CHANGELOG.md

Changelog

All notable changes to agentic-toolkit are recorded here. Format follows Keep a Changelog.

Versions are retro-fitted from commit history; dates are ballpark, grouped by milestone rather than per-commit.

[Unreleased]

Added

• ELI5 output style (ai/customize/config/ELI5.md) — a Claude Code output style that asks for plain, minimal answers (small words, short sentences, at most two options when a decision is needed, exact paths/commands preserved).
• STE100 output style (ai/customize/config/STE100.md) — a Claude Code output style based on ASD-STE100 Simplified Technical English: short sentences, one instruction each, active voice, common words, condition-first phrasing, exact paths/commands.

[1.12.1] — 2026-08-07

Repo-hygiene and tooling housekeeping — no functional changes.

Changed

• Statusline context indicator switched from an ASCII bar to a colored percentage. ai/customize/config/statusline-command.sh replaced the 20-char [==== ] context-usage bar with a compact | NN% readout that colors by pressure — green <25%, orange <35%, red otherwise. Synced from the live ~/.claude/statusline-command.sh.
• Agent/IDE scratch gitignored and de-tracked. .gitignore now default-denies every dot-directory (.*/), re-admitting only what ships (.github/). Per-machine agent/IDE state (.claude/, .litectx/, .idea/, …) regenerates locally and only added noise and churn; any already-committed copies are removed from tracking (local files kept on disk). Repo hygiene only.
• node_modules/ gitignored. No rule existed, unlike liteagents; a stray node_modules/ under the bundled live-canvas-channel plugin showed as untracked noise. Nothing was previously tracked, so no de-tracking was needed.

[1.12.0] — 2026-08-04

Dependency security fixes mirrored from liteagents@2.15.3.

Security

• Merges Dependabot #14/#13: hono 4.13.0 (closes residual moderate ReDoS, GHSA-8j4g-w8fx-2239), ip-address 10.4.0 (3 high SSRF/trust-boundary advisories). Also applies npm audit fix for fast-uri 3.1.5 (high, GHSA-7p8r-x3mc-p8w7) — not yet proposed as a separate Dependabot PR. Transitive deps of the bundled live-canvas-channel plugin's MCP SDK. Plugin audit: 3 → 0 vulnerabilities.

[1.11.0] — 2026-07-13

Hot-memory pipeline change mirrored from liteagents@2.15.0.

Changed

• /stash delegates the write-up to a mid-tier-model subagent (mirror from liteagents). The session drafts the raw content inline (only it holds conversation context), then hands off to a mid-tier-model subagent to expand and write the formatted stash file — dispatched in the background where the tool supports it, falling back to writing inline otherwise. Applied identically across all four packages.
• subagentic-manual.md documents the new mechanism in its Hot Memory section.

[1.10.0] — 2026-07-10

Hot-memory pipeline changes mirrored from liteagents@2.14.0 and liteagents@2.14.1.

Added

• /remember bootstraps a standards-guide template (mirror from liteagents). A new AGENT_RULES.md (an AI agent collaboration/coding-standards guide) ships bundled next to friction.js in all four packages. On first /remember run in a project, if <tool-dir>/remember/AGENT_RULES.md doesn't already exist, it's copied from the bundled template — never overwritten again after that, so local edits persist. When present, it's injected into the agent config via its own independent marker pair (<!-- AGENT_RULES:START/END -->), separate from the MEMORY.md block and framed as a guide to consult when building something new — not hot context loaded every session.

Changed

• /remember extraction is parallel and model-agnostic (mirror from liteagents). Step 2's per-stash extraction calls are now spawned as concurrent subagent calls instead of one at a time. Every hardcoded sonnet mention (steps 2, 3, 4a) is replaced with "the mid-tier model" — a new Guardrails note explains the intent (capable of semantic judgment, cheaper/faster than the top reasoning tier; Sonnet is the Claude example, not a requirement) so the instructions work unmodified across Claude/opencode/ampcode/droid regardless of which models each tool has configured.
• subagentic-manual.md documents both changes in its Hot Memory section.

[1.9.0] — 2026-07-08

Hot-memory pipeline changes mirrored from liteagents@2.13.0 (plus the 2.12.x-era mirrors that were sitting in Unreleased).

Added

• Antigen ledger (/remember step 4c) (mirror from liteagents). Every behavioral rule now carries an evidence trail in <tool-dir>/remember/ledger.json: which mistake-class it targets (class_hints dedup key), the evidence that promoted it, and every phrasing ever tried (attempts — failed attempts are the rejected-edit buffer, never re-proposed). A class that fires again while its rule is loaded increments recurred_while_hot: at 2 the phrasing is marked failed and rephrased; after 2 failed phrasings the antigen is ESCALATED — removed from hot, recorded as a Fact, flagged for a human decision. Failure detection without statistics; instructions-only (no new code), identical across all four packages.
• /remember writes its run report to <tool-dir>/remember/report.md (latest snapshot, overwritten each run).

Changed

• Pipeline consolidated to two dirs, each owned by its command (mirror from liteagents): <tool-dir>/stash/ (/stash) and <tool-dir>/remember/ (MEMORY.md, ledger.json, report.md, .processed, transient friction/ output). Was three (stash/, friction/, memory/). /remember performs a one-time loud migration: pipeline files move, user-owned files in the old memory/ stay put, stale friction output is discarded (always regenerated fresh).
• Claude's bundled dir joins the naming convention: claude/commands/friction/friction.js → claude/commands/remember/friction.js — a command's helper dir is now named after the owning command in all four packages (the other three were renamed earlier, still in Unreleased below).
• context-builder reads/injects <tool-dir>/remember/MEMORY.md (was .../memory/MEMORY.md) across all four packages; /stash's consolidation-nudge backlog count reads .../remember/.processed; subagentic-manual.md path references updated.
• Friction cluster ranking breaks equal-recurrence ties by intensity (mirror from liteagents). Antigen/episode clusters were ordered by tier then recurrence, with equal-recurrence ties left to incidental order — so a mild reaction could outrank a far more intense one that recurred equally. Added a final tiebreak on median peak friction (already computed): the more intense reaction ranks first. Ranking-only — never promotes across the severity × recurrence 2×2 (a loud one-off stays an episode). Applied across all four tool packages (claude, droid, opencode, ampcode).
• docs: friction-README.md → remember-README.md (mirror from liteagents) — refreshed for the two-dir layout + ledger; README pointer updated.

Fixed

• Non-Claude helper dir renamed friction/ → remember/ (opencode, ampcode, droid). These tools expect a command's bundled directory to share the command's name; friction.js is run by /remember, so it now lives in remember/friction.js (was the mismatched friction/). remember.md's bundle-path reference updated to match.

[1.8.0] — 2026-07-03

Added

• /release command across all four tool packages (claude, droid, opencode, ampcode). An end-to-end feature-delivery orchestrator that does not re-implement checks — it runs the existing gates (/ship, /security, /diff-review) under /verify-done discipline, then performs the release actions. Split by a hard gate: everything before is safe/read-only (preflight branch resolution, verify with fresh evidence + coverage table); everything after rewrites history and is confirmed step by step (docs update, semver bump, commit, push, PR, merge, tag). Publish stays manual (workflow_dispatch by design). Usage: /release [branch] — requires a feature branch, never releases main. Registered in opencode.jsonc; catalogs in each AGENT(S)/CLAUDE.md and subagentic-manual.md bumped (Claude/Amp 8→9 commands, Droid/OpenCode 17→18).
• live-canvas recategorized on Amp from a command to a subagent in ampcode/AGENT.md (batch mode only on Amp).

Changed

• .litectx/ (local litectx index DB) added to .gitignore.

[1.7.0] — 2026-06-16

Changed

• /friction collapsed into /remember; hot-memory pipeline is now two commands (/stash → /remember) (mirrored from liteagents). /remember runs friction.js itself (best-effort) against the tool's global sessions root — resolved from an editable, never-prompt probe list baked into remember.md (Claude/Droid/Amp/opencode + Codex/Antigravity; add your own at the top) — then consolidates. A no-sessions miss is surfaced loudly, never silently skipped. /stash now nudges toward /remember at ≥5 unprocessed stashes (derived count, no counter file). Each package's /remember writes the memory into its own agent config (CLAUDE.md / AGENTS.md / AGENT.md). Standalone /friction command removed across all four tool dirs (the friction.js script stays). Counts, catalogs, subagentic-manual.md, opencode.jsonc, docs/friction-README.md, and README updated.

Fixed

• friction.js no longer crashes on a single malformed JSONL line (mirrored from liteagents). The four tool copies under ai/subagentic/*/.../friction/ parsed session logs and friction_raw.jsonl with bare .map(line => JSON.parse(line)) — one corrupt line aborted the whole run. A new parseJsonl(raw, source) helper skips bad lines with a one-line stderr warning (line number + source) and keeps the good records; the whole-file friction_analysis.json read is now wrapped in try/catch (reports the file, exits 1).
• ai/subagentic/subagentic-manual.md restored after a range-sed corruption inherited from the 1.6.0 mirror. The corruption originated upstream in liteagents (a range-sed with end-pattern test-generate$ matched far past its intended scope and overwrote ~310 lines with duplicates of the "Simple Commands" bullet) and rode along to this repo when the file was copied in commit bda2c0c. Restored from the cleaned liteagents version; all renames and count updates intact.

Removed

• ai/customize/memcp/ — the experimental MCP memory server (semantic-search persistent memory for Claude Code). Internal tooling, not published. Removal clears the open Dependabot alert for uuid < 11.1.1 (the only direct/indirect dep using uuid in this repo). README references in the ai/customize/ description and the Structure block updated.

[1.6.0] — 2026-06-16

Subagentic command set consolidated and verify-then-fix loops baked into the claim-shaped commands. Five 3-word skill slugs renamed to 2-word slugs. /review renamed to /diff-review to avoid collision with the Anthropic-official code-review plugin. Synced from liteagents@2.9.0+.

Changed

• /review → /diff-review across all four tool packages (claude, droid, opencode, ampcode). Avoids the name collision with the Anthropic-official code-review plugin's review skill. The renamed command absorbed the old /code-review (which had been workflow ceremony for requesting reviews from a subagent) and now accepts a file, branch (e.g. /diff-review main diffs merge-base(main, HEAD)..HEAD), or explicit range (main..HEAD).
• Claim-shaped commands now verify → fix → ask. /diff-review, /security, /optimize, /test-generate all re-ground each cited file:line before acting, auto-fix only confirmed + unambiguous + no-contract-change findings, and stop to ask on uncertain claims, multi-shape fixes, downstream-affecting changes, security primitives, or "dead code" that may be intentionally kept. /diff-review ends with a Ready to merge? Yes / No / With fixes verdict.
• /refactor now runs the tests after the edit (lightweight pattern: the user's input is the claim; tests are the verifier). Detects the project's test command, runs it scoped to the affected area, reports pass/fail. On failure stops and asks (revert / patch / update test) rather than auto-reverting or pushing forward silently.
• /test-generate rewritten as a generate-and-verify-it-bites loop. Discovers existing framework (refuses to add a new runner), generates, runs the new tests, and marks each one biting or superficial by mentally swapping in a broken impl. Superficial tests (expect(true).toBe(true), mock-asserting-itself, setup-masked passes) count as a failure to ship.
• Skills renamed to short 2-word slugs. systematic-debugging → debug-method (4-phase framework, with its 4 pressure-test scenarios + creation log preserved). root-cause-tracing → trace-back (with its find-polluter.sh bisection helper preserved). testing-anti-patterns → test-traps (now includes timing/polling as Anti-Pattern 6 after the condition-based-waiting fold). test-driven-development → tdd-flow. verification-before-completion → verify-done. Resulting cluster reads scannably: debug-method / trace-back / verify-done and tdd-flow / test-generate / test-traps.
• condition-based-waiting folded into test-traps as Anti-Pattern 6: Timeout-Based Waiting. Promotes the timing/polling guidance to auto-trigger coverage (was previously manual-trigger only). Its example.ts helper (waitForEvent / waitForEventCount / waitForEventMatch) moves with it and is referenced from AP6.
• installer/cli.js banner derives counts from package.json.description — was hardcoded as "11 agents + 23 commands per tool" and drifted silently. Now parses 11 specialized agents and 18 commands from the description, so the banner auto-tracks whenever the count changes. Fallback values prevent crash on regex miss.

Removed

• /code-review (was workflow ceremony for requesting reviews; mostly overlapped /diff-review). Use /diff-review instead — /diff-review main for branch-vs-main, no args for staged/working-tree.
• /debug — was a thin echo of the systematic-debugging skill. The renamed skill (debug-method) carries the real workflow with its pressure-test scenarios.
• /explain — 11 lines of "explain this code" with no real constraints. The model does this naturally from a plain prompt.
• /git-commit — Claude Code has built-in commit handling; the other three tools don't need a thin wrapper either. Use natural-language prompts instead.

Counts

• Claude: 9 skills + 9 commands (was 11 skills + 12 commands).
• Droid / Opencode / Ampcode: 18 commands per tool (was 23 commands).
• package.json description updated to reflect the new totals.

[1.5.5] — 2026-06-16

Changed

• README simplified (210 → 75 lines) to match the repo — AI subagent kits for four tools plus per-distro Linux dev-tool setup (tools-debian/tools-fedora). Dropped the marketing hero and the Documentation / Support & Community / Contributing / License sections (kept a one-line license note), and fixed dead links: the removed tools/ paths, the deleted CONTRIBUTING.md, and the vibecoding-101.md filename typo.

[1.5.0] — 2026-05-26

Friction memory pipeline redesigned; ships the previously-unreleased /security, /ship, and /git-commit hardening; prunes stale docs.

Added

• docs/friction-README.md — canonical guide to the stash → friction → remember (hot-memory) pipeline.

Changed

• /friction redesigned: antigens come from observed user reactions, not machine proxies — clusters by what the user said (content/phrase overlap) instead of (signal, tool_pattern); recurrence × severity drives the suggested artifact; only patterns recurring across 5+ sessions load into hot memory. Pasted SSH/shell output is no longer mistaken for friction, and profanity counts only when it's aimed at the agent. Applied across all four platforms.
• /remember rewritten to consolidate from friction's short quotes, never raw logs — classifies agent-vs-self, drops self-corrections, merges paraphrases, tiers antigens by recurrence, and surfaces self_suspect for confirmation. MEMORY section renamed Preferences → Antigens.
• /security command hardened across all four platforms (Claude, Opencode, Ampcode, Droid). Replaces the five generic categories with "the recurring six" — secrets in the repo, tenant isolation, rate limiting, error handling past the happy path, authorization beyond authentication (IDOR), and inefficient data access (N+1) — plus an expanded "also scan for" list (injection, trust boundaries, config, dependencies). Output is now coverage-auditable: it reports which classes were checked clean and which are N/A, not just the hits. Adds read-only git/rg tooling (git log, git grep, rg) to allowed-tools.
• /ship command hardened across all four platforms. Now detects the stack first (npm / pyproject / go.mod / Cargo / Makefile) and runs only checks that exist, reporting each as pass / fail / N/A. Adds security-relevant gates (ownership + role authorization, rate limiting, data-access scoping/scaling) and broadens allowed-tools across stacks (pnpm, yarn, pytest, python, go, cargo, make).
• /git-commit (Claude) allowed-tools syntax fix — Bash(git *) → Bash(git:*) to match Claude Code's matcher format.
• README cleaned of emoji — removed decorative emoji from headings, inline links, and list markers for a plain-text, professional presentation. Emoji bullet lists ("Who Is This For?", "Support & Community") converted to standard markdown bullets; "Built with ❤️" now reads "Built with love"; fixed the #quick-start anchor link that had depended on the heading emoji.

Removed

• Pruned nine stale planning/spec docs from docs/ (agent consolidation/cleanup plans, split summary, digraph notes, subagents-and-skills, verification-and-isolation pattern, FAQ, CONTRIBUTING) — superseded or merged into the README/guide.
• Dead scaffolding in friction.js: the unused overlap() helper, the MIN_KW/MIN_INTER constants, the unread selfCount (superseded by anySelf), and the empty top_files field with its unreachable renderer.

Infrastructure

• Root package.json added (private; metadata only — this repo is multi-language, package.json is for the version badge and npm-style tooling integration).
• README badges (version + license, plato-style; #2a4f8c).

[1.4.0] — 2026-05-18

live-canvas matures: simpler overlay, robust channel server, JSON mode writes to disk. The live-canvas-channel MCP plugin (v0.5.0) gains explicit lifecycle tools, a capability gate that prevents silent feedback loss, and automatic takeover when a prior Claude session is holding the port.

Added

• Channel server: lazy port binding via MCP tools — channel_open, channel_close, and batch_open. Port 8788 is only bound when the skill explicitly calls one. Plain Claude sessions stay idle; multiple sessions can coexist with /mcp green without racing for the port.
• Channel server: parent-flag capability gate — channel_open inspects the parent claude process's command line and refuses to bind from sessions launched without --dangerously-load-development-channels. Before this, a plain claude could win the port and silently drop every notification (POST 200 but no <channel> tag) — the "nothing landed" black hole. Cross-platform: Linux /proc/<ppid>/cmdline, macOS/BSD ps -p <ppid> -o args=, Windows wmic.
• Channel server: automatic sibling takeover — when channel_open finds port 8788 held by another instance of the same plugin running as the same uid, it takes over (SIGTERM, rebind, SIGKILL fallback). The taken-over pid is returned as took_over so the skill can announce it. Foreign processes are still refused with {status: "in_use", holder_pid} — the plugin won't kill anything it doesn't own.
• JSON mode writes to disk — channel server gains a POST /feedback-jsonl route that appends submissions to <project>/.claude-design/feedback.jsonl. Skill calls batch_open (no flag gate) and sets the overlay's batchEndpoint accordingly. Falls back to the legacy browser download only when the MCP isn't running.
• SKILL.md Case D — explicit relaunch instructions — when the tool returns no_channel_capability, the skill prints the exact live-claude command instead of a generic "Live mode unavailable" error.
• Lab banner template — new templates/lab-banner.html ("temporary review surface") replaces the old per-mode banners. Mode-agnostic, paste-once.

Changed

• Vanilla overlay everywhere — deleted the React-specific feedback components (~2300 lines). overlay-vanilla.js (one file, plain DOM, zero deps) now works in every supported framework, including React/Next.js/Vite via <script> + useEffect.
• User-facing rename "Batch" → "JSON" — the non-Live mode is called "JSON mode" everywhere user-facing.
• Demo relocated to dev/ — templates/demo/post-variants.html was never copied during real runs. Moved to dev/post-variants.html at the skill root in each platform package.
• Explicit mode pick — skill asks Live vs JSON via AskUserQuestion every run instead of silently auto-detecting.
• SKILL.md mode-selection: replaced curl /health probe with the mcp__live-canvas__channel_open tool call. The tool's structured result (opened / already_listening / in_use / no_channel_capability / took_over) is authoritative.

Fixed

• Channel server shutdown race — server.close() is async but process.exit() was synchronous; stale processes held port 8788 indefinitely after the MCP host disconnected, breaking /reload-plugins and subsequent sessions. Now uses a closing guard and lets server.close() callback drive exit (500ms unref'd ceiling).
• Overlay mode badge stale on re-expand — collapsing and re-expanding the overlay showed the wrong mode name after a runtime live→batch fallback. Badge text now refreshes from state.mode on every re-expand.
• setup.sh sudo guard — bails when run with sudo instead of silently installing into /root/.claude/plugins/.
• Silent channel black-hole when a plain claude won the port race — fixed by the capability gate above; the failure mode can no longer occur.

Removed

• React-specific feedback templates (templates/feedback-react/) in all platform packages.
• INTEGRATION_NOTES.md (draft notes superseded by README + SKILL).

[1.3.0] — 2026-05-10

live-canvas skill lands across all four platforms — design UI variations with click-to-annotate browser feedback. Claude Code ships a companion live-canvas-channel MCP plugin that enables live mode (each overlay Save streams into the session in real time); other platforms run in batch mode only.

Added

• live-canvas skill/command across Claude, Ampcode, Droid, Opencode (templates: vanilla overlay, React feedback components, demo variants).
• Claude Code: live-canvas-channel MCP plugin (marketplace + setup script) for live streaming feedback.
• variants.json across all platform packages.

Changed

• Skills count: 10 → 11 (live-canvas).
• Droid/Opencode commands: 22 → 23 (live-canvas).
• friction command refreshed across all platforms.
• subagentic-manual.md and README.md updated to reflect new inventory + live-canvas channel plugin note.

[1.2.0] — 2026-02-13

Documentation hardening pass: tone moved from "preferences" to "standards"; testing strategy rewritten with battle-tested best practices; development philosophy (POC-first, vanilla-first, lightweight) added; tech-stack tables collapsed to a minimal environment section; agent rules restructured for coherence with a CLAUDE.md stub.

Changed

• Agent rules restructured: coherence pass + CLAUDE.md stub.
• Testing strategy rewritten with battle-tested best practices.
• Tone hardened from preferences to standards.
• Tech stack tables → minimal environment section.

Added

• Development philosophy doc: POC-first, vanilla-first, lightweight.

Removed

• Replit references from dev-workflow docs.
• AI tools section from README; Radix UI moved to correct table.

[1.1.0] — 2026-01-24

Post-1.0 polish: liteagents extracted to its own repo; stash command added across all platforms; Kitty terminal installer + config; minor frontmatter and provider-rename refinements.

Added

• stash command across all platforms (Claude / Opencode / Ampcode / Droid).
• Kitty terminal installer script + working config + menu integration.
• npm global install documented as recommended installation method.

Changed

• Subagentic frontmatter: id → name; title removed from agent frontmatter (continuation of 1.0's auto-discovery).
• Removed Superpowers / BMAD references; cleaned obsolete hooks and stale README files from the Claude package.
• Skill count corrected: 21 → 20 after subagent-spawning skill removed.

Extracted

• liteagents — extracted to its own repo (hamr0/liteagents) after the rename. agentic-toolkit and liteagents now sit alongside each other rather than nested.

[1.0.0] — 2026-01-18

Stable cut. Agent consolidation 15 → 14 → 11; skills consolidation 22 → 11; commands stable at 10. Phase 9 manifest update finalizes the inventory. Frontmatter-based auto-discovery replaces central manifests so agents/skills/commands self-register from their own metadata.

Changed

• Agents: 15 → 11. Each renamed; digraph flow definitions added.
• Skills: 22 → 11. Removed 13 non-core skills.
• Commands: 10 essential slash commands (debug, explain, git-commit, optimize, refactor, review, security, ship, stash, test-generate).
• Subagentic architecture: frontmatter-based auto-discovery.

Added

• Phase 9 manifest files: skills/commands inventory.
• Standardized command/skill descriptions with argument hints.

Removed

• Task-template references from all subagentic-manual.md files.

[0.9.0] — 2026-01-17

Phases 6-8. Self-verification, task-type detection, and TDD auto-trigger. Verification stops being optional — every non-trivial task triggers it.

Added

• Self-verification gate (Phase 6).
• Task-type detection (Phase 7).
• TDD auto-trigger (Phase 8).

[0.8.0] — 2026-01-17

Phases 2-5. Subagent isolation, TDD hints, document verification. Synced across all platforms.

Added

• Subagent isolation (Phase 2).
• TDD hints (Phase 3).
• Document verification (Phases 4-5).
• Per-package sync across Claude / Opencode / Ampcode / Droid.

[0.7.0] — 2026-01-17

Phase 1: Universal Verification Gate. Every agent runs through a verification step before completing tasks.

Added

• Universal verification gate.

[0.6.0] — 2026-01-17

Phase 0 cleanup. Removed 13 non-core skills + commands from disk; cleaned references; renamed all agents and added digraph flow definitions; added the essential slash-command set.

Removed

• 13 non-core skills + commands.
• Stale resources and references.

Changed

• All agents renamed; digraph flows added to each.

Added

• Essential slash commands (initial set).

[0.5.0] — 2025-Q4

Droid CLI support lands. 90+ droids configured for Droid CLI BYOK (bring-your-own-keys); custom-model config + README updates. Marketplace skills + plugins reorganized; BMAD references cleaned; npm package configuration completed (Task 1.0). Subagents directory structure simplified.

Added

• Droid CLI subagents package (subagents/droid/).
• 90+ droid agents + custom-model config (BYOK).
• Marketplace skills + plugins.
• "Awesome Claude tools" section + Droid + Synthetic references.
• Amp commands + skills documentation.
• npm package configuration (Task 1.0).

Changed

• Subagents directory structure simplified.
• BMAD references cleaned up.
• Resources path fixed for opencode.jsonc.
• AGENTS.md updated with command set.

[0.4.0] — 2025-10-28

Provider rename + bloat reduction + docs scaffolding. Subagents moved to provider-named directories (claude/, opencode/, ampcode/); ampcode kit reduced from ~113MB to ~80KB; AGENT_RULES relocated under ai/. Documentation: CONTRIBUTING, FAQ, vibecoding-101 placeholder, awesome-llm-skills link.

Changed

• Subagent layout: ai/subagents/{claude,opencode,ampcode}/.
• AGENT_RULES → ai/AGENT_RULES.md.
• Ampcode kit: 113MB → 80KB.
• README: equal treatment of all 3 kits; merge conflicts resolved; consistent naming.

Added

• docs/CONTRIBUTING.md.
• docs/FAQ.md.
• docs/vibecoding-101.md placeholder.
• Awesome LLM skills reference link.

[0.3.0] — 2025-10-28

Subagentic manual + multi-provider scaffolding. Templates and agent files optimized for cross-provider use; Ampcode kit added alongside Claude / Opencode.

Added

• subagentic-manual.md (cross-provider integration guide).
• Ampcode subagents kit.

Changed

• Path harmonization across providers.
• Templates and agent files optimized.

[0.2.0] — 2025-10-28

Initial cleanup of the migrated repo. Multi-provider scaffolding (Claude + Opencode), README rewrites, path normalization.

Changed

• Path normalization across Claude/Opencode kits.
• README rewritten for the agentic-toolkit shape.

[0.1.0] — 2025-10-26

Initial migration from forked agent-dev-suite repo. Restructured into the agentic-toolkit shape; README + initial docs in place.

Added

• Initial subagents structure migrated from agent-dev-suite.
• Initial README + docs.