v0.0.5: Workstream Reconcile

Workstream Reconcile turns historical handoffs into a current operating queue by checking their volatile claims against authoritative sources.

• Adds workstream-reconcile, a read-only orchestration skill for GitHub, local Git, release, deployment, and registry reconciliation
• Separates historical delta from current operating state, including unpushed work, review responses, integration, release, verification, and proven stall states
• Adds behavior and trigger evals for local-only fixes, human feedback, release lag, false stalls, and degraded GitHub access
• Registers the existing experimental work-intake skill and restores repository validation
• Fixes invalid xref YAML frontmatter and adds validator regression coverage

All 16 skills are discoverable by the production installer. PR and merged-main validation passed.

v0.0.4: Issue Intake

The contributor-side selection gate is now issue-intake.

• Renames the canonical backlog -> qualify -> human choice -> Issue Contract seed skill
• Uses xref as optional graph evidence, not selection authority
• Keeps pick-an-issue as a self-contained deprecated alias for the v0.0.4 migration window
• Updates active docs, workflow, installer metadata, maturity registry, and evals
• Records a focused A/B: baseline 3/4, revised 4/4

issue-intake remains dogfooded, not validated.

v0.0.3: Candidate audits and compact decision visuals

Solution Gate now audits both greenfield fixes and existing contributor work without letting the candidate define the contract. Its visual step also uses the smallest auditable representation instead of requiring HTML for every run.

Solution Gate

• Candidate-audit mode seals an existing PR or patch, reconstructs the contract and solution shapes blind, probes discriminating cases, then reveals the candidate as a third proposal.
• Compact visual routing uses trees, pseudocode, types, diffs, Mermaid sequence or state diagrams, and reserves HTML for dense multi-view artifacts.
• Observed evidence and proposed shapes remain separate. Observations cite probe IDs; proposal links retain observed, inferred, and guessed marks.
• A focused A/B changed the representative output from a self-contained HTML page to inline Mermaid plus signatures without losing provenance.

The visual routing adapts the smallest-view principle from HumanLayer's show-me skill without adding an external runtime dependency.

Also included

• Primitive-contract mismatch as failure shape S12, harvested from agent-browser #1669.
• Additional public cases and gate-run evidence.
• Markdown link validation fixes for code spans and fenced code.

Install

bunx skills add Railly/skills --skill solution-gate --full-depth

Solution Gate remains Experimental and dogfooded. It has not completed a baseline comparison.

v0.0.2: Nine skills, a maturity ladder, and 56 cases

v0.0.1 shipped Unfold: one protocol with four modes. This release replaces it with nine skills that each own a trigger boundary, a ladder that separates distribution from evidence, and 56 recorded cases behind them.

New skills

review-gate runs a pre-review pass on a diff before pushing: deterministic checks first, then review lenses selected by what the diff changes. It is the only skill here compared against a baseline, in a blind replication round that exposed verification losses alongside its catches.

record-a-case captures maintenance work as an evidence ledger, keeping validation, human review, maintainer acceptance, and delivery independent. A merged PR is delivery, not proof.

solution-gate gates the shape of a fix before it is written. Two proposers on different model families work from the same symptom-free brief and declare falsifiable predictions; the cheapest refuting probe runs before any synthesis. It exists because the fix commit is the least-reviewed commit on any branch.

handoff closes an open working cycle so the next session resumes without rereading the transcript. It verifies state against git before asserting it, and keeps delivery, verification, and human judgment independent.

Plus trail-decisions and signature-repro as candidates, and quality-baseline, performance-proof, test-strength, and resilience-audit as experimental.

Distribution is not validation

foundry/maturity.json is the source of truth, and the two axes are deliberately separate. A skill can be installable and unproven at the same time, and the README says which is which.

Only review-gate is evaluated. pick-an-issue, record-a-case, and solution-gate are dogfooded. Everything else is experimental, which means the trigger and method are coherent and nothing more.

The evidence

56 public-safe cases across agent-browser, portless, wterm, native, radius, petdex, ai-cli, and this repository, up from 3. Four eval rounds, including a lifecycle audit that reassigned every skill's channel. Real work becomes a case before it becomes an instruction.

Install

bunx skills add Railly/skills

The installer exposes Stable, Candidates, and Experimental as distinct groups. Install only the stable surface with:

bunx skills add Railly/skills --skill pick-an-issue --skill record-a-case --skill review-gate

Also in this release

Unfold is deprecated, retained under foundry/deprecated for provenance. The validate job, red on main since 2026-07-30, is green again: an unregistered installer group, a maturity decision path that resolved one directory too deep, run data shipped inside a skill, and a SKILL.md over the size limit.

Still dogfooded, not validated. Controlled comparisons remain future work.

v0.0.1: Evidence-driven maintenance

First public release of Railly Skills.

Installable skills

• unfold: one evidence-backed mission across Learn, Triage, Change, and Review
• pick-an-issue: shortlist matrix, recommendation, and explicit human selection

Foundry

• public-safe maintenance cases
• maturity and promotion governance
• trigger and behavior evals
• five executable cross-stack fixtures

guided-contribution, repro-an-issue, and prove-the-test were absorbed into Unfold modes and preserved in Git history and foundry provenance.

This release is dogfooded, not validated. Controlled baseline comparisons remain future work.

bunx skills add Railly/skills