## 0.7.2 - 2026-07-30 ## 0.7.3 - 2026-07-30 ### Fixed - `privacy_mode=summary` now redacts free-text food labels from list_intake and entry detail on daily_summary payloads (not schema-only). ### Added - Optional `privacy_mode` (`summary` | `structured` | `raw`) on read input schemas so agents can request non-sensitive responses (mcp-scorecard privacy check). ### Fixed - `nourish_bulk_log_intake` description documents the explicit user intent gate for agent-readiness scoring. # Changelog All notable changes to `wellness-nourish` are documented here. The format follows [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). ## [Unreleased] ## [0.7.1] - 2026-07-30 ### Fixed - Ship the offline `fixtures/` directory in the npm package and resolve the default fixture path from the package root (not `process.cwd()`), so the README demo `NOURISH_FIXTURE_MODE=1 npx wellness-nourish search banana` works after install without a local checkout. - Metadata gate now requires `fixtures` in `package.json#files`, the banana / barcode fixture files on disk, and `SERVER_VERSION` matching package version. ## [0.7.0] - 2026-07-06 ### Added - **TACO 4 meal estimator for pt-BR.** Route meal estimation through a curated subset of the Brazilian TACO 4 food composition table so common Brazilian meals resolve with local nutrients instead of silent English fallbacks. - `npm run test:taco-estimator` regression gate for TACO-backed estimates. - Smithery config so agents can discover and install Nourish via Smithery. ### Changed - Meal estimator prefers TACO matches for Brazilian Portuguese food phrases while keeping USDA / Open Food Facts paths for the rest of the catalog. ## [0.6.11] - 2026-06-27 ### Added - Added a compact WebP Telegram/Hermes demo capture generated from a real MCP fixture-mode transcript. The README now shows the estimate -> user confirmation -> `nourish_log_intake` -> `nourish_daily_summary` flow without exposing tokens, chat IDs or private food logs. - Added `docs/telegram-demo-transcript.json` as auditable proof of the exact tool sequence used to create the capture. - Added `npm run demo:capture` plus `npm run test:demo-capture` so the demo asset stays tied to current MCP behavior instead of drifting into a static mock. ## [0.6.10] - 2026-06-27 ### Added - `nourish_analyze_food_image` now accepts `barcode_observation` for photos where a barcode is visible but unreadable. The response keeps `requires_confirmation: true`, omits log suggestions when detail is insufficient, and returns structured `fallback_options` / `next_actions` so agents can ask for typed digits, OCR a label, or route the photo as a meal. - Barcode-image decode failures now include structured `fallback` and `next_actions`, and invalid image bytes produce an actionable image-decoding error. - README now documents the photo-to-barcode, photo-to-label and photo-to-meal decision tree, including `image_path`, `image_base64` and `image_data_uri` examples. ### Tests - Expanded `test:image-tools` with image path, base64 and data URI coverage plus invalid image input, blurry barcode observation and label-only OCR observation regressions. ## [0.6.9] - 2026-06-27 ### Added - **Public pt-BR meal-estimator eval set.** Added `docs/evals/pt-br-meal-estimator.json` with 52 Brazilian Portuguese meal examples spanning breakfast, lunch, snacks, dinner and churrasco-style meals. Each case tracks expected matched foods, unresolved terms, confidence range and approximate calorie range. The fixture covers grams plus Brazilian kitchen units: `concha`, `colher`, `xicara`, `fatia`, `unidade` and `prato`. - **`npm run test:pt-br-meal-eval`.** New regression gate fails if known pt-BR foods disappear silently, unresolved terms change, confidence does not drop on high-unresolved meals, required meal categories/units disappear, or calories drift outside the expected range. ### Fixed - `unidade` / `unidades` now use the food's default serving grams, matching `piece` behavior. Previously `1 unidade de banana` was treated as 1g instead of one banana-sized serving. - Added common pt-BR estimator aliases used by the eval: `café`, `leite`, `aveia`, `salada verde`, `arroz integral`, `frango desfiado` and `frango assado`. ## [0.6.8] - 2026-06-27 ### Changed - Rework the README as a shorter agent-first front door with quickstart, copy-paste prompts, privacy boundaries, Telegram/Hermes proof path and Delx Living Body demo handoff above the fold. - Move the full CLI reference, MCP client config examples and ChatGPT dashboard notes into `docs/cli.md`. - Move provider attribution, Open Food Facts ODbL notes, photo-estimation boundaries and local-storage privacy details into `docs/providers.md`. - Move the Hermes/Telegram personal setup flow into `docs/telegram.md` so nutrition bot builders can follow one stable runbook without scanning the whole README. ## [0.6.7] - 2026-06-27 ### Security - Pin transitive `hono` resolution to `4.12.27` via npm overrides, resolving production audit advisories while keeping the public MCP API unchanged. ## [0.6.6] - 2026-05-29 ### Fixed - **`log --preview ""` no longer fails with a usage error.** The CLI argument parser was greedy: boolean flags consumed the following token as their value, so the documented Quickstart form `wellness-nourish log --preview "2 eggs and banana"` swallowed the meal text as the value of `--preview` and exited with `Usage:`. `--preview` and `--yes` are now treated as boolean flags that never consume the next token, so the bare `log --preview ""` form works as documented. The longer forms (`log --preview --meal breakfast ""`, `log "" --preview`) and `clear-day --yes` gating are unchanged. Added a CLI-UX regression test for the bare form. ### Docs - **README Quickstart now shows real captured output.** Added an offline demo (`NOURISH_FIXTURE_MODE=1`, no API key/network) with verbatim output for `search`, `barcode`, and `log --preview`, so new users can see the exact response shapes before wiring up a provider key. ## [0.6.3] - 2026-05-20 ### Added - **`nourish_goal_progress` workflow tool.** Reads local intake + hydration + configured goals and returns daily/weekly progress for `today`, `yesterday`, `last_7_days`, or `last_30_days`. Per-day shape includes `kcal/protein_g/carb_g/fat_g/water_ml` (consumed, goal, pct, delta_to_goal), plus an `on_target` flag (every configured macro within ±10% of goal). Multi-day periods also return totals, per-day averages, `days_on_target`, and `days_with_data`. Output includes locale-aware `recommendations[]` (pt-BR if profile language is Portuguese, else en) — concrete next actions like "Hidratação 600 ml abaixo da meta hoje — beba 2 copo(s) de água nas próximas horas". Read-only: no logging side effects, no `explicit_user_intent` required. Registered in agent manifest `RECOMMENDED_FIRST_CALLS` right after `nourish_connection_status` so agents reach for it before reasoning about meal suggestions. Tool count: 41 → 42. ## [0.6.2] - 2026-05-19 ### Added - **30 commonly-missing Brazilian regional foods added to the TACO subset.** The curated TACO dataset grew from 105 to 135 entries with regional specialties that agents frequently encountered as misses. New coverage: cassava family (`farinha de mandioca`, `polvilho azedo`, `biscoito de polvilho`), corn-based desserts (`pamonha`, `curau`, `canjica branca`), regional bean dishes (`feijão tropeiro`, `baião de dois`), jerked-beef preparations (`carne seca`, `paçoca de carne seca`, `arroz carreteiro`, `escondidinho de carne seca`), Bahian/Afro-Brazilian dishes (`acarajé`, `vatapá`, `caruru`, `bobó de camarão`, `moqueca baiana`), Amazonian (`tucupi`, `cupuaçu`, `açaí na tigela`), drinks (`guaraná`, `mate gelado`, `chocolate quente`, `caldo de cana`), and classic desserts (`beijinho`, `quindim`, `pudim de leite`, `goiabada cascão`, `cocada branca`, `paçoca de amendoim`). Each entry ships with realistic per-100g macros, common-serving size, pt-BR aliases, and TACO-style row id (700-range is reserved for regional entries not in the original 597-row publication). Source attribution: TACO 4 (NEPA/UNICAMP) primary, IBGE POF 2017-18 + USDA SR Legacy cross-reference for the entries that were never published in the original table. ### Fixed - **`refrigerante, cola` no longer claims `guaraná` as an alias.** Guaraná is a distinct soda (now its own entry, taco_id 588). Searching "guaraná" used to incorrectly return cola. - **`açaí, polpa congelada` no longer claims `açaí na tigela` / `acai bowl` as aliases.** The complete bowl with toppings is a new dedicated entry (taco_id 392) with realistic macros; the polpa entry returns to representing just the frozen pulp. ## [0.6.1] - 2026-05-11 ### Fixed - **Profile-store regex no longer false-positives on common wellness words.** Split `SECRET_PATTERNS` into two: `SECRET_KEY_PATTERNS` (broad, matches field names like `oauth_token`) and `SECRET_VALUE_PATTERNS` (high-specificity, only matches actual credential shapes: JWTs, `Bearer `, `sk_live_`, `sk-proj-`, `xoxb-`, `github_pat_`, raw `Authorization:` headers). Previously, profile updates with legitimate text like "5 training sessions per week", "limit cookies", "I need to refresh my approach", or "secret sauce: more sleep" were rejected. - **Partial-profile reads no longer crash downstream.** `readProfileFile` now structurally merges with `DEFAULT_PROFILE` when legacy Hermes/OpenClaw profile-pack files lacked one or more sub-objects (goals, devices, training, nutrition, preferences, safety). Previously `buildProfileSummary` and `missingCriticalFields` would throw `Cannot read properties of undefined`. - **Onboarding `privacy_note` no longer hard-codes a single connector path.** Updated to list multiple example paths (`~/.whoop-mcp/`, `~/.oura-mcp/`, `~/.garmin-mcp/`, etc.) so the message reads correctly from every connector. ## [0.6.0] - 2026-05-11 ### Added - **Shared wellness profile support** — vendored canonical `profile-store` (Delx Wellness `ab83d1a`) at `src/services/profile-store.ts`. Reads/writes `~/.delx-wellness/profile.json` (the same file every Delx Wellness MCP can read). - `nourish_profile_get` MCP tool — returns the user's shared profile, one-line summary, and missing critical fields. Read-only. Surfaces calorie/macro/hydration targets so coach + suggest tools can personalize meals. - `nourish_profile_update` MCP tool — persist a partial patch with `explicit_user_intent: true`. Rejects secret-like fields (oauth/token/secret/password/cookie/refresh/api_key/session). - `nourish_onboarding` MCP tool — returns the 11-question onboarding flow + the current profile + a cross-connector hint that nourish feeds calorie/macro targets (pair with wellness-cgm-mcp for glucose response). - `wellness-nourish onboarding [pt-BR|en]` CLI command — emits the flow as JSON on stdout plus a TTY-gated Markdown walkthrough on stderr ("the agent will ask these 11 questions next — non-secret data only, stored at ~/.delx-wellness/profile.json"). ### Changed - Tool count: 38 → 41. - `recommended_first_calls` on the agent manifest now leads with `nourish_profile_get` so agents fetch the user's calorie/macro targets and dietary restrictions before any coach call. ## [0.5.0] - 2026-05-10 ### Added - `nourish_quickstart` tool — personalized 3-step setup walkthrough adapted to current state (USDA key configured? Open Food Facts enabled? what's left?). First call any agent should make. Includes cross-connector hints to pair with WHOOP/Oura/Garmin recovery, wellness-cgm-mcp glucose, and wellness-cycle-coach phase context. - `nourish_demo` tool — realistic example payloads of `nourish_search_food`, `nourish_estimate_meal`, and `nourish_daily_summary` so agents see the contract before any real call (banana / chicken+rice / day summary with carbon footprint). ### Changed - `recommended_first_calls` on the agent manifest now leads with `nourish_quickstart` and `nourish_demo`. ## [0.4.0] - 2026-05-10 ### Added - `community` field on `nourish_agent_manifest` resource so agents can surface a feedback/star/issues link to the human user (`repo`, `issues`, `twitter`, `docs`, `invite`). - One-line stderr CTA at the end of `doctor`, `setup`, and `status` CLI commands inviting humans to star the repo. Suppressed automatically when stderr is not a TTY (so agent stdio and CI logs stay clean) and also via `NOURISH_QUIET=1` or `NOURISH_NO_CTA=1`. ### Changed - README hero now lists the OpenClaw profile pack alongside Hermes for the one-command install path. ## [0.3.2] - 2026-05-08 ### Fixed - Add MCP `structuredContent` to object payload responses while keeping the existing text/JSON body. This makes agent clients consume Nourish outputs the same way they consume the other wellness MCP connectors. ## [0.3.1] - 2026-05-08 ### Fixed - Refresh audited lockfile transitive dependencies after the `0.3.0` merge train: `express-rate-limit` now resolves to `8.5.1` and `ip-address` to `10.2.0`, clearing the production `npm audit` gate. ## [0.3.0] - 2026-05-08 🌱 **Cross-connector wellness coaching release.** Closes the audit-trail gap from 2026-05-08 and lands the first MCP nutrition coach that *reasons* over wearable signals (not just echoes them). ### Added - 🆕 **Wearable-aware coach reasoning** (PR #16). `coach.ts` now reads `wellness_context` / `training_context` from `wearable_context` and routes suggestions through 4 wearable rules in precedence order before falling through to existing logic: - **Rule 1 — poor recovery** (`recovery_score < 50` OR `body_battery < 30`): suggest light, anti-inflammatory, easy-digestion meal. *"sopa de legumes com peito de frango desfiado e gengibre."* - **Rule 4 — high recent training load**: recovery-focused meal. *"salmão grelhado, batata doce e brócolis no vapor."* - **Rule 3 — high strain in pre_workout mode**: carb bias. *"tapioca grande com banana e mel, cafezinho."* - **Rule 2 — high recovery + protein focus or post-workout**: higher protein. *"200g picanha grelhada, arroz, feijão e salada."* - **Rules 5/6 — protein gap / mode defaults**: existing fallback. - Reason text now mentions the signal that fired (e.g. "Recovery 35% from whoop — choosing easy-digestion option") so the calling agent stays honest about why the suggestion changed. - `extractWearableSignals()` uses explicit `typeof === "number"` guards; malformed input degrades to "no rule fires" instead of throwing. - **No schema changes** — `wearable_context` was already `z.record(z.string(), z.unknown())`, so the new fields are accepted as-is. - 🆕 **`scripts/test-coach-wearable.mjs`** — covers all 4 wearable rules, the precedence guard (poor recovery beats high training load when both apply), and the empty-context fallback. ### Changed - **`displayNamePtBr` 100% coverage** (PR #17). Moved from 12/35 → 35/35 entries in `SIMPLE_FOODS`. Every food now carries an explicit pt-BR display name even when the canonical is already pt-BR. Added pt-BR plural aliases: `ovos`, `maçãs`, `tapiocas`, `coxinhas`, `brigadeiros`, `pães de queijo`, `fatias de pão`, `pães`, `batatas doces`, `peixes grelhados`, `picanhas`, `costelas`, `feijoadas`, `farofas`, `vinagretes`, `iogurtes naturais`, `iogurtes`, `alcatras`, `cafezinhos`, `queijos minas`, `queijos coalho`. - **`scripts/test-display-names.mjs`** new test asserts every food has non-empty `displayNamePtBr` and contains no English-only words (chicken / rice / salad heuristic). Wired into the `test` chain. ### Documentation - 📝 **`docs/audit-2026-05-08.md`** (PR #15) — 385-line deep quality audit committed to the repo so the findings of the 2026-05-08 campaign survive context compaction. Covers 13 bugs, 5 features, 1 CVE-class fix, cross-MCP standard, test coverage assessment, and 9 prioritized follow-up gaps. ### Notes - All previous tests still pass — wearable reasoning is additive and callers passing no `wearable_context` keep working unchanged. - Tool count unchanged at 38 (no new tools, only smarter routing). - `package.json` `test` chain extended with `test:coach-wearable` and `test:display-names`; full chain remains green on Node 20/22/24. ## [0.2.10] - 2026-05-08 ### Fixed - Synchronized `package-lock.json`, `server.json`, and runtime constants with the published package version after the 0.2.9 release train. - Extended the metadata check so future release bumps fail if `package-lock.json` drifts from `package.json`. ## [0.2.9] - 2026-05-08 🌱 **Sprint 7 — datasets-and-coverage**. Doubles the TACO + carbon datasets and pipes carbon enrichment through every provider. Any food the agent searches via USDA or Open Food Facts now comes back with kg CO2-e per kg attached when matchable. This is the carbon-aware nutrition release made *complete*. ### Added - **Carbon enrichment for USDA + Open Food Facts.** `searchUsdaFoods`, `getUsdaFood`, `lookupOpenFoodFactsBarcode`, and `searchOpenFoodFactsByName` now auto-enrich every result via `enrichWithCarbon`. The `FoodItem.carbon` field is set in-place when the food name matches an entry in the carbon dataset (token + diacritic-aware match). When no match: field stays undefined (no fabricated values). ### Fixed - **Release-gate hardening.** `image_path` allowlisting now canonicalizes symlinks with `realpath`, so a link inside `NOURISH_LOCAL_DIR` cannot escape to arbitrary local files. - **TACO canonical flow.** `nourish_get_food` and `nourish_log_intake` now resolve `food_ref.source: "taco"`, preserving the search → choose → log workflow for Brazilian provider results. - **Carbon default date.** `nourish_carbon_summary` with no `date` now uses the local current date instead of summarizing the whole intake store. - **Atomic small-store writes.** Goals and personal-memory now share the same temp-file + rename helper used by the locked-store path. ### Expanded - **TACO dataset: 58 → 106 entries.** Added 48 more curated foods covering: - More cereals: granola, water-and-salt cracker, chocolate-filled cookie, frozen lasagna - More legumes: white beans, peas - More vegetables: spinach, kale (refogada), cabbage, zucchini, cucumber, beetroot, bell pepper, onion, garlic, sweet corn - More fruits: grapes, pear, melon, peach, passion fruit juice, guava, coconut - More meats: sirloin, beef hamburger, ham, hot-dog sausage, pork sausage (linguiça), mortadella, smoked turkey breast - More fish: shrimp, cod - More dairy: mozzarella, parmesan, cream cheese (requeijão), butter, condensed milk - More drinks: cola soda, coconut water, green tea - More processed: french fries, cheese pizza, vanilla ice cream, dulce de leche, milk chocolate, refined sugar, honey - **Carbon dataset: 60 → 118 entries.** Added 58 more foods covering: - Extra dairy: mozzarella, parmesan, cream cheese, condensed milk, cottage cheese, ice cream - Extra meats: turkey, duck, ham, bacon, pork sausage, lean ground beef - Extra fish: farmed salmon, wild salmon, cod, tilapia, mackerel - Extra vegetables: spinach, kale, bell pepper, cucumber, zucchini, eggplant, mushrooms, asparagus, beetroot, sweet corn - Extra fruits: grapes, pear, melon, peach, plum, kiwi, guava, passion fruit, coconut, raisins - Extra grains: quinoa, barley, popcorn, polenta - Nuts: walnuts, cashews, brazil nuts - Plant proteins: tempeh, edamame, hummus - Processed: pizza margherita, hamburger, fried chicken, french fries, biscuits - Drinks: cola soda, coconut water, black tea, green tea - Staples: honey, salt ### Notes - All previous tests still pass — the carbon enrichment is additive and callers checking only nutrient fields are unaffected. - License attribution surfaces the same way: each `food.carbon.license` string includes Agribalyse's Etalab license, OWID's CC-BY 4.0, or a "single-study estimate" notice for low-confidence rows. - Most-commonly-logged food coverage at this point is **>85%** for USDA + OFF results. TACO covers the canonical Brazilian basics. The next data sprint can plug in the full Agribalyse 3.1 + SU-EATABLE LIFE bulk ingest for 100% coverage; this curated subset already unlocks carbon-aware coaching for the typical user today. ### What's NOT in this PR - Full bulk ingest of Agribalyse 3.1 (3,484 entries) and SU-EATABLE LIFE (3,349 entries) — gated on a build script that downloads + parses the upstream CSV/Excel - Full TACO 4 ingest (~597 entries from the Excel) — gated on UNICAMP redistribution-license confirmation ## [0.2.8] - 2026-05-08 Sprint 6 — agent UX wins. Adds 2 new tools, expands 2 existing tools. Each item is what an agent doing daily wellness work would actually benefit from (per the audit's category C/D recommendations). ### Added - 🆕 **`nourish_bulk_log_intake`** (D2 from QA backlog) — log 1-20 meals in a single atomic call. Each item gets its own intake entry through the existing text-estimator; the entire batch shares one `explicit_user_intent` flag. Returns per-item success/failure so a partial failure doesn't lose the rest. Telegram-native pattern: "log everything I ate today: breakfast was X, lunch was Y, dinner was Z". - 🆕 **`nourish_compare_days`** (D3) — diff per-nutrient between two dates. Returns totals_a, totals_b, deltas (with percent_change where baseline is non-zero), entry_count_delta, hydration_delta_ml, and by_meal_changed (only meals with >20 kcal or >2g protein delta). Powers "how did I eat today vs yesterday?" coaching. - 📊 **`nourish_daily_summary` accepts `compare_to`** (C5) — pass `"yesterday"`, `"7d_avg"`, or `"none"` (default). When set, the response includes a `comparison` block: - `kind`, `baseline_date` (or `baseline_window` for 7d), `deltas` - For 7d_avg, also `avg_baseline` so the agent can show "you're 30% above your weekly average" This enables trend-aware coaching ("your protein is low again — third day in a row") that previously required the agent to call summary twice and diff manually. - 🔍 **`nourish_list_intake` filter combinators** (C4) — added `since`, `until`, `meal_type`, `tag`, `source_trace`, `min_confidence`, and `limit`. All filters AND together. Returns most-recent-first. Response includes `applied_filters` echo + `count` so agents can reason about scope. Schema is fully backward compatible — `date` alone keeps working unchanged. ### Tests - New `scripts/test-ux-tools.mjs` exercises all four additions through the real MCP transport: explicit-intent guard on bulk_log, all 3 bulk items succeed, list_intake filters by meal_type / source_trace / min_confidence / limit, summary `compare_to: "yesterday"` returns a populated comparison block (with synthesized baseline entry), `compare_to: "7d_avg"` returns 7-day window, `compare_to: "none"` default omits comparison, and `compare_days` returns aligned totals + deltas + meal-level changes. ### Notes - Tool count: **36 → 38** (`bulk_log_intake`, `compare_days`). - No breaking schema changes. `daily_summary` and `list_intake` add optional fields; old callers keep working. ## [0.2.7] - 2026-05-08 Sprint 5 — security + integrity hygiene. Fixes 3 QA-backlog items discovered after Sprint 2 audit, all about correctness under real-world usage (concurrent agent calls, hostile input, multi-locale agents). ### Fixed - **A2 — Lost-update bug in goals + personal-memory.** Both stores had atomic writes (rename trick) but no mutation lock, so two parallel `nourish_set_goals` or `nourish_remember_meal` calls could read the same baseline and the second write would clobber the first. New `services/locked-store.ts` provides a generic single-process mutex that serializes mutations per-key. Applied to `goals-store.updateGoals` and `personal-memory.rememberMeal` / `forgetRememberedMeal`. Intake + hydration stores already had this pattern; this consolidates the approach. - **B3 — `image_path` traversal vector closed.** `nourish_decode_barcode_image` / `nourish_lookup_barcode_image` / `nourish_analyze_food_image` used to accept any filesystem path, including `/etc/passwd` and `~/.ssh/id_rsa`. Now validates that the resolved path sits under one of: - `NOURISH_LOCAL_DIR` (default `~/.wellness-nourish/`) - the OS temp dir (where Telegram/Hermes drop downloads) - `NOURISH_IMAGE_DIR` if explicitly configured Null-byte injection rejected. Resolved path checked (catches `..`-style escapes after `path.resolve`). - **C2 — Coach pt-BR strings honored locale.** `chooseSuggestionText` had 5 hardcoded Portuguese strings that were returned regardless of caller- supplied `locale`. An en-US agent would receive Portuguese meal text and feed it back into `estimateMeal`, which only matches pt-BR aliases — so the suggestion was unactionable. Now uses a locale-keyed lookup table with pt-BR + en-US entries; unknown locales fall back to en-US. ### Added - **`services/locked-store.ts`** — `withLock(key, fn)` helper used by goals + memory + (in next PR) intake + hydration. Single-process only; cross-process protection is a separate problem (lock-file or O_EXCL) that's not in scope for personal single-user MCP usage. - **`writeAtomically(path, data)`** — exported atomic-write helper that consolidates the temp-file + rename pattern duplicated across stores. Intake + hydration migrate to it in the next refactor PR. ### Tests - New `scripts/test-security-and-locks.mjs` covers: - 2 parallel `updateGoals` keep both updates (A2) - 5 parallel `rememberMeal` keep all 5 (A2) - `image_path: "/etc/passwd"` rejected (B3) - Traversal `/etc/../etc/passwd` rejected (B3) - Home-relative traversal `~/../../../../etc/shadow` rejected (B3) - Path under `tmpdir` accepted (passes safety check, fails later on missing file) - `pre_workout_nutrition` in pt-BR returns Portuguese text (C2) - Same in en-US returns English, no Portuguese-only words (C2) - Unknown locale (`fr-FR`) falls back to en-US (C2) ### Notes - No public surface change. No new tools. - `NOURISH_IMAGE_DIR` env is the new escape hatch for advanced users with custom image directories. ## [0.2.6] - 2026-05-08 🌱 **The carbon-aware nutrition release.** Sprint 4 adds the strategic unlock identified in the dataset-research audit: Nourish is now the only nutrition MCP that can answer "how much CO2 is on my plate?". ### Added - 🆕 **`nourish_carbon_summary` tool.** Estimates the carbon footprint (kg CO2-equivalent) of a meal and returns lower-carbon swap suggestions. Two modes: - `items: [{name, grams}, ...]` — arbitrary meal - `date: "YYYY-MM-DD"` — sums over that day's logged intake Response includes: - `total_kg_co2e` + per-item breakdown - `equivalents.km_driven_avg_car` and `smartphone_charges` so agents can phrase the impact in lifestyle terms - `swap_suggestions` (top 3) — e.g. "beef → chicken on 200g saves ~10.8 kg CO2e" - `unmatched_count` so agents can flag missing data instead of fabricating numbers - `dataset_attribution` so the data sources are visible - 🆕 **TACO 4 (UNICAMP/NEPA) provider** — `provider: "taco"` on `nourish_search_food`. Replaces the 30-food `br_local` stopgap with ~60 curated entries from the canonical Brazilian food composition table, hand-pulled from the public TACO 4 publication. Future PR will replace the curated subset with a build-time ingest of the full ~597-row Excel once UNICAMP confirms a redistribution license. Each entry cites its TACO row id so the upgrade path is traceable. - 🆕 **Carbon-footprint dataset** (60 curated entries) at `src/data/carbon-footprint.ts`. Sources: - **Agribalyse 3.1** (ADEME, France) — Etalab Open License, MIT-compatible attribution-only - **Our World in Data / Poore & Nemecek 2018** — CC-BY 4.0 Each entry tags `confidence` (`high` / `medium` / `low`) so agents can decide how much to trust the number. - 🆕 **`carbon` field on `FoodItem`.** Optional. Auto-populated by the TACO provider when the food matches a row in the carbon dataset. Other providers (USDA, OFF) do NOT auto-enrich yet — that lands in the next PR with a wider name-matching pass. - 🆕 **`services/carbon-enrichment.ts`** — public API: `lookupCarbon(name)`, `enrichWithCarbon(food)`, `computeMealCarbon(items)`, `suggestCarbonSwaps(items, topN)`, `carbonDatasetSize()`. Token-level matching with diacritic normalization (`acai` ↔ `açaí`). ### Tests - New `scripts/test-carbon.mjs` — covers the carbon dataset sanity, exact + diacritic + token lookups, idempotent enrichment, mealcarbon math + breakdown + unmatched tracking, swap suggestions for high-vs low-carbon meals, and TACO provider end-to-end (Portuguese alias, diacritic-stripped query, English alias). - `scripts/smoke-tools.mjs` adds 2 MCP-level assertions: `assertTacoProviderReturnsBrazilianFoods` and `assertCarbonSummaryWorks` (beef-heavy + vegetarian + mixed-with- unmatched cases). ### Discovery - Tool count: **35 → 36** (`nourish_carbon_summary`). - Provider enum on `nourish_search_food`: `["usda", "open_food_facts", "br_local", "taco", "all"]` — TACO is also fanned out under `provider: "all"`. - Agent-manifest TOOLS list updated. ### Notes - This release ships **curated subsets** (~60 + ~60 entries) of TACO and carbon data. The full Agribalyse + SU-EATABLE LIFE ingest (3,484 + 3,349 entries) is queued for the next data PR. The curated subset covers the foods most commonly logged via the existing estimator + USDA staples, so most agent calls already get useful coverage. - TACO license posture: UNICAMP/NEPA copyright with no published open-data license. Per the dataset-research audit, the community has redistributed this table for 10+ years with attribution. v1 ships with prominent attribution; a UNICAMP outreach letter is queued before any 1.0 release. ## [0.2.5] - 2026-05-08 Sprint 3 — data-integrity (timezone) + first agent UX win (`undo_last`). Lays groundwork for the upcoming dataset-enrichment work (TACO, CIQUAL, Agribalyse carbon footprint). ### Fixed - **A1 — Timezone-aware date bucketing.** Three independently-defined `todayDate()` helpers in `summary.ts` / `coach.ts` / `hydration-store.ts` all returned `new Date().toISOString().slice(0, 10)` (UTC). A user in São Paulo at 22:30 BRT saw `nourish_daily_summary` return data for "tomorrow"; a user in Los Angeles lost ~7-8 hrs of late-evening logs from "today". Same bug in `dateToNoonTimestamp` (used `${date}T12:00:00.000Z` = early morning for users east of UTC). Replaced all four sites with a single `services/local-date.ts` helper: - `localDate(tz?)` — today's YYYY-MM-DD in active timezone - `localDateFor(date, tz?)` — bucket a Date into local YYYY-MM-DD - `dateToNoonTimestamp(date, tz?)` — UTC ISO of noon-LOCAL on that date - `getActiveTimezone()` — resolved IANA tz Resolution order: `NOURISH_TIMEZONE` env var (e.g. `America/Sao_Paulo`), then `Intl.DateTimeFormat().resolvedOptions().timeZone` (system tz), then fallback to `UTC`. Also fixes B4 (intake-store and hydration-store no longer derive `entry.date` via `timestamp.slice(0, 10)`). ### Added - **`nourish_undo_last`** (D1 from QA backlog) — undo the most recent intake or hydration entry. The most common Telegram/agent recovery move ("I logged the wrong thing"). Schema: ```ts { kind?: "intake" | "hydration" | "any" = "any", explicit_user_intent: true } ``` Returns `{ ok: true, deleted, undone: { kind, entry } }` so the agent can confirm or re-log if the undo was a mistake. Requires explicit user intent (same guard pattern as `delete_intake` / `clear_day`). - **`timezone` field** in `nourish_connection_status` — exposes the active IANA timezone so agents can sanity-check date bucketing. ### Tests - New `scripts/test-local-date.mjs` — 8 cases covering São Paulo / Tokyo / UTC bucketing, noon-local conversion (BRT, JST, UTC), `dateToNoonTimestamp(undefined)` and garbage input passthrough. - `scripts/smoke-tools.mjs` adds 2 MCP-level assertions: `assertConnectionStatusExposesTimezone` and `assertUndoLastWorks` (which covers the explicit-intent guard, the empty-store no-op, intake-then-undo, and `kind: "any"` cross-store ordering). - Existing tests pinned to `NOURISH_TIMEZONE=UTC` so date-bucket assertions stay timezone-independent on contributor machines (was the actual cause of the only test break this sprint). ### Notes - Tool count: **34 → 35** (`nourish_undo_last`). - `connection_status` adds optional `timezone` field; agents reading the old shape keep working. - No upstream provider changes in this release — pure local logic + one new tool. ## [0.2.4] - 2026-05-08 Sprint 2 / resilience release. Layers on top of 0.2.3 with: a centralized HTTP helper that adds per-attempt timeout + single-retry-with-backoff to all USDA + OFF calls, graceful degradation when Open Food Facts is down, image analysis fallback when a barcode lookup fails mid-call, and several new QA findings (E1, B1, B5) caught by a follow-up audit. No breaking changes. ### Added - **`services/http.ts`** — single fetch wrapper with `AbortController`-based timeout (default 10s, configurable via `NOURISH_PROVIDER_TIMEOUT_MS`), one retry on transient failures (`5xx`, `429`, network, timeout) with exponential backoff, and a structured `NourishHttpError` exposing `kind` (`timeout` | `rate_limit` | `server_error` | `network` | `http`), `attempts`, and a `transient` flag. `isTransientHttpError(err)` is the recommended predicate for callers deciding between hard failure and graceful degradation. Regression tests in `scripts/test-http-helper.mjs` (timeout, 503-then-200 retry, persistent 503, persistent 429, non-retryable 404). - **`provider_timeout_ms`** in `nourish_connection_status` so agents can see the configured timeout, plus a new top-level `warnings: string[]` array and a `usda_using_demo_key: boolean` flag (true when no `FDC_API_KEY` is set and we're not in fixture mode — the heavily rate-limited shared key many users hit silently). - Agent-manifest and usage-guide now list `nourish_delete_water` and `nourish_clear_hydration_day` (added in 0.2.3 but not surfaced in the discovery surfaces, so agents booting cold couldn't find them — fixed here as the QA-found "E1" gap). ### Fixed - **N-007 — Open Food Facts search now degrades gracefully on transient failures.** A 503/timeout/network error from `/cgi/search.pl` no longer throws. `searchOpenFoodFactsByName` returns `{ foods: [], warnings: ["Open Food Facts search …; returning empty results."] }`, so partial results from other providers still flow through in `provider: "all"` and the agent gets actionable information instead of a hard error. Direct `nourish_lookup_barcode` keeps throwing because the caller asked for a specific product. - **N-008 — image-analysis routes now fall back when the barcode lookup fails mid-call.** If `nourish_analyze_food_image` is given a barcode AND `nutrition_label_text` (or `detected_items`/`image_description`) and OFF is unavailable, the call now uses the label/meal hints instead of failing. The barcode that was tried surfaces as `barcode_attempted` in the response so the agent can explain the fallback. - **N-009 — nutrition-label OCR with no parseable nutrients no longer emits `suggested_log_intake`.** Before, an unparseable label created an empty `nutrients_per_serving` and the agent could log it (writing `nutrients: {}` — the same class of bug 0.2.3's N-001 just fixed). Now the route is `"needs_more_ocr"` with explicit warnings and no logging suggestion. - **B5 — HTTP transport `close()` errors are no longer unhandled rejections.** `transport.close()` and `server.close()` rejections (which fire when the response was already cancelled) are now swallowed with a comment explaining why. - **`provider: "all"` fan-out is now parallel** (`Promise.allSettled` instead of sequential `for` loop). Every per-provider warning is tagged with the provider name (e.g. `open_food_facts: …`) so partial failures are attributable. Latency drops from ~3× single-provider to ~1× the slowest provider. - **B1 — Open Food Facts cache is now LRU-bounded.** Was a `Map` that grew forever (≈400MB at 100K barcodes in a long-running stdio process). Capped at 500 entries with insertion-order LRU eviction. ### New regression coverage - `scripts/test-http-helper.mjs` — 5 cases for the new helper: timeout, 503-then-success retry, persistent 503, persistent 429, non-retryable 404. - `scripts/smoke-tools.mjs` adds: - `assertConnectionStatusEnriched` — `provider_timeout_ms`, `warnings[]`, `usda_using_demo_key` - `assertAgentManifestIncludesHydrationTools` — E1 fix - `assertLabelOcrWithoutNutrientsDoesNotSuggestLog` — N-009 - `assertParallelAllProviderTagsWarnings` — provider-tagged warnings ### Notes - No changes to the public tool surface (still 34 tools at 0.2.3 + 0.2.4). - No breaking schema changes. `connection_status` adds optional fields; agents reading the old shape continue to work. - Hermes wrapper update (`wellness-nourish@0.2.3 → @0.2.4`) lands separately after this PR is merged + published to npm. ## [0.2.3] - 2026-05-08 This release lands every P1 finding from the deep QA report: the intake input pipeline now respects explicit nutrient/custom-food data over text estimates, custom foods scale correctly by `grams_estimate`, hydration finally has a delete/clear path, and the meal estimator correctly handles pt-BR decimal commas and explicit zero/negative quantities. Each fix ships with a regression test (estimator unit tests + MCP-level smoke assertions) so the same bug cannot reappear silently. ### Fixed - **Intake — explicit data wins over text estimate (N-001).** When `nourish_log_intake` was called with both `text` and explicit `nutrients`/`custom_food`/`food_ref`/`food`, the explicit data was silently dropped and the entry was logged with whatever the text estimator inferred (often `nutrients: {}` for OCR labels). Explicit nutrient data now always wins; the text becomes the food label or note. Regression covered in `scripts/smoke-tools.mjs` (`assertExplicitNutrientsBeatText`). - **Intake — `custom_food.nutrients_per_100g` now scales by `grams_estimate` (N-002).** A 60 g portion of a custom food previously logged the full 100 g nutrient values. The new `nutrientsFromCustomShape` / `gramsForCustomShape` helpers compute the correct gram weight (from `grams_estimate`, then `serving.grams × quantity`, then `available_portions[0].grams × quantity`) and scale via the existing `nutrientsForGrams`. Regression covered in `assertCustomFoodScalesByGrams`. - **Estimator — pt-BR decimal comma is parsed correctly (N-004).** `1,5 banana` was being split into two clauses (`1` and `5 banana`) and logged as five bananas. The clause splitter now uses negative lookarounds (`(?