Claude
d0a8340717
A code-grounded robustness audit of the strategic spine (MCP gateway, A2A, x402) plus the two Mu-discovered foundations (in-process dispatch, durable agentic workflow). Headline: all four are well-built internally and fragile at the edge — the exposure surface the strategy rests on is the least externally-proven part of the codebase (no test drives a real external MCP host, third-party A2A SDK, or real x402 facilitator/wallet). - internal/docs/GAP_AUDIT.md — the findings, file:line + severity + what "robust in practice" requires, and a prioritized build order. - .github/loop/PRIORITIES.md — repointed so the loop pulls the well-defined strategic-spine fixes first (#4813 MCP stdio JSON, #4814 x402 budget-bypass, #4815 A2A conformance, #4786 buyer wiring), with the architectural items (#4816 durable agentic workflow, #4817 dispatch fast-path, MCP transport unification) marked needs-human as 1:1 development the loop must not auto-build. Reframes the queue around "advances the strategy vs grooms a proxy" — and makes explicit that the planner ranks a human-curated backlog, it does not invent work. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CmdEY7pYmV5zzwCjNJ4ykL
3.7 KiB
Priorities
The ranked work queue for the autonomous improvement loop. The planner ranks it; the builder works the top item whose linked issue is still open. Direction comes from the human (the roadmap and the gap audit) — the planner does NOT invent work, it ranks the curated backlog. The builder executes well-defined issues; it does not improvise.
Advances the strategy vs. grooms a proxy. Rank by whether an item advances
the actual strategy — real integration/exposure, real capability — not by whether
it produces a green increment. Making MCP speak MCP to an external client, A2A
interoperate with a real external agent, and x402 actually settle is real work.
Guarding docs the loop wrote or chasing a weak provider's quirks is grooming —
needs-human it.
Reading / editing. An item is done when its linked issue closes (the PR adds
Closes #<issue>). The human reorders this list or the issues at any time.
Off-limits to the loop (never auto-merged): brand/positioning copy, breaking
public-API changes, architectural rewrites. Items labelled needs-human below are
1:1 development work — the loop must not auto-build them.
Work queue (ranked)
Strategic spine — make integration/exposure actually robust (gap audit)
These are the surfaces the strategy rests on, and they're the least externally-proven part of the codebase. Highest value.
- MCP: stdio/ws tool results must be JSON +
isError, with a stdio test (#4813) — the path Claude Desktop uses currently returns Go%vmap-syntax instead of JSON and misreports tool errors. Cheapest, highest-impact fix. - x402: fix the budget-cap bypass + require a real Settler (#4814) — a malformed amount defeats the spend cap; verify-only serves the resource for free. Hardens the safety the flagship relies on.
- A2A: conform to external clients — well-known path + spec SSE events (#4815) — an external A2A SDK 404s on discovery and can't parse the stream. Real cross-framework interop.
- Agents that pay — wire the x402 buyer into the agent runtime (#4786) — the flagship capability; the buyer
Clientexists but is wired into nothing (the agent's "spend budget" is bookkeeping that never pays).
Capability — reach & deployment (roadmap: Next; builds on the spine)
- gRPC-reflection MCP (#4796) — expose external reflected gRPC services as MCP tools.
- Kubernetes operator + CRDs foundation (#4797) —
Agent/Service/Flowas native K8s resources.
Human-led — real 1:1 development (needs-human; the loop must NOT auto-build these)
- MCP transport unification (gap audit item 2) — mount the JSON-RPC handler as the HTTP transport and run all transports through one pre-call pipeline (auth→rate→breaker→payment). Architectural.
- Durable agentic workflow: HITL pause + per-tool-call checkpointing (#4816) — the convergence leg; core primitive design.
- In-process dispatch fast-path (#4817) — a local transport so in-process calls skip codec + network hop.
Evidence base: internal/docs/GAP_AUDIT.md (this session's code audit) + the "requirements discovered from Mu" notes. Restocked by Claude Code; the planner ranks, it does not invent.