Generated with timbers draft sprint-report --since 2026-02-10 --until 2026-02-14 | claude -p --model opus

Sprint Report: Feb 10–14, 2026

Summary

Timbers shipped six releases (v0.5.0 through v0.10.0) across five days, anchored by two major architectural changes: a complete storage pivot from git notes to file-per-entry storage under .timbers/, and a new MCP server exposing 6 tools over stdio. The sprint also overhauled hook integration, added a notes field for capturing design deliberation, and launched a marketing landing page.

By Category

Storage Pivot

  • Replaced git notes with .timbers/<id>.json flat files — atomic writes, no merge conflicts, simpler GitOps interface (8 methods → 4)
  • Migrated 37 existing entries from git notes to file storage
  • Added YYYY/MM/DD directory layout for scale, with 14 integration tests proving merge safety
  • Filtered ledger-only commits from GetPendingCommits to break the chicken-and-egg loop where entry commits appeared as undocumented
  • Removed all git notes code and references

MCP Server

  • timbers serve subcommand using go-sdk v1.3.0 with 6 tools (log, pending, query, show, status, prime)
  • Handlers call internal/ packages directly — no CLI wrapper overhead
  • Added idempotentHint to read-only tools for client caching

Hooks & Agent Integration

  • Rewrote Claude Code hooks from shell scripts to JSON settings.json format (the old approach was a complete no-op)
  • Expanded from single SessionStart hook to 4 events: SessionStart, PreCompact, Stop, PostToolUse
  • Fixed PostToolUse hook to read stdin instead of empty $TOOL_INPUT env var
  • Switched hooks from global to project-level scope
  • Added graceful degradation — warns with install URL when timbers isn’t on PATH

Coaching & Entry Quality

  • Added --notes field for capturing deliberation (journey to a decision, distinct from --why verdict)
  • Rewrote coaching: motivated rules (explain why each rule exists), concrete 5-point notes trigger checklist, XML section tags, reduced imperative density
  • Tightened --why coaching to differentiate from --notes
  • Added PII/content safety guardrails to prime output

CLI & Developer Experience

  • --color flag (never/auto/always) for terminal color scheme compatibility (Solarized Dark fix)
  • Auto-commit entry files in timbers log — eliminates the staged-but-uncommitted gap
  • Renamed prompt → draft command; added ADR decision-log template
  • Added changelog command with tag-based grouping
  • Centralized config directory with cross-platform support (XDG_CONFIG_HOME, Windows AppData)
  • .env.local support for API keys (avoids Claude Code OAuth conflicts)
  • Enhanced doctor with config checks and GitHub release version check
  • Routed errors/warnings to stderr when piped

Architecture

  • AgentEnv interface with registry pattern — each agent environment (Claude, future Gemini/Codex) is self-contained via init() registration
  • Refactored init, doctor, setup, and uninstall to iterate AllAgentEnvs()

Site & Marketing

  • Built marketing landing page: dark theme, GSAP animations, terminal-styled code blocks
  • Published Hugo site to GitHub Pages with 5 example artifacts generated from real ledger data
  • Regenerated examples twice as ledger grew (55 → 69 entries)
  • Fixed Hugo baseURL after org migration; chained devblog → pages deploy

Releases & CI

  • Tagged v0.5.0 (multi-event hooks), v0.6.0 (storage pivot), v0.7.0 (MCP server), v0.8.0 (notes field), v0.9.0 (coaching rewrite), v0.10.0 (color flag + auto-commit + content safety)
  • Switched devblog CI from weekly to daily
  • Fixed stale git-notes fetch in CI

Highlights

  • Storage pivot was the highest-risk change — replacing the entire persistence layer across 26 files while maintaining backward compatibility. The file-per-entry approach proved its merit immediately: merge conflicts disappeared, and the simpler GitOps interface (4 methods vs 8) made the MCP server implementation straightforward.

  • The coaching rewrite validated a council decision: three perspectives independently concluded that quality problems were clarity problems, not model-specific problems. Adding motivation to rules (explaining why each exists) and concrete trigger checklists produced measurably better agent output — entries after the rewrite contain genuine design trade-offs instead of feature descriptions.