# Changelog All notable changes to this project will be documented in this file. The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html). ## [0.8.0] - 2026-05-17 ### Changed - **Extract and collate models now default to OpenRouter.** A single OpenRouter key covers all three pipeline stages. The separate MiniMax API key is no longer needed. Users upgrading from 0.7.0 whose model config matches the old default are auto-migrated with a notification. - **Settings now use a nested `pi-intelli-search` namespace.** Bare keys (e.g. `extractModel`) are preferred over flat `intelli*`-prefixed keys. Both formats still work; flat keys are deprecated and show a notification on every session start. - **README Settings section restructured.** A new Settings Reference table maps each setting to its pipeline stage, explains what it does, and gives guidance on when to change it. Context window considerations for small-window models (e.g. 256K) and cost-vs-speed trade-offs are called out. ### Breaking - **`maxUrls` split into `defaultUrls` and `maxUrls`.** The old `maxUrls` setting was always a fallback default, not a cap. It is now the hard cap (default 16). A new `defaultUrls` setting (default 8) provides the fallback when the agent does not pass `maxUrls` per call. Old settings containing `maxUrls` map automatically to the cap, matching what users always assumed it did. The pipeline now clamps agent requests with `Math.min(requested, maxUrls)`. The agent's SKILL.md heuristic (3/8/12) is now bounded by this setting. - **`llmsFullSites` setting removed.** The manual domain→URL map has been replaced by automatic discovery: every fetched domain is probed at `https://domain/llms-full.txt`. If the file exists (HTTP 200), it is downloaded raw to the cache. A small built-in list handles sites with non-standard paths (Cloudflare, Next.js, Vite). No configuration is needed. ### Added - **Auth warning on startup.** If no OpenRouter key is configured, a notification appears immediately rather than waiting for the first tool call to fail. - **Model validation before pipeline runs.** Typos in model names (e.g. `minimax/M3.7`) are caught before any API cost is incurred. - **Limit enforcement E2E tests.** Seven E2E test scripts now validate the full extension install experience. New tests cover: `defaultUrls` and `maxUrls` cap clamping, `extractMaxChars` and `extractionMaxTokens` enforcement (20× reduction proven), `collationMaxTokens` enforcement (9.5× reduction proven), model override via settings, and upgrade migration from 0.7.0. - **Automatic llms-full.txt discovery.** Every domain in the search results is probed at `https://domain/llms-full.txt`. If the file exists it is downloaded raw to the cache for offline grep. A small built-in list handles sites with non-standard paths (Cloudflare `/product/llms-full.txt`, Next.js `/docs/llms-full.txt`, Vite). The manual `llmsFullSites` setting is removed. A new `disableLlmsFullDiscovery` setting (default `false`) lets users opt out when the bandwidth cost of probing multiple domains is unwanted. - **Upgrade notification for `maxUrls` semantic change.** Users upgrading to 0.8.0 who had configured a custom `maxUrls` value see an in-product notification explaining it has become a hard cap (was a fallback default in 0.7.0). The `defaultUrls` setting is the new agent fallback. ### Fixed - **Settings from both locations are now read correctly.** Project-local `.pi/settings.json` takes precedence over global `~/.pi/agent/settings.json` (project overrides global). - **Version tracking survives directory changes.** Previously the version file was project-relative and could be missed when running `Pi` from different directories. - **Settings cache now correctly invalidated after default migration.** On first upgrade, tools previously read unmigrated defaults from a stale cache (the migration notification fired, but the pipeline itself used the old models). The cache is now rebuilt after migration so the pipeline immediately uses the new defaults. - **Rate-limit monitoring no longer goes dark after session replacement.** The `sessionActive` flag was never reset on `session_start`, so rate-limit status in the footer stopped updating after `/new` or `/fork`. - **Version marker written after migration completes.** Previously the version file was persisted before migration ran. If migration failed mid-session, the user was permanently stranded on stale defaults with no recovery path. - **Auth check tightened.** An empty `openrouter: {}` in `auth.json` no longer suppresses the missing-key warning. ## [0.7.0] - 2026-05-14 ### Changed - **npm scope migrated** from `@mariozechner` to `@earendil-works`: `pi-ai` and `pi-coding-agent` peer dependencies updated to 0.74.x. The `Pi` project moved from Mario Zechner's personal scope to an organization scope. No API changes. ### Fixed - `loadSettings()` and `getAgentDir()` now respect `PI_CODING_AGENT_DIR`, allowing the extension to read settings from isolated environments (e.g. E2E tests, CI). ### Compatibility - Minimum `Pi` version raised to 0.74.0 (previously 0.73.0). Earlier `Pi` releases bundle the deprecated `@mariozechner`-scoped packages and cannot resolve the new names. ## [0.6.0] - 2026-05-10 ### Changed - **Tool descriptions rewritten** to follow agent tool description best practices (Anthropic engineering guidance). Removed implementation details the agent cannot act on (Perplexity Sonar, Defuddle, LLM), replaced internal jargon with plain language, added cross-tool redirects so the agent picks the right tool, and clarified parameter descriptions. - Added `domains` usage guideline to `intelli_research` promptGuidelines. ## [0.5.1] - 2026-05-06 ### Fixed - **Images broken on pi.dev/packages:** replaced `.avif` images with `.png` equivalents. The `Pi` packages website does not support `.avif`. Also updated `pi.image` in `package.json`. ## [0.5.0] - 2026-05-06 ### Added - **Extension comparison guide** (`docs/COMPARISON.md`): feature-by-feature breakdown of `intelli-search` against 7 other `Pi` search extensions across search, fetch, extraction, collation, caching, and cost. ### Changed - **Documentation restructured:** SKILL.md reordered so decision logic precedes mechanism. README tightened with direct comparisons and consistent naming. All docs rewritten in assertive voice with hedged language removed. - **Pipeline infographic updated:** replaced `02.png` with a more detailed `06.png` showing the full 7-stage pipeline. ## [0.4.1] - 2026-05-06 ### Changed - Bumped `devDependencies` to `@mariozechner/pi-ai` 0.73.0 and `@mariozechner/pi-coding-agent` 0.73.0 to stay aligned with the latest `Pi` release. No runtime impact: peer dependencies remain `*` and `Pi` 0.73.0 introduces no breaking changes to extension APIs used by `intelli-search`. ### Compatibility - Verified against `Pi` 0.73.0: all 106 unit tests pass, E2E pipeline exercises the full search → fetch → extract → collate → cache flow in an isolated `Pi` environment. - Upstream `Pi` 0.73.0 fixes that benefit `intelli-search`: MiniMax M2.7 model metadata correction ([pi-mono#4110](https://github.com/badlogic/pi-mono/pull/4110)), and safer `models.json` provider override merging ([pi-mono#3651](https://github.com/badlogic/pi-mono/issues/3651)). ## [0.4.0] - 2026-05-05 ### Added - Schema.org JSON stripping in `cleanBrokenMetadata`: removes `