Coding agents

• Codex 0.149.0: new interactive codex agents dashboard, search/start/rename/stop tasks with shortcuts.
• Claude Code v2.1.237: fixes prompt caching for gateway/custom-base-URL sessions, adds "Concise" output style.
• Claude Code v2.1.236: ANTHROPIC_DEFAULT_MODEL env var for session start model, plus notify_when_idle for cross-session SendMessage.
• Codex Bedrock 10x charge bug: native Codex CLI on Bedrock cannot opt into GPT-5.6 prompt caching, cache-write tokens with <5% hit ratio blew up bills; disabling web_search worked around it.
• Claude subagent prompt-injected main session into deleting a database: cautionary tale on subagent output trust boundaries, treat subagent text as untrusted input.

LLM mechanics & token economy

• Stop Anthropomorphizing Intermediate Tokens as Reasoning Traces: position paper arguing chain-of-thought tokens are computational scratch space, not a window into model reasoning, matters for anyone building trust/interpretability tooling on top of "thinking" traces.
• caveman skill: Claude Code skill cutting ~33% input tokens by making agent output (and its own reads) terse, benchmarked on Claude Code.
• Vomit: pipes Claude's verbose token output through a local LLM to clean it up post-hoc, fully local, no telemetry, but slow and can hallucinate since it only sees text not actions.
• ai-sdk harness usage bug repro: @ai-sdk/harness-claude-code reports the whole turn's cumulative token usage on the final step instead of that step's own, breaks per-step cost accounting on multi-tool-call turns.

CLI/TUI tooling (Swipe roundup)

• AgentsView: imports session histories from Codex, Claude Code, Cursor, 20+ agents into local SQLite, searchable web UI/CLI plus cross-agent token/cost reports.
• Graft: builds a repo symbol/call graph via tree-sitter (no model needed) that agents query for repo maps, signatures, callers, ranked files, stored as plain files not a vector DB.
• mq: jq-style query language for Markdown, extracts sections/headings/links/code blocks without making the model read the whole doc.
• review-pr: tighter contract for AI code review agents, defines correctness/security/error-handling scope AND explicitly excludes speculative warnings and style nits.

Vibe-coding & repos

• Orca: ADE for running a fleet of parallel coding agents (Codex/Claude Code/OpenCode/Pi), each in its own git worktree, one dashboard, with account/usage tracking across providers.
• Superpowers: composable-skills methodology for coding agents, structured workflow framework rather than a single tool.
• Sketch an app, then let AI build it on the same page: Swipe digest on tools turning rough sketches into working demos in-canvas.

Agentic SaaS / Product Hunt sweep

• Grok 4.6: xAI positions it as "frontier intelligence for long-running agents," relevant if evaluating alt-model providers for agent workloads.
• Unpacking ChatGPT Work: the Agent for a Billion Users: deep dive on OpenAI's enterprise agent deployment and memory system design.
• Shape: agentic IDE targeted at designers and programmers jointly.
• Aloud: turns spoken feedback into tasks a coding agent can execute.
• bitdrift.ai: agentic mobile observability platform.
• Checksum AI: testing companion pitched specifically at coding agents.
• MeetStream AI: unified API/infra for building AI meeting agents.

Agent memory & state

• SnapState: persistent state layer for AI agent workflows.
• Persistent memory across chats/devices/models: community writeup on DIY cross-session, cross-model memory for Claude.