Skip to content

CHANGELOG.md

Latest commit

 

History

History
84 lines (68 loc) · 8.9 KB

CHANGELOG.md

File metadata and controls

84 lines (68 loc) · 8.9 KB
Raw

Changelog

All notable changes to weft are documented here. Format follows Keep a Changelog, and the project adheres to Semantic Versioning.

Unreleased

Added

  • weft doctor (/wf-doctor) — template currency. Reports whether a working copy (project-local or user-global) has drifted from the plugin-bundled canonical it was derived from — the one staleness git can't see, since the copies live in different trees. Compares by content hash (key-order/whitespace independent), so only real semantic drift is flagged. Also reports whether the plugin's own git clone is behind upstream (no network; HEAD vs the already-fetched @{u}). Read-only, no persisted state. --json for machine output.
  • weft analyze (/wf-analyze) — per-template run insight. Reads the event log into a per-template rollup: step durations (median/p90 from paired wf.step_changed deltas), loop counts, guard-block counts, abandonment + stall counts, slowest step, plus a recurring-friction section joined from workflow-qa.jsonl (by workflow_id). Turns the event log from a record into a feedback loop.
  • wf.guard_blocked events. The guard engine appends a wf.guard_blocked event (workflow_id, step_id, step_name, pattern, command, tool) when it blocks a tool call — previously blocks were silent, so "does this guard ever fire?" was unanswerable from the log. Matching logic unchanged.
  • wf.stalled annotation + real session-id capture. SessionStart writes the live session id to a per-project <weft_dir>/.session-id (read by cli.py as a fallback), fixing events logged with session_id="unknown"; a running workflow idle longer than WEFT_STALL_HOURS (default 12) gets a pure-annotation wf.stalled event (state untouched).
  • Context surfacing in context.md. generate_context_md renders the current step's description, a computed Next action: line (gated on running/waiting), and two optional per-step fields — suggest (commands that fit the step) and insights (💡 hard-won findings) — so the right command and the why-it-bites surface when a step goes active and survive compaction.
  • wf.step_changed paired start+end events. _advance_to_next now emits a wf.step_changed event with to_status="running" when each step begins. Auto-skipped optional+requires_skill steps now also emit a wf.step_changed with to_status="skipped". step_loop_back emits the target step's wf.step_changed (pending->running) alongside the existing wf.loop_iteration event. Dashboards reading events.jsonl can now compute true per-step duration (paired transition deltas) and per-iteration loop timing. Test updated: feature workflow expects 2 * step_count - 1 wf.step_changed events.
  • wf.failed terminal event. When a step's on_fail=block policy lands the workflow in failed state, step_fail emits a wf.failed event with workflow_id, blocked_at_step, reason, policy. Previously, blocked workflows had no terminal marker so downstream aggregators couldn't compute duration or distinguish "running" from "blocked-and-abandoned".
  • scripts/wf.py — interactive TUI over weft workflows. Curses-based: list view (running workflows + PR/CI status), detail view (per-workflow step + recent events), templates view (discovery). Falls back to the static wf-monitor.py table when stdout is not a TTY or invoked with --list / --json / --templates. Reads per-project .claude/weft/state.json + workflow-context.json and per-ticket subdirectories. Read-only.
  • scripts/wf-monitor.py — read-only viewer producing a plain-text aligned table over workflows + GitHub PR status. Standalone for scripting (wf-monitor.py --json | jq) or as the data layer behind the TUI.
  • pyproject.toml with weft console script entry point.
  • CONTRIBUTING.md and GitHub issue templates.
  • CI workflow running pytest on push and pull requests.
  • User-global templates tier — ~/.weft/templates/*.json (override with WEFT_USER_TEMPLATES_DIR). Discovery order: project-local > user-global > plugin-bundled.
  • README sections: "Custom templates" (three tiers), "Why not just use TODOs?" (positioning), "Requirements".
  • README GIFs (5 total): pitch (agentic chain — preview + start), walkthrough (first-run lifecycle), compose (heredoc → save-template → preview → start), extend (author SKILL.md + reference it from a template), audit (query + raw jsonl + rebuild from events). All recorded with asciinema → agg from the real weft CLI; reproducible scripts in docs/demos/.
  • README sections: "Compose your own workflow", "Extend with your own skills", "Auditable by design".
  • .claude-plugin/marketplace.json so /plugin marketplace add dioptx/weft works.
  • docs/demo.md — annotated end-to-end walkthrough captured from a real workflow run.
  • Makefile with install, test, lint, clean targets for one-command contributor onboarding.

Fixed

  • insights step field is now actually populated. generate_context_md rendered per-step insights (💡 lines) and the field was documented, but start_workflow/rebuild_from_events never copied it from the template into step state — so it silently rendered nothing. Both build sites now carry insights, matching the existing suggest-style optional fields. (suggest has the same latent gap; wire it the same way if a template ever uses it.)

Changed

  • Date-based workflow_id now carries second resolution (name-YYYYMMDD-HHMMSS, was name-YYYYMMDD). Two same-day runs of one template in one project no longer share an id, so /wf-analyze counts them as distinct workflows instead of conflating their step timings. PR-scoped ids (name-pr<N>) are unchanged — they stay stable by design. /wf-analyze already groups by the name field in wf.started, not by parsing the id, so the format change is backward-compatible for analytics.
  • README: corrected license, version, install instructions (was a fictional claude plugin install syntax — now uses real /plugin marketplace add flow), skill count (10→11, added wf-compose), and feature-workflow step count (12→11).
  • .claude-plugin/plugin.json bumped to 0.3.0 to match shipped feature set.

Removed

  • templates/publish-repo.json — was specific to the maintainer's portfolio publishing setup (hardcoded paths, repo names, deploy targets). Deployment-specific templates belong in user-global or project-local tiers, not bundled with the plugin.

0.3.0 — 2026-04-10

Added

  • Looping constructs in template schema and state machine: steps can declare loop_back_to, max_iterations, exit_condition.
  • New transitions loop-continue and loop-done with wf.loop_iteration events handled by rebuild_from_events.
  • Projections now surface loop count, skill directives, and exit conditions.
  • New /wf-compose skill — reads conversation context, scans the skill registry, and proposes v2 templates with loops + skill blocks (ASCII diagram preview before approval).
  • feature-workflow.json upgraded to schema v2: manual review-cycle replaced with an enforced 3-step loop (review → apply-fixes → run-tests).
  • 12 new tests covering loop lifecycle, rebuild, and behavioral mandate 10. Suite at 187 tests.

Changed

  • Renamed default feature template to feature-workflow (previously namespaced to a private workflow).

0.2.0 — 2026-04-09

Added

  • publish-repo workflow template — 10-step pipeline ending in a dry-run gate.
  • Two new template-management skills: /wf-new-template, /wf-edit-template.
  • MIT LICENSE file.
  • pyrightconfig.json for editor type-resolution.

Changed

  • Default feature template renamed to remove a private namespace prefix (now feature-workflow).
  • Plugin cache refreshed via version bump for new skills to surface.

Fixed

  • /wf-rebuild now preserves step definitions and current_step.
  • /wf-rebuild infers current_step correctly for completed and aborted workflows.

0.1.0 — 2026-04-08

Added

  • First public cut of weft: deterministic workflow tracking for Claude Code.
  • Event-sourced state machine with append-only JSON log under .claude/weft/.
  • Built-in templates: generic (3 steps) and feature-workflow (12 steps).
  • Eight smart skills: /wf-start, /wf-step, /wf-status, /wf-abort, /wf-preview, /wf-rebuild, /wf-dashboard, /ev-query.
  • Guard engine enforcing per-step allowed-tools and blocked-commands via PreToolUse.
  • Stop gate refusing session exit while steps are incomplete.
  • Compaction-safe PreCompact hook writing a context.md projection.
  • Live curses-based dashboard.
  • pytest suite covering event store, state machine, hooks, CLI, and behaviors.