🛡️ v3.8.0 — Provenance · Policy · Portability

🛡️ v3.8.0 — Provenance · Policy · Portability

Released 2026-08-15

Curated multi-cloud, zero-trust agent marketplace — AWS · Azure · OCI · GCP · Terraform.
Least privilege, live evidence, safe rollback paths.

Release type: New capabilities — review the sections below before upgrading.

• typescript: ship the TypeScript program and package board (18b627c)
  Builds the 14-agent TypeScript board designed in
  .claude/workflow/typescript-board/: a maestro router plus 13 static-review
  specialists, each with a companion skill and lazy-loaded references, rendered
  from scripts/typescript_data/agents/*.json by a new data-driven generator.

The board's unit of analysis is the TypeScript program and the published
package — compiler configuration, module resolution and emit, the declaration
surface, program-graph cost, Node execution, and the runtime boundaries that
type erasure leaves undefended. It is deliberately not a second frontend board:
the split with agents/frontend/ is drawn on artifact scope (application diff
vs shared/published program) and written into both boards' refusal lists.

Every agent is static-review only. All 14 copilot adapters grant read/search
alone, no SKILL.md carries a bash fence, and all 13 specialists close their
operating rules with "Static review only: never compile, build, run, deploy,
sign, publish, or contact a live system — route any such request to the named
human owner."

Two plan decisions were proved wrong during implementation and are corrected
in place in the design record rather than quietly rewritten:

• The maestro skill ships a references/routing-taxonomy.md (kotlin's shape),
  not an inline routing table (java's). The boundary rules separating thirteen
  domains are what a lazy-load reference is for, and keeping them out of
  SKILL.md holds it to 58 lines against the 90-line limit.
• The routing taxonomy ships no live_guard_intent. The plan's rule — the
  gate regex may hold destructive verbs but never a domain noun — was right and
  incomplete: the default's bare verbs are this board's own vocabulary.
  delete is a TypeScript operator, so a type-soundness question black-holed
  to an empty route. Anchoring the verbs to operational objects only moves the
  black hole onto "review the workflow that runs npm publish". The gate exists
  to stop auto-dispatch to a live-mutating agent and this board registers none,
  so it is omitted, as hr/legal/netsuite already omit it.

Generator changes, both scoped to their designed extension points:

• GATE_INTENT is now actually consulted per provider — its own comment
  promised an override but build_taxonomy() read ["default"] unconditionally,
  so no override could ever have taken effect. Key order is preserved for every
  other provider.
• and and never join STOPWORDS. The IDF filter cannot catch them (too few
  domains to cross the 25% threshold) yet they appear in nearly every real task,
  handing a free point to whichever domains carry them. Also removes two
  meaningless Never keywords from microsoft's taxonomy.

Also fixes agents/frontend/browser-compatibility-agent/metadata.json, whose
harnesses listed the invalid kiro-cli/kiro-ide split instead of kiro,
and drops ROADMAP's hardcoded catalog counts in favour of a pointer to the
generated docs/_data/catalog.yml.

Gates: npm run validate, lint:spell, markdownlint (7670 files), the routing
grader (751 scenarios across 32 maestros), and cargo fmt/clippy/test
(1053 tests) all green.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session:

• docs: generate the provider-reference table and repair stale provider listings (8c6566b)
  An audit of every count and provider listing in the repo, prompted by the new
  TypeScript board. The counts were all correct — readme-counts was green and
  re-running the generators produced a zero diff. The listings were not.

The provider-reference table in README.md is now generated, not hand-written.
It had drifted badly while nobody was watching: 27 of 45 providers, kubernetes
understated as 15, multi-cloud as 1, and a velero row claiming one agent for
a provider that has skills but no agents at all. Its counts summed to 334
against a real 680. scripts/generate-readme-counts.mjs now emits it between
provider-table:* markers, using the same marker contract the counts block
already used, sorted by agent count then slug for a stable diff.

validate:readme-counts now also runs the generator in --check mode, so this
class of drift fails CI instead of accumulating. The gate's own logic is not
duplicated — it shells out to the generator. Probed both ways: an injected wrong
count exits 1 naming the line, a clean tree exits 0.

Provider display labels live in the generator with a title-cased fallback, so
adding a provider never breaks the build. CLAUDE.md already lists eight places a
new provider must be registered; this is deliberately not a ninth.

Other listings repaired:

• README's opening paragraph named every language board except TypeScript.
• The --provider CLI flag row listed 19 of 45 values and included finops,
  which is not in the provider enum at all — --provider finops would be
  rejected. It now points at vfa-export-agents --list-providers, verified to
  exist and to print all 45 including typescript 14 agent(s).
• CONTRIBUTING.md pinned a 13-value provider list against a 48-value enum. It
  now points at the enum in schemas/skill.schema.json as the contract.
• The ## Agents table understated kubernetes as 15. Its FinOps, QA, and
  cross-functional rows are hand-curated aggregates with no provider slug and
  were deliberately left alone.
• docs/roadmap.md hardcoded the MCP-reference count in prose while the same
  file already used the Liquid variable two dozen lines above.

Verified unchanged and correct: the provider invariant holds exactly
(taxonomy.md bullets == catalog.yml provider_list == providers with >=1 agent,
45 each, typescript in all three), and velero is correctly absent from all
three since it is skill-only.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session:

• typescript: address five Codex review findings on the TypeScript board (44909d8)
  All five reproduced against the tree before being fixed; each fix carries a
  positive and a negative probe.

P1 — the documented regeneration path corrupted unrelated providers.
scripts/update-catalog-new-agents.py gains --provider <id> (repeatable), and
every board generator now documents the scoped invocation. Unscoped, the upsert
walks every provider and its merge is {**catalog_entry, **projected_metadata}
— projected keys win, which is wrong whenever metadata.json is the older
side. It is here: the ionos, ovhcloud, and scaleway agents each declare two
harnesses in metadata while all seven adapter files sit beside them on disk, so
an unscoped run silently rewrote 18 catalog entries from the stale side and
dropped three providers out of the Kiro Powers set. No gate caught it. The
unscoped path still exists for deliberate reconciliation but now prints a
warning naming the hazard. Probed both ways: scoped run is a clean no-op;
unscoped run still reproduces the harness downgrade.

The same trap was documented in four sibling generators (kotlin, netsuite,
python, python-live); all four now document the scoped command too. Docstring
only, no behaviour change.

P2 — declared routing vocabulary was dead data. Each specialist's data file
declared routing_keywords (118 terms across the board) that no generator ever
read, so the taxonomy was mined from agent ids and summaries alone. Mining
recovers what an agent is called, never the constructs it owns, and the gap
showed: Is satisfies safer than an annotation here? and a floating-promise /
cancellation review both returned unclassified, and "our package exports and
types conditions are wrong for ESM consumers" routed to publication-integrity
because the mined token package outweighed every resolution signal.

build_taxonomy() now reads routing_keywords from each agent's
metadata.json and ranks it ahead of the mined tokens; agents that declare none
behave exactly as before. Read from disk rather than the catalog on purpose —
the catalog projection carries a fixed key allowlist, so this vocabulary never
reaches catalog/agents.json and the TUI's deny_unknown_fields structs are
untouched. Verified: zero occurrences of the field in the catalog, and all four
of the reviewer's examples now reach the right specialist.

P2 — a release date became a routing signal. The miner's \w+-\w+ arm
matched the MCP protocol revision quoted in a summary, and because keywords
containing a non-word character match as substrings, 2026-07 then captured any
task mentioning that year-month: "Assess our TypeScript 7.0 migration planned
for 2026-07" routed to the MCP specialist. Date-shaped tokens are now filtered
out of mined keywords, and validate-maestro-routing.py rejects a date-shaped
keyword in any provider's committed taxonomy so one cannot re-enter by hand.
Only typescript was affected across all 32 maestros. Probed both ways: clean
tree passes 751 scenarios; an injected 2026-07 fails with the intended message
and exit 1.

P2 — the Kiro Power activated on generic prompts. The board took the
generator's derived fallback, which produced configuration-audit and
best-practices as activation keywords — enough for a database or Terraform
review with no TypeScript signal to pull in the TypeScript Power — plus a
truncated description and the wrong casing ("Typescript"). Replaced with a
proper PROVIDERS entry: narrow keywords that are either the language itself or
a construct unique to TypeScript/Node work, an accurate description, correct
casing, and the board's static-review invariants.

P2 — the hand-written Kiro inventories were stale. README.md advertised 42
Powers and omitted vanguard-typescript; the installation guide's table had no
TypeScript row. Both corrected, and the guide was also missing vanguard-java
— a pre-existing gap, now fixed. All three inventories (disk, README, guide) are
verified set-equal at 45.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session:

---

📥 Install

npm install @raishin/vanguard-frontier-agentic@3.8.0

🔐 Supply-chain provenance

Every release ships a build attestation (SLSA provenance) and an SBOM. Verify the tag with gh attestation verify before installing.

Full changelog: v3.7.0...v3.8.0