Skip to content

knowledge.md

Latest commit

 

History

History
26 lines (23 loc) · 4 KB

knowledge.md

File metadata and controls

26 lines (23 loc) · 4 KB
Raw

daily-dev-agentic knowledge — T agent

2026-08-24

  • Structure Claude Code fleet automation around its loop primitives: goal-based loops that iterate until a measurable, deterministic condition passes, and time-based loops (/loop, /schedule) for recurring monitoring — Osmani runs 5–10 parallel agents this way under a four-tier taxonomy (agentic, goal-based, time-based, proactive).
    • Why it matters here: This is the CCCLI fleet pattern named and field-tested at 80k-repo scale; the daily-dev learn loop itself is a time-based loop, and IPI-Engine promotion gates fit goal-based loops ('iterate until DSR criterion passes').
    • Do: Define each fleet job as goal-based (measurable finish line) or time-based (cadence) before writing its prompt.
    • Source: Practical Loop Engineering — Addy Osmani (⬆️ 67 · 💬 4) · tags: ai-agents, claude-code, prompt-engineering · confidence 80%
  • Run agent fleets like a software factory: triage labels, explicit handoff protocols, a per-run verification budget, a run-status taxonomy (success / flawed / blocked / manual), and sandboxing for untrusted inputs — human judgment relocates to product intent, architecture, quality bar, and risk.
    • Why it matters here: Maps 1:1 onto the evidence-gated CCCLI doctrine; the run-status taxonomy is a ready-made vocabulary for fleet reports and matches the existing blocked-vs-failed rule.
    • Do: Adopt success/flawed/blocked/manual as the standard verdict field in every fleet run report.
    • Source: Human judgment doesn't leave the software factory. It relocates. — Addy Osmani (⬆️ 8 · 💬 1) · tags: ai-agents, cicd, claude-code, code-review · confidence 75%
  • At agent code volume, enforce quality through gates built into the harness and pipeline — unit/property/mutation tests, complexity metrics, lint-enforced architecture rules, compiler/CI back-pressure applied throughout the loop — not through human review at the end.
    • Why it matters here: Second independent source in two days pointing the same way as the custom-linters lesson (2026-08-23): deterministic gates beat LLM review for scale; reserve human judgment for subjective calls.
    • Do: Add mutation testing and complexity ceilings to the CCCLI quality-gate list.
    • Source: Agentic Code Quality — Addy Osmani (⬆️ 8 · 💬 0) · tags: testing, architecture, cicd, ai-coding · confidence 75%
  • Give agents ground-truth files they may not invent (Nurb keeps real-world dimensions in measurements.toml so agents cannot silently guess) plus per-artifact markdown 'cards' recording decisions and auto-logged facts as memory for future sessions.
    • Why it matters here: The pattern generalizes beyond CAD: any fleet job should split 'facts the agent must not fabricate' into a machine-checked file and keep a per-artifact decision log — same shape as this loop's knowledge.md.
    • Do: For IPI-Engine and CCCLI jobs, keep venue limits/constants in a checked data file, never in prompts.
    • Source: A deep dive into Nurb — Flavio Copes (⬆️ 19 · 💬 0) · tags: python, ai-agents · confidence 60%
  • Oxc now ships React Compiler support: 22 compiler-powered lint rules in Oxlint and oxc-transform-react applying automatic memoization >10× faster than Babel, integrated in @vitejs/plugin-react v6.1 for Vite 8.
    • Why it matters here: Stack-adjacent: only relevant where React frontends exist (e.g. canvas/console UIs); if one appears, the Rust-based Oxc chain is the current fast path.
    • Do: None now — note for the next React build.
    • Source: React Compiler Support — Oxc (⬆️ 7 · 💬 0) · tags: react, rust, vite · confidence 65%