16 July 2026
Four releases that make the maturity dial a token dial. Frontier models are priced for judgment, not boilerplate — this arc rebuilds how ADD spends context: rules now load physically per maturity level at session start (a POC project drops ~70% of rule tokens), every sub-agent dispatch carries a MODEL tier (fast / editor / architect) and a maturity-scaled token BUDGET, and the heaviest skills lost 35–45% of their lines to on-demand templates.
v0.9.8 fixed cross-runtime correctness: the Codex compiler now rewrites paths and command namespaces in every emitted tree, and both runtimes share one autoload predicate so they can never disagree on always-loaded invariants. v0.9.9 shipped the SessionStart rule loader, the dispatch cost policy, and size-capped learning views. v0.9.10 slimmed skills, de-noised the injection patterns (closing a real false negative on the canonical injection phrase), and hardened CI.
Measured on the shipped v0.9.11 build (vs the v0.9.7 baseline of ~19.4k tokens loaded every session at every maturity): a POC project’s session baseline is now ~5.0k tokens (−74%), Alpha ~8.7k (−55%), Beta ~13.8k (−29%), and even GA pays ~14.0k (−28%) — the on-demand moves save tokens at every level, and the maturity dial scales the rest. The five heaviest skills invoke 30–45% lighter (init 1,093→638 lines, verify 761→432), and the learnings view read before every skill is hard-capped so it can’t creep back. The numbers are CI-guarded: the autoload set is derived from rule frontmatter, and contributors must state the context cost of any autoloaded-surface change.
v0.9.11 closed the loop on a sharp community catch by @tdmitruk: rules copied into projects by older /add:init runs never updated on plugin upgrade — and conflicted with the freshly injected ones. Rule copying is now retired entirely; the hook is the sole mechanism, stale copies trigger a per-session warning, and the migration chain offers a confirmed one-time cleanup.
Read the full release notes →
Release
Token economy
Community
30 April 2026
Doc-only release before the v1.0 credibility cycle starts. Removes drift between what the docs claim and what actually shipped through v0.9.4: ten spec frontmatters flipped Draft→Complete with Shipped-In recorded, eighteen plan files close-out-headed, specs/timeline-events.md superseded by the shipped telemetry + dashboard work.
PRD § 4 reframed to point at the new docs/milestones/v1.0-ga.md (canonical) and docs/v1.0-roadmap.md (synthesis of three parallel research swarms). The maturity cascade matrix now includes the four v0.9.0-shipped rules (cache-discipline, injection-defense, secrets-handling, telemetry). README rules count corrected (11→19); methodology-led intro seeded as the first step toward the methodology/runtime brand split.
The reworded v1.0 GA criteria drop the generic 60-day calendar floor and replace it with engineering evidence: at least three minor/patch releases past v0.9.0 with no post-tag rollback, at least one community PR merged during the beta cycle, and a release-evidence bundle generated for the v1.0 candidate. Criterion six rewords the unmeasurable “20+ projects” claim to maintainer attestation via private signal.
Architectural findings F-006 (host-neutral kernel) and F-007 (adapter contracts) deferred to v1.1.0; the v0.9.7 release will ship the Tier 1 substitution-only fix that closes ~80% of the leak in time for v1.0 GA. New rule core/rules/model-roles.md closes the M3 deferral with one paragraph distinguishing Architect (planning, spec authoring) from Editor (mechanical edits, frontmatter sweeps) work.
Read the full release notes →
Release
Hygiene
v1.0 prep
23 April 2026
Seven feature specs shipped in parallel by worktree-isolated agent swarms during a ten-hour /add:away session. 207 acceptance criteria, 93 tests, three signed releases back-to-back: v0.8.1 (hotfix), v0.9.0 (M3 ships, alpha→beta), v0.9.1 (beta polish, CI wired).
TDD guardrails that actually bite (the agent can’t delete a failing test to make green), secrets handling that runs (pre-commit gate, not prose), prompt-injection defense with scan hook and documented threat model, Codex-native skill emission with preserved frontmatter, OTel-aligned telemetry with cache-read/creation fields, stable-prefix cache discipline for sub-agent dispatch, and tool-portable AGENTS.md generation.
Maturity promoted alpha → beta against the v0.9 release. Readiness 92% against the cascade matrix; the single exemption (CI guardrail wiring) was time-boxed to v0.9.1 and is now closed. 93 tests run on every PR and push to main.
A plugin-family review from a separate agent surfaced the Codex install path mismatch that would have shipped broken Codex installs with v0.9.0 — caught in the v0.8.1 hotfix pivot, closed with a smoke test that installs into a temp CODEX_HOME and resolves every path.
Read the full post →
Release
M3
Beta
Multi-runtime
Security
22 April 2026
Learning files grow monotonically. A 6-month-old project’s learnings.json eats 14% of the context window before a skill starts its real work; at 1 year it hits Read-tool caps and entries silently drop. v0.8 moves the filtering out of Claude’s context and into a jq-based PostToolUse hook.
Agents read a compact learnings-active.md (top entries grouped by category + one-line index of the rest). The canonical JSON is never modified by the filter. A three-layer fallback — active view → run filter → full JSON — means no data is ever lost.
62–82% token reduction as projects age, measured with the real Claude tokenizer. Comes with a /add:learnings skill (migrate, archive, stats), an archived entry field, and configurable thresholds in .add/config.json. Migration runs automatically from any version ≥ 0.5.0.
Second consecutive community-driven release. Proposed, implemented, and iterated by @tdmitruk with two rounds of review feedback — the blocker and all four non-blocking nits addressed in a single follow-up push.
Read the full post →
Release
Learnings
Context
Community
12 April 2026
The biggest structural change since v0.1. ADD extracts from a Claude-specific plugin into a runtime-neutral core/ with per-runtime adapters. Codex CLI joins Claude Code as a first-class target, generated from the same source.
v0.7 was designed by dispatching 5 AI agents in parallel with non-overlapping review scopes (Anthropic spec, install UX, Codex portability, multi-runtime architecture, security). Each caught findings the others would have missed. The pattern is now Tier 1 knowledge in the plugin itself.
Security hardening moved from prose to CI: JSON Schema validation on frontmatter, rule-boundary diff checks, compile-drift verification, and GPG-signed release tags from v0.7.3. Full threat model in SECURITY.md.
Three community contributors shipped major additions in the same window: /add:docs (Caleb Dunn), /add:roadmap + /add:milestone + /add:promote (Piotr Pawluk), and /add:ux (David Giambarresi).
Read the full post →
Release
Multi-runtime
Codex
Security
February 2026
TDD gave us tests before code. BDD gave us behavior before tests. ADD gives us coordinated agent teams before everything.
Today we're launching ADD as a free, open-source Claude Code plugin. ADD is a structured SDLC methodology designed for the reality that AI agents do the development work while humans architect, decide, and verify. It brings the same discipline to AI-native development that TDD brought to testing two decades ago.
ADD ships with 9 commands, 9 skills, 11 rules, and 13 templates — all pure markdown and JSON, zero dependencies. Install it with one command and start building in 5 minutes.
What makes ADD different: Agents aren't solo contributors — they're team members. ADD treats them as a coordinated team with roles (test-writer, implementer, reviewer, deployer), scoped permissions, independent verification, and accumulated knowledge that compounds across projects.
The maturity dial (POC → Alpha → Beta → GA) governs all process rigor with a single control. Start lightweight for prototypes, dial up to exhaustive verification for production systems. The knowledge cascade ensures agents get smarter with every project.
Get started with ADD →
Launch