2026
-
Codex Source Dive (V): The Security Model
Codex security is a layered execution membrane: sandboxing defines the technical boundary, approvals decide review, exec policy classifies command risk, and OS backends enforce the result.
-
Codex Source Dive (IV): The Tool Runtime
Codex tools are model-visible schemas backed by a policy-bound runtime that routes shell commands, patches, MCP, skills, permissions, sandboxing, streams, and history through one managed execution path.
-
MiniMax Sparse Attention: Teaching Long-Context Models to Use an Index
MiniMax Sparse Attention turns long context into searchable memory: a learned index selects relevant key-value blocks, then exact softmax attention reads only those blocks.
-
Self-Harness: Harnesses That Improve Themselves
Self-Harness is best read as evidence-driven harness versioning: mine failed traces, propose bounded scaffold edits, and promote only changes that survive regression tests.
-
Codex Source Dive (I): the agentic loop is a runtime boundary
TLDR: A Codex turn is not one model call. It is a managed execution window where user input, tool calls, tool results, cancellation, compaction, and final answers are ordered by the runtime.
-
Codex Source Dive (II): goals are runtime state, not prompts
TLDR: A Codex Goal is a thread-level long-running task state machine. It stores objective, status, budget, usage, resume state, and continuation gates instead of relying on one remembered prompt.
-
Codex Source Dive (III): subagents are a thread tree
TLDR: A Codex subagent is not a background model call. It is a persistent child thread with identity, inherited runtime policy, forked context, mailbox communication, capacity limits, and resume behavior.
-
A reading stack for the AI-coding era
TLDR: As code generation gets cheaper, the scarce skills move toward judgment: architecture, strategy, systems thinking, safety, measurement, customer truth, organization, and cooperation.