Show HN: Log-decisions – an append-only decision journal for AI agents

A new open-source Agent Skill called log-decisions provides an append-only DECISIONS.md journal for AI agents to record consequential judgment calls, with decide/assume/escalate rules. The skill, compatible with Claude Code, Codex, Gemini CLI, and Cursor, was extracted from the swe-workflow repository (≤ v2.0.2) and is available under the MIT license. It aims to give agent-assisted work a durable, grep-able home for decisions that specs don't settle.

An append-only DECISIONS.md journal recording the consequential calls a spec didn't settle — what was chosen, by whom, and why, with decide / assume / escalate rules for what an agent may settle alone. General-purpose. The journal isn't coding-specific: any agent-assisted work — research, writing, ops, data — produces judgment calls someone will later ask "why was it done this way?" about. This skill gives those calls one durable, grep -able home at the repo root. It ships as an Agent Skill — pure markdown, no scripts — so it runs on Claude Code, Codex, Gemini CLI, Cursor , and any other skills-compatible agent. npx skills add swe-workflow/log-decisions /plugin marketplace add swe-workflow/log-decisions /plugin install log-decisions@log-decisions The skill skills/log-decisions/SKILL.md /swe-workflow/log-decisions/blob/main/skills/log-decisions/SKILL.md defines: The bar — if the spec already authorized the call it's not journal-worthy; if the agent had to invent the authorization, it is. The 2×2 — classify each call as determinable × reversible , then decide grounded by an artifact , assume safe default, logged for async review , or escalate stop and ask . A catastrophic floor — data loss, destructive migration, irreversible spend, public-interface breaks — always escalates. The entry — one append-only block per decision: Question, Options considered, Chosen, Decided-by, Justification, Outcome, with Supersedes: for revisions. Never edited, never reordered. Discussion — Show HN thread https://news.ycombinator.com/item?id=49264597 Claude Code plugin — install with the commands above; submitted to Anthropic's Plugin Directory https://code.claude.com/docs/en/discover-plugins pending review swe-workflow https://github.com/swe-workflow/swe-workflow — the idea → PRD → issues → ship suite — orchestrates this skill across every stage: spec-layer grills journal their assumed answers through it, and ship builds stage entries to a per-worktree DECISIONS.staged.md promoted at close-out. It was extracted from that repo ≤ v2.0.2 into this standalone one.MIT