Skip to content
AXIOMAXIOM

AXIOM integration research — report

Note on method: no terminal tool is available in this mode, so rg could not be run across the sibling repos; findings below come from directory listings and targeted file reads (VERIFIED where stated). Items requiring a true text sweep (3, 5) are marked as partial.

1. MCP configuration / consumption per repo

Section titled “1. MCP configuration / consumption per repo”
Repo MCP config file Servers declared MCP packages (server side)
brivio E:\gh\brivio\.vscode\mcp.json (VERIFIED) postgres-local (stdio, @modelcontextprotocol/server-postgres → dev DB on dragos:22432), sentry (http, https://mcp.sentry.dev/mcp) apps/mcp = @brivio/mcp v0.3.0 — uses @modelcontextprotocol/sdk (catalog), Hono http transport, OTel. 4 meta-tools over the gateway OpenAPI (brivio_api_search/details/read/write) + curated tools; writes disabled by default (BRIVIO_MCP_DISABLE_WRITE). README: E:\gh\brivio\apps\mcp\README.md
metu none (.vscode/ has extensions/launch/settings/tasks only) none apps/mcp-server = @metu/mcp-server 0.0.1, @modelcontextprotocol/sdk (catalog), depends on @metu/core (Conductor tools), @metu/db, @metu/auth
mmo none (.vscode/settings.json, tasks.json only) none packages/ai-mcp = @mmo/ai-mcp — stdio / HTTP+SSE / WS transports, PAT scopes (daw:read, generate:audio…); README says “Scaffolded in P0”
codai none (.vscode/: extensions, settings, tasks — settings.json has no mcp key) none No *mcp* package; but apps/desktop/mcp/ exists (codai-phone-mcp.mjs per packages/agent-core/spec/tools.json description — a phone-control MCP shim)
money none none none
vmui none (no .github/ dir at all) none none

User-level MCP config: C:\Users\vladu\.copilot\mcp-config.json declares tavily, sentry, context7, microsoft-learn, chrome-devtools, phone-chrome (+ more past line 60). This is the most plausible place to register AXIOM as a local stdio server ("type":"local","command":"npx","args":["-y","@codai/axiom-mcp"]). Note: that file contains live secrets inline (Tavily, Sentry tokens) — not reproduced here.

Repo-level instruction files that mention MCP: E:\gh\brivio\.copilot-ripple.json explicitly names apps/mcp/src/tools.ts (68 hand-written tools) as a surface that must be mirrored manually — a documented, unmet automation gap.

2. codai — tools / agents / sandbox concepts

Section titled “2. codai — tools / agents / sandbox concepts”

VERIFIED from listings + reads:

  • apps/sandbox (@codai/sandbox, Hono): “Frontier v3 B4 — the containment leg”. A Cloud Run service with zero secrets/IAM that runs UNTRUSTED code: POST /exec { language: 'python'|'shell', code, timeoutSeconds }, auth via x-sandbox-token. It is an execution sandbox, not a plugin/tool host — AXIOM does not fit inside it, but AXIOM’s apply step could target a sandbox workspace. Files: E:\gh\codai\apps\sandbox\src\server.ts, package.json.
  • packages/agent-core — shared agent-loop contract: spec/agent-core.json, spec/tools.json, spec/tools-v2.json, spec/capabilities.json. Defines RiskClass (READ|ACT|SENSITIVE), AutonomyMode (strict|assisted|autopilot), APPROVAL_MATRIX, safety prefixes (NEEDS_APPROVAL:, REFUSED:, BLOCKED:, DRY-RUN:). This is the closest thing to a tool registry (E:\gh\codai\packages\agent-core\src\index.ts, spec/tools.json). An AXIOM tool set (axiom.check, axiom.apply) would need a risk class here (apply = ACT).
  • packages/policy-engine — spend mandates for agent payments (Mandate, Intent, Decision = allow|deny|needs_approval, reason codes OVER_DAILY_CAP, DENYLISTED…). Money-only; not code policy. E:\gh\codai\packages\policy-engine\src\types.ts.
  • packages/rules-core — signed rules manifests: JCS canonicalization, DSSE envelope (application/vnd.codai.rules+json;v=1), floor.ts, catalogue.ts, glob.ts. This is the “rules layer” the user’s guard-rules-surface.ps1 hook enforces. AXIOM’s policy profiles (profiles/*.json) would ideally be expressed/signed the same way.
  • packages/swe-harness — SWE eval harness with file-tool.ts (text-protocol READ: requests, path-escape rejection), workspace.ts, sandbox-env.ts, verdict.ts, pipeline.ts; apps/swe-exec-runner executes it. Closest analogue to AXIOM’s “apply files to a repo” step.
  • packages/verify — eval verification (task-validity, verifier-score), not file verification.
  • Other relevant apps: apps/desktop (Tauri, has mcp/ dir), apps/mobile-agent, apps/phone-android, apps/agent-pay.
  • .github/copilot-instructions.md lists no MCP/tool-registry convention; .agents/ is empty.

3. Existing codegen / scaffolding

Section titled “3. Existing codegen / scaffolding”

Partial (no rg sweep). VERIFIED by listing:

  • brivio: scripts/sdk/generate-sdks.mjs + openapi-sync.mjs (OpenAPI → TS/PHP/Python/Go/Java/.NET SDKs, packages/sdk-*/generated/); scripts/codemods/ (4 jscodeshift-style codemods: font-size-to-scale, pulse-to-skeleton, suspense-reveal, z-index-to-scale); scripts/gen-tenant-tables.ts; generate-*.mjs (icons, OG PNG, inventory, VAPID keys). No plop/hygen/turbo-gen seen.
  • codai: packages/agent-core/scripts/ (gen-*.mjs generators reading spec/*.json per the index.ts comment).
  • metu: .github/skills/add-app-page, add-db-migration, add-agent-tool, add-sdk-endpoint, add-integration, add-inngest-function — prompt-driven scaffolding recipes (SKILL.md), not tooling. This is the direct conceptual overlap with AXIOM’s intent DSL → scaffold: metu already encodes “how to add X” as skills for agents to execute by hand.
  • money, mmo, vmui: none found in listings.

Verdict: no generic scaffolder exists; AXIOM would not duplicate tooling, but AXIOM’s emitters (Next.js 14 pages + React 18, node:http) are off-stack for every repo (Next 16 App Router, Hono 4, Drizzle) — per repo memory axiom-overview.md.

4. Existing policy / compliance / PII checks

Section titled “4. Existing policy / compliance / PII checks”
  • brivio packages/compliance: fiscal-only — bon-fiscal.ts, state-machine.ts, fiscal-printer/{datecs,tremol}.ts. Not PII/code policy.
  • brivio packages/trust: engine/{evidence,health,signature,timestamp,validation}.ts, crypto/, signers/, archive/, identity/, tenant/ — document trust/e-signature, not code policy.
  • brivio packages/safe-mode, scripts/safe-mode/scan-real-emails.mjs — env-gating + a real-PII scan of seed data (closest PII analogue).
  • brivio scripts/check-*.mjs (~75 guards, run via scripts/run-guards.mjs, pre-commit + CI): check-audit-coverage, check-org-context, check-consent-gating, check-ropa, check-sub-processors, check-hardcoded-strings, check-egress-guard, check-licenses, check-vacuous-assertions, check-sdk-coverage… This is the repo’s real policy layer; AXIOM’s check() would duplicate it unless AXIOM calls into these guards or emits results in their format.
  • codai packages/policy-engine — spend caps/mandates (money). packages/rules-core — signed rules manifests. Gateway spend caps per user (copilot-instructions). No code-PII scanner.
  • money: apps/quant/app/policy.py — trading acceptance gates, unrelated.
  • metu: packages/core/src/agent/ ACL policy via runTool() (per agent-and-conductor.instructions.md).

Verdict: AXIOM’s PII scan has no direct equivalent (only brivio’s seed scan); its SLA/budget policies are unrelated to codai’s spend policy-engine. Overlap risk is with brivio’s check-*.mjs guard suite.

5. Verified file apply / sha256 / atomic write

Section titled “5. Verified file apply / sha256 / atomic write”

Partial (no rg sweep). VERIFIED:

  • E:\gh\codai\apps\sandbox\src\server.ts — createHash('sha256') for constant-time token compare only.
  • E:\gh\codai\packages\rules-core\src\{canon,envelope}.ts — canonicalHash (JCS + sha256) and DSSE signature envelopes for manifests. Closest match to AXIOM’s manifest+sha256 verification pattern.
  • E:\gh\codai\packages\swe-harness\src\{file-tool,workspace}.ts — repo file read/apply for evals (SEARCH/REPLACE text protocol, path-escape rejection).
  • brivio: packages/trust/src/engine/{signature,timestamp}.ts (document hashes), audit_log hash chain (ADR-0005). No generic atomic-write/apply-patch utility found.
  • User-level: C:\Users\vladu\.copilot\hooks\guard-write.ps1 / guard-tooluse.ps1 block clobbering whole-file writes — the house “safe apply” gate is a hook, not a library.

6. Agent config — house patterns

Section titled “6. Agent config — house patterns”

C:\Users\vladu\.copilot\

  • agents/: brivio-feature.md, codai-gateway-ops.md, codai-training.md, code-reviewer.md, db-change.md, Deep Build.md, Goal.agent.md, incident-responder.md, plan.md, Prompt.md, release-manager.md, upgrade-migrator.md
  • instructions/: core.instructions.md
  • skills/: codai-byok-ops/, gcloud-auth-verification/, managing-python-dependencies/
  • hooks/: codai-hooks.json (registry), gates: guard-tooluse.ps1, guard-rules-surface.ps1, guard-command.ps1, guard-write.ps1, no-permission-stop.ps1, nudge-parallel.ps1, copilot-signal.ps1, session-start.ps1; utilities: run-build.ps1, who-owns-file.ps1, rules-ticket.ps1, check-ripple.ps1, sync-to-vm.ps1, audit-tools.ps1, measure-*.ps1, test-*.ps1 (one test per gate), lib/, known-tools.txt, README.md.

C:\Users\vladu\.claude\: rules/ (41 files listed above, e.g. agent-behavior.md, security.md, ripple-completeness.md), skills/ (42 dirs, e.g. code-review, debug-detective, security-audit, skill-creator).

Hook/gate house pattern (from codai-hooks.json + headers):

  • Registered in codai-hooks.json: {"version":1,"hooks":{"sessionStart":[…],"preToolUse":[…],"Stop":[…]}}, each {"type":"command","powershell":"C:/Users/vladu/.copilot/hooks/x.ps1","timeoutSec":N}. Event casing preToolUse/sessionStart/Stop is load-bearing.
  • Script contract: PowerShell <# .SYNOPSIS / .DESCRIPTION #> header stating the measured incident that motivated it; $ErrorActionPreference='Stop'; reads tool payload JSON on stdin; exit 0 = allow, exit 2 = deny with reason on stderr (RO+EN, naming a floor rule id like codai.floor.self-modify-rules); budget 4–15 s; each gate has a test-*.ps1 mutation test; test-config-integrity.ps1 after any change.
  • Repo-level equivalents: brivio .copilot-ripple.json (overrides for check-ripple.ps1: publicApi regex, ignorePaths) and scripts/run-guards.mjs running check-*.mjs in pre-commit. An AXIOM repo hook should follow: check-axiom-*.mjs guard in scripts/ + optional .copilot-*.json override file.

Frontmatter formats (VERIFIED):

Instructions — E:\gh\brivio\.github\instructions\brivio-conventions.instructions.md:

---
description: Brivio repo conventions every change must follow
applyTo: "**"
---

metu variant with scoped globs — E:\gh\metu\.github\instructions\agent-and-conductor.instructions.md:

---
applyTo: 'packages/core/src/agent/**,packages/core/src/goals/**,packages/ai/**,...'
description: Conductor agent — tools registry, ACL policy, planner, AI SDK v5, BYOK provider mesh.
---

Skills — E:\gh\metu\.github\skills\add-agent-tool\SKILL.md:

---
name: add-agent-tool
description: Register a new tool the Conductor (or any agent) can call — schema, kind, undo, ACL hookup. Use when ...
---

(brivio brivio-wiring-guards/SKILL.md uses the same name + folded description: >-.) Agents (~/.copilot/agents/*.md) add tools: [...] to the frontmatter.

Repos with these dirs: brivio (instructions/ ×2, skills/ ×21), metu (instructions/ ×14, skills/ ×10), money (skills/ ×3, no instructions). codai/mmo/vmui: none (codai has only copilot-instructions.md; vmui has AGENTS.md).