Engineering

Internals: the MCP server, graph extractors, hooks, and state model.

• Ablation Protocol — does the kernel improve output, or only token count?Pre-register the experiment that answers the one question the token
• Adapter Parity - Claude vs Codex CoverageDefine equivalent behavior across Claude and Codex without claiming that their runtime APIs are identical.
• Agent Token-Economy & Identity Roadmap
• Agent Hub & Multi-Agent OrchestrationCanonical contract for the task-system web UX, multi-agent concurrency stabilization, the session↔transcript id bridge, and the live-agents/traces surfaces. This is the doc-anchor for the `agent-hub` epic (TASK-167+).
• Agent SupervisionDefine the opt-in, adapter-neutral control plane that routes role work
• Bash 5.3.9 + Heredoc Deadlock — Forensics & Discipline
• boardos ↔ coding-os.db Coupling ContractDefines the read/write contract between board_os and the shared SQLite owned by thinking_os.
• CI Quality Gates — SSOT
• Codex GUI Presence Fallback
• Concept Graph (coedit / conceptlink)Contract for the `concept_graph` table — the lightweight file/concept
• Config Composition — per-stack .coding-os/ merge at cos init
• Context Budget — what coding-os costs before it saves anythingState, per project profile and from measurement, how many tokens the
• Derived-Store Coherence Audit — 2026-06-30
• Cortex Design System
• Destructive-Edit Guard — Friction Before DestructionCanonical contract for the `warn-destructive-edit` hook — a PreToolUse
• Dispatcher ContractDefine the provider-neutral request/result protocol and the adapter-owned execution responsibilities for formula agents.
• Doc-System Overhaul — Audit & Roadmap (2026-06)
• Extension Manager — Architecture & Security/Trust ModelSSOT contract for adding / removing / uploading **skills** and **MCP servers** to a coding-os project from the Hub panel — the unified `CatalogUnit` model, the per-project `extensions.json` manifest, the trust state machine, the fail-closed security gate, the API + UI surface, the adapter-parity fan-out, and the phased P0–P5 rollout.
• Graph-OS Query Guide
• Graph-OS — Hallucination Cures & Token Economics
• Graph-OS — Complete Use-Case Catalogue
• Hooks Reference — Every Hook, Every CategoryCanonical catalog of every hook coding-os ships. Lists what it fires on, what it blocks vs warns, and where its source lives. This is the file to read when `cos hooks-log` shows a hook you don't recognize, or when deciding whether to add a new hook vs extend an existing one.
• Hub ArchitectureCanonical contract for the singleton hub daemon (FastAPI + React SPA on :9188), per-project routing via `/api/p/<slug>/*`, and the propagation matrix for live symlinks (hooks/rules/skills/commands).
• Hub Threat Model — localhost trust boundary (TASK-363)
• Learning ExtractionCanonical contract for what `cos_learn_extract` produces — the
• loggingos — Central Log Helpers
• MCP Error Envelope — cos Tool Response ContractCanonical response shape for every `cos_*` MCP tool exposed by
• MCP Server Fast-Path Entry — cos-mcp-start
• MCP Schema Traps — Exact Types for Agent Callers
• Modularity / Auto-Sync Audit — June 2026
• Modularity / Auto-Sync Audit — July 2026
• Naming ContractDefine the canonical names for Coding OS subsystems so paths, config,
• Observability Eye — Enterprise Error & Log Pipeline
• Polyglot Extractor — Post-Ship Audit (2026-05-12)
• Project Anatomy — Polyglot Coexistence Contract
• Using RTK (Rust Token Killer) with coding-os
• Rules Loading — How src/core/rules/.md Actually Become ActiveExplicit, honest answer to "when is a rule file actually loaded, and by whom?" — because the files in `src/core/rules/` carry YAML frontmatter (`globs`, `alwaysApply`) that *looks* like auto-loading but neither Claude Code nor Codex consume that frontmatter.
• Scheduled JobsCanonical contract for the nightly maintenance pipeline — `src/core/scheduled/nightly.py` — and the optional CRON B agent (Claude Code `CronCreate`).
• Skill Architecture — Fundamentals + Stack SpecializationCanonical explanation of how coding-os organizes skills. Two-layer model: **agent-agnostic fundamentals** in `src/core/skills/` + **stack-specific specializations** in `src/templates/<stack>/skills/`. A stack skill declares `depends_on: [fundamentals...]` in its frontmatter so the agent loads both and gets DRY guidance that specializes where it matters.
• State Files — Design, Ownership, and Multi-Agent IsolationCanonical explanation of what lives under `.coding-os/`, how state files prove ownership, how two agents running against the same project stay isolated, and why the `.claude/.*` state files are gone.
• Template Enforcement — enforce-template.shCanonical reference for the PreToolUse hook that BLOCKS raw `Write` on four specific markdown classes, redirecting the agent to the right bootstrap flow. This is the "templates exist and must be used" policy — it does NOT touch ad-hoc docs (playbooks, engineering rules, runbook notes, session checkpoints).
• Template Location — In-Repo Files vs In-CLI vs HybridAnswer "should coding-os ship template files into every consumer project, or should templates live inside the `cos` CLI and be materialized on demand?" with explicit personas, scenarios, and a recommendation grounded in how modern scaffolders handle the same tradeoff (Rails, Next.js CLI, Nx, Cookiecutter, Yeoman).
• Test Governance — Multi-Agent-Safe, Laptop-Safe Suite Execution
• Third-Party Token-Cost Benchmark — graph envelope vs a graph-less agentMeasure, reproducibly and on public repos, what a `cos_graph_*` envelope
• Workflow Audit — V1 Flowchart vs Current Source