Skip to main content

​
AI Features

This repo ships two AI products from one codebase. Outward-facing: any MCP-compatible client (Claude, Cursor, Copilot, ChatGPT plugins) can query Livepeer documentation as freshness-stamped authoritative source instead of inferring from training data. Public surfaces are the hosted MCP endpoint, llms.txt, an AI-enriched sitemap, and the Mintlify chat assistant on the live site. Inward-facing: every supported AI coding tool has a thin native adapter rooted in one AGENTS.md baseline. 34 canonical AI skills cover session lifecycle (thread, pm, research, design, build, iterate, close), content pipelines, repo audits, and governance creation. The same governance gates apply to humans and agents – adapters surface canonical truth for specific consumers but hold no authority of their own (per locked D-DG-11). The third strand – the automation belt that keeps both honest – was partially broken until 2026-05-25: llms.txt and sitemap-ai.xml had drifted 47 days stale, root-caused by the cron-is-dry-run bug in the dispatch workflows. Fix shipped 2026-05-25 in commit e42946cdf on docs-v2-dev-draft; the next scheduled cron after merge to docs-v2 should regenerate both files. See Known gaps.

​
Public AI surfaces

Three AI artefacts publish three different counts: llms.txt = 188 entries, sitemap-ai.xml = 181 URLs, docs-index.json = 532. Different pipelines, no reconciliation.

​
Agent adapters (6 + the cross-agent baseline)

All adapters point UP into AGENTS.md per locked D-DG-11. Adapters surface canonical truth for specific consumers but hold no authority. Any rule stated only in an adapter file is invalid and must canonicalise into a T0–T3 source in docs-guide/. Validator: operations/scripts/validators/governance/compliance/check-agent-docs-freshness.js – proves every adapter exists, is non-empty, matches the approved set. Validator path drift: docs-guide/policies/agent-governance-framework.mdx lines 137 + 228 reference operations/scripts/validators/governance/check-agent-docs-freshness.js (missing /compliance/ subfolder). AGENTS.md line 98 has the correct path. Cleanup required.

​
AI skills (34 local + 53 portable + 53 templates)

The skill catalogue covers session lifecycle, content pipelines, audits, governance creation, and agent packaging. Live counts (2026-05-23):
  • 34 active local SKILL.md under ai-tools/ai-skills/<skill>/SKILL.md (+ 1 archived in x-archive/)
  • 53 portable skill exports under ai-tools/agent-packs/skills/
  • 53 templates under ai-tools/ai-skills/templates/*.template.md
  • 9 entries in skill-catalog.json (26% of 34 skills catalogued)
  • 1 entry in agent-packs/skills/manifest.json (the browser skill) – 52 portable skills exist next to it unregistered
  • 317 entries in ai-tools-registry.json

​
Skills by group (live, 2026-05-23)

Maturity: per SLICE-02, 18 of 34 local skills still carry status: draft with empty test logs – including most of the session lifecycle backbone (thread, research, design, build, iterate, close). The primary scaffold has never been validated in production despite daily use.

​
Templates and portable skills

  • 53 .template.md files at ai-tools/ai-skills/templates/ – canonical authoring templates.
  • 53 portable SKILL.md at ai-tools/agent-packs/skills/ – generated exports for distribution to per-agent packs.
  • 15 portable wrappers are stale v1.0 pointing to local canonical bodies at v1.3/v1.4 – needs regeneration.
  • 38 portable-only skills have NO local canonical body – referenced by dispatchers like page-ship.mdx but the dispatcher graph is partly fictional.
  • Templates 32/37/38 each appear twice with different content; 41/42 duplicate the content of 37/38 (renumbered without deleting originals).

​
Generators that keep the skill system honest

​
AI tools registry

ai-tools/registry/ai-tools-registry.json is the canonical 344 KB inventory of every ai-tools/** artefact. 317 entries. 7-lane model with lifecycle states. Coverage gap: 9 of 34 skills are in skill-catalog.json (26%); the rest are in the registry but absent from the catalogue. Two separate inventory systems (skill-catalog.json vs ai-tools-registry.json) with no automated cross-check.

​
ai-rules (ai-tools/ai-rules/)

Pre-governance rule docs flagged for retirement by the governance framework: Two AI_GUIDELINES.md files give contradictory --no-verify policy – the .augment/ shadow forbids it; the live + _retired versions allow it with metadata. Reconciliation required.

​
AI governance contracts

Documented in docs-guide/policies/agent-governance-framework.mdx and docs-guide/frameworks/ai-tools-governance.mdx.

​
Contributor rules

  • Do not edit generated AI artefacts (llms.txt, sitemap-ai.xml, docs-index.json, generated skill catalogs, generated agent packs) by hand. Run the named generator.
  • Do not add a new skill without a SKILL.md, registry entry, and validation story.
  • Do not treat AI-generated reports as source of truth unless their generator and source inputs are named.
  • Do not bypass branch or deletion rules for AI work. The Codex task-contract scope is checked at preflight.
  • Native adapters must be thin per D-DG-11. New rules go in AGENTS.md or the relevant docs-guide policy – not adapter files.

​
Known gaps

​
llms.txt + sitemap-ai.xml 47 days stale (P0) – root cause FIXED 2026-05-25

Both root artefacts last regenerated 2026-04-15 vs docs.json last touched 2026-05-22. Root cause: the cron-is-dry-run bug in dispatch-discoverability.yml + 5 sibling dispatchers. Fix shipped in commit e42946cdf on docs-v2-dev-draft (inverted FLAGS in all 6 scheduled jobs). Pending merge to docs-v2, after which the next 04:00 UTC cron should regenerate both files. See Cron-is-dry-run bug FIXED.

​
Veracity pipeline now has dispatchers (was: Phase 6 blocker)

Three new dispatchers under operations/scripts/dispatch/content/veracity/ (May 2026) close the long-standing “veracity-pass DRAFT blocks Phase 6” gap: docs-research-packet.js (engine), orchestrator-guides-research-review.js (scope wrapper), docs-page-research-pr-report.js (advisory PR artefact, experimental, non-blocking). All manual-pipeline; not yet on cron. See Automations § Recently shipped.

​
skill.md is documented but has no source

v2/resources/documentation-guide/ai-automations/ai-features.mdx advertises https://docs.livepeer.org/skill.md as a published AI artefact. The repo has no skill.md at root. Either Mintlify generates from another input, or the doc is aspirational. Resolve or strike from the public page.

​
docs-guide/tooling/ai-tools.mdx skill/template counts are stale

The page claims “42 templates” and “42 skills”. Live is 53 templates and 34 local skills (53 portable). Both numbers wrong. Replace with generator-emitted blocks from generate-ai-skills-indexes.js + CI assertion on drift.

​
.claude/CLAUDE.md violates D-DG-11 thin-adapter rule

326 lines vs AGENTS.md 119 lines (2.74×). Most of the bulk is a project-state table that belongs elsewhere. Trim to ≤200 lines; no policy duplicated from AGENTS.md.

​
18 of 34 local skills are status: draft with empty test logs

The entire session-lifecycle backbone (thread, research, design, build, iterate, close) has never been validated in production despite being the primary scaffold for AI-assisted work in this repo. Promote one at a time: pick a draft skill, run 3 production passes, record results, flip to status: current.

​
Local skills (34) vs portable skills (53) divergence is undocumented

agent-packs/skills/manifest.json registers ONLY ONE skill (browser) while 53 portable SKILLs live next to it. 38 portable-only skills have no local canonical body – referenced by dispatchers but the graph is partly fictional. Reconcile or document why portable-only skills exist.

​
Template numbering drift

Templates 32/37/38 each appear twice with different content; 41/42 duplicate the content of 37/38 (renumbered without deletion). Cleanup queue.

​
docs-copy/ has both reference/ and references/

Two ref directories under ai-tools/ai-skills/docs-copy/ (singular + plural) containing different files. Never reconciled. Pick one.

​
Validator path drift in agent-governance-framework.mdx

Lines 137 + 228 reference validators/governance/check-agent-docs-freshness.js; actual path is validators/governance/compliance/check-agent-docs-freshness.js. AGENTS.md line 98 has the right path. Cleanup required.

​
.github/augment-instructions.md at active path with archived-header

206 lines – the heaviest single adapter in the repo. Header says archived; path is active. Directory-walking agents still read it. Either move to _retired/ or actually delete.

​
ai-tools/ai-rules/.augment/ phantom-retired subtree

165 lines, 4 files, at active path. Should be in _retired/ per the governance framework. Currently invisible to the freshness validator.

​
ai-tools/registry/ai-tools-inventory.md lags the registry

Inventory mtime 2026-04-05; registry mtime 2026-04-08 (later). Regenerate the inventory.

​
codex/skills-manifest.json provenance drift

Claims generated_at: 2026-03-22 but file mtime is 2026-04-08. Either manual edit OR generator bug (failed to update its own generated_at field, same pattern as docs-index.json SLICE-12 finding).

​
robots.txt doesn’t advertise sitemap-ai.xml

Only Mintlify’s auto-generated sitemap.xml is listed. AI-enriched metadata in sitemap-ai.xml is undiscoverable by crawlers via robots.txt. Add the entry.

​
Three AI artefacts publish three different counts

llms.txt = 188 entries, sitemap-ai.xml = 181 URLs, docs-index.json = 532. No automated reconciliation check. Either pick one source-of-truth or document why three.

​
Mintlify chat assistant has no docs.json alignment check

.mintlify/Assistant.md carries an IA routing map that’s hand-maintained against docs.json. No validator confirms they agree. Wire one or commit to manual sync on every docs.json change.
  • AI Tools Governance Framework – architecture: 7-lane model, lifecycle states, skill/dispatcher/workflow taxonomy
  • Agent Governance Framework – adapter rules + retired-legacy table + migration map
  • AI Tools Registry – registry schema + per-skill detail (currently has stale counts – needs refresh per Known gaps)
  • Automations – the cron-is-dry-run bug + dispatch architecture that drives AI artefact regeneration
  • Content Writing Pipeline + Checks Framework – how the content-pipeline skills (pass-a, pass-b, tab-map) operationalise per-page review
  • Ownerless Governance Policy – the 4-part contract for governed surfaces (4 of 5 unified surfaces formally ownerless-ready post-D-ACT-10; legacy 8-entry file superseded)
  • AGENTS.md (root) – cross-agent baseline (119 lines); every adapter points here
  • .mintlify/Assistant.md – the Mintlify chat assistant configuration
  • ai-tools/registry/ai-tools-registry.json – canonical 317-entry inventory
  • workspace/thread-outputs/repo-consolidation-deep/SLICE-02-ai-tools.md – full 324-file ai-tools/ inventory with maturity per skill
  • workspace/thread-outputs/repo-consolidation-deep/SLICE-06-agent-adapters.md – full 49-file adapter inventory with thin-adapter compliance per adapter
Last modified on June 2, 2026