View

Generate a branded HTML artifact from Egregore data and open it in the browser.

Works in both connected and local mode — resolves data from memory files, not the graph.

When to invoke

User says: "show me visually", "render this", "view as artifact", "open in browser", "make this readable", "generate artifact", "show me [quest/handoff/plan]", "/view"

Not this: terminal formatting → just format in markdown · dashboard → /dashboard

Arguments: $ARGUMENTS (Optional: artifact type and/or name, or a file path)

Loom routing

Skip this section if your prompt contains LOOM-EXECUTOR — you are the executor; run the skill as specced below. Full protocol: .claude/context/loom.md.

Composition is the default for flagship documents — and it is main-loop-only. Compose (do NOT delegate; run the Composition path below inline, print bash bin/loom.sh footer view --override, set "override":true,"class":"composition" in telemetry, skip the rest of this routing section) whenever either:

• the doc is flagship — a document render that reads as a strategy / prep / board / briefing / explainer / analysis / decision doc, i.e. something meant to be read or presented, with multiple ## sections. This is now the DEFAULT for such docs (the floor disappointed too many times); OR
• an explicit cue is present — --compose, "compose this", "make it presentable / client-facing / flagship", "with the design trace / use the design trace / designed artifact / band 5".

Opt DOWN to the fast template floor (which may route to the cheap tier) ONLY when: the invocation carries --floor / --fast, or the ask is a quick/utility look ("just show me", "quick look", "rough render"), or the target is short/non-flagship (a stub, a single-section note), or it's a typed artifact (quest / handoff / activity / board / network — those keep their own templates and are not affected by this default). Composition is a frontier-authoring act; a cheap executor can only produce the floor.

1. Resolve: ROUTE=$(bash bin/loom.sh route view), then DECISION_ID=$(printf '%s\n' "$ROUTE" | jq -r '.decision_id // empty').
2. If mode ≠ delegate, or the user signalled depth ("deep", "think hard", --deep) → run this skill inline as normal. On a depth override, print bash bin/loom.sh footer view --override after the output and set "override":true in telemetry.
3. Otherwise delegate: spawn the Agent tool with subagent_type: "loom-executor", model = the route's tier, prompt = LOOM-DECISION-ID: $DECISION_ID on its own first line, then LOOM-EXECUTOR: Execute .claude/skills/view/SKILL.md, plus the user's arguments and any context the spec needs from the session. Print the executor's final output verbatim, then print the output of bash bin/loom.sh footer view.
4. If the spawn fails or the executor's first line is LOW_CONFIDENCE: — triage the reason: needs-user-interaction or a main-loop-only tool → take over and finish this skill inline (no escalation); genuine uncertainty or failure → reassign ROUTE=$(bash bin/loom.sh escalate view "<reason>"), refresh DECISION_ID from ROUTE, then re-spawn once on the new tier carrying the returned decision ID (sticky for this session).
5. Telemetry (fire-and-forget): bash bin/telemetry.sh emit "command" '{"command":"view","routed":true,"mode":"delegate","model":"<actual>","route_tier":"<table tier>","class":"<class>","escalated":<bool>,"override":<bool>,"source":"<source>"}' 2>/dev/null &

Supported artifact types

• quest — renders quest markdown from memory/quests/
• handoff — renders handoff markdown from memory/handoffs/
• activity — renders live team activity dashboard (no file needed)
• board — renders project board from memory/board/board.json (no file needed; interactive editor with paste-back loop, 5 tabs: Activity / Priority / Person / Timeline / Done). In connected mode, also publishes to a stable URL at egregore.xyz/view/{org}/board on every invocation — bookmark it and refresh to see the latest.
• network — renders people/relationship network (no file needed)
• document — renders any markdown file with branded styling (auto-detected fallback)

Resolution logic

The key job of /view is resolving what the user wants to see into a file path. This must work without a graph.

1. Parse the arguments

• /view quest artifact-generation → type=quest, name=artifact-generation
• /view handoff oss-security-audit → type=handoff, name=oss-security-audit
• /view activity → type=activity, no file needed
• /view board → type=board, reads memory/board/board.json, no file argument needed
• /view network → type=network, no file needed
• /view artifact-generation → no type specified, search all types
• /view memory/knowledge/decisions/some-decision.md → direct file path
• show me the security audit visually → extract keywords, search

2. Resolve the file

Direct file path: If the argument looks like a file path (contains / or ends in .md), resolve it directly. If it exists, use it — type is auto-detected from location or falls back to document.

Quest: Search memory/quests/ for {name}.md or partial match:

# Exact match
FILE="memory/quests/${name}.md"
# Partial match — find files containing the name
ls memory/quests/*.md | grep -i "$name" | head -1

Handoff: Search memory/handoffs/ recursively (files are in date subdirectories):

# Search all subdirectories
find memory/handoffs/ -name "*.md" -not -name "index*" | grep -i "$name" | head -1
# If multiple matches, prefer most recent (sorted by path which includes date)
find memory/handoffs/ -name "*.md" -not -name "index*" | grep -i "$name" | sort -r | head -1

Activity: No file resolution needed — runs bin/activity-data.sh live.

Board / Network: No file resolution needed. board reads memory/board/board.json automatically (via git root). network is generated from people data.

Auto-detect type (no type specified):

1. Search memory/quests/ first
2. Then memory/handoffs/ recursively
3. Then memory/knowledge/ recursively
4. If found, infer type from location (quest or handoff) — everything else is document

3. Generate and open

Resolve the renderer first — prefer the in-repo CLI. Running repo code avoids fetching an external npm package (which permission classifiers flag) and exercises local edits to packages/egregore-artifacts without waiting for an npm release:

RENDER="npx egregore-artifacts@latest"
if [ -f packages/egregore-artifacts/bin/cli.js ] && [ -d packages/egregore-artifacts/node_modules/react ]; then
  RENDER="node packages/egregore-artifacts/bin/cli.js"
fi

(If the local package exists but deps are missing, either run npm install --prefix packages/egregore-artifacts or fall back to npx.)

Design trace (documents). A document render should follow the design trace, not ship bare: auto-walk the UGI synthesis graph from the document's substance (five stage ids — objective · audience · register · palette · grammar; option ids and auto-walk rules in packages/design-system/generative-ui/skill/SKILL.md), resolve the brief, and pass it to the renderer:

node --input-type=module -e "
import { resolveBrief } from './packages/design-system/generative-ui/resolve-brief.js';
import fs from 'node:fs';
fs.mkdirSync('/tmp/egregore-artifacts', { recursive: true });
fs.writeFileSync('/tmp/egregore-artifacts/brief-{slug}.json',
  JSON.stringify(resolveBrief(['{objective}','{audience}','{register}','{palette}','{grammar}'])));
"
$RENDER document <file> --brief /tmp/egregore-artifacts/brief-{slug}.json

Pick the five ids from the substance, one line of judgment each (e.g. a strategy prep doc → decide · operators · editorial · vellum · decisive; a public explainer → persuade · newcomer · marketing · loam · quiet). The brief drives palette + grammar treatment; the designed layout (nav · hero · anchored sections) renders regardless. If the generative-ui layer is unavailable (pure-npx environment, no repo checkout), render without --brief — never block on the trace.

Composition path (--compose) — band 5, main-loop only

The template above is the floor. --compose is the ceiling: the design trace at full depth means COMPOSITION, not pass-through (D6 free-generative band — how the reference pages were made). The template renderer can never reach it; composition is the frontier model authoring the page from the substance. This path is what makes that reachable from the command instead of only by accident.

This is the DEFAULT for flagship documents (strategy / prep / board / briefing / explainer / analysis / decision docs — see the Loom-routing rule above). It also fires on explicit cues: --compose, "compose / make it presentable / client-facing / flagship / with the design trace / use the design trace / designed artifact / band 5". When the user names "the design trace," they mean this composed ceiling — never the floor. Opt DOWN to the fast template floor below only for quick/utility looks or --floor/--fast (again, see the routing rule). The floor is what disappoints when someone wanted the trace — so when unsure whether a document is flagship, compose.

Technical documents compose too — in a different register. A spec, RFC, protocol, architecture doc, API reference, evaluation report, or postmortem is flagship and gets the full chrome, but it does not get editorial voice. See step 4 below: the register decision comes before any content is written, and picking wrong is the single most common way a composed render lands badly.

Procedure (run inline — never delegate; see the compose note in Loom routing):

1. Read the source document in full.

2. Start from the scaffold: .claude/skills/view/compose-scaffold.html (copy it; it carries the complete Meridian chrome — vellum/nocturne/loam tokens, fonts, sticky nav, theme toggle, contours — and a documented component kit). You fill content; you do not rebuild the chrome or re-pick colors.

3. Walk the UGI graph for the register/palette/grammar (as above) and stamp the manifest register/grammar + the footer trail. Pick the palette by substance: vellum (strategy/decision), loam (warm/instructional), etc.

4. Decide the register FIRST — editorial or technical. This governs every sentence you then write, and it is not a style preference; it is what the document is. Ask: does a reader come here to be persuaded of a view, or to look something up and implement it?

   	Editorial	Technical
   Documents	strategy · prep · briefing · explainer · analysis · narrative recap · pitch	spec · RFC · protocol · architecture · API reference · evaluation report · runbook · postmortem
   Headings	statement titles — the section's finding as display copy ("Most of the zoom-out is already decided.")	descriptive, numbered — the section's name ("3.3 Toponym and cultivar collision")
   Hero	two-tone Lead. <em>accent phrase</em> + italic standfirst	plain title + a .meta grid (version · date · author · depends-on)
   Prose	argues, lands a point, carries voice	states, qualifies, cites; declarative and neutral
   Opens with	the claim	Abstract, then Scope
   Decisions	.hl / .readout — the thing to land	explicit Decision: blocks, individually citable
   Tables	.sumtable, status matrices	numbered with <caption> ("Table 2 — …") so prose can reference them
   Never	bury the finding in a neutral heading	editorialize a heading, or assert without the measurement behind it

   Signals for technical: numbered sections, a version field, "spec"/"protocol"/ "requirements" in the title, code blocks carrying invocations, tables of measurements, a References section. When the document is something someone will implement from, choose technical. Editorial voice on a spec reads as unserious and buries the lookup value — that is the failure mode this table exists to prevent.

   Stamp the choice into the manifest register and the footer trail (editorial / technical, with a matching grammar such as decisive or specification).

5. Transform, don't mirror — in the chosen register. Never reproduce the markdown structure verbatim; pick a component per section from the kit by what the content is:

   • .ledger for Q→A pairs · .tagcard for named claims · .claims for numbers/stats · .panels+.verdict-band for option sets · .hl for the one thing to land · .steps for sequences · .feat for capability+status rows · .threads for decision lists · .gap for negatives · .sumtable with .dots for a status matrix · .readout for the honest bottom line
   • technical renders lean on captioned tables, Decision: blocks, numbered rationale lists, and .note/.note.caution for limitations and hazards; they use .hl/.readout sparingly and never as a substitute for a heading
   • transformation in technical register means structure, not voice — split prose into tables, number the rationale, surface the decisions; do not rewrite the author's claims into slogans
   • nav links: one per composed section, to its id anchor

6. Never inline a hex in content — every theme-sensitive color is a var(--token), or the toggle breaks dark mode. The kit already obeys this. Verify before opening: every var(--x) used must be defined in both [data-theme="vellum"] and [data-theme="nocturne"], or dark mode breaks silently on that element.

7. Write to /tmp/egregore-artifacts/composed-{slug}.html and open directly (open <path>), then report the path + Renderer: composed (band 5, inline).

Composition is judgment, not a script — the scaffold is the vocabulary, the substance decides the sentence. When --compose is absent, use the fast template path below.

For typed artifacts with a file:

$RENDER <type> <resolved-file-path>

For auto-detected (just a file path):

$RENDER <resolved-file-path>

For activity (no file):

$RENDER activity

For board (connected mode only — publish to stable URL):

After opening locally, fire-and-forget a publish with a stable --id so the URL egregore.xyz/view/{org}/board always shows the latest board:

_API_URL=$(jq -r '.api_url // empty' egregore.json 2>/dev/null)
_MODE=$(jq -r '.mode // empty' egregore.json 2>/dev/null)
if [ -n "$_API_URL" ] && [ "$_MODE" != "local" ]; then
  ORG_NAME=$(jq -r '.org_name // .slug' egregore.json)
  bash bin/publish-artifact.sh board memory/board/board.json \
    --id board \
    --title "Project Board — $ORG_NAME" \
    --author "$ORG_NAME" \
    --description "Latest board for $ORG_NAME" 2>/dev/null &
fi

The publish script exits silently on failure, so the local open always succeeds regardless of API state. OSS/local mode skips this step entirely — the board stays local unless the user publishes explicitly.

Auto-linked references (connected mode). When publish-artifact.sh publishes a markdown file, any backtick-wrapped memory/**/*.{md,html} paths inside it are re-published in parallel at deterministic URLs (egregore.xyz/view/{slug}/{m|h}-{12 hex}) so the rendered parent view contains clickable links to each referenced file. See bin/publish-references.sh. No-op in OSS mode (the relay assigns random slugs, so the renderer falls back to plain <code>).

4. Report

✓ Artifact opened in browser
  File: /tmp/egregore-artifacts/{type}-{name}-{ts}.html

For /view board in connected mode, append:

◆ https://egregore.xyz/view/{org_slug}/board   (stable — refresh for latest)

Read org_slug from egregore.json.

Fallback

If the local CLI is unavailable and npx egregore-artifacts fails (not installed), install it first:

npm install -g egregore-artifacts

Ambiguity handling

If the name matches multiple files, use AskUserQuestion:

Found multiple matches for "security":
1. handoffs/2026-03/31-cem-oss-security-audit.md
2. quests/oss-security-review.md
Which one?

If no matches found, fall through to synthesis mode (see below).

Synthesis mode

When the input is a prompt or topic rather than a file name — or when file resolution finds nothing — synthesize an artifact from multiple sources.

1. Read relevant files — search memory/, codebase, and conversation context for material matching the prompt. Read as many files as needed.
2. Write a temporary markdown file — synthesize the findings into a well-structured document at /tmp/egregore-artifacts/synthesized-{slug}.md. Use headings, lists, code blocks — the renderer handles all standard markdown.
3. Render it — $RENDER document /tmp/egregore-artifacts/synthesized-{slug}.md (renderer resolved as in §3)
4. Report — same as normal: ✓ Artifact opened in browser

This is the default fallback — don't ask the user if they want synthesis. If /view auth architecture doesn't match a file, just do the research and render it.

When to synthesize vs. when to say "not found":

• Prompt is a topic/question ("auth architecture", "how does onboarding work") → synthesize
• Prompt looks like a filename that should exist but doesn't ("my-missing-doc") → say not found
• Use judgment — if the user clearly expects a specific file, don't synthesize a guess

Examples

> /view quest artifact-generation

✓ Artifact opened in browser
  File: /tmp/egregore-artifacts/quest-artifact-generation.html

> show me the security audit visually

Resolving "security audit"...
  Found: memory/handoffs/2026-03/31-cem-oss-security-audit.md

✓ Artifact opened in browser
  File: /tmp/egregore-artifacts/handoff-31-cem-oss-security-audit.html

> /view memory/knowledge/decisions/auth-redirect.md

✓ Artifact opened in browser
  File: /tmp/egregore-artifacts/document-auth-redirect.html

> /view activity

✓ Artifact opened in browser
  File: /tmp/egregore-artifacts/activity-2026-04-06.html

> /view auth architecture

No file match — synthesizing from codebase...
  Reading: api/main.py, api/auth.py, api/services/supabase.py, ...

✓ Artifact opened in browser
  File: /tmp/egregore-artifacts/document-auth-architecture.html