v0.3.0 — 2026-06-24

Release Notes

Adds a renderable-fenced-block pipeline (d2/csv compile-checked in lint and rendered in
PDF), a per-repo change-set validation engine, consolidates shared Python helpers into a
_shared/ package, and extends yf-beads-init repair with native runtime/hook/remote
cleanup.

Renderable fenced blocks: d2 + csv (plan-017, #33 + #34 + #37)

• _shared/renderable_fences.py — new canonical registry of renderable fenced-block
  kinds (d2, csv) shared across the lint, PDF, and diagram skills.
• yf-markdown-lint — new ML009 rule: d2 fenced blocks are compile-checked (caught
  malformed diagram source at authoring time, not render time).
• yf-markdown-pdf — scripts/blocks.lua renders fenced blocks during the PDF pipeline
  (d2 → embedded diagram, csv → table) with a glyph-fallback.tex for broad Unicode
  coverage; --no-render-fences opts out. _shared/sync.py gains a Python→Lua emitter
  that keeps blocks.lua mirrored from the canonical registry (DRIFT-CHECK edges enforce it).
• yf-diagram-authoring — embed/lift/inline round-trip between a .d2 source file and an
  inline fenced block.
• yf-research — new record-epic subcommand and the CLI-spec realignment that goes with it.

Helper consolidation + yf-beads-init native cleanup (plan-016, #15 + #39 + #36)

• _shared/ — two-helper vendoring sweep: manifest_update.py (whole-file, byte-identical
  copies across five skills) and json_extract.py (defensive JSON extractor, region-vendored
  into plan_manager.py + research_manager.py; json-get gains list int-indexing). New
  DRIFT-CHECK edges and test_sync.py coverage. Zero yf Rust change for this epic.
• yf-beads-init — repair() native cleanup: untrack-runtime, content-guarded
  remove-hook-shims, and a confirm-gated remove-remote (--remove-remote), wired into
  preflight-converge (the read-only invariant is preserved). Rust tests + SKILL/SPEC docs.
• audit — regression test pinning the audit --json-output json.dumps invariant (bug
  already absent; no production change).

New: yf-change-validation — per-repo change-set validation engine (plan-015, #27 + #25)

• yf-change-validation — new skill that validates a change set against a per-repo,
  operator-approved CHANGE-VALIDATION.md manifest: declared check tiers (fast/full),
  a changed-path → check-id trigger scope, and a signal fingerprint that detects when the
  manifest itself has gone stale relative to the repo's real test/build surface.
  • scripts/change_validation.py — the engine: infer (bootstrap a manifest by scanning
    CI workflows, Cargo.toml, test_*.py, build configs), run (execute the tier or
    the trigger-scoped affected subset for a changed path; fail-closed), and check-drift
    (re-propose the manifest
    when the signal fingerprint diverges). Self-maintaining: a stale manifest re-proposes
    itself rather than silently passing.
  • spec/{schema,engine,inference}.md + SPEC.md — the manifest schema, the
    validate/trigger-scope engine, and the inference/bootstrap heuristics.
  • templates/manifest.md — the seed CHANGE-VALIDATION.md template.
  • protocols/CHANGE-VALIDATION-TRIGGER.md — always-loaded companion rule binding the
    on-change trigger a description cannot fire (silent no-op unless the repo has an
    approved CHANGE-VALIDATION.md).
• yf-plan — §6.1.5 layer (b) now delegates change-set validation to
  yf-change-validation when an approved manifest exists; the static validate-cmd
  becomes the thin middle fallback and the not-checked notice the floor (3-tier
  precedence, output schema + exit-3 contract preserved).
• CHANGE-VALIDATION.md — dogfood manifest for this repo (approved): fast/full tiers
  wiring the cargo + per-skill uv/pytest suites and _shared/sync.py --check, with the
  website build rows trimmed (deploy-only, not a validation gate) and yf-drift-check
  excluded (prose/LLM trigger, not a runnable command).
• Docs (#25): added worktree-uv guidance to yf-plan's worktree address-space docs —
  prefix uv run … with env -u VIRTUAL_ENV inside a plan worktree so uv resolves the
  worktree's own environment, and do not follow uv's --active suggestion (it targets the
  primary venv, the wrong address space).

_shared/ package — retire the duplicated active-set classifier (plan-014, #15)

• _shared/ — new top-level package (outside the skills/ embed root, so yf never
  treats it as a skill) holding canonical Python helpers shared across skills. Ships zero
  yf Rust changes: the vendored copies live inside each consuming script and install copies
  them verbatim.
  • _shared/active_set.py — the single canonical active-set classifier (classify_active +
    Edge, ActiveSetReport, helpers, and the ACTIVE_*/PARENT_CHILD/OPEN/IN_PROGRESS/
    CLOSED_STATUSES/GATE_TYPE constants), delimited by BEGIN/END region markers.
  • _shared/sync.py — repo-time sync tool that regenerates the marker-fenced classifier
    region in-place in each consumer from canonical (no sibling import, no new file — each
    script stays self-contained); --check mode exits non-zero on divergence (CI/manual backstop).
  • _shared/README.md — documents the regenerate-the-fenced-region (vendoring shape (b)) pattern
    and the enforcement point.
• yf-beads-hygiene / yf-beads-upstream — the active-set classifier is now a generated,
  marker-fenced region in both beads_hygiene.py and upstream.py, regenerated from
  _shared/active_set.py. The plan-013 "do NOT edit one without the other" hand-maintenance
  banner is gone; both test suites stay green and the single-file test loaders are untouched.
• DRIFT-CHECK.md — migrated the classifier edges: classifier-canonical re-pointed to
  _shared/active_set.py (fixed authority); the old pairwise e-classifier-copy replaced by two
  derived value-equal region edges (e-active-set-copy-hygiene, e-active-set-copy-upstream).
  A tampered region now FAILs the copy, never the canonical.

Reconcile policy: local beads = active work only (plan-013, #38 + #17)

• yf-beads-upstream — implements custom.upstream.granularity (coarse|granular,
  default coarse; fills the previously-unimplemented REQ-BUP-043) and a default-deny
  custom.upstream.auto_hoist_followons knob. Adds a hoist operation (per-granularity
  create-or-map via External:, dry-run-first, reversible bd close -r — never bd delete),
  narrow-vs-broad follow-on detection, an un-hoist/restore path, and a land-the-plane flow
  that proposes follow-on hoists with a single confirm by default (no-prompt only under
  the opt-in key, narrow signal only). enumerate now uses the shared active-set definition.
  Companion rule UPSTREAM_TRACKING.md 1.0.2 → 1.1.0.
• yf-beads-hygiene — new read-only-first reconcile pass: classifies the active set
  (in_progress / claimed-open / open ancestors) and lists non-active local beads as hoist
  candidates plus obsolete upstream issues (mechanical delivered signal; flag-for-review
  fallback, never auto-close). Gated --apply/--yes/--record delegates the hoist to
  yf-beads-upstream (the carve: hygiene proposes, upstream executes).
• DRIFT-CHECK.md — new e-classifier-copy edge asserting the active-set classifier
  (authored once in yf-beads-hygiene, copied verbatim into yf-beads-upstream for install
  independence) stays identical across both.

yf-beads-init — prune empty .codex/config.toml residual (dqo)

• Repair now removes the bare [features] residual bd setup codex --remove leaves behind,
  deleting .codex/config.toml only when effectively empty (then pruning an empty .codex/).

Install yf 0.3.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/dixson3/yoshiko-flow/releases/download/v0.3.0/yf-installer.sh | sh

Install prebuilt binaries via Homebrew

brew install dixson3/tap/yf

Download yf 0.3.0

File	Platform	Checksum
yf-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
yf-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
yf-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
yf-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum