← back to AbramsEgo
research: dropped links + norma tour patterns
a48913f10a070e7c15963f9e07d311f2fc0040fc · 2026-07-01 16:55:17 -0700 · Steve
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Files touched
M memos/design-research.md
Diff
commit a48913f10a070e7c15963f9e07d311f2fc0040fc
Author: Steve <steve@designerwallcoverings.com>
Date: Wed Jul 1 16:55:17 2026 -0700
research: dropped links + norma tour patterns
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
memos/design-research.md | 196 +++++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 196 insertions(+)
diff --git a/memos/design-research.md b/memos/design-research.md
index 1bf6f6c..6944866 100644
--- a/memos/design-research.md
+++ b/memos/design-research.md
@@ -117,3 +117,199 @@ the aesthetic as a product; the tour itself is the demo.
| ffmpeg frame extraction (130 frames) | $0 (local) |
| Frame reads (10) + curl attempts | $0 (local) |
| Browserbase/Exa | $0 — not used (deferred to task 31) |
+
+---
+
+## dropped-links
+
+**Researched:** 2026-07-01 (task 31). **Cost: $0** — WebFetch and the Exa MCP
+were both permission-blocked in the session, but every source fell to free
+plumbing: all three X posts recovered via Twitter's public syndication CDN
+(`cdn.syndication.twimg.com/tweet-result?id=<id>&token=<derived>` — token is
+`((id/1e15)*π).toString(36)` minus zeros/dots), docs + READMEs via curl, and
+repo architecture via the unauthenticated DeepWiki MCP endpoint
+(`mcp.deepwiki.com/mcp`, JSON-RPC `ask_question`). No Browserbase, no Exa.
+
+### Nous Research — Hermes Agent web search & extraction
+
+Thread (2026-06-30): *"Hermes Agent now reads the web up to 60x faster and 49x
+cheaper. Scraping backends pass clean content straight to the agent without
+redundant processing steps; large pages are saved locally and paged on demand."*
+Linked docs (hermes-agent.nousresearch.com → web-search feature page):
+
+- Two model-callable tools — `web_search` + `web_extract` — behind a
+ **multi-provider backend matrix** (Firecrawl default, SearXNG, Brave, DDGS,
+ Tavily, Exa, Parallel, xAI/Grok), with a **per-capability split**: search
+ and extract can use different providers (free SearXNG search + paid
+ Firecrawl extract).
+- **Size-driven extract pipeline:** <5k chars → raw pass-through, no LLM call;
+ 5k–500k → single-pass summary via a cheap *auxiliary* model, capped ~5k out;
+ 500k–2M → 100k-char chunks summarized in parallel then synthesized; >2M →
+ refused with a "use a more focused URL" hint. On summarizer failure it falls
+ back to the first ~5k of raw content instead of erroring.
+- **Auxiliary-model cost routing:** extraction summaries default to the main
+ model but are explicitly routable to a cheap fast model so long-page reads
+ don't burn reasoning-model tokens.
+
+**Patterns → AbramsEgo panels:**
+
+1. **Tiered by-size processing** → *server.js snapshot collectors*: collectors
+ that read big JSONL ledgers (revenue-ledger, cost-ledger, activity) should
+ pass small files raw, summarize/aggregate above a threshold, and refuse
+ with a hint at silly sizes — never stream a 2M-char blob into a panel.
+2. **Provider fallback ladder, free-first** → *research/build tasks*: exactly
+ what this task lived — WebFetch → nitter mirrors → syndication CDN → (Exa
+ only if all free rungs fail). Codified here so future tasks don't start at
+ the metered rung.
+3. **Cheap-model auxiliary routing** → *build loop*: summarize/label jobs in
+ loop tasks should name a cheap model; decisions stay on the main model.
+ Pairs with the Always-Show-Costs rule.
+4. **Quantified before/after framing** ("60x faster, 49x cheaper") → *landing
+ page copy* (task 35): every claim carries a number.
+
+### @0xRaduan — DeepWiki study posts + the two repos
+
+- Post 2038922136375492835 (2026-03-31), replying to @Fried_rice: *"can study
+ it with deep wiki now: deepwiki.com/instructkr/claude-code"*. That repo has
+ since been **renamed → `ultraworkers/claw-code`** (Rust, ~194k stars):
+ self-described as *"an agent-managed museum exhibit … developed and
+ maintained with no human intervention"* — the same species as AbramsEgo's
+ autonomous build-queue loop, at internet scale.
+- Post 2043291064149438661 (2026-04-12) thread: *"how I research any topic on
+ the internet nowadays. TLDR on the best option for claude code statuslines:
+ **cship** by @MarieStephenLeo — OSS, written in Rust, shows you your 5h/7d
+ limits."* Repo: github.com/stephenleo/cship — Starship-style TOML-config
+ statusline for Claude Code: session cost, context usage, model, API usage
+ limits, per-module warn/critical thresholds, ≤10ms render budget.
+
+**claw-code agent-environment architecture** (via DeepWiki MCP; the
+deepwiki.com page itself is client-rendered JS — curl gets nothing):
+
+- **AgentOutput manifest per agent run**, persisted as JSON explicitly so
+ external viewers can poll it: `agent_id, name, status, derived_state,
+ current_blocker, error, output_file, lane_events`.
+- **LaneEvents** — lifecycle verbs `lane.started / lane.finished /
+ lane.blocked / lane.commit.created` — with reconciliation of contradictory
+ or out-of-order events into one deterministic truth.
+- **derived_state taxonomy:** `working / finished_cleanable /
+ blocked_background_job / interrupted_transport`.
+- **LaneOwnership `watcher_action` (Act / Observe / Ignore) + per-event
+ ConfidenceLevel** so downstream automation knows whether to act on an event.
+
+**Patterns → AbramsEgo panels:**
+
+1. **Manifest-per-run file contract** → *build-queue loop*: have
+ scripts/build-loop.sh write/update a per-task manifest JSON (status,
+ derived_state, current_blocker, commit hash, events) that the server.js
+ sessions/activity collectors read — far richer than inferring state from
+ tasks/ vs done/ vs failed/ directory moves.
+2. **derived_state on cards** → *Sessions & build-queue panel*: cards show
+ working / blocked(+named blocker string) / finished-cleanable — "blocked"
+ with the actual blocker text beats a generic red dot.
+3. **Lifecycle-verb event log** → *Activity feed*: normalize items to
+ started/finished/blocked/commit.created verbs; a confidence field keeps
+ noisy sources from triggering automation.
+4. **cship's module+threshold model** → *top metrics bar / cost panel*:
+ per-module warn/critical thresholds with per-state colors (already SPEC'd);
+ add the **5h/7d API-limit readout** as a first-class stat next to session
+ cost — limits are the scarcest resource the loop consumes.
+5. **DeepWiki MCP as a $0 research tool** for future repo-study tasks. Caveat
+ observed live: querying an unindexed/renamed repo name returns a
+ confidently wrong answer (the old `instructkr/claude-code` name produced a
+ garbage "Devin" citation) — always resolve the canonical repo name first
+ (GitHub API 301 follow), then ask.
+
+**What NOT to copy:** claw-code is a CLI harness, not a dashboard — steal the
+state/file contracts, not the Rust implementation. And its "no human
+intervention" posture stops at AbramsEgo's rails: gated actions stay
+Steve-gated; the manifest pattern must make approvals *more* visible, not
+automate past them.
+
+---
+
+## norma-videos
+
+**Researched:** 2026-07-01 (task 31). **Cost: $0 (own server, SSH read-only).**
+
+**Index status: the :9891 index is DEAD.** Port 9891 on Kamatera is loopback-
+only and now serves **recycledwallpaper.com** (port was reused); nothing
+externally answers http://45.61.58.125:9891/. The video assets themselves
+survive on Kamatera under `/root/.claude/skills/app-demo-video/output/`:
+`norma-panels-tour/{email-center, saved-analyses, clients, gmail-crm,
+outreach-playbook}`, `norma-analyzer-deep-usage`, `norma-email-analyzer`,
+`norma-email-outreach` (plus older cuts in `/root/Projects/Norma/demos/`).
+
+### The proven per-panel tour recipe (app-demo-video skill)
+
+Each panel directory is a self-contained build:
+
+```
+narration.txt ~300–370 chars, one paragraph
+voice.mp3 TTS of the narration (+ voice-padded.mp3)
+page@<hash>.webm Playwright screen recording of the live panel
+intro.mp4 title card
+body.mp4 recording cut to narration length
+captions.ass → body-cap.mp4 captions BURNED IN
+concat.txt intro.mp4 + body-cap.mp4
+<panel>.mp4 final artifact
+```
+
+Final files are tiny — **0.5–1.2 MB ≈ 30–60s per panel**; the long-form
+`norma-analyzer-deep-usage.mp4` is 4.3 MB. The narration formula is worth
+copying verbatim (email-center example): name the panel → one scope sentence
+(*"every send, every campaign and every reply … unified in one queue"*) →
+2–3 capability sentences (*"You can filter by status, by client … drill into
+any thread"*) → close on the always-on stat (*"reply rates, opens and clicks
+are tracked per message so you always know what is working"*).
+
+### What transfers to the AbramsEgo landing (task 35 consumes this)
+
+1. **One MP4 per SPEC panel**, 30–60s, named `<panel>.mp4` — a rack of short
+ per-panel tours, not one monolithic walkthrough.
+2. **intro + captioned-body concat** structure; captions burned in, so the
+ landing can autoplay muted and still communicate.
+3. **The narration formula** above; keep `narration.txt` beside each MP4 so
+ the landing renders the same text as copy under the player.
+4. **≤1.5 MB per panel video** — proven size, fine to self-host from :9773
+ (no CDN, no external dependency).
+5. **Regenerate against the live :9773 dashboard** with the app-demo-video
+ skill — the Norma set is the *format* reference, not content to reuse. And
+ since the :9891 index is dead, the AbramsEgo landing self-hosts its videos.
+
+---
+
+## Apply next (updated 2026-07-01, task 31 — supersedes the task-30 list)
+
+1. Segmented 10-cell gauges + bracketed monospace labels in the top metrics
+ bar (keep threshold colors).
+2. Per-agent golden-angle hue chips in Fleet health, reused everywhere that
+ agent appears.
+3. `[ PANEL NAME ]` boxed headers with a per-panel summary metrics row
+ (count / healthy% / live N).
+4. Focus state: click card → glow + dim siblings.
+5. One-stat caption line on session/sub-agent cards (model · tokens · $).
+6. Slide-in + source-panel pulse on new activity-feed items.
+7. Build-loop **per-task manifest JSON** (status / derived_state /
+ current_blocker / lifecycle events / commit) written by build-loop.sh,
+ read by server.js collectors. (claw-code)
+8. Activity feed normalized to lifecycle verbs
+ (started/finished/blocked/commit.created). (claw-code)
+9. Sessions panel: **5h/7d API-limit readout** next to session cost. (cship)
+10. Snapshot collectors: tiered by-size handling of big JSONL ledgers —
+ raw / aggregate / refuse-with-hint. (Hermes)
+11. **Task 35:** per-panel 30–60s MP4 tours via the app-demo-video skill —
+ narration formula, intro+captioned-body concat, ≤1.5 MB each, self-hosted
+ from :9773 with narration.txt beside each file. (Norma)
+12. (Deferred/stretch) spatial map toggle view. (androoagi)
+13. ~~Browserbase pass on x.com posts~~ **RESOLVED $0** — the syndication-CDN
+ trick got full post content; Browserbase not needed for X text posts.
+
+### Cost lines (task 31)
+
+| Item | Cost |
+|---|---|
+| X posts ×5 (syndication CDN) | $0 (free endpoint) |
+| Hermes docs + READMEs (curl) | $0 |
+| DeepWiki MCP queries ×3 | $0 (free endpoint) |
+| Kamatera Norma asset survey (SSH) | $0 (own server) |
+| Browserbase / Exa | $0 — not needed |
← 6e2e9c4 research: androoagi design north-star notes
·
back to AbramsEgo
·
queue: task 31 done — dropped-links + norma-videos research c6143c7 →