Skip to content

Latest commit

 

History

History
35 lines (25 loc) · 3.96 KB

CHANGELOG.md

File metadata and controls

35 lines (25 loc) · 3.96 KB

Changelog

Notable changes to the Build framework. Newest first.

Latest - agent-agnostic

Build was written in and for Claude Code. The skills were always written to the Agent Skills open standard, but the docs and file layout assumed one tool. This release makes the framework's portability real and explicit.

What this means for you: you can now run the whole workspace in Codex, opencode, Cursor, Gemini CLI, or any other agent that implements the Agent Skills and AGENTS.md standards - without editing a single skill. If you're on Claude Code, nothing changes; CLAUDE.md still works, it's just a symlink now.

  1. AGENTS.md is the canonical instruction file, with CLAUDE.md symlinked to it. AGENTS.md is the cross-tool convention (Codex, opencode, Cursor, ~20 others); Claude Code reads only CLAUDE.md, and Anthropic documents this symlink as the supported bridge. Edit AGENTS.md, never the symlink.
  2. Skills are reachable at both discovery paths. They still live at .claude/skills/, with .agents/skills symlinked to the same folder, because Codex scans .agents/skills and never .claude/skills. One copy, no drift.
  3. New docs/portability.md - what runs where, per-tool setup, and an honest tested-vs-assumed table. Verified: Claude Code and opencode (all 8 skills load unmodified, per opencode debug skill). Assumed: everything else on the spec's client list.
  4. templates/PROJECT-CLAUDE.md renamed to PROJECT-AGENTS.md, and onboard-project/tune-persona now generate a per-project AGENTS.md plus its CLAUDE.md symlink. Claude Code discovers nested CLAUDE.md but not nested AGENTS.md, so both are needed.
  5. Docs reworded tool-neutrally. The Claude Code specific parts (auto-memory, plugins, SessionStart hooks) are now flagged as such, with guidance on what to swap. None of them are load-bearing.
  6. Windows note: the repo uses symlinks, which need Administrator or Developer Mode. SETUP.md and docs/portability.md document the @AGENTS.md import fallback.

Reconciled from live use

Reconciled ahead of the public release:

  1. New review-prep skill ported from live use: dispatches a fresh-context agent to read a project's STATUS/docs/decisions and produce a pre-review punch list (pulse, decisions waiting, decisions closed, stale/at-risk flags, the 3 questions for the meeting). The fresh-context dispatch is the point - a cold reader catches STATUS drift and cross-artifact risks the operator's own working memory misses. Validated across two project shapes before porting.
  2. Optional templates from real projects: STRATEGY.md (durable vision/roadmap), EXPERIMENT.md (thesis-under-test), CONTEXT-BRIDGE.md (PM <-> separate engineering repo), and FREELANCE-ENGAGEMENT.md (billable engagement log).
  3. onboard-project now generates a per-project persona file from the persona interview as a first-class output, not an optional afterthought.
  4. process-document gained transcript handling - detects call transcripts and runs an ASR-cleaning sub-agent (incl. non-English translate-to-English) before summarizing.
  5. Richer DECISION.md - status vocabulary now aligns with the standard ADR lifecycle (proposed / accepted / deprecated / superseded / revisiting).

v2 - structural refresh

  1. CLAUDE.md is now workspace-level, not per-project. The per-project version moved to templates/PROJECT-CLAUDE.md.
  2. New WORKSPACE.md explains why Claude should run from the workspace root, not each project folder.
  3. New docs/ section covers the auto-memory system, Obsidian integration, and the plugin ecosystem.
  4. Two new skills ported from live use: tune-persona and git-week.
  5. templates/WAYS-OF-WORKING.md added (the triad of BRIEF/STATUS/WAYS was referenced but the template was missing).
  6. examples/adscale-case-study/ - a fully anonymized real project, included inline so you can see the framework in action.