changelog

Changelog.

What shipped, what's next, and what to do about it. Each skill's version field is the source of truth — this page rolls them up by release. Vote on what's coming to signal priority.

What shipped

Latest release is open; patches and earlier releases expand in place.
shippedJul 10, 202627 → 29 skills

Every PR now runs the documentation mesh in order: oc-docs-forge writes the documentation packet, oc-repo-ops verifies it and the repository around it, oc-bug-check runs the fast code gate, and only then does oc-git-ops open the PR. The skill that writes the docs never decides whether they're good enough.

Two new skills

oc-docs-forge · /oc-docs

Generates the PR documentation packet, PR-body documentation section, optional PR comment, README deltas, and product-doc updates from the actual change set.

oc-repo-ops · /oc-repo

Runs the repo-readiness gate: generated files, docs catalogs, link hygiene, ignore-rule consistency, dirty related files, and release-surface cleanup before every PR.

Every-PR contract

  • PR body and comments: every PR body carries a ## Documentation section — "no docs needed" requires evidence — plus, when the packet runs long, a durable PR comment under a stable marker.
  • README and product docs: the repo-facing and product-facing docs that changed behavior makes stale are updated from the actual diff, not from memory.
  • Repo readiness: generated files, docs catalogs, release surfaces, ignore rules, and dirty related files are verified before PR handoff — fail closed.
Back-compatible with v1.7. All 29 skills lockstep-bump to 1.8.0; the two new skills ship gated behind their skills.registry.<id>.enabled flags, and no documented cross-skill contract was removed — the every-PR gate is additive wiring in oc-git-ops.
checkpoint readsno repo-local CLIcorroborated status29 skills

The checkpoint protocol was leaning on an optional in-repo scaffolder that installed skill bundles don't ship. This patch makes a direct read of .checkpoints/*.checkpoint.json the authoritative path, so "where did I leave off?" answers the same in a consumer repo as it does here — and insists that answer be checked against reality, not taken on faith.

What shipped

  • oc-orchestrator treats a direct checkpoint-file read as the authoritative fallback and no longer recommends a repo-local scaffolder where it cannot exist.
  • oc-checkpoint-protocol requires agents to corroborate the checkpoint state they read against specs, git, tests, and release artifacts; missing CLI output is neither a blocker nor evidence of progress.
  • Lockstep patch bump — all 29 skills move to 1.8.1.
Back-compatible with v1.8.0. No routes, storage, environment variables, or user-facing commands changed, and existing checkpoint JSON requires no migration.
shippedJul 01, 202624 → 27 skills

oc-signal-forge ships first as 1.7; the oc-modularize-ops → oc-fleet-ops chain follows in 1.7.1. All three lockstep-bump the catalog to 1.7.0.

Three new skills

oc-signal-forge · /oc-signal 1.7

Question → trustworthy metric. Designs the instrumentation, builds the harvester + transform, and adversarially proves the signal answers the question before wiring it to oc-dash-forge. The product-analytics backend none of the instrumentation skills owned.

oc-modularize-ops · /oc-modularize 1.7.1

Decompose a live monolith with provably zero functionality/data loss — golden fixtures captured from real traffic as the equivalence oracle. Willing to say "don't." Hands the code-move + cutover to oc-migration-ops and deploy to oc-fleet-ops.

oc-fleet-ops · /oc-fleet 1.7.1

Provision, deploy, and operate one-or-more containers across self-managed infra (k8s/Nomad/Compose, IaC, on-prem VMs, GCE). Terraform when it fits, not by reflex. Claims the bare-metal/VPS territory oc-deploy-ops routes away — its peer, not its rival.

Resolves existing seams

  • oc-signal-forge finally executes oc-app-architect's 08-analytics.md plan and feeds validated metrics to oc-dash-forge; oc-monitoring-ops enforces each signal's freshness SLA.
  • oc-modularize-ops owns the decision plus the real-data equivalence oracle, then hands the bulk code-move and live cutover to oc-migration-ops (Structural type).
  • oc-fleet-ops claims the bare-metal / self-managed-fleet territory oc-deploy-ops routes away — the two are peers (managed app → deploy-ops; self-managed fleet → fleet-ops).
Shipped from the 1.7 "Seams & Signals" plan (docs/releases/1.7-plan.md).
Back-compatible with v1.6. All 27 skills lockstep-bump to 1.7.0; the three new skills ship gated behind their skills.registry.<id>.enabled flags, and no documented cross-skill contract was removed.
installed skillscheckpoint writesprotocol bundlescatalog guards

This patch fixes the gap between the repo-local experience and a user's installed skill bundle. Checkpoint writes no longer depend on a repo-only CLI, every bundle carries the shared orchestrator and checkpoint references it asks the model to read, and the catalog generator now blocks missing bootstraps, dangling reference citations, and overlong descriptions before a release can ship.

What shipped

  • Checkpoint writes are self-contained. The protocol now allows direct `.checkpoints/*.checkpoint.json` writes with file tools, with `scripts/checkpoint.mjs` kept as an optional in-repo fast path.
  • Shared protocol files are bundled everywhere. Missing `references/orchestrator.md` and `references/checkpoint-protocol.md` copies are created and checked rather than silently skipped.
  • Catalog validation catches portability drift. Missing bootstrap lines, dangling `references/...` citations, and overlong descriptions fail the build.
  • The audit trail is durable. The verified portability audit lives at docs/audits/2026-07-04-portability-audit.md.
Back-compatible with v1.7. No routes, storage, environment variables, or user-facing commands changed. Existing checkpoint JSON remains valid.
checkpoint auditwrite triggersdrift-aware nextCLI commands

The audit found that checkpoint infrastructure exists, but it detected more drift than it prevented. This patch closes the first reliability gaps instead of leaving them as a plan: checkpoint:next now checks completed/merged work before recommending an action, and the shared protocol now names the lifecycle events where checkpoint writes are mandatory.

What shipped

  • Lifecycle writes are mandatory. The protocol now requires writes on session resume/pause, phase or gate completion, user decisions, blockers, cross-skill handoffs, PR open/merge, deploys, release ship, conflict resolution, and doctor drift.
  • checkpoint:next is drift-aware. It skips queued actions that reference already-merged PRs or completed tickets; if every queued action is stale, it recommends checkpoint reconciliation instead.
  • The missing command surface is real. checkpoint list, checkpoint show [skill], and checkpoint reset <skill> now exist in the shared CLI.
  • Bug-check filename mismatch is resolved. The tracked wrong-name bug-check.checkpoint.json is removed; the canonical local-only path remains oc-bug-check.checkpoint.json.
  • Stale resume state is reconciled. oc-app-architect no longer points the next agent at merged PR #351 or the completed v1.6 launch queue.
  • Obsolete writer references are cleaned up. oc-reverse-spec now points to scripts/checkpoint.mjs, not the old scripts/checkpoint.sh.

Still queued

  • oc-orchestrator: trim stale oversized checkpoint state and deepen the v1.7 routing map.
  • oc-signal-forge, oc-modularize-ops, and oc-fleet-ops: add richer resume/action examples and context-primer guidance.
  • oc-code-auditor, oc-stack-forge, and oc-ux-engineer: shorten oversized summaries into structured checkpoint fields.
  • oc-telemetry-ops: add missing progress_table and context_primer.
  • oc-release-ops: add a release-complete guard requiring relevant skill checkpoints to match shipped state.
Back-compatible. Existing checkpoint JSON keeps validating; the patch adds stricter recommendation behavior and new utility commands without changing the wire schema.

Two new skills

oc-cost-ops · /oc-cost

Attributes Anthropic spend to each skill phase, sets per-phase and per-suite budget gates in the checkpoint, and recommends model-tier routing (Haiku for cheap phases, Opus for spec/audit/migration).

oc-telemetry-ops · /oc-telemetry

Opt-in, local-first usage metering to .checkpoints/usage.sqlite — anonymized aggregates only, default OFF. No prompt content or PII ever leaves the machine.

protocol wire 1.1

First bump since v1.2's pm_refs: additive optional cost, eval_scores, and telemetry_handle fields. Wire 1.0 and 1.1 both validate — fully backward compatible.

public /dashboard + /showcase

Anonymized aggregate usage and cost-per-feature, fed by opt-in telemetry, plus a /showcase of projects shipped with opchain.

Cross-skill ripples

  • oc-prompt-ops — the cost_per_eval contract it already advertised is now fulfilled by oc-cost-ops, pairing the cost-regression gate with the score gate.
  • oc-bug-check and oc-code-auditor now emit eval scores against a stable rubric, not just pass/fail.
  • New GTM surfaces: /dashboard and /showcase.
Back-compatible with v1.5. All 24 skills lockstep-bump to 1.6.0; the two new skills ship gated behind their skills.registry.<id>.enabled flags, and the checkpoint wire 1.0 → 1.1 is additive — both validate.

Four new skills

oc-claude-api · /oc-claude-api

Model routing (Opus/Sonnet/Haiku/Fable), prompt caching by default, streaming vs. batch, tool use, and migration playbooks for model-version bumps.

oc-rag-forge · /oc-rag

Tri-agent. Vector-DB decision tree, embedding choice, chunking, hybrid search, and retrieval eval with golden sets and recall@k / nDCG.

oc-agent-forge · /oc-agent

Tri-agent. Claude Agent SDK scaffold, subagent topology, tool-budget design, harness-loop shapes, and eval over task fixtures.

oc-prompt-ops · /oc-prompt

Prompts as artifacts: file-shaped versioning, eval datasets, regression detection. /oc-prompt regress blocks on a score drop.

Cross-skill ripples

  • oc-app-architect gained an “AI app?” branch in /oc-discover that routes to the four new skills.
  • oc-stack-forge added vector-DB packs (pgvector, Turbopuffer, Pinecone, Supabase Vectors) to the pack registry.
  • oc-code-auditor learned a prompt-injection + tool-use AI-safety sweep; opchain published its own dogfooding eval set under oc-prompt-ops.
  • New GTM surfaces: /ai-recipes, /blog, /for-agencies.
Back-compatible with v1.4. All 22 skills version-bump to 1.5.0; the four new skills ship gated behind their skills.registry.<id>.enabled flags.

What's new

MCP server

A POST /mcp endpoint exposes the catalog, intent routing, the orchestrator protocol, and checkpoints as MCP tools, prompts, and resources — for Codex and any MCP client.

two ways in

Point Codex at the hosted opchain.dev/mcp endpoint, or drop the skills into .codex/skills/ — Codex Agent Skills use the same SKILL.md format.

/oc-* as prompts

Every /oc-* command is re-exposed as an MCP prompt for clients that surface them (e.g. Claude Desktop). Codex doesn't consume MCP prompts — there you trigger skills by description, or use the .codex/skills drop-in.

checkpoints anywhere

Session state persists across MCP sessions, so you resume where you left off — the same contract Claude Code uses.

New on /install

  • Codex / any MCP agent — a fourth install flow with the ~/.codex/config.toml snippet and the .codex/skills drop-in.
  • local stdio server — an offline mcp/local-server.mjs for air-gapped setups.
Back-compatible with v1.4. The skills are unchanged — only their packaging — so Claude Code installs are untouched. All 18 skills lockstep-bump to 1.4.3.
earlier releases
complete bundlesoc- nameshonest checkpointsdrift detection

The “download all” bundle was dropping each skill's references/, scripts/, and examples/; it now ships full skill trees again. Every skill also gained an oc-<name> alias, the persistence docs were rewritten to the git-tracked checkpoint they actually use, and the validator, spec, and README were reconciled onto one checkpoint schema with drift detection. All 18 skills lockstep-bumped to 1.4.2.

pack registry32 packs5 skills bumped

v1.4 turns oc-stack-forge's coverage from hardcoded prose into a versioned pack registry — one pack.yml per language, framework, mobile platform, or deploy target, validated on every build. /coverage is the user-visible catalog; total coverage grows from 6 stacks in v1.3 to 32 packs.

What's new

  • Pack registry architecture. scripts/gen-stack-packs.mjs validates every pack on every build.
  • Hybrid dispatch. oc-deploy-ops reads packs at runtime; oc-api-dev reads them at build time for typed scaffold metadata.
  • Modern web + enterprise + mobile bulk. elixir / bun / deno / java / csharp / kotlin / php / swift packs; phoenix / remix / sveltekit / solid / spring / dotnet / laravel frameworks; iOS / Android / Flutter / React-Native mobile packs.
  • Hosting adapters. railway / netlify / heroku / aws-amplify deploy-target packs.
Back-compatible with v1.3. Five skills bump in lockstep to 1.4.0.
oc-release-ops18th skillPM-MCP live

v1.3 makes the v1.2 PM-MCP prose executable end-to-end and ships oc-release-ops — the 18th skill — to automate the scope → /changelog → bump → ship cadence opchain uses for itself.

What's new

  • The PM-MCP loop is real. Concrete tool-name registry, retry/backoff, idempotency markers, deferred-action queue.
  • oc-release-ops is the 18th skill. /oc-release plan|draft|bump|announce|ship: reads checkpoints to propose the next semver, drafts the changelog entry, bumps all skill versions atomically.
  • Platform menu grew. Django + Postgres + Render, Rails + Postgres + Heroku, Go + Fly.io, Rust + Axum + Shuttle.rs.
Back-compatible with v1.2. All 18 skills lockstep at 1.3.0.
PM-MCPLinearJira

The ticket becomes the thread of execution. oc-app-architect, oc-git-ops, oc-deploy-ops, and oc-monitoring-ops all become PM-tool aware.

Back-compatible with v1.1. The PM loop is enrichment, never a hard dependency.

Tri-agent harness across oc-app-architect, oc-code-auditor, oc-integrations-engineer; full-pipeline flow from /oc-discover to /oc-deploy; oc-checkpoint-protocol v1.0 across every skill; static Astro site; first set of six in-action scenarios.

First public cut of the opchain skill catalog as a set of Markdown files installable into Claude Code. No site, no scenarios — just the skills.