Skip to content
Casulo AI Labs AID AI Integrated Development
Releases

Changelog

AID v2.3.0

v2.3.0 ·

AID v2.3.0

Five per-profile tarballs + CLI bundle for the AID methodology installer.

Verify your download:

sha256sum -c SHA256SUMS

Offline bundles

AID v2.2.2

v2.2.2 ·

AID v2.2.2

Five per-profile tarballs + CLI bundle for the AID methodology installer.

Verify your download:

sha256sum -c SHA256SUMS

Offline bundles

AID v2.2.1

v2.2.1 ·

AID v2.2.1

Five per-profile tarballs + CLI bundle for the AID methodology installer.

Verify your download:

sha256sum -c SHA256SUMS

Offline bundles

AID v2.2.0

v2.2.0 ·

AID v2.2.0

Five per-profile tarballs + CLI bundle for the AID methodology installer.

Verify your download:

sha256sum -c SHA256SUMS

Offline bundles

AID v2.1.0

v2.1.0 ·

AID v2.1.0 — a feature release: a catalog-based connector registry with full lifecycle management, a verb-first "AID Lite" shortcut path for small changes, and a structural rewrite of dashboard state tracking. Builds on v2.0.6.

Highlights

  • Connector lifecycle + consumption. .aid/connectors/ is now a full catalog, not just a scan result: aid-discover's ELICIT state authors it, and two new on-demand skills — aid-set-connector / aid-unset-connector — let you add, update, or remove a connector at any time without re-running discovery. Pipeline skills now consume the catalog MCP-first: for a mcp-type connection they tell the agent to request it from your host tool's own MCP/plugin (Claude Code, Codex, Cursor, …) — AID resolves nothing and stores no credential for it. (#133, #143)
  • AID Lite: verb-first shortcuts. Already know your change? /aid-fix, /aid-create-api, /aid-change-cli, and dozens more — 51 verb-first shortcut skills (plus aid-add-* / aid-update-* aliases, 80 invocation names in all) spanning create, change, refactor, remove, deprecate, migrate, fix, test, prototype, document, report, review, and research — collapse Describe → Define → Specify → Plan → Detail into one fast, mostly-autonomous run that still produces the full graded artifact set, then halts for your approval. No interview required. /aid-triage routes a free-form description to the right entry point (a shortcut, /aid-describe, or /aid-ask) when you're not sure. (#134)
  • Structured STATE.md. Machine-parsed pipeline state (approval, grade, status, counts, paths) moves from free-form prose into a defined YAML-frontmatter schema, read deterministically by the dashboard — closing out a class of regex-parsing bugs, including the one where an approved Knowledge Base could render as still "Building" with a dead kb.html button. The dashboard's phase model was also rebuilt to match the real pipeline (Describe → Define → Specify → Plan → Detail → Execute), with the Lite path (Describe → Execute → Done) rendered distinctly. (#140)
  • Dashboard served from the CLI. home.html is now a data-free template served from the CLI's own install rather than a per-project committed copy, so every project's dashboard page is always current with the installed CLI version — no more drift, no manual refresh. (#142)

Improvements

  • Delivery-folder layout rationalized: the full path nests deliveries under deliveries/delivery-NNN/ (mirroring the existing features/feature-NNN/ pattern); the Lite path drops the extra folder and keeps the single delivery's lifecycle/gate directly in the work-root STATE.md. (#132)
  • kb.html (the Knowledge Base visual summary) is now a per-project generated artifact at .aid/knowledge/kb.html; the old .aid/dashboard/ folder is retired entirely. (#142)

Fixes

  • Security/correctness: a bounded-read guard (io_bounds.py, a 5 MB DoS-protection module used by the dashboard reader at 9 call sites) was missing from every install channel's manifest. A release built from that state would have shipped a dashboard reader that fails to import on npm, PyPI, and curl/bash installs — caught and fixed before publish. (#137)
  • Windows: aid dashboard stop now reliably kills the actual server process. Previously it could kill only the cmd/.bat wrapper, leaving an orphaned server bound to the port; a port-reap safety net now cleans up any older orphan still listening. This was the root cause of a dashboard that showed every pipeline as "0 tasks / 0 deliveries" after repeated start/stop cycles. (#136)
  • Docs site (Pages) deploy no longer runs a doomed job on tag-triggered GitHub Releases; it now deploys only from master pushes or a manual dispatch. (#131)
  • Stale skill/shortcut counts in docs/install.md corrected, with a new guard test to prevent recurrence. (#137)

Upgrade

  • Existing project: aid update refreshes all installed tools; aid update self updates the aid CLI itself — channel-aware: it re-runs the bootstrap script for curl/irm installs, or prints npm install -g aid-installer@latest / pipx upgrade aid-installer for those channels.
  • New install (bootstrap once per machine):
    curl -fsSL https://raw.githubusercontent.com/AndreVianna/aid-methodology/master/install.sh | bash

    irm https://raw.githubusercontent.com/AndreVianna/aid-methodology/master/install.ps1 | iex
    Then aid add <tool> inside your project.
  • aid-installer is also published to npm (npm i -g aid-installer) and PyPI (pipx install aid-installer). Per-profile tarballs and SHA256SUMS are attached to this GitHub Release for offline/air-gapped installs.

Offline bundles

AID v2.0.6

v2.0.6 ·

AID v2.0.6 — a housekeeping patch across the discovery/summarize toolchain: a config relocation with an automatic adopter migration, plus temp-file hygiene. No methodology or KB-schema changes; drop-in over v2.0.5 (aid update <tool>).

Changed

Discovery term exclusions now live in settings.yml. The user-confirmed term-exclusions list — the terms /aid-discover treats as already-known and never re-flags — moves out of the tracked hidden .aid/knowledge/.term-exclusions.md dotfile into .aid/settings.yml under discovery.term_exclusions, alongside discovery.doc_set. It is now a first-class, visible setting rather than a hidden file inside your Knowledge Base, and /aid-discover reads, creates, and updates it there.

aid update <tool> migrates you automatically. If your project still has the old dotfile, its terms are carried into settings.yml (under a discovery: section, created if absent) and the old file is retired to .aid/.trash/. The migration is gated on that file's presence — projects that never had one are untouched — and is idempotent. Implemented in both the bash and PowerShell installers.

Fixed

Transient scratch no longer contaminates the Knowledge Base. /aid-summarize and /aid-discover used to write build/scratch artifacts (the summary-src/ workspace, .manual-checklist.json, .spot-check-facts.txt, the Markdown-export build) inside .aid/knowledge/, mixing throwaway files with your tracked KB. They now write everything under the gitignored .aid/.temp/, so a discovery or summary run leaves .aid/knowledge/ clean.

KB document counts no longer include hidden files. The KB metrics scanner build-metrics.sh was missing the ! -name '.*' dotfile guard that its five sibling scanners already carried, so hidden files could skew the document counts. Fixed.

Unchanged

Methodology phases, KB schema, CLI flags, skill behavior, and all output contracts are identical to v2.0.0–v2.0.5. An internal clean-code pass removed redundant comments from the shipped skills, agents, and scripts with no behavior, API, or output change.

Verify your download

sha256sum -c SHA256SUMS

Offline bundles

AID v2.0.5

v2.0.5 ·

AID v2.0.5 — a /aid-discover review-scope fix, companion to v2.0.4's scan-scope work: v2.0.4 stopped the scanners from walking non-source trees; v2.0.5 stops the review from grading non-knowledge files. No methodology or KB-schema changes; drop-in over v2.0.4 (aid update <tool>).

Fixed

The /aid-discover review's two keystone gates — M3 (Essence / teach-back) and M4 (Assertiveness / act-back), which force grade ≤ D — read a raw .aid/knowledge/*.md glob with no category filter. So they ingested the KB's meta process/ledger docs (STATE.md, README.md, external-sources.md — kb-category: meta) and the generated INDEX.md as if they were hand-authored project knowledge, poisoning the reconstruction / work-simulation and, through those gates, the grade itself. (The M1/M2 mandates already routed by kb-category, so only the keystone gates leaked.)

v2.0.5 introduces a single, deterministic, tag-driven definition of the reviewed knowledge surface — list_reviewable: the KB .md docs whose frontmatter is kb-category != meta AND source != generated. The review orchestrator computes this once and passes the explicit list to M3/M4 instead of the glob:

  • Excludes the meta ledgers (STATE.md, README.md, external-sources.md) and generated docs (INDEX.md); keeps all hand-authored primary/extension knowledge.
  • Because it reads each doc's own frontmatter, it needs no hardcoded filename list and adapts if tags are corrected — a doc with no frontmatter defaults to reviewable, so the surface never silently shrinks below the primary docs.
  • M2's authoring checks were clarified to apply only to Full Primary docs; M3's blind-reconstruction stage was hardened to disregard ambient host/agent instruction files (CLAUDE.md, AGENTS.md); and build-project-index.sh's stale default output path was realigned to .aid/generated/ so a generated index can never sit in the review surface.

Deterministic and robust: one batched awk pass (no per-file spawn), LC_ALL=C ordering, and an empty/absent KB directory returns success with empty output (never aborts the review). Verified by a new regression suite plus the full canonical suite; profiles regenerated and byte-verified against the emission manifests.

Unchanged

Methodology phases, KB schema, CLI flags, and all output contracts are identical to v2.0.0–v2.0.4.

Verify your download

sha256sum -c SHA256SUMS

Offline bundles

AID v2.0.4

v2.0.4 ·

AID v2.0.4 — a KB-scanner scope & reproducibility patch. The two scanners that feed /aid-discover (harvest-coined-terms.sh → candidate concepts, build-project-index.sh → project index) now scope their walk to real, hand-authored target-project source, deterministically. No methodology or KB-schema changes; drop-in over v2.0.3 (aid update <tool>).

Fixed

Previously the scanners walked directories and files that aren't target-project source, which polluted candidate-concepts.md / project-index.md and — because those inputs change independently of the project — broke the byte-reproducibility the KB relies on (false /aid-housekeep KB-DELTA drift, non-deterministic concept-closure). Both scanners now exclude, in lockstep:

  • AID tool-install ("dogfood") trees at the repo root — .claude/, .cursor/, .codex/, .agent/. On a project that has AID installed these are the AID install itself, never the target's source (the same reason .aid/ was already pruned). .github/ is deliberately not pruned (it's a standard project directory).
  • .gitignored files — via git check-ignore with the global core.excludesFile neutralized, so only the project's committed .gitignore drives exclusions (reproducible across machines). Tracked files are never dropped. Git-repo-guarded, with a non-git fallback.
  • Generated / vendored / minified files — .gitattributes linguist-generated + linguist-vendored (project-declared; documentation is kept, since coined terms live in prose), @generated / DO NOT EDIT header markers, and *.min.js / *.min.css / *.map.
  • More build/cache/output dirs in the always-on prune set — .venv, venv, coverage, htmlcov, .nyc_output, logs, tmp, temp, bower_components, .cache, .turbo, .svelte-kit, .angular, Pods, .dart_tool, .terraform, and more.

All exclusion is deterministic and cross-OS reproducible: every check derives only from committed inputs (.gitignore, .gitattributes, file headers, paths) with machine-specific global excludes neutralized, and every sort runs under LC_ALL=C. Each check is a single batched process (no per-file subprocess). Verified byte-identical to v2.0.3 on projects without such content, on both git and non-git roots; regression tests assert every layer excludes noise while real source and docs are retained.

(Known residual: git check-ignore also consults the per-clone .git/info/exclude, which has no override flag; it is empty by default, so output is reproducible across machines in practice.)

Unchanged

Methodology phases, KB schema, CLI flags, and all output contracts are identical to v2.0.0–v2.0.3.

Verify your download

sha256sum -c SHA256SUMS

Offline bundles

AID v2.0.3

v2.0.3 ·

AID v2.0.3 — completes the cross-platform performance hardening begun in v2.0.1 (/aid-discover harvest) and v2.0.2 (project-index). This release extends the same fix across the entire KB / discovery / summarize toolchain: every helper an adopter runs is now fast on Windows Git Bash / MSYS. No methodology or KB-schema changes; drop-in over v2.0.2 (aid update <tool>).

The problem

On Windows Git Bash / MSYS (and cygwin), every fork()/exec costs 0.5–1.8s — thousands of times slower than on Linux (microseconds). Any loop that spawns a subprocess per file / per term / per line (echo | grep, $(...), per-item awk/sed/tr/wc/basename) therefore turned multi-second Linux runs into multi-minute — or hung — Windows runs. v2.0.1/v2.0.2 fixed the two hottest offenders; v2.0.3 sweeps the rest.

Fixed

Thirteen more KB/pipeline helpers were re-batched (per-item spawns → a single awk pass or bash builtins). Measured Windows speedups from testing:

  • build-kb-index.sh (RAG routing table, runs last in every discover cycle): ~8m46s → ~36s
  • closure-check.sh (concept-closure oracle): >10 min → ~14s
  • cleanup-classify.sh: ~213s → ~23s
  • kb-actback-task.sh: ~59s → ~8.5s
  • plus lint-frontmatter, build-kb-index, kb-freshness-check, kb-citation-lint, build-metrics, kb-dual-intent-probes, grade-summary, validate-html-output, build-md-export, complexity-score, and parse-recipe.

A follow-up hot-path polish pass removed the remaining incremental forks in the most-invoked scripts: read-setting.sh (lazy path resolution — called on every skill invocation), recon-classify.sh (dropped two /tmp round-trips on the discover triage path), assemble.sh, summarize-preflight.sh, and housekeep-state.sh.

Every change is byte-identical to prior output and cross-OS reproducible — verified by old-vs-new diffs on real inputs plus the full canonical test suite (Windows + Linux). No behavior, format, or contract changed; only the number of subprocesses did.

Unchanged

Methodology phases, KB schema, CLI flags, and all output contracts are identical to v2.0.0–v2.0.2. This is purely a performance patch.

Verify your download

sha256sum -c SHA256SUMS

Offline bundles

AID v2.0.2

v2.0.2 ·

AID v2.0.2 — a discovery-performance patch, companion to v2.0.1's /aid-discover fix. No methodology or KB-schema changes; drop-in over v2.0.1 (aid update <tool>).

Fixed

  • build-project-index.sh no longer forks a subprocess per file. The project-index step (run once before the discovery sub-agents) built its file inventory with lang=$(detect_lang "$path") inside a per-file loop — a command substitution that forks a subshell for every file. On 1400 files under Windows Git Bash / MSYS that cost ~1–2 min. Language detection now happens inside the single awk pass that builds the inventory, eliminating all per-file spawns (**1m15s → ~12s** on a ~280-file tree in testing). Output is byte-identical to v2.0.1.

Unchanged

Methodology phases, KB schema, and all output contracts are unchanged. Companion to the v2.0.1 harvest fix.

Verify your download

sha256sum -c SHA256SUMS

Offline bundles

AID v2.0.1

v2.0.1 ·

AID v2.0.1 — installer-upgrade correctness plus a Windows discovery-performance fix. Patch release; no methodology or KB-schema changes. Adopters upgrade with aid update <tool>.

Fixed

  • /aid-discover no longer hangs on Windows. The deterministic coined-term harvest spawned 10 processes per file plus per-token/per-row subprocesses; under Windows Git Bash / MSYS fork (0.5–1.8s per spawn) this took 15–40 min or hung outright on real brownfield repos. Extraction is now batched (ripgrep when available, coreutils-grep fallback — byte-identical output either way), turning it into tens of seconds. Output is byte-identical to v2.0.0 on every platform.
  • In-place installer upgrades apply completely. aid add / aid update over an older install now overwrite AID-owned files on diff (stale flat-path skill bodies were previously skipped silently), seed .aid/settings.yml if missing, and maintain a managed .gitignore region for AID's transient directories.
  • Installer Bash ↔ PowerShell parity + hardening: copy-failure surfacing, .gitignore CRLF idempotency, ## Workflow heading de-duplication, settings-format stamping, and related audit fixes.
  • Docs / reference cleanup: removed dangling /aid-init, setup.sh, and legacy-path references across profile READMEs and skill docs; refreshed kb.html; corrected the phase-model line in the profile root files.

Unchanged

Methodology phases, KB schema, and the closure/essence-capture contracts are unchanged. This is a drop-in patch over v2.0.0.

Verify your download

sha256sum -c SHA256SUMS

Offline bundles

AID v2.0.0

v2.0.0 ·

AID v2.0.0

Five per-profile tarballs + CLI bundle for the AID methodology installer.

Verify your download:

sha256sum -c SHA256SUMS

Offline bundles

AID v1.1.1

v1.1.1 ·

AID v1.1.1

Five per-profile tarballs + CLI bundle for the AID methodology installer.

Verify your download:

sha256sum -c SHA256SUMS

Offline bundles

AID v1.1.0

v1.1.0 ·

AID v1.1.0

Five per-profile tarballs + CLI bundle for the AID methodology installer.

Verify your download:

sha256sum -c SHA256SUMS

Offline bundles

AID v1.0.0

v1.0.0 ·

The first public release of AID — AI Integrated Development: a methodology and toolchain for building software with AI coding agents, human-gated at every phase. This release covers the full methodology currently available, grouped by area.

Pipeline & Skills

  • Six numbered phases as slash commands: Discover (/aid-discover), Interview (/aid-interview), Specify (/aid-specify), Plan (/aid-plan), Detail (/aid-detail), Execute (/aid-execute).
  • /aid-config — bootstrap a project and manage .aid/settings.yml.
  • Optional on-demand Deliver skills (not numbered phases): /aid-deploy, /aid-monitor.
  • /aid-summarize — offline knowledge-summary.html viewer.
  • /aid-housekeep — off-pipeline Knowledge Base maintenance. (11 user-facing skills in total.)

Knowledge Base

  • 14 standard Knowledge Base document types populated by Discover, with RAG-by-convention navigation.
  • Per-project configurable doc-set (discovery.doc_set).
  • Self-contained offline KB summary (Mermaid inlined, light/dark, WCAG AA).

Agent Model

  • 9 specialized agents across three tiers (large / medium / small).
  • Structural separation of duties — the reviewer's tier is always ≥ the executor's, and the agent that writes code never grades its own work.

Lite Path & Recipes

  • Description-first TRIAGE auto-routes small work to the lite path (skips Specify, Plan, and Detail).
  • 51 recipes (add- / change- / fix- families) with {{slot}} substitution.
  • Sub-paths LITE-BUG-FIX / LITE-REFACTOR / LITE-FEATURE, with mid-flight escalation to the full path.

Quality, Review & Feedback

  • Two-tier review: a per-task quick-check plus a per-delivery graded review/fix loop (A+ → F, deterministic grade).
  • Configurable minimum grade (global, with per-skill overrides).
  • 11 formal feedback loops for upstream revision (Q&A entries, IMPEDIMENT.md, MONITOR-STATE findings).

Execution Engine

  • 8 task types: RESEARCH, DESIGN, IMPLEMENT, TEST, DOCUMENT, MIGRATE, REFACTOR, CONFIGURE.
  • Parallel pool dispatch with configurable capacity (execution.max_parallel_tasks).
  • Branch-per-delivery isolation (aid/{work}-delivery-NNN).
  • Always-on traceability (L1 state markers, L2 ETA brackets, L3 heartbeat files).

Installation & Host-Tool Support

  • Persistent global aid CLI: add / status / update / remove / version.
  • Four install channels: curl/irm script, npm, PyPI, and offline bundle.
  • Five host tools: Claude Code, Codex CLI, Cursor, GitHub Copilot CLI, and Antigravity (byte-identical skill/agent bodies).
  • Protect-on-diff for root agent files, an invariant root AGENTS.md, and a tag-triggered release pipeline.

Documentation & Examples

  • The methodology guide, FAQ, glossary, repository map, and install/release runbooks.
  • This documentation site.
  • Worked examples: greenfield, brownfield full-path, and brownfield lite-path.

Install: see the Installation guide. Verify downloads with sha256sum -c SHA256SUMS.

Offline bundles

Report an issue with this page →