.scratch/ holds throwaway diagnostic scripts and captured fixtures from
debugging sessions against the live instance --- including a 200 kB screenshot.
None of it belongs in history, but it sat untracked in every `git status`,
one `git add -A` away from being committed by accident.
The files are left on disk; only the path is ignored.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The MCP server runs straight from this working tree, so whatever .venv holds
is what the import agent executes. uv.lock was untracked, which meant a
rebuilt .venv could resolve different versions than the ones the test suite
was last green against, with nothing in the repo recording the difference.
The lock file is unchanged by this commit --- `uv lock --check` resolves it
clean against pyproject.toml. It is only now under version control.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Recipes imported from a video caption have no nutrition data, so the values
have to be estimated. Mealie will accept anything: the fields are free-text
strings, a misspelt key is dropped silently, and a whole-recipe total looks
exactly like a per-serving one. The recipe page then renders whatever landed
as fact.
patch_recipe now validates a nutrition block first. Keys must be Mealie's own,
so a number cannot vanish into "carbs". Values are normalized to bare numbers,
matching how the library already stores them. Energy is checked against the
macros with the Atwater factors (4/9/4 kcal per gram) and refused if it is
more than 25% off, which is what catches an arithmetic slip.
Nutrition is per serving, and the text import left recipeServings at 0 --- it
set only the free-text recipeYield, so "4-6 personer" gave Mealie no number to
divide by or scale with. import_recipe_text now also sets recipeServings, from
the lower bound of a range, which is how this library already stores
"10-12 personer" (recipeServings 10). patch_recipe refuses nutrition while the
count is still missing, and accepts it when the same patch supplies it.
verify_recipe fails an import that has no nutrition, or has values without a
serving count.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two failures from today's TikTok import, both misread as faults on the
Mealie instance.
set_cover_image passed slug_or_id straight into /api/recipes/{slug}/image.
That path only resolves slugs: /api/recipes/{slug} itself accepts an id, the
routes beneath it do not. With the recipe's id the scrape POST answered 404
and the file upload PUT answered 500 — neither of which says "wrong key", so
the agent concluded the image endpoint was broken and gave up on the cover.
Verified on this instance (v3.22.0): GET /api/recipes/<id>/comments answers
500 while the same call with the slug answers 200. The recipe is now read
first and its stored slug used for the sub-route.
patch_recipe sent tags through untouched, so {"name": "Källa: TikTok"} went
out as-is and came back 422. RecipeTag and RecipeCategory both require name
*and* slug per the instance's OpenAPI schema. Tags and categories are now
resolved before the PATCH: an existing organizer is reused (case-insensitive
on name) so tagging cannot fork a vocabulary that already holds "Källa:
Instagram" and "Källa: YouTube", and a genuinely new name is created through
its own endpoint, which supplies the slug Mealie itself would pick.
Neither path had a test that could have caught this: the cover tests always
passed a slug, and the patch contract test asserted the 422-producing shape
as if it were correct.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
import_recipe_text posted plain Swedish text to
/api/recipes/create/html-or-json. That endpoint is a scraper entry point:
Mealie runs recipe_scrapers over whatever `data` holds and answers 400 for
anything without schema.org markup. A text block never had any, so the tool
returned 400 on every call it has ever made — including a minimal one-line
test recipe. The seven session logs of the importing agent contain no
successful text import; every recipe that landed came in through
import_recipe_url.
The existing test mocked _request away, so it asserted the payload shape and
never that Mealie would accept it. Green test, broken tool.
Create the recipe with POST /api/recipes and fill it with a PATCH instead.
Both are already exercised against this instance. Ingredients are stored as
readable `note` lines, which is what parse_ingredients expects to structure
afterwards.
Also accept a parenthetical on a section heading. `Gör så här (REKONSTRUERAD):`
marks where the steps came from, and the old exact-match check turned that
import away before it reached Mealie.
A PATCH that fails now leaves a name-only recipe, so the error names the slug
rather than leaving a silent empty stub.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Buzz's read-only Mealie agent cannot import recipes, and the full server
exposes delete_recipe and the Home Assistant shopping-list write to any
client that runs it. Neither profile fits an on-demand import agent.
Add IMPORT_TOOL_NAMES — the read-only allowlist plus import_recipe_url,
import_recipe_text, import_recipe_image, patch_recipe, parse_ingredients
and set_cover_image — and expose it through main_import(). Adding a recipe
is recoverable from the Mealie UI; removing one is not, so delete_recipe
stays out of the profile even though it writes.
Verified over stdio: the entrypoint lists exactly those 15 tools.
Co-authored-by: fredamn76 <fredrik.fallman@gmail.com>
Signed-off-by: fredamn76 <fredrik.fallman@gmail.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Remove every tool that is not explicitly approved for the private Buzz agent. This keeps future mutating tools closed by default instead of relying on a denylist that must be updated manually.
Co-authored-by: fredamn76 <fredrik.fallman@gmail.com>
Signed-off-by: fredamn76 <fredrik.fallman@gmail.com>
Expose the existing stdio server to a dedicated Buzz agent without granting mutation tools or leaking unrelated runtime credentials. Document the verified Hermes and Buzz client paths.
Co-authored-by: fredamn76 <fredrik.fallman@gmail.com>
Signed-off-by: fredamn76 <fredrik.fallman@gmail.com>
Mealie's POST /api/recipes/{slug}/image downloads the URL it is given and
rejects anything that is not an image, so handing it a recipe page answered
400 {"message": "Url is not an image"}. This was read as a source-specific
failure on ica.se, but that page's own og:image uploads fine (verified live,
200 + 74608 bytes of image/webp), and httpbin's JPEG works too — the endpoint
simply never accepted pages.
Callers naturally pass the recipe page they just imported from, so resolve a
page to its advertised og:image and retry rather than making every caller know
the distinction. A page with no discoverable image now raises instead of
leaving the recipe silently coverless.
Co-authored-by: fredamn76 <fredrik.fallman@gmail.com>
Signed-off-by: fredamn76 <fredrik.fallman@gmail.com>
suggest_recipes returned 422 for every call. The endpoint takes food UUIDs as
repeated query parameters, not a comma-joined string of names — the rule
inherited from the skill was never actually verified. Added resolve_foods to
map names to food records first, and report anything unresolved instead of
dropping it.
scale_ingredients silently returned unscaled lines. It read the structured
quantity field and string-replaced str(quantity) into the display text, which
failed twice over: unparsed recipes store quantity 0.0 with the real amount only
in the text, and str(800.0) never matches "800 g kycklinglårfilé". Amounts are
now read from the display text, ranges scale at both ends ("1-1,5 msk"), mixed
numbers scale as one value ("1 1/2 msk"), and lines with no amount are reported
in `unscaled` rather than presented as if they had been scaled.
looks_mangled flagged the readable "1 1/2 msk tomatpuré" because its duplicate-word
rule matched the repeated digit. Restricted to alphabetic words, so it still
catches "2 dl dl grädde".
Verified against all 238 recipes in the live library: no false positives remain.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Ports a working Hermes agent skill to an MCP server. The skill's value was not
its Mealie endpoints but ~25 operational rules found by running imports against
the live instance; those are now code with tests rather than prompt text.
The server owns deterministic mechanics — auth, the non-default User-Agent
Cloudflare requires, the `extension` field on cover uploads, PATCH instead of
the 500-ing bulk-action endpoints, decimal-comma normalization, and restoring
Swedish display text after parsing. Language and taste judgement stay with the
model, fed by the four reference notes shipped as MCP resources.
verify_recipe runs the finished-import definition as code so an agent cannot
report success on a recipe with an English ingredient line, a missing cover, or
ingredients still in the "Click Parse" state.
Home Assistant is optional and isolated; the Obsidian meal plan is out of scope.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
fredamn76
Fizz
Honey