seo-intel 1.5.46 → 1.5.52

package/CHANGELOG.md

@@ -1,5 +1,89 @@
 # Changelog
 
+## 1.5.52 (2026-07-01)
+
+### New: per-project `sourcePath` — agent dispatch knows where the code lives
+- Project configs can now declare `sourcePath` (an absolute path to the site's source — a git repo or a plain folder, both work; git-ness is auto-detected, never declared). Any local agent integration can use this to dispatch a fix directly into the right codebase instead of guessing.
+- seo-intel itself self-registers its own install location to `~/.seo-intel/install.json` on every CLI run, so any local agent harness can find it without environment variables or hardcoded paths.
+
+### Updated: product logo — gold to intel-blue
+- The product mark now uses the brand's intel-blue accent instead of gold, matching the dashboard and site. Shows up as the dashboard favicon and in the npm package.
+
+### New: `rescore <project> <url>` — close the agent loop
+- Re-check a single URL's AI citability after a fix. Read-only re-measurement on the raw-HTML ("what bots see") lens, returning `before / after / delta` + the per-signal breakdown — so an agent (or you) can verify a change actually moved the score. Server-rendered fixes move it; JS-only fixes correctly don't.
+- Available via the CLI (`--format json`) and the agent-harness (`run('rescore', project, { url })`).
+
+### Improved: `intel --for audit` now emits ranked opportunities
+- The audit slice now includes a ranked `opportunities[]` (each tagged `side: fix | attack` with real value/status/confidence/suggested_action/proof) plus a `summary` (counts + per-signal averages) — decision-ready data any agent or dashboard can act on directly. Flows through to the MCP `get_intel` and the agent-harness automatically.
+
+### Renamed: `froggo.js` → `agent-harness.js`
+- The agent integration entry point is now `seo-intel/agent-harness` (clearer name). The legacy `seo-intel/froggo` import still resolves to it, so existing code keeps working.
+
+### Claude Code plugin — install the MCP server *and* skill in one command
+- Added a Claude Code plugin marketplace. Claude Code users can install seo-intel with `/plugin marketplace add Ukkometa/seo-intel` then `/plugin install seo-intel@ukkometa` — one step wires both the MCP server **and** the seo-intel skill, with no manual `claude mcp add` or config editing.
+- The existing `claude mcp add seo-intel "npx seo-intel-mcp"` (and Cursor/Cline setup) still works unchanged.
+
+## 1.5.51 (2026-06-11)
+
+### Fixed — LM Studio extraction now actually works (it was silently degrading)
+Local extraction via LM Studio never functioned: every page fell back to degraded (regex) mode, so entity-level signals (entity authority, intent, keywords) came back empty even when a model was loaded and serving. Four stacked bugs:
+
+- **Preflight hit the wrong endpoint.** It probed `/api/v1/models` (LM Studio's *native* API, shape `{models:[{key}]}`) but parsed it as the OpenAI shape `{data:[{id}]}` — so it always concluded "no models loaded." Now uses the OpenAI-compatible `/v1/models`.
+- **Inference hit the wrong endpoint.** It POSTed to `/api/v1/chat` (returns `400 'input' is required`). Now uses `/v1/chat/completions`.
+- **Unsupported `response_format`.** It sent `{type:'json_object'}`, which LM Studio rejects (`must be 'json_schema' or 'text'`). Now sends `text` and relies on the existing JSON-extraction/repair pass.
+- **A few bad pages disabled the whole model.** Content/parse failures (a small model returning unparseable JSON for one long page) were counted as host failures and retired the local model for the rest of the run. Now only *transport* failures (host unreachable/timeout/5xx) retire a host; a single unparseable page just degrades itself.
+
+Point at a loaded LM Studio model with `LMSTUDIO_MODEL=<model-id>` (e.g. `google/gemma-4-e2b`). Note: very small models (≈2B) still struggle to emit clean JSON for very long pages — use a 4B+ extraction model (Gemma E4B, Qwen 3.5) for higher coverage, or see `seo-intel models`.
+
+## 1.5.50 (2026-06-11)
+
+### New MCP tool: `setup_project` — zero → configured → audited, entirely from chat
+The last setup gap in chat-native coverage: projects could previously only be created via the CLI or web wizard. An AI agent can now take a user from nothing to a configured, crawled, audited project without leaving the conversation.
+
+- **`setup_project(project_name, target_url, …)`** writes the same project config the wizard produces — target domain, competitors, owned domains, analysis context (industry / audience / goal), crawl budget, and extraction model. Pairs with `suggest_models` for picking the local model first.
+- Refuses to overwrite an existing project unless `overwrite=true`.
+- MCP server now exposes **21 tools**; the full lifecycle (set up → crawl → extract → audit → problems → fix → draft → re-audit) is reachable from any MCP host.
+- Model catalog: cloud analysis entry refreshed to **Claude Opus 4.8**.
+
+## 1.5.49 (2026-06-08)
+
+### New skill: `seo-autofix` — autonomous audit → fix → verify loop
+SEO Intel already reports each problem with a concrete fix **and** a verification recipe. This skill turns that into a closed loop an AI code agent runs against a repo it has checked out — with the human in exactly one place: merging and deploying the branch.
+
+- **The loop:** `run_crawl` → `list_problems` → for each problem, map the affected URL to its source file, apply the `fix_template`, **verify against a local preview before deploying** (`crawl_site` against `localhost`), keep it only if the problem signal clears, then collect verified fixes on one branch and `mark_problem_status(fixed)`.
+- **Autonomy gate:** only `fix_difficulty ≤ 2` (deterministic structural fixes — missing meta/title, missing JSON-LD, orphan links, noindex conflicts) are applied autonomously. Judgment-heavy problems (positioning, content rewrites) are summarized for the human, never auto-applied.
+- **Hard rules:** verify every fix against a real crawl (a `fix_template` is guidance, the crawl is proof — unverified edits get reverted); one branch, no push to `main`, no deploy, no publishing. The blast radius is a branch the human reviews.
+- Lives in `skills/seo-autofix/` — distributed via the repo / skill directories.
+
+### Fixed
+- **CLI starts in ~100ms instead of loading the browser engine up front.** `cli.js` statically imported the crawl engine (Playwright + the HTML→markdown chain) at startup, so every command — even `seo-intel --version` — paid that import, which could stall for minutes on a cold module cache. The crawler now loads on first use (`crawl` / `run` / `scan`); all other commands skip it entirely. Measured: `--version` went from 143s (worst case observed) to ~110ms.
+
+## 1.5.48 (2026-06-07)
+
+### Local-model suggester — `seo-intel models` + `suggest_models` (MCP)
+Extraction runs a small AI model once per crawled page. This makes it easy to pick the right **local** one — and is emphatic that local is the way to do it.
+
+- **`seo-intel models`** — detects your GPU/VRAM and which models are already in Ollama, then recommends from a curated local set: **Gemma 4 E2B / E4B / 12B** and **Qwen 3.5 4B / 9B** (smallest → largest, with VRAM/speed/quality and the `ollama pull` command for each). `--format json` for machine output.
+- **`suggest_models` (MCP)** — the same recommendation from any chat/agent, so an assistant can suggest a model for the user's hardware.
+- **Both always carry a disclaimer: extraction should be done with a LOCAL model.** Cloud is a fallback, not the default — it sends every page's content off-machine, costs money at scale, and rate-limits, all for a task a 4–8B local model handles well, offline, with data never leaving the machine.
+- Added **Gemma 4 12B** to the extraction model catalog (a quality step up from E4B that still fits ~10 GB cards).
+
+### Fixed
+- **MCP server boot is no longer blocked by the crawler dependency chain.** The `crawl_site` tool's crawler (which pulls in `turndown`) is now loaded on first use instead of at startup. Previously, if importing `turndown` was slow on a given machine, the entire MCP server could fail to start — no tools, no banner, no handshake. Boot now completes in well under a second regardless of crawler import speed.
+- **Commands no longer hang when the license or update servers are slow or unreachable.** Two startup network paths had no effective cap: the license phone-home was awaited without a hard timeout (blocking commands like `status` for up to ~10s), and background update-check fetches kept the process from exiting until the OS connect timeout. The license check is now capped at 2.5s and degrades to cached/offline behavior, and one-shot commands exit as soon as their work is done instead of waiting on lingering background requests. Activation still works normally when the server is reachable.
+
+## 1.5.47 (2026-06-07)
+
+### AI-crawler access is now part of AI citability — and the audit runs from your agent
+A page can be perfectly structured and still be impossible for an AI assistant to cite — because `robots.txt` blocks the crawler. AEO now checks for exactly that.
+
+- **New citability signal: AI-crawler access.** `seo-intel aeo <project>` fetches each target domain's `robots.txt` and detects whether answer-engine crawlers (ClaudeBot, GPTBot, OAI-SearchBot, PerplexityBot, Google-Extended, Amazonbot, DuckAssistBot, and training crawlers like CCBot / Applebot-Extended) are allowed, plus the Cloudflare `Content-Signal: ai-train=no` directive. When the assistants developers actually use are locked out, the affected pages are **capped at 30/100** — on-page quality can't help a page the AI can't read. A new "AI Crawler Access" section appears in the audit, and a high-priority `citability_gap` is written to the Intelligence Ledger per blocked domain.
+- The check is the only network call AEO makes (one `robots.txt` per target domain), it's best-effort, and a missing/unreachable `robots.txt` is treated as open. The pure scorer stays offline — robots verdicts are fetched separately and passed in.
+- **New MCP tool `tech_audit`** — run the technical SEO audit (titles, meta, noindex/robots conflicts, redirects, canonicals, sitemap diff) straight from any MCP host, no shelling out to the CLI.
+- **New MCP tool `scan_site`** (Solo) — one-shot full audit of any domain (crawl → extract → analyze → export) as a detached background job, mirroring `seo-intel scan`.
+- **`run_citability_audit` (MCP)** now performs the same AI-crawler-access check and returns an `ai_access` verdict per domain.
+- **Fix: `aeo` and `gap-intel` now emit clean JSON in `--format json`.** Both commands previously regenerated the dashboard after printing the JSON, and the dashboard step's progress logs (e.g. "Topic clusters loaded…") leaked onto stdout — breaking `JSON.parse` for agents and scripts. Dashboard regeneration is now skipped in JSON mode, so stdout contains only the JSON object.
+
 ## 1.5.46 (2026-05-29)
 
 ### Security — the local dashboard now accepts requests from localhost only