Amaleh / architecture atlas / 19 Sep 2026

One coordinator.
A visible path to done.

Who acts, what gets called, and where every result returns. Start with the implemented command path; other tabs explain the wider design policy.

IMPLEMENTED · HOST-DRIVEN LOOP

Select a card to trace its routes. Numbers match the connection list.
100%
Scroll the diagram in both directions. All nodes are also keyboard accessible.
Work / evidenceReturn to coordinatorFallback / repair / resume
Author → reviewer DeepSeek ↔ GLM FlashDefault review one fresh pass with relevant lensesDefault repair limit 2 Flash cycles → Kimi → main model
Review policy: borrowed questions, controlled execution

Matt Pocock’s Spec and Standards axes remain separate in the findings. Fresh-review contributes omission tracing, concrete failure scenarios, conditional lenses, and evidence-based refutation. Amaleh adapts those prompts: one opposite-family reviewer normally covers the baseline lenses; sensitive changes can add a focused parallel reviewer. It does not invoke both original pipelines unchanged.

Initial review → repair 1 + targeted recheck → repair 2 + targeted recheck → Kimi root-cause diagnosis before a third Flash repair. After one unresolved Kimi diagnosis/repair cycle, the main Astra/Fable takes over. No automatic full re-review of unchanged work; required coverage and real defects remain visible. Counters survive restart and superficial task renames. These strategy limits never authorize shipping a known failure.

Matt Pocock primary source · Review questions are bundled in the skill.

The contract for a completed result
  • The invoking host owns the original outcome until delivery. A worker's “done” is a claim, not proof.
  • Inspect artifacts and run required checks. Verify money calculations and reachable payment behavior against explicit expected outcomes.
  • Each defensive change needs a stated requirement, observed failure, or reachable failure path. No invented future requirements.
  • Independent work progresses when its own dependencies are met. Coupled changes and conflicting resources stay serialized.
  • Repair findings, inspect only affected changes again, and escalate a repeating failure to a different diagnosis or plan.
What is deliberately not a mandatory pipeline

Superpowers and BMAD are research references, not required chains of phases. Small work can remain with the host. Research, architectural review, specialist review, and user questions happen only when the actual task requires them. No nested reviewer tree or universal one-agent-per-file rule.

Sources and evidence limits

Jev decision primitives · Jev limitations · Compaction reference · Provider routing · Frontend evaluations · Kimi usage contract

Candidate strengths are sourced starting hypotheses. Live end-to-end fixtures and controlled regression tests have been performed; they are not a general model-quality benchmark. See bundled verification notes for tested paths and remaining host/platform limits.

Runtime, installation, and permission boundaries
  • TypeScript core. Bun first, Node fallback. Platform APIs handle paths and subprocess arguments.
  • One canonical amaleh skill, linked into Claude and Codex. Global pi model settings stay intact.
  • Skill links do not install lifecycle hooks. Hook and pi extension registration is a separate, verified installation step.
  • No artificial total spending limit. Real credit exhaustion checkpoints the run and requests intervention; transient failures receive bounded retries.
  • Jev uncertainty falls back to the invoking host. Permission, graph legality, resource ownership, and test results remain deterministic.
How we will test this design

Prove independent task progress, conflicting-resource serialization, host fallback, restart without duplicate work, repeated compaction without losing constraints, image-capable routing, provider metadata preservation, and refusal to mark failed checks complete. Check Bun/Node parity and report platforms not exercised.

Measure total time, tokens, retry cost, repeated checks, and verified outcomes. A smaller transcript or a cheaper token rate alone does not establish an improvement.

Interpret the host selector correctly

Astra and Fable illustrate the two requested host families. The installed adapter inherits the actual invoking model and effort when available; it does not switch every Claude session to Fable or every Codex session to Astra. Missing identity requires an explicit configuration rather than a guessed model.