# Changelog

> **Language / Dil:** **English** · [Turkce](CHANGELOG.tr.md)

This project follows the [Keep a Changelog](https://keepachangelog.com/en/1.0.0/) format and [Semantic Versioning](https://semver.org/).

## [Unreleased]

## [1.36.0] - 2026-07-05

> Minor release. A 2026 currency pass across the whole advisory surface — the ads,
> market, SEO, ASO, and content commands/agents are re-grounded in live-verified
> 2026 platform mechanics and algorithms — plus a real cost-reporting bug fix. No
> new commands/agents/skills (surface unchanged: 30/86/63); the freeze held (all
> maintenance/currency). Every claim was researched live against current provider
> docs, folded in as verify-live checks with no baked numbers, then adversarially
> re-verified (6 corrections applied — incl. dropping the "ad_storage sole gate"
> overstatement and an unverifiable Meta webhook claim).

### Changed — advisory currency (2026)

- **`/meta-review`, `/ads-review`, `ads-strategist`** — grounded in live-verified
  Meta/Google measurement mechanics (Pixel↔CAPI one-dataset dedup on
  `event_id`+`event_name`, `action_source` per channel, PII hash scope, CTWA +
  Lead Ads funnels) and 2026 delivery-algorithm reality (AI creative-reading
  retrieval + unified ranking → creative-is-targeting, Advantage+ default,
  cross-surface learning, incrementality over last-click). Platform changes: the
  one-click-CAPI second-pipeline dedup trap, Google Consent Mode decoupling from
  Google Signals, offline-import → Data Manager API, Enhanced Conversions unified
  toggle, AI Max Final-URL 404 risk, Call Ads retired, PMax campaign-level
  negatives.
- **`/market`, `market-researcher`** — 2026 signal reliability: search volume is a
  floor, not a market size (zero-click + AI-Mode query fan-out); a multi-signal
  stack (TikTok trend velocity, community pain-points, marketplace SQP); AI answer
  engines as a discovery surface; AI-citation presence ≠ SERP rank.
- **`/seo`, `seo-strategist`** — aligned to Google's official 2026 AI-search
  guidance: GEO = SEO (same index via RAG + query fan-out); `llms.txt` is ignored
  (removed the prior recommendation); FAQ rich results deprecated; Preferred
  Sources; the Search Console Gen-AI impression report; `Google-Extended` ≠ AI
  Overviews.
- **`/aso`** — 2026 App Store algorithm reality: Apple LLM semantic-relevance
  ranking, Apple Search Ads' second in-organic slot, App Intents/Spotlight
  discovery, indexed screenshot captions, Custom Product Pages, and Google Play
  "Ask Play".
- **`content-creator`, `visual-director`, `/content-generate`** — 2026 platform
  reality: originality enforced (reposts/third-party watermarks de-recommended),
  AI-assist is fine but mass-produced "slop" is penalized, realistic-AI-media
  disclosure (C2PA/SynthID; EU AI Act Art. 50), engineer sends-per-reach +
  watch-through, interest-graph through-lines for LinkedIn/X/Threads.

### Fixed

- **`badi stats` cost reporting** — the transcript pricing table charged Opus at
  the retired Opus-3 rate (`$15/$75`); corrected to the current `$5/$25` (was
  over-costing every Opus session ~3×). Added `claude-fable-5` ($10/$50),
  `claude-opus-4-8` ($5/$25), and `claude-sonnet-5` ($3/$15), and taught the
  model-family classifier the `fable` family (Fable 5 / Mythos 5).

### Maintenance

- Dependency bumps: `node-html-parser` 7 → 8.0.4 (behaviour-verified),
  `actions/checkout` 6 → 7, `@biomejs/biome` 2.5.1 (+ `biome.json` schema).
- README flagship line refreshed to Opus 4.8 / Sonnet 5 / Fable 5. Agent `model:`
  aliases (`sonnet`/`haiku`) auto-track to Sonnet 5 / Haiku 4.5 — left unchanged.

## [1.35.0] - 2026-06-20

> Minor release. Internal hardening (branch-guard, semver, doctor/release gates,
> command-file splits, deterministic tests) + two CEO-approved additions
> (`/market`, `/tasks`). Also rolls up v1.34.2 (the subdirectory hook-path fix,
> tagged but never npm-published) — npm users move 1.34.1 → 1.35.0.

### Fixed — post-review hardening (multi-lens review of the v1.35.0 diff)

A 6-lens review (code/security/eng/ceo/audit) with adversarial verification of
`v1.34.1..HEAD` caught real defects — most in this cycle's own new code:

- **branch-guard `<<<` here-string bypass (HIGH).** `stripHeredocs` matched the
  bash here-string `<<<` as a heredoc opener and swallowed the rest of the
  command, so `cat <<< x && git commit` on a protected branch was NOT blocked.
  The heredoc regex now rejects a 3rd `<` (lookbehind/ahead).
- **branch-guard quoted branch names (MEDIUM).** `git switch -c "feat space"`
  was not tracked (the capture stopped at the space), wrongly blocking a
  legitimate commit. The switch/checkout matcher now handles quoted names and
  the capital force-create forms `-C`/`-B`.
- **`npm run lint` was red** — `biome check .` flagged 3 cycle-authored test
  files that `biome check` (no path) skipped. Formatted; the gate is green.
- **`checkScoopManifest`** now matches the version as a delimited tarball token
  (`badi-${v}.tgz`) instead of a bare substring (avoids `badi-11.3.2` ≅ `1.3.2`).
- **Stale runtime/doc counts** corrected: `badi skills` help (62→63), README
  "command modules" (36→38) + "workflow commands" (84→86), scoop description
  (84/62→86/63), `badi mobile` unknown-command usage (now lists all 8 subs),
  `command-profiles.js` `core` comment (20→21), `.claude/memory.md` trimmed back
  under 100 lines (cleared the doctor warning).
- New regression tests: `<<<` non-heredoc, quoted/space + `-C`/`-B` branch names.

### Fixed — `badi --help` subcommand lists were stale

- Audited every CLI command, slash command, and skill against its registry: all
  86 commands (COMMAND_PROFILES / command-index / vault) and 63 skills (INDEX vs
  disk) are consistent — nothing orphaned. The drift was only in the `badi --help`
  subcommand summaries, now corrected: `market` (+ `gaps`/`wishlist`), `mobile`
  (+ `version`/`assets`), `aso` (+ `metadata`/`compete`/`search`), `wp`
  (`plugin`/`theme` → `plugins`/`themes`, + `update`).

### Added — `/tasks` dependency-aware sequencing command (freeze exception: engine)

- **New `/tasks` slash command** — turns a plan/spec into an ordered, parallel-
  executable `tasks.md` (numbered `T001…`, `[P]` parallel-safe markers, explicit
  file paths, phases) and runs it: `[P]` tasks fan out concurrently (mapping onto
  the Workflow engine's `parallel()`), dependent tasks pipeline. The execution
  layer between `/eng-review` (the plan) and the work. Commands 85 → 86.
- This is the single piece distilled from github/spec-kit after a `/ceo-review`
  pass — the rest (constitution agent, `/clarify`, `/specify`, `/plan`) was KILLED
  as overlapping `/architect` / `/brief` / `/spec-check` / `/team`. Approved as an
  **engine exception** to the freeze; logged in `.claude/workspace/freeze-exceptions.md`.

### Added — `/market` command + `market-research` skill (freeze exception: wiring)

- **New `/market` slash command** — niche discovery, opportunity sizing, and
  competitor/gap analysis via the existing `market-researcher` agent, with
  `badi market` for App Store data. Wraps capability that already shipped (the
  agent v1.34, the CLI v1.15); commands 84 → 85.
- **New `market-research` vault skill** (general family) — the methodology the
  command/agent follow (demand validation, niche discovery, opportunity sizing,
  competitor/gap, positioning/pricing, go/shrink/pass). Skill categories 62 → 63
  (general 25 → 26).
- Approved through `/ceo-review` as a **wiring exception** to the feature freeze
  (completes existing market infra, no new capability); logged in
  `.claude/workspace/freeze-exceptions.md`.

### Changed — seo command split (v1.35.0 hardening, PR-F)

- **`lib/commands/seo.js` (1060 lines) split into a `seo/` directory-module** —
  shared HTML/fetch/DDG helpers in `_shared.js` + one file per subcommand
  (audit/meta/sitemap/speed/backlinks/rank/compare) + `index.js` dispatch +
  `help.js`. CLI grammar and all output unchanged (verbatim move); `parseDDGResults`
  re-exported from `index.js`. help-doctor audits it as a directory-module.
- **NOT done — `stripTags` → node-html-parser** (the original P5 "no regex-HTML"
  item): empirically a behavior **regression**, so deliberately skipped.
  `node-html-parser`'s `.text` inserts a newline for `<br>`/block elements
  (`<h1>a<br>b</h1>` → `"a\nb"` vs the regex's `"ab"`) and does not replicate the
  fixpoint loop's nested-injection collapse (`<scr<script>ipt>` → `"<script>…"`
  vs `"ipt>…"`). The security concern was already resolved by the fixpoint loop
  (audit reclassified it to style-only), so the regex stays — it is the correct
  tool here.

### Changed — mobile command split + help-doctor coverage (v1.35.0 hardening, PR-E)

- **`lib/commands/mobile.js` (1220 lines) split into a `mobile/` directory-module**
  following the existing `icerik/` pattern: one file per subcommand
  (init/version/build/release/assets/crash/deeplink/ota) + `index.js` dispatch +
  `help.js`. CLI grammar and all output strings are unchanged (verbatim move).
- **`help-doctor` now audits directory-modules** (a dir with `index.js`), combining
  the `index.js` dispatch with the `help.js` help text. This restores help-drift
  coverage that was silently dropped when `icerik/` and `plugin/` were split, and
  covers the new `mobile/` — 35 → 38 audited commands, all drift-free. New exported
  `detectDriftSource(source, opts)`.

### Fixed — deterministic stats tests (v1.35.0 hardening, PR-D)

- **The `stats v1.29+` tests no longer flake (timeout) under load.** They ran the
  real CLI against the machine's actual `~/.claude/projects`, which on a dev box
  is large enough to exceed the 15s timeout when the box is busy. `transcriptsRoot`
  gained a `BADI_TRANSCRIPTS_ROOT` env seam (precedence: override arg > env >
  default); the test now points it at an empty fixture dir, so the scan is instant
  and deterministic regardless of load. Run time dropped from ~15-30s to <0.5s.

### Fixed — branch-guard cwd/cd awareness (v1.35.0 hardening, PR-C)

- **The `branch-guard` hook no longer false-positives on legitimate commits and
  no longer checks the wrong repo.** It used to (a) flat-regex the raw command
  against the process branch evaluated once — so a compound command that does
  `git switch -c feature/x && git commit` was blocked because the branch was
  still `main` at evaluation time; (b) trip on a `git commit` string inside a
  heredoc body; and (c) check the hook's own repo even when the command targets
  another via `cd`/`git -C`. Now it resolves the target directory, reads that
  repo's branch, strips heredoc bodies, and walks command segments so a
  `switch`/`checkout -b` before the commit takes effect. New exported pure
  helpers in `_util.mjs`: `branchOf`, `stripHeredocs`, `splitSegments`,
  `resolveTargetDir`, `effectiveBranchForCommit`, `isGitCommit`. Fail-closed on
  parse/path ambiguity (evaluates the base branch).

### Changed — internal refactor (v1.35.0 hardening, PR-B)

- **Consolidated semver helpers into `lib/helpers.js`.** `parseVersion` (was in
  `lib/data/plugin-manifest.js`), the byte-identical `bumpVersion` duplicated in
  `release.js` + `mobile.js`, and the ad-hoc `semverGt` split in `update-check.js`
  now share one strict 3-part implementation (re-exported from the old locations
  so importers are unaffected). `bumpVersion` now throws on an unparseable
  version instead of silently producing `1.2.NaN`; `semverGt` tolerates
  prerelease tails and is conservative on bad input.

### Changed — doctor & release-gate robustness (v1.35.0 hardening, PR-A)

- **`doctor` derives the expected hooks from `settings.json`** instead of a
  hardcoded list — adding/removing a hook no longer needs a hand-edit here (the
  52→53 hand-bump pain). New exported pure helper `hookFilesFromSettings`.
- **`release check` now verifies the README suite count too** (the "N tests
  across M suites" line), not just the test count — `parseTestSummary` parses
  `suites`, fed forward as `ctx.actualSuites`.
- **`release check --skip-test` now WARNs** that the docs-sync reality check was
  skipped (README counts verified for internal consistency only, not against the
  live suite) instead of reporting a clean reality-verified pass.
- **New vault guard test:** `INDEX.md`'s advertised category counts must match
  the categories on disk (catches the 25-vs-21 drift class).

### Fixed — release tooling

- `dist/scoop/badi.json` carried a stale download `url` (`badi-1.34.1.tgz`) after the
  v1.34.2 version bump — the file claimed two versions at once. Corrected to 1.34.2, and
  a new `release check` gate (`checkScoopManifest`) now asserts the scoop `version` and
  `url` agree so this drift can't recur silently.

### Changed — dev tooling

- `@biomejs/biome` 2.4.16 → 2.5.0 (dev-dependency). Config migrated to the 2.5.0
  schema; adapted to new lint rules: `release.js` `findIndex(x => x === v)` →
  `indexOf(v)` (`useIndexOf`), and `assets/` (static OG/social images) excluded
  from the linter so SVG assets aren't lint-gated. Lint stays clean.

## [1.34.2] - 2026-06-20

### Fixed — hooks break when Claude is launched from a subdirectory

The generated `.claude/settings.json` registered all 14 hooks with **relative**
commands (`node .claude/hooks/X.mjs`). A relative path resolves against the
current working directory at hook-execution time, not the project root. So
launching Claude Code from any **subdirectory** of a badi-managed project (for
example a sub-package inside a monorepo) made every hook fail with
`Error: Cannot find module '.../<subdir>/.claude/hooks/X.mjs'` — the hook file
lives at the project root, which the relative path can't reach.

- **Anchor every hook to `$CLAUDE_PROJECT_DIR`** in the `settings.json` template
  (`node "$CLAUDE_PROJECT_DIR/.claude/hooks/X.mjs"`) so hooks resolve regardless
  of the launch directory. This matches what `statusline.js` and the
  plugin-variant manifest (`$CLAUDE_PLUGIN_ROOT`) already do.
- **`badi skills auto on`** now writes the skill-router hook with the same
  `$CLAUDE_PROJECT_DIR` anchoring.
- **New doctor check** — "hook commands are project-root anchored" warns when a
  `settings.json` still carries relative hook paths (catches installs created by
  an older badi before they break in a subdirectory).
- **Migration:** existing installs pick up the fix by running `badi update`.

Tests: 1269 → 1274 (+5: `findRelativeHookCommands` producer + the shipped
template is fully anchored). Doctor: 52 → 53 checks.

## [1.34.1] - 2026-06-14

### Fixed — self-review of the hygiene round (multi-agent code review of #275-#277)

A recall-mode multi-agent review of the same-day merge caught two regressions the
hygiene PRs themselves introduced, plus a design hole in the new gate:

- **The docs-sync gate didn't catch the drift it exists to stop.** It compared the
  three README test-count surfaces only against *each other*, never against the real
  suite. #275 set them to 1191; #276's +64 vault tests pushed main to 1260 — so the
  README was stale again and `checkDocsSync` still passed. Now `checkNpmTest` feeds the
  real pass count into `ctx.actualTests` and the gate fails on any surface that disagrees
  with reality, with a floor that fails when the suite ran but no surface parsed (a README
  reword can no longer pass while verifying nothing). README counts corrected to 1269.
  A second self-review pass then caught that the reality feed was itself **inert in
  production**: `runNpmTest` parsed only the TAP `# pass N` format while the runner emits
  the spec reporter's `ℹ pass N`, so the count was always null and the new block never
  ran. Extracted as the exported, unit-tested `parseTestSummary` that reads both formats.
- **A stray active skill shipped to npm — closed at the source.** `.claude/skills/expo-app-config/SKILL.md`
  (255 lines) — untracked in #275, re-staged by #276's `git add -A` from working-tree
  user-local state — rode `package.json` `files[]` to npm. Untracked again; a `.gitignore`
  rule (`.claude/skills/*/`) blocks the `git add -A` vector, and `files[]` was narrowed
  from the whole `.claude/skills/` dir to the two scaffold files (`.gitkeep`, `README.md`)
  so a locally-activated skill can't ride a local `npm publish` either.
- **SECURITY.md version check is now row-aware.** Was a bare `sec.includes("1.34.x")`
  substring that false-passed a version listed Unsupported/EOL and could prefix-collide;
  now the minor must appear in a row whose **adjacent** status cell reads active/supported
  (scanning any cell, plus a `✅` branch that a stray `\b` made dead, were both fixed).
- README.tr.md: stale "915 Onaylanmis Test (51 test)" feature row (badge already
  removed in #277) replaced with a pointer to the canonical English count.

### Fixed — docs credibility + vault validation (post-v1.34 hygiene round)

A three-lens project review (product / engineering / QA on fresh evidence) found the
code healthy but the marketing surfaces drifting one release behind, every release.

- README.md: test count unified to 1191 (badge said 1161, feature table 1161, dev
  section 1184 — three contradictory values on one page); harness table subagents
  27 → 30; modular-architecture row 22 → 36 command modules; directory-structure
  block rewritten to match the actual tree (14 hooks, commands-vault, skills-vault);
  Homebrew/Scoop install rows demoted to "Planned" (tap/bucket repos do not exist yet
  — the old rows gave Windows users a dead install command).
- SECURITY.md: supported versions 1.33.x → 1.34.x.
- dist/scoop/badi.json: 1.30.1-era description ("22 AI agents, 77 commands") → current
  30/84/62; version/url bumped to 1.34.0 (CI copies the description verbatim to the
  published bucket).
- .claude/skills-vault/INDEX.md: regenerated from disk — previously claimed 25
  categories and rendered 21; now lists all 62 (25 general + 25 pentest-* + 12 expo-*),
  making the pentest/expo families visible in their own index for the first time.
- New release gate **docs-sync** (`checkDocsSync` in `badi release check`): README
  test-count internal consistency + harness-table subagent count vs disk + SECURITY.md
  supported-version coverage. Root cause fix — v1.32/33/34 each shipped stale numbers
  because no gate compared docs against reality.
- Vault validation: `metadata.homepage` backfilled in all 37 pentest-*/expo-* SKILL.md
  files (37/62 categories failed `validateSkillFile`; the bundler silently auto-filled);
  new vault-walking test prevents a third silent recurrence. `badi doctor` now also
  verifies the 14th hook (`inject-active-plan.mjs` was missing from its checklist).
- memory.md consolidated (98 → under budget; older entries moved to memory-archive.md).
- README.tr.md: archived-snapshot banner added (TR + EN) pointing to the canonical
  English README; the stale static test badge removed (dynamic npm badges stay).
  Closes the last open thread of the English-only migration tracking issue (#207) —
  full TR re-sync only if demand materializes. CHANGELOG.tr.md remains actively
  maintained as the intentional Turkish variant.

## [1.34.0] - 2026-06-06

### Added — harness-compatible security artifact chain (security-check v1.2.0)

The `security-check` skill family now chains its phases through file contracts using
the artifact names of Anthropic's [defending-code-reference-harness](https://github.com/anthropics/defending-code-reference-harness)
(Apache-2.0): `THREAT_MODEL.md → VULN-FINDINGS.json/.md → TRIAGE.json/.md`.

- `sc-orchestrator` (v1.1.0): Phase 2 now closes by emitting `VULN-FINDINGS.json/.md`
  at the project root — upstream-compatible field names and shape (`F-NNN` ids,
  uppercase `HIGH|MEDIUM|LOW` severity); `confidence`/`confidence_reason` are `null`
  at the producer stage by design (upstream fills `confidence` as a 0.0-1.0 float via
  second-opinion scoring; badi defers all confidence to the verify stage on a 0-10 scale).
- `sc-verifier` (v1.1.0): new Step 10 emits `TRIAGE.json/.md` (verdict mapping
  `TRUE_POSITIVE`/`FALSE_POSITIVE`/`CANNOT_VERIFY`, `verify_verdict`, dedupe links,
  0-10 confidence).
- `pentest-threat-model`: new file-output convention — `THREAT_MODEL.md` at the
  project root with numbered sections; sections 3-4 are the scoping contract the
  pipeline reads.
- README (EN + TR): "Harness-Compatible Artifact Chain" usage section with quickstart,
  one-way-interop framing, and a re-activation note for users who enabled the skills
  before v1.34 (`badi skills add` never overwrites an active skill).
- `.gitignore`: generated `VULN-FINDINGS.*` / `TRIAGE.*` ignored (`THREAT_MODEL.md`
  stays committed — it is a durable design document).
- README.tr.md headline counts reconciled (22/77/13 → 30/84/14); a full TR parity
  pass remains tracked separately.
- Provenance: artifact-name/field contract adapted from the Apache-2.0 upstream;
  skill text and the sc-* verification methodology are independently authored (MIT).
  Deliberately deferred: an advisory `patch` stage (conflicts with the security
  family's no-write contract) and the upstream autonomous pipeline (gVisor + ASAN).
- New CLI subcommand **`badi security pipeline [--json]`** — read-only status of the
  artifact chain: per-stage exists/missing/stale (mtime-based: upstream newer than
  downstream = stale), suggested next step, and a triage-naming disambiguation note
  (`badi security triage` = deterministic severity filter; `sc-verifier` = agentic
  verify writing `TRIAGE.json`). Always exits 0 — informational, not a CI gate.
  Tests: 1185 → 1191.

## [1.33.2] - 2026-06-05

### Fixed — network transparency: the dependency-audit hook's registry call is now disclosed and opt-out-able

An independent `evaluate-repository` self-review (run against the awesome-claude-code maintainer's verbatim evaluation prompt) found that `dependency-audit.mjs` runs `npm/yarn/pnpm audit` on every SessionStart — the only automatic network call Badi makes — while the README "Network Usage" table omitted it and claimed "Nothing is sent in the background."

- README "Network Usage" table: new row for the SessionStart dependency-audit hook; lead sentence corrected to name the single automatic call.
- New **`BADI_NO_DEP_AUDIT=1`** environment opt-out — the hook exits before any cache write or registry call. Regression test added (the opt-out must exit before the cache dir is created).
- `SECURITY.md` refreshed: supported versions `1.3.x` → `1.33.x`, `12 Hooks` → `14 Hooks`, `48 Security Skills` → `62 opt-in skill categories` (incl. 25 advisory `pentest-*`); the dependency-audit line now notes the network call + opt-out.
- README dev section's stale `251 tests` (a v1.12-era figure) → `1184 tests across 219 suites` (same fix in README.tr.md; the v1.12.0 version-history row keeps its then-accurate 251).

### Changed — discoverability

- npm `keywords`: added `chatgpt`, `codex`, `openai`, `copilot`, `ai-cli`, `coding-assistant` (20 → 26) so the package surfaces in adjacent-ecosystem searches; all existing keywords kept. GitHub topics updated in the same pass.
- `badi --help`: fixed `events`/`security` printing on one line; added missing sections (Commands Profile, Schedule, Agent/Watcher, Transcript & Plan, GitHub & KB, AI/Dev subcommands); `commands` description now lists the `pentest` profile.
- `.claude/command-index.md`: header corrected to 4 profiles (pentest reserved); `badi commands` footer lists `pentest`.
- `badi doctor` (claude harness): hardcoded agent check list 21 → 30 — the 9 agents added since v1.18 (tasarim-kurator, the virtual eng team, ads-strategist, and the market/SEO/data advisory trio) are now verified.
- README Version History: added the missing `v1.19.0`, `v1.28.0`, `v1.29.0`, `v1.30.x`, `v1.31.0` rows; the `badi market wishlist` claim moved from the v1.20.0 row to its accurate v1.19.0 home.

Tests: 1184 → 1185.

## [1.33.1] - 2026-06-05

### Fixed — `badi skills auto on` registered a dead hook

The prompt-aware auto-router opt-in was broken: `badi skills auto on` wrote `bash .claude/hooks/skill-router.sh` into `settings.json`, but the hooks were migrated `.sh → .mjs` (Node) back in v1.22 — `skill-router.sh` doesn't exist and a `.mjs` file can't run under `bash`. Any user who enabled the auto-router got a hook that silently never ran. (The registration in `lib/commands/skills.js` was the one spot the v1.22 migration missed, and a test had pinned the buggy `.sh` string so it went unnoticed.)

- `badi skills auto on` now registers `node .claude/hooks/skill-router.mjs` (matcher `""`, `timeout: 5000`), matching the `inject-active-plan` UserPromptSubmit hook; on/off detection updated accordingly.
- Test hardened to assert the command is `node ...skill-router.mjs` and to reject any `bash` / `.sh` form (the regression guard that would have caught this).
- Context: 14 hook files ship; 13 are wired into the default `settings.json`, and `skill-router` is the opt-in 14th — so the documented "14 hooks" count is correct.

## [1.33.0] - 2026-06-05

The English-only migration goes all the way down — and is independently verified. v1.32 renamed the CLI grammar; **v1.33 translates every remaining body, comment, and string across 60+ surfaces and proves zero residue through 7 rounds of adversarial multi-agent audit.** Plus three new advisory agents bring the fleet to 30.

### Added — research / SEO / data advisory agents (atoms.dev gap-fill)

Three dedicated advisory subagents that wrap existing capability into focused, isolated contexts (the `security-scanner` / `performance-profiler` pattern). A gap analysis against atoms.dev's agent roster found these three roles had tooling in badi but no dedicated agent; the other five roles were already covered.

- `market-researcher` — demand/niche discovery, competitor + market signals, opportunity sizing *before* you build (fills the gap between `product-strategist` and `ads-strategist`). WebSearch/WebFetch + the App Store market tooling.
- `seo-strategist` — SEO audit + keyword/architecture strategy + organic-growth plan; owns the `/seo` command and the `seo` / `seo-crawl-budget` skills as one subagent.
- `data-analyst` — dataset analysis → growth insight; wraps the `data-analytics` skill (62 procedures).
- All three are READ_ONLY/advisory (`disallowedTools: [Write, Edit, NotebookEdit]`). Fleet: 27 → 30 agents.

### Added — ads-strategist agent + /meta-review + /ads-review (advisory paid-ads layer)

Project-aware paid advertising review, modeled on the /ceo-review pattern: badi knows the project (memory, code, target user) and turns that context plus live market research into a platform-specific ad strategy with a launch-readiness verdict (READY TO LAUNCH / FIX FIRST / DON'T ADVERTISE YET).

- New agent `ads-strategist` (READ_ONLY, advisory; first agent with WebSearch/WebFetch).
- `/meta-review` (Meta: audiences, CBO/ABO funnel, creative angles, policy risk) and `/ads-review` (Google Ads: keyword universe, Search/PMax, RSA assets, Quality Score, conversion tracking) — both in the `content` profile.
- Hard boundary: advisory-only — no ad-platform APIs, no credentials, no spend automation. Fleet: 26 → 27 agents, 82 → 84 commands.

### Changed — English-only migration completed at the body level + VERIFIED CLEAN

v1.32 was the CLI-grammar rename; v1.33 finishes the job by translating every remaining Turkish *body* and proving it. The translation ran in phases (agent/hook/CLAUDE.md bodies, all 84 command bodies, all 59 SKILL.md bodies, lib comments/errors/CLI strings, test titles/messages/comments, CHANGELOG, the TaskBoard sub-system) — then an independent adversarial multi-agent audit ran **7 times** until it returned `VERIFIED CLEAN` (zero residue, high confidence). Each round caught either a deeper layer or an entirely un-swept surface; residue dropped 171 → 55 → 6 → 4 → 7 → watchers → 0.

- All 84 slash-command description lines and all agent `description:` fields are English; `.claude/command-index.md` regenerated (was stale at 50 Turkish entries → 84 English entries grouped by profile).
- Translated bodies: 30 agents, 14 hooks (runtime block/inject/incident messages included), CLAUDE.md, 84 commands, 59 SKILL.md (general 22 / pentest 25 / expo 12), `lib/**` comments + error messages + CLI output (completion scripts, update banner), the generated `.windsurfrules` / `GEMINI.md` / `AGENTS.md` headers, 62 test files (titles + assert messages + comments), `dist/` distribution templates, `.claude/watchers/` definitions, and demo/meta files.
- Intentional Turkish remains only in the data/user layer: tokenizer stopword Sets, the `icerik-helpers` TR→ASCII normalization table, internal identifiers, `README.tr.md` / `CHANGELOG.tr.md`, user memory, SKILL.md `Triggers on:` bilingual routing keywords, and historical CHANGELOG version-history records.
- README/docs currency: counts now 30 agents / 84 commands / 14 hooks / 62 skill categories across README, manifests, docs, and `package.json`.

### Fixed — clean English seed templates for new installs (no maintainer-data leak)

`badi init` was copying the maintainer's own `.claude/memory.md`, `knowledge-base.md`, and `workspace/TaskBoard.md` (Turkish, with internal issue references) into every new project, and `package.json` `files[]` shipped them in the npm tarball.

- New `lib/seed/` holds clean, empty English templates; `seedUserFiles()` writes them on install and never overwrites existing user data.
- `copyRecursive` gained an `exclude` option so the maintainer's working `.claude/` data is no longer copied; `package.json` `files[]` no longer ships `memory.md` / `knowledge-base.md` / `knowledge-nominations.md` / `workspace/`.
- Verified end-to-end with a real `npm pack` tarball install: a new user gets empty English memory/KB/TaskBoard and none of the maintainer's data.

## [1.32.0] - 2026-06-03

> **English-only migration completed + virtual engineering team.** The entire user-facing CLI surface — command output, command grammar, and Claude Code slash commands — is now English. A managerial agent layer (`/ceo-review`, `/eng-review`, `/qa`, `/ship`, `/team`) was added, inspired by the "virtual eng team" pattern.

### Added — Virtual engineering team (managerial agent layer)

- **4 agents:** `product-strategist` (CEO lens), `engineering-manager`, `release-manager`, `qa-lead`.
- **5 commands:** `/ceo-review`, `/eng-review`, `/qa`, `/ship`, and `/team` — a single orchestrator that runs the whole team end-to-end as a gate chain (strategy → plan → build → QA → ship).
- Fleet: **22 → 26 agents**, **77 → 82 commands**.

### Changed (BREAKING) — English command grammar

- Top-level: `badi icerik` → **`badi content`**, `badi tasarim` → **`badi design`**.
- `content` subcommands: `basla`→`start`, `durum`→`status`, `fikir`→`idea`, `kapat`→`close`, `ac`→`open`, `ara`→`search`, `sablon`→`template` (`olustur`/`sil` → `create`/`delete`), `gorsel`→`visual`, `takvim`→`calendar`, `marka`→`brand`.
- Flags: `--sablon`→`--template`, `--tur`→`--type`, `--son`→`--last`, `--ornek`→`--example`; idea type `genel`→`general`.
- `karousel` → **`carousel`** everywhere (CLI type, templates, completion).
- **Breaking:** the old Turkish commands/flags are rejected. Internal source filenames, function names, and workspace data dirs (`takvim/`, `gorseller/`, `marka-sesi.md`) are unchanged (invisible to users).

### Changed (BREAKING) — English slash commands

- 14 Turkish-named Claude Code slash commands renamed under a unified `content-` prefix: `/icerik-uret`→`/content-generate`, `/icerik-basla`→`/content-start`, `/icerik-durum`→`/content-status`, `/icerik-plan`→`/content-plan`, `/icerik-kapat`→`/content-close`, `/icerik-fikir`→`/content-idea`, `/icerik-ara`→`/content-search`, `/icerik-perf`→`/content-perf`, `/icerik-sablon`→`/content-template`, `/icerik-takvimi`→`/content-calendar`, `/gorsel-brief`→`/content-visual-brief`, `/marka-sesi`→`/content-brand-voice`, `/video-senaryo`→`/content-video-script`, `/karousel`→`/content-carousel`.

### Changed — English-only CLI output (i18n phases 2p–2s)

- `list`, `plan`, plugin subcommands; `events` + `mcp`; `bin/badi.js` (`--help` + dispatch/errors); shared lib helpers, harness adapters (claude/cursor/agents), and data manifests — all output strings, doctor-check labels, and code comments translated.
- Intentionally **kept Turkish**: the stopword set in `aso-helpers.js` (keyword-analysis data, not UI — translating it would degrade Turkish ASO keyword extraction).

### Changed — tooling

- `@biomejs/biome` 2.4.15 → 2.4.16 (config migrated to the new schema, codebase reformatted).
- `actions/checkout` 4 → 6 in CI.

---

_The entries below were previously under `[Unreleased]` and ship as part of 1.32.0:_

### Changed — `badi icerik` English-only (TR content removed · English-only migration phase 1)

Content generation is now **English only**. The first phase of the full English-only migration:
- `lib/templates/tr.js` removed; `en.js` is the only content template.
- `--lang` is now a **no-op** (swallowed for backward compatibility but always generates EN); `tr`/`tr,en` is no longer supported.
- Generated files have no language suffix (`-en`/`-tr` dropped); brand voice is `marka-sesi.md`.
- `release-notes` + `marka` outputs are English (header/body/footer translated).
- 3 content tests updated to en-only.

**Breaking**: `badi icerik ... --lang tr` no longer generates TR (generates English).

### Added — `badi release check` lint gate (audit O1)

`badi release check` now runs a `biome check` (lint) stage — a lint error **blocks** publish (level: fail), and can be skipped with `--skip-lint`. Reason: lint/format was not gated in the CI release and rot was accumulating (19 accumulated errors surfaced from this during a maintenance pass). The new `checkLint` was added to the `CHECKS` array; `--skip-lint` is documented in help + parser (help-doctor clean).

### Fixed — `badi publish` workflow auto-sync manifest (#196)

`badi publish --version <bump>` now, after the package.json bump, automatically regenerates the `.claude-plugin/{plugin,marketplace}.json` files and adds them to the same commit. In the v1.30.1 and v1.31.0 publishes the manifest was left stale, requiring a manual `badi release sync-manifest` — this closes that gap.

- New step: **5. Plugin Manifest Sync** (version bump → manifest sync → single commit)
- New flag: `--skip-manifest-sync` (escape hatch, for manual sync)
- `lastUpdated` is informational only (marketplace "last updated" display); `isManifestStale` now **excludes** it (staleness is structural-only) — a toISOString format/instant difference does not produce a false stale (review K1 fix)
- If the `.claude-plugin/` directory does not exist, the step is silently skipped
- Help text + step list updated (1-10)

**Known issue (v1.30.1 + v1.31.0 publish)**: Before these versions the `chore(release)` commit only bumps `package.json`, leaving the manifest stale. The hotfix was done with a manual `badi release sync-manifest` + a separate commit (v1.31.0: PR #195).

## [1.31.0] - 2026-05-22

> Anthropic Claude Code 2.1.126-2.1.147 (May 1-22, 2026) compatibility release. Adds `/security-review` (built-in slash command in 2.1.140+) bridge, `/code-review` (2.1.147+) feature-parity, hook terminal-isolation audit (2.1.139+), marketplace manifest `lastUpdated` field (2.1.144+), and a GitHub Action scaffold wrapping Anthropic's official `claude-code-security-review`.

### Added — Security orchestration (`badi security`)

New CLI command bridging Anthropic's native `/security-review` (built-in in Claude Code 2.1.140+) with badi's deterministic baseline + CI scaffold:

```bash
badi security baseline [--json]      # secret-scan + npm audit deterministic baseline
badi security triage [report]        # filter /security-review report by severity
badi security init --ci [--force]    # GitHub Action scaffold (.github/workflows/security-review.yml)
```

What Badi does not do: AI semantic vulnerability hunt — that is done by Anthropic's native `/security-review`. Badi's bridge: deterministic baseline + post-scan triage + CI orchestration.

Skill/slash cross-refs added: `security-scan.md`, `secret-scan.md`, `security-check/SKILL.md` now point to the native `/security-review` command as the entry point.

### Added — `/review` feature parity (Anthropic `/code-review` 2.1.147+)

`.claude/commands-vault/review.md` argument format extended:

- `effort`: `low` | `medium` (default) | `high` — analysis depth
- `--comment`: post as inline comments on the active PR (via gh CLI)
- `--correctness-only`: focus only on correctness bugs (skip Channel B+C)
- Auto PR context: the active branch's PR is detected via `gh pr view`

Badi `/review` is a **superset** of Anthropic's `/code-review` — 3 channels (security+performance+architecture) + TR report + KRITIK/YUKSEK/ORTA/DUSUK classification + effort tuning + `--comment` + `--correctness-only`.

### Added — Plugin marketplace `lastUpdated` (Anthropic 2.1.144+)

`lib/data/marketplace-manifest.js`:
- `lastPackageJsonCommitDate()` — last version bump date via `git log -1 --format=%cI -- package.json` (ISO 8601)
- `buildPluginManifest()` and `buildMarketplaceManifest()` produce a `lastUpdated` field
- In Anthropic Claude Code 2.1.144+ the plugin's last update time appears in the `/plugin` Browse pane

**Stale-check semantics**: `lastUpdated` is tied to the `package.json` commit and does not change on every commit — only after a version bump. This keeps `badi release check` from reporting "noisy stale". `badi publish --version <bump>` automatically runs `release sync-manifest` after the package.json bump.

`badi release sync-manifest` updates it automatically; if the string is empty the field is not added (stale-check compatibility when there is no git in CI clones).

### Fixed — v1.31.0 internal review hotfix (13 finding)

PR #193 internal `/review` found 13 findings; all findings were closed in the same release:

- **K1**: `badi security baseline` was not running secret-scan (the module CLI entry was missing) — a CLI entry point was added to `secret-scan.js`; a baseline integration test was written (O1)
- **K2**: `runTriage` regexes lacked word-boundaries — they were counting words like "below/follow/yellow"; `\b(...)\b` + markdown heading-first parsing
- **Y1**: `dist/github-actions/security-review.yml` `@main` floating ref — supply chain risk; SHA-pinned (`0c6a49f...`, periodic update)
- **Y2**: `runBaseline` line 105 dead ternary (both branches were identical) — simplified
- **Y3**: `runBaseline`/`runTriage` cwd-relative path — a `projectRoot()` helper was added, fixing usage when called from a subdirectory
- **O2**: secret-scan/npm-audit parse error silent catch — warning to the user via `console.error`
- **O3**: `tests/hooks-isolation.test.js` skill-router fixture early-exit only — updated to a full-flow test prompt
- **O4**: `.claude/commands-vault/review.md` argument parse documentation note — clarify that Claude does the prompt interpretation
- **D1**: `dependency-audit.mjs` had no inject rate limit — `lastInjectedAt` added to the cache, does not inject if younger than 1 hour (reduce context noise)
- **D2**: `docs/enterprise.md` dead outbound link — updated to `code.claude.com/docs/en/server-managed-settings`
- **D3**: TaskBoard 5 issues (#188-192) moved to Completed
- **D4**: This CHANGELOG entry — internal review hotfix record history

### Added — GitHub Action scaffold (`dist/github-actions/security-review.yml`)

Opt-in workflow template:
- Wraps Anthropic's official `anthropics/claude-code-security-review` action
- `permissions: pull-requests: write, contents: read`
- `ANTHROPIC_API_KEY` secret ref (not input)
- Badi-specific default excludes: `node_modules, coverage, dist, _bootstrap, .claude/skills-vault, .claude/commands-vault`
- `pull_request` (head SHA) — NOT `pull_request_target` (prompt injection hardening)
- Custom prompt + FP filter optional

Setup: `badi security init --ci` scaffolds into the project root.

### Fixed — Hook terminal-isolation (Anthropic 2.1.139+)

Anthropic Claude Code 2.1.139 (11 May 2026) started running hooks without terminal access. Badi's 15 hooks (14 + `_util.mjs`) were audited:

- **Category 1 (JSON output protocol / log-only)**: 13 hooks — safe, no change
- **Category 2 (plain text stdout protocol violation)**: 2 hooks — **FIX applied**
  - `dependency-audit.mjs`: `process.stdout.write(plain)` → `writeContextInjection()`
  - `post-compact-resume.mjs`: `process.stdout.write(plain)` → `writeContextInjection()`
- **Category 3 (terminal manipulation / ANSI)**: 0 hooks — clean

This fix was technically wrong before v1.31.0 too: Category 2 outputs did not enter Claude's context, falling to the terminal or getting lost. Now they are genuinely injected via the `additionalContext` JSON protocol.

Audit report: `docs/hooks/isolation-audit.md`.

### Added — Documentation

- **`docs/enterprise.md`** — Anthropic managed-settings compliance guide (`forceLoginOrgUUID`, `allowManagedDomainsOnly`, `parentSettingsBehavior`, hook isolation, plugin marketplace, telemetry, SSO, audit log)
- **`docs/hooks/isolation-audit.md`** — categorized report of 15 hooks + Category 2 fix documentation
- **README** "Security Notes (v1.31.0+)" section:
  - `--dangerously-skip-permissions` Claude Code 2.1.126+ scope extension warning
  - Hook isolation crossref
  - Enterprise managed-settings crossref

### Added — Tests

54 new tests:
- `tests/hooks-isolation.test.js` (45 tests) — for each hook: JSON-only stdout + no ANSI escape + empty stderr
- `tests/security.test.js` (6 tests) — `badi security baseline/triage/init --ci` subcommands
- `tests/marketplace-manifest.test.js` (+3) — `lastUpdated` field + dist/github-actions scaffold verification

Test count: 1074 → 1128.

### Changed

- `bin/badi.js` — `security` command added to the dispatcher, "Security orchestration (v1.31+)" section in the help text
- `.gitignore` — `!dist/github-actions` exception (tracked workflow template)
- `dist/README.md` — "GitHub Action templates (v1.31.0+)" section

### Closes

#188 (security-review integration), #189 (/review parity), #190 (marketplace lastUpdated), #191 (CI scaffold), #192 (hook isolation audit).

## [1.30.1] - 2026-05-19

> Includes review-driven refinements (see `### Refinements (post-#185 internal review)` below) before npm publish. All 9 review findings closed in the same release.

### Added — Multi-channel distribution mirrors

Badi is now installable from **2 live channels + 2 in-progress channels** (skeletoned in `dist/`, awaiting tap/bucket repo creation):

```bash
npm i -g @fatihkan/badi                                       # primary (canonical) ✅ live
/plugin install fatihkan/badi                                 # Claude Code marketplace ✅ live
brew tap fatihkan/badi && brew install badi                   # Homebrew (macOS/Linux) ⏳ coming soon
scoop bucket add badi <repo> && scoop install badi            # Scoop (Windows) ⏳ coming soon
```

The Homebrew tap (`fatihkan/homebrew-badi`) and Scoop bucket (`fatihkan/scoop-bucket`) repos are **not yet published**; this release ships the manifest skeletons and CI workflow. Until those repos exist, use npm or the Claude Code marketplace.

- **`.claude-plugin/plugin.json` + `marketplace.json`** brought up-to-date with v1.30.0 reality (was stale since v1.16.5 — 21 → 22 agents, .sh → .mjs hooks, added `skills` field pointing to `./.claude/skills-vault`, added UserPromptSubmit hook block for the new plan-inject hook).
- **`dist/homebrew/badi.rb`** — npm-backed Homebrew formula skeleton. Sha256 is populated by the release workflow after npm publish.
- **`dist/scoop/badi.json`** — Scoop manifest skeleton with `autoupdate` block pointing to npm registry.
- **`dist/README.md`** — Setup guide for the separate `homebrew-badi` tap and `scoop-bucket` mirror repos.
- **`.github/workflows/dist-publish.yml`** — Opt-in workflow (`workflow_dispatch` only, no auto-trigger) that downloads the npm tarball, computes sha256, and pushes updated formulae to the tap/bucket repos when `DIST_PUBLISH_TOKEN` secret is configured. Matches Badi's "no auto-CI" policy.

### Added — `badi release sync-manifest`

```bash
badi release sync-manifest          # regenerate .claude-plugin/*.json from package.json + .claude/
badi release sync-manifest --dry-run # show what would be written, don't touch disk
```

Reads `package.json` + walks `.claude/{agents,commands,hooks,skills-vault}` to produce stable, deterministic plugin manifests. Alphabetical agent listing → no spurious git diffs across runs.

A new `release check` validation surfaces drift:

```
XX  .claude-plugin/plugin.json present      (generate with badi release sync-manifest)
!!  .claude-plugin/plugin.json current      (stale — badi release sync-manifest)
```

Block-on-stale ensures every npm publish ships with marketplace metadata that actually matches the package.

### Module additions

- `lib/data/marketplace-manifest.js` — generator + staleness checker (`buildPluginManifest`, `buildMarketplaceManifest`, `collectAgents`, `countHooks`, `countCommands`, `countSkillCategories`, `isManifestStale`, `writeManifests`). Pure functions; deterministic output.
- `lib/commands/release.js` — `checkMarketplaceManifest` added to `CHECKS` array; `runSyncManifest` added as a new subcommand.

### Refinements (post-#185 internal review)

Pre-publish hotfix closing 9 review findings:

**Security**
- **K1** — Workflow `${{ inputs.version }}` and `${{ steps.X.outputs.* }}` no longer flow directly into shell scripts. All user-controlled values pass through an explicit `env:` block, then are referenced as quoted `$VAR` inside the shell. Defends against shell injection from `workflow_dispatch` inputs even if a collaborator account is compromised. Additionally, the version field is semver-validated before being used downstream.
- **Y1** — Git operations no longer embed `DIST_PUBLISH_TOKEN` in the clone URL (`https://x-access-token:$TOKEN@...`). Instead, an `http.https://github.com/.extraHeader` config injects `Authorization: Basic <base64(x-access-token:$TOKEN)>`. Token never appears in git URLs, error logs, or process listings.
- **O2** — Bot commits now use the GitHub Actions standard identity `github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>` (verified-bot-eligible) instead of the generic `noreply@github.com`.

**Correctness**
- **O3** — Marketplace manifest staleness check uses a recursive `deepEqualJson` instead of `JSON.stringify` comparison. Hand-edited or differently-formatted JSON files with the same semantic content are no longer reported as stale. Generator output remains deterministic; this only affects the false-positive case when users (or tooling) reorder keys.
- **D2** — Scoop installer script now `throw`s on `npm install` failure (`$LASTEXITCODE -ne 0`) and on missing `badi` binary post-install. Previously `Write-Error` was used, which does not change the exit code — Scoop treated failed installs as successful.

**Docs**
- **Y2** — `README.md` install matrix gains a **Status** column. Homebrew and Scoop rows show `⏳ Coming soon (tap/bucket repo setup pending)` until the mirror repos (`fatihkan/homebrew-badi`, `fatihkan/scoop-bucket`) are created. CHANGELOG entry also notes the in-progress status.

**Quality**
- **O1** — Unused `statSync` import removed from `lib/data/marketplace-manifest.js`.
- **D1** — `.gitignore` now covers `.DS_Store` (macOS) and `Thumbs.db` (Windows) at the project root. Prevents OS metadata from slipping into newly-tracked `dist/` directories.
- **D3** — Workflow header now documents implicit `python3` runtime dependency (provided by `ubuntu-latest` default).

### Tests

1054 → **1074** (+20):
- `tests/marketplace-manifest.test.js` — generator unit (11) + on-disk stale detection (1) + dist skeleton + hardened workflow regex assertions + `deepEqualJson` (6)

### Distribution channels NOT in this release

- AUR (Arch User Repository) — community-driven, no in-repo PKGBUILD.
- deb / rpm — too much overhead for an npm-backed CLI.
- Cargo / Go modules — N/A.

## [1.30.0] - 2026-05-19

> Includes review-driven refinements (see `### Refinements (post-#182 internal review)` below) before npm publish. All 11 review findings closed in the same release; consumers always see the polished version.

### Added — agent-agnostic context export (`init` / `update`)

Two new harness adapters extend Badi's multi-agent reach. The canonical
CLAUDE.md + memory + knowledge-base content can now be compiled into:

- **Windsurf** — `.windsurfrules` single-file rules. Detected via
  `existsSync(".windsurfrules")`. `badi init --harness windsurf`.
- **AGENTS.md (Generic)** — neutral fallback for OpenAI Codex CLI, Aider,
  and any tool that reads project-level AGENTS.md. `badi init --harness agents`.

Combine: `badi init --harness all` now emits 5 outputs (Claude Code, Cursor,
Gemini, Windsurf, AGENTS.md). Other harnesses (Claude/Cursor/Gemini) were
already shipped in earlier versions; this release adds the missing two.

### Added — `badi release check` pre-flight verifier

New standalone command that validates publish readiness without mutating
state. Performs 7-9 checks: git cleanliness, branch, `package.json`
presence, CHANGELOG.md + CHANGELOG.tr.md version entry, `npm test`
exit status, `gh` CLI presence, `npm pack --dry-run` tarball size.

```bash
badi release check                       # full check (incl. npm test)
badi release check --bump minor          # compute target version from package.json
badi release check --version 1.30.0      # check for explicit version
badi release check --strict              # warnings → errors (CI mode)
badi release check --skip-test           # quick check, no npm test
```

Reduces the publish-fails-late feedback cycle (dirty tree / missing
CHANGELOG / failing tests) from "discovered during `badi publish`" to
"caught before publish".

### Added — `inject-active-plan` hook (UserPromptSubmit)

New hook at `.claude/hooks/inject-active-plan.mjs`. Registered in
`.claude/settings.json` as a UserPromptSubmit hook. On every user prompt:

1. Scans `.claude/plans/<slug>.approved` markers
2. Loads each approved plan's `.md` body
3. Injects up to 5 plans into Claude's context as
   `<active-plan slug="X" state="approved">…</active-plan>` blocks
4. Hard cap: 200KB total per injection

Pending/denied plans are not injected. No-op when no approved plan
exists. Couples `badi plan approve` to active Claude context — the plan
isn't just gated, it's recalled every prompt.

### Added — plugin manifest `apiVersion` + dependency graph

`badi-plugin.json` gains a `badi` field for forward-compat:

```json
{
  "name": "my-plugin",
  "version": "0.3.0",
  "badi": {
    "apiVersion": "1.x",
    "dependsOn": ["other-plugin@>=0.2"]
  }
}
```

Range syntax supports `*`, `X.x`, `X.Y.x`, `X.Y.Z`, `>=X.Y[.Z]`. Default
when omitted: `1.x` (covers existing plugins). On install, mismatched
apiVersion surfaces a warning (non-blocking — empirical compat preferred
over hard gate).

Two new subcommands:

- **`badi plugin doctor`** — audits all installed plugins: manifest
  validity, apiVersion compat, missing/version-mismatched deps. Exits 1
  on any issue (CI-friendly).
- **`badi plugin graph`** — prints topologically sorted load order:
  ```
  ├─ base-plugin v1.0.0
  └─ derived-plugin v0.3.0 ← base-plugin
  ```

Internals: new `lib/data/plugin-manifest.js` (`parseRange`,
`validateManifest`, `checkApiCompat`, `topoSort`, `findUnsatisfied`).
Cycle detection in `topoSort` throws explicit error.

### Added — `badi events` self-telemetry

Badi now emits its own typed events alongside reading Claude Code
transcripts. Every CLI invocation emits:

- `badi.command.started` (cmd, args_count)
- `badi.command.completed` (cmd, duration_ms, exit_code: 0)
- `badi.command.failed` (cmd, duration_ms, error_message, exit_code: 1)

Events land in `~/.claude/projects/<project>/badi-events.jsonl` (same
directory Badi already reads for stats/session). **Privacy:**

- Append-only JSONL, **local only** — no network calls
- Whitelisted event types (`ALLOWED_TYPES`); unknown types dropped
- Arg values are NOT stored (only `args_count`), preventing accidental
  secret leakage
- String fields capped at 200 chars, arrays at 50 elements
- `BADI_TELEMETRY=off` (or `0`/`false`) completely disables emission

New reader command:

```bash
badi events list [--limit N]            # last N events
badi events stats                       # per-command count + avg duration + fail count
badi events tail                        # list --limit 10
badi events status                      # telemetry on/off + log size
badi events path                        # log file path
```

Filters: `--since DATE`, `--until DATE`, `--cmd <name>`, `--type
<badi.command.completed>`.

Hot reload pattern: pure-function emitter (`emit(type, data)`) → file
append; reader (`readEvents()`) → parse + reverse. Wired into
`bin/badi.js` dispatcher so all commands instrument identically without
per-command changes.

### Refinements (post-#182 internal review)

Pre-publish hotfix closing 11 findings from an internal `/review`:

**Performance / observability**
- **B2** — Plan inject hook context budget reduced from 200KB (~25% of Claude context) to **50KB default + 3 plans default**. Configurable via `BADI_PLAN_INJECT_MAX_BYTES` / `BADI_PLAN_INJECT_MAX_PLANS` env. `BADI_PLAN_INJECT_OFF=1` disables.
- **B4** — `badi events list` now uses tail-reader (`readEventsTail`) — last K bytes via fd-based `readSync`, O(K) instead of O(N). Filter use (`--since/--until/--cmd/--type`) still falls back to full read.
- **C6** — Dispatcher now uses `process.on("exit", code => emit(...))` so `process.exit(1)` paths in commands correctly emit `badi.command.failed` (previously only `badi.command.started` made it to the log for those).

**Correctness**
- **B3** — `runNpmTest` regex anchored to TAP `^# pass N$` / `^# fail N$` to prevent miscounting "passed" / "failed" in test names or timings.
- **A3** — Plan injection escapes `</active-plan>` literals inside plan body, preventing the XML-ish wrapper from closing early when a plan's content happens to contain that exact tag.

**Diagnostic clarity**
- **A2/B1** — `parseRange` now returns `{ matcher, recognized }`. `checkApiCompat` surfaces a `warning` field when the apiVersion string is in an unrecognized format. `badi plugin doctor` shows the warning and counts it as an issue (previously the permissive fallback misled operators by silently passing unknown formats).

**Architecture (zero behavior change)**
- **C1** — Single-file harness factory `lib/harnesses/_single-file.js`. `gemini/windsurf/agents` refactored to use it. **551 → 339 lines (−38%)**; a new single-file harness now needs ~26 lines instead of ~170.
- **C2** — `runReleaseCheck` refactored from a 113-line monolithic procedure into a `CHECKS` array of pure check functions. Each check is independently unit-testable; future plugins can `CHECKS.push(...)` to extend.
- **C3** — `lib/commands/plugin.js` (437 lines) split into `lib/commands/plugin/{install,remove,list,show,doctor,graph,help,_shared}.js` following the `icerik/` pattern.
- **C5** — `ALLOWED_TYPES` whitelist now also accepts `plugin.<owner>.<event>` namespace (regex `/^plugin\.[a-z0-9-]+\.[a-z0-9.-]+$/`), enabling plugin-emitted events without modifying core. `badi.*` remains a closed list.

**Plan inject hook env knobs**
- `BADI_PLAN_INJECT_MAX_BYTES` — total injection cap (default 50KB; min 1KB)
- `BADI_PLAN_INJECT_MAX_PLANS` — max plans to inject (default 3; min 1)
- `BADI_PLAN_INJECT_OFF` — `1`/`true`/`off` disables hook
- `BADI_HOOK_DEBUG=1` — `[plan-inject]` stderr trace

### Tests

967 → **1054** (+87 across the new feature set + refinements):
- `tests/harness-extras.test.js` (10) — windsurf/agents harness
- `tests/cli.release.test.js` (3) — release help / subcommand routing
- `tests/cli.events.test.js` (6) — events status/list/path CLI
- `tests/plugin-manifest.test.js` (24) — parseRange/validateManifest/
  checkApiCompat/topoSort/findUnsatisfied
- `tests/event-emitter.test.js` (4) — ALLOWED_TYPES, emit safety
- `tests/cli.plugin-doctor.test.js` (4) — post-split doctor + apiVersion recognized warning
- `tests/release-checks.test.js` (9) — CHECKS array, individual check functions, bumpVersion
- `tests/event-emitter-extras.test.js` (13) — isAllowedType, plugin.* wildcard, readEventsTail, parseRange recognized
- `tests/harness-factory.test.js` (7) — buildSingleFileHarness, extraWriter, detect, force semantics

Existing test updates: harness registry test now expects 5 ids (was 3),
hooks fail-safe test now expects 14 hooks (was 13).

## [1.29.0] - 2026-05-19

### Added — observability v1.29 (Claude Code transcript bazli)

Five+ new commands and flags that read `~/.claude/projects/*.jsonl`
transcripts directly. Privacy-preserving: nothing leaves your machine.

- **`badi stats --session [--limit N]`** — last-N sessions: id (8-char),
  project, branch, model class, duration, prompts, tools, $cost.
- **`badi stats --models`** — per-model session count + USD breakdown +
  global cache hit rate.
- **`badi stats --cost`** — total USD across all transcripts + top-10
  projects.
- **`badi stats --since DATE --until DATE`** — ISO/YYYY-MM-DD range filter
  (applies to `--session`/`--models`/`--cost`).
- **`badi stats --branch <name>`** — git branch filter using transcript
  `gitBranch` field.
- **`badi search "<query>"`** — multi-token AND search across all
  transcripts (user prompts + last-prompt events). Supports
  `--since/--until/--branch/--limit`.
- **`badi session <id-or-prefix>`** — single-session detail: header
  (model/tokens/cost/duration), tool breakdown, files touched, chronological
  timeline (prompts + tool_use + thinking). `--full` for full timeline,
  `--tools` / `--files` for focused views.
- **`badi plan list/new/show/status/approve/deny/reset`** — local
  file-based plan approval workflow. Markers in
  `.claude/plans/<slug>.{approved,denied}`.
  `badi plan status <slug> --format json` exits 0 only if approved (hook-friendly).
  Slug validated against `/^[a-z0-9][a-z0-9-]{0,63}$/` (path traversal safe).
- **`badi plugin show <name>`** — manifest detail: version, description,
  agent/command/hook/skill counts with full per-item listing.
- **`badi list --mcp`** — aggregate MCP server invocations across all
  transcripts; per-server call count + unique tool count.

Internals: new `lib/data/transcript-reader.js` (parseSession,
parseSessionWithEvents, applyFilters, MODEL_PRICING, costForUsage,
findSession, formatDuration, shortSessionId). Pricing table for
Opus/Sonnet/Haiku 4.x with family-name fallback for unknown variants.

Tests: 934 → **967** (+33 across `tests/transcript-reader.test.js`,
`tests/cli.plan.test.js`, `tests/cli.observability.test.js`).

## [1.28.1] - 2026-05-16

### Added — help-doctor: automated drift detector

A regression test that walks every `lib/commands/*.js` and verifies that every
subcommand the parser accepts (`case "x":`, `args[0] === "x"`) and every flag
it parses (`args.includes("--x")`, `a === "--x"`) is documented in the
file's user-facing help text. Catches the v1.27.1-class drift (route flags
missing from `commands --help`) at PR time instead of after release.

- `lib/help-doctor.js` — pure `detectDrift(filePath)` / `auditFiles(files)` /
  `loadAllowlist(path)` helpers. Console.log-based help extraction handles
  inline help, separate `showHelp()` functions, and files with multiple
  exports (e.g. `commit.js` defining both `runCommit` and `runChangelog`).
- `tests/help-doctor.test.js` — 8 tests: full repo audit + 5 detector units
  + allowlist schema validation. Hooks into `npm test`; CI fails on drift.
- `.claude/help-doctor.allow.json` — false-positive allowlist with a
  mandatory `_why:` rationale per entry. Used for flags documented in
  `bin/badi.js` top-level help (init/update/doctor/list) and internal
  aliases (market `full`).
- `badi doctor help` — new CLI subcommand that runs the same audit
  interactively. `--format json` for CI consumption, `--strict` to fail on
  any drift.

### Fixed — drift exposed by the new detector

- `lib/commands/market.js` — `--days`, `--query`, `--json` flags added to
  the `Options:` block (were parsed but never documented).
- `lib/commands/stats.js` — `--week` flag added to help (default behaviour
  but the explicit form was undocumented).
- `lib/commands/publish.js` — `--source` flag added to the skill-bundle
  help (existed in the code path, not in the help text).

### Fixed — security hardening pass (`/security-scan` 6 findings)

A `/security-scan` run on v1.28.0 surfaced 1 YUKSEK + 3 ORTA + 2 DUSUK
findings. All 6 closed in this release.

#### Y1 — `badi skills add/remove <name>` path traversal

The `name` argument flowed directly from CLI into `join(vault, name)` and
`join(active, name)` with no validation. Empirically reproduced: `badi
skills add ../../<existing-dir>` reached `copySkill` because
`isSkillCategory` saw the resolved path existed. A malicious script could
copy arbitrary directories into `.claude/skills/`, exposing their content
to the next Claude Code session.

Fix: every name passes through `/^[a-z0-9][a-z0-9-]*$/` (`isValidSkillName`,
exported). Names with `/`, `\`, `..`, leading `.`, or uppercase are
rejected at `copySkill`, `removeSkill`, and `isSkillCategory`.

#### O1 — `badi plugin install <source>` git argument injection

`git clone --depth 1 <source> <dest>` interpreted any `source` starting
with `-` as a flag. Vector: `--upload-pack=<command>` (git option
injection) or `-u <command>` could execute arbitrary commands.

Fix: explicit `source.startsWith("-")` rejection + `--` separator on the
git argv (defense-in-depth): `["clone", "--depth", "1", "--", source, dest]`.

#### O2 — `tests/cli.secret-scan.test.js` fixtures triggered the scanner itself

The deliberate sample strings for `private-key`, `mongodb-uri`, and
`postgres-uri` were stored as raw string literals; running `badi
secret-scan` on the repo working tree returned 3 self-findings. CI noise
that could mask real leaks.

Fix: split-string concatenation for the three samples (`"mongo" + "db://..."`
etc.). Runtime behaviour unchanged; static source no longer contains the
literal patterns.

#### O3a — `badi tasarim export --write <path>` no project-root scope

User-supplied `--write` path resolved without containment check; could
write outside the project root.

#### O3b — `badi secret-scan --ignore-file <path>` / `--patterns <path>`

Same class as O3a. A security tool should not silently read arbitrary
absolute paths.

Fix for both: new `assertWithinProject(baseDir, candidate, flag)` /
`isWithinProject` helper. Resolves the path, checks `relative()` doesn't
start with `..` or `/`. Reject with a clear error message otherwise.

#### D1 — `npm audit`: 3 moderate (dev-dependency only)

3 moderate advisories in `esbuild` (transitively `vite` → `vitepress`).
All dev-only — used to build the docs site, never shipped to npm consumers.
`fixAvailable: false` upstream. Documented here, no code change; will be
resolved when the `vitepress` chain ships a fix.

### Added — `tests/security-hardening.test.js`

11 new regression tests: `isValidSkillName` positive/negative, CLI
path-traversal rejection, plugin `--` flag rejection, tasarim `--write`
scope guard, secret-scan `--ignore-file`/`--patterns` scope guard, and the
O2 meta-check that `badi secret-scan` on the repo itself returns 0
findings.

### Stats

- Tests: 915 → 934 (+19: +8 help-doctor, +11 security hardening)
- help-doctor: 16 files apparent drift → 2 real fixes + 5 allowlist entries with rationale; 0 unexplained drift remaining.
- security hardening: 5 files changed (skills.js, plugin.js, tasarim.js, secret-scan.js, test fixture); audit post-fix 0/0 KRITIK/YUKSEK, ORTA→1, DUSUK→1 (npm audit dev-only)

## [1.28.0] - 2026-05-16

### Fixed — secret-scan: critical CI silent-pass

A `/review` audit followed by empirical probes (planted secrets in a sandbox
project) uncovered five real bugs in the secret scanner. Two are merge-blocker
critical; without these fixes a CI pipeline using `badi secret-scan --format
json` would have **passed green** even with leaked credentials.

**K1 — JSON mode never exited non-zero on critical findings.** The JSON output
branch returned before the `process.exit(1)` call. Reproduced live: planted
Anthropic key → text mode `exit 1`, JSON mode `exit 0`. CI pipelines wired to
`if ! badi secret-scan --format json` saw success despite KRITIK findings.

**K2 — Dedup key collapsed distinct secrets that shared the same masked prefix
+ suffix.** Two genuine OpenAI keys with identical first/last 4 characters
collapsed into a single finding; the second leak was silently dropped. Reproduced
live in `/tmp/badi-probe`. Dedup key now includes file path + raw match.

**Y1 — `statSync` followed symbolic links** — risk of cycles (`node_modules/.cache
-> ../`) and path traversal outside the project. Symlinks are now skipped
unconditionally with an `entry.isSymbolicLink()` early-continue; the count
appears in JSON output as `scanned.symlinksSkipped`.

**Y2 — `github-classic` regex `[a-f0-9]{40}` matched every SHA-1 hash.** Every git
commit hash in code comments triggered a DUSUK false positive. GitHub deprecated
classic tokens in 2021; replaced with `github_pat_[A-Za-z0-9_]{82}` (fine-grained
PAT) at KRITIK severity.

**Y3 — Test coverage** expanded from 4 to 51 tests including K1/K2 regression
guards, symlink handling, the `--git` history scan, every PATTERNS entry's
canonical sample, and the dedup-collision empirical reproduction.

### Added — secret-scan: configurability + transparency

- `--exit-code <critical|strict|never>` — explicit CI contract:
  - `critical` (default): KRITIK + YUKSEK → exit 1 (previous behavior, now also in JSON)
  - `strict`: any finding → exit 1
  - `never`: report only, always exit 0
- `--max-commits N` — git history depth limit (default 100). Emits a stderr
  warning when truncation occurs (previously silent).
- `--max-files N` — file scan limit (default 5000).
- `--ignore id1,id2,...` — comma-separated pattern-id allowlist.
- `--ignore-file <path>` — load pattern-ids from a `.secretignore`-style file
  (auto-discovers `.secretignore` in CWD).
- `--patterns <path>` — load custom organization-specific patterns from JSON
  (merged with the built-in 17).
- JSON output now includes `scanned.totalCommits`, `scanned.truncated`, and
  `scanned.symlinksSkipped` for transparency.
- `--help` documents exit codes, every flag, and out-of-scope items (stash,
  reflog, packed-refs).

### Changed — secret-scan: internal restructure

- Pattern registry externalized to `lib/data/secret-patterns.js` (canonical),
  enabling reuse + future user customization.
- `runSecretScan` split into pure helpers (`scanContent`, `dedupFindings`,
  `applyIgnore`, `groupBySeverity`, `computeExitCode`, `printText`, `printJson`,
  `parseArgs`) — exported for direct unit testing.
- `MAX_FILE_SIZE_BYTES`, `MAX_COMMITS_DEFAULT`, `GIT_SHOW_MAX_BUFFER` etc.
  promoted to named constants.

### Changed — `/security-scan` slash command

- Typographic drift fixed (`Bagiml ilik`, `konfigur asyon`, `Taramas i`, etc.).
- Documents new flags + CI exit-code contract + out-of-scope items.
- `/secret-scan` slash command updated similarly.

### Breaking?

This is technically additive — exit-code behavior in text mode is unchanged
(KRITIK/YUKSEK → 1, default). **JSON mode changes from always-0 to matching
text mode**. Any CI pipeline that relied on JSON mode returning 0 on findings
was relying on a bug; this release surfaces those findings correctly. To
restore the old always-0 behavior, pass `--exit-code never`.

### Stats

- Tests: 868 → 915 (+47 secret-scan tests, including 51 in the rewritten suite)
- 4 files changed in `lib/`; 1 new (`lib/data/secret-patterns.js`)
- 1 file changed in `.claude/commands/`

## [1.27.1] - 2026-05-16

### Fixed — help completeness across CLI surfaces

Deep audit of `badi <cmd> --help` outputs revealed three classes of gaps;
all are now closed.

- **Top-level `badi --help`**: the `aso` / `market` / `tasarim` lines were
  rendered on a single concatenated row because `console.log` was called
  once with three arguments (which inserts spaces, not newlines). Split
  into three separate calls. (`bin/badi.js`)
- **`badi commands --help`** (v1.26+): the `route "<prompt>"` subcommand,
  stdin form, profile flags (`--yes` / `--dry-run` / `--force` /
  `--verbose`), and route flags (`--top N` / `--inject` / `--json`) were
  all missing from the help text — only the switch statement knew about
  them. Help text now mirrors what the code actually accepts, with five
  example invocations and a note that user-authored commands are
  preserved on profile switch. (`lib/commands/commands.js`)
- **`badi skills --help`** (v1.20+): the `--top N` and `--json` route
  flags were undocumented; the categories footnote was frozen at
  "v1.25 / 50 categories" even though v1.27 ships 62 categories (25
  general + 25 `pentest-*` + 12 `expo-*`). Both updated.
  (`lib/commands/skills.js`)

No behavior change — these are doc-string corrections. The actual
subcommand and flag parsing already worked; the help text was lagging
behind. Tests: 868 still passing.

### Changed — README drift

The 1.27 release notes table listed test count as `805 → 815` (a typo —
real count jumped to 868 because of the 53 hook fail-safe resilience
tests added in PR #163). README hero rows still said "50 opt-in skill
categories" and "805 passing tests". All four spots corrected across
`README.md` and `README.tr.md`. A new "Profile-Based Commands + Command
Router" subsection was added that documents `badi commands profile/route`
end-to-end (the v1.26 release shipped without a dedicated README section).

## [1.27.0] - 2026-05-15

### Added — expo-* skill family (12 categories, advisory)

Mobile development lifecycle skill family covering the full Expo + React
Native cross-platform workflow — from project setup to App Store / Play
Store release. Twelve new opt-in categories under the `expo-*` namespace.

**Scope**: advisory only. Badi guides through configuration, command
sequences, and trade-offs; the user runs the actual build/submit/update
commands themselves.

#### Categories (12)

- `expo-orchestrator` — workflow choice (managed/bare/dev-client), project
  setup, package selection, EAS account linking, release planning, delegates
  to sub-skills.
- `expo-router` — file-based routing (`app/` directory), `_layout.tsx`,
  dynamic routes, parallel routes, deep linking, prefetch, tab/stack/drawer
  patterns.
- `expo-eas-build` — `eas.json` profiles (development/preview/production),
  credentials (iOS provisioning + push cert; Android keystore + service
  account), build cache, secrets, monorepo support.
- `expo-eas-submit` — App Store Connect + Google Play submit, metadata
  management, build artifact selection, review notes, phased release,
  screenshot upload rules.
- `expo-eas-update` — OTA update fundamentals, channels, runtime versions,
  branch management, rollback strategy, embedded vs OTA payload, asset
  selection.
- `expo-config-plugin` — `withInfoPlist`, `withAndroidManifest`,
  `withDangerousMod`, `withEntitlementsPlist`, mod compose, plugin testing.
- `expo-prebuild` — managed → bare migration, `npx expo prebuild`,
  ios/android directory authority, `.easignore`, native upgrade discipline.
- `expo-modules` — Expo Modules API (Swift/Kotlin), `expo-module-scripts`,
  `expo.modules.json`, `requireNativeModule`, async function definition,
  view module, event emitter pattern.
- `expo-dev-client` — `expo-dev-client` setup, build profile, custom dev
  menu, runtime version compatibility, EAS Update integration, debug build
  vs Expo Go comparison.
- `expo-notifications` — push token retrieval, FCM (Android) + APNs (iOS)
  credentials, notification categories, action buttons, scheduled
  notifications, channels (Android 8+), permission flow.
- `expo-app-config` — `app.json` vs `app.config.ts`/`app.config.js`,
  environment variables (.env + EAS Secrets), variants (dev/staging/prod),
  extra fields, plugin chain, slug/scheme/bundle/version discipline.
- `expo-troubleshooting` — Metro cache, version mismatch
  (`expo install --check`, `expo-doctor`), Pod install errors, Gradle
  daemon, native module conflicts, EAS Build log reading, dependency
  hoisting issues.

### Changed

- `lib/skills/stack-map.js` — existing `expo` entry expanded to suggest
  `expo-orchestrator` + `expo-app-config` + `expo-troubleshooting`. Six new
  detection entries added: `expo-eas` (eas.json), `expo-router`
  (expo-router package or `app/` dir), `expo-modules` (expo-modules-core
  package or `modules/` dir), `expo-notifications`, `expo-dev-client`,
  `expo-config-plugin`.

### Fixed — hook defensive fail-safe (#162)

All 13 `.claude/hooks/*.mjs` files now register top-level
`uncaughtException` + `unhandledRejection` handlers that force `exit(0)`
on runtime errors. Claude Code's session will no longer surface
"Failed with non-blocking status code" warnings for transient hook
failures. Set `BADI_HOOK_DEBUG=1` in your environment to emit error
messages to stderr for diagnosis.

This addresses the symptom reported on Node v25.x where Stop hook
async invocation surfaced module resolution errors. Note: this fix
hardens runtime error recovery but cannot catch entry-point load
failures (e.g. missing files); use `badi doctor` to verify hook files
exist on your install.

### Tests

- Test count: **805 → 868 (+63)**, all green.
- New `describe("detectStack: expo-* family (v1.27)")` block in
  `tests/stack-detector.test.js`: 10 tests covering package, directory,
  and config file detection signals.
- New `tests/hooks-failsafe.test.js`: 53 tests verifying every hook
  exits 0 on empty / malformed / valid stdin, plus marker validation.

### Vault

- Total opt-in categories: **50 → 62** (25 general + 25 pentest-* + 12
  expo-*).

## [1.26.0] - 2026-05-15

### Added — Profile-based command management + prompt-aware command routing

Token efficiency and DX hygiene release. Three integrated changes that
make Badi's 77 slash commands prompt-aware and profile-tagged.

#### A. Session profiles (commands-vault)

- `.claude/commands-vault/` parallel canonical store for command files.
- `badi commands migrate` copies existing `.claude/commands/` to vault.
- `badi commands profile <core|dev|content|pentest|all>` switches active
  profile. Removes non-profile commands from `.claude/commands/` (vault
  untouched); restoring is `badi commands profile all`.
- 77 commands tagged across four profiles:
  - **core** (21): session, measurement, audit — always active.
  - **dev** (39): development, devops, security, audit tooling.
  - **content** (17): social media, brand, content calendar.
  - **pentest** (0): authorized pentest engagement (future).
- User-defined commands (not in COMMAND_PROFILES) are never touched.

#### B. Top 10 command slim

Mechanical refactor of the ten largest commands to remove padding while
preserving every information point:

- `gorsel-brief`, `video-senaryo`, `icerik-takvimi`, `system-audit`,
  `karousel`, `icerik-uret`, `marka-sesi`, `api-doc`, `playbook`,
  `project-architect`.
- Per-invocation cost reduced (each command body shrinks on average ~30%).
- Bodies remain copy-paste ready; no value loss.

#### C. Prompt-aware command routing

- `badi commands route "<prompt>"` scores commands by keyword match.
- `--inject` flag emits a lightweight hint blob (command name + 1-line
  description; ~30–50 tokens per match, vs ~1.3K for full SKILL.md bodies).
- `UserPromptSubmit` hook (`.claude/hooks/skill-router.mjs`) now calls
  both `skills route` and `commands route`. Result: when a prompt
  mentions a topic (e.g. "deploy", "Instagram post"), Badi quietly
  suggests the relevant command(s) without loading their full body.

### Added

- `lib/command-profiles.js` — profile mapping (`COMMAND_PROFILES`,
  `commandsForProfile`, `isInProfile`, `profileCounts`).
- `lib/commands/commands.js` — CLI module (migrate, profile, route,
  list, available, profiles, restore).
- `lib/skills-router.js` — `indexCommandFile`, `buildCommandIndex`,
  `buildCommandHint` for command-vault routing.
- `.claude/commands.profile.json` — persisted active profile state.

### Changed

- `.claude/hooks/skill-router.mjs` now dispatches to both skills and
  commands routers (4s timeout per call, fail-soft).
- `bin/badi.js` — `commands` subcommand wired in.

### Tests

- Test count: **774 → 805 (+31)**, all green.
- New: `tests/command-profiles.test.js` (16 tests), `tests/commands-cli.test.js`
  (15 tests).

### Notes — Token economics correction

`badi ai token` reports the **POTENTIAL** ceiling (sum of all file sizes),
not the actual per-turn loaded tokens. Claude Code loads `.claude/commands/`
only on slash invocation; only short descriptions appear in the system
prompt. v1.26 optimizes both:
- **Baseline (system prompt description list)**: profile switch removes
  the descriptions of out-of-profile commands.
- **Invocation cost**: top 10 commands now ~30% smaller bodies.
- **DX**: `/` menu less cluttered; only relevant commands suggested.

## [1.25.0] - 2026-05-15

### Added — pentest-* skill family (25 categories, advisory/defensive)

A complete penetration testing engagement discipline family added to the
skills vault. Twenty-five new opt-in categories under the `pentest-*`
namespace, covering recon-to-report workflow with engagement-level
discipline (scope declaration, OPSEC tagging, evidence chain of custody).

**Important**: this family is **advisory/defensive only** — no live
exploit execution, no payload crafting, no C2 operation. Methodology,
analysis, planning, detection rule authoring, and reporting are the
scope. For active exploitation, use a dedicated pentest toolkit.

The family follows the discipline model from the [pentest-ai-agents]
project (MIT licensed) by adapting their engagement-level guardrails
(scope-guard, OPSEC tagging QUIET/MODERATE/LOUD, hard refusal list) into
the Badi opt-in skill vault format.

[pentest-ai-agents]: https://github.com/0xSteph/pentest-ai-agents

#### Categories (25)

**Orchestration + planning**: `pentest-orchestrator`, `pentest-engagement`,
`pentest-threat-model`, `pentest-opsec-evidence`

**Reconnaissance + assessment**: `pentest-recon`, `pentest-bugbounty`,
`pentest-ctf`

**Domain methodology** (advisory): `pentest-web`, `pentest-api`,
`pentest-bizlogic`, `pentest-ad`, `pentest-cloud`, `pentest-mobile`,
`pentest-wireless`, `pentest-cicd`, `pentest-social`, `pentest-llm`,
`pentest-privesc`, `pentest-credentials`

**Defensive engineering**: `pentest-detection`, `pentest-forensics`,
`pentest-malware`, `pentest-stig`

**Chain + report**: `pentest-exploit-chain`, `pentest-report`

#### Stack detection integration

`lib/skills/stack-map.js` extended with three new entries that activate
pentest skills when the project directory contains engagement artifacts:

- `pentest-engagement` — `scope.md` / `scope.txt` / `roe.md` /
  `findings.db` / `targets.txt` / `engagements/` / `evidence/` →
  proposes orchestrator + engagement + report + threat-model +
  opsec-evidence
- `pentest-recon-output` — `.nmap` / `.gnmap` / `.nessus` extensions →
  proposes orchestrator + recon
- `pentest-web-tooling` — `nuclei.yaml` / `nuclei-templates/` /
  `burp-project.json` → proposes orchestrator + web + api + recon

Run `badi skills detect` inside a pentest engagement directory and the
family auto-surfaces without manual configuration.

#### Tests

`tests/stack-detector.test.js` adds 6 new test cases covering the
pentest signal detection (scope file, engagements dir, findings.db,
.nmap output, nuclei config, negative case).

#### Vault size

Vault grew from 25 to 50 categories. All `pentest-*` categories follow
the standard opt-in model — zero token cost until activated via
`badi skills add pentest-<name>` or `badi skills auto-install`.

## [1.24.0] - 2026-05-14

### Added — `badi skills detect` + `badi skills auto-install` (#152)

Stack-aware skill curation inspired by
[midudev/autoskills](https://github.com/midudev/autoskills). Two new
subcommands under the existing `badi skills` family:

- `badi skills detect` — scans the project root (package.json deps,
  config file existence, file extensions, manifest keys, npm scripts)
  and lists recommended skill categories. Read-only.
- `badi skills auto-install [--yes|--dry-run]` — detect + interactive
  confirmation + opt-in activation. Idempotent; only copies categories
  not already active.

Architecture note: Badi already has a runtime auto-router
(`lib/skills-router.js`, v1.20+) that selects skills per-prompt. This
release adds an **install-time** counterpart: the auto-router answers
"which skill fits *this* prompt", while `auto-install` answers "which
skills should I keep installed *for this project*". The two layers are
independent.

Supported stacks (35+): React, Vue, Svelte, Angular, Next.js, Nuxt,
Astro, Remix, Tailwind, shadcn/ui, React Native, Expo, Flutter, Swift,
Kotlin, Express, Hono, NestJS, Fastify, TypeScript, Go, Rust, Python,
Ruby, Prisma, Drizzle, Supabase, Stripe, Clerk, Better Auth, Vitest,
Jest, Playwright, Cypress, Docker, Terraform, GitHub Actions, Vercel
AI SDK, OpenAI, Anthropic, LangChain, Shopify, WooCommerce, Resend,
Nodemailer, Remotion.

### Security/UX hardening — review fixes (#152)

In-PR review (`/review`) raised 8 findings, all closed before merge:

- Non-TTY `auto-install` without `--yes` now exits `1` (was silently
  `0`, CI bypass risk)
- "Stack detected but no matching vault category" now shows a distinct
  message instead of the misleading "all already active"
- `configDirs` is a new field separate from `configFiles`; root-level
  directory matches require `isDirectory()` and file matches require
  `isFile()` (was permissive `readdirSync` only)
- `prisma` and `.github` migrated from `configFiles` (incorrect) to
  `configDirs` (correct)
- `detectStack` caches `readdirSync(root)` once per call instead of
  per-rule (~80 → 1 syscalls on a typical scan)
- Empty `Enter` no longer auto-confirms; prompt is now `[e/H]`
  (explicit default no)
- Swift/Kotlin detection extended beyond `.swift`/`.kt` root files to
  Podfile/Package.swift/AndroidManifest.xml + `ios/`/`android/` dirs

### Tests

- Total: 727 → 768 (+41 green)
- `tests/stack-detector.test.js` (32 tests) — unit: globToRegex,
  readPackageJson, matchAnyFile, evaluateDetect, detectStack including
  10 stack scenarios + idempotency + configDirs/configFiles kind
  enforcement
- `tests/cli.skills.auto.test.js` (9 test) — subprocess: detect, dry-run,
  --yes, idempotency, non-TTY exit-1, --yes+--dry-run, no-vault-match,
  help

## [1.23.0] - 2026-05-11

### Security — XSS guard + tooltip DOM API in `badi kb graph` (#149)

`renderHtml` was embedding `JSON.stringify` output directly into an inline
`<script>` block. A markdown H1 title containing `</script>` could close
the script tag early and inject DOM. Now `<` is unicode-escaped before
embedding, and the hover tooltip uses `textContent` + DOM API instead of
`innerHTML`. Includes a regression test that asserts no unescaped
`</script>` ever lands in the output.

### Performance — `buildGraph` O(n²) → O(1) link resolution (#149)

`resolveLink` was scanning the full file list per link (`includes` + per-call
`basename` map). On 1000+ file workspaces this caused visible latency.
`buildGraph` now pre-computes `Set<file>` and `Map<basename → paths>` once,
giving constant-time lookups. Same pass also caches `safeRead` results so
each file is read exactly once instead of twice.

### Changed — review hardening (#149)

Other small follow-ups from the #147 review:

- `gh` `spawnSync` `maxBuffer` raised to 50 MB (was 1 MB default)
- `detectRepo` collapsed to a single regex with 6 variant tests
- `extractLinks` tolerates nested parens in URLs
  (`[Wiki](https://x.com/Foo_(bar))` is now a single link)
- `subSync` dry-run notice moved up under the header
- `subSync` repo fallback string: `(auto-detect failed)`

### Added — `badi kb` knowledge graph (#12 MVP)

New `badi kb` subcommand family that builds a directed graph from
markdown links and wikilinks in `.claude/`:

- `badi kb graph` — emits a self-contained HTML at
  `.claude/workspace/knowledge-graph.html` with a vanilla SVG
  force-directed layout (drag-to-pin nodes, mouse-wheel zoom, pan,
  hover tooltip). **Zero CDN dependencies** — the file works offline.
- `badi kb backlinks <file>` — lists incoming references to a file
- `badi kb orphans` — files with no inbound references
- `badi kb stats` — totals + top-5 most-referenced + per-type counts

Parses wikilinks (`[[file]]`, `[[file|alias]]`) and markdown links
(`[text](./path.md)`). External URLs (`https:`, `mailto:`, `tel:`, `#`)
and anchors are stripped. Files are classified into 11 types
(memory, knowledge, workspace, agent, command, skill, output-style,
agent-memory, daily, adr, other) for color coding.

Skip directories: `skills-vault`, `backups`, `node_modules`, `.git`,
`logs`. `knowledge-graph.html` is git-ignored.

Out of scope for MVP (tracked under #12): `--topic` filter, auto-open
browser, `--strict` broken-link exit, incremental cache.

### Added — `badi gh sync` GitHub issue → TaskBoard (#11 MVP)

New `badi gh` subcommand family. The first command, `badi gh sync`, fetches
open GitHub issues (via `gh` CLI) and merges them into
`.claude/workspace/TaskBoard.md` categorized by priority labels:

- `priority:p1-high` → `## Bugun`
- `priority:p2-medium` → `## Bu Hafta`
- `priority:p3-large`, `priority:p4-future`, unlabeled → `## Bekleyen Isler`

Idempotent — re-running adds nothing if no new issues exist. Manual tasks
in TaskBoard are preserved. Supports `--dry-run`, `--repo`, `--state`,
`--limit`. Zero runtime dependencies (uses the `gh` CLI users already have).

Out of scope for this MVP (tracked under #11): `badi gh pr draft`,
`badi gh release draft <version>`, bi-directional status sync.

### Changed — subLint refactor + subExport --write guard (#138)

`lib/commands/tasarim.js`:

- **`subLint`**: split into a pure `resolveLintExit(stdout, status) -> code`
  function plus a thin orchestrator. `process.exit` is now called at a single
  site. Behaviour unchanged; testability vastly improved.
- **`subExport --write`**: empty-on-error guard. If the underlying package
  exits non-zero or produces empty stdout, the target file is **no longer
  written**. Previously a partial/empty file could land on disk. The
  command now exits with the package's status (or `1` for empty stdout)
  and prints a clear stderr message.

### Added — Test coverage for tasarim + frontmatter (#137)

- `tests/frontmatter.test.js` (12 tests) — `parseFrontmatter` unit coverage:
  typical, CRLF tolerance, URL/ratio values, edge cases (empty string,
  unclosed frontmatter, body trim).
- `tests/cli.tasarim-lint.test.js` (14 tests) — `badi tasarim lint`/`export`
  wrapper error paths + the new `resolveLintExit` matrix (8 cases) +
  `--write` guard subprocess tests (2 cases).

Test: 642 → 668 (+26). Lint: 0 issue.

### Changed — Biome 2.4.14 → 2.4.15 (#136)

Dependabot patch bump. `biome.json` `$schema` URL aligned.
Found and fixed a latent test bug: `tests/cli.hooks-node.test.js::branch-guard`
was running with the main repo as `cwd`, so on `main` it would self-block.
Tests now use an isolated temp project with `git init -b feature/test`.

## [1.22.1] - 2026-05-11

### Fixed — Windows ESM URL scheme + chmod assertion (#126 phase 3)

The two remaining Windows test failure categories after Phase 2 were cleaned up:

**ESM URL scheme**: two dynamic imports in `tests/harness.test.js` were
being called with a Windows absolute path (`D:\...`). The Node 22 ESM loader
rejected it (`Received protocol 'd:'`). Converted to a `file://`
URL with `pathToFileURL`.

**chmod assertion**: the old test `install makes hooks +x` checked chmod mode
bits (a no-op on Windows). In Phase 2, once the hooks became
.mjs the +x bit was no longer needed; the test was updated to only check file
existence (the `statSync` import was cleaned up).

### Added — Windows install section (README) (#126 phase 4)

A new "Windows install" section was added to README.md and README.tr.md:
- `npm install -g @fatihkan/badi` + `badi init --harness claude`
- `chcp 65001` tip (Turkish character support)
- WSL auto-detection note

### Changed — Bash hooks ported to Node.js (#126 phase 2)

All 13 hook scripts were converted from `.sh` -> `.mjs`. Bash is no longer required;
Windows users can run the hooks without installing WSL/Git Bash.

**Affected files:**
- 13 new `.claude/hooks/<name>.mjs` (old `.sh` files deleted)
- `lib/hooks/util.js` — shared helpers (readStdinJson, projectRoot,
  appendLog, writeDecision, truncateLog, etc.)
- `.claude/settings.json` — commands `bash X.sh` -> `node X.mjs`
- `lib/harnesses/{claude,cursor,gemini}.js` — `.sh` -> `.mjs` or
  `.mjs|.sh` dual filter (backward compatibility)
- `tests/cli.hooks-node.test.js` — 22 new tests (smoke + contract for each
  hook). Old `tests/hooks.test.js` deleted.
- `biome.json` — 2-space override for `package.json` (npm convention)

**Behavior preserved:**
- `guard-bash`: HARD_BLOCK / SOFT_BLOCK / LOG_WARNING three-tier permission
- `branch-guard`: main/master/release/* protection
- `completeness-gate`: secret patterns (Stripe, GitHub, AWS, Slack, JWT)
  + knowledge-base TBD/TODO + memory.md 100 lines + settings.json JSON
- `dependency-audit`: 24-hour cache + lock file hash + npm/yarn/pnpm
- `session-reset`: create directory structure + counter/marker cleanup +
  log rotation
- `skill-router`: prompt -> matched skill SKILL.md body context
  injection
- log-changes / log-failures / log-stop-verdict / track-usage / pre &
  post-compact: same record/cleanup flow

**Test:** 624 -> 642 (+18 net, 22 new hook + some cross-platform fixes)
**Lint:** 0 issue.

### Added — Windows compatibility baseline (#126 phase 1)

`lib/platform.js` capabilities API + Windows Task Scheduler backend +
`badi doctor` Windows-aware section + CI `windows-latest` matrix.

**`lib/platform.js`** — single source of truth for OS detection:

```js
import { isWindows, bashAvailable, getOpener, getSchedulerKind, osSummary } from "./platform.js";
```

Exposes: `platform`, `isMac`, `isLinux`, `isWindows`, `isWsl()`,
`commandExists()`, `bashAvailable()` (handles WSL + Git Bash + native),
`getOpener()` (cmd-args tuple per platform), `getSchedulerKind()`,
`osSummary()`, `utf8Console()`, `utf8Hint()`.

**`badi doctor`** now prints OS / bash / scheduler / UTF-8 status and
warns if Windows + no bash detected (hooks won't run).

**Windows Task Scheduler** (`lib/schedulers/taskscheduler.js`) — full
adapter matching launchd/systemd shape: `id`, `platform`, `isAvailable`,
`install`, `uninstall`, `isInstalled`, `describe`. Auto-selected by
`pickScheduler()` on Windows.

**File opener** (`lib/commands/icerik/ac.js`) — Windows uses
`cmd /c start`, no longer falls back silently.

**CI matrix** — `.github/workflows/test.yml` adds `windows-latest`
alongside ubuntu/macos for Node 20.x and 22.x.

**Follow-up phases** (still in #126):

- Convert bash hooks to Node.js (so they run cross-platform without
  needing bash)
- README Windows install section
- Cmd/PowerShell hook fallback for users without WSL/Git Bash

## [1.22.0] - 2026-05-09

### Added — `badi mcp serve` Model Context Protocol server (#120)

Badi can now expose itself as an MCP server over stdio. External AI
clients (Claude Code, Cursor, Continue.dev, Claude Desktop) can plug
Badi in via `claude mcp add badi -- npx -y @fatihkan/badi mcp serve`
and call Badi tools/resources from any session.

**Tools** (read-only, safe):

- `badi.skills.route` — score matching skills for a given prompt
- `badi.skills.list` — list active opt-in skills
- `badi.skills.available` — list all skill categories in the vault
- `badi.skills.inject` — return concatenated SKILL.md bodies for a prompt

**Resources**:

- `badi://memory` — `.claude/memory.md`
- `badi://knowledge-base` — `.claude/knowledge-base.md`
- `badi://taskboard` — `.claude/workspace/TaskBoard.md`

```bash
# One-time install for Claude Code
claude mcp add badi -- npx -y @fatihkan/badi mcp serve

# Or write the snippet manually
badi mcp config > .mcp.json

# Inspect what's exposed
badi mcp tools
badi mcp resources
```

Zero external dependencies — JSON-RPC over stdio implemented in ~100
LOC; no `@modelcontextprotocol/sdk` required (which pulls express +
hono + zod). MCP spec version `2024-11-05`.

### Added — `outputstyle` and `statusline` commands (#118)

Two new opt-in commands for Claude Code harness customization:

**`badi outputstyle`** — manage `.claude/output-styles/<name>.md`
profiles that Claude Code activates with `/output-style <name>`. Three
built-ins: `terse` (concise, no fluff), `verbose` (detailed rationale +
trade-offs), `eli5` (plain-language explanations).

```bash
badi outputstyle available     # list built-ins
badi outputstyle add terse     # install profile
badi outputstyle list          # show installed
badi outputstyle remove terse
badi outputstyle clear
```

Then in Claude Code: `/output-style terse`

**`badi statusline`** — write `statusLine` config to
`.claude/settings.json` and emit the helper script under
`.claude/status-line/`. Two built-ins: `git` (branch + dirty mark),
`skill-chip` (branch + active skill count).

```bash
badi statusline available
badi statusline set git
badi statusline list
badi statusline reset
```

Restart Claude Code to see the status line update.

### Fixed — lint baseline + dep bump (#119)

- biome lint: 28 errors → 0; refactored 6 `noAssignInExpressions` to
  `for (const m of str.matchAll(regex))` (modern idiom)
- biome.json schema 2.4.8 → 2.4.14 (synced with CLI)
- @biomejs/biome 2.4.13 → 2.4.14
- vitepress 1.6.3 → 1.6.4
- Tests: 580 → 595 (15 new for outputstyle + statusline)

### Notes

- Component count: 79 commands (was 77), 22 agents, 13 hooks, 25 skills
- Three open `npm audit` moderate (transitive `esbuild` via vitepress)
  remain — upstream "no fix available", dev-only

## [1.21.0] - 2026-05-03

### Added — `seo-crawl-budget` skill (#109)

New opt-in skill: a methodology for getting indexed in 6-24 hours for
low-competition long-tail keywords. A 20-article campaign, a cyclic
internal-link matrix, manual Search Console triggering, a 6-phase structure
(keyword generation, brief templates, link matrix, publication schedule,
GSC actions, tracking metrics).

Adapted from the [moneyvadi-prog/crawl-budget-manipulation](https://github.com/moneyvadi-prog/crawl-budget-manipulation)
(MIT, Gulsah Arslan) repo.

Integrated with the v1.20 auto-router: when `badi skills auto on` is active,
the SKILL.md body is automatically injected when TR/EN triggers like
"crawl budget", "long-tail", "search console", "indexleme",
"internal linking" appear in the prompt.

#### Usage — Automatic mode (recommended)

```bash
# One-time setup
badi skills auto on
```

After that you ask directly inside Claude Code:

```
You ▸ My new blog posts aren't getting indexed, how do I manage crawl budget?
[Badi auto-router]
  - seo-crawl-budget (score 12) — triggers: crawl budget, indexing
Claude ▸ {SKILL.md embedded} Planning a 20-article campaign...
```

#### Usage — Manual mode

```bash
badi skills available | grep seo-crawl-budget   # available in the list
badi skills add seo-crawl-budget                  # permanent opt-in
badi skills list                                  # see active skills
```

#### Usage — One-off (without router, without opt-in)

```bash
badi skills route --inject "long-tail keyword indexing problem"
# writes the SKILL.md body to stdout — you can pipe it to another tool
```

#### What the skill produces

When active, the agent suggests / provides the following files as templates:

```
seo-campaign-<slug>/
├── keywords-A.json          # 10 published evenly
├── keywords-B.json          # 10 scheduled
├── briefs/                  # 20 article briefs
├── linking-matrix.md        # Cyclic internal-link graph
├── publication-schedule.csv # Date + time
├── search-console-checklist.md
└── tracking-template.md     # 14-28 day metrics
```

**New files**: `.claude/skills-vault/seo-crawl-budget/SKILL.md`.
Test: `tests/cli.skills-router.test.js` TR + EN trigger match
cases (3 new tests, 583 total).

Skill category count: 24 -> 25.

## [1.20.0] - 2026-05-02

### Added — auto skill router (`badi skills route` + `auto on/off`)

Prompt-based automatic skill activation. A keyword index is built from the
SKILL.md descriptions in the vault, and scored against the user prompt.
When integrated with the UserPromptSubmit hook, the body of the skills matching
the prompt type is injected into the context
(per-turn, no filesystem writes).

```
badi skills route "add schema markup for SEO"     # score the matches
badi skills route --inject "..."                   # writes the SKILL.md body
badi skills auto on                                # activate the hook
badi skills auto off                               # turn off
badi skills auto status                            # status
```

**Score formula**: a trigger token (`triggers on:` list) match is weighted 3x,
a description token match 1x. The threshold defaults to `minScore=2`,
top defaults to 3 skills. A stopword filter (TR + EN) reduces false
matches.

**Auto-router opt-in**: `badi skills auto on` adds a UserPromptSubmit
hook to settings.json. The hook reads the prompt and passes the SKILL.md
body of the matching skills from the vault to Claude as additionalContext.
The token tax is only when there is a match — if the prompt is shorter
than 3 words or there is no match, the hook passes silently.

**New files**: `lib/skills-router.js` (matcher),
`.claude/hooks/skill-router.sh` (hook), `tests/cli.skills-router.test.js`
(22 tests). doctor also checks the `skill-router.sh` hook.

### Added — `badi market gaps` (#84 phase 2)

New `gaps` subcommand cross-references the existing market signals into
ranked opportunity findings:

```
badi market gaps 284882215
badi market gaps 284882215 --query "facebook alternative" --json
```

For every cross-competitor complaint code the helper computes a
`gapScore = coverage% * volume * (1 - difficulty/100)`, optionally
boosted by Reddit demand when `--query` is supplied. Each finding
carries a coverage percentage (how many competitors carry the
complaint), volume (negative review count), severity (`HIGH` /
`MEDIUM` / `LOW`), and a one-line human-readable rationale.

Output sorted by `gapScore` descending — top entries are the strongest
signals for "this market segment has a broadly under-served need".
`--json` mode emits the structured report (target + difficulty +
demand + ranked findings).

This closes the third Phase-2 capability for #84. SensorTower revenue
(#84-1) stays blocked on paid API access; `gaps` does not depend on it.

## [1.19.0] - 2026-05-02

### Added — `badi market wishlist` (#84 phase 2)

New subcommand for the demand × supply matrix:

```
badi market wishlist "habit tracker"
badi market wishlist "ai journaling" --json --days 60
```

Demand signal: Reddit anonymous JSON search (no API key, last 30 days
by default). Supply signal: App Store search result count for the
same query. The matrix maps four quadrants: `BLUE_OCEAN` (high demand
× low supply), `COMPETITIVE`, `NICHE`, `SATURATED`.

`--json` flag emits a structured report with the demand sample (top
posts + subreddits) and supply sample (top apps with rating + count)
for piping into other tools.

Implements the second of three Phase-2 capabilities. SensorTower
revenue (#84-1) stays blocked on paid API access; opportunity gaps
cross-analysis (#84-3) follows up.

## [1.18.0] - 2026-05-02

### Added — agent frontmatter audit (#90)

Each of the 21 agents now carries an explicit `permissionMode: default`.
The 15 read-only / advisor agents (archaeologist, api-designer,
architecture-advisor, debt-collector, error-whisperer, migration-pilot,
onboarding-sherpa, performance-profiler, pr-ghostwriter,
refactoring-advisor, rubber-duck, security-scanner, test-strategist,
unsticker, yak-shave-detector) now declare
`disallowedTools: [Write, Edit, NotebookEdit]` for defense-in-depth
under Claude Code 2.1.119+ headless/`--print` execution where these
fields are honored.

A new `tests/agent-frontmatter.test.js` validates the policy on every
run (122 assertions covering all 21 agents).

### Added — `badi tasarim` Phase 2 (#85)

- New `tasarim-kurator` agent — interactive DESIGN.md producer that
  questions brand identity, color psychology, typography character,
  and component decisions, then writes a rationale-rich DESIGN.md
- New `design-tokens` skill (vault) — when active, agents producing
  UI/components/visuals consult the project's DESIGN.md frontmatter
  for canonical tokens instead of inventing ad-hoc values
- `visual-director` agent now declares DESIGN.md delegation: reads
  tokens via `design-tokens`, surfaces brand-drift warnings, hands
  off new color/typography decisions to `tasarim-kurator`

The skill is opt-in (per v1.17 model). Activate via:
```
badi skills add design-tokens
```

## [1.17.0] - 2026-04-29

### Changed — opt-in skills (BREAKING)

Skills are no longer auto-loaded. All 23 skill categories now live in
`.claude/skills-vault/` (Claude Code does NOT scan this directory).
The active `.claude/skills/` folder starts empty; users opt in to
exactly the skills they want via the new `badi skills` command.

**Why:** auto-loading 23 skill categories cost ~10–15k tokens per
turn even when none were used. Opt-in cuts that to zero by default
and lets users dial in the cost they want to pay.

**Plugin path also drops the `skills` field** in `plugin.json`. Plugin
users get the same opt-in flow once they install the npm CLI for
`badi skills` — the plugin itself stops shipping skills entirely so
no auto-load tax for plugin consumers either.

### Added — `badi skills` command

```
badi skills                  # status table + interactive picker
badi skills available        # list every skill in the vault
badi skills list             # list active skills
badi skills add <name…>      # opt one or more in
badi skills remove <name…>   # opt one or more out
badi skills clear            # reset all active skills
badi skills reset            # alias for clear
```

Interactive mode (TTY only) shows a numbered checklist and accepts
selections like `1,3,5-7`, `all`, or `none` to set the active set in
one operation.

### Migration notes

Existing installs are protected: the update path treats
`.claude/skills/` as user-customizable, so anything already there
stays intact. To migrate to the lean default, run
`badi skills clear` and then `badi skills add <only what you need>`.
Token analyzer (`badi ai token`) reports vault size separately as
"Vault (yuklenmez)" so you can see the savings.

## [1.16.5] - 2026-04-29

### Added — plugin-level safety hooks

The Claude Code plugin path now ships two universal safety hooks
inline in `plugin.json`, so `/plugin install badi@badi-marketplace`
gives users the same Bash-level protections as the npm path:

- **`guard-bash.sh`** (PreToolUse, matcher `Bash`) — blocks
  destructive command patterns (`rm -rf /`, force-push to
  main/master, `chmod 777`, `curl | bash`, secret exfiltration,
  `dd of=/dev/`, etc.). Three-tier system: hard-block, prompt,
  allow.
- **`branch-guard.sh`** (PreToolUse, matcher `Bash`) — refuses
  direct `git commit` on `main` / `master` / `production`, and
  `git push --force` on `main` / `master` / `release/*`. Pushes
  to feature branches pass through.

Hooks reference scripts via `${CLAUDE_PLUGIN_ROOT}/.claude/hooks/`
so they resolve correctly from the plugin cache regardless of the
user's project layout.

### Skipped on purpose

Project-state hooks (`pre-compact-handoff`, `post-compact-resume`,
`session-reset`, `track-usage`, `log-*`, `backup-before-write`,
`completeness-gate`, `dependency-audit`) stay npm-only. They depend
on a writable project-local `.claude/` tree (memory.md,
TaskBoard.md, logs/, backups/) that the plugin cache can't provide.
Plugin users who want the full hook suite still get it via
`npx @fatihkan/badi init`.

## [1.16.4] - 2026-04-29

### Added — Claude Code plugin distribution

Badi can now be installed as a Claude Code plugin, in addition to the
existing npm path:

```
/plugin marketplace add fatihkan/badi
/plugin install badi@badi-marketplace
```

Two new manifests under `.claude-plugin/`:

- `plugin.json` — declares 21 agents (`./.claude/agents/*.md`),
  commands (`./.claude/commands`), and skills (`./.claude/skills`) via
  custom paths, so the existing `.claude/` tree stays canonical with
  no duplication. The `agents` field rejects directory paths in the
  current schema, so all 21 agent files are listed individually.
- `marketplace.json` — single-plugin marketplace entry under
  `badi-marketplace` (category: productivity, strict: true).

Both manifests pass `claude plugin validate`. End-to-end smoke-tested
from a real Claude Code session: `marketplace add fatihkan/badi` →
`install badi@badi-marketplace` lands version 1.16.3 in the plugin
cache with the full agent/command/skill tree intact.

`.claude-plugin/` was added to `package.json#files` so the npm tarball
also carries the manifests — a single checkout serves both
distribution channels.

### Distribution scope

| Component | Plugin (`/plugin install`) | npm CLI (`npx @fatihkan/badi init`) |
|-----------|:--------------------------:|:----------------------------------:|
| 21 agents | ✓ | ✓ |
| 77 slash commands | ✓ | ✓ |
| 23 skill categories | ✓ | ✓ |
| 12 automation hooks | — | ✓ |
| Multi-harness compiler (Cursor / Gemini CLI) | — | ✓ |
| `badi` CLI toolchain (icerik / market / tasarim / publish / schedule / list / stats / doctor) | — | ✓ |

Hooks stay npm-only because they need a project-local writable
`.claude/` tree the plugin cache cannot provide (plugin contents are
read-only and cached separately from the user's project).

## [1.16.3] - 2026-04-29

### Changed — discoverability

English-first metadata across npm and GitHub surfaces, plus explicit
references to Anthropic Claude Opus 4.7 and Sonnet 4.6 to broaden
search visibility.

- `package.json` description rewritten in English with brand + model
  versions: "Workflow management CLI for Claude Code, Cursor, and
  Gemini CLI — 21 AI agents, 77 commands, OWASP security scans. Built
  for Anthropic Claude Opus 4.7 and Sonnet 4.6."
- `package.json` keywords curated for SEO: dropped narrow terms
  (`turkish`, `business-operating-system`); added `anthropic`,
  `claude`, `claude-opus`, `claude-sonnet`, `ai-agents`, `subagents`,
  `cli`, `developer-tools`, `security-scanner`, `owasp`,
  `code-review`, `cursor`, `gemini-cli` (20 total).
- README hero (EN + TR) rewritten to lead with Anthropic + Claude
  Opus 4.7 / Sonnet 4.6 + multi-harness support.
- GitHub repo description and topics updated via API: 6 low-value
  topics removed (`typescript`, `nodejs`, `npm-package`,
  `workflow-automation`, `open-source`, `security-scanning`),
  11 high-value topics added (full set above).

### Fixed — count drift

CLAUDE.md and README files were citing inconsistent component counts.
Cross-checked against the filesystem and standardized on:

- 21 agents (`.claude/agents/*.md`)
- 77 commands (`.claude/commands/*.md`)
- 12 hooks (`.claude/hooks/*.sh`)
- 23 skill categories (`.claude/skills/*/`)
- 398 passing tests

CLAUDE.md hero was reading "50 commands, 21 skill categories"; README
table was citing "25 skill categories" / "395 tests".

### Fixed — node badge

README.md and README.tr.md badge said `node >=18` but `engines.node`
has been `>=20.11.0` since v1.13. Badge now matches.

## [1.16.2] - 2026-04-26

### Fixed — security

8 open CodeQL alerts (1 high-severity error + 7 warnings) closed in
a single PR. Highlights:

- `domain.js`: strict-first TLS connection — only fall back to
  insecure mode on known cert-verify errors (expired, self-signed,
  hostname mismatch). Returned data includes `validForDomain` flag
  surfaced as a new check line in `badi ssl-check` output.
- `skills-bundler.js#formatScalar`: now escapes backslash before
  double-quote — Windows paths and regex-literal-bearing strings
  round-trip losslessly.
- `seo.js#countWords`: switched from regex-based script/style strip
  (CodeQL kept finding bypasses across nesting / attribute-bearing
  closings) to `node-html-parser` DOM tree. New runtime dependency
  but eliminates an entire class of sanitization issues.
- `plugin.js`: `source.includes("github.com")` replaced with
  `new URL(source).hostname` exact match — `evil.com/?github.com=1`
  no longer counted as a GitHub URL.
- `.github/workflows/test.yml`: `permissions: { contents: read }`
  added (other 4 workflows already had explicit permissions).

New rule recorded in `.claude/memory.md`: HTML processing for
sanitization or text extraction must use a parser, not regex.

### Fixed — `badi tasarim`

Manual smoke-test of `badi tasarim` (v1.16.0) revealed three usability
issues. All three closed in this hotfix; tests 395 → 397.

- **Lint exit code now reflects findings.** `badi tasarim lint` was
  returning exit 0 even when the underlying `@google/design.md` lint
  reported errors — silent failure for CI/automation. Now the JSON
  output is parsed; exit 1 if `summary.errors > 0`.
- **`--write` flag for export.** `badi tasarim export` previously
  had no way to direct output to a file (stdout only). Added
  `--write <path>` for explicit file output.
- **Help text clarified.** `--out` documentation now states that it
  is the DESIGN.md file path (init: write target; lint/show/export:
  source). The previous "Cikti dosya yolu" wording mis-implied a
  destination across all subcommands.

Two upstream issues remain (filed as GitHub issues, not addressable
in this repo):
- `@google/design.md@0.1.1` lint emits `raw.match is not a function`
  on the default `badi tasarim init` skeleton.
- `@google/design.md@0.1.1` export returns empty token categories
  for that same skeleton — same parse defect.

## [1.16.0] - 2026-04-26

### Added — `badi tasarim`

Visual identity command. Closes #58. Wraps Google's `@google/design.md`
CLI (pinned to `0.1.1`, called via `npx`) for lint/export and provides
local init/show.

Subcommands:
- `badi tasarim init [--ornek <name>] [--out PATH] [--force]` — write a
  fresh `DESIGN.md` with frontmatter tokens (colors / typography /
  spacing / radius / elevation / components) + 8-section prose
  skeleton. `--ornek paws-and-paths|atmospheric-glass|totality-festival`
  writes a stub pointing to the upstream example.
- `badi tasarim lint [--strict]` — validates `DESIGN.md` via
  `@google/design.md lint`.
- `badi tasarim export --format tailwind|dtcg [--out PATH]` — exports
  tokens to a Tailwind config or DTCG JSON.
- `badi tasarim show [--tokens|--prose]` — prints frontmatter only,
  prose only, or both (default).

Default location: `.claude/workspace/DESIGN.md`. Internet required for
the first `npx` invocation; subsequent runs use cache.

### Tests

379 → 395 (+16). 7 suites covering help, init (default + --ornek +
--force + --out), show (--tokens/--prose/default/missing-file),
export validation, and unknown-subcommand handling.

## [1.15.3] - 2026-04-26

Documentation polish. No code or test changes.

## [1.15.2] - 2026-04-26

Documentation polish. No code or test changes.

## [1.15.1] - 2026-04-26

Documentation polish. No code or test changes.

## [1.15.0] - 2026-04-26

### Added — `badi market`

App Store market-research command. MVP scope: competitor discovery + multi-
region review aggregation + 11-code complaint categorization + difficulty score
(BLUE_OCEAN / COMPETITIVE / HARD / SATURATED).

Subcommands:
- `badi market <appId>` — full report (5-stage pipeline)
- `badi market discover <appId>` — competitor discovery (genre filter)
- `badi market reviews <appId>` — multi-region multi-page reviews +
  rating distribution + complaint categories
- `badi market difficulty <appId>` — 0-100 score + categorical verdict

Flags: `--country <c1,c2,...>`, `--pages <N>`, `--limit <N>`. No API keys
required (iTunes Lookup, iTunes Search, Apple RSS endpoints).

New `lib/market-helpers.js`: `discoverCompetitors`,
`aggregateReviewsMultiRegion`, `categorizeReviewIssue` (11 codes,
TR+EN stem patterns), `summarizeReviews`, `calculateDifficulty`,
`findCrossCompetitorComplaints`. Reuses existing `aso-helpers.js` for
the iTunes API surface.

### Phase 2 (separate issues, not in 1.15.0)

- SensorTower revenue scrape (real $/downloads)
- Wishlist demand×supply matrix (`✓✓✓ missing` / `✓✓ buggy` / `✓ present but
  bad` / `✓✓✓ present` notation)
- Opportunity gap report (complaint ∩ wishlist intersection + verbatim
  quotes)
- `--format json` for piping
- Region-aware difficulty (US vs TR vs JP)

### Tests

359 → 379 (+20). 7 helper unit-test suites + 4 CLI smoke tests.

## [1.14.1] - 2026-04-26

CI / engines bump. No runtime-code changes.

### Changed
- **CI matrix dropped Node 18.** `tests/cli.*.test.js` (14 files) and
  `scripts/enrich-skills.js` use `import.meta.dirname`, which requires
  Node 20.11+. The Node 18 row in CI had been failing silently for
  these files; bumping the matrix surfaces the rest of the suite cleanly.
- **`engines.node` bumped to `>=20.11.0`** (was `>=18.0.0`). Aligns
  package metadata with the actual runtime requirement of the test suite.
- README install instructions updated to require Node 20.11+.

### Notes
- Production CLI surface (`bin/badi.js` and `lib/`) already used the
  cross-version-safe `fileURLToPath(import.meta.url)` pattern; users
  on Node 20+ are unaffected. Node 18 users were never able to run the
  test suite, but the CLI itself worked.

## [1.14.0] - 2026-04-26

**Skill ecosystem MVP.** Closes issue #56 (skill-bundle infrastructure)
and #57 (badi-discipline behavioral skill).

### Added — Skill bundle pipeline

- `lib/skills/schema.js` — Badi skill bundle frontmatter validator
  (required/optional fields, allowed-tools allowlist, semver-shaped
  badi-version, known categories). Two modes: warn (default) and
  strict (CI). Includes `parseRichFrontmatter` for nested metadata
  blocks.
- `lib/harnesses/skills-bundler.js` — compiles `.claude/skills/<name>/`
  into `<target>/skills/<name>/` with auto-enriched frontmatter,
  copies `references/` subdirs, and generates a router skill grouping
  bundled skills by category.
- `badi publish --skill-bundle [--target <dir>] [--source <dir>]
  [--strict] [--dry-run]` — wires the bundler into the publish
  orchestrator. The actual git/npm push for the badi-skills repo is
  out of scope; the orchestrator stops after writing the bundle and
  prints next-step git commands.
- `scripts/enrich-skills.js` — in-place enrichment for the source
  `.claude/skills/<name>/SKILL.md` files. Idempotent. Reads
  `scripts/skill-descriptions.json` for curated trigger-rich
  descriptions.
- All 23 top-level `.claude/skills/<name>/SKILL.md` files now carry
  the full Badi skill bundle frontmatter (name, description, license,
  compatibility, allowed-tools, metadata.{author, homepage,
  badi-version, category}).

### Added — `badi-skills` bootstrap kit

- `_bootstrap/badi-skills/` — bare repo skeleton for the separate
  `badi-skills` GitHub repo. LICENSE (MIT), README install
  instructions for skills CLI / Claude Code marketplace / Cursor
  Remote Rule / OpenAI Codex, CLAUDE.md and AGENTS.md routing,
  `.claude-plugin/plugin.json` and `.cursor-plugin/manifest.json`
  manifests, `.github/workflows/validate.yml` CI that pulls the
  schema from this repo and validates every SKILL.md in strict mode.
- `_bootstrap/README.md` walks through one-time repo creation:
  `gh repo create` → push bootstrap → `badi publish --skill-bundle`
  → copy bundle → tag `v1.0.0`.

### Added — `badi-discipline` behavioral skill

- `_bootstrap/badi-skills/skills/badi-discipline/SKILL.md` — 8
  principles: Think Before Coding, Simplicity First, Surgical
  Changes, Goal-Driven Execution, Yak-Shave Detection, TaskBoard
  Discipline, Knowledge-Base Source Requirement, and Destructive
  Action Gate (extracted from Badi's own subagents and hooks).
- 4 progressive-disclosure references: task-discipline,
  destructive-actions, yak-shave-patterns, knowledge-base-sources.
- Carries an explicit tradeoff note: "These are prompt-level
  guidelines, not enforcement. Bias toward caution over speed; for
  trivial tasks, use judgment."

### Tests

- 307 → 359 (+52). Schema 32, bundler 17, publish 3.

## [1.13.2] - 2026-04-25

Code-review follow-up to v1.13.1. Seven findings from the post-merge review
addressed in a single hotfix.

### Fixed
- **Local-time "today" calculation.** `badi icerik durum` and `badi icerik kapat`
  were comparing the today count between `mtime.toISOString()` (UTC) and the local
  `getDateString()`. When the UTC boundary differed from the local boundary (example: UTC+3
  at 01:00), "today" was counted wrong. Now the comparison is done with the local
  `startOfToday`.
- **`runTemplate` switch** was protected against possible type drift with `default: throw`.
  If the `TEMPLATE_TYPES` list is out of sync with the switch it fails early
  (a generated file is not written with empty content).
- **`badi icerik` unknown-subcommand message** now shows not only the template types
  but also the session commands (`list`, `basla`, `durum`, ...).

### Refactored
- The `lib/commands/icerik.js` shim was removed; `bin/badi.js` now
  imports `lib/commands/icerik/index.js` directly (unnecessary
  indirection deleted).
- The `subInstall` comment in `lib/commands/agent.js` explains
  *why* the shell-watcher confirmation is required.

### Tests
- 307/307 green — behavior preserved.

## [1.13.1] - 2026-04-25

### Fixed — `badi agent`
- **`agent install` confirmation was fake.** When a watcher contained a `shell`
  check, install printed "shall I register this with the scheduler?" and then
  installed anyway, without reading any input. The prompt now actually waits
  for `y/N`. Added `--yes` / `-y` to opt out for non-interactive use.
- **`agent install <unknown>` and `agent tail <unknown>` threw raw stack
  traces** from `WatcherError`. Both now print a clean `Watcher not found: <path>`
  message and exit `1`.

### Refactored
- `lib/commands/icerik.js` was a single 1,667-line if-chain over 13
  subcommands. Split into per-subcommand modules under `lib/commands/icerik/`
  (issue #41). `icerik.js` was a one-line re-export shim — kept for
  backwards compatibility with imports, removed in 1.13.2.

### Tests
- 304 → 307 (+3): `--yes` flag visible in help, clean errors for missing
  watchers in `install` and `tail`. The +3 came entirely from the `agent.js`
  polish; the icerik refactor itself was test-equivalent.

## [1.13.0] - 2026-04-24

### Added — Background Agents (issue #55)

Badi now has a **background watcher** system. Users can define YAML-frontmatter watchers in `.claude/watchers/` that run on an OS-native scheduler and produce reports consumed by the next `/start` session.

New commands:
- `badi agent create <name> [--template project-health|deploy-watchdog]` — scaffold a watcher.
- `badi agent list` — enumerate installed watchers + scheduler state.
- `badi agent run <name>` — execute once manually (great for dev).
- `badi agent install <name> [--scheduler launchd|systemd|cron] [--dry-run]` — register with the best available OS scheduler.
- `badi agent uninstall <name>` — remove scheduler registration (keeps the `.md`).
- `badi agent tail <name> [-n N]` — stream the watcher's report.
- `badi agent status [--since 24h|7d] [--format text|json]` — aggregate recent alerts across all watchers.
- `badi agent remove <name>` — full cleanup (scheduler + watcher file).

### 5 built-in watch types

| Type | Checks |
|------|--------|
| `git` | `git log` scoped by `last-N-commits` / `since:<ref>` / `all`, regex pattern match |
| `shell` | arbitrary command + `alert_on: exit-nonzero / stdout-match:<re> / stderr-match:<re>` + timeout |
| `file` | file change detection (mtime+size) and `package.json` dependency-added/removed |
| `log` | offset-tracked tail with `new-entry` or `pattern-match:<re>` |
| `http` | HEAD/GET with SSRF guard, `status-nonok`, `latency>Ns`, `body-match:<re>` (composite via `|`) |

### 3 OS scheduler adapters

- **launchd** (macOS) — `~/Library/LaunchAgents/com.badi.watcher.<name>.plist` + `launchctl bootstrap/bootout`.
- **systemd** (Linux) — `~/.config/systemd/user/badi-watcher-<name>.{service,timer}` + `systemctl --user enable --now`.
- **cron** (universal fallback) — marked lines in user crontab, cleanly added/removed.

`pickScheduler()` auto-selects by platform; `--scheduler` flag overrides; `--dry-run` shows the plan + full unit content without writing.

### /start integration

The `/start` slash command now calls `badi agent status --since 24h` before the daily briefing. Alerts from the last 24h show up in the Brifing block as a "Watcher" line, and Claude will offer to investigate.

### Templates shipped

- `.claude/watchers/project-health.md` — git + npm test + package.json + failures log (15m).
- `.claude/watchers/deploy-watchdog.md` — http health + deploy error log (5m, `active: false` by default — edit URL and flip to `true`).

### Technical
- New `lib/watchers/{index,parse}.js` + `lib/watchers/types/{git,shell,file,log,http}.js`.
- New `lib/schedulers/{index,launchd,systemd,cron}.js`.
- New `lib/commands/agent.js` + wired into `bin/badi.js`.
- 38 new tests in `tests/watcher.test.js` (parse 13, types 10, schedulers 5, runWatcher e2e 2, agent CLI 6 + misc).
- Total suite: **304/304 green** (266 → 304).

### Security notes
- Watchers with `type: shell` execute arbitrary commands. The install flow prints a warning in TTY mode. Treat `.claude/watchers/*.md` as trusted code — do not install third-party watchers blindly.
- `http` type uses `validateUrl()` (helpers.js) for SSRF protection — localhost / private IPs blocked.
- Scheduler unit files are installed at user scope (no sudo). `cron` edits only the user's crontab.

## [1.12.1] - 2026-04-24

Post-release review hotfix. Addresses all 10 findings from the v1.12.0 code review in a single PR.

### Fixed — High

- **Test isolation** — `preferences.js` now honors `BADI_PREFS_HOME` (base dir) and `BADI_PREFS_PATH` (full override). Previously the env var passed by the test harness was a no-op, which meant any test that forgot `--no-save` would write to the real `~/.config/badi/preferences.json`.
- **Non-TTY silent save** — `badi init` no longer writes `defaultHarness` to preferences when running headless without a `--harness` flag. CI pipelines can no longer accidentally pin a harness choice.

### Fixed — Medium

- **Cursor content transform** — Every Cursor-bound file (`.cursor/commands/*.md` and `.cursor/rules/badi-main.mdc`) now opens with a preface that warns about Claude-specific path references (`.claude/hooks/`, subagents, skills). Bodies are still preserved 1:1; the preface is idempotent (re-running install does not stack it).
- **Interactive menu testability** — `parseMenuAnswer()` extracted from `selectHarnessInteractive()` as a pure function and exported. 9 offline tests cover number / id / Enter / invalid / out-of-range / negative / unknown-id paths.

### Fixed — Low

- **Cursor rule header** — dropped the redundant `globs: **/*` line (irrelevant when `alwaysApply: true`).
- **`setPreference` validation** — `defaultHarness` values are validated against the harness registry before writing to disk.
- **Case-insensitive `--harness`** — `badi init --harness CURSOR` / `Gemini` / `ALL` now resolve correctly.
- **Brittle test assertions** — hard-coded `>10` / `>30` thresholds replaced with checks against actual source counts (`readdirSync(SRC/commands).length`, `pass + warn + fail === checks.length`).
- **Cursor directory counting** — removed duplicate `result.created++` for `.cursor/` so the install summary reflects real file/dir creation counts.

### Technical
- `lib/preferences.js` — env-var support + `VALIDATORS` registry.
- `lib/commands/init.js` — `parseMenuAnswer` exported; `selectHarnessInteractive` now delegates parsing; non-TTY branch disables `saveDefault`.
- `lib/harnesses/cursor.js` — `transformCommand()` exported; rule header trimmed; `.cursor/` creation accounting fixed.
- `lib/harnesses/index.js` — `resolveHarnesses` lowercases inputs before lookup.
- `tests/harness.test.js` — 15 new tests (case-insensitive resolve, preface idempotency, rule header shape, `parseMenuAnswer` 9 cases, env-var isolation 2 cases). Total: 266/266 green.

## [1.12.0] - 2026-04-24

### Added — Multi-harness support (issue #54)

Badi now supports multiple LLM CLIs. `badi init` can generate assets for Claude Code, Cursor, or Gemini CLI.

- **Interactive selector** — running `badi init` without flags shows a harness selection menu.
- **`--harness <id>`** — non-interactive install: `claude`, `cursor`, `gemini`, `all`, or comma-separated (`claude,cursor`).
- **Harness-aware update/doctor** — `badi update` and `badi doctor` auto-detect the installed harness and target it.
- **Preferences** — `~/.config/badi/preferences.json` stores `defaultHarness` (the next `badi init` will suggest it). `--no-save` opts out.

### Harness matrix

| Harness | Rules | Commands | MCP | Subagents | Hooks | Skills |
|---------|:-----:|:--------:|:---:|:---------:|:-----:|:------:|
| Claude Code | `CLAUDE.md` | 77 | `.mcp.json` | 21 | 12 | 23 |
| Cursor | `.cursor/rules/badi-main.mdc` | 77 | `.cursor/mcp.json` | — | — | — |
| Gemini CLI | `GEMINI.md` (merged) | inline | `.gemini/settings.json` | — | — | — |

On Cursor 12 hooks + 23 skills are skipped; on Gemini 21 subagents + 77 slash commands are skipped as well — the `badi init` output lists them in a `skippedComponents` report.

### Technical
- New `lib/harnesses/` dir: `claude.js`, `cursor.js`, `gemini.js`, `index.js` (registry + `resolveHarnesses` + `detectHarness`).
- New `lib/preferences.js` — `~/.config/badi/preferences.json` read/write.
- `lib/commands/init.js`, `update.js`, `doctor.js` — refactored to dispatch through the adapter registry. Claude-only behavior preserved when no other harness detected.
- 44 new tests in `tests/harness.test.js` (7 suites: registry, 3 adapters, init/update/doctor CLI flows). Full suite: 251/251 green.

## [1.11.0] - 2026-04-22

### Added — Content Types
- **`badi icerik newsletter [topic]`** — weekly email newsletter scaffold (subject A/B, preview text, hook, CTA, footer, HTML config).
- **`badi icerik podcast [topic]`** — episode notes + show notes skeleton (hook, draft flow by timecode, transcript scaffold, platform metadata, social promo clips).
- **`badi icerik thread [topic]`** — 10-post X/LinkedIn thread (hook → problem → story → 3 key points → counterargument → takeaway → application → CTA, plus engagement strategy).
- **`badi icerik case-study [topic]`** — customer success story with one-liner, headline metric, problem/solution/results table, testimonial, distribution plan.
- All four types support TR + EN via `--lang tr,en`.

### Added — ASO extensions
- **`badi aso playstore <app-id>`** — Google Play audit using existing `lookupPlayStore` scraper.
- **`badi aso reviews <app-id>`** — fetches real iTunes RSS reviews, runs keyword-based sentiment classifier (positive / negative / bug / feature_request / neutral), surfaces top 5 critical reviews + top 5 feature requests.
- **`badi aso screenshots <app-id>`** — app-specific asset dump: orientation split, resolution distribution, sample URLs, sizing advice. The old `badi aso screenshots` (no id) still shows the generic sizing guide.

### Added — SEO extensions
- **`badi seo backlinks <domain>`** — best-effort, free tooling only. Combines a DuckDuckGo mention search (site-exclude filter) with Wayback Machine CDX snapshot presence. Clearly labelled as directional, not a full backlink profile.
- **`badi seo rank <domain> <keyword>`** — DuckDuckGo organic rank check (top ~30 results), highlights position when the domain is found.
- **`badi seo compare <url1> <url2>`** — side-by-side SEO audit across HTTPS, title/meta lengths, OG tags, canonical, heading structure, image alt coverage, word count, schema, HTML size, script count, compression.

### Added — Mobile extensions
- **`badi mobile crash-setup <framework> <provider>`** — Sentry or Crashlytics scaffolding for React Native, Flutter, iOS, and Android with paste-ready config snippets.
- **`badi mobile deeplink [domain|scheme://]`** — validates URL schemes (RFC 3986), fetches `apple-app-site-association` + `assetlinks.json`, reports appIDs, package names, SHA256 fingerprint previews, and tester URL commands.
- **`badi mobile ota [codepush|expo]`** — step-by-step OTA update setup for App Center CodePush (with deprecation notice) and Expo EAS Update (configure → release → rollback).

### Added — Publish Orchestrator
- **`badi publish`** — release orchestrator that runs the full sequence in one command: clean-git check → branch verification → CHANGELOG gate → `package.json` version bump (also `package-lock.json` if present) → commit → tag → push main + tag → `gh release create --generate-notes` → `npm publish --access public`.
- Flags: `--version patch|minor|major`, `--dry-run`, `--skip-npm`, `--skip-github`, `--skip-changelog`, `-m/--message`.
- **`badi publish check`** — pre-flight readiness: git cleanliness, branch, package metadata, CHANGELOG presence, `gh` CLI, `npm whoami`.

### Technical
- `lib/templates/tr.js` + `lib/templates/en.js` — 4 new template functions (newsletter, podcast, thread, caseStudy) per language.
- `lib/aso-helpers.js` — new exports: `fetchAppStoreReviews`, `analyzeSentiment`, `parseScreenshotUrl`.
- `lib/commands/seo.js` — DuckDuckGo uses POST + browser UA so results actually come back (GET returns a shell page).
- `lib/commands/publish.js` — new ~330-line orchestrator module.
- 27 new tests across icerik, aso, seo, mobile, publish test files (total suite 202/202 green).
- No new runtime dependencies.

## [1.10.0] - 2026-04-22

### Added — Frontend Taste (Premium UI Skills)
- **9 bundled design variants** under `.claude/skills/frontend-taste/` — stops Claude Code from producing generic "AI-looking" UI:
  - `default` (design-taste-frontend) — all-rounder with DESIGN_VARIANCE / MOTION_INTENSITY / VISUAL_DENSITY dials
  - `gpt-taste` — strict editorial, AIDA structure, mandatory GSAP ScrollTriggers
  - `minimalist` (minimalist-ui) — Notion / Linear editorial feel
  - `brutalist` (industrial-brutalist-ui) — Swiss typography + raw grid
  - `soft` (high-end-visual-design) — high-end agency feel, spring motion
  - `redesign` (redesign-existing-projects) — audit + fix existing UIs
  - `output` (full-output-enforcement) — anti-truncation, stacks with others
  - `stitch` (stitch-design-taste) — Google Stitch `DESIGN.md` generator
  - `images-first` (image-taste-frontend) — reference-led visual workflow
- **`badi taste` CLI** — inspect, show, prompt, status the bundled variants
  - `badi taste` — list all 9 variants with usage hints
  - `badi taste show <id>` — print a variant's full SKILL.md
  - `badi taste prompt <id>` — show an example Claude Code trigger prompt
  - `badi taste status` — verify 9/9 variants installed
- Trigger in Claude Code by naming the variant in the prompt (e.g. "Use the frontend-taste/brutalist skill.")

### Technical
- `lib/commands/taste.js` — new command module (~170 lines)
- `bin/badi.js` — help and command map updated
- No new dependencies; skills are copied via existing `badi init` / `badi update`
- `package.json` version bumped to 1.10.0

## [1.9.0] - 2026-04-21

### Added — Global Documentation + Built-in Skill
- **English-first documentation**: `README.md` and `CHANGELOG.md` are now English (primary for global npm discovery)
- **Turkish docs preserved**: `README.tr.md` and `CHANGELOG.tr.md` — linked from both sides
- **Mobile App Store Screenshots skill**: `.claude/skills/mobile/app-store-screenshots/` bundled skill
  - Auto-triggers in Claude Code when the user asks for App Store / Google Play screenshots
  - Scaffolds a Next.js project and exports at all Apple / Google required resolutions (html-to-image)
  - Installed automatically via `badi init` — no extra command needed
- `badi mobile assets screenshots` output now references the skill and how to trigger it

### Technical
- New files in `files` array: `README.tr.md`, `CHANGELOG.tr.md`
- `package.json` version bumped to 1.9.0
- CLI output text still Turkish (backward compatible) — full CLI i18n is planned for v1.10

## [1.8.2] - 2026-04-19

### Added
- **`badi update --force`** — Force-refresh slash/agent/hook files (overwrites leftovers from older versions)
- User files are preserved: `memory.md`, `knowledge-base.md`, `workspace/`, `plugins/`, `logs/`, `backups/`
- Clearer update summary — overwritten vs preserved files shown separately
- If no new files were added, the user now sees a hint to try `--force`

### Previous behavior
`badi update` only added new files and did not refresh existing slash/agent content.
With `--force`, a full update is now possible (user customizations are preserved only in the memory/workspace areas).

### Usage
```bash
badi update              # Safe — only missing files
badi update --force      # Full — refresh all content except memory/workspace
badi update --dry-run    # Preview
```

## [1.8.1] - 2026-04-19

### Improved
- `badi doctor` produces more actionable suggestions when issues are detected
- Added a "Troubleshooting" section to the README — common errors + fixes
- Explicit guidance for the missing-hook error (`guard-bash.sh: No such file or directory`)

## [1.8.0] - 2026-04-19

### Added — AI/LLM + DevOps

**badi ai (5 subcommands):**
- `ai token` — `.claude/` token-usage analysis (by category + largest files)
- `ai prompt-test` — Regression tests for slash/agent files
- `ai memory-diff` — `memory.md` + `knowledge-base` limit check
- `ai review` — Staged-diff code review via Claude API (Haiku 4.5)
- `ai translate [file] --to [lang]` — Markdown translation

**badi dev (5 subcommands):**
- `dev deps` — Dependency update analysis (patch/minor/major)
- `dev deps --apply-patch` — Automatic patch-level updates
- `dev bundle` — Bundle size + framework detection + largest assets
- `dev docker-lint` — Dockerfile best practices (FROM/USER/HEALTHCHECK etc.)
- `dev env-check` — `.env` validation (missing/extra/placeholder)
- `dev api-test [url]` — HTTP endpoint tester (method/body/header/expect)

**Slash commands (10 new):**
- `/ai-token`, `/ai-review`, `/ai-translate`, `/prompt-test`, `/memory-diff`
- `/deps-update`, `/bundle-analyze`, `/docker-lint`, `/env-check`, `/api-test`

### Technical
- `lib/commands/ai.js` — Claude API integration (`ANTHROPIC_API_KEY`)
- `lib/commands/dev.js` — npm/yarn/pnpm detect + native tooling
- 11 new tests (total 169)
- **76 slash commands** total (66 + 10)

## [1.7.0] - 2026-04-19

### Added — Missing Slash Commands
CLI commands existed but slash commands were missing — 9 new `.claude/commands/`:
- `/wp` — WordPress site management (v1.4+)
- `/seo` — SEO audit (v1.4+)
- `/aso` — App Store Optimization (v1.5+)
- `/mobile` — Mobile project management (v1.5+)
- `/stats` — Usage analytics (v1.1+)
- `/schedule` — Scheduled reminders (v1.2+)
- `/icerik-ara` — Archive search (v1.2+)
- `/icerik-sablon` — Template inheritance (v1.2+)
- `/icerik-perf` — Content performance (v1.1+)

### Changed — Existing Slash Integrations
Integrated v1.6 CLI tools into existing slash commands:
- `/health` — `badi secret-scan`, `ssl`, `dns`, `lighthouse`, `a11y`
- `/security-scan` — Runs `badi secret-scan` first
- `/audit` — T4-tier Badi CLI suite
- `/deploy` — Pre-deploy `secret-scan --git` (block on critical)
- `/perf-check` — `badi lighthouse` production metrics
- `/changelog` — Quick generation via `badi changelog --write`
- `/release` — 4-command workflow (scan/changelog/check/release)

### Result
- **66 slash commands** (57 + 9)
- **21 CLI commands** (each with a slash counterpart)
- **12 hooks** (all active)
- Layered architecture: CLI → Slash → Agent

## [1.6.0] - 2026-04-19

### Added — Domain Health + Security + Git Workflow

- **`badi ssl [domain]`** — SSL certificate analysis (expiry, TLS version, cipher strength)
- **`badi dns [domain]`** — DNS record audit (A/AAAA/MX/TXT/SPF/DMARC/CAA) + email-security score
- **`badi whois [domain]`** — Domain registration + expiry + transfer lock
- **`badi lighthouse [url]`** — Core Web Vitals + Perf/A11y/SEO/BP via PageSpeed Insights
- **`badi secret-scan`** — 17 patterns (AWS/GCP/GitHub/OpenAI/Stripe/npm/DB URI/private keys); `--git` scans history
- **`badi a11y [url]`** — WCAG 2.1 accessibility audit (axe-core)
- **`badi commit`** — Conventional-commit helper + format validation (`--check`, `--message`)
- **`badi changelog`** — Grouped CHANGELOG.md generation from git log (`--from`, `--to`, `--version`, `--write`)

### Added — Slash Commands (.claude/commands/)
- `/ssl-check` — Badi CLI + interpretation guide
- `/dns-audit` — DNS + email security analysis
- `/whois` — Domain health
- `/lighthouse` — Performance/A11y/SEO audit
- `/secret-scan` — Secret scanning + action plan
- `/a11y-audit` — WCAG compliance + manual-test reminders
- `/conv-commit` — Staged-change analysis + commit
- `/changelog-gen` — Release workflow

### Technical
- `lib/commands/domain.js` — SSL (TLS socket), DNS (`node:dns`), WHOIS (TCP socket)
- `lib/commands/lighthouse.js` — PSI API integration
- `lib/commands/secret-scan.js` — 17 regex patterns, false-positive filter
- `lib/commands/a11y.js` — PSI accessibility category
- `lib/commands/commit.js` — Conventional format regex, git log parsing
- 15 new tests (total 158)

## [1.5.0] - 2026-04-18

### Added — Mobile and ASO

- **`badi aso`** — App Store Optimization command set (closes #47)
  - `audit` — iOS app listing audit + score calculation
  - `keywords` — Title/subtitle/description keyword analysis
  - `metadata` — iOS/Android character-limit guide
  - `review` — Review response templates
  - `compete` — App-to-app comparison + common/unique keywords
  - `screenshots` — iOS/Android size guide
  - `search` — iTunes API app search
- **`badi mobile`** — Mobile project lifecycle (closes #49, #50, #51)
  - `init` — React Native / Flutter / Expo / Swift / Kotlin templates
  - `version bump` — iOS/Android/Flutter version sync (package.json + Info.plist + build.gradle + pubspec.yaml)
  - `build` — iOS/Android release build (RN + Flutter)
  - `release` — TestFlight, Play Internal, App Store, Play guides
  - `assets icon/splash/screenshots` — Size and design guides
- **`badi icerik release-notes`** — App Store/Play Store release notes (closes #48)
  - `--platform ios|android` — 4000 / 500 character limits
  - `--lang tr,en` — Parallel generation
- **`badi icerik post --platform`** — Mobile platform variants (closes #53)
  - `appstore`, `playstore`, `mobile` CTA blocks

### Technical
- `lib/aso-helpers.js` — iTunes Lookup/Search + Play Store scrape
- `lib/commands/aso.js` — 7 subcommands
- `lib/commands/mobile.js` — 5 subcommand groups
- 28 new tests (total 143)

## [1.4.3] - 2026-04-18

### Changed
- External repo references cleaned up (52 files)
- External links removed from README, SECURITY, CHANGELOG
- Metadata simplified in 49 SKILL.md files under `.claude/skills/security-check/`
- Badi-focused metadata (dropped author, homepage, organization fields)

## [1.4.2] - 2026-04-17

### Performance
- **Startup time reduced by ~96%** (813ms → 26ms) — lazy command loading
- `bin/badi.js` now imports commands dynamically — only the invoked command is loaded
- Template lazy loading: TR/EN templates (~800 lines) only loaded during template generation
- `levenshteinDistance` memory reduced from O(m·n) to O(n) (single-row DP)
- Early-exit optimization (length-difference check)

### Measurements
| Command | Before | After | Improvement |
|---------|--------|-------|-------------|
| `badi --version` | 813ms | 26ms | ~97% |
| `badi list --agents` | ~800ms | 29ms | ~96% |

## [1.4.1] - 2026-04-17

### Fixed
- **Security**: SSRF protection added to SEO commands (localhost, private IP, non-http blocked)
- **Security**: WordPress `appPassword` base64 obfuscation + `wp-sites.json` file mode 0600
- **Bug**: `seo sitemap` operator precedence error (considered sitemap found even on 404)
- Increased `wp update` timeout to 120s (previously too short for large sites)
- WP test cleanup improvements

## [1.4.0] - 2026-04-17

### Added
- **`badi wp`** — WordPress site management (digital-agency feature)
  - `wp add/list/remove` — site configuration (WP-CLI or REST API)
  - `wp status` — WP version, active theme, plugin status
  - `wp plugins/themes` — detailed plugin and theme listing
  - `wp update` — bulk core/plugins/themes update
  - `wp security` — 6-point security scan
- **`badi seo`** — SEO audits and analysis
  - `seo audit` — 20+ checks, score calculation
  - `seo meta` — meta tag analysis and missing-tag detection
  - `seo sitemap` — robots.txt + sitemap.xml validation
  - `seo speed` — TTFB, HTML size, resource analysis, compression
- 10 new tests (total 115)

### Fixed
- CI test script compatibility (`tests/` → `tests/*.test.js`)

## [1.3.2] - 2026-04-16

### Fixed
- **CRITICAL**: Command injection — replaced `execSync` with `execFileSync` (plugin.js)
- **CRITICAL**: `guard-bash.sh` pipe chain bug (out-of-project write detection was broken)
- `VERSION` now read from `package.json` (single source, avoids sync drift)
- Habit streak calculation logic bug (stats.js)
- CSV-export formula-injection hardening (stats.js)
- Schedule wrap-around day ranges (sat-sun, fri-mon now work)
- Chalk fallback via Proxy (all chain combinations supported)
- `perf add` atomic append (race condition prevented)
- `session-reset.sh` macOS-incompatible `-printf` removed
- `dependency-audit.sh` cross-platform date + pnpm support
- `badi list` now lists the user's project (not `PKG_ROOT`)
- `badi update` adds `CLAUDE.md` if missing
- `badi icerik ac --open` opens file in editor
- `checkDuplicates` warning code set to `exit(2)` (warning, not error)
- Removed unused imports (helpers.js)
- `schedule parseTimeSpec` invalid-time validation (0-23:0-59)

### Added
- GitHub Actions CI workflow (Node 18/20/22 x ubuntu/macos)
- GitHub Actions publish workflow (npm provenance)
- Dependabot weekly npm + actions updates
- `FUNDING.yml` (GitHub Sponsors + Buy Me a Coffee)
- npm downloads + CI status badges in README
- v1.0.0, v1.1.0, v1.2.0, v1.3.0 GitHub Releases

## [1.3.1] - 2026-04-13

### Added
- 48 security-skill integration
- Full OWASP Top 10 coverage: SQLi, XSS, CSRF, SSRF, RCE, XXE and more
- 7 language-specific security scanners: Go, TypeScript, Python, PHP, Rust, Java, C#
- 3000+ security checklist items
- 4-phase security pipeline: Discover → Scan → Verify → Report
- Confidence scoring to reduce false positives

## [1.3.0] - 2026-04-12

### Changed
- `bin/badi.js` reduced from 3,812 lines to 135 lines (15 ESM modules)
- `CLAUDE.md` simplified from 6.8KB to 1.2KB (~82% reduction)
- Skills optimized from 676KB to 256KB (~62% reduction)
- Commands optimized from 264KB to 236KB
- `track-usage.sh` matcher narrowed (all tools → Bash|Write|Edit)
- 24-hour cache added to `dependency-audit.sh`
- Smart filtering in hooks (skips test/tmp files)
- Log rotation added (usage 1000, incident/failure 500 lines)

## [1.2.0] - 2026-04-12

### Added
- `badi icerik ara` — archive search + similarity detection + `--force`
- `--lang tr,en` — multilingual content generation (TR/EN parallel)
- `badi icerik sablon` — template inheritance (create/list/delete + `--sablon`)
- `badi schedule` — scheduled reminders (add/list/remove/check)
- EN templates: English versions for 6 content types
- Levenshtein + Jaccard similarity algorithms
- Frontmatter parse support
- Default language setting via `preferences.json`
- 33 new tests (total 105)

## [1.1.0] - 2026-04-12

### Added
- `badi stats` — usage statistics (bar chart, trend, habit streak, CSV export)
- `badi completion bash|zsh|fish` — shell-completion script generation
- `badi icerik perf` — content performance tracking (add/list/trend/roi/platform)
- Update notifier — npm registry check, 24-hour cache
- `track-usage.sh` hook — usage logging via PostToolUse
- 24 new tests (total 72)

## [1.0.0] - 2026-04-09

### Added
- 21 expert agents (security, performance, test, API, architecture, content, project planning)
- 50 workflow commands (session, quality, deployment, strategy, content)
- 12 security hooks (guard-bash, branch-guard, backup, completeness-gate)
- 21 skill categories (1,000+ procedures)
- CLI: init, update, doctor, list, plugin subcommands
- Content production engine: post, karousel, video, gorsel, takvim, marka
- Plugin system
- 6-layer memory architecture
- 48 tests