# Changelog All notable changes to liteagents will be documented in this file. The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html). --- ## [Unreleased] ### Planned - Community marketplace submissions - Additional skills for data analysis - Enhanced testing capabilities - Performance optimizations ## [2.15.2] - 2026-07-24 ### Security - **Cleared the last plugin audit finding via an `overrides` pin: `@hono/node-server` → 2.0.11.** The advisory [GHSA-frvp-7c67-39w9](https://github.com/advisories/GHSA-frvp-7c67-39w9) (moderate, Windows-only serve-static path traversal) is patched only in `@hono/node-server@2.0.5`+ — a **major** bump. The plugin's `@modelcontextprotocol/sdk@1.29.0` (latest) pins `@hono/node-server: ^1.19.9`, a `1.x` range that can never resolve to the fix, so `npm audit fix` couldn't clear it (upstream tracked in [modelcontextprotocol/typescript-sdk#2531](https://github.com/modelcontextprotocol/typescript-sdk/issues/2531), still open). Added the community-confirmed `overrides` block to the plugin's `package.json` forcing `@hono/node-server` to `^2.0.11` within the SDK subtree. Plugin `npm audit` is now **0 findings**. The path was already unreachable here (stdio server, no `@hono/node-server`/HTTP-transport usage); this just removes the audit noise. Note: `@hono/node-server@2.x` needs Node 20+, but the plugin never loads it, so its own `>=18` runtime floor is unchanged. 138 root tests green; remove the override once the SDK widens its range. ## [2.15.1] - 2026-07-24 ### Security - **Bumped the bundled `live-canvas-channel` plugin's transitive deps to patched versions.** `fast-uri` 3.1.2 → 3.1.4 (clears a **high**-severity host-confusion advisory), `body-parser` 2.2.2 → 2.3.0 (DoS via a silently-disabled size limit), and `hono` 4.12.25 → 4.12.31 (moderate). These come in via the plugin's `@modelcontextprotocol/sdk` and ship in the tarball's `package-lock.json`. All sit on unused transport paths — the plugin server is stdio (`StdioServerTransport`) + raw `node:http`, with no `@hono/node-server`/express/Streamable-HTTP surface — so this is defense-in-depth hardening of the bundled lockfile, not a runtime fix. `npm audit` in the plugin drops from 4 findings (incl. 1 high) to 1 residual moderate (`@hono/node-server` serve-static), which is gated on the MCP SDK's own dependency range and has no upstream fix yet. 138 root tests green. ### Fixed - **Publish workflow pinned to `npm@11` — npm 12.0.0's `npm publish --provenance` is broken.** The job ran `npm install -g npm@latest`, which started resolving to npm 12.0.0 (released 2026-07-09) on the Node 22 runner. npm 12's `libnpmpublish` provenance code does `require('sigstore')`, but the tarball bundles only the `@sigstore/*` scoped packages — so `--provenance` dies with `MODULE_NOT_FOUND` and the publish fails outright. npm@11 bundles `sigstore` and publishes fine. Pinned to the major rather than floating on `@latest`. Revisit once npm ships a provenance fix. CI only — no runtime or published-artifact change. ### Changed - **Agent/IDE scratch is gitignored and de-tracked (`.claude/`, `.litectx/`, `.idea/`).** Per-machine agent and IDE state is no part of the package — it regenerates locally and only added noise and churn. Now ignored, and any already-committed copies removed from tracking (local files kept on disk). Functional dot-paths (`.github/`, `.gitignore`, `.npmignore`, `.mcp.json`) stay tracked. Repo hygiene only. --- ## [2.15.0] - 2026-07-13 ### Changed - `/stash` delegates the write-up to a mid-tier-model subagent, dispatched in the background where the tool supports it, instead of writing inline on the top-tier session model — drafting stays inline (only the running session has conversation context), but formatting and file I/O move off the main turn. Falls back to inline writing if subagent/background dispatch isn't available. --- ## [2.14.1] - 2026-07-10 ### Changed - **`/remember` extraction is parallel and model-agnostic.** 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 your 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. Identical across all four packages. --- ## [2.14.0] - 2026-07-10 ### Added - **`/remember` bootstraps a standards-guide template.** 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 `/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 CLAUDE.md/AGENTS.md/ AGENT.md via its own independent marker pair (``), separate from the MEMORY.md block and framed as a guide to consult when building something new — not hot context loaded every session. This repo dogfoods it: its own copy moved from `.claude/memory/AGENT_RULES.md` to `.claude/remember/AGENT_RULES.md`. Design + pipeline walkthrough: `docs/remember-README.md`. ### Docs - **Antigen-gate PRD (§10):** deferred entry for a local classifier model as a paraphrase-blocking *proposer* between friction's shingle clustering and `/remember`'s LLM merge (LLM always disposes each shortlisted merge). Un-defer condition: offline measurement on the existing candidate corpus shows shingle-missed paraphrase merges would move at least one class across a recurrence tier. --- ## [2.13.0] - 2026-07-08 ### Added - **Antigen ledger (`/remember` step 4c).** Every behavioral rule now carries an evidence trail in `/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. Design + the POC evidence that killed the statistical ON/OFF gate (deferred, un-defer condition named): `docs/antigen-gate-prd.md`. - **`/remember` writes its run report** to `/remember/report.md` (latest snapshot, overwritten each run). ### Changed - **Pipeline consolidated to two dirs, each owned by its command:** `/stash/` (`/stash`) and `/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). Validated live on this repo's own memory. - **Claude's bundled dir joins the naming convention:** `packages/claude/commands/friction/` → `commands/remember/` — a command's helper dir is named after the owning command in **all four** packages now (2.12.1 did the other three). - **docs: `friction-README.md` → `remember-README.md`** — updated for the new layout + ledger, linked from README as the pipeline explainer. ### Fixed - **Hot memory was silently not loading in Claude Code.** The managed CLAUDE.md section injected a bare `@MEMORY.md`, which resolves relative to the containing file — i.e. a nonexistent root-level file. Now an explicit `@/remember/MEMORY.md` path in all four packages' injection instructions. - **`/remember` step 4b tier ambiguity:** LLM-merged antigen groups now explicitly obey the recurrence tiers (merging consolidates evidence, never elevates it) — surfaced by dogfooding the ledger's first live run. --- ## [2.12.1] - 2026-07-07 ### Fixed - **Non-Claude helper dir renamed `friction/` → `remember/`** (opencode, ampcode, droid). These tools expect a command's bundled directory to share the owning command's name; `friction.js` is run by `/remember`, so it now lives in `remember/friction.js` instead of the mismatched `friction/`. `remember.md`'s bundle-path reference updated to match. Claude keeps `friction/` — its command loader has no such naming constraint. `opencode.jsonc` needed no change (its only `friction` mention is a description string, not a path). --- ## [2.12.0] - 2026-07-07 ### Changed - **Friction cluster ranking now breaks ties by intensity.** Antigen/episode clusters were ordered by tier then recurrence, with equal-recurrence ties left to incidental feed order — so a mild reaction could outrank a far more intense one that recurred equally. Added a final tiebreak on median peak friction (a value already computed), so the more intense reaction ranks first. Ranking-only: it reorders within what recurrence already gated and never promotes across the severity × recurrence 2×2 (a loud one-off stays an episode). Applied identically across all four packages. ### Fixed - **`validate-packages` failed 0/4 valid on every package.** `validatePackage` demanded a `.md` for each selected command, but a command can legitimately ship as a helper subdirectory with no doc — `commands/friction/friction.js`, run by `/remember` after `/friction` was collapsed into it. The installer already bundles such subdirs; validation now matches that instead of requiring a resurrected `friction.md`. No user-facing command is re-added. --- ## [2.11.1] - 2026-07-03 ### Changed - **`/release` merge step now documents the solo-repo path.** Its "stop and ask the user to approve" guard assumed a second reviewer exists — a dead end on a solo repo, where you cannot approve your own PR. Added: the expected path is then an owner-authorized admin-merge (`gh pr merge --admin`), run only on explicit say-so — a sanctioned owner action, not a silent bypass. Mirrored across all four packages + the global copy. Surfaced by dogfooding `/release` on its own v2.11.0 release. --- ## [2.11.0] - 2026-07-03 ### Added - **`/release` — end-to-end feature-delivery orchestrator, added across all four packages.** A thin orchestrator (it delegates, never re-implements): resolves a feature branch (arg → `git switch` to it; on `main` with no arg → creates `feat/` and carries the work over; otherwise the current branch), then runs `/ship` + `/security` + `/diff-review` under `/verify-done` discipline — reading each sibling command's real checklist so nothing is hand-waved. A hard gate splits the safe/read-only verify half from the irreversible release half (docs → version bump → commit → push → PR → merge → tag), which confirms each step. It **stops before `npm publish`** — `publish.yml` is manual `workflow_dispatch` by design. Counts: claude commands 8→9; droid/opencode/ampcode 17→18. ### Fixed - **Doc count drift corrected while adding `/release`.** `packages/ampcode/AGENT.md` had mis-filed `live-canvas` (a skill) under its Commands table, leaving Skills at 8 rows beneath a "9 total" header; moved it to Skills so both tables reconcile at 9. README's "Commands/Skills" section still listed a standalone `/friction` bullet (collapsed into `/remember` back in 2.9.0) — removed, and the section total corrected to match "9 skills + 9 commands". --- ## [2.10.0] - 2026-06-16 ### Changed - **`/friction` collapsed into `/remember`; the hot-memory pipeline is now two commands (`/stash → /remember`), down from three.** `/remember` now runs `friction.js` itself as a best-effort first step before consolidating — so the antigen data is always fresh and friction can't be forgotten. Friction targets the tool's **global sessions root** (all projects, since behavioral patterns are cross-project), resolved from an editable, never-prompt probe list baked into `remember.md` (Claude Code, Droid/Factory, Amp, opencode, plus Codex and Antigravity roots; add your own at the top). A no-sessions miss is surfaced **loudly** and degrades to stash-only — never a silent skip. The standalone `/friction` command was removed across all four packages (the `friction.js` script stays, directly runnable for inspection). Counts: claude commands 9→8, droid/opencode/ampcode 17 each. - **`/stash` now nudges toward consolidation.** After saving, it derives the unprocessed backlog (`stash files − .processed manifest entries`) and, at ≥5, emits a one-line prompt to run `/remember`. No counter is stored — the count is derived from ground truth, and running `/remember` clears it. The nudge is informational; `/remember` never runs automatically. ### Fixed - **`friction.js` no longer crashes on a single malformed JSONL line.** The four mirrored copies 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 a try/catch that reports the file and bails with exit 1 instead of throwing. Mirrored identically (modulo `.claude`/`.factory`/`.opencode`/`.amp` branding) across all four tool packages. - **`packages/subagentic-manual.md` restored after a range-sed corrupted ~310 lines.** A `sed '/start/,/end/{s/.../...}'` earlier this session had `test-generate$` as the end pattern; the range matched far past its intended scope (later occurrences in tree diagrams and category bullets), overwriting the entire tail of the document — Subagents reference, Commands reference, Hot Memory, Usage Patterns, Platform Architecture, Frontmatter Architecture, Contributing — with ~310 duplicate copies of the "Simple Commands" bullet. Restored from the pre-corruption snapshot and re-applied all the renames + count updates that should have happened cleanly. Same fix mirrored to `agentic-toolkit/ai/subagentic/subagentic-manual.md`. ### Changed - **Root `package.json` `engines.node` bumped `>=14.0.0` → `>=18.0.0`.** Node 14 has been EOL since 2023-04; the floor now matches the bundled `live-canvas-channel` plugin (`>=18`) and sits well under what CI publishes on (Node 22). As a minimum it excludes no one currently on a supported runtime. - **`/review` renamed to `/diff-review` across all four tool packages.** Avoids the name collision with the Anthropic-official `code-review` plugin (which also ships a skill named `review` that operates on PRs). `/diff-review` is more accurate to what the command does — it operates on a diff (staged, working tree, branch range, or against a ref), not on a remote PR. Mirrored into `~/.claude/commands/` and all docs/agents/`opencode.jsonc` references swept. - **`/diff-review` absorbed `/code-review`; collapsed to a single command across all four tool packages.** `/diff-review` now accepts a file, a branch (`/diff-review main` diffs `merge-base(main, HEAD)..HEAD` — the common "review my branch before merging" path), or an explicit range (`main..HEAD`). It bakes in the user's standing review focus: bugs needing a fix, dead code, loose ends (added TODO/FIXME, swallowed errors, stubs, abandoned flags), correctness, security, performance, maintainability. - **`/diff-review` and `/security` now verify findings, selectively auto-fix, and stop to ask only when needed.** After listing findings, each cited `file:line` is re-grounded in context (and `git grep`'d for dead-code claims) and marked confirmed / false-positive / uncertain. Confirmed + unambiguous + no-contract-change fixes apply directly; the changed region is re-read after the edit. HITL gates fire only for: uncertain findings, multiple reasonable fix shapes, downstream-affecting changes (signatures / response shape / schema / public symbol removal), security primitives (auth / crypto / session / token), or "dead code" that looks intentionally kept. `/diff-review` ends with a one-line **Ready to merge? Yes / No / With fixes** verdict. - **Skills renamed to short 2-word slugs (round 2).** `testing-anti-patterns` → **`test-traps`** (now includes timing/polling as AP6 after the fold). `test-driven-development` → **`tdd-flow`** (slug short, H1 short, `TDD` prose preserved as the industry term). `verification-before-completion` → **`verify-done`** (rhythmically mirrors `test-first`). Round 2 paired with round 1 (`systematic-debugging` → `debug-method`, `root-cause-tracing` → `trace-back`) gives a scannable cluster: `tdd-flow / test-generate / test-traps` and `debug-method / trace-back / verify-done`. All references swept across docs, agent files, opencode.jsonc, and the debug-method skill's cross-refs. - **Skills renamed to short 2-word slugs (round 1).** `systematic-debugging` → **`debug-method`** (the 4-phase framework with its 4 pressure-test scenarios + creation log preserved). `root-cause-tracing` → **`trace-back`** (the backward-tracing technique with its `find-polluter.sh` bisection helper preserved). Names are shorter, cluster alphabetically under `debug-`, and the "method vs technique" split is now obvious at a glance. All references swept across docs, agent files, and the debug-method skill's internal cross-refs to trace-back. - **`/test-generate` rewritten as a generate-and-verify loop, not just a generator.** New flow: discover the existing test framework (refuses to add a new runner) → mirror nearby tests for style/fixtures → generate happy / edge / error cases → **run the new tests** with the project's real test command → **verify each test bites** (mentally swap a broken impl — does the assertion catch it?). Superficial tests (`expect(true).toBe(true)`, mock-asserting-itself, setup-masked passes) count as a failure to ship. Same claim → verify → report shape as `/diff-review` and `/security`. HITL gates: a meaningful test would require a non-obvious design change in production code, ambiguous existing test patterns, or a mock style the project doesn't currently use. - **`/optimize` now verifies bottleneck claims before optimizing.** Each cited `file:line` must have at least one of: a profile / benchmark / log line showing call frequency or duration, an obvious hot loop / per-request handler, or user-provided evidence. Unverified claims are marked **uncertain — don't optimize on speculation**. Auto-fixes only when confirmed + unambiguous + no behavior/API change. HITL gates: uncertain (no profile), multiple reasonable shapes (cache vs precompute vs batch vs paginate vs index), public-API / response / schema changes, correctness-for-speed trades, or concurrency primitives. - **`/refactor` now runs the tests after the edit.** The "existing tests must pass" constraint was load-bearing but unverified — `/refactor` now detects the project's test command (`package.json` scripts, `pytest`, `go test`, `cargo test`, `Makefile`), runs it scoped to the affected area when possible, and reports pass/fail. If tests fail it **stops and asks** with three options (revert / patch the refactor / update the test with reasoning) rather than auto-reverting (destroys work) or pushing forward (breaks the invariant). Also stops on scope creep and public-API-boundary changes. - **`condition-based-waiting` folded into `test-traps` as Anti-Pattern 6: Timeout-Based Waiting.** The two skills covered the same domain (test quality) but only one auto-triggered; folding promotes the timing/polling guidance to auto-trigger coverage. The `example.ts` helper (domain-specific `waitForEvent` / `waitForEventCount` / `waitForEventMatch`) moves with it and is referenced from AP6. Counts: claude skills 10 → 9; droid/opencode/ampcode commands 19 → 18. - **CI:** the publish workflow now polls the npm registry for ~2 min (was ~15s; `--prefer-online` skips npm's view cache) and accepts an `exit 0` publish even if the registry hasn't reflected it yet, so a successful-but-slow-to-reflect publish no longer reports a false failure. - **`publish.yml` is now manual-only (`workflow_dispatch`) — npm OIDC trusted publishing with provenance, idempotent, and verifies the registry end-state.** - **`publish.yml` install step `npm ci` → `npm install`.** This toolkit is dependency-free (no `package-lock.json`), so `npm ci` failed with `EUSAGE`; `npm install` is a fast no-op here and still works if deps are ever added. Removed the superseded manual `scripts/publish.sh` (NPM_TOKEN-via-`pass` flow) — publishing now goes solely through the `publish.yml` GitHub Actions workflow. ### Removed - `/code-review` (was: workflow ceremony about *when* to request a review, mostly overlapping `/diff-review`'s purpose). Use `/diff-review` instead — `/diff-review main` for branch-vs-main, `/diff-review` with no args for staged/working-tree. - **`/debug`** — was a thin 17-line echo of the `systematic-debugging` skill. The skill (now `debug-method`) carries the real workflow with its pressure-test scenarios; the command added nothing. - **`/explain`** — was 11 lines of "explain this code" with no real constraints or workflow. The model does this naturally from a plain prompt. - **`/git-commit`** — Claude Code has built-in commit handling and the other three tools don't need a thin wrapper around `git diff --staged` + a templated message either. Use natural-language prompts instead. --- ## [2.9.0] - 2026-05-26 Redesign of the `/friction` → `/remember` memory pipeline so friction stops poisoning hot memory and antigens come from what the user actually said. Applied identically across all four tool packages (claude, opencode, ampcode, droid). ### Changed - **`/friction` now seeds antigens from observed user reactions, not machine proxies.** Antigen candidates are anchored only on real user reactions (corrections, curses, interrupts); inferred signals survive only as corroborating severity. Clustering is by what the user *said* (content/phrase overlap) instead of `(signal, tool_pattern)`, and recurrence × severity drives a `suggested_artifact` — only patterns recurring across 5+ sessions are meant to load into hot memory. On a 253-session corpus this took false hot preferences from 15 → 0. - **`/remember` rewritten to consolidate from friction's short quotes, never raw logs.** It classifies each reaction's target (agent vs. self), drops self-corrections, semantically merges paraphrases that lexical clustering left split, and tiers antigens by recurrence. The generated `MEMORY.md` section is renamed `Preferences` → `Antigens` (High loads hot / Medium recorded / Low = episode). ### Fixed - **Terminal pastes are no longer mistaken for friction.** Pasted SSH/shell dumps (prompt lines like `> sudo …` and `root@host:~#`, command output) were captured as user reactions, polluting antigen keywords with shell vocabulary (`postconf`, `qemu`). They are now detected and excluded from both signal detection and keyword extraction, while genuine short corrections that merely mention a command are preserved. - **Profanity only counts when it's aimed at the agent.** Narrative/rhetorical curses ("does anyone search any shit?", a pasted reddit story) no longer raise a `user_curse` signal; a curse is kept only in a short reaction turn or when an agent-directed word sits next to it. - **Self-corrections are now surfaced to the consolidation step.** The `self_suspect` hint ("wrong project", "nevermind") is propagated from candidate to cluster and rendered in `antigen_review.md`, so `/remember` is told to confirm agent-vs-self target before treating a cluster as an antigen. ### Removed - Dead scaffolding left over from the redesign in `friction.js`: the unused `overlap()` helper, the `MIN_KW`/`MIN_INTER` constants, the unread `selfCount` counter (superseded by the `anySelf` flag), and the always-empty `top_files` field with its unreachable renderer block. --- ## [2.8.3] - 2026-05-24 ### Security - **Path containment uses a boundary match.** `installer/path-manager.js` confined installs to the home directory via `startsWith(homeDir)`, which would also accept a sibling like `${homeDir}-evil`. Now matches on a path separator (`=== homeDir || startsWith(homeDir + path.sep)`), and likewise for the temp dir. Defense-in-depth for a local installer; verified install/backup/uninstall still work. ### Changed - **`docs/INSTALLER_GUIDE.md` custom-path docs now match the installer.** It advertised `/opt/...`, `/mnt/...`, and team/external locations, but the installer confines paths to the home directory (or temp dir). Documented the real rule and corrected the examples. - Removed decorative emoji from `README.md` headings (kept the friction traffic-light indicators). - CI: bumped `actions/checkout` and `actions/setup-node` to v6 (Node 24) ahead of the June 2026 Node 20 deprecation. --- ## [2.8.2] - 2026-05-24 Maintenance release: removes the last of the legacy 3-variant system, fixes the broken `liteag` alias and post-install message, and trims the shipped documentation down to the README plus an accurate installer guide. No changes to the installed agents/commands/skills. ### Fixed - **`npm test` (the CI publish gate) was failing, blocking releases.** The installer test suites still assumed the removed 3-variant system (Lite/Standard/Pro). The multi-tool suite has been rewritten for the single-variant installer (one `pro` package per tool), and the cross-platform suite's terminal checks no longer assert raw environment presence (`stdout` TTY, `TERM`, `SHELL`) — those failed whenever output is piped (i.e. always under the runner and in CI). They now verify the installer's graceful fallback instead. `npm test` passes 138/138, including under a minimal CI environment. - **The `liteag` short alias was broken.** Its `cli.js` was a leftover 3-variant wrapper that defaulted to a non-existent `standard` variant, looked for a `.claude-plugin/plugin-standard.json` that no longer exists, and exited 1. It now simply forwards all arguments to the real interactive installer (`installer/cli.js`), so `liteag` and `liteagents` behave identically (check existing installs, backup, install, uninstall). The actual installer was not modified. - **`postinstall` pointed at a command that doesn't exist.** The post-install message told users to run `$ agentic-kit`; the published bin is `liteagents`. Corrected. ### Added - **Content-integrity check in the multi-tool test suite.** It now pins the expected per-tool counts of agents, commands, skills, and plugins (counting `.md` dispatch entries and skill/plugin directories, not raw files). Accidentally adding or removing a command/skill/agent fails the publish gate with the exact delta until the expected number is updated deliberately — so the CI gate now protects what actually ships, not just installer plumbing. ### Changed - **Rewrote `docs/INSTALLER_GUIDE.md` to match the actual installer** (1586 → ~430 lines). Removed the entire fictional "Command-Line Flags" section (the installer takes no flags — it is interactive only), the Lite/Standard/Pro variant system, invented agent lists and component counts, and corrected the default install paths (`~/.config/opencode`, `~/.config/amp`, `~/.factory` — the old guide listed `~/.opencode`/`~/.ampcode`/`~/.droid`, none of which the installer uses). Added an accurate interactive walkthrough and an Uninstalling section, and dropped links to deleted/nonexistent docs. ### Removed - **Trimmed `docs/` to one guide.** Removed 20 internal/maintainer/stale docs (~7,500 lines) that shipped to users — implementation notes, QA/test reports, `pass`/publishing guides, the variant-era `MIGRATION.md`/`RELEASE_NOTES_1.2.0.md`, `KNOWLEDGE_BASE.md`, `CONTRIBUTING.md`, `PRIVACY.md` (telemetry is a no-op stub), `SECURITY.md`, etc. The README plus the rewritten `docs/INSTALLER_GUIDE.md` are the documentation now. - **Stopped shipping `docs/` in the npm package** (dropped from `package.json` `files`, along with the broken root `QUICK-START.md`/`TROUBLESHOOTING.md` entries that never existed). End users get the README; the installer guide lives in the repo. Fixed the README's broken doc links and pointed `CLAUDE.md` and the package validator at the surviving guide. - Stale `variant-system` npm keyword — there is one package per tool, not a variant matrix. - Dead 3-variant migration code in `installer/path-manager.js` (`detectLegacyInstallation`, `countLegacyComponents`, `classifyVariantFromComponents`, `createManifestForLegacy`). These classified pre-1.2.0 installs into lite/standard/pro by counting `resources`/`hooks` dirs that no longer exist, and nothing in the installer ever called them. Verified the installer still installs, backs up, and uninstalls correctly after removal; `npm test` 138/138 and the installation-engine suite 60/60 still pass. - Obsolete documentation describing removed features: `docs/VARIANT_CONFIGURATION.md`, `docs/UPDATED_VARIANT_CONFIGURATION.md` (the 3-variant matrix), `docs/SILENT_MODE_GUIDE.md` and `docs/INSTALLATION_DEMO.md` (a `--variant`/`--tools`/`--silent` flag CLI the installer never had — it is interactive only). Dropped the stale `VARIANT_CONFIGURATION.md` entry from `scripts/validate-package.js`. --- ## [2.8.1] - 2026-05-22 ### Changed - **`/security` and `/ship` rewritten across all four tools (claude, opencode, ampcode, droid).** Both were thin stubs; they're now substantive, stack-agnostic gates that apply to libraries, CLIs, web apps, and services alike. - `/security` leads with the six failure classes that recur in nearly every quickly-built app — secrets committed to the repo, data-access / tenant isolation, rate limiting (including authenticated write routes), error handling past the happy path, authorization-beyond-authentication (IDOR / privilege), and N+1 / unindexed data access — plus a trust-boundary pass (spoofable headers like `X-Forwarded-For`, services bound to `0.0.0.0`, unvalidated untrusted input) and severity-ranked, coverage-auditable output. - `/ship` is now stack-adaptive: it detects the toolchain (`package.json`, `pyproject.toml`, `go.mod`, `Cargo.toml`, `Makefile`) and runs only the checks that exist instead of assuming `npm run lint`/`build`/`migrate`, and adds gates for authorization, rate limiting, data-access scoping, error handling, and secret-scanning before deploy. ### Fixed - **`allowed-tools` permission syntax normalized to the canonical colon form** (`Bash(git:*)`) in the claude package's `ship.md`, `security.md`, and `git-commit.md` — the space form (`Bash(git *)`) is not a valid Claude Code permission wildcard. The opencode/ampcode/droid packages retain their existing space-form syntax (their runners parse `allowed-tools` differently, if at all), so there is no behavior change there. --- ## [2.8.0] - 2026-05-18 ### Added - **live-canvas channel: lazy port binding via MCP tools** — `server.js` (v0.5.0) now exposes `channel_open`, `channel_close`, and `batch_open` tools and only binds port 8788 when the skill explicitly calls one. Plain Claude sessions stay idle by default; multiple sessions can have the plugin loaded with `/mcp` green without racing for the port. - **live-canvas channel: automatic sibling takeover** — when `channel_open` finds port 8788 held by another instance of the live-canvas plugin running as the same uid, it takes over (SIGTERM the sibling, rebind, SIGKILL fallback if needed). The taken-over pid is returned as `took_over` in the response so the skill can surface it to the user. Foreign processes are still refused with `{status: "in_use", holder_pid, ...}` — the plugin won't kill anything it doesn't own. Removes the dead-end "port busy, go close it yourself in some other terminal" prompt the user was hitting on every second `/live-canvas`. - **live-canvas JSON mode: writes to disk instead of browser download** — channel server gains a `POST /feedback-jsonl` route that appends submissions to `/.claude-design/feedback.jsonl`. The skill calls a new `batch_open` MCP tool (no flag gate — JSON mode doesn't use channels) and sets the overlay's `batchEndpoint` to `/feedback-jsonl`. Falls back to the legacy browser-download path only when the MCP isn't available or another session holds the port. - **live-canvas channel: parent-flag capability gate** — `channel_open` inspects the parent `claude` process's command line and refuses to bind if `--dangerously-load-development-channels` is missing, returning `{status: "no_channel_capability", message: ...}`. Without this, plain `claude` sessions could win the port race and silently drop every notification (POST 200, no `` tag — the "nothing landed" black hole). - **Cross-platform parent-cmdline detection** — Linux reads `/proc//cmdline` (fast, no subprocess); macOS/BSD falls back to `ps -p -o args=`; Windows falls back to `wmic process where processid= get commandline`. If none work the gate fails closed. - **SKILL.md Case D — explicit relaunch block** — when `channel_open` returns `no_channel_capability`, the skill prints the exact `live-claude --continue` command (and the literal `--dangerously-load-development-channels` long form) and stops, instead of proceeding into a non-functional Live mode. ### Changed - **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`) is authoritative — no more curl-vs-marketplace-dir branch table. Mirrors in `packages/{droid,ampcode,opencode}/commands/live-canvas.md` synced. ### Fixed - **live-canvas: silent channel black-hole when a plain `claude` won the port race.** Before, the first session to start (often a plain `claude` without the experimental channels flag) would bind 8788 first; subsequent `live-claude` sessions hit EADDRINUSE and the user's browser feedback would POST 200 into a session that silently discarded notifications. The capability gate + lazy bind together eliminate this: only flagged sessions can claim the port. --- ## [2.7.0] - 2026-05-17 ### Added - **live-canvas: one-shot installer** — `setup.sh` now copies the marketplace to `~/.claude/plugins/`, runs `npm install`, and writes a `live-claude` shell function to `~/.zshrc` and `~/.bashrc` so the user can launch a Live-mode session with one command. Idempotent. - **live-canvas: collapsible overlay** — a "−" button next to "Add Feedback" hides the bar to a 36px corner bubble (sessionStorage-persisted). Mobile-friendly: comment popup goes full-width below 640px. - **live-canvas: lab banner** — generated lab pages now include a "this is a temporary review surface" banner template (`templates/lab-banner.html`), mode-agnostic, paste-once. - **live-canvas: explicit mode pick** — the skill asks Live vs JSON every run via `AskUserQuestion` instead of silently auto-detecting. If Live is unreachable, the skill diagnoses (installed-but-not-Live vs first-time setup) and prints targeted next steps instead of failing opaquely. - **CLAUDE.md inline dev rules** — must-know rules from `.claude/memory/AGENT_RULES.md` (Simple > clever, surgical changes, dependency hierarchy, mobile-first UI, POC-first) are now inlined in CLAUDE.md so every agent session sees them. - **Installer banner reads `package.json`** — the ASCII logo's version string is no longer hardcoded; `UPDATE_VERSION.sh` only needs to touch `package.json` to keep it current. README version badge already auto-pulled from `package.json` via shields.io. ### Changed - **live-canvas: vanilla overlay everywhere** — deleted the React-specific overlay (`templates/feedback-react/`, 5 files, ~2300 lines). `overlay-vanilla.js` (one file, plain DOM, zero deps) now works in every supported framework, including React/Next.js/Vite via a `