 with recommended defaults; five dated waves (W0 commit + KC-2 → L1–L8 code residue → L9–L11 retention + ratification kits at KC-3 → Ivan’s measurements and 14 days of real use → release candidate → tag) and the descope ladder. Its lanes are mirrored as the Closing board in ../../TODO.md. |
A self-hosted, local-first cockpit for Claude Code agent/subagent activity on a Mac Mini M4 — persisted subagent DAG + dollar-cost/delegation-savings + Telegram alerts + owned persistence — differentiated from the 28.4k★ baseline (claude-code-templates) by the four things it lacks, and from the whole field by a loopback-only, no-spawner, mandatory-token security posture.
Build it — the vision is A-grade and the security thinking is best-in-class — and v2 now resolves the three specification gaps that would otherwise have sunk it in code into the ten canonical decisions CD-1…CD-10: (1) ingest source-of-truth / reconciliation → a single immutable substrate + deterministic projection contingent on the Phase-0 spike; (2) licensing hygiene (cast/disler/nirdiamant are all-rights-reserved — reimplement, don’t copy) → a CI-enforced per-artifact rule; (3) scope discipline → personal-first / commercial-clean, with the vector-DB leap carved to an experimental track.
Test coverage (CI-gated) · README badges + donation · GitHub Pages docs site — while the app itself stays loopback-only.
Where the bar actually stands (2026-08-15, updated 2026-08-25). The historical wording here was “>90% coverage, CI-gated”. As shipped the threshold is 100% for statements, branches, functions and lines in all five packages, and every one of them carries a static guard that fails if a
v8 ignore/c8 ignore/istanbul ignorepragma appears insrc/— so the figure cannot be bought back by removing code from the denominator. Four of the guards live in a file namedcoverage-honesty.test.ts;apps/web’s is thecoverage honestyblock oftest/honesty.test.tsx. This paragraph used to continue"CI-gated" is not the same as merge-blocking, and today nothing is merge-blocking— true until 2026-08-25, whilemainwas unprotected and a red coverage run could not physically stop a merge. Since 2026-08-25mainis branch-protected: the required status check isci(lowercase — the job id in.github/workflows/ci.yml, not the workflow’sCIdisplay name), and force-pushes and deletion ofmainare refused for everyone. So a red run now withholds the merge button from a contributor, but not from the repository owner —enforce_adminsis deliberately off, because agenthropic has exactly one maintainer whose normal working mode is a direct push tomain, and admin enforcement would lock the sole maintainer out of their own repository. The canonical write-up of this correction is the standing correction; development-plan.md §2c records the other places the plan and the tree disagree.The GitHub Pages leg of the bar is met as well, as of the same day: Pages was enabled on 2026-08-25 by an owner-credentialled call and the site is live at https://ivanbbaev.github.io/agenthropic/. Verify with
gh api repos/IvanBBaev/agenthropic --jq .has_pages→true.Re-measured 2026-09-18 (the Wave 4 R1 rehearsal — every scriptable
RELEASE.mdcommand run, every exit code 0, no box ticked): 131 test files / 2428 tests, the 100 thresholds held in all five packages, 412 packages under the license gate with one documented exception, Pages answeringHTTP/2 200. The bar is met on the working tree; the release commit re-measures it.Re-measured 2026-09-26 (second R1 rehearsal, on
af10c87, clean tree): 167 test files / 3060 tests, the 100 thresholds held in all five packages, 406 packages / 428 installed versions under the license gate with one documented exception (the earlier 412 was a double count, corrected 2026-09-24). Pages not re-checked in that session.
Reviews the design of record in docs/ai/DESIGN.md (internal design basis, kept local-only
— not published in this repo). Implementation began on 2026-07-11 under an explicit owner
override of CD-8, which had required Gate A to be signed and the WP-S7 verdict to read
GO before any production code. The override covers dispatching only: it relaxes no
security invariant, signs no Gate A, ratifies no LABEL-ME numbers (they remain
PROVISIONAL) and does not authorise a commit or push without an explicit ask. Everything in
this corpus written in the future tense about “when code starts” is describing a gate that
opened on 2026-07-11.